diff --git a/.github/workflows/test.yml b/.github/workflows/test.yml
index 6afbf0a068..6261859663 100644
--- a/.github/workflows/test.yml
+++ b/.github/workflows/test.yml
@@ -704,6 +704,8 @@ jobs:
       # TODO: run all once emulator is faster
       - name: Run RDNA3 ops tests
         run: SKIP_SLOW_TEST=1 AMD_LLVM=0 pytest -n=auto test/test_ops.py -k "test_sparse_categorical_crossentropy or test_tril or test_nonzero or test_softmax_argmax" --durations 20
+      - name: Run RDNA4 emulator tests
+        run: MOCKGPU_ARCH=rdna4 python -m pytest test/test_tiny.py -v --durations 20
 
   testnvidia:
     strategy:
diff --git a/examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_run.sh b/examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_run.sh
index 87e303851e..756774b7b0 100755
--- a/examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_run.sh
+++ b/examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_run.sh
@@ -10,6 +10,7 @@ export DEBUG=${DEBUG:-0}
 export FLASH_ATTENTION=${FLASH_ATTENTION:-1}
 export ALL2ALL=${ALL2ALL:-1}
 export USE_ATOMICS=${USE_ATOMICS:-1}
+export ASM_GEMM=${ASM_GEMM:-1}
 
 export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
 export DP=${DP:-8} BS=${BS:-8} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
diff --git a/extra/assembly/amd/emu.py b/extra/assembly/amd/emu.py
index b7aa412764..f7e8caac43 100644
--- a/extra/assembly/amd/emu.py
+++ b/extra/assembly/amd/emu.py
@@ -49,10 +49,11 @@ from tinygrad.helpers import Context, DEBUG, colored
 from tinygrad.engine.realize import get_runner
 
 from extra.assembly.amd import decode_inst
-from extra.assembly.amd.autogen.rdna3.str_pcode import PCODE
-from extra.assembly.amd.autogen.rdna3.ins import (SOP1, SOP2, SOPC, SOPK, SOPP, SMEM, VOP1, VOP1_SDST, VOP2, VOP3, VOP3_SDST, VOP3SD, VOP3P, VOPC,
-  DS, FLAT, GLOBAL, SCRATCH, VOPD, SOPPOp, SMEMOp, VOP1Op, VOP2Op, VOP3Op, VOPDOp)
-from extra.assembly.amd.dsl import VCC_LO, EXEC_LO, SCC
+from extra.assembly.amd.autogen.rdna3.str_pcode import PCODE as PCODE_RDNA3
+from extra.assembly.amd.autogen.rdna4.str_pcode import PCODE as PCODE_RDNA4
+from extra.assembly.amd.autogen.rdna3 import ins as ir3
+from extra.assembly.amd.autogen.rdna4 import ins as ir4
+from extra.assembly.amd.dsl import VCC_LO, EXEC_LO, SCC, ttmp
 from extra.assembly.amd.autogen.common import Fmt, OpType
 from extra.assembly.amd.pcode import parse_block, _FUNCS
 
@@ -79,15 +80,23 @@ def _apply_src_mods(val: UOp, mod_bit: int, abs_bits: int, neg_bits: int, bits:
   if neg_bits & (1 << mod_bit): fv = fv.neg()
   return fv.bitcast(ut).cast(dtypes.uint32) if bits == 16 else fv.bitcast(ut)
 
-# Map VOPD ops to VOP2 ops for pcode lookup
+# Map VOPD ops to VOP2 ops for pcode lookup (both RDNA3 and RDNA4)
 VOPD_TO_VOP2 = {
-  VOPDOp.V_DUAL_FMAC_F32: VOP2Op.V_FMAC_F32_E32, VOPDOp.V_DUAL_MUL_F32: VOP2Op.V_MUL_F32_E32,
-  VOPDOp.V_DUAL_ADD_F32: VOP2Op.V_ADD_F32_E32, VOPDOp.V_DUAL_SUB_F32: VOP2Op.V_SUB_F32_E32,
-  VOPDOp.V_DUAL_SUBREV_F32: VOP2Op.V_SUBREV_F32_E32, VOPDOp.V_DUAL_MAX_F32: VOP2Op.V_MAX_F32_E32,
-  VOPDOp.V_DUAL_MIN_F32: VOP2Op.V_MIN_F32_E32, VOPDOp.V_DUAL_ADD_NC_U32: VOP2Op.V_ADD_NC_U32_E32,
-  VOPDOp.V_DUAL_LSHLREV_B32: VOP2Op.V_LSHLREV_B32_E32, VOPDOp.V_DUAL_AND_B32: VOP2Op.V_AND_B32_E32,
-  VOPDOp.V_DUAL_MOV_B32: VOP1Op.V_MOV_B32_E32, VOPDOp.V_DUAL_CNDMASK_B32: VOP2Op.V_CNDMASK_B32_E32,
-  VOPDOp.V_DUAL_FMAAK_F32: VOP2Op.V_FMAAK_F32_E32, VOPDOp.V_DUAL_FMAMK_F32: VOP2Op.V_FMAMK_F32_E32,
+  ir3.VOPDOp.V_DUAL_FMAC_F32: ir3.VOP2Op.V_FMAC_F32_E32, ir3.VOPDOp.V_DUAL_MUL_F32: ir3.VOP2Op.V_MUL_F32_E32,
+  ir3.VOPDOp.V_DUAL_ADD_F32: ir3.VOP2Op.V_ADD_F32_E32, ir3.VOPDOp.V_DUAL_SUB_F32: ir3.VOP2Op.V_SUB_F32_E32,
+  ir3.VOPDOp.V_DUAL_SUBREV_F32: ir3.VOP2Op.V_SUBREV_F32_E32, ir3.VOPDOp.V_DUAL_MAX_F32: ir3.VOP2Op.V_MAX_F32_E32,
+  ir3.VOPDOp.V_DUAL_MIN_F32: ir3.VOP2Op.V_MIN_F32_E32, ir3.VOPDOp.V_DUAL_ADD_NC_U32: ir3.VOP2Op.V_ADD_NC_U32_E32,
+  ir3.VOPDOp.V_DUAL_LSHLREV_B32: ir3.VOP2Op.V_LSHLREV_B32_E32, ir3.VOPDOp.V_DUAL_AND_B32: ir3.VOP2Op.V_AND_B32_E32,
+  ir3.VOPDOp.V_DUAL_MOV_B32: ir3.VOP1Op.V_MOV_B32_E32, ir3.VOPDOp.V_DUAL_CNDMASK_B32: ir3.VOP2Op.V_CNDMASK_B32_E32,
+  ir3.VOPDOp.V_DUAL_FMAAK_F32: ir3.VOP2Op.V_FMAAK_F32_E32, ir3.VOPDOp.V_DUAL_FMAMK_F32: ir3.VOP2Op.V_FMAMK_F32_E32,
+  # RDNA4 mappings (same VOP1/VOP2 targets, RDNA4 uses _NUM_ suffix for min/max)
+  ir4.VOPDOp.V_DUAL_FMAC_F32: ir3.VOP2Op.V_FMAC_F32_E32, ir4.VOPDOp.V_DUAL_MUL_F32: ir3.VOP2Op.V_MUL_F32_E32,
+  ir4.VOPDOp.V_DUAL_ADD_F32: ir3.VOP2Op.V_ADD_F32_E32, ir4.VOPDOp.V_DUAL_SUB_F32: ir3.VOP2Op.V_SUB_F32_E32,
+  ir4.VOPDOp.V_DUAL_SUBREV_F32: ir3.VOP2Op.V_SUBREV_F32_E32, ir4.VOPDOp.V_DUAL_MAX_NUM_F32: ir3.VOP2Op.V_MAX_F32_E32,
+  ir4.VOPDOp.V_DUAL_MIN_NUM_F32: ir3.VOP2Op.V_MIN_F32_E32, ir4.VOPDOp.V_DUAL_ADD_NC_U32: ir3.VOP2Op.V_ADD_NC_U32_E32,
+  ir4.VOPDOp.V_DUAL_LSHLREV_B32: ir3.VOP2Op.V_LSHLREV_B32_E32, ir4.VOPDOp.V_DUAL_AND_B32: ir3.VOP2Op.V_AND_B32_E32,
+  ir4.VOPDOp.V_DUAL_MOV_B32: ir3.VOP1Op.V_MOV_B32_E32, ir4.VOPDOp.V_DUAL_CNDMASK_B32: ir3.VOP2Op.V_CNDMASK_B32_E32,
+  ir4.VOPDOp.V_DUAL_FMAAK_F32: ir3.VOP2Op.V_FMAAK_F32_E32, ir4.VOPDOp.V_DUAL_FMAMK_F32: ir3.VOP2Op.V_FMAMK_F32_E32,
 }
 WAVE_SIZE = 32
 # Special registers stored after inline constants (256-259)
@@ -146,11 +155,15 @@ _pcode_fixes = {
   'V_TRIG_PREOP_F64': ("result = 64'F((1201'B(2.0 / PI)[1200 : 0] << shift.u32) & 1201'0x1fffffffffffff)", "result = trig_preop_result(shift)"),
 }
 
+def _get_pcode_dict(op) -> dict:
+  """Return the PCODE dictionary for the given opcode based on its architecture."""
+  return PCODE_RDNA4 if 'rdna4' in type(op).__module__ else PCODE_RDNA3
+
 # Pcode parser
 @functools.cache
 def get_pcode(op) -> str:
   op_name = op.name
-  pcode = PCODE[op]
+  pcode = _get_pcode_dict(op)[op]
   if op_name in _pcode_fixes: pcode = pcode.replace(*_pcode_fixes[op_name])
   if 'V_DIV_SCALE' in op_name:
     dt, exp_lim, ldexp_val = ('f32', '23', '64') if 'F32' in op_name else ('f64', '52', '128')
@@ -174,7 +187,12 @@ def get_pcode(op) -> str:
 def parse_pcode(pcode: str, srcs: dict[str, UOp] | None = None) -> tuple[dict, list[tuple[str, UOp]]]:
   vars: dict = srcs.copy() if srcs else {}
   assigns: list[tuple[str, UOp]] = []
-  lines = [l.strip().rstrip(';') for l in pcode.split('\n') if l.strip() and not l.strip().startswith('//')]
+  raw_lines = [l.strip().rstrip(';') for l in pcode.split('\n') if l.strip() and not l.strip().startswith('//')]
+  # TODO: pcode.py should tokenize full pcode string instead of line-by-line, then this hack can be removed
+  lines: list[str] = []
+  for l in raw_lines:
+    if lines and lines[-1].endswith('&&'): lines[-1] = lines[-1] + ' ' + l
+    else: lines.append(l)
   _, final, _ = parse_block(lines, 0, vars, assigns=assigns)
   sliced = set(d.split('[')[0] for d, _ in assigns if '[' in d)
   for var, val in final.items():
@@ -317,9 +335,9 @@ class _Ctx:
     return base, mask, size
 
   # Dynamic register access (takes UOp index instead of int)
-  def rsgpr_dyn(self, reg: UOp) -> UOp:
+  def rsgpr_dyn(self, reg: UOp, valid: UOp | None = None) -> UOp:
     """Read SGPR with dynamic register index."""
-    return self.sgpr.index(reg.cast(dtypes.int), ptr=True).load()
+    return self.sgpr.index(reg.cast(dtypes.int), valid, ptr=True).load() if valid is not None else self.sgpr.index(reg.cast(dtypes.int), ptr=True).load()
 
   def wsgpr_dyn(self, reg: UOp, val: UOp) -> UOp:
     """Write SGPR with dynamic register index. Writes to NULL (124) are discarded."""
@@ -341,15 +359,18 @@ class _Ctx:
     If lane is None, only scalar access is supported (off must be < 256).
     is_f64: True for F64 operations where 64-bit literals go in high 32 bits."""
     is_float_const = (off >= _c(240)) & (off <= _c(248))
-    sgpr_lo = self.rsgpr_dyn(off)
+    is_vgpr = off >= _c(256)
+    is_sgpr = is_vgpr.ne(True)
+    sgpr_lo = self.rsgpr_dyn(off, is_sgpr)
 
     if lane is not None:
-      is_vgpr, vgpr_reg = off >= _c(256), off - _c(256)
+      vgpr_reg = off - _c(256)
       vgpr_lo = self.rvgpr_dyn(vgpr_reg, lane, is_vgpr)
       vgpr_val = _u64(vgpr_lo, self.rvgpr_dyn(vgpr_reg + _c(1), lane, is_vgpr)) if bits == 64 else vgpr_lo
 
     if bits == 64:
-      sgpr_val = _u64(sgpr_lo, self.rsgpr_dyn(off + _c(1)))
+      sgpr_hi = self.rsgpr_dyn(off + _c(1), is_sgpr)
+      sgpr_val = _u64(sgpr_lo, sgpr_hi)
       # Integer inline constants: sign-extend 32-bit value from buffer to 64-bit
       # Float constants: cast F32 to F64
       int_inline = sgpr_lo.cast(dtypes.int32).cast(dtypes.int64)
@@ -402,17 +423,19 @@ class _Ctx:
     return UOp.sink(*self.scalar_stores(assigns, sdst_reg, sdst_size), *self.inc_pc())
 
   def compile_lane_pcode(self, op, inst) -> UOp:
-    """Compile READLANE/READFIRSTLANE/WRITELANE using pcode parser."""
+    """Compile cross-lane ops (READLANE/WRITELANE/PERMLANE) using pcode parser."""
     pcode = get_pcode(op)
     op_name = op.name if hasattr(op, 'name') else str(op)
     src0_off, vdst_off = self.inst_field(type(inst).src0), self.inst_field(type(inst).vdst)
     src0_reg = (src0_off >= _c(256)).where(src0_off - _c(256), _c(0))  # VGPR index or 0
     src1_off = self.inst_field(type(inst).src1) if hasattr(type(inst), 'src1') else None
+    src2_off = self.inst_field(type(inst).src2) if hasattr(type(inst), 'src2') else None
     exec_lo = self.rsgpr_dyn(_c(EXEC_LO.offset))
     srcs = {
       'SRC0': src0_reg, 'VDST': vdst_off, 'EXEC_LO': exec_lo, 'EXEC': exec_lo.cast(dtypes.uint64), '_vgpr': self.vgpr,
       'S0': self.rsrc_dyn(src0_off, _c(0, dtypes.int)) if 'WRITELANE' in op_name else src0_reg,
       'S1': self.rsrc_dyn(src1_off, _c(0, dtypes.int)) if src1_off is not None else _c(0),
+      'S2': self.rsrc_dyn(src2_off, _c(0, dtypes.int)) if src2_off is not None else _c(0),
     }
     _, assigns = parse_pcode(pcode, srcs)
     stores = []
@@ -480,13 +503,14 @@ class _Ctx:
 # INSTRUCTION HANDLERS
 # ═══════════════════════════════════════════════════════════════════════════════
 
-def _compile_sopp(inst: SOPP, ctx: _Ctx) -> UOp:
-  simm16 = ctx.inst_field_signed(SOPP.simm16).cast(dtypes.int16)
-  if inst.op == SOPPOp.S_ENDPGM:
+def _compile_sopp(inst: ir3.SOPP | ir4.SOPP, ctx: _Ctx) -> UOp:
+  simm16 = ctx.inst_field_signed(type(inst).simm16).cast(dtypes.int16)
+  if inst.op in (ir3.SOPPOp.S_ENDPGM, ir4.SOPPOp.S_ENDPGM):
     return UOp.sink(ctx.wsgpr_dyn(_c(PC_LO_IDX), UOp.const(dtypes.uint32, 0xFFFFFFFF)),
                           ctx.wsgpr_dyn(_c(PC_HI_IDX), UOp.const(dtypes.uint32, 0xFFFFFFFF)))
+  if inst.op in (ir3.SOPPOp.S_NOP, ir4.SOPPOp.S_NOP): return UOp.sink(*ctx.inc_pc())  # S_NOP is a no-op
   # NOTE: we ignore SOPPs without PCODE
-  if inst.op in PCODE:
+  if inst.op in _get_pcode_dict(inst.op):
     pcode = get_pcode(inst.op)
     pc_bytes = ctx.rpc()  # PC is already 64-bit byte address
     vcc, exec_lo = ctx.rsgpr_dyn(_c(VCC_LO.offset)), ctx.rsgpr_dyn(_c(EXEC_LO.offset))
@@ -498,50 +522,57 @@ def _compile_sopp(inst: SOPP, ctx: _Ctx) -> UOp:
         return UOp.sink(ctx.wsgpr_dyn(_c(PC_LO_IDX), lo), ctx.wsgpr_dyn(_c(PC_HI_IDX), hi))
   return UOp.sink(*ctx.inc_pc())
 
-def _compile_smem(inst: SMEM, ctx: _Ctx) -> UOp:
+def _compile_smem(inst: ir3.SMEM | ir4.SMEM, ctx: _Ctx) -> UOp:
   # Cache invalidation instructions are no-ops in the emulator (we don't model caches)
-  if inst.op in (SMEMOp.S_GL1_INV, SMEMOp.S_DCACHE_INV):
+  cache_inv_ops = [ir3.SMEMOp.S_GL1_INV, ir3.SMEMOp.S_DCACHE_INV, ir4.SMEMOp.S_DCACHE_INV]
+  if hasattr(ir4.SMEMOp, 'S_GL1_INV'): cache_inv_ops.append(ir4.SMEMOp.S_GL1_INV)
+  if inst.op in cache_inv_ops:
     return UOp.sink(*ctx.inc_pc())
   # Dynamic sbase field (bits 5:0) - SGPR pair, field value * 2 = register offset
-  sbase = ctx.inst_field(SMEM.sbase) * _c(2)
+  sbase = ctx.inst_field(type(inst).sbase) * _c(2)
   # Dynamic sdata field (bits 12:6) - destination SGPR
-  sdata_reg = ctx.inst_field(SMEM.sdata)
-  offset = ctx.inst_field_signed(SMEM.offset)  # 21-bit signed immediate
-  # Dynamic soffset field (bits 63:57) - SGPR for additional offset (NULL=124 reads as 0)
-  soffset = ctx.inst_field(SMEM.soffset)
+  sdata_reg = ctx.inst_field(type(inst).sdata)
+  # RDNA4 uses 'ioffset', RDNA3 uses 'offset' - use type(inst) to get correct field
+  offset_field = type(inst).ioffset if hasattr(type(inst), 'ioffset') else type(inst).offset
+  offset = ctx.inst_field_signed(offset_field)  # signed immediate
+  # Dynamic soffset field - SGPR for additional offset (NULL=124 reads as 0)
+  soffset = ctx.inst_field(type(inst).soffset)
   addr = _u64(ctx.rsgpr_dyn(sbase), ctx.rsgpr_dyn(sbase + _c(1))) + offset.cast(dtypes.uint64) + ctx.rsgpr_dyn(soffset).cast(dtypes.uint64)
-  ndwords = {SMEMOp.S_LOAD_B32: 1, SMEMOp.S_LOAD_B64: 2, SMEMOp.S_LOAD_B128: 4, SMEMOp.S_LOAD_B256: 8, SMEMOp.S_LOAD_B512: 16}.get(inst.op, 1)
+  _SMEM_NDWORDS = {ir3.SMEMOp.S_LOAD_B32: 1, ir3.SMEMOp.S_LOAD_B64: 2, ir3.SMEMOp.S_LOAD_B128: 4,
+    ir3.SMEMOp.S_LOAD_B256: 8, ir3.SMEMOp.S_LOAD_B512: 16, ir4.SMEMOp.S_LOAD_B32: 1, ir4.SMEMOp.S_LOAD_B64: 2,
+    ir4.SMEMOp.S_LOAD_B96: 3, ir4.SMEMOp.S_LOAD_B128: 4, ir4.SMEMOp.S_LOAD_B256: 8, ir4.SMEMOp.S_LOAD_B512: 16}
+  ndwords = _SMEM_NDWORDS[inst.op]
   stores = [ctx.wsgpr_dyn(sdata_reg + _c(i), ctx.vmem.index((addr + UOp.const(dtypes.uint64, i * 4) >> UOp.const(dtypes.uint64, 2)).cast(dtypes.int)))
             for i in range(ndwords)]
   return UOp.sink(*stores, *ctx.inc_pc())
 
-def _compile_sop(inst: SOP1 | SOP2 | SOPC | SOPK, ctx: _Ctx) -> UOp:
+def _compile_sop(inst: ir3.SOP1 | ir3.SOP2 | ir3.SOPC | ir3.SOPK | ir4.SOP1 | ir4.SOP2 | ir4.SOPC | ir4.SOPK, ctx: _Ctx) -> UOp:
   bits = inst.canonical_op_bits
   literal = ctx.inst_field(type(inst).literal) if hasattr(type(inst), 'literal') else None
 
-  if isinstance(inst, SOPK):
-    sdst_off = ctx.inst_field(SOPK.sdst)
-    simm16 = ctx.inst_field(SOPK.simm16)
+  if isinstance(inst, (ir3.SOPK, ir4.SOPK)):
+    sdst_off = ctx.inst_field(type(inst).sdst)
+    simm16 = ctx.inst_field(type(inst).simm16)
     # Sign-extend simm16
     simm16_sext = simm16.cast(dtypes.int16).cast(dtypes.int32)
     srcs = {'S0': ctx.rsgpr_dyn(sdst_off), 'SIMM16': simm16_sext, 'D0': ctx.rsgpr_dyn(sdst_off)}
     dst_off, dst_size = sdst_off, 1
-  elif isinstance(inst, SOP1):
-    sdst_off = ctx.inst_field(SOP1.sdst)
-    ssrc0_off = ctx.inst_field(SOP1.ssrc0)
+  elif isinstance(inst, (ir3.SOP1, ir4.SOP1)):
+    sdst_off = ctx.inst_field(type(inst).sdst)
+    ssrc0_off = ctx.inst_field(type(inst).ssrc0)
     srcs = {'S0': ctx.rsrc_dyn(ssrc0_off, None, bits['s0'], literal)}
     dst_off, dst_size = sdst_off, bits['d'] // 32
-  elif isinstance(inst, SOP2):
-    sdst_off = ctx.inst_field(SOP2.sdst)
-    ssrc0_off = ctx.inst_field(SOP2.ssrc0)
-    ssrc1_off = ctx.inst_field(SOP2.ssrc1)
+  elif isinstance(inst, (ir3.SOP2, ir4.SOP2)):
+    sdst_off = ctx.inst_field(type(inst).sdst)
+    ssrc0_off = ctx.inst_field(type(inst).ssrc0)
+    ssrc1_off = ctx.inst_field(type(inst).ssrc1)
     srcs = {'S0': ctx.rsrc_dyn(ssrc0_off, None, bits['s0'], literal),
             'S1': ctx.rsrc_dyn(ssrc1_off, None, bits['s1'], literal)}
     if literal is not None: srcs['SIMM32'] = literal
     dst_off, dst_size = sdst_off, bits['d'] // 32
-  elif isinstance(inst, SOPC):
-    ssrc0_off = ctx.inst_field(SOPC.ssrc0)
-    ssrc1_off = ctx.inst_field(SOPC.ssrc1)
+  elif isinstance(inst, (ir3.SOPC, ir4.SOPC)):
+    ssrc0_off = ctx.inst_field(type(inst).ssrc0)
+    ssrc1_off = ctx.inst_field(type(inst).ssrc1)
     srcs = {'S0': ctx.rsrc_dyn(ssrc0_off, None, bits['s0'], literal),
             'S1': ctx.rsrc_dyn(ssrc1_off, None, bits['s1'], literal)}
     dst_off, dst_size = _c(0), 0  # SOPC writes to SCC, not sdst
@@ -550,18 +581,18 @@ def _compile_sop(inst: SOP1 | SOP2 | SOPC | SOPK, ctx: _Ctx) -> UOp:
 
   return ctx.compile_sop_pcode(inst.op, srcs, dst_off, dst_size)
 
-def _compile_vop12(inst: VOP1 | VOP1_SDST | VOP2, ctx: _Ctx) -> UOp:
+def _compile_vop12(inst: ir3.VOP1 | ir3.VOP1_SDST | ir3.VOP2 | ir4.VOP1 | ir4.VOP1_SDST | ir4.VOP2, ctx: _Ctx) -> UOp:
   op_name = _op_name(inst)
-  if op_name == 'V_READFIRSTLANE_B32_E32': return ctx.compile_lane_pcode(inst.op, inst)
+  if op_name in ('V_READFIRSTLANE_B32_E32', 'V_PERMLANE64_B32_E32'): return ctx.compile_lane_pcode(inst.op, inst)
   lane, exec_mask, bits = ctx.range(), ctx.rsgpr_dyn(_c(EXEC_LO.offset)), inst.canonical_op_bits
   literal = ctx.inst_field(type(inst).literal) if hasattr(type(inst), 'literal') else None
-  vdst_reg = ctx.inst_field(VOP1.vdst)
+  vdst_reg = ctx.inst_field(type(inst).vdst)
   write_hi_half = bits['d'] == 16 and (vdst_reg >= _c(128))
   if isinstance(write_hi_half, UOp): vdst_reg = write_hi_half.where(vdst_reg - _c(128), vdst_reg)
   elif write_hi_half: vdst_reg -= 128
-  if isinstance(inst, VOP1):
+  if isinstance(inst, (ir3.VOP1, ir4.VOP1)):
     # Handle VOP1 hi-half source operand (src0 >= v[128] for 16-bit ops)
-    src0_off = ctx.inst_field(VOP1.src0)
+    src0_off = ctx.inst_field(type(inst).src0)
     s0 = ctx.rsrc_dyn(src0_off, lane, bits['s0'], literal)
     if bits['s0'] == 16:
       src0_hi = src0_off >= _c(384)
@@ -570,13 +601,13 @@ def _compile_vop12(inst: VOP1 | VOP1_SDST | VOP2, ctx: _Ctx) -> UOp:
       s0 = src0_hi.where(_hi16(ctx.rvgpr_dyn(src0_reg, lane)), s0)
     srcs = {'S0': s0}
   else:
-    vsrc1_reg = ctx.inst_field(VOP2.vsrc1)
+    vsrc1_reg = ctx.inst_field(type(inst).vsrc1)
     vsrc1_hi = bits['s0'] == 16 and (vsrc1_reg >= _c(128))
     vsrc1_actual = _cond(vsrc1_hi, vsrc1_reg - _c(128), vsrc1_reg)
     s1 = _cond_hi16(vsrc1_hi, ctx.rvgpr_dyn(vsrc1_actual, lane))
     d0 = _cond_hi16(write_hi_half, ctx.rvgpr_dyn(vdst_reg, lane))  # FMAC/FMAMK hi-half dest needs hi-half accumulator
     # Handle VOP2 hi-half src0 operand (src0 >= v[128] for 16-bit ops)
-    src0_off = ctx.inst_field(VOP2.src0)
+    src0_off = ctx.inst_field(type(inst).src0)
     s0 = ctx.rsrc_dyn(src0_off, lane, bits['s0'], literal)
     if bits['s0'] == 16:
       src0_hi = src0_off >= _c(384)
@@ -584,19 +615,20 @@ def _compile_vop12(inst: VOP1 | VOP1_SDST | VOP2, ctx: _Ctx) -> UOp:
       src0_reg = src0_hi.where(src0_off - _c(384), _c(0))
       s0 = src0_hi.where(_hi16(ctx.rvgpr_dyn(src0_reg, lane)), s0)
     srcs = {'S0': s0, 'S1': s1, 'D0': d0}
-    if inst.op in (VOP2Op.V_FMAAK_F32_E32, VOP2Op.V_FMAMK_F32_E32, VOP2Op.V_FMAAK_F16_E32, VOP2Op.V_FMAMK_F16_E32):
+    if inst.op in (ir3.VOP2Op.V_FMAAK_F32_E32, ir3.VOP2Op.V_FMAMK_F32_E32, ir3.VOP2Op.V_FMAAK_F16_E32,
+                   ir3.VOP2Op.V_FMAMK_F16_E32):
       assert literal is not None
       srcs['SIMM32'] = literal
   return ctx.compile_vop_pcode(inst.op, srcs, lane, vdst_reg, exec_mask, opsel_dst_hi=write_hi_half)
 
-def _compile_vopc(inst: VOPC | VOP3, ctx: _Ctx, opsel: int = 0, abs_bits: int = 0, neg_bits: int = 0) -> UOp:
+def _compile_vopc(inst: ir3.VOPC | ir3.VOP3 | ir4.VOPC | ir4.VOP3, ctx: _Ctx, opsel: int = 0, abs_bits: int = 0, neg_bits: int = 0) -> UOp:
   exec_mask, op_name, bits = ctx.rsgpr_dyn(_c(EXEC_LO.offset)), _op_name(inst), inst.canonical_op_bits
   is_cmpx, is_vopc = 'CMPX' in op_name, hasattr(inst, 'vsrc1')  # is_vopc: e32 vs e64
 
   # Handle both VOPC (vsrc1) and VOP3 (src1) instruction formats - read operands dynamically
   if is_vopc:
-    src0_off = ctx.inst_field(VOPC.src0)
-    vsrc1_off = ctx.inst_field(VOPC.vsrc1)
+    src0_off = ctx.inst_field(type(inst).src0)
+    vsrc1_off = ctx.inst_field(type(inst).vsrc1)
     # For 16-bit ops, vsrc1 >= 128 means hi-half of v[vsrc1-128]
     if bits['s0'] == 16:
       vsrc1_hi = vsrc1_off >= _c(128)
@@ -605,9 +637,9 @@ def _compile_vopc(inst: VOPC | VOP3, ctx: _Ctx, opsel: int = 0, abs_bits: int =
       vsrc1_hi = False
       src1_off = _c(256) + vsrc1_off
   else:
-    src0_off = ctx.inst_field(VOP3.src0)
-    src1_off = ctx.inst_field(VOP3.src1)
-    dst_off = ctx.inst_field(VOP3.vdst)
+    src0_off = ctx.inst_field(type(inst).src0)
+    src1_off = ctx.inst_field(type(inst).src1)
+    dst_off = ctx.inst_field(type(inst).vdst)
     vsrc1_hi = False
   literal = ctx.inst_field(type(inst).literal) if hasattr(type(inst), 'literal') else None
 
@@ -636,7 +668,7 @@ def _compile_vopc(inst: VOPC | VOP3, ctx: _Ctx, opsel: int = 0, abs_bits: int =
     stores = [ctx.wsgpr_dyn(dst_off, new_result)] if not is_vopc else [ctx.wsgpr_dyn(_c(VCC_LO.offset), new_result)]
   return UOp.sink(*stores, *ctx.inc_pc())
 
-def _compile_vop3(inst: VOP3, ctx: _Ctx) -> UOp:
+def _compile_vop3(inst: ir3.VOP3 | ir4.VOP3, ctx: _Ctx) -> UOp:
   exec_mask = ctx.rsgpr_dyn(_c(EXEC_LO.offset))
   bits = inst.canonical_op_bits
   opsel, op_name = getattr(inst, 'opsel', 0) or 0, _op_name(inst)
@@ -645,18 +677,22 @@ def _compile_vop3(inst: VOP3, ctx: _Ctx) -> UOp:
   if op_name in ('V_READLANE_B32', 'V_READFIRSTLANE_B32', 'V_READFIRSTLANE_B32_E64', 'V_WRITELANE_B32'):
     return ctx.compile_lane_pcode(inst.op, inst)
 
+  # V_PERMLANE16_B32 / V_PERMLANEX16_B32: cross-lane swizzle via pcode
+  if 'PERMLANE16' in op_name or 'PERMLANEX16' in op_name:
+    return ctx.compile_lane_pcode(inst.op, inst)
+
   # VOP3 VOPC (v_cmp_*_e64) - delegate to unified VOPC handler
   if 'V_CMP' in op_name or 'V_CMPX' in op_name:
     return _compile_vopc(inst, ctx, opsel=opsel, abs_bits=getattr(inst, 'abs', 0) or 0, neg_bits=getattr(inst, 'neg', 0) or 0)
 
   # Regular VOP3 - read operands dynamically
   lane = ctx.range()
-  vdst_reg = ctx.inst_field(VOP3.vdst)
+  vdst_reg = ctx.inst_field(type(inst).vdst)
   literal = ctx.inst_field(type(inst).literal) if hasattr(type(inst), 'literal') else None
   ops = inst.canonical_operands
-  src0 = ctx.rsrc_dyn(ctx.inst_field(VOP3.src0), lane, bits['s0'], literal, 's0' in ops and ops['s0'][0] == Fmt.FMT_NUM_F64)
-  src1 = ctx.rsrc_dyn(ctx.inst_field(VOP3.src1), lane, bits['s1'], literal, 's1' in ops and ops['s1'][0] == Fmt.FMT_NUM_F64)
-  src2 = ctx.rsrc_dyn(ctx.inst_field(VOP3.src2), lane, bits['s2'], literal, 's2' in ops and ops['s2'][0] == Fmt.FMT_NUM_F64)
+  src0 = ctx.rsrc_dyn(ctx.inst_field(type(inst).src0), lane, bits['s0'], literal, 's0' in ops and ops['s0'][0] == Fmt.FMT_NUM_F64)
+  src1 = ctx.rsrc_dyn(ctx.inst_field(type(inst).src1), lane, bits['s1'], literal, 's1' in ops and ops['s1'][0] == Fmt.FMT_NUM_F64)
+  src2 = ctx.rsrc_dyn(ctx.inst_field(type(inst).src2), lane, bits['s2'], literal, 's2' in ops and ops['s2'][0] == Fmt.FMT_NUM_F64)
   if bits['s0'] == 16:
     src0 = _apply_opsel(src0, 0, opsel)
     src1 = _apply_opsel(src1, 1, opsel)
@@ -666,19 +702,19 @@ def _compile_vop3(inst: VOP3, ctx: _Ctx) -> UOp:
   src1 = _apply_src_mods(src1, 1, abs_bits, neg_bits, bits['s1'])
   src2 = _apply_src_mods(src2, 2, abs_bits, neg_bits, bits['s2'])
   srcs = {'S0': src0, 'S1': src1, 'S2': src2}
-  if inst.op in (VOP3Op.V_CNDMASK_B32_E64, VOP3Op.V_CNDMASK_B16) and src2 is not None: srcs['VCC'] = src2
+  if inst.op in (ir3.VOP3Op.V_CNDMASK_B32_E64, ir3.VOP3Op.V_CNDMASK_B16) and src2 is not None: srcs['VCC'] = src2
   # FMAC instructions need D0 (accumulator) from destination register
   if 'FMAC' in op_name: srcs['D0'] = ctx.rvgpr_dyn(vdst_reg, lane)
   opsel_dst_hi = bool(opsel & 0b1000) and bits['d'] == 16
   return ctx.compile_vop_pcode(inst.op, srcs, lane, vdst_reg, exec_mask, opsel_dst_hi=opsel_dst_hi, clmp=getattr(inst, 'clmp', 0))
 
-def _compile_vop3sd(inst: VOP3SD, ctx: _Ctx) -> UOp:
+def _compile_vop3sd(inst: ir3.VOP3SD | ir4.VOP3SD, ctx: _Ctx) -> UOp:
   exec_mask = ctx.rsgpr_dyn(_c(EXEC_LO.offset))
   bits, pcode, ops = inst.canonical_op_bits, get_pcode(inst.op), inst.canonical_operands
 
   # Read operands dynamically from instruction encoding
-  vdst_reg, sdst_off = ctx.inst_field(VOP3SD.vdst), ctx.inst_field(VOP3SD.sdst)
-  src0_off, src1_off, src2_off = ctx.inst_field(VOP3SD.src0), ctx.inst_field(VOP3SD.src1), ctx.inst_field(VOP3SD.src2)
+  vdst_reg, sdst_off = ctx.inst_field(type(inst).vdst), ctx.inst_field(type(inst).sdst)
+  src0_off, src1_off, src2_off = ctx.inst_field(type(inst).src0), ctx.inst_field(type(inst).src1), ctx.inst_field(type(inst).src2)
   literal = ctx.inst_field(type(inst).literal) if hasattr(type(inst), 'literal') else None
 
   has_carry_in = 's2' in ops and ops['s2'][2] == OpType.OPR_SREG
@@ -724,13 +760,13 @@ def _compile_vop3sd(inst: VOP3SD, ctx: _Ctx) -> UOp:
   else:
     return ctx.compile_vop_pcode(inst.op, srcs, lane, vdst_reg, exec_mask, sdst_reg=inst.sdst.offset)
 
-def _compile_wmma(inst: VOP3P, ctx: _Ctx) -> UOp:
+def _compile_wmma(inst: ir3.VOP3P | ir4.VOP3P, ctx: _Ctx) -> UOp:
   op_name = _op_name(inst)
   exec_mask = ctx.rsgpr_dyn(_c(EXEC_LO.offset))
-  vdst_reg = ctx.inst_field(VOP3P.vdst)
-  src0_r = ctx.inst_field(VOP3P.src0) - _c(256)
-  src1_r = ctx.inst_field(VOP3P.src1) - _c(256)
-  src2_r = ctx.inst_field(VOP3P.src2) - _c(256)
+  vdst_reg = ctx.inst_field(type(inst).vdst)
+  src0_r = ctx.inst_field(type(inst).src0) - _c(256)
+  src1_r = ctx.inst_field(type(inst).src1) - _c(256)
+  src2_r = ctx.inst_field(type(inst).src2) - _c(256)
   is_f16_output = 'F16_16X16X16_F16' in op_name or 'BF16_16X16X16_BF16' in op_name  # F16/BF16 output vs F32 output
   is_bf16 = 'BF16' in op_name
   cvt = _FUNCS['bf16_to_f32'] if is_bf16 else _FUNCS['f16_to_f32']
@@ -757,16 +793,16 @@ def _compile_wmma(inst: VOP3P, ctx: _Ctx) -> UOp:
     stores = [ctx.wvgpr_dyn(vdst_reg + _c(i // 32), UOp.const(dtypes.int, i % 32), mat_d[i].bitcast(dtypes.uint32), exec_mask) for i in range(256)]
   return UOp.sink(*stores, *ctx.inc_pc())
 
-def _compile_vop3p(inst: VOP3P, ctx: _Ctx) -> UOp:
+def _compile_vop3p(inst: ir3.VOP3P | ir4.VOP3P, ctx: _Ctx) -> UOp:
   op_name = _op_name(inst)
   if 'WMMA' in op_name and ('16X16X16_F16' in op_name or '16X16X16_BF16' in op_name): return _compile_wmma(inst, ctx)
 
   lane = ctx.range()
   exec_mask = ctx.rsgpr_dyn(_c(EXEC_LO.offset))
-  vdst_reg = ctx.inst_field(VOP3P.vdst)
-  src0 = ctx.rsrc_dyn(ctx.inst_field(VOP3P.src0), lane, 16)
-  src1 = ctx.rsrc_dyn(ctx.inst_field(VOP3P.src1), lane, 16)
-  src2 = ctx.rsrc_dyn(ctx.inst_field(VOP3P.src2), lane, 16)
+  vdst_reg = ctx.inst_field(type(inst).vdst)
+  src0 = ctx.rsrc_dyn(ctx.inst_field(type(inst).src0), lane, 16)
+  src1 = ctx.rsrc_dyn(ctx.inst_field(type(inst).src1), lane, 16)
+  src2 = ctx.rsrc_dyn(ctx.inst_field(type(inst).src2), lane, 16)
   opsel, opsel_hi = getattr(inst, 'opsel', 0) or 0, getattr(inst, 'opsel_hi', 3) if getattr(inst, 'opsel_hi', 3) is not None else 3
   opsel_hi2 = getattr(inst, 'opsel_hi2', 1) if getattr(inst, 'opsel_hi2', 1) is not None else 1
   neg, neg_hi = getattr(inst, 'neg', 0) or 0, getattr(inst, 'neg_hi', 0) or 0
@@ -788,7 +824,7 @@ def _compile_vop3p(inst: VOP3P, ctx: _Ctx) -> UOp:
     s0_mod = apply_neg_mix(apply_abs(src0, 1, 1, 1), 1, 1, 1)
     s1_mod = apply_neg_mix(apply_abs(src1, 2, 2, 2), 2, 2, 2)
     s2_mod = apply_neg_mix(apply_abs(src2, 4, 4, 4), 4, 4, 4)
-    srcs = {'S0': s0_mod, 'S1': s1_mod, 'S2': s2_mod,
+    srcs = {'S@0': s0_mod, 'S@1': s1_mod, 'S@2': s2_mod,
             'OPSEL_HI': UOp.const(dtypes.uint32, combined_opsel_hi), 'OPSEL': UOp.const(dtypes.uint32, opsel)}
   else:
     def get_half_bits(val: UOp, use_hi: bool, apply_neg: bool = False) -> UOp:
@@ -806,18 +842,19 @@ def _compile_vop3p(inst: VOP3P, ctx: _Ctx) -> UOp:
     if is_dot_iu: srcs['NEG'] = UOp.const(dtypes.uint32, neg)
   return ctx.compile_vop_pcode(inst.op, srcs, lane, vdst_reg, exec_mask)
 
-def _compile_vopd(inst: VOPD, ctx: _Ctx) -> UOp:
+def _compile_vopd(inst: ir3.VOPD | ir4.VOPD, ctx: _Ctx) -> UOp:
   exec_mask = ctx.rsgpr_dyn(_c(EXEC_LO.offset))
-  # Read operands dynamically
-  vdstx_reg = ctx.inst_field(VOPD.vdstx)
+  # Read operands dynamically - use type(inst) to get correct field descriptors
+  inst_type = type(inst)
+  vdstx_reg = ctx.inst_field(inst_type.vdstx)
   # vdsty has complex encoding: actual = (raw << 1) | ((vdstx & 1) ^ 1)
-  vdsty_raw = ctx.inst_field(VOPD.vdsty)
+  vdsty_raw = ctx.inst_field(inst_type.vdsty)
   vdsty_reg = (vdsty_raw << _c(1)) | ((vdstx_reg & _c(1)) ^ _c(1))
-  srcx0_off = ctx.inst_field(VOPD.srcx0)
-  srcy0_off = ctx.inst_field(VOPD.srcy0)
-  vsrcx1_reg = ctx.inst_field(VOPD.vsrcx1)
-  vsrcy1_reg = ctx.inst_field(VOPD.vsrcy1)
-  literal = ctx.inst_field(type(inst).literal) if hasattr(type(inst), 'literal') else None
+  srcx0_off = ctx.inst_field(inst_type.srcx0)
+  srcy0_off = ctx.inst_field(inst_type.srcy0)
+  vsrcx1_reg = ctx.inst_field(inst_type.vsrcx1)
+  vsrcy1_reg = ctx.inst_field(inst_type.vsrcy1)
+  literal = ctx.inst_field(inst_type.literal) if hasattr(inst_type, 'literal') else None
 
   lane = ctx.range()
   srcy0, srcy1 = ctx.rsrc_dyn(srcy0_off, lane, literal=literal), ctx.rvgpr_dyn(vsrcy1_reg, lane)
@@ -828,49 +865,64 @@ def _compile_vopd(inst: VOPD, ctx: _Ctx) -> UOp:
     assert vop is not None, f"no VOP mapping for VOPD {label}: {op}"
     if label == 'Y': srcs = {'S0': srcy0, 'S1': srcy1, 'D0': ctx.rvgpr_dyn(vdst_reg, lane)}
     else: srcs = {'S0': ctx.rsrc_dyn(src0_off, lane, literal=literal), 'S1': ctx.rvgpr_dyn(vsrc1_reg, lane), 'D0': ctx.rvgpr_dyn(vdst_reg, lane)}
-    if op in (VOPDOp.V_DUAL_FMAAK_F32, VOPDOp.V_DUAL_FMAMK_F32):
+    if op in (ir3.VOPDOp.V_DUAL_FMAAK_F32, ir3.VOPDOp.V_DUAL_FMAMK_F32, ir4.VOPDOp.V_DUAL_FMAAK_F32, ir4.VOPDOp.V_DUAL_FMAMK_F32):
       assert literal is not None
       srcs['SIMM32'] = literal
-    if op == VOPDOp.V_DUAL_CNDMASK_B32: srcs['VCC'] = ctx.rsgpr_dyn(_c(VCC_LO.offset))
+    if op in (ir3.VOPDOp.V_DUAL_CNDMASK_B32, ir4.VOPDOp.V_DUAL_CNDMASK_B32): srcs['VCC'] = ctx.rsgpr_dyn(_c(VCC_LO.offset))
     pcode = get_pcode(vop)
     srcs.update({'VCC': ctx.rsgpr_dyn(_c(VCC_LO.offset)), 'EXEC': exec_mask, 'SCC': ctx.rsgpr_dyn(_c(SCC.offset)), 'laneId': lane})
     for dest, val in parse_pcode(pcode, srcs)[1]:
       if dest.startswith('D0'): all_stores.append(ctx.wvgpr_dyn(vdst_reg, lane, _val_to_u32(val), exec_mask, after=srcy1))
   return UOp.sink(UOp.group(*all_stores).end(lane), *ctx.inc_pc())
 
-def _compile_mem_op(inst: DS | FLAT | GLOBAL | SCRATCH, ctx: _Ctx) -> UOp:
+def _compile_mem_op(inst: ir3.DS | ir3.FLAT | ir3.GLOBAL | ir3.SCRATCH | ir4.DS | ir4.VFLAT | ir4.VGLOBAL | ir4.VSCRATCH, ctx: _Ctx) -> UOp:
   """Unified memory operation compiler for DS, FLAT, GLOBAL, SCRATCH."""
   exec_mask, op_name = ctx.rsgpr_dyn(_c(EXEC_LO.offset)), _op_name(inst)
   pcode = get_pcode(inst.op)
 
-  is_lds = isinstance(inst, DS)
-  is_scratch = isinstance(inst, SCRATCH)
+  is_lds = isinstance(inst, (ir3.DS, ir4.DS))
+  is_scratch = isinstance(inst, (ir3.SCRATCH, ir4.VSCRATCH))
   mem = ctx.lds if is_lds else ctx.scratch if is_scratch else ctx.vmem
   addr_shift = UOp.const(dtypes.uint32 if is_lds else dtypes.uint64, 2)
 
   # Extract register info - all dynamic for deduplication
   if is_lds:
-    addr_reg = ctx.inst_field(DS.addr)
-    vdata_reg = ctx.inst_field(DS.data0)
-    vdst_reg = ctx.inst_field(DS.vdst)
-    offset0 = ctx.inst_field(DS.offset0)
-    offset1 = ctx.inst_field(DS.offset1)
+    addr_reg = ctx.inst_field(type(inst).addr)
+    vdata_reg = ctx.inst_field(type(inst).data0)
+    vdst_reg = ctx.inst_field(type(inst).vdst)
+    offset0 = ctx.inst_field(type(inst).offset0)
+    offset1 = ctx.inst_field(type(inst).offset1)
     offset = offset0  # DS uses offset0 as primary offset
     saddr_reg = None
-  else:
+  elif isinstance(inst, (ir4.VGLOBAL, ir4.VSCRATCH, ir4.VFLAT)):  # RDNA4: vaddr, vsrc, ioffset
+    addr_reg = ctx.inst_field(type(inst).vaddr)
+    vdata_reg = ctx.inst_field(type(inst).vsrc)
+    vdst_reg = ctx.inst_field(type(inst).vdst)
+    offset = ctx.inst_field_signed(type(inst).ioffset)
+    offset0, offset1 = _c(0), _c(0)
+    saddr_reg = ctx.inst_field(type(inst).saddr) if hasattr(type(inst), 'saddr') else None
+  else:  # RDNA3: addr, data, offset
     addr_reg = ctx.inst_field(type(inst).addr)
     vdata_reg = ctx.inst_field(type(inst).data)
     vdst_reg = ctx.inst_field(type(inst).vdst)
     offset = ctx.inst_field_signed(type(inst).offset)
     offset0, offset1 = _c(0), _c(0)
-    # Dynamic saddr - read field, NULL (124) or >= 128 means no saddr
-    saddr_reg = ctx.inst_field(type(inst).saddr) if hasattr(inst, 'saddr') else None
+    saddr_reg = ctx.inst_field(type(inst).saddr) if hasattr(type(inst), 'saddr') else None
 
   # Data width from canonical_op_bits (32/64/96/128), default to 32 for untyped ops
   data_bits_mem = inst.canonical_op_bits.get('data', 32)
   is_atomic, glc = 'ATOMIC' in op_name, getattr(inst, 'glc', 0)
   has_data1 = is_lds and hasattr(inst, 'data1') and inst.data1 is not None
-  data1_reg = ctx.inst_field(DS.data1) if is_lds else _c(0)
+  data1_reg = ctx.inst_field(type(inst).data1) if is_lds else _c(0)
+
+  # DS_PERMUTE/DS_BPERMUTE: cross-lane VGPR access via pcode
+  if is_lds and 'PERMUTE' in op_name:
+    pcode = get_pcode(inst.op)
+    srcs = {'ADDR': addr_reg, 'DATA0': vdata_reg, 'VDST': vdst_reg, 'OFFSET': offset,
+            'EXEC': exec_mask.cast(dtypes.uint64), '_vgpr': ctx.vgpr}
+    _, assigns = parse_pcode(pcode, srcs)
+    stores = [ctx.vgpr.index(val[0].cast(dtypes.int)).store(val[1].cast(dtypes.uint32)) for dest, val in assigns if dest.startswith('VGPR[')]
+    return UOp.sink(*stores, *ctx.inc_pc())
 
   def make_addr(lane: UOp) -> UOp:
     if is_lds: return ctx.rvgpr_dyn(addr_reg, lane)
@@ -912,14 +964,26 @@ def _compile_mem_op(inst: DS | FLAT | GLOBAL | SCRATCH, ctx: _Ctx) -> UOp:
       else:
         data = {'DATA': _u64(ctx.rvgpr_dyn(vdata_reg, lane), ctx.rvgpr_dyn(vdata_reg + _c(1), lane)),
                 'DATA2': _u64(ctx.rvgpr_dyn(data1_reg, lane), ctx.rvgpr_dyn(data1_reg + _c(1), lane)) if has_data1 else UOp.const(dtypes.uint64, 0)}
-      return {'ADDR': addr, 'ADDR_BASE': addr, 'OFFSET': offset, 'OFFSET0': offset0, 'OFFSET1': offset1, '_lds': mem, 'laneId': lane, **data}
+      # RDNA3 uses ADDR/OFFSET, RDNA4 uses vgpr_a/offset (lowercase) + CalcDsAddr function
+      return {'ADDR': addr, 'ADDR_BASE': addr, 'OFFSET': offset, 'OFFSET0': offset0, 'OFFSET1': offset1, '_lds': mem, 'laneId': lane,
+              'vgpr_a': ctx.rvgpr_dyn(addr_reg, lane), 'offset': offset, **data}
     active = _lane_active(exec_mask, lane)
+    # saddr < 124 means valid SGPR pair, otherwise use 0 (NULL means no saddr contribution)
+    use_saddr = (saddr_reg < _c(124)) if saddr_reg is not None else UOp.const(dtypes.bool, False)
+    saddr_raw = _u64(ctx.rsgpr_dyn(saddr_reg), ctx.rsgpr_dyn(saddr_reg + _c(1))) if saddr_reg is not None else UOp.const(dtypes.uint64, 0)
+    saddr_base = use_saddr.where(saddr_raw, UOp.const(dtypes.uint64, 0))
+    # Sign-extend offset to 64-bit for the final address calculation
+    ioffset64 = offset.cast(dtypes.int64).cast(dtypes.uint64)
+    # v_addr for CalcGlobalAddr: when saddr valid, use low 32 bits as offset; otherwise full 64-bit address. Include ioffset.
+    vaddr_full = _u64(ctx.rvgpr_dyn(addr_reg, lane), ctx.rvgpr_dyn(addr_reg + _c(1), lane))
+    vaddr_lo = ctx.rvgpr_dyn(addr_reg, lane).cast(dtypes.uint64)
+    vaddr_base = use_saddr.where(vaddr_lo + ioffset64, vaddr_full + ioffset64)
     if is_atomic:
       return {'ADDR': addr, 'DATA': _u64(ctx.rvgpr_dyn(vdata_reg, lane), ctx.rvgpr_dyn(vdata_reg + _c(1), lane)) if data_bits_mem == 64 else ctx.rvgpr_dyn(vdata_reg, lane),
-              '_vmem': mem, '_active': active, 'laneId': lane}
+              '_vmem': mem, '_active': active, 'laneId': lane, 'v_addr': vaddr_base, 's_saddr': saddr_base}
     vdata = ctx.rvgpr_dyn(vdata_reg, lane).cast(dtypes.uint64) if 'STORE' in op_name else ctx.rvgpr_dyn(vdst_reg, lane) if 'D16' in op_name else UOp.const(dtypes.uint32, 0)
     if 'STORE' in op_name and data_bits_mem >= 64: vdata = vdata | (ctx.rvgpr_dyn(vdata_reg + _c(1), lane).cast(dtypes.uint64) << UOp.const(dtypes.uint64, 32))
-    srcs = {'ADDR': addr, 'VDATA': vdata, '_vmem': mem, '_active': active, 'laneId': lane}
+    srcs = {'ADDR': addr, 'VDATA': vdata, '_vmem': mem, '_active': active, 'laneId': lane, 'v_addr': vaddr_base, 's_saddr': saddr_base}
     for i in range(data_bits_mem // 32): srcs[f'VDATA{i}'] = ctx.rvgpr_dyn(vdata_reg + _c(i), lane) if 'STORE' in op_name else UOp.const(dtypes.uint32, 0)
     return srcs
 
@@ -970,10 +1034,15 @@ def _compile_mem_op(inst: DS | FLAT | GLOBAL | SCRATCH, ctx: _Ctx) -> UOp:
 
 # Dispatch table: instruction type -> handler function
 _INST_HANDLERS: dict[type, Callable[..., UOp]] = {
-  SOPP: _compile_sopp, SMEM: _compile_smem, SOP1: _compile_sop, SOP2: _compile_sop, SOPC: _compile_sop, SOPK: _compile_sop,
-  VOP1: _compile_vop12, VOP1_SDST: _compile_vop12, VOP2: _compile_vop12, VOPC: _compile_vopc, VOP3: _compile_vop3, VOP3_SDST: _compile_vop3,
-  VOP3SD: _compile_vop3sd, VOP3P: _compile_vop3p, VOPD: _compile_vopd,
-  DS: _compile_mem_op, FLAT: _compile_mem_op, GLOBAL: _compile_mem_op, SCRATCH: _compile_mem_op,
+  ir3.SOPP: _compile_sopp, ir3.SMEM: _compile_smem, ir3.SOP1: _compile_sop, ir3.SOP2: _compile_sop, ir3.SOPC: _compile_sop, ir3.SOPK: _compile_sop,
+  ir3.VOP1: _compile_vop12, ir3.VOP1_SDST: _compile_vop12, ir3.VOP2: _compile_vop12, ir3.VOPC: _compile_vopc, ir3.VOP3: _compile_vop3,
+  ir3.VOP3_SDST: _compile_vop3, ir3.VOP3SD: _compile_vop3sd, ir3.VOP3P: _compile_vop3p, ir3.VOPD: _compile_vopd,
+  ir3.DS: _compile_mem_op, ir3.FLAT: _compile_mem_op, ir3.GLOBAL: _compile_mem_op, ir3.SCRATCH: _compile_mem_op,
+  # RDNA4 instruction classes
+  ir4.SOPP: _compile_sopp, ir4.SMEM: _compile_smem, ir4.SOP1: _compile_sop, ir4.SOP2: _compile_sop, ir4.SOPC: _compile_sop, ir4.SOPK: _compile_sop,
+  ir4.VOP1: _compile_vop12, ir4.VOP1_SDST: _compile_vop12, ir4.VOP2: _compile_vop12, ir4.VOPC: _compile_vopc, ir4.VOP3: _compile_vop3,
+  ir4.VOP3_SDST: _compile_vop3, ir4.VOP3SD: _compile_vop3sd, ir4.VOP3P: _compile_vop3p, ir4.VOPD: _compile_vopd,
+  ir4.DS: _compile_mem_op, ir4.VFLAT: _compile_mem_op, ir4.VGLOBAL: _compile_mem_op, ir4.VSCRATCH: _compile_mem_op,
 }
 
 # ═══════════════════════════════════════════════════════════════════════════════
@@ -983,9 +1052,9 @@ _INST_HANDLERS: dict[type, Callable[..., UOp]] = {
 _canonical_runner_cache: list[tuple[int, int, int, object]] = []  # [(base, mask, size, runner), ...]
 
 @functools.cache
-def _get_runner(inst_bytes: bytes):
+def _get_runner(inst_bytes: bytes, arch: str = "rdna3"):
   """Build and compile instruction to CompiledRunner. Cached by instruction bytes, with canonical dedup."""
-  inst = decode_inst(inst_bytes)
+  inst = decode_inst(inst_bytes, arch)
   inst_size = inst.size()
   inst_int = int.from_bytes(inst_bytes[:inst_size], 'little')
 
@@ -1014,15 +1083,15 @@ def _get_runner(inst_bytes: bytes):
   return runner, True
 
 @functools.cache
-def decode_program(data: bytes) -> dict[int, tuple[str, Callable, list[int], Any]]:
+def decode_program(data: bytes, arch: str = "rdna3") -> dict[int, tuple[str, Callable, list[int], Any]]:
   """Decode program to {pc: (name, fxn, globals, runner)}."""
   result: dict[int, tuple[str, Callable, list[int], Any]] = {}
   i = 0
   while i < len(data):
-    inst = decode_inst(data[i:])
-    if isinstance(inst, SOPP) and inst.op == SOPPOp.S_CODE_END: break
+    inst = decode_inst(data[i:], arch)
+    if hasattr(inst, 'op') and inst.op in (ir3.SOPPOp.S_CODE_END, ir4.SOPPOp.S_CODE_END): break
     try:
-      runner, is_new = _get_runner(bytes(data[i:i + inst.size() + 4]))
+      runner, is_new = _get_runner(bytes(data[i:i + inst.size() + 4]), arch)
       if DEBUG >= 3:
         try: inst_str = repr(inst)
         except Exception: inst_str = f"<{type(inst).__name__} at PC={i}>"
@@ -1081,9 +1150,9 @@ class WaveState:
 # ═══════════════════════════════════════════════════════════════════════════════
 
 def run_asm(lib: int, lib_sz: int, gx: int, gy: int, gz: int, lx: int, ly: int, lz: int, args_ptr: int, rsrc2: int = 0x19c,
-            scratch_size: int = 0) -> int:
+            scratch_size: int = 0, arch: str = "rdna3") -> int:
   """Execute AMD assembly program. scratch_size is private_segment_fixed_size from kernel descriptor (per-lane)."""
-  program_raw = decode_program(bytes((ctypes.c_char * lib_sz).from_address(lib).raw))
+  program_raw = decode_program(bytes((ctypes.c_char * lib_sz).from_address(lib).raw), arch)
   program = {lib + offset: val for offset, val in program_raw.items()}  # Remap to actual addresses
   lds_size = ((rsrc2 & hsa.AMD_COMPUTE_PGM_RSRC_TWO_GRANULATED_LDS_SIZE) >> hsa.AMD_COMPUTE_PGM_RSRC_TWO_GRANULATED_LDS_SIZE_SHIFT) * 512
   total_threads = lx * ly * lz
@@ -1111,6 +1180,12 @@ def run_asm(lib: int, lib_sz: int, gx: int, gy: int, gz: int, lx: int, ly: int,
                                  (hsa.AMD_COMPUTE_PGM_RSRC_TWO_ENABLE_SGPR_WORKGROUP_ID_Z, gidz)]:
               if rsrc2 & enabled: st._write_sgpr(sgpr_idx, gid); sgpr_idx += 1
 
+            # RDNA4 uses TTMP registers for workgroup IDs: ttmp[9]=gidx, ttmp[10]=gidy, ttmp[11]=gidz
+            if arch == "rdna4":
+              st._write_sgpr(ttmp[9].offset, gidx)
+              st._write_sgpr(ttmp[10].offset, gidy)
+              st._write_sgpr(ttmp[11].offset, gidz)
+
             # v0 = packed workitem IDs, scratch stride in secret SGPR
             for lane in range(n_lanes):
               tid = wave_start + lane
@@ -1127,7 +1202,7 @@ def run_asm(lib: int, lib_sz: int, gx: int, gy: int, gz: int, lx: int, ly: int,
               assert fxn is not None, f"[emu] No fxn for {name} at PC={pc}"
               assert 4 not in globals_list or scratch_buf, f"SCRATCH instruction {name} but scratch_size=0"
               if DEBUG >= 6:
-                inst = decode_inst(bytes((ctypes.c_char * 12).from_address(pc).raw))
+                inst = decode_inst(bytes((ctypes.c_char * 12).from_address(pc).raw), arch)
                 print(f"[emu] exec PC={pc:X}: {inst!r}")
               fxn(*[c_bufs[g] for g in globals_list])
             else: raise RuntimeError("exceeded 1M instructions, likely infinite loop")
diff --git a/extra/assembly/amd/pcode.py b/extra/assembly/amd/pcode.py
index 3f826f3260..166f2c5e48 100644
--- a/extra/assembly/amd/pcode.py
+++ b/extra/assembly/amd/pcode.py
@@ -200,6 +200,17 @@ def _ff1(val: UOp, bits: int) -> UOp:
     result = cond.where(_const(dtypes.int, i), result)
   return result
 
+def _sad_u8(a: UOp, b: UOp, acc: UOp, masked: bool = False) -> UOp:
+  """Sum of absolute differences of 4 unsigned bytes + accumulator. If masked, skips bytes where a == 0."""
+  a, b, acc = a.cast(dtypes.uint32), b.cast(dtypes.uint32), acc.cast(dtypes.uint32)
+  result = acc
+  for i in range(4):
+    a_byte = (a >> _u32(i * 8)) & _u32(0xFF)
+    b_byte = (b >> _u32(i * 8)) & _u32(0xFF)
+    diff = (a_byte > b_byte).where(a_byte - b_byte, b_byte - a_byte)
+    result = result + (a_byte.ne(_u32(0)).where(diff, _u32(0)) if masked else diff)
+  return result
+
 _FUNCS: dict[str, Callable[..., UOp]] = {
   'sqrt': lambda a: UOp(Ops.SQRT, a.dtype, (a,)), 'trunc': lambda a: UOp(Ops.TRUNC, a.dtype, (a,)),
   'log2': lambda a: UOp(Ops.LOG2, a.dtype, (a,)), 'sin': lambda a: _trig_reduce(a),
@@ -254,6 +265,15 @@ _FUNCS: dict[str, Callable[..., UOp]] = {
   # Float to int16 conversions
   'v_cvt_i16_f32': lambda a: UOp(Ops.TRUNC, dtypes.float32, (a.bitcast(dtypes.float32),)).cast(dtypes.int16),
   'v_cvt_u16_f32': lambda a: _f_to_u(a.bitcast(dtypes.float32), dtypes.uint16),
+  # SAD (Sum of Absolute Differences) - sum |a_i - b_i| for 4 bytes + accumulator
+  'v_sad_u8': lambda a, b, c: _sad_u8(a, b, c),
+  'v_msad_u8': lambda a, b, c: _sad_u8(a, b, c, masked=True),
+  # System NOPs - these are scheduling hints, no effect on emulation
+  'MIN': lambda a, b: (a < b).where(a, b),
+  's_nop': lambda a: _u32(0),
+  # Address calculation for memory operations
+  'CalcDsAddr': lambda a, o, *r: a.cast(dtypes.uint32) + o.cast(dtypes.uint32),
+  'CalcGlobalAddr': lambda v, s, *r: v.cast(dtypes.uint64) + s.cast(dtypes.uint64),
 }
 for is_max, name in [(False, 'min'), (True, 'max')]:
   for dt, sfx in [(dtypes.float32, 'f32'), (dtypes.int, 'i32'), (dtypes.uint32, 'u32'), (dtypes.int16, 'i16'), (dtypes.uint16, 'u16')]:
@@ -435,7 +455,7 @@ class Parser:
         self.eat('DOT')
         dt_name = self.eat('IDENT').val
         return self._handle_mem_load(addr, DTYPES.get(dt_name, dtypes.uint32))
-      if name == 'VGPR':
+      if name == 'VGPR' and self.at('LBRACKET'):
         self.eat('LBRACKET')
         lane = self.parse()
         self.eat('RBRACKET')
@@ -462,7 +482,21 @@ class Parser:
       if self.try_eat('LBRACE'):
         idx = self.eat('NUM').val
         self.eat('RBRACE')
-        elem = self.vars.get(f'{name}{idx}', _u32(0))
+        # Handle VGPR{lane}[reg] - 2D array access after loop unrolling
+        if name == 'VGPR' and self.at('LBRACKET'):
+          self.eat('LBRACKET')
+          reg = self.parse()
+          self.eat('RBRACKET')
+          vgpr = self.vars.get('_vgpr')
+          if vgpr is None: return _u32(0)
+          return vgpr.index(_to_u32(reg) * _u32(32) + _u32(int(idx)), ptr=True).load()
+        elem = self.vars.get(f'{name}@{idx}', self.vars.get(f'{name}{idx}'))
+        if elem is None:
+          # Extract bit idx from base variable (like var[idx])
+          base = self.vars.get(name)
+          assert isinstance(base, UOp), f"unknown variable: {name}{idx}"
+          dt = dtypes.uint64 if base.dtype in (dtypes.uint64, dtypes.int64) else dtypes.uint32
+          elem = (base.cast(dt) >> _const(dt, int(idx))) & _const(dt, 1)
         if self.try_eat('DOT'):
           dt_name = self.eat('IDENT').val
           return _cast_to(elem, DTYPES.get(dt_name, dtypes.uint32))
@@ -475,15 +509,13 @@ class Parser:
         return self._handle_bracket_rest(first, _u32(0), name)
       if name in self.vars:
         v = self.vars[name]
-        return v if isinstance(v, UOp) else _u32(0) if isinstance(v, dict) else _u32(0)
+        assert isinstance(v, UOp), f"expected UOp for {name}, got {type(v)}"
+        return v
       raise RuntimeError(f"unknown variable: {name}")
     raise RuntimeError(f"unexpected token in primary: {self.peek()}")
 
   def _handle_dot(self, base, field: str) -> UOp:
-    if isinstance(base, str): return _u32(0)
-    if not isinstance(base, UOp):
-      if isinstance(base, dict): return base.get(field, _u32(0))
-      return _u32(0)
+    assert isinstance(base, UOp), f"expected UOp for dot access, got {type(base)}"
     if field == 'u64' and self.at('LBRACKET') and self.peek(1).type == 'IDENT' and self.peek(1).val == 'laneId':
       self.eat('LBRACKET')
       self.eat_val('laneId', 'IDENT')
@@ -541,9 +573,11 @@ class Parser:
       var_name = self._find_var_name(base)
     if first.op == Ops.CONST:
       idx = int(first.arg)
+      # Check for array element (var@idx)
       if var_name and f'{var_name}@{idx}' in self.vars:
         v = self.vars[f'{var_name}@{idx}']
         return _cast_to(v, dt_suffix) if dt_suffix else v
+      # Bit extraction
       dt = dtypes.uint64 if base.dtype in (dtypes.uint64, dtypes.int64) else dtypes.uint32
       base_cast = base.cast(dt) if base.dtype != dt else base
       result = ((base_cast >> _const(dt, idx)) & _const(dt, 1))
@@ -631,13 +665,14 @@ class Parser:
     raise RuntimeError(f"unexpected token after {bits}': {self.peek()}")
 
   def _parse_number(self, num: str) -> UOp:
-    if num.startswith('0x') or num.startswith('0X'): return _const(dtypes.uint64, int(num.rstrip('ULul'), 16))
-    suffix, num = _strip_suffix(num)
-    if '.' in num or suffix in ('F', 'f'):
-      return _const(dtypes.float32 if suffix in ('F', 'f') else dtypes.float64, float(num))
-    val = int(num)
-    if 'ULL' in suffix: return _const(dtypes.uint64, val)
-    if 'LL' in suffix or 'L' in suffix: return _const(dtypes.uint64, val)
+    if num.startswith('0x') or num.startswith('0X'):
+      is_u64 = num.upper().endswith('ULL') or num.upper().endswith('LL') or num.upper().endswith('UL')
+      return _const(dtypes.uint64 if is_u64 else dtypes.uint32, int(num.rstrip('ULul'), 16))
+    suffix, num_str = _strip_suffix(num)
+    if '.' in num_str or suffix in ('F', 'f'):
+      return _const(dtypes.float32 if suffix in ('F', 'f') else dtypes.float64, float(num_str))
+    val = int(num_str)
+    if 'ULL' in suffix or 'LL' in suffix or 'L' in suffix: return _const(dtypes.uint64, val)
     if 'U' in suffix: return _const(dtypes.uint32, val)
     return _const(dtypes.int if val < 0 else dtypes.uint32, val)
 
@@ -655,7 +690,8 @@ class Parser:
       if ';' in body or '\n' in body or 'return' in body.lower():
         lines = [l.strip() for l in body.replace(';', '\n').split('\n') if l.strip() and not l.strip().startswith('//')]
         _, _, result = parse_block(lines, 0, lv, self.funcs)
-        return result if result is not None else _u32(0)
+        assert result is not None, f"lambda {name} must return a value"
+        return result
       return parse_expr(body, lv, self.funcs)
     if name in self.funcs:
       return self.funcs[name](*args)
@@ -663,7 +699,7 @@ class Parser:
 
   def _handle_mem_load(self, addr: UOp, dt) -> UOp:
     mem = self.vars.get('_vmem') if '_vmem' in self.vars else self.vars.get('_lds')
-    if mem is None: return _const(dt, 0)
+    assert mem is not None, "memory load requires _vmem or _lds"
     adt = dtypes.uint64 if addr.dtype == dtypes.uint64 else dtypes.uint32
     active = self.vars.get('_active')
     gate = (active,) if active is not None else ()
@@ -725,29 +761,9 @@ def parse_tokens(toks: list[Token], vars: dict[str, VarVal], funcs: dict | None
 
 # Unified block parser for pcode
 def _subst_loop_var(line: str, loop_var: str, val: int) -> str:
-  """Substitute loop variable and evaluate bracket expressions.
-  Converts var[loop_var] to var{val} for array element access (like the old regex parser)."""
+  """Substitute loop variable with its value."""
   toks = tokenize(line)
-  # First pass: convert var[loop_var] to var{loop_var} to mark for array element assignment
-  result_toks: list[Token] = []
-  j = 0
-  while j < len(toks):
-    t = toks[j]
-    # Check for pattern: IDENT[loop_var] where it's not preceded by a dot (not .type[...])
-    if t.type == 'IDENT' and j+3 < len(toks) and toks[j+1].type == 'LBRACKET' and toks[j+2].type == 'IDENT' and toks[j+2].val == loop_var and toks[j+3].type == 'RBRACKET':
-      # Check that it's not .type[loop_var]
-      if not result_toks or result_toks[-1].type != 'DOT':
-        result_toks.append(t)
-        result_toks.append(Token('LBRACE', '{'))
-        result_toks.append(Token('NUM', str(val)))
-        result_toks.append(Token('RBRACE', '}'))
-        j += 4
-        continue
-    result_toks.append(t)
-    j += 1
-  # Second pass: substitute loop variable in remaining positions
-  subst_parts = [str(val) if t.type == 'IDENT' and t.val == loop_var else t.val for t in result_toks if t.type != 'EOF']
-  return ' '.join(subst_parts)
+  return ' '.join(str(val) if t.type == 'IDENT' and t.val == loop_var else t.val for t in toks if t.type != 'EOF')
 
 def _set_bits(old: UOp, val: UOp, width: int, offset: int) -> UOp:
   """Set bits [offset:offset+width) in old to val, masking and shifting appropriately."""
@@ -797,8 +813,9 @@ def parse_block(lines: list[str], start: int, vars: dict[str, VarVal], funcs: di
       def parse_bound():
         if p.at('NUM') and p.peek(1).type == 'QUOTE': p.eat('NUM'); p.eat('QUOTE')
         if p.at('NUM'): return int(p.eat('NUM').val.rstrip('UuLl'))
-        expr = p.parse()
-        return int(expr.arg) if expr.op == Ops.CONST else 0
+        expr = p.parse().simplify()
+        assert expr.op == Ops.CONST, f"loop bound must be constant, got {expr}"
+        return int(expr.arg)
       start_val = parse_bound()
       p.eat('COLON')
       end_val = parse_bound()
@@ -844,7 +861,9 @@ def parse_block(lines: list[str], start: int, vars: dict[str, VarVal], funcs: di
 
     # declare
     if first == 'declare':
-      if '[' not in line and len(toks) >= 2 and toks[1].type == 'IDENT': vars[toks[1].val] = _u32(0)
+      # Initialize scalar declarations (skip arrays and vars already passed as srcs)
+      if '[' not in line and len(toks) >= 2 and toks[1].type == 'IDENT':
+        vars.setdefault(toks[1].val, _u32(0))
       i += 1; continue
 
     # lambda definition
@@ -902,6 +921,7 @@ def parse_block(lines: list[str], start: int, vars: dict[str, VarVal], funcs: di
       j, lane_toks = _match_bracket(toks, 1)
       if j < len(toks) and toks[j].type == 'LBRACKET':
         j, reg_toks = _match_bracket(toks, j)
+        if j < len(toks) and toks[j].type == 'DOT': j += 2  # skip .type suffix
         if j < len(toks) and toks[j].type == 'EQUALS': j += 1
         ln, rg, val = parse_tokens(lane_toks, vars, funcs), parse_tokens(reg_toks, vars, funcs), parse_tokens(toks[j:], vars, funcs)
         if assigns is not None: assigns.append((f'VGPR[{_tok_str(lane_toks)}][{_tok_str(reg_toks)}]', (_to_u32(rg) * _u32(32) + _to_u32(ln), val)))
@@ -965,19 +985,32 @@ def parse_block(lines: list[str], start: int, vars: dict[str, VarVal], funcs: di
             block_assigns[var] = vars[var] = _set_bit(existing, _to_u32(parse_tokens(bit_toks, vars, funcs)), parse_tokens(toks[j+1:], vars, funcs))
             i += 1; continue
 
-    # Array element: var{idx} = value
-    if len(toks) >= 5 and toks[0].type == 'IDENT' and toks[1].type == 'LBRACE' and toks[2].type == 'NUM':
-      var, idx = toks[0].val, int(toks[2].val)
-      j = 4
-      while j < len(toks) and toks[j].type != 'EQUALS': j += 1
-      if j < len(toks):
-        val = parse_tokens(toks[j+1:], vars, funcs)
-        existing = block_assigns.get(var, vars.get(var))
-        if existing is not None and isinstance(existing, UOp):
-          block_assigns[var] = vars[var] = _set_bit(existing, _u32(idx), val)
-        else:
-          block_assigns[f'{var}@{idx}'] = vars[f'{var}@{idx}'] = val
-        i += 1; continue
+    # Array element: var[idx] = value (static index) or var[expr] = value (dynamic)
+    if len(toks) >= 4 and toks[0].type == 'IDENT' and toks[1].type == 'LBRACKET':
+      var = toks[0].val
+      j, idx_toks = _match_bracket(toks, 1)
+      if j < len(toks) and toks[j].type == 'EQUALS':
+        # Static index: var[NUM] = value
+        if len(idx_toks) == 1 and idx_toks[0].type == 'NUM':
+          idx = int(idx_toks[0].val.rstrip('UuLl'))
+          val = parse_tokens(toks[j+1:], vars, funcs)
+          existing = block_assigns.get(var, vars.get(var))
+          if existing is not None and isinstance(existing, UOp):
+            block_assigns[var] = vars[var] = _set_bit(existing, _u32(idx), val)
+          else:
+            block_assigns[f'{var}@{idx}'] = vars[f'{var}@{idx}'] = val
+          i += 1; continue
+        # Dynamic index: var[expr] = value where var has @-elements
+        elems = [(k.split('@')[1], v) for k, v in {**vars, **block_assigns}.items() if k.startswith(f'{var}@') and isinstance(v, UOp)]
+        if elems:
+          idx_expr = parse_tokens(idx_toks, vars, funcs)
+          val = parse_tokens(toks[j+1:], vars, funcs)
+          for elem_idx_str, old_elem in elems:
+            elem_idx = int(elem_idx_str)
+            cond = _to_u32(idx_expr).eq(_u32(elem_idx))
+            new_val = cond.where(val.cast(old_elem.dtype) if val.dtype != old_elem.dtype else val, old_elem)
+            block_assigns[f'{var}@{elem_idx}'] = vars[f'{var}@{elem_idx}'] = new_val
+          i += 1; continue
 
     # Compound assignment: var += or var -=
     assign_op = next((j for j, t in enumerate(toks) if t.type == 'ASSIGN_OP'), None)
@@ -1024,13 +1057,14 @@ def parse_block(lines: list[str], start: int, vars: dict[str, VarVal], funcs: di
       def parse_cond(s, kw):
         ll = s.lower()
         return _to_bool(parse_expr(s[ll.find(kw) + len(kw):ll.rfind('then')].strip(), vars, funcs))
-      def not_static_false(c): return c.op != Ops.CONST or c.arg is not False
+      def is_const(c, v): return c.op == Ops.CONST and c.arg is v
       cond = parse_cond(line, 'if')
-      conditions: list[tuple[UOp, UOp | dict[str, VarVal] | None]] = [(cond, None)] if not_static_false(cond) else []
+      conditions: list[tuple[UOp, UOp | dict[str, VarVal] | None]] = [(cond, None)] if not is_const(cond, False) else []
       else_branch: tuple[UOp | None, dict[str, VarVal]] = (None, {})
       vars_snap = dict(vars)
+      static_true = is_const(cond, True)  # track if any condition is statically true
       i += 1
-      i, branch, ret = parse_block(lines, i, vars, funcs, assigns)
+      i, branch, ret = parse_block(lines, i, vars, funcs, assigns if not is_const(cond, False) else None)
       if conditions: conditions[0] = (cond, ret if ret is not None else branch)
       vars.clear(); vars.update(vars_snap)
       while i < len(lines):
@@ -1039,12 +1073,16 @@ def parse_block(lines: list[str], start: int, vars: dict[str, VarVal], funcs: di
         lf = ltoks[0].val.lower()
         if lf == 'elsif':
           c = parse_cond(lines[i], 'elsif')
-          i += 1; i, branch, ret = parse_block(lines, i, vars, funcs, assigns)
-          if not_static_false(c): conditions.append((c, ret if ret is not None else branch))
+          take = not static_true and not is_const(c, False)
+          i += 1; i, branch, ret = parse_block(lines, i, vars, funcs, assigns if take else None)
+          if take:
+            conditions.append((c, ret if ret is not None else branch))
+            if is_const(c, True): static_true = True
           vars.clear(); vars.update(vars_snap)
         elif lf == 'else':
-          i += 1; i, branch, ret = parse_block(lines, i, vars, funcs, assigns)
-          else_branch = (ret, branch)
+          i += 1
+          i, branch, ret = parse_block(lines, i, vars, funcs, assigns if not static_true else None)
+          if not static_true: else_branch = (ret, branch)
           vars.clear(); vars.update(vars_snap)
         elif lf == 'endif': i += 1; break
         else: break
@@ -1056,17 +1094,21 @@ def parse_block(lines: list[str], start: int, vars: dict[str, VarVal], funcs: di
             if rv.dtype != result.dtype and rv.dtype.itemsize == result.dtype.itemsize: result = result.cast(rv.dtype)
             result = c.where(rv, result)
         return i, block_assigns, result
-      # Main style: merge variable assignments with WHERE
-      else_assigns = else_branch[1]
-      all_vars = set().union(*[ba.keys() for _, ba in conditions if isinstance(ba, dict)], else_assigns.keys())
-      for var in all_vars:
-        res: Any = else_assigns.get(var, block_assigns.get(var, vars.get(var, _u32(0))))
-        for cond, ba in reversed(conditions):
-          if isinstance(ba, dict) and var in ba:
-            tv = ba[var]
-            if isinstance(tv, UOp) and isinstance(res, UOp):
-              res = cond.where(tv, res.cast(tv.dtype) if tv.dtype != res.dtype and tv.dtype.itemsize == res.dtype.itemsize else res)
-        block_assigns[var] = vars[var] = res
+      # If statically true, use that branch directly; otherwise merge with WHERE
+      if static_true:
+        ba = next((b for c, b in conditions if is_const(c, True) and isinstance(b, dict)), {})
+        block_assigns.update(ba); vars.update(ba)
+      else:
+        else_assigns = else_branch[1]
+        all_vars = set().union(*[ba.keys() for _, ba in conditions if isinstance(ba, dict)], else_assigns.keys())
+        for var in all_vars:
+          res: Any = else_assigns.get(var, block_assigns.get(var, vars.get(var, _u32(0))))
+          for cond, ba in reversed(conditions):
+            if isinstance(ba, dict) and var in ba:
+              tv = ba[var]
+              if isinstance(tv, UOp) and isinstance(res, UOp):
+                res = cond.where(tv, res.cast(tv.dtype) if tv.dtype != res.dtype and tv.dtype.itemsize == res.dtype.itemsize else res)
+          block_assigns[var] = vars[var] = res
       continue
 
     # Regular assignment: var = value
diff --git a/extra/assembly/amd/sqttmap.py b/extra/assembly/amd/sqttmap.py
index 6cb7b6b8eb..addb481eec 100644
--- a/extra/assembly/amd/sqttmap.py
+++ b/extra/assembly/amd/sqttmap.py
@@ -2,11 +2,8 @@
 from dataclasses import dataclass
 from typing import Iterator
 
-from tinygrad.runtime.support.elf import elf_loader
-
 from extra.assembly.amd.sqtt import decode, print_packets, INST, VALUINST, IMMEDIATE, WAVESTART, WAVEEND, InstOp, PacketType, IMMEDIATE_MASK
 from extra.assembly.amd.dsl import Inst
-from extra.assembly.amd import decode_inst
 from extra.assembly.amd.autogen.rdna3.ins import SOPP, s_endpgm
 from extra.assembly.amd.autogen.rdna3.enum import SOPPOp
 
@@ -16,19 +13,11 @@ class InstructionInfo:
   wave: int
   inst: Inst
 
-def map_insts(data:bytes, lib:bytes) -> Iterator[tuple[PacketType, InstructionInfo|None]]:
+def map_insts(data:bytes, lib:bytes, target:int) -> Iterator[tuple[PacketType, InstructionInfo|None]]:
   """maps SQTT packets to instructions, yields (packet, instruction_info or None)"""
   # map pcs to insts
-  pc_map:dict[int, Inst] = {}
-  image, sections, _ = elf_loader(lib)
-  text = next((sh for sh in sections if sh.name == ".text"), None)
-  assert text is not None, "no .text section found"
-  text_off, text_size = text.header.sh_addr, text.header.sh_size
-  offset = text_off
-  while offset < text_off + text_size:
-    inst = decode_inst(image[offset:])
-    pc_map[offset-text_off] = inst
-    offset += inst.size()
+  from tinygrad.viz.serve import amd_decode
+  pc_map = amd_decode(lib, target)
 
   wave_pc:dict[int, int] = {}
   # only processing packets on one [CU, SIMD] unit
@@ -37,7 +26,7 @@ def map_insts(data:bytes, lib:bytes) -> Iterator[tuple[PacketType, InstructionIn
     if not simd_select(p): continue
     if isinstance(p, WAVESTART):
       assert p.wave not in wave_pc, "only one inflight wave per unit"
-      wave_pc[p.wave] = 0
+      wave_pc[p.wave] = next(iter(pc_map))
       continue
     if isinstance(p, WAVEEND):
       pc = wave_pc.pop(p.wave)
@@ -80,22 +69,22 @@ def map_insts(data:bytes, lib:bytes) -> Iterator[tuple[PacketType, InstructionIn
 # test to compare every packet with the rocprof decoder
 
 def test_rocprof_inst_traces_match(sqtt, prg, target):
-  from tinygrad.viz.serve import llvm_disasm
+  from tinygrad.viz.serve import amd_decode
   from extra.sqtt.roc import decode as roc_decode, InstExec
-  disasm = {addr+prg.base:inst_disasm for addr, inst_disasm in llvm_disasm(target, prg.lib).items()}
-  rctx = roc_decode([sqtt], {prg.name:disasm})
-  rwaves = rctx.inst_execs[(sqtt.kern, sqtt.exec_tag)]
+  addr_table = amd_decode(prg.lib, target)
+  disasm = {addr+prg.base:(inst.disasm(), inst.size()) for addr,inst in addr_table.items()}
+  rctx = roc_decode([sqtt], {prg.tag:disasm})
+  rwaves = rctx.inst_execs.get((sqtt.kern, sqtt.exec_tag), [])
   rwaves_iter:dict[int, list[Iterator[InstExec]]] = {} # wave unit (0-15) -> list of inst trace iterators for all executions on that unit
   for w in rwaves: rwaves_iter.setdefault(w.wave_id, []).append(w.unpack_insts())
-  rwaves_base = next(iter(disasm)) # base program counter
 
   passed_insts = 0
-  for pkt, info in map_insts(sqtt.blob, prg.lib):
+  for pkt, info in map_insts(sqtt.blob, prg.lib, target):
     if DEBUG >= 2: print_packets([pkt])
     if info is None: continue
     if DEBUG >= 2: print(f"{' '*29}{info.inst.disasm()}")
     rocprof_inst = next(rwaves_iter[info.wave][0])
-    ref_pc = rocprof_inst.pc-rwaves_base
+    ref_pc = rocprof_inst.pc-prg.base
     # always check pc matches
     assert ref_pc == info.pc, f"pc mismatch {ref_pc}:{disasm[rocprof_inst.pc][0]} != {info.pc}:{info.inst.disasm()}"
     # special handling for s_endpgm, it marks the wave completion.
@@ -110,7 +99,8 @@ def test_rocprof_inst_traces_match(sqtt, prg, target):
   for k,v in rwaves_iter.items():
     assert len(v) == 0, f"incomplete wave {k}"
 
-  print(f"passed for {passed_insts} instructions across {len(rwaves)} waves scheduled on {len(rwaves_iter)} wave units")
+  if len(rwaves):
+    print(f"passed for {passed_insts} instructions across {len(rwaves)} waves scheduled on {len(rwaves_iter)} wave units")
 
 if __name__ == "__main__":
   import argparse, pickle, pathlib
@@ -123,7 +113,7 @@ if __name__ == "__main__":
   with open(args.profile, "rb") as f:
     data = pickle.load(f)
   sqtt_events = [e for e in data if type(e).__name__ == "ProfileSQTTEvent"]
-  kern_events = {e.name:e for e in data if type(e).__name__ == "ProfileProgramEvent"}
+  kern_events = {e.tag:e for e in data if type(e).__name__ == "ProfileProgramEvent"}
   target = next((e for e in data if type(e).__name__ == "ProfileDeviceEvent" and e.device.startswith("AMD"))).props["gfx_target_version"]
   for e in sqtt_events:
     if args.kernel is not None and args.kernel != e.kern: continue
diff --git a/extra/assembly/amd/test/hw/helpers.py b/extra/assembly/amd/test/hw/helpers.py
index 5bc1227c46..73ce9cb01b 100644
--- a/extra/assembly/amd/test/hw/helpers.py
+++ b/extra/assembly/amd/test/hw/helpers.py
@@ -43,6 +43,23 @@ VCC = VCC_LO  # For VOP3SD sdst field (VCC_LO is exported from dsl)
 USE_HW = os.environ.get("USE_HW", "0") == "1"
 FLOAT_TOLERANCE = 1e-5
 
+def get_gpu_target() -> tuple[int, int, int]:
+  """Get the GPU target as (major, minor, stepping) tuple."""
+  if not USE_HW: return (0, 0, 0)
+  from tinygrad.device import Device
+  return Device["AMD"].target
+
+def skip_unless_gfx(min_major: int, min_minor: int = 0, reason: str = ""):
+  """Skip test if GPU target is below the minimum required version."""
+  import unittest
+  def decorator(test_func):
+    if not USE_HW: return test_func
+    target = get_gpu_target()
+    if target[0] < min_major or (target[0] == min_major and target[1] < min_minor):
+      return unittest.skip(reason or f"requires gfx{min_major}{min_minor}0+")(test_func)
+    return test_func
+  return decorator
+
 # Output buffer layout: vgpr[16][32], sgpr[16], vcc, scc, exec
 N_VGPRS, N_SGPRS, WAVE_SIZE = 16, 16, 32
 VGPR_BYTES = N_VGPRS * WAVE_SIZE * 4  # 16 regs * 32 lanes * 4 bytes = 2048
@@ -212,8 +229,12 @@ amdhsa.kernels:
 
   return parse_output(bytes(out_buf), n_lanes)
 
-def compare_wave_states(emu_st: WaveState, hw_st: WaveState, n_lanes: int, n_vgprs: int = N_VGPRS) -> list[str]:
-  """Compare two WaveStates and return list of differences."""
+def compare_wave_states(emu_st: WaveState, hw_st: WaveState, n_lanes: int, n_vgprs: int = N_VGPRS, ulp_tolerance: int = 0) -> list[str]:
+  """Compare two WaveStates and return list of differences.
+
+  Args:
+    ulp_tolerance: Allow up to this many ULPs difference for float comparisons (0 = exact match required)
+  """
   import math
   diffs = []
   for i in range(n_vgprs):
@@ -224,6 +245,11 @@ def compare_wave_states(emu_st: WaveState, hw_st: WaveState, n_lanes: int, n_vgp
         emu_f, hw_f = _f32(emu_val), _f32(hw_val)
         if math.isnan(emu_f) and math.isnan(hw_f):
           continue
+        # Check ULP difference for floats (only for same-sign values)
+        if ulp_tolerance > 0 and (emu_val < 0x80000000) == (hw_val < 0x80000000):
+          ulp_diff = abs(int(emu_val) - int(hw_val))
+          if ulp_diff <= ulp_tolerance:
+            continue
         diffs.append(f"v[{i}] lane {lane}: emu=0x{emu_val:08x} ({emu_f:.6g}) hw=0x{hw_val:08x} ({hw_f:.6g})")
   for i in range(N_SGPRS):
     emu_val = emu_st.sgpr[i]
@@ -236,16 +262,19 @@ def compare_wave_states(emu_st: WaveState, hw_st: WaveState, n_lanes: int, n_vgp
     diffs.append(f"scc: emu={emu_st.scc} hw={hw_st.scc}")
   return diffs
 
-def run_program(instructions: list, n_lanes: int = 1) -> WaveState:
+def run_program(instructions: list, n_lanes: int = 1, ulp_tolerance: int = 0) -> WaveState:
   """Run instructions and return WaveState.
 
   If USE_HW=1, runs on both emulator and hardware, compares results, and raises if they differ.
   Otherwise, runs only on emulator.
+
+  Args:
+    ulp_tolerance: Allow up to this many ULPs difference for float comparisons (0 = exact match required)
   """
   emu_st = run_program_emu(instructions, n_lanes)
   if USE_HW:
     hw_st = run_program_hw(instructions, n_lanes)
-    diffs = compare_wave_states(emu_st, hw_st, n_lanes)
+    diffs = compare_wave_states(emu_st, hw_st, n_lanes, ulp_tolerance=ulp_tolerance)
     if diffs:
       raise AssertionError(f"Emulator vs Hardware mismatch:\n" + "\n".join(diffs))
     return hw_st
diff --git a/extra/assembly/amd/test/hw/test_ds.py b/extra/assembly/amd/test/hw/test_ds.py
index 4f307bcf67..45e9080d27 100644
--- a/extra/assembly/amd/test/hw/test_ds.py
+++ b/extra/assembly/amd/test/hw/test_ds.py
@@ -719,5 +719,47 @@ class TestAtomicOrdering(unittest.TestCase):
     self.assertEqual(st.vgpr[0][4], 150, "Final value should be 150")
 
 
+class TestDsPermute(unittest.TestCase):
+  """Tests for DS_PERMUTE_B32 and DS_BPERMUTE_B32 instructions."""
+
+  def test_ds_permute_b32_identity(self):
+    """DS_PERMUTE_B32 with identity permutation (lane 0 sends to lane 0)."""
+    # For simplicity, test with single lane
+    instructions = [
+      v_mov_b32_e32(v[0], 0),  # addr = 0 (lane 0)
+      v_mov_b32_e32(v[1], 0xDEADBEEF),  # data
+      ds_permute_b32(v[2], v[0], v[1]),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    # Lane 0 sends to lane 0, so lane 0 gets 0xDEADBEEF
+    self.assertEqual(st.vgpr[0][2], 0xDEADBEEF)
+
+  def test_ds_bpermute_b32_identity(self):
+    """DS_BPERMUTE_B32 with identity permutation (each lane reads from itself)."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0),  # addr = 0 (read from lane 0)
+      v_mov_b32_e32(v[1], 0xCAFEBABE),  # data in lane 0
+      ds_bpermute_b32(v[2], v[0], v[1]),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    # Lane 0 reads from lane 0's v[1]
+    self.assertEqual(st.vgpr[0][2], 0xCAFEBABE)
+
+  def test_ds_permute_b32_broadcast(self):
+    """DS_PERMUTE_B32 broadcast - all lanes send to lane 0."""
+    # With 4 lanes, all sending to lane 0, highest lane wins
+    instructions = [
+      v_mov_b32_e32(v[0], 0),  # All lanes send to addr 0 (lane 0)
+      v_mov_b32_e32(v[1], 0x11111111),  # All lanes send same data
+      ds_permute_b32(v[2], v[0], v[1]),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=4)
+    # Lane 0 receives data (highest numbered active lane wins)
+    self.assertEqual(st.vgpr[0][2], 0x11111111)
+
+
 if __name__ == '__main__':
   unittest.main()
diff --git a/extra/assembly/amd/test/hw/test_sop.py b/extra/assembly/amd/test/hw/test_sop.py
index b8f340ad7b..4f51e67abc 100644
--- a/extra/assembly/amd/test/hw/test_sop.py
+++ b/extra/assembly/amd/test/hw/test_sop.py
@@ -62,6 +62,7 @@ class TestBasicScalar(unittest.TestCase):
     st = run_program(instructions, n_lanes=1)
     self.assertEqual(st.sgpr[1], 0x80000000)
 
+  @skip_unless_gfx(11, 5, "SALU FP ops require gfx1150+")
   def test_s_fmamk_f32(self):
     """S_FMAMK_F32: D = S0 * literal + S1."""
     # 2.0 * 3.0 + 1.0 = 7.0
@@ -73,6 +74,7 @@ class TestBasicScalar(unittest.TestCase):
     st = run_program(instructions, n_lanes=1)
     self.assertEqual(st.sgpr[2], f2i(7.0))
 
+  @skip_unless_gfx(11, 5, "SALU FP ops require gfx1150+")
   def test_s_fmamk_f32_negative(self):
     """S_FMAMK_F32 with negative values."""
     # -2.0 * 4.0 + 10.0 = 2.0
diff --git a/extra/assembly/amd/test/hw/test_vop1.py b/extra/assembly/amd/test/hw/test_vop1.py
index 12d518a066..10e7f24c52 100644
--- a/extra/assembly/amd/test/hw/test_vop1.py
+++ b/extra/assembly/amd/test/hw/test_vop1.py
@@ -1561,5 +1561,24 @@ class TestCvtNormF16(unittest.TestCase):
     self.assertAlmostEqual(result, 32768, delta=1)
 
 
+class TestPermlane64(unittest.TestCase):
+  """Tests for V_PERMLANE64_B32 instruction (wave64 cross-half swap)."""
+
+  def test_v_permlane64_b32_is_nop_in_wave32(self):
+    """V_PERMLANE64_B32 is a NOP in wave32 mode.
+
+    Per AMD pcode: "if WAVE32 then s_nop(...) else ... endif"
+    The emulator runs in wave32 mode, so this instruction should not modify registers.
+    """
+    instructions = [
+      v_mov_b32_e32(v[0], 0xCAFEBABE),  # source
+      v_mov_b32_e32(v[1], 0x12345678),  # dest (should be preserved)
+      v_permlane64_b32_e32(v[1], v[0]),  # NOP in wave32
+    ]
+    st = run_program(instructions, n_lanes=1)
+    # Dest register should be unchanged (NOP behavior in wave32)
+    self.assertEqual(st.vgpr[0][1], 0x12345678)
+
+
 if __name__ == '__main__':
   unittest.main()
diff --git a/extra/assembly/amd/test/hw/test_vop3.py b/extra/assembly/amd/test/hw/test_vop3.py
index 596fcd4711..f798519f91 100644
--- a/extra/assembly/amd/test/hw/test_vop3.py
+++ b/extra/assembly/amd/test/hw/test_vop3.py
@@ -3262,5 +3262,81 @@ class TestMinMaxF16Vop3(unittest.TestCase):
     self.assertAlmostEqual(result, 4.0, delta=0.01)
 
 
+class TestSadHi(unittest.TestCase):
+  """Tests for V_SAD_HI_U8 instruction."""
+
+  def test_v_sad_hi_u8_basic(self):
+    """V_SAD_HI_U8: (sad << 16) + acc."""
+    # |1-5| + |2-6| + |3-7| + |4-8| = 16, << 16 = 0x100000, + 100 = 0x100064
+    instructions = [
+      v_mov_b32_e32(v[0], 0x04030201),
+      v_mov_b32_e32(v[1], 0x08070605),
+      v_mov_b32_e32(v[2], 100),
+      v_sad_hi_u8(v[3], v[0], v[1], v[2]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][3], (16 << 16) + 100)
+
+  def test_v_sad_hi_u8_zero_diff(self):
+    """V_SAD_HI_U8: identical inputs gives acc only."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0x12345678),
+      v_mov_b32_e32(v[2], 50),
+      v_sad_hi_u8(v[3], v[0], v[0], v[2]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][3], 50)
+
+
+class TestPermlane(unittest.TestCase):
+  """Tests for V_PERMLANE16_B32 and V_PERMLANEX16_B32 instructions."""
+
+  def test_v_permlane16_b32_identity(self):
+    """V_PERMLANE16_B32 with identity permutation (lane i reads from lane i within row)."""
+    # lanesel encodes 4 bits per position: position i gets lanesel[i*4+3:i*4]
+    # Identity: position 0->0, 1->1, ..., 15->15
+    # lanesel = 0xFEDCBA9876543210 (positions 15-0 in nibbles)
+    instructions = [
+      v_mov_b32_e32(v[0], 0xDEADBEEF),  # source data
+      s_mov_b32(s[0], 0x76543210),       # lanesel low (positions 0-7)
+      s_mov_b32(s[1], 0xFEDCBA98),       # lanesel high (positions 8-15)
+      v_permlane16_b32(v[1], v[0], s[0], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    # Lane 0 reads from lane 0 (position 0 -> lanesel[3:0] = 0)
+    self.assertEqual(st.vgpr[0][1], 0xDEADBEEF)
+
+  def test_v_permlane16_b32_broadcast(self):
+    """V_PERMLANE16_B32 broadcast lane 0 to all lanes in row."""
+    # lanesel = all zeros -> all positions read from lane 0 within row
+    instructions = [
+      v_mov_b32_e32(v[0], 0xCAFEBABE),  # source data
+      s_mov_b32(s[0], 0),                # lanesel low = 0 (all read lane 0)
+      s_mov_b32(s[1], 0),                # lanesel high = 0
+      v_permlane16_b32(v[1], v[0], s[0], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=4)
+    # All lanes read from lane 0 of their row
+    for lane in range(4):
+      self.assertEqual(st.vgpr[lane][1], 0xCAFEBABE)
+
+  def test_v_permlanex16_b32_identity(self):
+    """V_PERMLANEX16_B32 cross-row read with identity selection."""
+    # In wave32: row 0 (lanes 0-15) reads from row 1 (lanes 16-31) and vice versa
+    # With single lane in row 0, it reads from lane 0 of row 1 (lane 16)
+    # But lane 16 doesn't exist in 1-lane test, so use 32 lanes
+    instructions = [
+      v_mov_b32_e32(v[0], 0x11111111),  # All lanes have this initially
+      s_mov_b32(s[0], 0x76543210),       # lanesel low
+      s_mov_b32(s[1], 0xFEDCBA98),       # lanesel high
+      v_permlanex16_b32(v[1], v[0], s[0], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=32)
+    # Lane 0 in row 0 reads from lane 0 of row 1 (lane 16)
+    self.assertEqual(st.vgpr[0][1], 0x11111111)
+    # Lane 16 in row 1 reads from lane 0 of row 0 (lane 0)
+    self.assertEqual(st.vgpr[16][1], 0x11111111)
+
+
 if __name__ == '__main__':
   unittest.main()
diff --git a/extra/assembly/amd/test/hw/test_vop3p.py b/extra/assembly/amd/test/hw/test_vop3p.py
index 62671b4ad8..185c61f7bb 100644
--- a/extra/assembly/amd/test/hw/test_vop3p.py
+++ b/extra/assembly/amd/test/hw/test_vop3p.py
@@ -767,6 +767,7 @@ class TestDot2F32F16(unittest.TestCase):
     """V_DOT2_F32_F16 with negative f16 values."""
     # src0 = {hi=-2.0, lo=3.0}, src1 = {hi=1.0, lo=2.0}
     # result = 3.0*2.0 + (-2.0)*1.0 + 0 = 6 - 2 = 4.0
+    # NOTE: Hardware DOT2 may have up to 1 ULP difference due to internal implementation
     src0 = (f32_to_f16(-2.0) << 16) | f32_to_f16(3.0)
     src1 = (f32_to_f16(1.0) << 16) | f32_to_f16(2.0)
     instructions = [
@@ -777,7 +778,7 @@ class TestDot2F32F16(unittest.TestCase):
       v_mov_b32_e32(v[2], 0),
       v_dot2_f32_f16(v[3], v[0], v[1], v[2], opsel_hi=3, opsel_hi2=1),
     ]
-    st = run_program(instructions, n_lanes=1)
+    st = run_program(instructions, n_lanes=1, ulp_tolerance=1)
     result = i2f(st.vgpr[0][3])
     self.assertAlmostEqual(result, 4.0, places=2)
 
diff --git a/extra/assembly/amd/test/test_rdna4_emu.py b/extra/assembly/amd/test/test_rdna4_emu.py
new file mode 100644
index 0000000000..57e427e643
--- /dev/null
+++ b/extra/assembly/amd/test/test_rdna4_emu.py
@@ -0,0 +1,98 @@
+import unittest, ctypes
+from extra.assembly.amd.autogen.rdna4 import ins as ir4
+from extra.assembly.amd.dsl import v, s
+from extra.assembly.amd.emu import WaveState, decode_program
+from tinygrad.device import Buffer, BufferSpec
+from tinygrad.dtype import dtypes
+
+class TestRDNA4Emu(unittest.TestCase):
+  def _run(self, insts: list, sgprs: dict[int, int] = None, vgprs: dict[tuple[int, int], int] = None) -> WaveState:
+    """Run instructions and return final WaveState."""
+    # Add S_ENDPGM if not present
+    if not any(isinstance(i, ir4.SOPP) and i.op == ir4.SOPPOp.S_ENDPGM for i in insts):
+      insts = list(insts) + [ir4.SOPP(ir4.SOPPOp.S_ENDPGM, simm=0)]
+
+    # Assemble and decode
+    code = b''.join(i.to_bytes() for i in insts)
+    code_buf = (ctypes.c_uint8 * len(code)).from_buffer_copy(code)
+    code_addr = ctypes.addressof(code_buf)
+    program_raw = decode_program(code, "rdna4")
+    program = {code_addr + offset: val for offset, val in program_raw.items()}
+
+    # Setup wave state
+    st = WaveState(n_lanes=1)
+    st.pc = code_addr
+    if sgprs:
+      for idx, val in sgprs.items(): st._write_sgpr(idx, val)
+    if vgprs:
+      for (reg, lane), val in vgprs.items(): st._write_vgpr(reg, lane, val)
+
+    # Setup vmem buffer with external_ptr=0 (maps to address 0, allows any pointer access)
+    vmem_buf = Buffer('CPU', 1 << 40, dtypes.uint32, options=BufferSpec(external_ptr=0)).ensure_allocated()
+
+    # Execute
+    c_bufs = [ctypes.c_uint64(st.sgpr_buf._buf.va_addr), ctypes.c_uint64(st.vgpr_buf._buf.va_addr),
+              ctypes.c_uint64(vmem_buf._buf.va_addr), ctypes.c_uint64(0), ctypes.c_uint64(0)]
+    for _ in range(100):
+      if (pc := st.pc) == 0xFFFFFFFFFFFFFFFF or pc not in program: break
+      _, fxn, globals_list, _ = program[pc]
+      fxn(*[c_bufs[g] for g in globals_list])
+    return st
+
+  def test_vopd_dual_mov(self):
+    """Test VOPD with two V_DUAL_MOV_B32 operations: v[1]=s[1], v[2]=s[2]."""
+    insts = [ir4.VOPD(ir4.VOPDOp.V_DUAL_MOV_B32, ir4.VOPDOp.V_DUAL_MOV_B32,
+                      vdstx=v[1], vdsty=v[2], srcx0=s[1], srcy0=s[2], vsrcx1=v[0], vsrcy1=v[0])]
+    st = self._run(insts, sgprs={1: 0x40e00000, 2: 0x41100000})  # 7.0f, 9.0f
+    self.assertEqual(st._read_vgpr(1, 0), 0x40e00000)  # v[1] = 7.0
+    self.assertEqual(st._read_vgpr(2, 0), 0x41100000)  # v[2] = 9.0
+
+  def test_vopd_dual_mov_after_other_vopd(self):
+    """Test VOPD reuse: first VOPD(v[3]=0, v[0]=?), then VOPD(v[1]=s[1], v[2]=s[2])."""
+    # This matches the BEAM kernel sequence that fails
+    insts = [
+      ir4.VOPD(ir4.VOPDOp.V_DUAL_MOV_B32, ir4.VOPDOp.V_DUAL_MOV_B32,
+               vdstx=v[3], vdsty=v[0], srcx0=0, srcy0=s[0], vsrcx1=v[0], vsrcy1=v[0]),  # v[3]=0, v[0]=s[0]
+      ir4.VOPD(ir4.VOPDOp.V_DUAL_MOV_B32, ir4.VOPDOp.V_DUAL_MOV_B32,
+               vdstx=v[1], vdsty=v[2], srcx0=s[1], srcy0=s[2], vsrcx1=v[0], vsrcy1=v[0]),  # v[1]=s[1], v[2]=s[2]
+    ]
+    st = self._run(insts, sgprs={0: 0x40a00000, 1: 0x40e00000, 2: 0x41100000})  # 5.0f, 7.0f, 9.0f
+    self.assertEqual(st._read_vgpr(1, 0), 0x40e00000)  # v[1] = 7.0
+    self.assertEqual(st._read_vgpr(2, 0), 0x41100000)  # v[2] = 9.0
+
+  def test_vopd_with_s_add_f32_sequence(self):
+    """Test full BEAM kernel sequence: s_add_f32 then VOPD."""
+    # This is the exact sequence from the failing BEAM kernel
+    insts = [
+      ir4.SOP2(ir4.SOP2Op.S_ADD_F32, sdst=s[0], ssrc0=s[0], ssrc1=s[8]),   # s[0] = s[0] + s[8]
+      ir4.SOP2(ir4.SOP2Op.S_ADD_F32, sdst=s[1], ssrc0=s[1], ssrc1=s[9]),   # s[1] = s[1] + s[9]
+      ir4.SOP2(ir4.SOP2Op.S_ADD_F32, sdst=s[2], ssrc0=s[2], ssrc1=s[10]),  # s[2] = s[2] + s[10]
+      ir4.VOPD(ir4.VOPDOp.V_DUAL_MOV_B32, ir4.VOPDOp.V_DUAL_MOV_B32,
+               vdstx=v[3], vdsty=v[0], srcx0=0, srcy0=s[0], vsrcx1=v[0], vsrcy1=v[0]),
+      ir4.VOPD(ir4.VOPDOp.V_DUAL_MOV_B32, ir4.VOPDOp.V_DUAL_MOV_B32,
+               vdstx=v[1], vdsty=v[2], srcx0=s[1], srcy0=s[2], vsrcx1=v[0], vsrcy1=v[0]),
+    ]
+    # Input: s[0:2] = [1,2,3], s[8:10] = [4,5,6]
+    # After s_add_f32: s[0:2] = [5,7,9]
+    st = self._run(insts, sgprs={0: 0x3f800000, 1: 0x40000000, 2: 0x40400000,  # 1.0, 2.0, 3.0
+                                  8: 0x40800000, 9: 0x40a00000, 10: 0x40c00000})  # 4.0, 5.0, 6.0
+    self.assertEqual(st._read_vgpr(1, 0), 0x40e00000)  # v[1] = 7.0
+    self.assertEqual(st._read_vgpr(2, 0), 0x41100000)  # v[2] = 9.0
+
+  def test_s_mov_b32_then_vopd(self):
+    """Test s_mov_b32 followed by VOPD - simulates BEAM kernel sequence."""
+    # Use s_mov_b32 with SGPR source (copy from pre-initialized SGPRs)
+    # s[10:12] will have values set by test harness, copy to s[0:2], then VOPD to VGPRs
+    insts = [
+      ir4.SOP1(ir4.SOP1Op.S_MOV_B32, sdst=s[0], ssrc0=s[10]),  # s[0] = s[10]
+      ir4.SOP1(ir4.SOP1Op.S_MOV_B32, sdst=s[1], ssrc0=s[11]),  # s[1] = s[11]
+      ir4.SOP1(ir4.SOP1Op.S_MOV_B32, sdst=s[2], ssrc0=s[12]),  # s[2] = s[12]
+      ir4.VOPD(ir4.VOPDOp.V_DUAL_MOV_B32, ir4.VOPDOp.V_DUAL_MOV_B32,
+               vdstx=v[1], vdsty=v[2], srcx0=s[1], srcy0=s[2], vsrcx1=v[0], vsrcy1=v[0]),
+    ]
+    st = self._run(insts, sgprs={10: 0x40a00000, 11: 0x40e00000, 12: 0x41100000})  # 5.0, 7.0, 9.0
+    self.assertEqual(st._read_vgpr(1, 0), 0x40e00000)  # v[1] = 7.0
+    self.assertEqual(st._read_vgpr(2, 0), 0x41100000)  # v[2] = 9.0
+
+if __name__ == '__main__':
+  unittest.main()
diff --git a/extra/assembly/amd/test/test_sqtt_examples.py b/extra/assembly/amd/test/test_sqtt_examples.py
index e50b0d6bd5..161664675a 100644
--- a/extra/assembly/amd/test/test_sqtt_examples.py
+++ b/extra/assembly/amd/test/test_sqtt_examples.py
@@ -203,12 +203,12 @@ class SQTTExamplesTestBase(unittest.TestCase):
 class TestSQTTExamplesRDNA3(SQTTExamplesTestBase):
   target = "gfx1100"
   expected = {
-    "profile_empty_run_0": [1803, 1908, 1928, 1979, 2006, 1912],
-    "profile_empty_run_1": [1803, 1908, 1928, 1979, 2006, 1912],
-    "profile_gemm_run_0": [2531, 1844, 1864, 1915, 1942, 1848, 3074, 1919, 1939, 1990, 2017, 1923, 19026, 1919, 1939, 1990, 2017, 1929],
-    "profile_gemm_run_1": [2554, 1844, 1864, 1915, 1942, 1848, 3084, 1919, 1939, 1990, 2017, 1923, 19010, 1919, 1939, 1990, 2017, 1923],
-    "profile_plus_run_0": [1900, 1908, 1928, 1979, 2006, 1912],
-    "profile_plus_run_1": [1856, 1908, 1928, 1979, 2006, 1912],
+    "profile_empty_run_0": [1844, 1885, 1905, 1956, 1983, 1889],
+    "profile_empty_run_1": [1780, 1885, 1905, 1956, 1983, 1889],
+    "profile_gemm_run_0": [2656, 2025, 2045, 2096, 2123, 2029, 3183, 2019, 2039, 2090, 2117, 2023, 19119, 2013, 2033, 2084, 2111, 2017],
+    "profile_gemm_run_1": [2662, 2025, 2045, 2096, 2123, 2029, 3179, 2019, 2039, 2090, 2117, 2023, 19113, 2071, 2091, 2142, 2169, 2075],
+    "profile_plus_run_0": [1886, 2013, 2033, 2084, 2111, 2017],
+    "profile_plus_run_1": [1988, 2071, 2091, 2142, 2169, 2075],
   }
 
 class TestSQTTExamplesRDNA4(SQTTExamplesTestBase): target = "gfx1200"
diff --git a/extra/gemm/amd_asm_matmul.py b/extra/gemm/amd_asm_matmul.py
index 0eb581c160..c2dc9f7456 100644
--- a/extra/gemm/amd_asm_matmul.py
+++ b/extra/gemm/amd_asm_matmul.py
@@ -471,7 +471,7 @@ THREADS = 128
 
 def test_matmul():
   dev = Device[Device.DEFAULT]
-  print(f"Device arch: {dev.arch}")
+  print(f"Device arch: {dev.renderer.arch}")
 
   if getenv("STOCK", 0):
     # Load the stock kernel from amd_seb/kernel8_batched_gmem.s
@@ -479,7 +479,7 @@ def test_matmul():
     asm = stock_path.read_text()
     print(f"Loaded stock kernel from {stock_path}")
   else:
-    asm = build_kernel(dev.arch)
+    asm = build_kernel(dev.renderer.arch)
 
   binary = dev.compiler.compile(asm)
   print(f"Compiled! Binary size: {len(binary)} bytes")
diff --git a/extra/gemm/asm/cdna/asm.py b/extra/gemm/asm/cdna/asm.py
new file mode 100644
index 0000000000..96611dddc9
--- /dev/null
+++ b/extra/gemm/asm/cdna/asm.py
@@ -0,0 +1,11517 @@
+from extra.assembly.amd.autogen.cdna.ins import *
+from tinygrad.dtype import dtypes
+
+# M0 is encoded with 124 (NULL in RDNA) in CDNA
+M0 = NULL
+
+# (M, N, K) -> (numWG, iters, total)
+GEMM_ARGS = {
+  (8192, 4096, 4096): (256, 64, 32768),
+  (8192, 14336, 4096): (256, 64, 114688),
+  (8192, 4096, 14336): (256, 224, 114688),
+  (8192, 128256, 4096): (16032, 64, 1026048),
+  (8192, 8192, 8192): (256, 128, 131072),
+  (4096, 4096, 4096): (256, 64, 16384),
+  (4096, 14336, 4096): (256, 64, 57344),
+  (4096, 4096, 14336): (256, 224, 57344),
+}
+ITERS_ARGS = {64: (67108864, 0), 128: (33554432, 0), 224: (613566757, 2147483656)}
+
+class Kernel:
+  def __init__(self, name="gemm"): self.name, self.instructions, self.labels, self.label_at_pos, self.pos = name, [], {}, {}, 0
+
+  def label(self, name):
+    self.labels[name] = self.pos
+    self.label_at_pos[self.pos] = name
+
+  def emit(self, inst, target=None):
+    self.instructions.append(inst)
+    inst._target, inst._pos = target, self.pos
+    self.pos += inst.size()
+    return inst
+
+  def waitcnt(self, lgkm=None, vm=None):
+    vmcnt, lgkmcnt, expcnt = vm if vm is not None else 63, lgkm if lgkm is not None else 15, 7
+    waitcnt = (vmcnt & 0xF) | ((expcnt & 0x7) << 4) | ((lgkmcnt & 0xF) << 8) | (((vmcnt >> 4) & 0x3) << 14)
+    self.emit(s_waitcnt(waitcnt))
+
+  def to_asm(self):
+    # patch branches
+    for inst in self.instructions:
+      if inst._target is None: continue
+      inst.simm16 = (self.labels[inst._target] - inst._pos - inst.size()) // 4
+    # convert instructions to bytes, pack hsa
+    inst_bytes = b"".join(inst.to_bytes() for inst in self.instructions)
+    body = "\n".join("  .byte " + ",".join(f"0x{b:02x}" for b in inst_bytes[i:i+16]) for i in range(0, len(inst_bytes), 16))
+    hsa = [('group_segment_fixed_size', 133120), ('private_segment_fixed_size', 0), ('kernarg_size', 24),
+           ('next_free_vgpr', 512), ('next_free_sgpr', 96), ('system_sgpr_workgroup_id_x', 1),
+           ('system_sgpr_workgroup_id_y', 1), ('system_sgpr_workgroup_id_z', 1), ('user_sgpr_kernarg_segment_ptr', 1),
+           ('user_sgpr_count', 2), ('user_sgpr_kernarg_preload_length', 0), ('user_sgpr_kernarg_preload_offset', 0),
+           ('accum_offset', 256), ('uses_dynamic_stack', 0), ('tg_split', 0), ('float_round_mode_32', 0),
+           ('float_round_mode_16_64', 0), ('float_denorm_mode_32', 3), ('float_denorm_mode_16_64', 3),
+           ('ieee_mode', 1), ('fp16_overflow', 0), ('dx10_clamp', 1)]
+    args = '\n'.join(f'      - .address_space: generic\n        .name: {n}\n        .offset: {i*8}\n'
+                     f'        .size: 8\n        .value_kind: global_buffer' for i,n in enumerate(['C', 'A', 'B']))
+    n = self.name
+    return '\n'.join(['.text', '.section\t.text.', f'.global\t{n}', '.p2align\t8', f'.type\t{n},@function', '', f'{n}:',
+      body, '', '.section .rodata,"a",@progbits', '.p2align 6, 0x0', f'.amdhsa_kernel {n}',
+      *[f'  .amdhsa_{k} {v}' for k, v in hsa], '.end_amdhsa_kernel', '', '.amdgpu_metadata', '---', 'amdhsa.kernels:',
+      '  - .args:', args, '    .group_segment_fixed_size: 133120', '    .kernarg_segment_align: 8',
+      '    .kernarg_segment_size: 24', '    .max_flat_workgroup_size: 256', f'    .name: {n}',
+      '    .private_segment_fixed_size: 0', '    .sgpr_count: 95', '    .sgpr_spill_count: 0', f'    .symbol: {n}.kd',
+      '    .vgpr_count: 249', '    .vgpr_spill_count: 0', '    .wavefront_size: 64', 'amdhsa.version:', '  - 1',
+      '  - 1', '...', '.end_amdgpu_metadata', ''])
+
+  # outputs readable source code for this kernel
+  def to_text(self) -> str:
+    lines, pos = [], 0
+    for inst in self.instructions:
+      if (label := self.label_at_pos.get(pos)) is not None: lines.append(f"{label}:")
+      lines.append(f"  {inst.disasm()}" if inst._target is None else f" {inst.op_name.lower()} {inst._target}")
+      pos += inst.size()
+    return "\n".join(lines)
+
+def build_kernel(batch, M, N, K, dtype):
+  numWG, iters, total = GEMM_ARGS[(M, N, K)]
+  total *= batch
+  magic, shift = ITERS_ARGS[iters]
+  v_mfma_16x16x32 = {dtypes.half:v_mfma_f32_16x16x32_f16, dtypes.bfloat16:v_mfma_f32_16x16x32_bf16}[dtype]
+  v_cvt_pk = {dtypes.half:v_cvt_pk_f16_f32, dtypes.bfloat16:v_cvt_pk_bf16_f32}[dtype]
+  v_cvt = {dtypes.half:v_cvt_f32_f16_e32, dtypes.bfloat16:v_cvt_f32_bf16_e32}[dtype]
+  k = Kernel(f"gemm_{batch}_{M}_{N}_{K}")
+  # load D, A, B pointers
+  k.emit(s_load_dwordx2(s[24:25], s[0:1], s[0], 0, 0, 0, 0, 1))
+  k.emit(s_load_dwordx2(s[30:31], s[0:1], s[0], 8, 0, 0, 0, 1))
+  k.emit(s_load_dwordx2(s[28:29], s[0:1], s[0], 16, 0, 0, 0, 1))
+  k.waitcnt(lgkm=0)
+  # params as constants
+  k.emit(s_mov_b32(s[69], numWG))
+  k.emit(s_mov_b32(s[20], N))
+  k.emit(s_mov_b32(s[21], batch * M))
+  k.emit(s_mov_b32(s[22], 1))
+  k.emit(s_mov_b32(s[23], K))
+  k.emit(s_mov_b32(s[36], N))
+  k.emit(s_mov_b32(s[37], 0))
+  k.emit(s_mov_b32(s[40], N))
+  k.emit(s_mov_b32(s[41], 0))
+  k.emit(s_mov_b32(s[42], K))
+  k.emit(s_mov_b32(s[43], 0))
+  k.emit(s_mov_b32(s[46], iters))
+  k.emit(s_mov_b32(s[47], magic))
+  k.emit(s_mov_b32(s[48], shift))
+  k.emit(s_mov_b32(s[49], total))
+  k.emit(s_mov_b32(s[62], 0))
+  k.emit(s_mov_b32(s[68], 0))
+  # kernel size is 256x256
+  k.emit(s_mov_b32(s[51], 256)); k.emit(s_mov_b32(s[52], 256))
+  k.emit(s_mov_b32(s[38], s[36]))
+  k.emit(s_mov_b32(s[39], s[37]))
+  k.emit(s_mov_b64(s[26:27], s[24:25]))
+  k.emit(s_and_b32(s[6], s[68], 4294901760))
+  k.emit(s_lshr_b32(s[6], s[6], 16))
+  k.emit(s_mov_b32(s[63], 0))
+  k.emit(s_setprio(3))
+  k.emit(s_mov_b32(M0, 133120))
+  k.emit(v_mov_b32_e32(v[180], v[0]))
+  # XCCG=256
+  # labels are named based on function:
+  # PGR = Prefetch Global Read (the global→LDS pipeline stage)
+  # SK = Stream-K (work partitioning by K-iterations, not tiles)
+  # WGM = WorkGroup Mapping (tile assignment scheme for cache locality)
+  # GLVW = Global Load Vector Width (edge tile width handling)
+  # BM0 = Block M offset 0 (register block position)
+  # OrdNLL = Ordered No-Load-Loop (final iteration without prefetch loads)
+  k.emit(s_mov_b32(s[75], 256))
+  k.emit(v_cvt_f32_u32_e32(v[18], s[75]))
+  k.emit(v_rcp_iflag_f32_e32(v[18], v[18]))
+  k.emit(v_cvt_f32_u32_e32(v[19], s[2]))
+  k.emit(v_mul_f32_e32(v[18], v[18], v[19]))
+  k.emit(v_cvt_u32_f32_e32(v[18], v[18]))
+  k.emit(v_mul_u32_u24_e64(v[19], v[18], s[75]))
+  k.emit(v_sub_u32_e32(v[19], s[2], v[19]))
+  k.emit(v_cmpx_eq_u32_e64(EXEC, v[19], s[75]))
+  k.emit(v_add_u32_e32(v[18], 1, v[18]))
+  k.emit(v_mov_b32_e32(v[19], 0))
+  k.emit(s_mov_b64(EXEC, -1))
+  k.emit(v_cmpx_gt_u32_e64(EXEC, v[19], s[75]))
+  k.emit(v_sub_u32_e64(v[18], v[18], 1))
+  k.emit(v_mul_u32_u24_e64(v[19], v[18], s[75]))
+  k.emit(v_sub_u32_e32(v[19], s[2], v[19]))
+  k.emit(s_mov_b64(EXEC, -1))
+  k.emit(v_readfirstlane_b32_e32(v[71], v[18]))
+  k.emit(v_readfirstlane_b32_e32(v[72], v[19]))
+  k.emit(s_mul_i32(s[71], s[71], s[75]))
+  k.emit(s_lshr_b32(s[72], s[72], 1))
+  k.emit(s_add_u32(s[71], s[71], s[72]))
+  k.emit(v_cvt_f32_u32_e32(v[18], s[75]))
+  k.emit(v_rcp_iflag_f32_e32(v[18], v[18]))
+  k.emit(v_cvt_f32_u32_e32(v[19], s[69]))
+  k.emit(v_mul_f32_e32(v[18], v[18], v[19]))
+  k.emit(v_cvt_u32_f32_e32(v[18], v[18]))
+  k.emit(v_mul_u32_u24_e64(v[19], v[18], s[75]))
+  k.emit(v_sub_u32_e32(v[19], s[69], v[19]))
+  k.emit(v_cmpx_eq_u32_e64(EXEC, v[19], s[75]))
+  k.emit(v_add_u32_e32(v[18], 1, v[18]))
+  k.emit(s_mov_b64(EXEC, -1))
+  k.emit(v_cmpx_gt_u32_e64(EXEC, v[19], s[75]))
+  k.emit(v_sub_u32_e64(v[18], v[18], 1))
+  k.emit(s_mov_b64(EXEC, -1))
+  k.emit(v_readfirstlane_b32_e32(v[72], v[18]))
+  k.emit(s_mul_i32(s[72], s[72], s[75]))
+  k.emit(s_sub_u32(s[73], s[69], s[72]))
+  k.emit(s_cmp_gt_u32(s[2], s[72]))
+  k.emit(s_cselect_b32(s[72], s[73], s[75]))
+  k.emit(s_lshr_b32(s[72], s[72], 1))
+  k.emit(s_bfm_b32(s[73], 1, 0))
+  k.emit(s_and_b32(s[73], s[2], s[73]))
+  k.emit(s_mul_i32(s[72], s[72], s[73]))
+  k.emit(s_add_u32(s[2], s[71], s[72]))
+  k.label('skip_WGMXCC')
+  k.emit(v_mov_b32_e32(v[20], 256))
+  k.emit(v_mov_b32_e32(v[19], s[20]))
+  k.emit(v_cvt_f32_u32_e32(v[18], v[20]))
+  k.emit(v_rcp_iflag_f32_e32(v[18], v[18]))
+  k.emit(v_cvt_f32_u32_e32(v[21], v[19]))
+  k.emit(v_mul_f32_e32(v[18], v[18], v[21]))
+  k.emit(v_cvt_u32_f32_e32(v[18], v[18]))
+  k.emit(v_mul_u32_u24_e32(v[21], v[18], v[20]))
+  k.emit(v_sub_u32_e32(v[21], v[19], v[21]))
+  k.emit(v_cmp_ne_u32_e64(VCC, v[21], 0))
+  k.emit(v_addc_co_u32(v[18], VCC, v[18], 0, VCC))
+  k.emit(v_mov_b32_e32(v[20], 256))
+  k.emit(v_mov_b32_e32(v[19], s[21]))
+  k.emit(v_readfirstlane_b32_e32(v[10], v[18]))
+  k.emit(v_cvt_f32_u32_e32(v[18], v[20]))
+  k.emit(v_rcp_iflag_f32_e32(v[18], v[18]))
+  k.emit(v_cvt_f32_u32_e32(v[21], v[19]))
+  k.emit(v_mul_f32_e32(v[18], v[18], v[21]))
+  k.emit(v_cvt_u32_f32_e32(v[18], v[18]))
+  k.emit(v_mul_u32_u24_e32(v[21], v[18], v[20]))
+  k.emit(v_sub_u32_e32(v[21], v[19], v[21]))
+  k.emit(v_cmp_ne_u32_e64(VCC, v[21], 0))
+  k.emit(v_addc_co_u32(v[18], VCC, v[18], 0, VCC))
+  k.emit(s_nop())
+  k.emit(v_readfirstlane_b32_e32(v[11], v[18]))
+  k.waitcnt(lgkm=0)
+  k.emit(s_mov_b32(s[85], 84148480))
+  k.emit(s_mov_b32(s[86], 117834498))
+  k.emit(s_sub_u32(s[28], s[28], 16))
+  k.emit(s_subb_u32(s[29], s[29], 0))
+  k.emit(s_sub_u32(s[30], s[30], 16))
+  k.emit(s_subb_u32(s[31], s[31], 0))
+  k.label('AlphaNonZero')
+  k.emit(s_mov_b32(s[57], s[2]))
+  k.emit(s_mul_i32(s[58], s[57], s[46]))
+  k.emit(s_mov_b32(s[59], s[49]))
+  k.emit(s_mul_i32(s[87], s[52], s[46]))
+  k.emit(s_cmp_lt_u32(s[87], s[49]))
+  k.emit(s_cbranch_scc1(), target='SK_InitDone')
+  k.emit(s_mul_i32(s[87], s[52], s[46]))
+  k.emit(s_mul_i32(s[88], s[46], s[51]))
+  k.emit(s_sub_u32(s[87], s[87], s[88]))
+  k.emit(s_mul_i32(s[58], s[57], s[46]))
+  k.emit(s_add_u32(s[58], s[58], s[87]))
+  k.emit(s_add_u32(s[59], s[58], s[46]))
+  k.emit(s_add_u32(s[89], s[46], 1))
+  k.emit(s_mul_i32(s[88], s[57], s[89]))
+  k.emit(s_add_u32(s[89], s[88], s[89]))
+  k.emit(s_cmp_lt_u32(s[57], s[87]))
+  k.emit(s_cselect_b32(s[58], s[88], s[58]))
+  k.emit(s_cselect_b32(s[59], s[89], s[59]))
+  k.emit(s_mul_i32(s[87], s[52], s[46]))
+  k.emit(s_min_u32(s[59], s[59], s[87]))
+  k.label('SK_InitDone')
+  k.emit(s_cmp_ge_u32(s[58], s[49]))
+  k.emit(s_cbranch_scc1(), target='KernelEnd')
+  k.label('PersistentLoopStart')
+  k.emit(v_xor_b32_e32(v[18], v[178], v[16]))
+  k.emit(v_min_i32_e32(v[16], v[16], v[18]))
+  k.emit(v_xor_b32_e32(v[18], v[179], v[17]))
+  k.emit(v_min_i32_e32(v[17], v[17], v[18]))
+  k.emit(s_mul_hi_u32(s[89], s[58], s[47]))
+  k.emit(s_lshr_b32(s[90], s[48], 31))
+  k.emit(s_mul_i32(s[88], s[58], s[90]))
+  k.emit(s_add_u32(s[88], s[88], s[89]))
+  k.emit(s_and_b32(s[90], s[48], 2147483647))
+  k.emit(s_lshr_b32(s[88], s[88], s[90]))
+  k.emit(s_mul_i32(s[89], s[88], s[46]))
+  k.emit(s_add_u32(s[90], s[89], s[46]))
+  k.emit(s_sub_u32(s[60], s[58], s[89]))
+  k.emit(s_min_u32(s[61], s[59], s[90]))
+  k.emit(s_sub_u32(s[61], s[61], s[89]))
+  k.emit(s_mul_i32(s[91], s[52], s[46]))
+  k.emit(s_sub_u32(s[91], s[49], s[91]))
+  k.emit(s_mul_i32(s[89], s[51], s[46]))
+  k.emit(s_add_u32(s[89], s[89], s[58]))
+  k.emit(s_cmp_lt_u32(s[89], s[91]))
+  k.emit(s_cbranch_scc1(), target='NoBranch_8G3ZEUE1ZDJOP9IU')
+  k.emit(s_mov_b32(s[89], s[90]))
+  k.emit(s_cmp_le_u32(s[91], s[58]))
+  k.emit(s_cbranch_scc1(), target='NoBranch_8G3ZEUE1ZDJOP9IU')
+  k.emit(s_mul_i32(s[87], s[52], s[46]))
+  k.emit(s_mul_i32(s[92], s[46], s[51]))
+  k.emit(s_sub_u32(s[87], s[87], s[92]))
+  k.emit(s_mul_i32(s[58], s[57], s[46]))
+  k.emit(s_add_u32(s[58], s[58], s[87]))
+  k.emit(s_add_u32(s[59], s[58], s[46]))
+  k.emit(s_add_u32(s[93], s[46], 1))
+  k.emit(s_mul_i32(s[92], s[57], s[93]))
+  k.emit(s_add_u32(s[93], s[92], s[93]))
+  k.emit(s_cmp_lt_u32(s[57], s[87]))
+  k.emit(s_cselect_b32(s[58], s[92], s[58]))
+  k.emit(s_cselect_b32(s[59], s[93], s[59]))
+  k.emit(s_add_u32(s[89], s[58], s[91]))
+  k.emit(s_add_u32(s[59], s[59], s[91]))
+  k.emit(s_min_u32(s[59], s[59], s[49]))
+  k.emit(s_cmp_ge_u32(s[58], s[49]))
+  k.emit(s_cbranch_scc1(), target='KernelEnd')
+  k.label('NoBranch_8G3ZEUE1ZDJOP9IU')
+  k.emit(s_mov_b32(s[58], s[89]))
+  k.emit(s_mul_i32(s[89], s[10], s[11]))
+  k.emit(v_cvt_f32_u32_e32(v[18], s[89]))
+  k.emit(v_rcp_iflag_f32_e32(v[18], v[18]))
+  k.emit(v_cvt_f32_u32_e32(v[19], s[88]))
+  k.emit(v_mul_f32_e32(v[18], v[18], v[19]))
+  k.emit(v_cvt_u32_f32_e32(v[18], v[18]))
+  k.emit(v_mul_u32_u24_e64(v[19], v[18], s[89]))
+  k.emit(v_sub_u32_e32(v[19], s[88], v[19]))
+  k.emit(v_cmpx_eq_u32_e64(EXEC, v[19], s[89]))
+  k.emit(v_add_u32_e32(v[18], 1, v[18]))
+  k.emit(v_mov_b32_e32(v[19], 0))
+  k.emit(s_mov_b64(EXEC, -1))
+  k.emit(v_cmpx_gt_u32_e64(EXEC, v[19], s[89]))
+  k.emit(v_sub_u32_e64(v[18], v[18], 1))
+  k.emit(v_mul_u32_u24_e64(v[19], v[18], s[89]))
+  k.emit(v_sub_u32_e32(v[19], s[88], v[19]))
+  k.emit(s_mov_b64(EXEC, -1))
+  k.emit(v_readfirstlane_b32_e32(v[4], v[18]))
+  k.emit(v_readfirstlane_b32_e32(v[90], v[19]))
+  k.emit(v_cvt_f32_u32_e32(v[18], s[10]))
+  k.emit(v_rcp_iflag_f32_e32(v[18], v[18]))
+  k.emit(v_cvt_f32_u32_e32(v[19], s[90]))
+  k.emit(v_mul_f32_e32(v[18], v[18], v[19]))
+  k.emit(v_cvt_u32_f32_e32(v[18], v[18]))
+  k.emit(v_mul_u32_u24_e64(v[19], v[18], s[10]))
+  k.emit(v_sub_u32_e32(v[19], s[90], v[19]))
+  k.emit(v_cmpx_eq_u32_e64(EXEC, v[19], s[10]))
+  k.emit(v_add_u32_e32(v[18], 1, v[18]))
+  k.emit(v_mov_b32_e32(v[19], 0))
+  k.emit(s_mov_b64(EXEC, -1))
+  k.emit(v_cmpx_gt_u32_e64(EXEC, v[19], s[10]))
+  k.emit(v_sub_u32_e64(v[18], v[18], 1))
+  k.emit(v_mul_u32_u24_e64(v[19], v[18], s[10]))
+  k.emit(v_sub_u32_e32(v[19], s[90], v[19]))
+  k.emit(s_mov_b64(EXEC, -1))
+  k.emit(v_readfirstlane_b32_e32(v[3], v[18]))
+  k.emit(v_readfirstlane_b32_e32(v[2], v[19]))
+  k.label('SKAlphaCheck')
+  k.emit(s_mov_b32(s[91], 16))
+  k.emit(v_cvt_f32_u32_e32(v[18], s[91]))
+  k.emit(v_rcp_iflag_f32_e32(v[18], v[18]))
+  k.emit(v_cvt_f32_u32_e32(v[19], s[3]))
+  k.emit(v_mul_f32_e32(v[18], v[18], v[19]))
+  k.emit(v_cvt_u32_f32_e32(v[18], v[18]))
+  k.emit(v_mul_u32_u24_e64(v[19], v[18], s[91]))
+  k.emit(v_sub_u32_e32(v[19], s[3], v[19]))
+  k.emit(v_cmpx_eq_u32_e64(EXEC, v[19], s[91]))
+  k.emit(v_add_u32_e32(v[18], 1, v[18]))
+  k.emit(s_mov_b64(EXEC, -1))
+  k.emit(v_cmpx_gt_u32_e64(EXEC, v[19], s[91]))
+  k.emit(v_sub_u32_e64(v[18], v[18], 1))
+  k.emit(s_mov_b64(EXEC, -1))
+  k.emit(v_readfirstlane_b32_e32(v[87], v[18]))
+  k.emit(s_mul_i32(s[90], s[87], s[91]))
+  k.emit(s_sub_u32(s[90], s[3], s[90]))
+  k.emit(s_mul_i32(s[90], s[90], s[10]))
+  k.emit(s_add_u32(s[90], s[90], s[2]))
+  k.emit(v_cvt_f32_u32_e32(v[18], s[91]))
+  k.emit(v_rcp_iflag_f32_e32(v[18], v[18]))
+  k.emit(v_cvt_f32_u32_e32(v[19], s[11]))
+  k.emit(v_mul_f32_e32(v[18], v[18], v[19]))
+  k.emit(v_cvt_u32_f32_e32(v[18], v[18]))
+  k.emit(v_mul_u32_u24_e64(v[19], v[18], s[91]))
+  k.emit(v_sub_u32_e32(v[19], s[11], v[19]))
+  k.emit(v_cmpx_eq_u32_e64(EXEC, v[19], s[91]))
+  k.emit(v_add_u32_e32(v[18], 1, v[18]))
+  k.emit(s_mov_b64(EXEC, -1))
+  k.emit(v_cmpx_gt_u32_e64(EXEC, v[19], s[91]))
+  k.emit(v_sub_u32_e64(v[18], v[18], 1))
+  k.emit(s_mov_b64(EXEC, -1))
+  k.emit(v_readfirstlane_b32_e32(v[88], v[18]))
+  k.emit(s_mul_i32(s[89], s[91], s[88]))
+  k.emit(s_sub_u32(s[89], s[11], s[89]))
+  k.emit(s_cmp_eq_u32(s[89], 0))
+  k.emit(s_cmov_b32(s[89], s[91]))
+  k.emit(s_cmp_ge_u32(s[87], s[88]))
+  k.emit(s_cselect_b32(s[88], s[89], s[91]))
+  k.emit(v_cvt_f32_u32_e32(v[18], s[88]))
+  k.emit(v_rcp_iflag_f32_e32(v[18], v[18]))
+  k.emit(v_cvt_f32_u32_e32(v[19], s[90]))
+  k.emit(v_mul_f32_e32(v[18], v[18], v[19]))
+  k.emit(v_cvt_u32_f32_e32(v[18], v[18]))
+  k.emit(v_mul_u32_u24_e64(v[19], v[18], s[88]))
+  k.emit(v_sub_u32_e32(v[19], s[90], v[19]))
+  k.emit(v_cmpx_eq_u32_e64(EXEC, v[19], s[88]))
+  k.emit(v_add_u32_e32(v[18], 1, v[18]))
+  k.emit(v_mov_b32_e32(v[19], 0))
+  k.emit(s_mov_b64(EXEC, -1))
+  k.emit(v_cmpx_gt_u32_e64(EXEC, v[19], s[88]))
+  k.emit(v_sub_u32_e64(v[18], v[18], 1))
+  k.emit(v_mul_u32_u24_e64(v[19], v[18], s[88]))
+  k.emit(v_sub_u32_e32(v[19], s[90], v[19]))
+  k.emit(s_mov_b64(EXEC, -1))
+  k.emit(v_readfirstlane_b32_e32(v[2], v[18]))
+  k.emit(v_readfirstlane_b32_e32(v[3], v[19]))
+  k.emit(s_mul_i32(s[3], s[2], s[88]))
+  k.emit(s_sub_u32(s[3], s[90], s[3]))
+  k.emit(s_mul_i32(s[87], s[87], s[91]))
+  k.emit(s_add_u32(s[3], s[3], s[87]))
+  k.label('WGM')
+  k.emit(v_and_b32_e32(v[19], 63, v[180]))
+  k.emit(v_and_b32_e32(v[18], 15, v[19]))
+  k.emit(v_lshlrev_b32_e32(v[18], 3, v[18]))
+  k.emit(v_lshrrev_b32_e32(v[19], 4, v[19]))
+  k.emit(v_lshl_add_u32_e64(v[18], v[19], 11, v[18]))
+  k.emit(v_lshrrev_b32_e32(v[22], 6, v[180]))
+  k.emit(v_and_b32_e32(v[22], 1, v[22]))
+  k.emit(v_lshl_add_u32_e64(v[18], v[22], 7, v[18]))
+  k.emit(v_and_b32_e32(v[20], 63, v[180]))
+  k.emit(v_and_b32_e32(v[19], 15, v[20]))
+  k.emit(v_lshlrev_b32_e32(v[19], 6, v[19]))
+  k.emit(v_lshlrev_b32_e32(v[19], 3, v[19]))
+  k.emit(v_lshrrev_b32_e32(v[20], 4, v[20]))
+  k.emit(v_lshl_add_u32_e64(v[19], v[20], 3, v[19]))
+  k.emit(v_lshrrev_b32_e32(v[21], 7, v[180]))
+  k.emit(v_and_b32_e32(v[21], 1, v[21]))
+  k.emit(v_lshl_add_u32_e64(v[19], v[21], 13, v[19]))
+  k.emit(v_lshrrev_b32_e32(v[20], 6, v[180]))
+  k.emit(v_lshrrev_b32_e32(v[20], 2, v[20]))
+  k.emit(s_mov_b32(s[87], 16384))
+  k.emit(v_mul_lo_u32(v[20], s[87], v[20]))
+  k.emit(v_add_lshl_u32_e64(v[16], v[20], v[18], 1))
+  k.emit(v_lshrrev_b32_e32(v[18], 6, v[180]))
+  k.emit(v_lshrrev_b32_e32(v[18], 2, v[18]))
+  k.emit(s_mov_b32(s[87], 64))
+  k.emit(v_mul_lo_u32(v[18], s[87], v[18]))
+  k.emit(v_add_lshl_u32_e64(v[17], v[18], v[19], 1))
+  k.emit(v_lshrrev_b32_e32(v[20], 10, v[17]))
+  k.emit(v_lshl_add_u32_e64(v[17], v[20], 5, v[17]))
+  k.emit(v_add_co_u32_e32(v[17], 32768, v[17]))
+  k.emit(v_add_u32_e32(v[178], 66560, v[16]))
+  k.emit(v_xor_b32_e32(v[178], v[178], v[16]))
+  k.emit(v_add_u32_e32(v[179], 66560, v[17]))
+  k.emit(v_xor_b32_e32(v[179], v[179], v[17]))
+  k.emit(v_lshrrev_b32_e32(v[19], 5, v[180]))
+  k.emit(v_and_b32_e32(v[18], 31, v[180]))
+  k.emit(v_lshlrev_b32_e32(v[18], 3, v[18]))
+  k.emit(v_mov_b32_e32(v[22], v[19]))
+  k.emit(v_lshrrev_b32_e32(v[20], 3, v[180]))
+  k.emit(v_and_b32_e32(v[21], 7, v[180]))
+  k.emit(v_lshlrev_b32_e32(v[21], 3, v[21]))
+  k.emit(v_mov_b32_e32(v[23], v[21]))
+  k.emit(v_mul_u32_u24_e32(v[24], 256, v[22]))
+  k.emit(v_add_lshl_u32_e64(v[24], v[18], v[24], 1))
+  k.emit(s_nop())
+  k.emit(v_readfirstlane_b32_e32(v[53], v[24]))
+  k.emit(s_nop())
+  k.emit(s_add_u32(s[55], s[53], 66560))
+  k.emit(s_xor_b32(s[55], s[55], s[53]))
+  k.emit(v_mul_u32_u24_e32(v[24], 64, v[20]))
+  k.emit(v_add_lshl_u32_e64(v[24], v[23], v[24], 1))
+  k.emit(v_lshrrev_b32_e32(v[26], 10, v[24]))
+  k.emit(v_lshl_add_u32_e64(v[24], v[26], 5, v[24]))
+  k.emit(v_add_co_u32_e32(v[24], 32768, v[24]))
+  k.emit(s_nop())
+  k.emit(v_readfirstlane_b32_e32(v[54], v[24]))
+  k.emit(s_nop())
+  k.emit(s_add_u32(s[56], s[54], 66560))
+  k.emit(s_xor_b32(s[56], s[56], s[54]))
+  k.emit(v_mov_b32_e32(v[24], v[18]))
+  k.emit(v_mov_b32_e32(v[25], v[20]))
+  k.emit(v_add_co_u32_e32(v[26], 32, v[25]))
+  k.emit(v_add_co_u32_e32(v[27], 32, v[26]))
+  k.emit(v_add_co_u32_e32(v[28], 32, v[27]))
+  k.emit(v_add_co_u32_e32(v[29], 32, v[28]))
+  k.emit(v_add_co_u32_e32(v[30], 32, v[29]))
+  k.emit(v_add_co_u32_e32(v[31], 32, v[30]))
+  k.emit(v_add_co_u32_e32(v[32], 32, v[31]))
+  k.emit(v_mov_b32_e32(v[33], v[19]))
+  k.emit(v_add_co_u32_e32(v[34], 8, v[33]))
+  k.emit(v_add_co_u32_e32(v[35], 8, v[34]))
+  k.emit(v_add_co_u32_e32(v[36], 8, v[35]))
+  k.emit(v_add_co_u32_e32(v[37], 8, v[36]))
+  k.emit(v_add_co_u32_e32(v[38], 8, v[37]))
+  k.emit(v_add_co_u32_e32(v[39], 8, v[38]))
+  k.emit(v_add_co_u32_e32(v[40], 8, v[39]))
+  k.emit(v_mov_b32_e32(v[41], v[21]))
+  k.emit(s_mul_i32(s[87], s[2], 256))
+  k.emit(s_sub_u32(s[87], s[20], s[87]))
+  k.emit(s_sub_u32(s[87], s[87], 8))
+  k.emit(v_mov_b32_e32(v[42], s[87]))
+  k.emit(v_min_i32_e32(v[24], v[42], v[24]))
+  k.emit(v_mul_lo_u32(v[42], s[40], v[33]))
+  k.emit(v_add_co_u32_e32(v[0], v[24], v[42]))
+  k.emit(v_add_u32_e32(v[0], 8))
+  k.emit(v_lshlrev_b32_e32(v[0], 1))
+  k.emit(v_mul_lo_u32(v[42], s[40], v[34]))
+  k.emit(v_add_co_u32_e32(v[1], v[24], v[42]))
+  k.emit(v_add_u32_e32(v[1], 8, v[1]))
+  k.emit(v_lshlrev_b32_e32(v[1], 1, v[1]))
+  k.emit(v_mul_lo_u32(v[42], s[40], v[35]))
+  k.emit(v_add_co_u32_e32(v[2], v[24], v[42]))
+  k.emit(v_add_u32_e32(v[2], 8, v[2]))
+  k.emit(v_lshlrev_b32_e32(v[2], 1, v[2]))
+  k.emit(v_mul_lo_u32(v[42], s[40], v[36]))
+  k.emit(v_add_co_u32_e32(v[3], v[24], v[42]))
+  k.emit(v_add_u32_e32(v[3], 8, v[3]))
+  k.emit(v_lshlrev_b32_e32(v[3], 1, v[3]))
+  k.emit(v_mul_lo_u32(v[42], s[40], v[37]))
+  k.emit(v_add_co_u32_e32(v[4], v[24], v[42]))
+  k.emit(v_add_u32_e32(v[4], 8, v[4]))
+  k.emit(v_lshlrev_b32_e32(v[4], 1, v[4]))
+  k.emit(v_mul_lo_u32(v[42], s[40], v[38]))
+  k.emit(v_add_co_u32_e32(v[5], v[24], v[42]))
+  k.emit(v_add_u32_e32(v[5], 8, v[5]))
+  k.emit(v_lshlrev_b32_e32(v[5], 1, v[5]))
+  k.emit(v_mul_lo_u32(v[42], s[40], v[39]))
+  k.emit(v_add_co_u32_e32(v[6], v[24], v[42]))
+  k.emit(v_add_u32_e32(v[6], 8, v[6]))
+  k.emit(v_lshlrev_b32_e32(v[6], 1, v[6]))
+  k.emit(v_mul_lo_u32(v[42], s[40], v[40]))
+  k.emit(v_add_co_u32_e32(v[7], v[24], v[42]))
+  k.emit(v_add_u32_e32(v[7], 8, v[7]))
+  k.emit(v_lshlrev_b32_e32(v[7], 1, v[7]))
+  k.emit(v_mul_lo_u32(v[33], s[42], v[25]))
+  k.emit(v_add_co_u32_e32(v[8], v[41], v[33]))
+  k.emit(v_add_u32_e32(v[8], 8, v[8]))
+  k.emit(v_lshlrev_b32_e32(v[8], 1, v[8]))
+  k.emit(v_mul_lo_u32(v[33], s[42], v[26]))
+  k.emit(v_add_co_u32_e32(v[9], v[41], v[33]))
+  k.emit(v_add_u32_e32(v[9], 8, v[9]))
+  k.emit(v_lshlrev_b32_e32(v[9], 1, v[9]))
+  k.emit(v_mul_lo_u32(v[33], s[42], v[27]))
+  k.emit(v_add_co_u32_e32(v[10], v[41], v[33]))
+  k.emit(v_add_u32_e32(v[10], 8, v[10]))
+  k.emit(v_lshlrev_b32_e32(v[10], 1, v[10]))
+  k.emit(v_mul_lo_u32(v[33], s[42], v[28]))
+  k.emit(v_add_co_u32_e32(v[11], v[41], v[33]))
+  k.emit(v_add_u32_e32(v[11], 8, v[11]))
+  k.emit(v_lshlrev_b32_e32(v[11], 1, v[11]))
+  k.emit(v_mul_lo_u32(v[33], s[42], v[29]))
+  k.emit(v_add_co_u32_e32(v[12], v[41], v[33]))
+  k.emit(v_add_u32_e32(v[12], 8, v[12]))
+  k.emit(v_lshlrev_b32_e32(v[12], 1, v[12]))
+  k.emit(v_mul_lo_u32(v[33], s[42], v[30]))
+  k.emit(v_add_co_u32_e32(v[13], v[41], v[33]))
+  k.emit(v_add_u32_e32(v[13], 8, v[13]))
+  k.emit(v_lshlrev_b32_e32(v[13], 1, v[13]))
+  k.emit(v_mul_lo_u32(v[33], s[42], v[31]))
+  k.emit(v_add_co_u32_e32(v[14], v[41], v[33]))
+  k.emit(v_add_u32_e32(v[14], 8, v[14]))
+  k.emit(v_lshlrev_b32_e32(v[14], 1, v[14]))
+  k.emit(v_mul_lo_u32(v[33], s[42], v[32]))
+  k.emit(v_add_co_u32_e32(v[15], v[41], v[33]))
+  k.emit(v_add_u32_e32(v[15], 8, v[15]))
+  k.emit(v_lshlrev_b32_e32(v[15], 1, v[15]))
+  k.emit(s_mul_hi_u32(s[91], s[2], 256))
+  k.emit(s_mul_i32(s[90], s[2], 256))
+  k.emit(s_mul_i32(s[88], s[60], 64))
+  k.emit(s_mul_hi_u32(s[89], s[88], s[40]))
+  k.emit(s_mul_i32(s[88], s[88], s[40]))
+  k.emit(s_add_u32(s[90], s[90], s[88]))
+  k.emit(s_addc_u32(s[91], s[91], s[89]))
+  k.emit(s_mov_b64(s[62:63], 1))
+  k.emit(s_sub_u32(s[88], s[20], 1))
+  k.emit(s_mul_hi_u32(s[89], 1, s[88]))
+  k.emit(s_mul_i32(s[88], 1, s[88]))
+  k.emit(s_add_u32(s[62], s[62], s[88]))
+  k.emit(s_addc_u32(s[63], s[63], s[89]))
+  k.emit(s_sub_u32(s[88], s[23], 1))
+  k.emit(s_mul_hi_u32(s[89], s[40], s[88]))
+  k.emit(s_mul_i32(s[88], s[40], s[88]))
+  k.emit(s_add_u32(s[62], s[62], s[88]))
+  k.emit(s_addc_u32(s[63], s[63], s[89]))
+  k.emit(s_sub_u32(s[62], s[62], s[90]))
+  k.emit(s_subb_u32(s[63], s[63], s[91]))
+  k.emit(s_lshl_b64(s[62:63], s[62:63], 1))
+  k.emit(s_add_u32(s[62], s[62], 16))
+  k.emit(s_addc_u32(s[63], s[63], 0))
+  k.emit(s_cmp_eq_u32(s[63], 0))
+  k.emit(s_cselect_b32(s[70], s[62], -1))
+  k.emit(s_mul_hi_u32(s[89], s[41], s[4]))
+  k.emit(s_mul_i32(s[88], s[41], s[4]))
+  k.emit(s_add_u32(s[90], s[90], s[88]))
+  k.emit(s_addc_u32(s[91], s[91], s[89]))
+  k.emit(s_lshl_b64(s[90:91], s[90:91], 1))
+  k.emit(s_add_u32(s[68], s[28], s[90]))
+  k.emit(s_addc_u32(s[69], s[29], s[91]))
+  k.emit(s_mov_b32(s[71], 131072))
+  k.emit(s_mul_hi_u32(s[91], s[3], 256))
+  k.emit(s_mul_i32(s[90], s[3], 256))
+  k.emit(s_mul_hi_u32(s[91], s[90], s[42]))
+  k.emit(s_mul_i32(s[90], s[90], s[42]))
+  k.emit(s_mul_i32(s[88], s[60], 64))
+  k.emit(s_mul_hi_u32(s[89], s[88], 1))
+  k.emit(s_mul_i32(s[88], s[88], 1))
+  k.emit(s_add_u32(s[90], s[90], s[88]))
+  k.emit(s_addc_u32(s[91], s[91], s[89]))
+  k.emit(s_mov_b64(s[76:77], 1))
+  k.emit(s_sub_u32(s[88], s[23], 1))
+  k.emit(s_mul_hi_u32(s[89], 1, s[88]))
+  k.emit(s_mul_i32(s[88], 1, s[88]))
+  k.emit(s_add_u32(s[76], s[76], s[88]))
+  k.emit(s_addc_u32(s[77], s[77], s[89]))
+  k.emit(s_sub_u32(s[88], s[21], 1))
+  k.emit(s_mul_hi_u32(s[89], s[42], s[88]))
+  k.emit(s_mul_i32(s[88], s[42], s[88]))
+  k.emit(s_add_u32(s[76], s[76], s[88]))
+  k.emit(s_addc_u32(s[77], s[77], s[89]))
+  k.emit(s_sub_u32(s[76], s[76], s[90]))
+  k.emit(s_subb_u32(s[77], s[77], s[91]))
+  k.emit(s_lshl_b64(s[76:77], s[76:77], 1))
+  k.emit(s_add_u32(s[76], s[76], 16))
+  k.emit(s_addc_u32(s[77], s[77], 0))
+  k.emit(s_cmp_eq_u32(s[77], 0))
+  k.emit(s_cselect_b32(s[74], s[76], -1))
+  k.emit(s_mul_hi_u32(s[89], s[43], s[4]))
+  k.emit(s_mul_i32(s[88], s[43], s[4]))
+  k.emit(s_add_u32(s[90], s[90], s[88]))
+  k.emit(s_addc_u32(s[91], s[91], s[89]))
+  k.emit(s_lshl_b64(s[90:91], s[90:91], 1))
+  k.emit(s_add_u32(s[72], s[30], s[90]))
+  k.emit(s_addc_u32(s[73], s[31], s[91]))
+  k.emit(s_mov_b32(s[75], 131072))
+  k.emit(s_mul_i32(s[83], 128, s[40]))
+  k.emit(s_mov_b32(s[84], 128))
+  k.emit(s_sub_u32(s[8], s[61], s[60]))
+  k.label('SKAlphaCheck2')
+  k.emit(s_and_b32(s[89], 63, s[23]))
+  k.emit(s_cmp_eq_u32(s[89], 0))
+  k.emit(s_cselect_b32(s[88], 0, 1))
+  k.emit(s_cmp_eq_u32(s[61], s[46]))
+  k.emit(s_cselect_b32(s[88], s[88], 0))
+  k.emit(s_sub_u32(s[8], s[8], s[88]))
+  k.emit(s_mov_b32(s[9], s[8]))
+  k.emit(s_and_b32(s[90], s[6], 7936))
+  k.emit(s_lshr_b32(s[90], s[90], 8))
+  k.emit(s_and_b32(s[91], s[6], 57344))
+  k.emit(s_and_b32(s[6], s[6], 255))
+  k.emit(s_mov_b32(s[88], s[6]))
+  k.label('beginStaggerUIter')
+  k.emit(s_lshl_b32(s[89], s[88], s[90]))
+  k.emit(s_cmp_ge_u32(s[9], s[89]))
+  k.emit(s_cbranch_scc1(), target='endStaggerUIter')
+  k.emit(s_lshr_b32(s[88], s[88], 1))
+  k.emit(s_branch(), target='beginStaggerUIter')
+  k.label('endStaggerUIter')
+  k.emit(s_sub_u32(s[89], s[88], 1))
+  k.emit(s_cmp_ge_u32(s[88], 1))
+  k.emit(s_cselect_b32(s[78], s[89], 0))
+  k.emit(s_cmp_eq_u32(s[91], 0))
+  k.emit(s_cbranch_scc1(), target='StaggerUMapping_1')
+  k.emit(s_mov_b32(s[88], s[2]))
+  k.emit(s_branch(), target='staggerInputEnd')
+  k.label('StaggerUMapping_1')
+  k.emit(s_cmp_eq_u32(s[91], 8192))
+  k.emit(s_cbranch_scc1(), target='StaggerUMapping_2')
+  k.emit(s_mov_b32(s[88], s[3]))
+  k.emit(s_branch(), target='staggerInputEnd')
+  k.label('StaggerUMapping_2')
+  k.emit(s_cmp_eq_u32(s[91], 16384))
+  k.emit(s_cbranch_scc1(), target='StaggerUMapping_3')
+  k.emit(s_mov_b32(s[88], -1))
+  k.emit(s_branch(), target='staggerInputEnd')
+  k.label('StaggerUMapping_3')
+  k.emit(s_cmp_eq_u32(s[91], 24576))
+  k.emit(s_cbranch_scc1(), target='StaggerUMapping_4')
+  k.emit(s_mul_i32(s[89], s[10], s[3]))
+  k.emit(s_add_u32(s[88], s[88], s[89]))
+  k.emit(s_add_u32(s[88], s[88], s[2]))
+  k.emit(s_branch(), target='staggerInputEnd')
+  k.label('StaggerUMapping_4')
+  k.emit(s_cmp_eq_u32(s[91], 32768))
+  k.emit(s_cbranch_scc1(), target='staggerInputEnd')
+  k.emit(s_mov_b32(s[88], -1))
+  k.emit(s_branch(), target='staggerInputEnd')
+  k.label('staggerInputEnd')
+  k.emit(s_and_b32(s[78], s[78], s[88]))
+  k.emit(s_lshl_b32(s[78], s[78], s[90]))
+  k.emit(s_cmp_gt_u32(s[60], 0))
+  k.emit(s_cmov_b32(s[78], 0))
+  k.emit(s_cmp_lt_u32(s[61], s[46]))
+  k.emit(s_cmov_b32(s[78], 0))
+  k.emit(s_mul_hi_i32(s[89], s[78], s[83]))
+  k.emit(s_mul_i32(s[88], s[78], s[83]))
+  k.emit(s_mul_hi_i32(s[80], s[8], s[83]))
+  k.emit(s_mul_i32(s[79], s[8], s[83]))
+  k.emit(s_sub_u32(s[79], s[83], s[79]))
+  k.emit(s_subb_u32(s[80], 0, s[80]))
+  k.emit(s_add_u32(s[68], s[68], s[88]))
+  k.emit(s_addc_u32(s[69], s[69], s[89]))
+  k.emit(s_sub_u32(s[62], s[62], s[88]))
+  k.emit(s_subb_u32(s[63], s[63], s[89]))
+  k.emit(s_cmp_eq_u32(s[63], 0))
+  k.emit(s_cselect_b32(s[70], s[62], -1))
+  k.emit(s_mul_hi_i32(s[89], s[78], s[84]))
+  k.emit(s_mul_i32(s[88], s[78], s[84]))
+  k.emit(s_mul_hi_i32(s[82], s[8], s[84]))
+  k.emit(s_mul_i32(s[81], s[8], s[84]))
+  k.emit(s_sub_u32(s[81], s[84], s[81]))
+  k.emit(s_subb_u32(s[82], 0, s[82]))
+  k.emit(s_add_u32(s[72], s[72], s[88]))
+  k.emit(s_addc_u32(s[73], s[73], s[89]))
+  k.emit(s_sub_u32(s[76], s[76], s[88]))
+  k.emit(s_subb_u32(s[77], s[77], s[89]))
+  k.emit(s_cmp_eq_u32(s[77], 0))
+  k.emit(s_cselect_b32(s[74], s[76], -1))
+  k.emit(s_add_u32(s[78], s[78], 2))
+  k.emit(s_cmp_eq_u32(s[8], 0))
+  k.emit(s_setprio())
+  k.emit(s_cbranch_scc1(), target='ShadowInitStart')
+  k.emit(s_mov_b32(M0, s[53]))
+  k.waitcnt(lgkm=0)
+  k.emit(s_barrier())
+  k.emit(buffer_load_dwordx4(v[0:3], v[0], s[68:71], 0, 0, 1, 0, 0, 1, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4096))
+  k.emit(buffer_load_dwordx4(v[0:3], v[1], s[68:71], 0, 0, 1, 0, 0, 1, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4096))
+  k.emit(buffer_load_dwordx4(v[0:3], v[2], s[68:71], 0, 0, 1, 0, 0, 1, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4096))
+  k.emit(buffer_load_dwordx4(v[0:3], v[3], s[68:71], 0, 0, 1, 0, 0, 1, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4096))
+  k.emit(buffer_load_dwordx4(v[0:3], v[4], s[68:71], 0, 0, 1, 0, 0, 1, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4096))
+  k.emit(buffer_load_dwordx4(v[0:3], v[5], s[68:71], 0, 0, 1, 0, 0, 1, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4096))
+  k.emit(buffer_load_dwordx4(v[0:3], v[6], s[68:71], 0, 0, 1, 0, 0, 1, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4096))
+  k.emit(buffer_load_dwordx4(v[0:3], v[7], s[68:71], 0, 0, 1, 0, 0, 1, 0, 1))
+  k.emit(s_mov_b32(M0, 133120))
+  k.emit(s_mov_b32(M0, s[54]))
+  k.emit(buffer_load_dwordx4(v[0:3], v[8], s[72:75], 0, 0, 1, 0, 0, 0, 1, 1, 1))
+  k.emit(s_add_u32(M0, M0, 4224))
+  k.emit(buffer_load_dwordx4(v[0:3], v[9], s[72:75], 0, 0, 1, 0, 0, 0, 1, 1, 1))
+  k.emit(s_add_u32(M0, M0, 4224))
+  k.emit(buffer_load_dwordx4(v[0:3], v[10], s[72:75], 0, 0, 1, 0, 0, 0, 1, 1, 1))
+  k.emit(s_add_u32(M0, M0, 4224))
+  k.emit(buffer_load_dwordx4(v[0:3], v[11], s[72:75], 0, 0, 1, 0, 0, 0, 1, 1, 1))
+  k.emit(s_add_u32(M0, M0, 4224))
+  k.emit(buffer_load_dwordx4(v[0:3], v[12], s[72:75], 0, 0, 1, 0, 0, 0, 1, 1, 1))
+  k.emit(s_add_u32(M0, M0, 4224))
+  k.emit(buffer_load_dwordx4(v[0:3], v[13], s[72:75], 0, 0, 1, 0, 0, 0, 1, 1, 1))
+  k.emit(s_add_u32(M0, M0, 4224))
+  k.emit(buffer_load_dwordx4(v[0:3], v[14], s[72:75], 0, 0, 1, 0, 0, 0, 1, 1, 1))
+  k.emit(s_add_u32(M0, M0, 4224))
+  k.emit(buffer_load_dwordx4(v[0:3], v[15], s[72:75], 0, 0, 1, 0, 0, 0, 1, 1, 1))
+  k.emit(s_mov_b32(M0, 133120))
+  k.emit(s_add_u32(s[90], s[8], 1))
+  k.emit(s_cmp_eq_u32(s[78], s[90]))
+  k.emit(s_cselect_b32(s[88], s[79], s[83]))
+  k.emit(s_cselect_b32(s[89], s[80], 0))
+  k.emit(s_add_u32(s[68], s[68], s[88]))
+  k.emit(s_addc_u32(s[69], s[69], s[89]))
+  k.emit(s_sub_u32(s[62], s[62], s[88]))
+  k.emit(s_subb_u32(s[63], s[63], s[89]))
+  k.emit(s_cmp_eq_u32(s[63], 0))
+  k.emit(s_cselect_b32(s[70], s[62], -1))
+  k.emit(s_add_u32(s[90], s[8], 1))
+  k.emit(s_cmp_eq_u32(s[78], s[90]))
+  k.emit(s_cselect_b32(s[88], s[81], s[84]))
+  k.emit(s_cselect_b32(s[89], s[82], 0))
+  k.emit(s_add_u32(s[72], s[72], s[88]))
+  k.emit(s_addc_u32(s[73], s[73], s[89]))
+  k.emit(s_sub_u32(s[76], s[76], s[88]))
+  k.emit(s_subb_u32(s[77], s[77], s[89]))
+  k.emit(s_cmp_eq_u32(s[77], 0))
+  k.emit(s_cselect_b32(s[74], s[76], -1))
+  k.label('ShadowInitStart')
+  k.emit(s_mov_b64(s[12:13], s[24:25]))
+  k.emit(s_mov_b32(s[14], 2147483648))
+  k.emit(s_mov_b32(s[15], 131072))
+  k.emit(s_mov_b64(s[16:17], s[24:25]))
+  k.emit(s_mov_b32(s[18], 2147483648))
+  k.emit(s_mov_b32(s[19], 131072))
+  k.emit(s_mov_b32(s[87], 1))
+  k.emit(s_mov_b32(s[88], 1))
+  k.emit(s_mul_i32(s[92], 256, s[3]))
+  k.emit(s_mul_hi_u32(s[91], s[92], s[38]))
+  k.emit(s_mul_i32(s[90], s[92], s[38]))
+  k.emit(s_lshl_b64(s[90:91], s[90:91], s[87]))
+  k.emit(s_add_u32(s[16], s[26], s[90]))
+  k.emit(s_addc_u32(s[17], s[27], s[91]))
+  k.emit(s_mul_hi_u32(s[91], s[92], s[36]))
+  k.emit(s_mul_i32(s[90], s[92], s[36]))
+  k.emit(s_lshl_b64(s[90:91], s[90:91], s[88]))
+  k.emit(s_add_u32(s[12], s[24], s[90]))
+  k.emit(s_addc_u32(s[13], s[25], s[91]))
+  k.emit(s_mul_hi_u32(s[91], s[4], s[39]))
+  k.emit(s_mul_i32(s[90], s[4], s[39]))
+  k.emit(s_lshl_b64(s[90:91], s[90:91], s[87]))
+  k.emit(s_add_u32(s[16], s[16], s[90]))
+  k.emit(s_addc_u32(s[17], s[17], s[91]))
+  k.emit(s_mul_hi_u32(s[91], s[4], s[37]))
+  k.emit(s_mul_i32(s[90], s[4], s[37]))
+  k.emit(s_lshl_b64(s[90:91], s[90:91], s[88]))
+  k.emit(s_add_u32(s[12], s[12], s[90]))
+  k.emit(s_addc_u32(s[13], s[13], s[91]))
+  k.emit(v_mov_b64_e32(v[182:183], 0))
+  k.emit(v_accvgpr_write(v[0], 0))
+  k.emit(v_accvgpr_write(v[1], 0))
+  k.emit(v_accvgpr_write(v[2], 0))
+  k.emit(v_accvgpr_write(v[3], 0))
+  k.emit(v_accvgpr_write(v[4], 0))
+  k.emit(v_accvgpr_write(v[5], 0))
+  k.emit(v_accvgpr_write(v[6], 0))
+  k.emit(v_accvgpr_write(v[7], 0))
+  k.emit(v_accvgpr_write(v[8], 0))
+  k.emit(v_accvgpr_write(v[9], 0))
+  k.emit(v_accvgpr_write(v[10], 0))
+  k.emit(v_accvgpr_write(v[11], 0))
+  k.emit(v_accvgpr_write(v[12], 0))
+  k.emit(v_accvgpr_write(v[13], 0))
+  k.emit(v_accvgpr_write(v[14], 0))
+  k.emit(v_accvgpr_write(v[15], 0))
+  k.emit(v_mfma_i32_32x32x16_i8(v[16:31], v[182:183], v[182:183], v[0:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_i32_32x32x16_i8(v[32:47], v[182:183], v[182:183], v[0:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_i32_32x32x16_i8(v[48:63], v[182:183], v[182:183], v[0:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_i32_32x32x16_i8(v[64:79], v[182:183], v[182:183], v[0:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_i32_32x32x16_i8(v[80:95], v[182:183], v[182:183], v[0:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_i32_32x32x16_i8(v[96:111], v[182:183], v[182:183], v[0:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_i32_32x32x16_i8(v[112:127], v[182:183], v[182:183], v[0:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_i32_32x32x16_i8(v[128:143], v[182:183], v[182:183], v[0:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_i32_32x32x16_i8(v[144:159], v[182:183], v[182:183], v[0:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_i32_32x32x16_i8(v[160:175], v[182:183], v[182:183], v[0:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_i32_32x32x16_i8(v[176:191], v[182:183], v[182:183], v[0:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_i32_32x32x16_i8(v[192:207], v[182:183], v[182:183], v[0:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_i32_32x32x16_i8(v[208:223], v[182:183], v[182:183], v[0:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_i32_32x32x16_i8(v[224:239], v[182:183], v[182:183], v[0:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_i32_32x32x16_i8(v[240:255], v[182:183], v[182:183], v[0:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_cmp_eq_u32(s[8], 0))
+  k.emit(s_cbranch_scc1(), target='toPGR1end_OrdNLL')
+  k.waitcnt(vm=0)
+  k.emit(s_barrier())
+  k.emit(s_xor_b32(s[53], s[55], s[53]))
+  k.emit(s_xor_b32(s[54], s[56], s[54]))
+  k.emit(s_cmp_eq_u32(s[8], 1))
+  k.emit(s_cbranch_scc1(), target='skipPGR2')
+  k.emit(s_mov_b32(M0, s[53]))
+  k.waitcnt(lgkm=0)
+  k.emit(s_barrier())
+  k.emit(buffer_load_dwordx4(v[0:3], v[0], s[68:71], 0, 0, 1, 0, 0, 1, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4096))
+  k.emit(buffer_load_dwordx4(v[0:3], v[1], s[68:71], 0, 0, 1, 0, 0, 1, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4096))
+  k.emit(buffer_load_dwordx4(v[0:3], v[2], s[68:71], 0, 0, 1, 0, 0, 1, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4096))
+  k.emit(buffer_load_dwordx4(v[0:3], v[3], s[68:71], 0, 0, 1, 0, 0, 1, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4096))
+  k.emit(buffer_load_dwordx4(v[0:3], v[4], s[68:71], 0, 0, 1, 0, 0, 1, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4096))
+  k.emit(buffer_load_dwordx4(v[0:3], v[5], s[68:71], 0, 0, 1, 0, 0, 1, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4096))
+  k.emit(buffer_load_dwordx4(v[0:3], v[6], s[68:71], 0, 0, 1, 0, 0, 1, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4096))
+  k.emit(buffer_load_dwordx4(v[0:3], v[7], s[68:71], 0, 0, 1, 0, 0, 1, 0, 1))
+  k.emit(s_mov_b32(M0, 133120))
+  k.emit(s_mov_b32(M0, s[54]))
+  k.waitcnt(lgkm=0)
+  k.emit(s_barrier())
+  k.emit(buffer_load_dwordx4(v[0:3], v[8], s[72:75], 0, 0, 1, 0, 0, 0, 1, 1, 1))
+  k.emit(s_add_u32(M0, M0, 4224))
+  k.emit(buffer_load_dwordx4(v[0:3], v[9], s[72:75], 0, 0, 1, 0, 0, 0, 1, 1, 1))
+  k.emit(s_add_u32(M0, M0, 4224))
+  k.emit(buffer_load_dwordx4(v[0:3], v[10], s[72:75], 0, 0, 1, 0, 0, 0, 1, 1, 1))
+  k.emit(s_add_u32(M0, M0, 4224))
+  k.emit(buffer_load_dwordx4(v[0:3], v[11], s[72:75], 0, 0, 1, 0, 0, 0, 1, 1, 1))
+  k.emit(s_add_u32(M0, M0, 4224))
+  k.emit(buffer_load_dwordx4(v[0:3], v[12], s[72:75], 0, 0, 1, 0, 0, 0, 1, 1, 1))
+  k.emit(s_add_u32(M0, M0, 4224))
+  k.emit(buffer_load_dwordx4(v[0:3], v[13], s[72:75], 0, 0, 1, 0, 0, 0, 1, 1, 1))
+  k.emit(s_add_u32(M0, M0, 4224))
+  k.emit(buffer_load_dwordx4(v[0:3], v[14], s[72:75], 0, 0, 1, 0, 0, 0, 1, 1, 1))
+  k.emit(s_add_u32(M0, M0, 4224))
+  k.emit(buffer_load_dwordx4(v[0:3], v[15], s[72:75], 0, 0, 1, 0, 0, 0, 1, 1, 1))
+  k.emit(s_mov_b32(M0, 133120))
+  k.emit(s_xor_b32(s[53], s[55], s[53]))
+  k.emit(s_xor_b32(s[54], s[56], s[54]))
+  k.label('skipPGR2')
+  k.emit(s_barrier())
+  k.emit(ds_read_b128(v[82:85], v[16]))
+  k.emit(ds_read_b128(v[86:89], v[16], v[0], v[0], 0, 0, 2))
+  k.emit(ds_read_b128(v[90:93], v[16], v[0], v[0], 0, 0, 4))
+  k.emit(ds_read_b128(v[94:97], v[16], v[0], v[0], 0, 0, 6))
+  k.emit(ds_read_b128(v[98:101], v[16], v[0], v[0], 0, 0, 8))
+  k.emit(ds_read_b128(v[102:105], v[16], v[0], v[0], 0, 0, 10))
+  k.emit(ds_read_b128(v[106:109], v[16], v[0], v[0], 0, 0, 12))
+  k.emit(ds_read_b128(v[110:113], v[16], v[0], v[0], 0, 0, 14))
+  k.emit(ds_read_b128(v[114:117], v[17]))
+  k.emit(ds_read_b128(v[118:121], v[17], v[0], v[0], 0, 128))
+  k.emit(ds_read_b128(v[122:125], v[17], v[0], v[0], 0, 0, 1))
+  k.emit(ds_read_b128(v[126:129], v[17], v[0], v[0], 0, 128, 1))
+  k.emit(ds_read_b128(v[130:133], v[17], v[0], v[0], 0, 0, 2))
+  k.emit(ds_read_b128(v[134:137], v[17], v[0], v[0], 0, 128, 2))
+  k.emit(ds_read_b128(v[138:141], v[17], v[0], v[0], 0, 0, 3))
+  k.emit(ds_read_b128(v[142:145], v[17], v[0], v[0], 0, 128, 3))
+  k.waitcnt(lgkm=0)
+  k.emit(v_perm_b32_e64(v[18], v[86], v[82], s[85]))
+  k.emit(v_perm_b32_e64(v[19], v[94], v[90], s[85]))
+  k.emit(v_perm_b32_e64(v[20], v[102], v[98], s[85]))
+  k.emit(v_perm_b32_e64(v[21], v[110], v[106], s[85]))
+  k.emit(v_perm_b32_e64(v[22], v[86], v[82], s[86]))
+  k.emit(v_perm_b32_e64(v[23], v[94], v[90], s[86]))
+  k.emit(v_perm_b32_e64(v[24], v[102], v[98], s[86]))
+  k.emit(v_perm_b32_e64(v[25], v[110], v[106], s[86]))
+  k.emit(v_perm_b32_e64(v[26], v[87], v[83], s[85]))
+  k.emit(v_perm_b32_e64(v[27], v[95], v[91], s[85]))
+  k.emit(v_perm_b32_e64(v[28], v[103], v[99], s[85]))
+  k.emit(v_perm_b32_e64(v[29], v[111], v[107], s[85]))
+  k.emit(v_perm_b32_e64(v[30], v[87], v[83], s[86]))
+  k.emit(v_perm_b32_e64(v[31], v[95], v[91], s[86]))
+  k.emit(v_perm_b32_e64(v[32], v[103], v[99], s[86]))
+  k.emit(v_perm_b32_e64(v[33], v[111], v[107], s[86]))
+  k.emit(v_perm_b32_e64(v[34], v[88], v[84], s[85]))
+  k.emit(v_perm_b32_e64(v[35], v[96], v[92], s[85]))
+  k.emit(v_perm_b32_e64(v[36], v[104], v[100], s[85]))
+  k.emit(v_perm_b32_e64(v[37], v[112], v[108], s[85]))
+  k.emit(v_perm_b32_e64(v[38], v[88], v[84], s[86]))
+  k.emit(v_perm_b32_e64(v[39], v[96], v[92], s[86]))
+  k.emit(v_perm_b32_e64(v[40], v[104], v[100], s[86]))
+  k.emit(v_perm_b32_e64(v[41], v[112], v[108], s[86]))
+  k.emit(v_perm_b32_e64(v[42], v[89], v[85], s[85]))
+  k.emit(v_perm_b32_e64(v[43], v[97], v[93], s[85]))
+  k.emit(v_perm_b32_e64(v[44], v[105], v[101], s[85]))
+  k.emit(v_perm_b32_e64(v[45], v[113], v[109], s[85]))
+  k.emit(v_perm_b32_e64(v[46], v[89], v[85], s[86]))
+  k.emit(v_perm_b32_e64(v[47], v[97], v[93], s[86]))
+  k.emit(v_perm_b32_e64(v[48], v[105], v[101], s[86]))
+  k.emit(v_perm_b32_e64(v[49], v[113], v[109], s[86]))
+  k.label('openLoopL')
+  k.emit(s_cmp_eq_u32(s[8], 1))
+  k.emit(s_cbranch_scc1(), target='toPGR1')
+  k.emit(s_cmp_le_u32(s[8], 2))
+  k.emit(s_cbranch_scc1(), target='LoopEndL')
+  k.label('LoopBeginL')
+  k.emit(s_getreg_b32(s[87], 260))
+  k.waitcnt(lgkm=0)
+  k.emit(s_cmp_eq_u32(s[87], 0))
+  k.emit(s_cbranch_scc1(), target='LoopBeginL_0')
+  k.emit(s_cmp_eq_u32(s[87], 1))
+  k.emit(s_cbranch_scc1(), target='LoopBeginL_1')
+  k.label('LoopBeginL_0')
+  k.emit(v_mfma_16x16x32(v[0:3], v[114:117], v[18:21], v[0:3], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_cmp_eq_u32(s[8], s[78]))
+  k.emit(ds_read_b128(v[82:85], v[16], v[0], v[0], 0, 0, 64))
+  k.emit(ds_read_b128(v[86:89], v[16], v[0], v[0], 0, 0, 66))
+  k.emit(v_mfma_16x16x32(v[4:7], v[114:117], v[22:25], v[4:7], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_cselect_b32(s[88], s[79], s[83]))
+  k.emit(v_mfma_16x16x32(v[8:11], v[114:117], v[26:29], v[8:11], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_cselect_b32(s[89], s[80], 0))
+  k.emit(ds_read_b128(v[90:93], v[16], v[0], v[0], 0, 0, 68))
+  k.emit(ds_read_b128(v[94:97], v[16], v[0], v[0], 0, 0, 70))
+  k.emit(v_mfma_16x16x32(v[12:15], v[114:117], v[30:33], v[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(s[68], s[68], s[88]))
+  k.emit(v_mfma_16x16x32(v[16:19], v[114:117], v[34:37], v[16:19], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_addc_u32(s[69], s[69], s[89]))
+  k.emit(ds_read_b128(v[98:101], v[16], v[0], v[0], 0, 0, 72))
+  k.emit(ds_read_b128(v[102:105], v[16], v[0], v[0], 0, 0, 74))
+  k.emit(v_mfma_16x16x32(v[20:23], v[114:117], v[38:41], v[20:23], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_sub_u32(s[62], s[62], s[88]))
+  k.emit(v_mfma_16x16x32(v[24:27], v[114:117], v[42:45], v[24:27], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_subb_u32(s[63], s[63], s[89]))
+  k.emit(ds_read_b128(v[106:109], v[16], v[0], v[0], 0, 0, 76))
+  k.emit(ds_read_b128(v[110:113], v[16], v[0], v[0], 0, 0, 78))
+  k.emit(v_mfma_16x16x32(v[28:31], v[114:117], v[46:49], v[28:31], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_cmp_eq_u32(s[63], 0))
+  k.emit(v_mfma_16x16x32(v[32:35], v[118:121], v[18:21], v[32:35], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.waitcnt(lgkm=4)
+  k.emit(s_cselect_b32(s[70], s[62], -1))
+  k.emit(v_perm_b32_e64(v[50], v[86], v[82], s[85]))
+  k.emit(v_perm_b32_e64(v[51], v[94], v[90], s[85]))
+  k.emit(v_mfma_16x16x32(v[36:39], v[118:121], v[22:25], v[36:39], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[146:149], v[17], v[0], v[0], 0, 64))
+  k.emit(v_mfma_16x16x32(v[40:43], v[118:121], v[26:29], v[40:43], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[44:47], v[118:121], v[30:33], v[44:47], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[150:153], v[17], v[0], v[0], 0, 192))
+  k.emit(v_mfma_16x16x32(v[48:51], v[118:121], v[34:37], v[48:51], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.waitcnt(lgkm=1)
+  k.emit(v_mfma_16x16x32(v[52:55], v[118:121], v[38:41], v[52:55], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_barrier())
+  k.emit(v_mfma_16x16x32(v[56:59], v[118:121], v[42:45], v[56:59], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_mov_b32(M0, s[53]))
+  k.emit(buffer_load_dwordx4(v[0:3], v[0], s[68:71], 0, 0, 1, 0, 0, 1, 0, 1))
+  k.emit(v_mfma_16x16x32(v[60:63], v[118:121], v[46:49], v[60:63], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[154:157], v[17], v[0], v[0], 0, 64, 1))
+  k.emit(v_mfma_16x16x32(v[64:67], v[122:125], v[18:21], v[64:67], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[52], v[102], v[98], s[85]))
+  k.emit(v_perm_b32_e64(v[53], v[110], v[106], s[85]))
+  k.emit(v_mfma_16x16x32(v[68:71], v[122:125], v[22:25], v[68:71], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4096))
+  k.emit(buffer_load_dwordx4(v[0:3], v[1], s[68:71], 0, 0, 1, 0, 0, 1, 0, 1))
+  k.emit(v_mfma_16x16x32(v[72:75], v[122:125], v[26:29], v[72:75], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[158:161], v[17], v[0], v[0], 0, 192, 1))
+  k.emit(v_mfma_16x16x32(v[76:79], v[122:125], v[30:33], v[76:79], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[54], v[86], v[82], s[86]))
+  k.emit(v_perm_b32_e64(v[55], v[94], v[90], s[86]))
+  k.emit(v_mfma_16x16x32(v[80:83], v[122:125], v[34:37], v[80:83], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4096))
+  k.emit(buffer_load_dwordx4(v[0:3], v[2], s[68:71], 0, 0, 1, 0, 0, 1, 0, 1))
+  k.emit(v_mfma_16x16x32(v[84:87], v[122:125], v[38:41], v[84:87], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[162:165], v[17], v[0], v[0], 0, 64, 2))
+  k.emit(v_mfma_16x16x32(v[88:91], v[122:125], v[42:45], v[88:91], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[56], v[102], v[98], s[86]))
+  k.emit(v_perm_b32_e64(v[57], v[110], v[106], s[86]))
+  k.emit(v_mfma_16x16x32(v[92:95], v[122:125], v[46:49], v[92:95], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4096))
+  k.emit(buffer_load_dwordx4(v[0:3], v[3], s[68:71], 0, 0, 1, 0, 0, 1, 0, 1))
+  k.emit(v_mfma_16x16x32(v[96:99], v[126:129], v[18:21], v[96:99], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[166:169], v[17], v[0], v[0], 0, 192, 2))
+  k.emit(v_mfma_16x16x32(v[100:103], v[126:129], v[22:25], v[100:103], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[58], v[87], v[83], s[85]))
+  k.emit(v_perm_b32_e64(v[59], v[95], v[91], s[85]))
+  k.emit(v_mfma_16x16x32(v[104:107], v[126:129], v[26:29], v[104:107], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4096))
+  k.emit(buffer_load_dwordx4(v[0:3], v[4], s[68:71], 0, 0, 1, 0, 0, 1, 0, 1))
+  k.emit(v_mfma_16x16x32(v[108:111], v[126:129], v[30:33], v[108:111], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[170:173], v[17], v[0], v[0], 0, 64, 3))
+  k.emit(v_mfma_16x16x32(v[112:115], v[126:129], v[34:37], v[112:115], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_cmp_eq_u32(s[8], s[78]))
+  k.emit(v_perm_b32_e64(v[60], v[103], v[99], s[85]))
+  k.emit(v_perm_b32_e64(v[61], v[111], v[107], s[85]))
+  k.emit(v_mfma_16x16x32(v[116:119], v[126:129], v[38:41], v[116:119], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_cselect_b32(s[88], s[81], s[84]))
+  k.emit(v_perm_b32_e64(v[62], v[87], v[83], s[86]))
+  k.emit(v_perm_b32_e64(v[63], v[95], v[91], s[86]))
+  k.emit(v_mfma_16x16x32(v[120:123], v[126:129], v[42:45], v[120:123], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_cselect_b32(s[89], s[82], 0))
+  k.emit(ds_read_b128(v[174:177], v[17], v[0], v[0], 0, 192, 3))
+  k.emit(v_mfma_16x16x32(v[124:127], v[126:129], v[46:49], v[124:127], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(s[72], s[72], s[88]))
+  k.emit(v_perm_b32_e64(v[64], v[103], v[99], s[86]))
+  k.emit(v_perm_b32_e64(v[65], v[111], v[107], s[86]))
+  k.emit(v_mfma_16x16x32(v[128:131], v[130:133], v[18:21], v[128:131], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_addc_u32(s[73], s[73], s[89]))
+  k.emit(v_perm_b32_e64(v[66], v[88], v[84], s[85]))
+  k.emit(v_perm_b32_e64(v[67], v[96], v[92], s[85]))
+  k.emit(v_mfma_16x16x32(v[132:135], v[130:133], v[22:25], v[132:135], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_sub_u32(s[76], s[76], s[88]))
+  k.emit(v_perm_b32_e64(v[68], v[104], v[100], s[85]))
+  k.emit(v_perm_b32_e64(v[69], v[112], v[108], s[85]))
+  k.emit(v_mfma_16x16x32(v[136:139], v[130:133], v[26:29], v[136:139], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_subb_u32(s[77], s[77], s[89]))
+  k.emit(v_perm_b32_e64(v[70], v[88], v[84], s[86]))
+  k.emit(v_perm_b32_e64(v[71], v[96], v[92], s[86]))
+  k.emit(v_mfma_16x16x32(v[140:143], v[130:133], v[30:33], v[140:143], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_cmp_eq_u32(s[77], 0))
+  k.emit(v_perm_b32_e64(v[72], v[104], v[100], s[86]))
+  k.emit(v_perm_b32_e64(v[73], v[112], v[108], s[86]))
+  k.emit(v_mfma_16x16x32(v[144:147], v[130:133], v[34:37], v[144:147], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.waitcnt(lgkm=0)
+  k.emit(s_cselect_b32(s[74], s[76], -1))
+  k.emit(v_perm_b32_e64(v[74], v[89], v[85], s[85]))
+  k.emit(v_perm_b32_e64(v[75], v[97], v[93], s[85]))
+  k.emit(v_mfma_16x16x32(v[148:151], v[130:133], v[38:41], v[148:151], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[76], v[105], v[101], s[85]))
+  k.emit(v_perm_b32_e64(v[77], v[113], v[109], s[85]))
+  k.emit(v_mfma_16x16x32(v[152:155], v[130:133], v[42:45], v[152:155], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[78], v[89], v[85], s[86]))
+  k.emit(v_perm_b32_e64(v[79], v[97], v[93], s[86]))
+  k.emit(v_mfma_16x16x32(v[156:159], v[130:133], v[46:49], v[156:159], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[80], v[105], v[101], s[86]))
+  k.emit(v_perm_b32_e64(v[81], v[113], v[109], s[86]))
+  k.emit(v_mfma_16x16x32(v[160:163], v[134:137], v[18:21], v[160:163], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[164:167], v[134:137], v[22:25], v[164:167], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[168:171], v[134:137], v[26:29], v[168:171], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[172:175], v[134:137], v[30:33], v[172:175], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[176:179], v[134:137], v[34:37], v[176:179], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_barrier())
+  k.emit(v_mfma_16x16x32(v[180:183], v[134:137], v[38:41], v[180:183], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4096))
+  k.emit(buffer_load_dwordx4(v[0:3], v[5], s[68:71], 0, 0, 1, 0, 0, 1, 0, 1))
+  k.emit(v_mfma_16x16x32(v[184:187], v[134:137], v[42:45], v[184:187], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[188:191], v[134:137], v[46:49], v[188:191], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[192:195], v[138:141], v[18:21], v[192:195], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4096))
+  k.emit(buffer_load_dwordx4(v[0:3], v[6], s[68:71], 0, 0, 1, 0, 0, 1, 0, 1))
+  k.emit(v_mfma_16x16x32(v[196:199], v[138:141], v[22:25], v[196:199], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[200:203], v[138:141], v[26:29], v[200:203], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[204:207], v[138:141], v[30:33], v[204:207], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4096))
+  k.emit(buffer_load_dwordx4(v[0:3], v[7], s[68:71], 0, 0, 1, 0, 0, 1, 0, 1))
+  k.emit(v_mfma_16x16x32(v[208:211], v[138:141], v[34:37], v[208:211], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[212:215], v[138:141], v[38:41], v[212:215], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[216:219], v[138:141], v[42:45], v[216:219], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_mov_b32(M0, s[54]))
+  k.emit(buffer_load_dwordx4(v[0:3], v[8], s[72:75], 0, 0, 1, 0, 0, 0, 1, 1, 1))
+  k.emit(v_mfma_16x16x32(v[220:223], v[138:141], v[46:49], v[220:223], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[224:227], v[142:145], v[18:21], v[224:227], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.waitcnt(vm=17)
+  k.emit(v_mfma_16x16x32(v[228:231], v[142:145], v[22:25], v[228:231], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4224))
+  k.emit(buffer_load_dwordx4(v[0:3], v[9], s[72:75], 0, 0, 1, 0, 0, 0, 1, 1, 1))
+  k.emit(v_mfma_16x16x32(v[232:235], v[142:145], v[26:29], v[232:235], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[236:239], v[142:145], v[30:33], v[236:239], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_barrier())
+  k.emit(v_xor_b32_e32(v[16], v[178], v[16]))
+  k.emit(v_xor_b32_e32(v[17], v[179], v[17]))
+  k.emit(v_mfma_16x16x32(v[240:243], v[142:145], v[34:37], v[240:243], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[82:85], v[16]))
+  k.emit(ds_read_b128(v[86:89], v[16], v[0], v[0], 0, 0, 2))
+  k.emit(v_mfma_16x16x32(v[244:247], v[142:145], v[38:41], v[244:247], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[248:251], v[142:145], v[42:45], v[248:251], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[90:93], v[16], v[0], v[0], 0, 0, 4))
+  k.emit(ds_read_b128(v[94:97], v[16], v[0], v[0], 0, 0, 6))
+  k.emit(v_mfma_16x16x32(v[252:255], v[142:145], v[46:49], v[252:255], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[0:3], v[146:149], v[50:53], v[0:3], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[98:101], v[16], v[0], v[0], 0, 0, 8))
+  k.emit(ds_read_b128(v[102:105], v[16], v[0], v[0], 0, 0, 10))
+  k.emit(v_mfma_16x16x32(v[4:7], v[146:149], v[54:57], v[4:7], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[8:11], v[146:149], v[58:61], v[8:11], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.waitcnt(vm=9)
+  k.emit(ds_read_b128(v[106:109], v[16], v[0], v[0], 0, 0, 12))
+  k.emit(ds_read_b128(v[110:113], v[16], v[0], v[0], 0, 0, 14))
+  k.emit(v_mfma_16x16x32(v[12:15], v[146:149], v[62:65], v[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[16:19], v[146:149], v[66:69], v[16:19], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_barrier())
+  k.emit(ds_read_b128(v[114:117], v[17]))
+  k.emit(v_mfma_16x16x32(v[20:23], v[146:149], v[70:73], v[20:23], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[24:27], v[146:149], v[74:77], v[24:27], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[118:121], v[17], v[0], v[0], 0, 128))
+  k.emit(v_mfma_16x16x32(v[28:31], v[146:149], v[78:81], v[28:31], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[32:35], v[150:153], v[50:53], v[32:35], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[122:125], v[17], v[0], v[0], 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[36:39], v[150:153], v[54:57], v[36:39], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.waitcnt(lgkm=4)
+  k.emit(v_mfma_16x16x32(v[40:43], v[150:153], v[58:61], v[40:43], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[126:129], v[17], v[0], v[0], 0, 128, 1))
+  k.emit(v_mfma_16x16x32(v[44:47], v[150:153], v[62:65], v[44:47], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[18], v[86], v[82], s[85]))
+  k.emit(v_perm_b32_e64(v[19], v[94], v[90], s[85]))
+  k.emit(v_mfma_16x16x32(v[48:51], v[150:153], v[66:69], v[48:51], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[130:133], v[17], v[0], v[0], 0, 0, 2))
+  k.emit(v_mfma_16x16x32(v[52:55], v[150:153], v[70:73], v[52:55], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[20], v[102], v[98], s[85]))
+  k.emit(v_perm_b32_e64(v[21], v[110], v[106], s[85]))
+  k.emit(v_mfma_16x16x32(v[56:59], v[150:153], v[74:77], v[56:59], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[134:137], v[17], v[0], v[0], 0, 128, 2))
+  k.emit(v_mfma_16x16x32(v[60:63], v[150:153], v[78:81], v[60:63], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[22], v[86], v[82], s[86]))
+  k.emit(v_perm_b32_e64(v[23], v[94], v[90], s[86]))
+  k.emit(v_mfma_16x16x32(v[64:67], v[154:157], v[50:53], v[64:67], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[138:141], v[17], v[0], v[0], 0, 0, 3))
+  k.emit(v_mfma_16x16x32(v[68:71], v[154:157], v[54:57], v[68:71], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[24], v[102], v[98], s[86]))
+  k.emit(v_perm_b32_e64(v[25], v[110], v[106], s[86]))
+  k.emit(v_mfma_16x16x32(v[72:75], v[154:157], v[58:61], v[72:75], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[142:145], v[17], v[0], v[0], 0, 128, 3))
+  k.emit(v_mfma_16x16x32(v[76:79], v[154:157], v[62:65], v[76:79], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[26], v[87], v[83], s[85]))
+  k.emit(v_perm_b32_e64(v[27], v[95], v[91], s[85]))
+  k.emit(v_mfma_16x16x32(v[80:83], v[154:157], v[66:69], v[80:83], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[28], v[103], v[99], s[85]))
+  k.emit(v_perm_b32_e64(v[29], v[111], v[107], s[85]))
+  k.emit(v_mfma_16x16x32(v[84:87], v[154:157], v[70:73], v[84:87], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[30], v[87], v[83], s[86]))
+  k.emit(v_perm_b32_e64(v[31], v[95], v[91], s[86]))
+  k.emit(v_mfma_16x16x32(v[88:91], v[154:157], v[74:77], v[88:91], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[32], v[103], v[99], s[86]))
+  k.emit(v_perm_b32_e64(v[33], v[111], v[107], s[86]))
+  k.emit(v_mfma_16x16x32(v[92:95], v[154:157], v[78:81], v[92:95], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4224))
+  k.emit(buffer_load_dwordx4(v[0:3], v[10], s[72:75], 0, 0, 1, 0, 0, 0, 1, 1, 1))
+  k.emit(v_mfma_16x16x32(v[96:99], v[158:161], v[50:53], v[96:99], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[34], v[88], v[84], s[85]))
+  k.emit(v_perm_b32_e64(v[35], v[96], v[92], s[85]))
+  k.emit(v_mfma_16x16x32(v[100:103], v[158:161], v[54:57], v[100:103], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[36], v[104], v[100], s[85]))
+  k.emit(v_perm_b32_e64(v[37], v[112], v[108], s[85]))
+  k.emit(v_mfma_16x16x32(v[104:107], v[158:161], v[58:61], v[104:107], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4224))
+  k.emit(buffer_load_dwordx4(v[0:3], v[11], s[72:75], 0, 0, 1, 0, 0, 0, 1, 1, 1))
+  k.emit(v_mfma_16x16x32(v[108:111], v[158:161], v[62:65], v[108:111], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[38], v[88], v[84], s[86]))
+  k.emit(v_perm_b32_e64(v[39], v[96], v[92], s[86]))
+  k.emit(v_mfma_16x16x32(v[112:115], v[158:161], v[66:69], v[112:115], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[40], v[104], v[100], s[86]))
+  k.emit(v_perm_b32_e64(v[41], v[112], v[108], s[86]))
+  k.emit(v_mfma_16x16x32(v[116:119], v[158:161], v[70:73], v[116:119], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4224))
+  k.emit(buffer_load_dwordx4(v[0:3], v[12], s[72:75], 0, 0, 1, 0, 0, 0, 1, 1, 1))
+  k.emit(v_mfma_16x16x32(v[120:123], v[158:161], v[74:77], v[120:123], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[42], v[89], v[85], s[85]))
+  k.emit(v_perm_b32_e64(v[43], v[97], v[93], s[85]))
+  k.emit(v_mfma_16x16x32(v[124:127], v[158:161], v[78:81], v[124:127], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[44], v[105], v[101], s[85]))
+  k.emit(v_perm_b32_e64(v[45], v[113], v[109], s[85]))
+  k.emit(v_mfma_16x16x32(v[128:131], v[162:165], v[50:53], v[128:131], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4224))
+  k.emit(buffer_load_dwordx4(v[0:3], v[13], s[72:75], 0, 0, 1, 0, 0, 0, 1, 1, 1))
+  k.emit(v_mfma_16x16x32(v[132:135], v[162:165], v[54:57], v[132:135], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[46], v[89], v[85], s[86]))
+  k.emit(v_perm_b32_e64(v[47], v[97], v[93], s[86]))
+  k.emit(v_mfma_16x16x32(v[136:139], v[162:165], v[58:61], v[136:139], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[48], v[105], v[101], s[86]))
+  k.emit(v_perm_b32_e64(v[49], v[113], v[109], s[86]))
+  k.emit(v_mfma_16x16x32(v[140:143], v[162:165], v[62:65], v[140:143], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4224))
+  k.emit(buffer_load_dwordx4(v[0:3], v[14], s[72:75], 0, 0, 1, 0, 0, 0, 1, 1, 1))
+  k.emit(v_mfma_16x16x32(v[144:147], v[162:165], v[66:69], v[144:147], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[148:151], v[162:165], v[70:73], v[148:151], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[152:155], v[162:165], v[74:77], v[152:155], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[156:159], v[162:165], v[78:81], v[156:159], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[160:163], v[166:169], v[50:53], v[160:163], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[164:167], v[166:169], v[54:57], v[164:167], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[168:171], v[166:169], v[58:61], v[168:171], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[172:175], v[166:169], v[62:65], v[172:175], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[176:179], v[166:169], v[66:69], v[176:179], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[180:183], v[166:169], v[70:73], v[180:183], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[184:187], v[166:169], v[74:77], v[184:187], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[188:191], v[166:169], v[78:81], v[188:191], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[192:195], v[170:173], v[50:53], v[192:195], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[196:199], v[170:173], v[54:57], v[196:199], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[200:203], v[170:173], v[58:61], v[200:203], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[204:207], v[170:173], v[62:65], v[204:207], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[208:211], v[170:173], v[66:69], v[208:211], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[212:215], v[170:173], v[70:73], v[212:215], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[216:219], v[170:173], v[74:77], v[216:219], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[220:223], v[170:173], v[78:81], v[220:223], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[224:227], v[174:177], v[50:53], v[224:227], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[228:231], v[174:177], v[54:57], v[228:231], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[232:235], v[174:177], v[58:61], v[232:235], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[236:239], v[174:177], v[62:65], v[236:239], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4224))
+  k.emit(buffer_load_dwordx4(v[0:3], v[15], s[72:75], 0, 0, 1, 0, 0, 0, 1, 1, 1))
+  k.emit(v_mfma_16x16x32(v[240:243], v[174:177], v[66:69], v[240:243], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[244:247], v[174:177], v[70:73], v[244:247], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_xor_b32(s[53], s[55], s[53]))
+  k.emit(s_xor_b32(s[54], s[56], s[54]))
+  k.emit(v_mfma_16x16x32(v[248:251], v[174:177], v[74:77], v[248:251], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_sub_u32(s[8], s[8], 1))
+  k.emit(s_cmp_eq_i32(s[8], 2))
+  k.emit(v_mfma_16x16x32(v[252:255], v[174:177], v[78:81], v[252:255], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_cbranch_scc0(), target='LoopBeginL_0')
+  k.emit(s_branch(), target='LoopEndL')
+  k.label('LoopBeginL_1')
+  k.emit(v_mfma_16x16x32(v[0:3], v[114:117], v[18:21], v[0:3], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_cmp_eq_u32(s[8], s[78]))
+  k.emit(v_mfma_16x16x32(v[4:7], v[114:117], v[22:25], v[4:7], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_cselect_b32(s[88], s[79], s[83]))
+  k.emit(ds_read_b128(v[82:85], v[16], v[0], v[0], 0, 0, 64))
+  k.emit(ds_read_b128(v[86:89], v[16], v[0], v[0], 0, 0, 66))
+  k.emit(v_mfma_16x16x32(v[8:11], v[114:117], v[26:29], v[8:11], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_cselect_b32(s[89], s[80], 0))
+  k.emit(v_mfma_16x16x32(v[12:15], v[114:117], v[30:33], v[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(s[68], s[68], s[88]))
+  k.emit(ds_read_b128(v[90:93], v[16], v[0], v[0], 0, 0, 68))
+  k.emit(ds_read_b128(v[94:97], v[16], v[0], v[0], 0, 0, 70))
+  k.emit(v_mfma_16x16x32(v[16:19], v[114:117], v[34:37], v[16:19], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_addc_u32(s[69], s[69], s[89]))
+  k.emit(v_mfma_16x16x32(v[20:23], v[114:117], v[38:41], v[20:23], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_sub_u32(s[62], s[62], s[88]))
+  k.emit(ds_read_b128(v[98:101], v[16], v[0], v[0], 0, 0, 72))
+  k.emit(ds_read_b128(v[102:105], v[16], v[0], v[0], 0, 0, 74))
+  k.emit(v_mfma_16x16x32(v[24:27], v[114:117], v[42:45], v[24:27], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_subb_u32(s[63], s[63], s[89]))
+  k.emit(v_mfma_16x16x32(v[28:31], v[114:117], v[46:49], v[28:31], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_cmp_eq_u32(s[63], 0))
+  k.emit(ds_read_b128(v[106:109], v[16], v[0], v[0], 0, 0, 76))
+  k.emit(ds_read_b128(v[110:113], v[16], v[0], v[0], 0, 0, 78))
+  k.emit(v_mfma_16x16x32(v[32:35], v[118:121], v[18:21], v[32:35], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.waitcnt(lgkm=4)
+  k.emit(s_cselect_b32(s[70], s[62], -1))
+  k.emit(v_perm_b32_e64(v[50], v[86], v[82], s[85]))
+  k.emit(v_perm_b32_e64(v[51], v[94], v[90], s[85]))
+  k.emit(v_mfma_16x16x32(v[36:39], v[118:121], v[22:25], v[36:39], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[40:43], v[118:121], v[26:29], v[40:43], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[146:149], v[17], v[0], v[0], 0, 64))
+  k.emit(v_mfma_16x16x32(v[44:47], v[118:121], v[30:33], v[44:47], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[48:51], v[118:121], v[34:37], v[48:51], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.waitcnt(lgkm=1)
+  k.emit(ds_read_b128(v[150:153], v[17], v[0], v[0], 0, 192))
+  k.emit(v_mfma_16x16x32(v[52:55], v[118:121], v[38:41], v[52:55], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_barrier())
+  k.emit(v_mfma_16x16x32(v[56:59], v[118:121], v[42:45], v[56:59], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[154:157], v[17], v[0], v[0], 0, 64, 1))
+  k.emit(v_mfma_16x16x32(v[60:63], v[118:121], v[46:49], v[60:63], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_mov_b32(M0, s[53]))
+  k.emit(buffer_load_dwordx4(v[0:3], v[0], s[68:71], 0, 0, 1, 0, 0, 1, 0, 1))
+  k.emit(v_mfma_16x16x32(v[64:67], v[122:125], v[18:21], v[64:67], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[52], v[102], v[98], s[85]))
+  k.emit(v_perm_b32_e64(v[53], v[110], v[106], s[85]))
+  k.emit(v_mfma_16x16x32(v[68:71], v[122:125], v[22:25], v[68:71], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[158:161], v[17], v[0], v[0], 0, 192, 1))
+  k.emit(v_mfma_16x16x32(v[72:75], v[122:125], v[26:29], v[72:75], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4096))
+  k.emit(buffer_load_dwordx4(v[0:3], v[1], s[68:71], 0, 0, 1, 0, 0, 1, 0, 1))
+  k.emit(v_mfma_16x16x32(v[76:79], v[122:125], v[30:33], v[76:79], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[54], v[86], v[82], s[86]))
+  k.emit(v_perm_b32_e64(v[55], v[94], v[90], s[86]))
+  k.emit(v_mfma_16x16x32(v[80:83], v[122:125], v[34:37], v[80:83], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[162:165], v[17], v[0], v[0], 0, 64, 2))
+  k.emit(v_mfma_16x16x32(v[84:87], v[122:125], v[38:41], v[84:87], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4096))
+  k.emit(buffer_load_dwordx4(v[0:3], v[2], s[68:71], 0, 0, 1, 0, 0, 1, 0, 1))
+  k.emit(v_mfma_16x16x32(v[88:91], v[122:125], v[42:45], v[88:91], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[56], v[102], v[98], s[86]))
+  k.emit(v_perm_b32_e64(v[57], v[110], v[106], s[86]))
+  k.emit(v_mfma_16x16x32(v[92:95], v[122:125], v[46:49], v[92:95], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[166:169], v[17], v[0], v[0], 0, 192, 2))
+  k.emit(v_mfma_16x16x32(v[96:99], v[126:129], v[18:21], v[96:99], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4096))
+  k.emit(buffer_load_dwordx4(v[0:3], v[3], s[68:71], 0, 0, 1, 0, 0, 1, 0, 1))
+  k.emit(v_mfma_16x16x32(v[100:103], v[126:129], v[22:25], v[100:103], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[58], v[87], v[83], s[85]))
+  k.emit(v_perm_b32_e64(v[59], v[95], v[91], s[85]))
+  k.emit(v_mfma_16x16x32(v[104:107], v[126:129], v[26:29], v[104:107], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[170:173], v[17], v[0], v[0], 0, 64, 3))
+  k.emit(v_mfma_16x16x32(v[108:111], v[126:129], v[30:33], v[108:111], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4096))
+  k.emit(buffer_load_dwordx4(v[0:3], v[4], s[68:71], 0, 0, 1, 0, 0, 1, 0, 1))
+  k.emit(v_mfma_16x16x32(v[112:115], v[126:129], v[34:37], v[112:115], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_cmp_eq_u32(s[8], s[78]))
+  k.emit(v_perm_b32_e64(v[60], v[103], v[99], s[85]))
+  k.emit(v_perm_b32_e64(v[61], v[111], v[107], s[85]))
+  k.emit(v_mfma_16x16x32(v[116:119], v[126:129], v[38:41], v[116:119], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_cselect_b32(s[88], s[81], s[84]))
+  k.emit(v_perm_b32_e64(v[62], v[87], v[83], s[86]))
+  k.emit(v_perm_b32_e64(v[63], v[95], v[91], s[86]))
+  k.emit(ds_read_b128(v[174:177], v[17], v[0], v[0], 0, 192, 3))
+  k.emit(v_mfma_16x16x32(v[120:123], v[126:129], v[42:45], v[120:123], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_cselect_b32(s[89], s[82], 0))
+  k.emit(v_mfma_16x16x32(v[124:127], v[126:129], v[46:49], v[124:127], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(s[72], s[72], s[88]))
+  k.emit(v_perm_b32_e64(v[64], v[103], v[99], s[86]))
+  k.emit(v_perm_b32_e64(v[65], v[111], v[107], s[86]))
+  k.emit(v_mfma_16x16x32(v[128:131], v[130:133], v[18:21], v[128:131], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_addc_u32(s[73], s[73], s[89]))
+  k.emit(v_perm_b32_e64(v[66], v[88], v[84], s[85]))
+  k.emit(v_perm_b32_e64(v[67], v[96], v[92], s[85]))
+  k.emit(v_mfma_16x16x32(v[132:135], v[130:133], v[22:25], v[132:135], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_sub_u32(s[76], s[76], s[88]))
+  k.emit(v_perm_b32_e64(v[68], v[104], v[100], s[85]))
+  k.emit(v_perm_b32_e64(v[69], v[112], v[108], s[85]))
+  k.emit(v_mfma_16x16x32(v[136:139], v[130:133], v[26:29], v[136:139], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_subb_u32(s[77], s[77], s[89]))
+  k.emit(v_perm_b32_e64(v[70], v[88], v[84], s[86]))
+  k.emit(v_perm_b32_e64(v[71], v[96], v[92], s[86]))
+  k.emit(v_mfma_16x16x32(v[140:143], v[130:133], v[30:33], v[140:143], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_cmp_eq_u32(s[77], 0))
+  k.emit(v_perm_b32_e64(v[72], v[104], v[100], s[86]))
+  k.emit(v_perm_b32_e64(v[73], v[112], v[108], s[86]))
+  k.emit(v_mfma_16x16x32(v[144:147], v[130:133], v[34:37], v[144:147], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.waitcnt(lgkm=0)
+  k.emit(s_cselect_b32(s[74], s[76], -1))
+  k.emit(v_perm_b32_e64(v[74], v[89], v[85], s[85]))
+  k.emit(v_perm_b32_e64(v[75], v[97], v[93], s[85]))
+  k.emit(v_mfma_16x16x32(v[148:151], v[130:133], v[38:41], v[148:151], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[76], v[105], v[101], s[85]))
+  k.emit(v_perm_b32_e64(v[77], v[113], v[109], s[85]))
+  k.emit(v_mfma_16x16x32(v[152:155], v[130:133], v[42:45], v[152:155], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[78], v[89], v[85], s[86]))
+  k.emit(v_perm_b32_e64(v[79], v[97], v[93], s[86]))
+  k.emit(v_mfma_16x16x32(v[156:159], v[130:133], v[46:49], v[156:159], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[80], v[105], v[101], s[86]))
+  k.emit(v_perm_b32_e64(v[81], v[113], v[109], s[86]))
+  k.emit(v_mfma_16x16x32(v[160:163], v[134:137], v[18:21], v[160:163], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[164:167], v[134:137], v[22:25], v[164:167], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[168:171], v[134:137], v[26:29], v[168:171], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[172:175], v[134:137], v[30:33], v[172:175], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[176:179], v[134:137], v[34:37], v[176:179], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_barrier())
+  k.emit(v_mfma_16x16x32(v[180:183], v[134:137], v[38:41], v[180:183], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[184:187], v[134:137], v[42:45], v[184:187], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4096))
+  k.emit(buffer_load_dwordx4(v[0:3], v[5], s[68:71], 0, 0, 1, 0, 0, 1, 0, 1))
+  k.emit(v_mfma_16x16x32(v[188:191], v[134:137], v[46:49], v[188:191], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[192:195], v[138:141], v[18:21], v[192:195], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[196:199], v[138:141], v[22:25], v[196:199], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4096))
+  k.emit(buffer_load_dwordx4(v[0:3], v[6], s[68:71], 0, 0, 1, 0, 0, 1, 0, 1))
+  k.emit(v_mfma_16x16x32(v[200:203], v[138:141], v[26:29], v[200:203], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[204:207], v[138:141], v[30:33], v[204:207], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[208:211], v[138:141], v[34:37], v[208:211], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4096))
+  k.emit(buffer_load_dwordx4(v[0:3], v[7], s[68:71], 0, 0, 1, 0, 0, 1, 0, 1))
+  k.emit(v_mfma_16x16x32(v[212:215], v[138:141], v[38:41], v[212:215], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[216:219], v[138:141], v[42:45], v[216:219], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[220:223], v[138:141], v[46:49], v[220:223], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_mov_b32(M0, s[54]))
+  k.emit(buffer_load_dwordx4(v[0:3], v[8], s[72:75], 0, 0, 1, 0, 0, 0, 1, 1, 1))
+  k.emit(v_mfma_16x16x32(v[224:227], v[142:145], v[18:21], v[224:227], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.waitcnt(vm=17)
+  k.emit(v_mfma_16x16x32(v[228:231], v[142:145], v[22:25], v[228:231], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[232:235], v[142:145], v[26:29], v[232:235], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4224))
+  k.emit(buffer_load_dwordx4(v[0:3], v[9], s[72:75], 0, 0, 1, 0, 0, 0, 1, 1, 1))
+  k.emit(v_mfma_16x16x32(v[236:239], v[142:145], v[30:33], v[236:239], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_barrier())
+  k.emit(v_xor_b32_e32(v[16], v[178], v[16]))
+  k.emit(v_xor_b32_e32(v[17], v[179], v[17]))
+  k.emit(v_mfma_16x16x32(v[240:243], v[142:145], v[34:37], v[240:243], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[244:247], v[142:145], v[38:41], v[244:247], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[82:85], v[16]))
+  k.emit(ds_read_b128(v[86:89], v[16], v[0], v[0], 0, 0, 2))
+  k.emit(v_mfma_16x16x32(v[248:251], v[142:145], v[42:45], v[248:251], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[252:255], v[142:145], v[46:49], v[252:255], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[90:93], v[16], v[0], v[0], 0, 0, 4))
+  k.emit(ds_read_b128(v[94:97], v[16], v[0], v[0], 0, 0, 6))
+  k.emit(v_mfma_16x16x32(v[0:3], v[146:149], v[50:53], v[0:3], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[4:7], v[146:149], v[54:57], v[4:7], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[98:101], v[16], v[0], v[0], 0, 0, 8))
+  k.emit(ds_read_b128(v[102:105], v[16], v[0], v[0], 0, 0, 10))
+  k.emit(v_mfma_16x16x32(v[8:11], v[146:149], v[58:61], v[8:11], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.waitcnt(vm=9)
+  k.emit(v_mfma_16x16x32(v[12:15], v[146:149], v[62:65], v[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[106:109], v[16], v[0], v[0], 0, 0, 12))
+  k.emit(ds_read_b128(v[110:113], v[16], v[0], v[0], 0, 0, 14))
+  k.emit(v_mfma_16x16x32(v[16:19], v[146:149], v[66:69], v[16:19], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_barrier())
+  k.emit(v_mfma_16x16x32(v[20:23], v[146:149], v[70:73], v[20:23], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[114:117], v[17]))
+  k.emit(v_mfma_16x16x32(v[24:27], v[146:149], v[74:77], v[24:27], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[28:31], v[146:149], v[78:81], v[28:31], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[118:121], v[17], v[0], v[0], 0, 128))
+  k.emit(v_mfma_16x16x32(v[32:35], v[150:153], v[50:53], v[32:35], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[36:39], v[150:153], v[54:57], v[36:39], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.waitcnt(lgkm=4)
+  k.emit(ds_read_b128(v[122:125], v[17], v[0], v[0], 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[40:43], v[150:153], v[58:61], v[40:43], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[18], v[86], v[82], s[85]))
+  k.emit(v_perm_b32_e64(v[19], v[94], v[90], s[85]))
+  k.emit(v_mfma_16x16x32(v[44:47], v[150:153], v[62:65], v[44:47], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[126:129], v[17], v[0], v[0], 0, 128, 1))
+  k.emit(v_mfma_16x16x32(v[48:51], v[150:153], v[66:69], v[48:51], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[20], v[102], v[98], s[85]))
+  k.emit(v_perm_b32_e64(v[21], v[110], v[106], s[85]))
+  k.emit(v_mfma_16x16x32(v[52:55], v[150:153], v[70:73], v[52:55], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[130:133], v[17], v[0], v[0], 0, 0, 2))
+  k.emit(v_mfma_16x16x32(v[56:59], v[150:153], v[74:77], v[56:59], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[22], v[86], v[82], s[86]))
+  k.emit(v_perm_b32_e64(v[23], v[94], v[90], s[86]))
+  k.emit(v_mfma_16x16x32(v[60:63], v[150:153], v[78:81], v[60:63], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[134:137], v[17], v[0], v[0], 0, 128, 2))
+  k.emit(v_mfma_16x16x32(v[64:67], v[154:157], v[50:53], v[64:67], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[24], v[102], v[98], s[86]))
+  k.emit(v_perm_b32_e64(v[25], v[110], v[106], s[86]))
+  k.emit(v_mfma_16x16x32(v[68:71], v[154:157], v[54:57], v[68:71], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[138:141], v[17], v[0], v[0], 0, 0, 3))
+  k.emit(v_mfma_16x16x32(v[72:75], v[154:157], v[58:61], v[72:75], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[26], v[87], v[83], s[85]))
+  k.emit(v_perm_b32_e64(v[27], v[95], v[91], s[85]))
+  k.emit(v_mfma_16x16x32(v[76:79], v[154:157], v[62:65], v[76:79], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[142:145], v[17], v[0], v[0], 0, 128, 3))
+  k.emit(v_mfma_16x16x32(v[80:83], v[154:157], v[66:69], v[80:83], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[28], v[103], v[99], s[85]))
+  k.emit(v_perm_b32_e64(v[29], v[111], v[107], s[85]))
+  k.emit(v_mfma_16x16x32(v[84:87], v[154:157], v[70:73], v[84:87], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[30], v[87], v[83], s[86]))
+  k.emit(v_perm_b32_e64(v[31], v[95], v[91], s[86]))
+  k.emit(v_mfma_16x16x32(v[88:91], v[154:157], v[74:77], v[88:91], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[32], v[103], v[99], s[86]))
+  k.emit(v_perm_b32_e64(v[33], v[111], v[107], s[86]))
+  k.emit(v_mfma_16x16x32(v[92:95], v[154:157], v[78:81], v[92:95], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[34], v[88], v[84], s[85]))
+  k.emit(v_perm_b32_e64(v[35], v[96], v[92], s[85]))
+  k.emit(v_mfma_16x16x32(v[96:99], v[158:161], v[50:53], v[96:99], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4224))
+  k.emit(buffer_load_dwordx4(v[0:3], v[10], s[72:75], 0, 0, 1, 0, 0, 0, 1, 1, 1))
+  k.emit(v_mfma_16x16x32(v[100:103], v[158:161], v[54:57], v[100:103], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[36], v[104], v[100], s[85]))
+  k.emit(v_perm_b32_e64(v[37], v[112], v[108], s[85]))
+  k.emit(v_mfma_16x16x32(v[104:107], v[158:161], v[58:61], v[104:107], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[38], v[88], v[84], s[86]))
+  k.emit(v_perm_b32_e64(v[39], v[96], v[92], s[86]))
+  k.emit(v_mfma_16x16x32(v[108:111], v[158:161], v[62:65], v[108:111], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4224))
+  k.emit(buffer_load_dwordx4(v[0:3], v[11], s[72:75], 0, 0, 1, 0, 0, 0, 1, 1, 1))
+  k.emit(v_mfma_16x16x32(v[112:115], v[158:161], v[66:69], v[112:115], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[40], v[104], v[100], s[86]))
+  k.emit(v_perm_b32_e64(v[41], v[112], v[108], s[86]))
+  k.emit(v_mfma_16x16x32(v[116:119], v[158:161], v[70:73], v[116:119], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[42], v[89], v[85], s[85]))
+  k.emit(v_perm_b32_e64(v[43], v[97], v[93], s[85]))
+  k.emit(v_mfma_16x16x32(v[120:123], v[158:161], v[74:77], v[120:123], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4224))
+  k.emit(buffer_load_dwordx4(v[0:3], v[12], s[72:75], 0, 0, 1, 0, 0, 0, 1, 1, 1))
+  k.emit(v_mfma_16x16x32(v[124:127], v[158:161], v[78:81], v[124:127], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[44], v[105], v[101], s[85]))
+  k.emit(v_perm_b32_e64(v[45], v[113], v[109], s[85]))
+  k.emit(v_mfma_16x16x32(v[128:131], v[162:165], v[50:53], v[128:131], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[46], v[89], v[85], s[86]))
+  k.emit(v_perm_b32_e64(v[47], v[97], v[93], s[86]))
+  k.emit(v_mfma_16x16x32(v[132:135], v[162:165], v[54:57], v[132:135], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4224))
+  k.emit(buffer_load_dwordx4(v[0:3], v[13], s[72:75], 0, 0, 1, 0, 0, 0, 1, 1, 1))
+  k.emit(v_mfma_16x16x32(v[136:139], v[162:165], v[58:61], v[136:139], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[48], v[105], v[101], s[86]))
+  k.emit(v_perm_b32_e64(v[49], v[113], v[109], s[86]))
+  k.emit(v_mfma_16x16x32(v[140:143], v[162:165], v[62:65], v[140:143], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[144:147], v[162:165], v[66:69], v[144:147], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4224))
+  k.emit(buffer_load_dwordx4(v[0:3], v[14], s[72:75], 0, 0, 1, 0, 0, 0, 1, 1, 1))
+  k.emit(v_mfma_16x16x32(v[148:151], v[162:165], v[70:73], v[148:151], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[152:155], v[162:165], v[74:77], v[152:155], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[156:159], v[162:165], v[78:81], v[156:159], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[160:163], v[166:169], v[50:53], v[160:163], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[164:167], v[166:169], v[54:57], v[164:167], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[168:171], v[166:169], v[58:61], v[168:171], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[172:175], v[166:169], v[62:65], v[172:175], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[176:179], v[166:169], v[66:69], v[176:179], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[180:183], v[166:169], v[70:73], v[180:183], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[184:187], v[166:169], v[74:77], v[184:187], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[188:191], v[166:169], v[78:81], v[188:191], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[192:195], v[170:173], v[50:53], v[192:195], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[196:199], v[170:173], v[54:57], v[196:199], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[200:203], v[170:173], v[58:61], v[200:203], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[204:207], v[170:173], v[62:65], v[204:207], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[208:211], v[170:173], v[66:69], v[208:211], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[212:215], v[170:173], v[70:73], v[212:215], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[216:219], v[170:173], v[74:77], v[216:219], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[220:223], v[170:173], v[78:81], v[220:223], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[224:227], v[174:177], v[50:53], v[224:227], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[228:231], v[174:177], v[54:57], v[228:231], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[232:235], v[174:177], v[58:61], v[232:235], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[236:239], v[174:177], v[62:65], v[236:239], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[240:243], v[174:177], v[66:69], v[240:243], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(M0, M0, 4224))
+  k.emit(buffer_load_dwordx4(v[0:3], v[15], s[72:75], 0, 0, 1, 0, 0, 0, 1, 1, 1))
+  k.emit(v_mfma_16x16x32(v[244:247], v[174:177], v[70:73], v[244:247], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_xor_b32(s[53], s[55], s[53]))
+  k.emit(s_xor_b32(s[54], s[56], s[54]))
+  k.emit(v_mfma_16x16x32(v[248:251], v[174:177], v[74:77], v[248:251], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_sub_u32(s[8], s[8], 1))
+  k.emit(s_cmp_eq_i32(s[8], 2))
+  k.emit(v_mfma_16x16x32(v[252:255], v[174:177], v[78:81], v[252:255], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_cbranch_scc0(), target='LoopBeginL_1')
+  k.emit(s_branch(), target='LoopEndL')
+  k.label('LoopEndL')
+  k.emit(s_waitcnt())
+  k.emit(s_barrier())
+  k.emit(v_mfma_16x16x32(v[0:3], v[114:117], v[18:21], v[0:3], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_cmp_eq_u32(s[8], s[78]))
+  k.emit(ds_read_b128(v[82:85], v[16], v[0], v[0], 0, 0, 64))
+  k.emit(ds_read_b128(v[86:89], v[16], v[0], v[0], 0, 0, 66))
+  k.emit(v_mfma_16x16x32(v[4:7], v[114:117], v[22:25], v[4:7], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_cselect_b32(s[88], s[79], s[83]))
+  k.emit(v_mfma_16x16x32(v[8:11], v[114:117], v[26:29], v[8:11], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_cselect_b32(s[89], s[80], 0))
+  k.emit(ds_read_b128(v[90:93], v[16], v[0], v[0], 0, 0, 68))
+  k.emit(ds_read_b128(v[94:97], v[16], v[0], v[0], 0, 0, 70))
+  k.emit(v_mfma_16x16x32(v[12:15], v[114:117], v[30:33], v[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(s[68], s[68], s[88]))
+  k.emit(v_mfma_16x16x32(v[16:19], v[114:117], v[34:37], v[16:19], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_addc_u32(s[69], s[69], s[89]))
+  k.emit(ds_read_b128(v[98:101], v[16], v[0], v[0], 0, 0, 72))
+  k.emit(ds_read_b128(v[102:105], v[16], v[0], v[0], 0, 0, 74))
+  k.emit(v_mfma_16x16x32(v[20:23], v[114:117], v[38:41], v[20:23], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_sub_u32(s[62], s[62], s[88]))
+  k.emit(v_mfma_16x16x32(v[24:27], v[114:117], v[42:45], v[24:27], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_subb_u32(s[63], s[63], s[89]))
+  k.emit(ds_read_b128(v[106:109], v[16], v[0], v[0], 0, 0, 76))
+  k.emit(ds_read_b128(v[110:113], v[16], v[0], v[0], 0, 0, 78))
+  k.emit(v_mfma_16x16x32(v[28:31], v[114:117], v[46:49], v[28:31], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_cmp_eq_u32(s[63], 0))
+  k.emit(v_mfma_16x16x32(v[32:35], v[118:121], v[18:21], v[32:35], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.waitcnt(lgkm=4)
+  k.emit(s_cselect_b32(s[70], s[62], -1))
+  k.emit(v_perm_b32_e64(v[50], v[86], v[82], s[85]))
+  k.emit(v_perm_b32_e64(v[51], v[94], v[90], s[85]))
+  k.emit(v_mfma_16x16x32(v[36:39], v[118:121], v[22:25], v[36:39], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[146:149], v[17], v[0], v[0], 0, 64))
+  k.emit(v_mfma_16x16x32(v[40:43], v[118:121], v[26:29], v[40:43], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[44:47], v[118:121], v[30:33], v[44:47], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[150:153], v[17], v[0], v[0], 0, 192))
+  k.emit(v_mfma_16x16x32(v[48:51], v[118:121], v[34:37], v[48:51], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.waitcnt(lgkm=1)
+  k.emit(v_mfma_16x16x32(v[52:55], v[118:121], v[38:41], v[52:55], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_barrier())
+  k.emit(v_mfma_16x16x32(v[56:59], v[118:121], v[42:45], v[56:59], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[60:63], v[118:121], v[46:49], v[60:63], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[154:157], v[17], v[0], v[0], 0, 64, 1))
+  k.emit(v_mfma_16x16x32(v[64:67], v[122:125], v[18:21], v[64:67], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[52], v[102], v[98], s[85]))
+  k.emit(v_perm_b32_e64(v[53], v[110], v[106], s[85]))
+  k.emit(v_mfma_16x16x32(v[68:71], v[122:125], v[22:25], v[68:71], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[72:75], v[122:125], v[26:29], v[72:75], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[158:161], v[17], v[0], v[0], 0, 192, 1))
+  k.emit(v_mfma_16x16x32(v[76:79], v[122:125], v[30:33], v[76:79], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[54], v[86], v[82], s[86]))
+  k.emit(v_perm_b32_e64(v[55], v[94], v[90], s[86]))
+  k.emit(v_mfma_16x16x32(v[80:83], v[122:125], v[34:37], v[80:83], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[84:87], v[122:125], v[38:41], v[84:87], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[162:165], v[17], v[0], v[0], 0, 64, 2))
+  k.emit(v_mfma_16x16x32(v[88:91], v[122:125], v[42:45], v[88:91], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[56], v[102], v[98], s[86]))
+  k.emit(v_perm_b32_e64(v[57], v[110], v[106], s[86]))
+  k.emit(v_mfma_16x16x32(v[92:95], v[122:125], v[46:49], v[92:95], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[96:99], v[126:129], v[18:21], v[96:99], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[166:169], v[17], v[0], v[0], 0, 192, 2))
+  k.emit(v_mfma_16x16x32(v[100:103], v[126:129], v[22:25], v[100:103], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[58], v[87], v[83], s[85]))
+  k.emit(v_perm_b32_e64(v[59], v[95], v[91], s[85]))
+  k.emit(v_mfma_16x16x32(v[104:107], v[126:129], v[26:29], v[104:107], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[108:111], v[126:129], v[30:33], v[108:111], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[170:173], v[17], v[0], v[0], 0, 64, 3))
+  k.emit(v_mfma_16x16x32(v[112:115], v[126:129], v[34:37], v[112:115], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_cmp_eq_u32(s[8], s[78]))
+  k.emit(v_perm_b32_e64(v[60], v[103], v[99], s[85]))
+  k.emit(v_perm_b32_e64(v[61], v[111], v[107], s[85]))
+  k.emit(v_mfma_16x16x32(v[116:119], v[126:129], v[38:41], v[116:119], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_cselect_b32(s[88], s[81], s[84]))
+  k.emit(v_perm_b32_e64(v[62], v[87], v[83], s[86]))
+  k.emit(v_perm_b32_e64(v[63], v[95], v[91], s[86]))
+  k.emit(v_mfma_16x16x32(v[120:123], v[126:129], v[42:45], v[120:123], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_cselect_b32(s[89], s[82], 0))
+  k.emit(ds_read_b128(v[174:177], v[17], v[0], v[0], 0, 192, 3))
+  k.emit(v_mfma_16x16x32(v[124:127], v[126:129], v[46:49], v[124:127], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_add_u32(s[72], s[72], s[88]))
+  k.emit(v_perm_b32_e64(v[64], v[103], v[99], s[86]))
+  k.emit(v_perm_b32_e64(v[65], v[111], v[107], s[86]))
+  k.emit(v_mfma_16x16x32(v[128:131], v[130:133], v[18:21], v[128:131], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_addc_u32(s[73], s[73], s[89]))
+  k.emit(v_perm_b32_e64(v[66], v[88], v[84], s[85]))
+  k.emit(v_perm_b32_e64(v[67], v[96], v[92], s[85]))
+  k.emit(v_mfma_16x16x32(v[132:135], v[130:133], v[22:25], v[132:135], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_sub_u32(s[76], s[76], s[88]))
+  k.emit(v_perm_b32_e64(v[68], v[104], v[100], s[85]))
+  k.emit(v_perm_b32_e64(v[69], v[112], v[108], s[85]))
+  k.emit(v_mfma_16x16x32(v[136:139], v[130:133], v[26:29], v[136:139], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_subb_u32(s[77], s[77], s[89]))
+  k.emit(v_perm_b32_e64(v[70], v[88], v[84], s[86]))
+  k.emit(v_perm_b32_e64(v[71], v[96], v[92], s[86]))
+  k.emit(v_mfma_16x16x32(v[140:143], v[130:133], v[30:33], v[140:143], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_cmp_eq_u32(s[77], 0))
+  k.emit(v_perm_b32_e64(v[72], v[104], v[100], s[86]))
+  k.emit(v_perm_b32_e64(v[73], v[112], v[108], s[86]))
+  k.emit(v_mfma_16x16x32(v[144:147], v[130:133], v[34:37], v[144:147], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.waitcnt(lgkm=0)
+  k.emit(s_cselect_b32(s[74], s[76], -1))
+  k.emit(v_perm_b32_e64(v[74], v[89], v[85], s[85]))
+  k.emit(v_perm_b32_e64(v[75], v[97], v[93], s[85]))
+  k.emit(v_mfma_16x16x32(v[148:151], v[130:133], v[38:41], v[148:151], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[76], v[105], v[101], s[85]))
+  k.emit(v_perm_b32_e64(v[77], v[113], v[109], s[85]))
+  k.emit(v_mfma_16x16x32(v[152:155], v[130:133], v[42:45], v[152:155], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[78], v[89], v[85], s[86]))
+  k.emit(v_perm_b32_e64(v[79], v[97], v[93], s[86]))
+  k.emit(v_mfma_16x16x32(v[156:159], v[130:133], v[46:49], v[156:159], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[80], v[105], v[101], s[86]))
+  k.emit(v_perm_b32_e64(v[81], v[113], v[109], s[86]))
+  k.emit(v_mfma_16x16x32(v[160:163], v[134:137], v[18:21], v[160:163], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[164:167], v[134:137], v[22:25], v[164:167], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[168:171], v[134:137], v[26:29], v[168:171], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[172:175], v[134:137], v[30:33], v[172:175], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[176:179], v[134:137], v[34:37], v[176:179], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_barrier())
+  k.emit(v_mfma_16x16x32(v[180:183], v[134:137], v[38:41], v[180:183], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[184:187], v[134:137], v[42:45], v[184:187], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[188:191], v[134:137], v[46:49], v[188:191], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[192:195], v[138:141], v[18:21], v[192:195], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[196:199], v[138:141], v[22:25], v[196:199], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[200:203], v[138:141], v[26:29], v[200:203], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[204:207], v[138:141], v[30:33], v[204:207], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[208:211], v[138:141], v[34:37], v[208:211], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[212:215], v[138:141], v[38:41], v[212:215], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[216:219], v[138:141], v[42:45], v[216:219], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[220:223], v[138:141], v[46:49], v[220:223], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[224:227], v[142:145], v[18:21], v[224:227], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.waitcnt(vm=17)
+  k.emit(v_mfma_16x16x32(v[228:231], v[142:145], v[22:25], v[228:231], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[232:235], v[142:145], v[26:29], v[232:235], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[236:239], v[142:145], v[30:33], v[236:239], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_barrier())
+  k.emit(v_xor_b32_e32(v[16], v[178], v[16]))
+  k.emit(v_xor_b32_e32(v[17], v[179], v[17]))
+  k.emit(v_mfma_16x16x32(v[240:243], v[142:145], v[34:37], v[240:243], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[82:85], v[16]))
+  k.emit(ds_read_b128(v[86:89], v[16], v[0], v[0], 0, 0, 2))
+  k.emit(v_mfma_16x16x32(v[244:247], v[142:145], v[38:41], v[244:247], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[248:251], v[142:145], v[42:45], v[248:251], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[90:93], v[16], v[0], v[0], 0, 0, 4))
+  k.emit(ds_read_b128(v[94:97], v[16], v[0], v[0], 0, 0, 6))
+  k.emit(v_mfma_16x16x32(v[252:255], v[142:145], v[46:49], v[252:255], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[0:3], v[146:149], v[50:53], v[0:3], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[98:101], v[16], v[0], v[0], 0, 0, 8))
+  k.emit(ds_read_b128(v[102:105], v[16], v[0], v[0], 0, 0, 10))
+  k.emit(v_mfma_16x16x32(v[4:7], v[146:149], v[54:57], v[4:7], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[8:11], v[146:149], v[58:61], v[8:11], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.waitcnt(vm=9)
+  k.emit(ds_read_b128(v[106:109], v[16], v[0], v[0], 0, 0, 12))
+  k.emit(ds_read_b128(v[110:113], v[16], v[0], v[0], 0, 0, 14))
+  k.emit(v_mfma_16x16x32(v[12:15], v[146:149], v[62:65], v[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[16:19], v[146:149], v[66:69], v[16:19], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_barrier())
+  k.emit(ds_read_b128(v[114:117], v[17]))
+  k.emit(v_mfma_16x16x32(v[20:23], v[146:149], v[70:73], v[20:23], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[24:27], v[146:149], v[74:77], v[24:27], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[118:121], v[17], v[0], v[0], 0, 128))
+  k.emit(v_mfma_16x16x32(v[28:31], v[146:149], v[78:81], v[28:31], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[32:35], v[150:153], v[50:53], v[32:35], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[122:125], v[17], v[0], v[0], 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[36:39], v[150:153], v[54:57], v[36:39], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.waitcnt(lgkm=4)
+  k.emit(v_mfma_16x16x32(v[40:43], v[150:153], v[58:61], v[40:43], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[126:129], v[17], v[0], v[0], 0, 128, 1))
+  k.emit(v_mfma_16x16x32(v[44:47], v[150:153], v[62:65], v[44:47], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[18], v[86], v[82], s[85]))
+  k.emit(v_perm_b32_e64(v[19], v[94], v[90], s[85]))
+  k.emit(v_mfma_16x16x32(v[48:51], v[150:153], v[66:69], v[48:51], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[130:133], v[17], v[0], v[0], 0, 0, 2))
+  k.emit(v_mfma_16x16x32(v[52:55], v[150:153], v[70:73], v[52:55], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[20], v[102], v[98], s[85]))
+  k.emit(v_perm_b32_e64(v[21], v[110], v[106], s[85]))
+  k.emit(v_mfma_16x16x32(v[56:59], v[150:153], v[74:77], v[56:59], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[134:137], v[17], v[0], v[0], 0, 128, 2))
+  k.emit(v_mfma_16x16x32(v[60:63], v[150:153], v[78:81], v[60:63], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[22], v[86], v[82], s[86]))
+  k.emit(v_perm_b32_e64(v[23], v[94], v[90], s[86]))
+  k.emit(v_mfma_16x16x32(v[64:67], v[154:157], v[50:53], v[64:67], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[138:141], v[17], v[0], v[0], 0, 0, 3))
+  k.emit(v_mfma_16x16x32(v[68:71], v[154:157], v[54:57], v[68:71], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[24], v[102], v[98], s[86]))
+  k.emit(v_perm_b32_e64(v[25], v[110], v[106], s[86]))
+  k.emit(v_mfma_16x16x32(v[72:75], v[154:157], v[58:61], v[72:75], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[142:145], v[17], v[0], v[0], 0, 128, 3))
+  k.emit(v_mfma_16x16x32(v[76:79], v[154:157], v[62:65], v[76:79], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[26], v[87], v[83], s[85]))
+  k.emit(v_perm_b32_e64(v[27], v[95], v[91], s[85]))
+  k.emit(v_mfma_16x16x32(v[80:83], v[154:157], v[66:69], v[80:83], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[28], v[103], v[99], s[85]))
+  k.emit(v_perm_b32_e64(v[29], v[111], v[107], s[85]))
+  k.emit(v_mfma_16x16x32(v[84:87], v[154:157], v[70:73], v[84:87], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[30], v[87], v[83], s[86]))
+  k.emit(v_perm_b32_e64(v[31], v[95], v[91], s[86]))
+  k.emit(v_mfma_16x16x32(v[88:91], v[154:157], v[74:77], v[88:91], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[32], v[103], v[99], s[86]))
+  k.emit(v_perm_b32_e64(v[33], v[111], v[107], s[86]))
+  k.emit(v_mfma_16x16x32(v[92:95], v[154:157], v[78:81], v[92:95], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[96:99], v[158:161], v[50:53], v[96:99], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[34], v[88], v[84], s[85]))
+  k.emit(v_perm_b32_e64(v[35], v[96], v[92], s[85]))
+  k.emit(v_mfma_16x16x32(v[100:103], v[158:161], v[54:57], v[100:103], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[36], v[104], v[100], s[85]))
+  k.emit(v_perm_b32_e64(v[37], v[112], v[108], s[85]))
+  k.emit(v_mfma_16x16x32(v[104:107], v[158:161], v[58:61], v[104:107], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[108:111], v[158:161], v[62:65], v[108:111], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[38], v[88], v[84], s[86]))
+  k.emit(v_perm_b32_e64(v[39], v[96], v[92], s[86]))
+  k.emit(v_mfma_16x16x32(v[112:115], v[158:161], v[66:69], v[112:115], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[40], v[104], v[100], s[86]))
+  k.emit(v_perm_b32_e64(v[41], v[112], v[108], s[86]))
+  k.emit(v_mfma_16x16x32(v[116:119], v[158:161], v[70:73], v[116:119], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[120:123], v[158:161], v[74:77], v[120:123], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[42], v[89], v[85], s[85]))
+  k.emit(v_perm_b32_e64(v[43], v[97], v[93], s[85]))
+  k.emit(v_mfma_16x16x32(v[124:127], v[158:161], v[78:81], v[124:127], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[44], v[105], v[101], s[85]))
+  k.emit(v_perm_b32_e64(v[45], v[113], v[109], s[85]))
+  k.emit(v_mfma_16x16x32(v[128:131], v[162:165], v[50:53], v[128:131], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[132:135], v[162:165], v[54:57], v[132:135], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[46], v[89], v[85], s[86]))
+  k.emit(v_perm_b32_e64(v[47], v[97], v[93], s[86]))
+  k.emit(v_mfma_16x16x32(v[136:139], v[162:165], v[58:61], v[136:139], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[48], v[105], v[101], s[86]))
+  k.emit(v_perm_b32_e64(v[49], v[113], v[109], s[86]))
+  k.emit(v_mfma_16x16x32(v[140:143], v[162:165], v[62:65], v[140:143], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[144:147], v[162:165], v[66:69], v[144:147], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[148:151], v[162:165], v[70:73], v[148:151], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[152:155], v[162:165], v[74:77], v[152:155], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[156:159], v[162:165], v[78:81], v[156:159], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[160:163], v[166:169], v[50:53], v[160:163], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[164:167], v[166:169], v[54:57], v[164:167], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[168:171], v[166:169], v[58:61], v[168:171], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[172:175], v[166:169], v[62:65], v[172:175], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[176:179], v[166:169], v[66:69], v[176:179], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[180:183], v[166:169], v[70:73], v[180:183], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[184:187], v[166:169], v[74:77], v[184:187], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[188:191], v[166:169], v[78:81], v[188:191], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[192:195], v[170:173], v[50:53], v[192:195], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[196:199], v[170:173], v[54:57], v[196:199], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[200:203], v[170:173], v[58:61], v[200:203], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[204:207], v[170:173], v[62:65], v[204:207], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[208:211], v[170:173], v[66:69], v[208:211], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[212:215], v[170:173], v[70:73], v[212:215], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[216:219], v[170:173], v[74:77], v[216:219], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[220:223], v[170:173], v[78:81], v[220:223], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[224:227], v[174:177], v[50:53], v[224:227], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[228:231], v[174:177], v[54:57], v[228:231], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[232:235], v[174:177], v[58:61], v[232:235], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[236:239], v[174:177], v[62:65], v[236:239], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[240:243], v[174:177], v[66:69], v[240:243], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[244:247], v[174:177], v[70:73], v[244:247], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[248:251], v[174:177], v[74:77], v[248:251], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[252:255], v[174:177], v[78:81], v[252:255], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.label('toPGR1')
+  k.emit(s_waitcnt())
+  k.emit(s_barrier())
+  k.emit(v_mfma_16x16x32(v[0:3], v[114:117], v[18:21], v[0:3], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[82:85], v[16], v[0], v[0], 0, 0, 64))
+  k.emit(ds_read_b128(v[86:89], v[16], v[0], v[0], 0, 0, 66))
+  k.emit(v_mfma_16x16x32(v[4:7], v[114:117], v[22:25], v[4:7], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[8:11], v[114:117], v[26:29], v[8:11], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[90:93], v[16], v[0], v[0], 0, 0, 68))
+  k.emit(ds_read_b128(v[94:97], v[16], v[0], v[0], 0, 0, 70))
+  k.emit(v_mfma_16x16x32(v[12:15], v[114:117], v[30:33], v[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[16:19], v[114:117], v[34:37], v[16:19], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[98:101], v[16], v[0], v[0], 0, 0, 72))
+  k.emit(ds_read_b128(v[102:105], v[16], v[0], v[0], 0, 0, 74))
+  k.emit(v_mfma_16x16x32(v[20:23], v[114:117], v[38:41], v[20:23], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[24:27], v[114:117], v[42:45], v[24:27], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[106:109], v[16], v[0], v[0], 0, 0, 76))
+  k.emit(ds_read_b128(v[110:113], v[16], v[0], v[0], 0, 0, 78))
+  k.emit(v_mfma_16x16x32(v[28:31], v[114:117], v[46:49], v[28:31], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[32:35], v[118:121], v[18:21], v[32:35], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.waitcnt(lgkm=4)
+  k.emit(v_perm_b32_e64(v[50], v[86], v[82], s[85]))
+  k.emit(v_perm_b32_e64(v[51], v[94], v[90], s[85]))
+  k.emit(v_mfma_16x16x32(v[36:39], v[118:121], v[22:25], v[36:39], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[146:149], v[17], v[0], v[0], 0, 64))
+  k.emit(v_mfma_16x16x32(v[40:43], v[118:121], v[26:29], v[40:43], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[44:47], v[118:121], v[30:33], v[44:47], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[150:153], v[17], v[0], v[0], 0, 192))
+  k.emit(v_mfma_16x16x32(v[48:51], v[118:121], v[34:37], v[48:51], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.waitcnt(lgkm=1)
+  k.emit(v_mfma_16x16x32(v[52:55], v[118:121], v[38:41], v[52:55], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_barrier())
+  k.emit(v_mfma_16x16x32(v[56:59], v[118:121], v[42:45], v[56:59], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[60:63], v[118:121], v[46:49], v[60:63], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[154:157], v[17], v[0], v[0], 0, 64, 1))
+  k.emit(v_mfma_16x16x32(v[64:67], v[122:125], v[18:21], v[64:67], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[52], v[102], v[98], s[85]))
+  k.emit(v_perm_b32_e64(v[53], v[110], v[106], s[85]))
+  k.emit(v_mfma_16x16x32(v[68:71], v[122:125], v[22:25], v[68:71], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[72:75], v[122:125], v[26:29], v[72:75], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[158:161], v[17], v[0], v[0], 0, 192, 1))
+  k.emit(v_mfma_16x16x32(v[76:79], v[122:125], v[30:33], v[76:79], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[54], v[86], v[82], s[86]))
+  k.emit(v_perm_b32_e64(v[55], v[94], v[90], s[86]))
+  k.emit(v_mfma_16x16x32(v[80:83], v[122:125], v[34:37], v[80:83], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[84:87], v[122:125], v[38:41], v[84:87], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[162:165], v[17], v[0], v[0], 0, 64, 2))
+  k.emit(v_mfma_16x16x32(v[88:91], v[122:125], v[42:45], v[88:91], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[56], v[102], v[98], s[86]))
+  k.emit(v_perm_b32_e64(v[57], v[110], v[106], s[86]))
+  k.emit(v_mfma_16x16x32(v[92:95], v[122:125], v[46:49], v[92:95], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[96:99], v[126:129], v[18:21], v[96:99], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[166:169], v[17], v[0], v[0], 0, 192, 2))
+  k.emit(v_mfma_16x16x32(v[100:103], v[126:129], v[22:25], v[100:103], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[58], v[87], v[83], s[85]))
+  k.emit(v_perm_b32_e64(v[59], v[95], v[91], s[85]))
+  k.emit(v_mfma_16x16x32(v[104:107], v[126:129], v[26:29], v[104:107], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[108:111], v[126:129], v[30:33], v[108:111], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[170:173], v[17], v[0], v[0], 0, 64, 3))
+  k.emit(v_mfma_16x16x32(v[112:115], v[126:129], v[34:37], v[112:115], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[60], v[103], v[99], s[85]))
+  k.emit(v_perm_b32_e64(v[61], v[111], v[107], s[85]))
+  k.emit(v_mfma_16x16x32(v[116:119], v[126:129], v[38:41], v[116:119], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[62], v[87], v[83], s[86]))
+  k.emit(v_perm_b32_e64(v[63], v[95], v[91], s[86]))
+  k.emit(v_mfma_16x16x32(v[120:123], v[126:129], v[42:45], v[120:123], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(ds_read_b128(v[174:177], v[17], v[0], v[0], 0, 192, 3))
+  k.emit(v_mfma_16x16x32(v[124:127], v[126:129], v[46:49], v[124:127], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[64], v[103], v[99], s[86]))
+  k.emit(v_perm_b32_e64(v[65], v[111], v[107], s[86]))
+  k.emit(v_mfma_16x16x32(v[128:131], v[130:133], v[18:21], v[128:131], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[66], v[88], v[84], s[85]))
+  k.emit(v_perm_b32_e64(v[67], v[96], v[92], s[85]))
+  k.emit(v_mfma_16x16x32(v[132:135], v[130:133], v[22:25], v[132:135], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[68], v[104], v[100], s[85]))
+  k.emit(v_perm_b32_e64(v[69], v[112], v[108], s[85]))
+  k.emit(v_mfma_16x16x32(v[136:139], v[130:133], v[26:29], v[136:139], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[70], v[88], v[84], s[86]))
+  k.emit(v_perm_b32_e64(v[71], v[96], v[92], s[86]))
+  k.emit(v_mfma_16x16x32(v[140:143], v[130:133], v[30:33], v[140:143], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[72], v[104], v[100], s[86]))
+  k.emit(v_perm_b32_e64(v[73], v[112], v[108], s[86]))
+  k.emit(v_mfma_16x16x32(v[144:147], v[130:133], v[34:37], v[144:147], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.waitcnt(lgkm=0)
+  k.emit(v_perm_b32_e64(v[74], v[89], v[85], s[85]))
+  k.emit(v_perm_b32_e64(v[75], v[97], v[93], s[85]))
+  k.emit(v_mfma_16x16x32(v[148:151], v[130:133], v[38:41], v[148:151], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[76], v[105], v[101], s[85]))
+  k.emit(v_perm_b32_e64(v[77], v[113], v[109], s[85]))
+  k.emit(v_mfma_16x16x32(v[152:155], v[130:133], v[42:45], v[152:155], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[78], v[89], v[85], s[86]))
+  k.emit(v_perm_b32_e64(v[79], v[97], v[93], s[86]))
+  k.emit(v_mfma_16x16x32(v[156:159], v[130:133], v[46:49], v[156:159], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[80], v[105], v[101], s[86]))
+  k.emit(v_perm_b32_e64(v[81], v[113], v[109], s[86]))
+  k.emit(v_mfma_16x16x32(v[160:163], v[134:137], v[18:21], v[160:163], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[164:167], v[134:137], v[22:25], v[164:167], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[168:171], v[134:137], v[26:29], v[168:171], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[172:175], v[134:137], v[30:33], v[172:175], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[176:179], v[134:137], v[34:37], v[176:179], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_barrier())
+  k.emit(v_mfma_16x16x32(v[180:183], v[134:137], v[38:41], v[180:183], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[184:187], v[134:137], v[42:45], v[184:187], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[188:191], v[134:137], v[46:49], v[188:191], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[192:195], v[138:141], v[18:21], v[192:195], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[196:199], v[138:141], v[22:25], v[196:199], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[200:203], v[138:141], v[26:29], v[200:203], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[204:207], v[138:141], v[30:33], v[204:207], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[208:211], v[138:141], v[34:37], v[208:211], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[212:215], v[138:141], v[38:41], v[212:215], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[216:219], v[138:141], v[42:45], v[216:219], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[220:223], v[138:141], v[46:49], v[220:223], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[224:227], v[142:145], v[18:21], v[224:227], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.waitcnt(vm=17)
+  k.emit(v_mfma_16x16x32(v[228:231], v[142:145], v[22:25], v[228:231], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[232:235], v[142:145], v[26:29], v[232:235], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[236:239], v[142:145], v[30:33], v[236:239], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_barrier())
+  k.emit(v_mfma_16x16x32(v[240:243], v[142:145], v[34:37], v[240:243], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[244:247], v[142:145], v[38:41], v[244:247], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[248:251], v[142:145], v[42:45], v[248:251], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[252:255], v[142:145], v[46:49], v[252:255], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[0:3], v[146:149], v[50:53], v[0:3], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[4:7], v[146:149], v[54:57], v[4:7], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[8:11], v[146:149], v[58:61], v[8:11], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.waitcnt(vm=9)
+  k.emit(v_mfma_16x16x32(v[12:15], v[146:149], v[62:65], v[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[16:19], v[146:149], v[66:69], v[16:19], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_barrier())
+  k.emit(v_mfma_16x16x32(v[20:23], v[146:149], v[70:73], v[20:23], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[24:27], v[146:149], v[74:77], v[24:27], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[28:31], v[146:149], v[78:81], v[28:31], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[32:35], v[150:153], v[50:53], v[32:35], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[36:39], v[150:153], v[54:57], v[36:39], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.waitcnt(lgkm=4)
+  k.emit(v_mfma_16x16x32(v[40:43], v[150:153], v[58:61], v[40:43], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[44:47], v[150:153], v[62:65], v[44:47], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[18], v[86], v[82], s[85]))
+  k.emit(v_perm_b32_e64(v[19], v[94], v[90], s[85]))
+  k.emit(v_mfma_16x16x32(v[48:51], v[150:153], v[66:69], v[48:51], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[52:55], v[150:153], v[70:73], v[52:55], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[20], v[102], v[98], s[85]))
+  k.emit(v_perm_b32_e64(v[21], v[110], v[106], s[85]))
+  k.emit(v_mfma_16x16x32(v[56:59], v[150:153], v[74:77], v[56:59], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[60:63], v[150:153], v[78:81], v[60:63], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[22], v[86], v[82], s[86]))
+  k.emit(v_perm_b32_e64(v[23], v[94], v[90], s[86]))
+  k.emit(v_mfma_16x16x32(v[64:67], v[154:157], v[50:53], v[64:67], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[68:71], v[154:157], v[54:57], v[68:71], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[24], v[102], v[98], s[86]))
+  k.emit(v_perm_b32_e64(v[25], v[110], v[106], s[86]))
+  k.emit(v_mfma_16x16x32(v[72:75], v[154:157], v[58:61], v[72:75], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[76:79], v[154:157], v[62:65], v[76:79], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[26], v[87], v[83], s[85]))
+  k.emit(v_perm_b32_e64(v[27], v[95], v[91], s[85]))
+  k.emit(v_mfma_16x16x32(v[80:83], v[154:157], v[66:69], v[80:83], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[28], v[103], v[99], s[85]))
+  k.emit(v_perm_b32_e64(v[29], v[111], v[107], s[85]))
+  k.emit(v_mfma_16x16x32(v[84:87], v[154:157], v[70:73], v[84:87], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[30], v[87], v[83], s[86]))
+  k.emit(v_perm_b32_e64(v[31], v[95], v[91], s[86]))
+  k.emit(v_mfma_16x16x32(v[88:91], v[154:157], v[74:77], v[88:91], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[32], v[103], v[99], s[86]))
+  k.emit(v_perm_b32_e64(v[33], v[111], v[107], s[86]))
+  k.emit(v_mfma_16x16x32(v[92:95], v[154:157], v[78:81], v[92:95], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[96:99], v[158:161], v[50:53], v[96:99], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[34], v[88], v[84], s[85]))
+  k.emit(v_perm_b32_e64(v[35], v[96], v[92], s[85]))
+  k.emit(v_mfma_16x16x32(v[100:103], v[158:161], v[54:57], v[100:103], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[36], v[104], v[100], s[85]))
+  k.emit(v_perm_b32_e64(v[37], v[112], v[108], s[85]))
+  k.emit(v_mfma_16x16x32(v[104:107], v[158:161], v[58:61], v[104:107], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[108:111], v[158:161], v[62:65], v[108:111], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[38], v[88], v[84], s[86]))
+  k.emit(v_perm_b32_e64(v[39], v[96], v[92], s[86]))
+  k.emit(v_mfma_16x16x32(v[112:115], v[158:161], v[66:69], v[112:115], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[40], v[104], v[100], s[86]))
+  k.emit(v_perm_b32_e64(v[41], v[112], v[108], s[86]))
+  k.emit(v_mfma_16x16x32(v[116:119], v[158:161], v[70:73], v[116:119], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[120:123], v[158:161], v[74:77], v[120:123], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[42], v[89], v[85], s[85]))
+  k.emit(v_perm_b32_e64(v[43], v[97], v[93], s[85]))
+  k.emit(v_mfma_16x16x32(v[124:127], v[158:161], v[78:81], v[124:127], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[44], v[105], v[101], s[85]))
+  k.emit(v_perm_b32_e64(v[45], v[113], v[109], s[85]))
+  k.emit(v_mfma_16x16x32(v[128:131], v[162:165], v[50:53], v[128:131], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[132:135], v[162:165], v[54:57], v[132:135], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[46], v[89], v[85], s[86]))
+  k.emit(v_perm_b32_e64(v[47], v[97], v[93], s[86]))
+  k.emit(v_mfma_16x16x32(v[136:139], v[162:165], v[58:61], v[136:139], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_perm_b32_e64(v[48], v[105], v[101], s[86]))
+  k.emit(v_perm_b32_e64(v[49], v[113], v[109], s[86]))
+  k.emit(v_mfma_16x16x32(v[140:143], v[162:165], v[62:65], v[140:143], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[144:147], v[162:165], v[66:69], v[144:147], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[148:151], v[162:165], v[70:73], v[148:151], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[152:155], v[162:165], v[74:77], v[152:155], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[156:159], v[162:165], v[78:81], v[156:159], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[160:163], v[166:169], v[50:53], v[160:163], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[164:167], v[166:169], v[54:57], v[164:167], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[168:171], v[166:169], v[58:61], v[168:171], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[172:175], v[166:169], v[62:65], v[172:175], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[176:179], v[166:169], v[66:69], v[176:179], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[180:183], v[166:169], v[70:73], v[180:183], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[184:187], v[166:169], v[74:77], v[184:187], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[188:191], v[166:169], v[78:81], v[188:191], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[192:195], v[170:173], v[50:53], v[192:195], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[196:199], v[170:173], v[54:57], v[196:199], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[200:203], v[170:173], v[58:61], v[200:203], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[204:207], v[170:173], v[62:65], v[204:207], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[208:211], v[170:173], v[66:69], v[208:211], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[212:215], v[170:173], v[70:73], v[212:215], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[216:219], v[170:173], v[74:77], v[216:219], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[220:223], v[170:173], v[78:81], v[220:223], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[224:227], v[174:177], v[50:53], v[224:227], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[228:231], v[174:177], v[54:57], v[228:231], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[232:235], v[174:177], v[58:61], v[232:235], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[236:239], v[174:177], v[62:65], v[236:239], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[240:243], v[174:177], v[66:69], v[240:243], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[244:247], v[174:177], v[70:73], v[244:247], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[248:251], v[174:177], v[74:77], v[248:251], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[252:255], v[174:177], v[78:81], v[252:255], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.label('toPGR1end_OrdNLL')
+  k.emit(s_xor_b32(s[87], s[55], s[53]))
+  k.emit(s_min_u32(s[53], s[53], s[87]))
+  k.emit(s_xor_b32(s[87], s[56], s[54]))
+  k.emit(s_min_u32(s[54], s[54], s[87]))
+  k.emit(s_and_b32(s[8], 63, s[23]))
+  k.emit(s_cmp_lt_u32(s[61], s[46]))
+  k.emit(s_cmov_b32(s[8], 0))
+  k.emit(s_cmp_eq_u32(s[8], 0))
+  k.emit(s_mov_b32(s[9], 0))
+  k.emit(s_cbranch_scc1(), target='SkipTailLoopL')
+  k.emit(s_sub_i32(s[88], 3, s[78]))
+  k.emit(s_cmp_ge_i32(s[88], 0))
+  k.emit(s_cbranch_scc0(), target='Negative_LHNOKZ26V2FLOONQ')
+  k.emit(s_mul_hi_u32(s[89], s[88], s[83]))
+  k.emit(s_mul_i32(s[88], s[88], s[83]))
+  k.emit(s_branch(), target='MultiplyDone_L9DK3KJL31S8WWGN')
+  k.label('Negative_LHNOKZ26V2FLOONQ')
+  k.emit(s_abs_i32(s[88], s[88]))
+  k.emit(s_mul_hi_u32(s[89], s[88], s[83]))
+  k.emit(s_mul_i32(s[88], s[88], s[83]))
+  k.emit(s_xor_b32(s[88], s[88], -1))
+  k.emit(s_xor_b32(s[89], s[89], -1))
+  k.emit(s_add_u32(s[88], s[88], 1))
+  k.emit(s_addc_u32(s[89], s[89], 0))
+  k.label('MultiplyDone_L9DK3KJL31S8WWGN')
+  k.emit(s_sub_u32(s[88], s[88], s[79]))
+  k.emit(s_subb_u32(s[89], s[89], s[80]))
+  k.emit(s_add_u32(s[68], s[68], s[88]))
+  k.emit(s_addc_u32(s[69], s[69], s[89]))
+  k.emit(s_sub_u32(s[62], s[62], s[88]))
+  k.emit(s_subb_u32(s[63], s[63], s[89]))
+  k.emit(s_cmp_eq_u32(s[63], 0))
+  k.emit(s_cselect_b32(s[70], s[62], -1))
+  k.emit(s_sub_i32(s[88], 3, s[78]))
+  k.emit(s_cmp_ge_i32(s[88], 0))
+  k.emit(s_cbranch_scc0(), target='Negative_3U2TZUPK3AVX5ODG')
+  k.emit(s_mul_hi_u32(s[89], s[88], s[84]))
+  k.emit(s_mul_i32(s[88], s[88], s[84]))
+  k.emit(s_branch(), target='MultiplyDone_NW6XNGOG77EAT0NM')
+  k.label('Negative_3U2TZUPK3AVX5ODG')
+  k.emit(s_abs_i32(s[88], s[88]))
+  k.emit(s_mul_hi_u32(s[89], s[88], s[84]))
+  k.emit(s_mul_i32(s[88], s[88], s[84]))
+  k.emit(s_xor_b32(s[88], s[88], -1))
+  k.emit(s_xor_b32(s[89], s[89], -1))
+  k.emit(s_add_u32(s[88], s[88], 1))
+  k.emit(s_addc_u32(s[89], s[89], 0))
+  k.label('MultiplyDone_NW6XNGOG77EAT0NM')
+  k.emit(s_sub_u32(s[88], s[88], s[81]))
+  k.emit(s_subb_u32(s[89], s[89], s[82]))
+  k.emit(s_add_u32(s[72], s[72], s[88]))
+  k.emit(s_addc_u32(s[73], s[73], s[89]))
+  k.emit(s_sub_u32(s[76], s[76], s[88]))
+  k.emit(s_subb_u32(s[77], s[77], s[89]))
+  k.emit(s_cmp_eq_u32(s[77], 0))
+  k.emit(s_cselect_b32(s[74], s[76], -1))
+  k.emit(s_mov_b32(M0, s[53]))
+  k.waitcnt(lgkm=0)
+  k.emit(s_barrier())
+  k.emit(buffer_load_short_d16(v[18], v[0], s[68:71], 0, 0, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[84], v[0], s[68:71], 0, 2, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16(v[19], v[0], s[68:71], 0, 4, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[85], v[0], s[68:71], 0, 6, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16(v[20], v[0], s[68:71], 0, 8, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[86], v[0], s[68:71], 0, 10, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16(v[21], v[0], s[68:71], 0, 12, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[87], v[0], s[68:71], 0, 14, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16(v[22], v[1], s[68:71], 0, 0, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[88], v[1], s[68:71], 0, 2, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16(v[23], v[1], s[68:71], 0, 4, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[89], v[1], s[68:71], 0, 6, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16(v[24], v[1], s[68:71], 0, 8, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[90], v[1], s[68:71], 0, 10, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16(v[25], v[1], s[68:71], 0, 12, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[91], v[1], s[68:71], 0, 14, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16(v[26], v[2], s[68:71], 0, 0, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[92], v[2], s[68:71], 0, 2, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16(v[27], v[2], s[68:71], 0, 4, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[93], v[2], s[68:71], 0, 6, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16(v[28], v[2], s[68:71], 0, 8, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[94], v[2], s[68:71], 0, 10, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16(v[29], v[2], s[68:71], 0, 12, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[95], v[2], s[68:71], 0, 14, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16(v[30], v[3], s[68:71], 0, 0, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[96], v[3], s[68:71], 0, 2, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16(v[31], v[3], s[68:71], 0, 4, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[97], v[3], s[68:71], 0, 6, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16(v[32], v[3], s[68:71], 0, 8, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[98], v[3], s[68:71], 0, 10, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16(v[33], v[3], s[68:71], 0, 12, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[99], v[3], s[68:71], 0, 14, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16(v[34], v[4], s[68:71], 0, 0, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[100], v[4], s[68:71], 0, 2, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16(v[35], v[4], s[68:71], 0, 4, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[101], v[4], s[68:71], 0, 6, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16(v[36], v[4], s[68:71], 0, 8, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[102], v[4], s[68:71], 0, 10, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16(v[37], v[4], s[68:71], 0, 12, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[103], v[4], s[68:71], 0, 14, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16(v[38], v[5], s[68:71], 0, 0, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[104], v[5], s[68:71], 0, 2, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16(v[39], v[5], s[68:71], 0, 4, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[105], v[5], s[68:71], 0, 6, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16(v[40], v[5], s[68:71], 0, 8, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[106], v[5], s[68:71], 0, 10, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16(v[41], v[5], s[68:71], 0, 12, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[107], v[5], s[68:71], 0, 14, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16(v[42], v[6], s[68:71], 0, 0, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[108], v[6], s[68:71], 0, 2, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16(v[43], v[6], s[68:71], 0, 4, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[109], v[6], s[68:71], 0, 6, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16(v[44], v[6], s[68:71], 0, 8, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[110], v[6], s[68:71], 0, 10, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16(v[45], v[6], s[68:71], 0, 12, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[111], v[6], s[68:71], 0, 14, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16(v[46], v[7], s[68:71], 0, 0, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[112], v[7], s[68:71], 0, 2, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16(v[47], v[7], s[68:71], 0, 4, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[113], v[7], s[68:71], 0, 6, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16(v[48], v[7], s[68:71], 0, 8, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[114], v[7], s[68:71], 0, 10, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16(v[49], v[7], s[68:71], 0, 12, 1, 0, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[115], v[7], s[68:71], 0, 14, 1, 0, 0, 1))
+  k.waitcnt(vm=0)
+  k.emit(v_or_b32_e32(v[18], v[18], v[84]))
+  k.emit(v_or_b32_e32(v[19], v[19], v[85]))
+  k.emit(v_or_b32_e32(v[20], v[20], v[86]))
+  k.emit(v_or_b32_e32(v[21], v[21], v[87]))
+  k.emit(v_or_b32_e32(v[22], v[22], v[88]))
+  k.emit(v_or_b32_e32(v[23], v[23], v[89]))
+  k.emit(v_or_b32_e32(v[24], v[24], v[90]))
+  k.emit(v_or_b32_e32(v[25], v[25], v[91]))
+  k.emit(v_or_b32_e32(v[26], v[26], v[92]))
+  k.emit(v_or_b32_e32(v[27], v[27], v[93]))
+  k.emit(v_or_b32_e32(v[28], v[28], v[94]))
+  k.emit(v_or_b32_e32(v[29], v[29], v[95]))
+  k.emit(v_or_b32_e32(v[30], v[30], v[96]))
+  k.emit(v_or_b32_e32(v[31], v[31], v[97]))
+  k.emit(v_or_b32_e32(v[32], v[32], v[98]))
+  k.emit(v_or_b32_e32(v[33], v[33], v[99]))
+  k.emit(v_or_b32_e32(v[34], v[34], v[100]))
+  k.emit(v_or_b32_e32(v[35], v[35], v[101]))
+  k.emit(v_or_b32_e32(v[36], v[36], v[102]))
+  k.emit(v_or_b32_e32(v[37], v[37], v[103]))
+  k.emit(v_or_b32_e32(v[38], v[38], v[104]))
+  k.emit(v_or_b32_e32(v[39], v[39], v[105]))
+  k.emit(v_or_b32_e32(v[40], v[40], v[106]))
+  k.emit(v_or_b32_e32(v[41], v[41], v[107]))
+  k.emit(v_or_b32_e32(v[42], v[42], v[108]))
+  k.emit(v_or_b32_e32(v[43], v[43], v[109]))
+  k.emit(v_or_b32_e32(v[44], v[44], v[110]))
+  k.emit(v_or_b32_e32(v[45], v[45], v[111]))
+  k.emit(v_or_b32_e32(v[46], v[46], v[112]))
+  k.emit(v_or_b32_e32(v[47], v[47], v[113]))
+  k.emit(v_or_b32_e32(v[48], v[48], v[114]))
+  k.emit(v_or_b32_e32(v[49], v[49], v[115]))
+  k.emit(s_mov_b32(M0, 133120))
+  k.emit(s_mov_b32(M0, s[54]))
+  k.emit(buffer_load_short_d16(v[50], v[8], s[72:75], 0, 0, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[84], v[8], s[72:75], 0, 2, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16(v[51], v[8], s[72:75], 0, 4, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[85], v[8], s[72:75], 0, 6, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16(v[52], v[8], s[72:75], 0, 8, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[86], v[8], s[72:75], 0, 10, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16(v[53], v[8], s[72:75], 0, 12, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[87], v[8], s[72:75], 0, 14, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16(v[54], v[9], s[72:75], 0, 0, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[88], v[9], s[72:75], 0, 2, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16(v[55], v[9], s[72:75], 0, 4, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[89], v[9], s[72:75], 0, 6, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16(v[56], v[9], s[72:75], 0, 8, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[90], v[9], s[72:75], 0, 10, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16(v[57], v[9], s[72:75], 0, 12, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[91], v[9], s[72:75], 0, 14, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16(v[58], v[10], s[72:75], 0, 0, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[92], v[10], s[72:75], 0, 2, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16(v[59], v[10], s[72:75], 0, 4, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[93], v[10], s[72:75], 0, 6, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16(v[60], v[10], s[72:75], 0, 8, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[94], v[10], s[72:75], 0, 10, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16(v[61], v[10], s[72:75], 0, 12, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[95], v[10], s[72:75], 0, 14, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16(v[62], v[11], s[72:75], 0, 0, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[96], v[11], s[72:75], 0, 2, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16(v[63], v[11], s[72:75], 0, 4, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[97], v[11], s[72:75], 0, 6, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16(v[64], v[11], s[72:75], 0, 8, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[98], v[11], s[72:75], 0, 10, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16(v[65], v[11], s[72:75], 0, 12, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[99], v[11], s[72:75], 0, 14, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16(v[66], v[12], s[72:75], 0, 0, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[100], v[12], s[72:75], 0, 2, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16(v[67], v[12], s[72:75], 0, 4, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[101], v[12], s[72:75], 0, 6, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16(v[68], v[12], s[72:75], 0, 8, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[102], v[12], s[72:75], 0, 10, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16(v[69], v[12], s[72:75], 0, 12, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[103], v[12], s[72:75], 0, 14, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16(v[70], v[13], s[72:75], 0, 0, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[104], v[13], s[72:75], 0, 2, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16(v[71], v[13], s[72:75], 0, 4, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[105], v[13], s[72:75], 0, 6, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16(v[72], v[13], s[72:75], 0, 8, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[106], v[13], s[72:75], 0, 10, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16(v[73], v[13], s[72:75], 0, 12, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[107], v[13], s[72:75], 0, 14, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16(v[74], v[14], s[72:75], 0, 0, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[108], v[14], s[72:75], 0, 2, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16(v[75], v[14], s[72:75], 0, 4, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[109], v[14], s[72:75], 0, 6, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16(v[76], v[14], s[72:75], 0, 8, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[110], v[14], s[72:75], 0, 10, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16(v[77], v[14], s[72:75], 0, 12, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[111], v[14], s[72:75], 0, 14, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16(v[78], v[15], s[72:75], 0, 0, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[112], v[15], s[72:75], 0, 2, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16(v[79], v[15], s[72:75], 0, 4, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[113], v[15], s[72:75], 0, 6, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16(v[80], v[15], s[72:75], 0, 8, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[114], v[15], s[72:75], 0, 10, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16(v[81], v[15], s[72:75], 0, 12, 1, 0, 0, 0, 1, 0, 1))
+  k.emit(buffer_load_short_d16_hi(v[115], v[15], s[72:75], 0, 14, 1, 0, 0, 0, 1, 0, 1))
+  k.waitcnt(vm=0)
+  k.emit(v_or_b32_e32(v[50], v[50], v[84]))
+  k.emit(v_or_b32_e32(v[51], v[51], v[85]))
+  k.emit(v_or_b32_e32(v[52], v[52], v[86]))
+  k.emit(v_or_b32_e32(v[53], v[53], v[87]))
+  k.emit(v_or_b32_e32(v[54], v[54], v[88]))
+  k.emit(v_or_b32_e32(v[55], v[55], v[89]))
+  k.emit(v_or_b32_e32(v[56], v[56], v[90]))
+  k.emit(v_or_b32_e32(v[57], v[57], v[91]))
+  k.emit(v_or_b32_e32(v[58], v[58], v[92]))
+  k.emit(v_or_b32_e32(v[59], v[59], v[93]))
+  k.emit(v_or_b32_e32(v[60], v[60], v[94]))
+  k.emit(v_or_b32_e32(v[61], v[61], v[95]))
+  k.emit(v_or_b32_e32(v[62], v[62], v[96]))
+  k.emit(v_or_b32_e32(v[63], v[63], v[97]))
+  k.emit(v_or_b32_e32(v[64], v[64], v[98]))
+  k.emit(v_or_b32_e32(v[65], v[65], v[99]))
+  k.emit(v_or_b32_e32(v[66], v[66], v[100]))
+  k.emit(v_or_b32_e32(v[67], v[67], v[101]))
+  k.emit(v_or_b32_e32(v[68], v[68], v[102]))
+  k.emit(v_or_b32_e32(v[69], v[69], v[103]))
+  k.emit(v_or_b32_e32(v[70], v[70], v[104]))
+  k.emit(v_or_b32_e32(v[71], v[71], v[105]))
+  k.emit(v_or_b32_e32(v[72], v[72], v[106]))
+  k.emit(v_or_b32_e32(v[73], v[73], v[107]))
+  k.emit(v_or_b32_e32(v[74], v[74], v[108]))
+  k.emit(v_or_b32_e32(v[75], v[75], v[109]))
+  k.emit(v_or_b32_e32(v[76], v[76], v[110]))
+  k.emit(v_or_b32_e32(v[77], v[77], v[111]))
+  k.emit(v_or_b32_e32(v[78], v[78], v[112]))
+  k.emit(v_or_b32_e32(v[79], v[79], v[113]))
+  k.emit(v_or_b32_e32(v[80], v[80], v[114]))
+  k.emit(v_or_b32_e32(v[81], v[81], v[115]))
+  k.emit(s_mov_b32(M0, 133120))
+  k.waitcnt(vm=0)
+  k.emit(s_barrier())
+  k.emit(v_and_b32_e32(v[82], 63, v[180]))
+  k.emit(v_lshlrev_b32_e32(v[82], 4, v[82]))
+  k.emit(v_add_u32_e32(v[82], s[53], v[82]))
+  k.emit(v_and_b32_e32(v[83], 63, v[180]))
+  k.emit(v_lshlrev_b32_e32(v[83], 4, v[83]))
+  k.emit(v_add_u32_e32(v[83], s[54], v[83]))
+  k.emit(ds_write_b128(v[0], v[82], v[18:21]))
+  k.emit(ds_write_b128(v[0], v[82], v[22:25], v[0], 0, 0, 16))
+  k.emit(ds_write_b128(v[0], v[82], v[26:29], v[0], 0, 0, 32))
+  k.emit(ds_write_b128(v[0], v[82], v[30:33], v[0], 0, 0, 48))
+  k.emit(ds_write_b128(v[0], v[82], v[34:37], v[0], 0, 0, 64))
+  k.emit(ds_write_b128(v[0], v[82], v[38:41], v[0], 0, 0, 80))
+  k.emit(ds_write_b128(v[0], v[82], v[42:45], v[0], 0, 0, 96))
+  k.emit(ds_write_b128(v[0], v[82], v[46:49], v[0], 0, 0, 112))
+  k.emit(ds_write_b128(v[0], v[83], v[50:53]))
+  k.emit(ds_write_b128(v[0], v[83], v[54:57], v[0], 0, 128, 16))
+  k.emit(ds_write_b128(v[0], v[83], v[58:61], v[0], 0, 0, 33))
+  k.emit(ds_write_b128(v[0], v[83], v[62:65], v[0], 0, 128, 49))
+  k.emit(ds_write_b128(v[0], v[83], v[66:69], v[0], 0, 0, 66))
+  k.emit(ds_write_b128(v[0], v[83], v[70:73], v[0], 0, 128, 82))
+  k.emit(ds_write_b128(v[0], v[83], v[74:77], v[0], 0, 0, 99))
+  k.emit(ds_write_b128(v[0], v[83], v[78:81], v[0], 0, 128, 115))
+  k.waitcnt(lgkm=0)
+  k.emit(s_barrier())
+  k.emit(v_xor_b32_e32(v[181], v[178], v[16]))
+  k.emit(v_min_i32_e32(v[16], v[16], v[181]))
+  k.emit(v_xor_b32_e32(v[181], v[179], v[17]))
+  k.emit(v_min_i32_e32(v[17], v[17], v[181]))
+  k.label('TailLoopBeginL')
+  k.emit(ds_read_b128(v[82:85], v[16]))
+  k.emit(ds_read_b128(v[86:89], v[16], v[0], v[0], 0, 0, 2))
+  k.emit(ds_read_b128(v[90:93], v[16], v[0], v[0], 0, 0, 4))
+  k.emit(ds_read_b128(v[94:97], v[16], v[0], v[0], 0, 0, 6))
+  k.emit(ds_read_b128(v[98:101], v[16], v[0], v[0], 0, 0, 8))
+  k.emit(ds_read_b128(v[102:105], v[16], v[0], v[0], 0, 0, 10))
+  k.emit(ds_read_b128(v[106:109], v[16], v[0], v[0], 0, 0, 12))
+  k.emit(ds_read_b128(v[110:113], v[16], v[0], v[0], 0, 0, 14))
+  k.emit(ds_read_b128(v[114:117], v[17]))
+  k.emit(ds_read_b128(v[118:121], v[17], v[0], v[0], 0, 128))
+  k.emit(ds_read_b128(v[122:125], v[17], v[0], v[0], 0, 0, 1))
+  k.emit(ds_read_b128(v[126:129], v[17], v[0], v[0], 0, 128, 1))
+  k.emit(ds_read_b128(v[130:133], v[17], v[0], v[0], 0, 0, 2))
+  k.emit(ds_read_b128(v[134:137], v[17], v[0], v[0], 0, 128, 2))
+  k.emit(ds_read_b128(v[138:141], v[17], v[0], v[0], 0, 0, 3))
+  k.emit(ds_read_b128(v[142:145], v[17], v[0], v[0], 0, 128, 3))
+  k.emit(s_mov_b32(s[87], 16384))
+  k.emit(v_add_co_u32_e32(v[16], s[87], v[16]))
+  k.emit(s_mov_b32(s[87], 64))
+  k.emit(v_add_co_u32_e32(v[17], s[87], v[17]))
+  k.waitcnt(lgkm=0)
+  k.emit(v_perm_b32_e64(v[18], v[86], v[82], s[85]))
+  k.emit(v_perm_b32_e64(v[19], v[94], v[90], s[85]))
+  k.emit(v_perm_b32_e64(v[20], v[102], v[98], s[85]))
+  k.emit(v_perm_b32_e64(v[21], v[110], v[106], s[85]))
+  k.emit(v_perm_b32_e64(v[22], v[86], v[82], s[86]))
+  k.emit(v_perm_b32_e64(v[23], v[94], v[90], s[86]))
+  k.emit(v_perm_b32_e64(v[24], v[102], v[98], s[86]))
+  k.emit(v_perm_b32_e64(v[25], v[110], v[106], s[86]))
+  k.emit(v_perm_b32_e64(v[26], v[87], v[83], s[85]))
+  k.emit(v_perm_b32_e64(v[27], v[95], v[91], s[85]))
+  k.emit(v_perm_b32_e64(v[28], v[103], v[99], s[85]))
+  k.emit(v_perm_b32_e64(v[29], v[111], v[107], s[85]))
+  k.emit(v_perm_b32_e64(v[30], v[87], v[83], s[86]))
+  k.emit(v_perm_b32_e64(v[31], v[95], v[91], s[86]))
+  k.emit(v_perm_b32_e64(v[32], v[103], v[99], s[86]))
+  k.emit(v_perm_b32_e64(v[33], v[111], v[107], s[86]))
+  k.emit(v_perm_b32_e64(v[34], v[88], v[84], s[85]))
+  k.emit(v_perm_b32_e64(v[35], v[96], v[92], s[85]))
+  k.emit(v_perm_b32_e64(v[36], v[104], v[100], s[85]))
+  k.emit(v_perm_b32_e64(v[37], v[112], v[108], s[85]))
+  k.emit(v_perm_b32_e64(v[38], v[88], v[84], s[86]))
+  k.emit(v_perm_b32_e64(v[39], v[96], v[92], s[86]))
+  k.emit(v_perm_b32_e64(v[40], v[104], v[100], s[86]))
+  k.emit(v_perm_b32_e64(v[41], v[112], v[108], s[86]))
+  k.emit(v_perm_b32_e64(v[42], v[89], v[85], s[85]))
+  k.emit(v_perm_b32_e64(v[43], v[97], v[93], s[85]))
+  k.emit(v_perm_b32_e64(v[44], v[105], v[101], s[85]))
+  k.emit(v_perm_b32_e64(v[45], v[113], v[109], s[85]))
+  k.emit(v_perm_b32_e64(v[46], v[89], v[85], s[86]))
+  k.emit(v_perm_b32_e64(v[47], v[97], v[93], s[86]))
+  k.emit(v_perm_b32_e64(v[48], v[105], v[101], s[86]))
+  k.emit(v_perm_b32_e64(v[49], v[113], v[109], s[86]))
+  k.emit(v_and_b32_e32(v[181], 63, v[180]))
+  k.emit(v_lshrrev_b32_e32(v[181], 4, v[181]))
+  k.emit(v_lshlrev_b32_e32(v[181], 3, v[181]))
+  k.emit(v_add_u32_e64(v[182], v[181], 0))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[18], v[18], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[22], v[22], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[26], v[26], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[30], v[30], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[34], v[34], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[38], v[38], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[42], v[42], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[46], v[46], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[19], v[19], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[23], v[23], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[27], v[27], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[31], v[31], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[35], v[35], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[39], v[39], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[43], v[43], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[47], v[47], 0, s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[20], v[20], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[24], v[24], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[28], v[28], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[32], v[32], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[36], v[36], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[40], v[40], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[44], v[44], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[48], v[48], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[21], v[21], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[25], v[25], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[29], v[29], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[33], v[33], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[37], v[37], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[41], v[41], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[45], v[45], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[49], v[49], 0, s[88:89]))
+  k.emit(v_and_b32_e32(v[181], 63, v[180]))
+  k.emit(v_lshrrev_b32_e32(v[181], 4, v[181]))
+  k.emit(v_lshlrev_b32_e32(v[181], 3, v[181]))
+  k.emit(v_add_u32_e64(v[182], v[181], 0))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[114], v[114], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[118], v[118], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[122], v[122], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[126], v[126], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[130], v[130], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[134], v[134], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[138], v[138], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[142], v[142], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[115], v[115], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[119], v[119], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[123], v[123], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[127], v[127], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[131], v[131], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[135], v[135], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[139], v[139], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[143], v[143], 0, s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[116], v[116], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[120], v[120], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[124], v[124], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[128], v[128], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[132], v[132], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[136], v[136], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[140], v[140], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[144], v[144], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[117], v[117], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[121], v[121], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[125], v[125], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[129], v[129], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[133], v[133], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[137], v[137], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[141], v[141], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[145], v[145], 0, s[88:89]))
+  k.emit(s_and_b32(s[87], s[23], 7))
+  k.emit(s_cmp_eq_u32(s[87], 0))
+  k.emit(s_cbranch_scc1(), target='TailLoop_SkipZeroOutMask_0FMPG10PI1CDGWZ9')
+  k.emit(s_and_b32(s[87], s[8], 7))
+  k.emit(s_sub_u32(s[87], 8, s[87]))
+  k.emit(s_lshl_b32(s[87], s[87], 4))
+  k.emit(v_lshlrev_b64(v[184:185], s[87], v[18:19]))
+  k.emit(v_lshlrev_b64(v[186:187], s[87], v[20:21]))
+  k.emit(v_add_u32_e64(v[182], v[181], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[18], v[18], v[184], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[19], v[19], v[185], s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[20], v[20], v[186], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[21], v[21], v[187], s[88:89]))
+  k.emit(v_lshlrev_b64(v[184:185], s[87], v[22:23]))
+  k.emit(v_lshlrev_b64(v[186:187], s[87], v[24:25]))
+  k.emit(v_add_u32_e64(v[182], v[181], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[22], v[22], v[184], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[23], v[23], v[185], s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[24], v[24], v[186], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[25], v[25], v[187], s[88:89]))
+  k.emit(v_lshlrev_b64(v[184:185], s[87], v[26:27]))
+  k.emit(v_lshlrev_b64(v[186:187], s[87], v[28:29]))
+  k.emit(v_add_u32_e64(v[182], v[181], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[26], v[26], v[184], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[27], v[27], v[185], s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[28], v[28], v[186], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[29], v[29], v[187], s[88:89]))
+  k.emit(v_lshlrev_b64(v[184:185], s[87], v[30:31]))
+  k.emit(v_lshlrev_b64(v[186:187], s[87], v[32:33]))
+  k.emit(v_add_u32_e64(v[182], v[181], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[30], v[30], v[184], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[31], v[31], v[185], s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[32], v[32], v[186], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[33], v[33], v[187], s[88:89]))
+  k.emit(v_lshlrev_b64(v[184:185], s[87], v[34:35]))
+  k.emit(v_lshlrev_b64(v[186:187], s[87], v[36:37]))
+  k.emit(v_add_u32_e64(v[182], v[181], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[34], v[34], v[184], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[35], v[35], v[185], s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[36], v[36], v[186], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[37], v[37], v[187], s[88:89]))
+  k.emit(v_lshlrev_b64(v[184:185], s[87], v[38:39]))
+  k.emit(v_lshlrev_b64(v[186:187], s[87], v[40:41]))
+  k.emit(v_add_u32_e64(v[182], v[181], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[38], v[38], v[184], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[39], v[39], v[185], s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[40], v[40], v[186], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[41], v[41], v[187], s[88:89]))
+  k.emit(v_lshlrev_b64(v[184:185], s[87], v[42:43]))
+  k.emit(v_lshlrev_b64(v[186:187], s[87], v[44:45]))
+  k.emit(v_add_u32_e64(v[182], v[181], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[42], v[42], v[184], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[43], v[43], v[185], s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[44], v[44], v[186], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[45], v[45], v[187], s[88:89]))
+  k.emit(v_lshlrev_b64(v[184:185], s[87], v[46:47]))
+  k.emit(v_lshlrev_b64(v[186:187], s[87], v[48:49]))
+  k.emit(v_add_u32_e64(v[182], v[181], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[46], v[46], v[184], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[47], v[47], v[185], s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[48], v[48], v[186], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[49], v[49], v[187], s[88:89]))
+  k.emit(v_lshlrev_b64(v[184:185], s[87], v[114:115]))
+  k.emit(v_lshlrev_b64(v[186:187], s[87], v[116:117]))
+  k.emit(v_add_u32_e64(v[182], v[181], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[114], v[114], v[184], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[115], v[115], v[185], s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[116], v[116], v[186], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[117], v[117], v[187], s[88:89]))
+  k.emit(v_lshlrev_b64(v[184:185], s[87], v[118:119]))
+  k.emit(v_lshlrev_b64(v[186:187], s[87], v[120:121]))
+  k.emit(v_add_u32_e64(v[182], v[181], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[118], v[118], v[184], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[119], v[119], v[185], s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[120], v[120], v[186], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[121], v[121], v[187], s[88:89]))
+  k.emit(v_lshlrev_b64(v[184:185], s[87], v[122:123]))
+  k.emit(v_lshlrev_b64(v[186:187], s[87], v[124:125]))
+  k.emit(v_add_u32_e64(v[182], v[181], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[122], v[122], v[184], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[123], v[123], v[185], s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[124], v[124], v[186], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[125], v[125], v[187], s[88:89]))
+  k.emit(v_lshlrev_b64(v[184:185], s[87], v[126:127]))
+  k.emit(v_lshlrev_b64(v[186:187], s[87], v[128:129]))
+  k.emit(v_add_u32_e64(v[182], v[181], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[126], v[126], v[184], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[127], v[127], v[185], s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[128], v[128], v[186], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[129], v[129], v[187], s[88:89]))
+  k.emit(v_lshlrev_b64(v[184:185], s[87], v[130:131]))
+  k.emit(v_lshlrev_b64(v[186:187], s[87], v[132:133]))
+  k.emit(v_add_u32_e64(v[182], v[181], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[130], v[130], v[184], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[131], v[131], v[185], s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[132], v[132], v[186], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[133], v[133], v[187], s[88:89]))
+  k.emit(v_lshlrev_b64(v[184:185], s[87], v[134:135]))
+  k.emit(v_lshlrev_b64(v[186:187], s[87], v[136:137]))
+  k.emit(v_add_u32_e64(v[182], v[181], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[134], v[134], v[184], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[135], v[135], v[185], s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[136], v[136], v[186], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[137], v[137], v[187], s[88:89]))
+  k.emit(v_lshlrev_b64(v[184:185], s[87], v[138:139]))
+  k.emit(v_lshlrev_b64(v[186:187], s[87], v[140:141]))
+  k.emit(v_add_u32_e64(v[182], v[181], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[138], v[138], v[184], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[139], v[139], v[185], s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[140], v[140], v[186], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[141], v[141], v[187], s[88:89]))
+  k.emit(v_lshlrev_b64(v[184:185], s[87], v[142:143]))
+  k.emit(v_lshlrev_b64(v[186:187], s[87], v[144:145]))
+  k.emit(v_add_u32_e64(v[182], v[181], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[142], v[142], v[184], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[143], v[143], v[185], s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[144], v[144], v[186], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[145], v[145], v[187], s[88:89]))
+  k.label('TailLoop_SkipZeroOutMask_0FMPG10PI1CDGWZ9')
+  k.emit(s_nop(1))
+  k.emit(v_mfma_16x16x32(v[0:3], v[114:117], v[18:21], v[0:3], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[4:7], v[114:117], v[22:25], v[4:7], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[8:11], v[114:117], v[26:29], v[8:11], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[12:15], v[114:117], v[30:33], v[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[16:19], v[114:117], v[34:37], v[16:19], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[20:23], v[114:117], v[38:41], v[20:23], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[24:27], v[114:117], v[42:45], v[24:27], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[28:31], v[114:117], v[46:49], v[28:31], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[32:35], v[118:121], v[18:21], v[32:35], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[36:39], v[118:121], v[22:25], v[36:39], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[40:43], v[118:121], v[26:29], v[40:43], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[44:47], v[118:121], v[30:33], v[44:47], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[48:51], v[118:121], v[34:37], v[48:51], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[52:55], v[118:121], v[38:41], v[52:55], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[56:59], v[118:121], v[42:45], v[56:59], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[60:63], v[118:121], v[46:49], v[60:63], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[64:67], v[122:125], v[18:21], v[64:67], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[68:71], v[122:125], v[22:25], v[68:71], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[72:75], v[122:125], v[26:29], v[72:75], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[76:79], v[122:125], v[30:33], v[76:79], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[80:83], v[122:125], v[34:37], v[80:83], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[84:87], v[122:125], v[38:41], v[84:87], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[88:91], v[122:125], v[42:45], v[88:91], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[92:95], v[122:125], v[46:49], v[92:95], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[96:99], v[126:129], v[18:21], v[96:99], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[100:103], v[126:129], v[22:25], v[100:103], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[104:107], v[126:129], v[26:29], v[104:107], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[108:111], v[126:129], v[30:33], v[108:111], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[112:115], v[126:129], v[34:37], v[112:115], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[116:119], v[126:129], v[38:41], v[116:119], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[120:123], v[126:129], v[42:45], v[120:123], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[124:127], v[126:129], v[46:49], v[124:127], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[128:131], v[130:133], v[18:21], v[128:131], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[132:135], v[130:133], v[22:25], v[132:135], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[136:139], v[130:133], v[26:29], v[136:139], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[140:143], v[130:133], v[30:33], v[140:143], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[144:147], v[130:133], v[34:37], v[144:147], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[148:151], v[130:133], v[38:41], v[148:151], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[152:155], v[130:133], v[42:45], v[152:155], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[156:159], v[130:133], v[46:49], v[156:159], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[160:163], v[134:137], v[18:21], v[160:163], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[164:167], v[134:137], v[22:25], v[164:167], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[168:171], v[134:137], v[26:29], v[168:171], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[172:175], v[134:137], v[30:33], v[172:175], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[176:179], v[134:137], v[34:37], v[176:179], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[180:183], v[134:137], v[38:41], v[180:183], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[184:187], v[134:137], v[42:45], v[184:187], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[188:191], v[134:137], v[46:49], v[188:191], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[192:195], v[138:141], v[18:21], v[192:195], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[196:199], v[138:141], v[22:25], v[196:199], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[200:203], v[138:141], v[26:29], v[200:203], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[204:207], v[138:141], v[30:33], v[204:207], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[208:211], v[138:141], v[34:37], v[208:211], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[212:215], v[138:141], v[38:41], v[212:215], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[216:219], v[138:141], v[42:45], v[216:219], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[220:223], v[138:141], v[46:49], v[220:223], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[224:227], v[142:145], v[18:21], v[224:227], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[228:231], v[142:145], v[22:25], v[228:231], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[232:235], v[142:145], v[26:29], v[232:235], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[236:239], v[142:145], v[30:33], v[236:239], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[240:243], v[142:145], v[34:37], v[240:243], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[244:247], v[142:145], v[38:41], v[244:247], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[248:251], v[142:145], v[42:45], v[248:251], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[252:255], v[142:145], v[46:49], v[252:255], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_sub_i32(s[8], s[8], 32))
+  k.emit(s_add_u32(s[9], s[9], 32))
+  k.emit(s_cmp_le_i32(s[8], 0))
+  k.emit(s_cbranch_scc1(), target='TailLoopEndL')
+  k.emit(ds_read_b128(v[82:85], v[16]))
+  k.emit(ds_read_b128(v[86:89], v[16], v[0], v[0], 0, 0, 2))
+  k.emit(ds_read_b128(v[90:93], v[16], v[0], v[0], 0, 0, 4))
+  k.emit(ds_read_b128(v[94:97], v[16], v[0], v[0], 0, 0, 6))
+  k.emit(ds_read_b128(v[98:101], v[16], v[0], v[0], 0, 0, 8))
+  k.emit(ds_read_b128(v[102:105], v[16], v[0], v[0], 0, 0, 10))
+  k.emit(ds_read_b128(v[106:109], v[16], v[0], v[0], 0, 0, 12))
+  k.emit(ds_read_b128(v[110:113], v[16], v[0], v[0], 0, 0, 14))
+  k.emit(ds_read_b128(v[146:149], v[17]))
+  k.emit(ds_read_b128(v[150:153], v[17], v[0], v[0], 0, 128))
+  k.emit(ds_read_b128(v[154:157], v[17], v[0], v[0], 0, 0, 1))
+  k.emit(ds_read_b128(v[158:161], v[17], v[0], v[0], 0, 128, 1))
+  k.emit(ds_read_b128(v[162:165], v[17], v[0], v[0], 0, 0, 2))
+  k.emit(ds_read_b128(v[166:169], v[17], v[0], v[0], 0, 128, 2))
+  k.emit(ds_read_b128(v[170:173], v[17], v[0], v[0], 0, 0, 3))
+  k.emit(ds_read_b128(v[174:177], v[17], v[0], v[0], 0, 128, 3))
+  k.emit(s_mov_b32(s[87], 16384))
+  k.emit(v_add_co_u32_e32(v[16], s[87], v[16]))
+  k.emit(s_mov_b32(s[87], 64))
+  k.emit(v_add_co_u32_e32(v[17], s[87], v[17]))
+  k.waitcnt(lgkm=0)
+  k.emit(v_perm_b32_e64(v[50], v[86], v[82], s[85]))
+  k.emit(v_perm_b32_e64(v[51], v[94], v[90], s[85]))
+  k.emit(v_perm_b32_e64(v[52], v[102], v[98], s[85]))
+  k.emit(v_perm_b32_e64(v[53], v[110], v[106], s[85]))
+  k.emit(v_perm_b32_e64(v[54], v[86], v[82], s[86]))
+  k.emit(v_perm_b32_e64(v[55], v[94], v[90], s[86]))
+  k.emit(v_perm_b32_e64(v[56], v[102], v[98], s[86]))
+  k.emit(v_perm_b32_e64(v[57], v[110], v[106], s[86]))
+  k.emit(v_perm_b32_e64(v[58], v[87], v[83], s[85]))
+  k.emit(v_perm_b32_e64(v[59], v[95], v[91], s[85]))
+  k.emit(v_perm_b32_e64(v[60], v[103], v[99], s[85]))
+  k.emit(v_perm_b32_e64(v[61], v[111], v[107], s[85]))
+  k.emit(v_perm_b32_e64(v[62], v[87], v[83], s[86]))
+  k.emit(v_perm_b32_e64(v[63], v[95], v[91], s[86]))
+  k.emit(v_perm_b32_e64(v[64], v[103], v[99], s[86]))
+  k.emit(v_perm_b32_e64(v[65], v[111], v[107], s[86]))
+  k.emit(v_perm_b32_e64(v[66], v[88], v[84], s[85]))
+  k.emit(v_perm_b32_e64(v[67], v[96], v[92], s[85]))
+  k.emit(v_perm_b32_e64(v[68], v[104], v[100], s[85]))
+  k.emit(v_perm_b32_e64(v[69], v[112], v[108], s[85]))
+  k.emit(v_perm_b32_e64(v[70], v[88], v[84], s[86]))
+  k.emit(v_perm_b32_e64(v[71], v[96], v[92], s[86]))
+  k.emit(v_perm_b32_e64(v[72], v[104], v[100], s[86]))
+  k.emit(v_perm_b32_e64(v[73], v[112], v[108], s[86]))
+  k.emit(v_perm_b32_e64(v[74], v[89], v[85], s[85]))
+  k.emit(v_perm_b32_e64(v[75], v[97], v[93], s[85]))
+  k.emit(v_perm_b32_e64(v[76], v[105], v[101], s[85]))
+  k.emit(v_perm_b32_e64(v[77], v[113], v[109], s[85]))
+  k.emit(v_perm_b32_e64(v[78], v[89], v[85], s[86]))
+  k.emit(v_perm_b32_e64(v[79], v[97], v[93], s[86]))
+  k.emit(v_perm_b32_e64(v[80], v[105], v[101], s[86]))
+  k.emit(v_perm_b32_e64(v[81], v[113], v[109], s[86]))
+  k.emit(v_and_b32_e32(v[181], 63, v[180]))
+  k.emit(v_lshrrev_b32_e32(v[181], 4, v[181]))
+  k.emit(v_lshlrev_b32_e32(v[181], 3, v[181]))
+  k.emit(v_add_u32_e64(v[182], v[181], 0))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[50], v[50], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[54], v[54], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[58], v[58], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[62], v[62], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[66], v[66], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[70], v[70], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[74], v[74], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[78], v[78], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[51], v[51], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[55], v[55], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[59], v[59], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[63], v[63], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[67], v[67], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[71], v[71], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[75], v[75], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[79], v[79], 0, s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[52], v[52], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[56], v[56], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[60], v[60], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[64], v[64], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[68], v[68], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[72], v[72], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[76], v[76], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[80], v[80], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[53], v[53], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[57], v[57], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[61], v[61], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[65], v[65], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[69], v[69], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[73], v[73], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[77], v[77], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[81], v[81], 0, s[88:89]))
+  k.emit(v_and_b32_e32(v[181], 63, v[180]))
+  k.emit(v_lshrrev_b32_e32(v[181], 4, v[181]))
+  k.emit(v_lshlrev_b32_e32(v[181], 3, v[181]))
+  k.emit(v_add_u32_e64(v[182], v[181], 0))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[146], v[146], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[150], v[150], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[154], v[154], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[158], v[158], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[162], v[162], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[166], v[166], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[170], v[170], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[174], v[174], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[147], v[147], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[151], v[151], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[155], v[155], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[159], v[159], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[163], v[163], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[167], v[167], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[171], v[171], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[175], v[175], 0, s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[148], v[148], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[152], v[152], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[156], v[156], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[160], v[160], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[164], v[164], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[168], v[168], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[172], v[172], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[176], v[176], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[149], v[149], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[153], v[153], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[157], v[157], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[161], v[161], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[165], v[165], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[169], v[169], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[173], v[173], 0, s[88:89]))
+  k.emit(v_cndmask_b32_e64(v[177], v[177], 0, s[88:89]))
+  k.emit(s_and_b32(s[87], s[23], 7))
+  k.emit(s_cmp_eq_u32(s[87], 0))
+  k.emit(s_cbranch_scc1(), target='TailLoop_SkipZeroOutMask_YVWB1RHZO1Z7SCZY')
+  k.emit(s_and_b32(s[87], s[8], 7))
+  k.emit(s_sub_u32(s[87], 8, s[87]))
+  k.emit(s_lshl_b32(s[87], s[87], 4))
+  k.emit(v_lshlrev_b64(v[184:185], s[87], v[50:51]))
+  k.emit(v_lshlrev_b64(v[186:187], s[87], v[52:53]))
+  k.emit(v_add_u32_e64(v[182], v[181], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[50], v[50], v[184], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[51], v[51], v[185], s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[52], v[52], v[186], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[53], v[53], v[187], s[88:89]))
+  k.emit(v_lshlrev_b64(v[184:185], s[87], v[54:55]))
+  k.emit(v_lshlrev_b64(v[186:187], s[87], v[56:57]))
+  k.emit(v_add_u32_e64(v[182], v[181], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[54], v[54], v[184], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[55], v[55], v[185], s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[56], v[56], v[186], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[57], v[57], v[187], s[88:89]))
+  k.emit(v_lshlrev_b64(v[184:185], s[87], v[58:59]))
+  k.emit(v_lshlrev_b64(v[186:187], s[87], v[60:61]))
+  k.emit(v_add_u32_e64(v[182], v[181], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[58], v[58], v[184], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[59], v[59], v[185], s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[60], v[60], v[186], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[61], v[61], v[187], s[88:89]))
+  k.emit(v_lshlrev_b64(v[184:185], s[87], v[62:63]))
+  k.emit(v_lshlrev_b64(v[186:187], s[87], v[64:65]))
+  k.emit(v_add_u32_e64(v[182], v[181], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[62], v[62], v[184], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[63], v[63], v[185], s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[64], v[64], v[186], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[65], v[65], v[187], s[88:89]))
+  k.emit(v_lshlrev_b64(v[184:185], s[87], v[66:67]))
+  k.emit(v_lshlrev_b64(v[186:187], s[87], v[68:69]))
+  k.emit(v_add_u32_e64(v[182], v[181], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[66], v[66], v[184], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[67], v[67], v[185], s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[68], v[68], v[186], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[69], v[69], v[187], s[88:89]))
+  k.emit(v_lshlrev_b64(v[184:185], s[87], v[70:71]))
+  k.emit(v_lshlrev_b64(v[186:187], s[87], v[72:73]))
+  k.emit(v_add_u32_e64(v[182], v[181], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[70], v[70], v[184], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[71], v[71], v[185], s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[72], v[72], v[186], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[73], v[73], v[187], s[88:89]))
+  k.emit(v_lshlrev_b64(v[184:185], s[87], v[74:75]))
+  k.emit(v_lshlrev_b64(v[186:187], s[87], v[76:77]))
+  k.emit(v_add_u32_e64(v[182], v[181], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[74], v[74], v[184], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[75], v[75], v[185], s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[76], v[76], v[186], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[77], v[77], v[187], s[88:89]))
+  k.emit(v_lshlrev_b64(v[184:185], s[87], v[78:79]))
+  k.emit(v_lshlrev_b64(v[186:187], s[87], v[80:81]))
+  k.emit(v_add_u32_e64(v[182], v[181], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[78], v[78], v[184], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[79], v[79], v[185], s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[80], v[80], v[186], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[81], v[81], v[187], s[88:89]))
+  k.emit(v_lshlrev_b64(v[184:185], s[87], v[146:147]))
+  k.emit(v_lshlrev_b64(v[186:187], s[87], v[148:149]))
+  k.emit(v_add_u32_e64(v[182], v[181], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[146], v[146], v[184], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[147], v[147], v[185], s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[148], v[148], v[186], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[149], v[149], v[187], s[88:89]))
+  k.emit(v_lshlrev_b64(v[184:185], s[87], v[150:151]))
+  k.emit(v_lshlrev_b64(v[186:187], s[87], v[152:153]))
+  k.emit(v_add_u32_e64(v[182], v[181], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[150], v[150], v[184], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[151], v[151], v[185], s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[152], v[152], v[186], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[153], v[153], v[187], s[88:89]))
+  k.emit(v_lshlrev_b64(v[184:185], s[87], v[154:155]))
+  k.emit(v_lshlrev_b64(v[186:187], s[87], v[156:157]))
+  k.emit(v_add_u32_e64(v[182], v[181], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[154], v[154], v[184], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[155], v[155], v[185], s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[156], v[156], v[186], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[157], v[157], v[187], s[88:89]))
+  k.emit(v_lshlrev_b64(v[184:185], s[87], v[158:159]))
+  k.emit(v_lshlrev_b64(v[186:187], s[87], v[160:161]))
+  k.emit(v_add_u32_e64(v[182], v[181], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[158], v[158], v[184], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[159], v[159], v[185], s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[160], v[160], v[186], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[161], v[161], v[187], s[88:89]))
+  k.emit(v_lshlrev_b64(v[184:185], s[87], v[162:163]))
+  k.emit(v_lshlrev_b64(v[186:187], s[87], v[164:165]))
+  k.emit(v_add_u32_e64(v[182], v[181], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[162], v[162], v[184], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[163], v[163], v[185], s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[164], v[164], v[186], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[165], v[165], v[187], s[88:89]))
+  k.emit(v_lshlrev_b64(v[184:185], s[87], v[166:167]))
+  k.emit(v_lshlrev_b64(v[186:187], s[87], v[168:169]))
+  k.emit(v_add_u32_e64(v[182], v[181], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[166], v[166], v[184], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[167], v[167], v[185], s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[168], v[168], v[186], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[169], v[169], v[187], s[88:89]))
+  k.emit(v_lshlrev_b64(v[184:185], s[87], v[170:171]))
+  k.emit(v_lshlrev_b64(v[186:187], s[87], v[172:173]))
+  k.emit(v_add_u32_e64(v[182], v[181], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[170], v[170], v[184], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[171], v[171], v[185], s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[172], v[172], v[186], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[173], v[173], v[187], s[88:89]))
+  k.emit(v_lshlrev_b64(v[184:185], s[87], v[174:175]))
+  k.emit(v_lshlrev_b64(v[186:187], s[87], v[176:177]))
+  k.emit(v_add_u32_e64(v[182], v[181], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[174], v[174], v[184], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[175], v[175], v[185], s[88:89]))
+  k.emit(v_add_u32_e64(v[182], v[182], 4))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[176], v[176], v[186], s[88:89]))
+  k.emit(v_cmp_ge_i32_e64(s[88:89], v[182], s[8]))
+  k.emit(v_cndmask_b32_e64(v[177], v[177], v[187], s[88:89]))
+  k.label('TailLoop_SkipZeroOutMask_YVWB1RHZO1Z7SCZY')
+  k.emit(s_nop(1))
+  k.emit(v_mfma_16x16x32(v[0:3], v[146:149], v[50:53], v[0:3], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[4:7], v[146:149], v[54:57], v[4:7], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[8:11], v[146:149], v[58:61], v[8:11], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[12:15], v[146:149], v[62:65], v[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[16:19], v[146:149], v[66:69], v[16:19], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[20:23], v[146:149], v[70:73], v[20:23], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[24:27], v[146:149], v[74:77], v[24:27], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[28:31], v[146:149], v[78:81], v[28:31], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[32:35], v[150:153], v[50:53], v[32:35], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[36:39], v[150:153], v[54:57], v[36:39], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[40:43], v[150:153], v[58:61], v[40:43], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[44:47], v[150:153], v[62:65], v[44:47], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[48:51], v[150:153], v[66:69], v[48:51], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[52:55], v[150:153], v[70:73], v[52:55], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[56:59], v[150:153], v[74:77], v[56:59], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[60:63], v[150:153], v[78:81], v[60:63], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[64:67], v[154:157], v[50:53], v[64:67], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[68:71], v[154:157], v[54:57], v[68:71], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[72:75], v[154:157], v[58:61], v[72:75], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[76:79], v[154:157], v[62:65], v[76:79], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[80:83], v[154:157], v[66:69], v[80:83], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[84:87], v[154:157], v[70:73], v[84:87], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[88:91], v[154:157], v[74:77], v[88:91], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[92:95], v[154:157], v[78:81], v[92:95], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[96:99], v[158:161], v[50:53], v[96:99], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[100:103], v[158:161], v[54:57], v[100:103], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[104:107], v[158:161], v[58:61], v[104:107], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[108:111], v[158:161], v[62:65], v[108:111], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[112:115], v[158:161], v[66:69], v[112:115], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[116:119], v[158:161], v[70:73], v[116:119], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[120:123], v[158:161], v[74:77], v[120:123], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[124:127], v[158:161], v[78:81], v[124:127], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[128:131], v[162:165], v[50:53], v[128:131], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[132:135], v[162:165], v[54:57], v[132:135], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[136:139], v[162:165], v[58:61], v[136:139], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[140:143], v[162:165], v[62:65], v[140:143], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[144:147], v[162:165], v[66:69], v[144:147], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[148:151], v[162:165], v[70:73], v[148:151], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[152:155], v[162:165], v[74:77], v[152:155], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[156:159], v[162:165], v[78:81], v[156:159], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[160:163], v[166:169], v[50:53], v[160:163], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[164:167], v[166:169], v[54:57], v[164:167], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[168:171], v[166:169], v[58:61], v[168:171], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[172:175], v[166:169], v[62:65], v[172:175], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[176:179], v[166:169], v[66:69], v[176:179], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[180:183], v[166:169], v[70:73], v[180:183], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[184:187], v[166:169], v[74:77], v[184:187], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[188:191], v[166:169], v[78:81], v[188:191], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[192:195], v[170:173], v[50:53], v[192:195], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[196:199], v[170:173], v[54:57], v[196:199], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[200:203], v[170:173], v[58:61], v[200:203], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[204:207], v[170:173], v[62:65], v[204:207], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[208:211], v[170:173], v[66:69], v[208:211], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[212:215], v[170:173], v[70:73], v[212:215], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[216:219], v[170:173], v[74:77], v[216:219], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[220:223], v[170:173], v[78:81], v[220:223], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[224:227], v[174:177], v[50:53], v[224:227], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[228:231], v[174:177], v[54:57], v[228:231], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[232:235], v[174:177], v[58:61], v[232:235], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[236:239], v[174:177], v[62:65], v[236:239], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[240:243], v[174:177], v[66:69], v[240:243], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[244:247], v[174:177], v[70:73], v[244:247], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[248:251], v[174:177], v[74:77], v[248:251], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mfma_16x16x32(v[252:255], v[174:177], v[78:81], v[252:255], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_sub_i32(s[8], s[8], 32))
+  k.emit(s_add_u32(s[9], s[9], 32))
+  k.emit(s_cmp_le_i32(s[8], 0))
+  k.emit(s_cbranch_scc0(), target='TailLoopBeginL')
+  k.label('TailLoopEndL')
+  k.emit(s_mov_b32(s[87], 512))
+  k.emit(s_mul_i32(s[87], s[9], s[87]))
+  k.emit(v_sub_u32_e64(v[16], v[16], s[87]))
+  k.emit(s_mov_b32(s[87], 2))
+  k.emit(s_mul_i32(s[87], s[9], s[87]))
+  k.emit(v_sub_u32_e64(v[17], v[17], s[87]))
+  k.label('SkipTailLoopL')
+  k.emit(s_setprio())
+  k.emit(s_mov_b64(s[68:69], 0))
+  k.emit(s_mov_b32(s[72], 0))
+  k.emit(v_mov_b32_e32(v[21], s[2]))
+  k.emit(v_mul_i32_i24_e32(v[21], 4294967040, v[21]))
+  k.emit(v_add_co_u32_e32(v[21], s[20], v[21]))
+  k.emit(v_mov_b32_e32(v[22], 256))
+  k.emit(v_cmp_lt_u32_e64(s[8:9], v[21], v[22]))
+  k.emit(v_cndmask_b32_e64(v[21], v[22], v[21], s[8:9]))
+  k.emit(v_lshrrev_b32_e32(v[23], 6, v[180]))
+  k.emit(v_and_b32_e32(v[23], 1, v[23]))
+  k.emit(v_lshrrev_b32_e32(v[24], 7, v[21]))
+  k.emit(v_and_b32_e32(v[24], 1, v[24]))
+  k.emit(v_cmp_eq_u32_e64(s[8:9], v[24], v[23]))
+  k.emit(v_cndmask_b32_e64(v[21], v[22], v[21], s[8:9]))
+  k.emit(v_lshrrev_b32_e32(v[22], 7, v[21]))
+  k.emit(v_lshlrev_b32_e32(v[24], 0, v[23]))
+  k.emit(v_sub_u32_e32(v[22], v[22], v[24]))
+  k.emit(v_lshrrev_b32_e32(v[24], 3, v[21]))
+  k.emit(v_lshrrev_b32_e32(v[25], 0, v[180]))
+  k.emit(v_and_b32_e32(v[25], 15, v[25]))
+  k.emit(v_lshlrev_b32_e32(v[25], 3, v[25]))
+  k.emit(v_lshrrev_b32_e32(v[25], 3, v[25]))
+  k.emit(v_lshlrev_b32_e32(v[23], 4, v[23]))
+  k.emit(v_add_co_u32_e32(v[25], v[23], v[25]))
+  k.emit(v_sub_u32_e32(v[24], v[24], v[25]))
+  k.emit(v_and_b32_e32(v[23], 7, v[21]))
+  k.emit(v_lshrrev_b32_e32(v[23], 3, v[23]))
+  k.emit(v_and_b32_e32(v[25], 7, v[21]))
+  k.emit(v_cmp_eq_u32_e64(VCC, v[25], 1))
+  k.emit(s_cbranch_vccnz(), target='ShiftVectorComponents0_GLVW1')
+  k.emit(v_cmp_eq_u32_e64(VCC, v[25], 2))
+  k.emit(s_cbranch_vccnz(), target='ShiftVectorComponents0_GLVW2')
+  k.emit(v_cmp_eq_u32_e64(VCC, v[25], 3))
+  k.emit(s_cbranch_vccnz(), target='ShiftVectorComponents0_GLVW3')
+  k.emit(v_cmp_eq_u32_e64(VCC, v[25], 4))
+  k.emit(s_cbranch_vccnz(), target='ShiftVectorComponents0_GLVW4')
+  k.emit(v_cmp_eq_u32_e64(VCC, v[25], 5))
+  k.emit(s_cbranch_vccnz(), target='ShiftVectorComponents0_GLVW5')
+  k.emit(v_cmp_eq_u32_e64(VCC, v[25], 6))
+  k.emit(s_cbranch_vccnz(), target='ShiftVectorComponents0_GLVW6')
+  k.emit(v_cmp_eq_u32_e64(VCC, v[25], 7))
+  k.emit(s_cbranch_vccnz(), target='ShiftVectorComponents0_GLVW7')
+  k.emit(s_branch(), target='ShiftVectorComponents0_GLVW0')
+  k.label('ShiftVectorComponents0_GLVW1')
+  k.emit(v_cmp_eq_u32_e64(VCC, v[22], 0))
+  k.emit(s_cbranch_vccnz(), target='ShiftVectorComponents0_GLVW1_BM0')
+  k.label('ShiftVectorComponents0_GLVW2')
+  k.emit(v_cmp_eq_u32_e64(VCC, v[22], 0))
+  k.emit(s_cbranch_vccnz(), target='ShiftVectorComponents0_GLVW2_BM0')
+  k.label('ShiftVectorComponents0_GLVW3')
+  k.emit(v_cmp_eq_u32_e64(VCC, v[22], 0))
+  k.emit(s_cbranch_vccnz(), target='ShiftVectorComponents0_GLVW3_BM0')
+  k.label('ShiftVectorComponents0_GLVW4')
+  k.emit(v_cmp_eq_u32_e64(VCC, v[22], 0))
+  k.emit(s_cbranch_vccnz(), target='ShiftVectorComponents0_GLVW4_BM0')
+  k.label('ShiftVectorComponents0_GLVW5')
+  k.emit(v_cmp_eq_u32_e64(VCC, v[22], 0))
+  k.emit(s_cbranch_vccnz(), target='ShiftVectorComponents0_GLVW5_BM0')
+  k.label('ShiftVectorComponents0_GLVW6')
+  k.emit(v_cmp_eq_u32_e64(VCC, v[22], 0))
+  k.emit(s_cbranch_vccnz(), target='ShiftVectorComponents0_GLVW6_BM0')
+  k.label('ShiftVectorComponents0_GLVW7')
+  k.emit(v_cmp_eq_u32_e64(VCC, v[22], 0))
+  k.emit(s_cbranch_vccnz(), target='ShiftVectorComponents0_GLVW7_BM0')
+  k.label('ShiftVectorComponents0_GLVW1_BM0')
+  k.emit(v_cmp_eq_u32_e64(VCC, v[23], 0))
+  k.emit(s_cbranch_vccnz(), target='ShiftVectorComponents0_GLVW1_BM0_VW0')
+  k.label('ShiftVectorComponents0_GLVW2_BM0')
+  k.emit(v_cmp_eq_u32_e64(VCC, v[23], 0))
+  k.emit(s_cbranch_vccnz(), target='ShiftVectorComponents0_GLVW2_BM0_VW0')
+  k.label('ShiftVectorComponents0_GLVW3_BM0')
+  k.emit(v_cmp_eq_u32_e64(VCC, v[23], 0))
+  k.emit(s_cbranch_vccnz(), target='ShiftVectorComponents0_GLVW3_BM0_VW0')
+  k.label('ShiftVectorComponents0_GLVW4_BM0')
+  k.emit(v_cmp_eq_u32_e64(VCC, v[23], 0))
+  k.emit(s_cbranch_vccnz(), target='ShiftVectorComponents0_GLVW4_BM0_VW0')
+  k.label('ShiftVectorComponents0_GLVW5_BM0')
+  k.emit(v_cmp_eq_u32_e64(VCC, v[23], 0))
+  k.emit(s_cbranch_vccnz(), target='ShiftVectorComponents0_GLVW5_BM0_VW0')
+  k.label('ShiftVectorComponents0_GLVW6_BM0')
+  k.emit(v_cmp_eq_u32_e64(VCC, v[23], 0))
+  k.emit(s_cbranch_vccnz(), target='ShiftVectorComponents0_GLVW6_BM0_VW0')
+  k.label('ShiftVectorComponents0_GLVW7_BM0')
+  k.emit(v_cmp_eq_u32_e64(VCC, v[23], 0))
+  k.emit(s_cbranch_vccnz(), target='ShiftVectorComponents0_GLVW7_BM0_VW0')
+  k.label('ShiftVectorComponents0_GLVW1_BM0_VW0')
+  k.emit(s_mov_b32(s[8], 0))
+  k.emit(v_cmpx_eq_u32_e64(s[8:9], v[24], s[8]))
+  k.emit(v_and_b32_e32(v[18], 63, v[180]))
+  k.emit(v_lshlrev_b32_e32(v[18], 2, v[18]))
+  k.emit(v_accvgpr_read(v[25], v[28]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[0], v[25]))
+  k.emit(v_accvgpr_read(v[25], v[60]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[32], v[25]))
+  k.emit(v_accvgpr_read(v[25], v[92]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[64], v[25]))
+  k.emit(v_accvgpr_read(v[25], v[124]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[96], v[25]))
+  k.emit(v_accvgpr_read(v[25], v[156]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[128], v[25]))
+  k.emit(v_accvgpr_read(v[25], v[188]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[160], v[25]))
+  k.emit(v_accvgpr_read(v[25], v[220]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[192], v[25]))
+  k.emit(v_accvgpr_read(v[25], v[252]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[224], v[25]))
+  k.emit(v_accvgpr_read(v[25], v[29]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[1], v[25]))
+  k.emit(v_accvgpr_read(v[25], v[61]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[33], v[25]))
+  k.emit(v_accvgpr_read(v[25], v[93]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[65], v[25]))
+  k.emit(v_accvgpr_read(v[25], v[125]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[97], v[25]))
+  k.emit(v_accvgpr_read(v[25], v[157]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[129], v[25]))
+  k.emit(v_accvgpr_read(v[25], v[189]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[161], v[25]))
+  k.emit(v_accvgpr_read(v[25], v[221]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[193], v[25]))
+  k.emit(v_accvgpr_read(v[25], v[253]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[225], v[25]))
+  k.emit(v_accvgpr_read(v[25], v[30]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[2], v[25]))
+  k.emit(v_accvgpr_read(v[25], v[62]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[34], v[25]))
+  k.emit(v_accvgpr_read(v[25], v[94]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[66], v[25]))
+  k.emit(v_accvgpr_read(v[25], v[126]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[98], v[25]))
+  k.emit(v_accvgpr_read(v[25], v[158]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[130], v[25]))
+  k.emit(v_accvgpr_read(v[25], v[190]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[162], v[25]))
+  k.emit(v_accvgpr_read(v[25], v[222]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[194], v[25]))
+  k.emit(v_accvgpr_read(v[25], v[254]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[226], v[25]))
+  k.emit(v_accvgpr_read(v[25], v[31]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[3], v[25]))
+  k.emit(v_accvgpr_read(v[25], v[63]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[35], v[25]))
+  k.emit(v_accvgpr_read(v[25], v[95]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[67], v[25]))
+  k.emit(v_accvgpr_read(v[25], v[127]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[99], v[25]))
+  k.emit(v_accvgpr_read(v[25], v[159]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[131], v[25]))
+  k.emit(v_accvgpr_read(v[25], v[191]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[163], v[25]))
+  k.emit(v_accvgpr_read(v[25], v[223]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[195], v[25]))
+  k.emit(v_accvgpr_read(v[25], v[255]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[227], v[25]))
+  k.emit(s_mov_b64(s[8:9], -1))
+  k.emit(s_or_saveexec_b64(VCC, s[8:9]))
+  k.emit(s_branch(), target='ShiftVectorComponents0_GLVW0')
+  k.label('ShiftVectorComponents0_GLVW2_BM0_VW0')
+  k.emit(s_mov_b32(s[8], 0))
+  k.emit(v_cmpx_eq_u32_e64(s[8:9], v[24], s[8]))
+  k.emit(v_and_b32_e32(v[18], 63, v[180]))
+  k.emit(v_lshlrev_b32_e32(v[18], 2, v[18]))
+  k.emit(v_accvgpr_read(v[25], v[24]))
+  k.emit(v_accvgpr_read(v[26], v[28]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[0], v[25]))
+  k.emit(v_accvgpr_write(v[4], v[26]))
+  k.emit(v_accvgpr_read(v[25], v[56]))
+  k.emit(v_accvgpr_read(v[26], v[60]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[32], v[25]))
+  k.emit(v_accvgpr_write(v[36], v[26]))
+  k.emit(v_accvgpr_read(v[25], v[88]))
+  k.emit(v_accvgpr_read(v[26], v[92]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[64], v[25]))
+  k.emit(v_accvgpr_write(v[68], v[26]))
+  k.emit(v_accvgpr_read(v[25], v[120]))
+  k.emit(v_accvgpr_read(v[26], v[124]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[96], v[25]))
+  k.emit(v_accvgpr_write(v[100], v[26]))
+  k.emit(v_accvgpr_read(v[25], v[152]))
+  k.emit(v_accvgpr_read(v[26], v[156]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[128], v[25]))
+  k.emit(v_accvgpr_write(v[132], v[26]))
+  k.emit(v_accvgpr_read(v[25], v[184]))
+  k.emit(v_accvgpr_read(v[26], v[188]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[160], v[25]))
+  k.emit(v_accvgpr_write(v[164], v[26]))
+  k.emit(v_accvgpr_read(v[25], v[216]))
+  k.emit(v_accvgpr_read(v[26], v[220]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[192], v[25]))
+  k.emit(v_accvgpr_write(v[196], v[26]))
+  k.emit(v_accvgpr_read(v[25], v[248]))
+  k.emit(v_accvgpr_read(v[26], v[252]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[224], v[25]))
+  k.emit(v_accvgpr_write(v[228], v[26]))
+  k.emit(v_accvgpr_read(v[25], v[25]))
+  k.emit(v_accvgpr_read(v[26], v[29]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[1], v[25]))
+  k.emit(v_accvgpr_write(v[5], v[26]))
+  k.emit(v_accvgpr_read(v[25], v[57]))
+  k.emit(v_accvgpr_read(v[26], v[61]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[33], v[25]))
+  k.emit(v_accvgpr_write(v[37], v[26]))
+  k.emit(v_accvgpr_read(v[25], v[89]))
+  k.emit(v_accvgpr_read(v[26], v[93]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[65], v[25]))
+  k.emit(v_accvgpr_write(v[69], v[26]))
+  k.emit(v_accvgpr_read(v[25], v[121]))
+  k.emit(v_accvgpr_read(v[26], v[125]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[97], v[25]))
+  k.emit(v_accvgpr_write(v[101], v[26]))
+  k.emit(v_accvgpr_read(v[25], v[153]))
+  k.emit(v_accvgpr_read(v[26], v[157]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[129], v[25]))
+  k.emit(v_accvgpr_write(v[133], v[26]))
+  k.emit(v_accvgpr_read(v[25], v[185]))
+  k.emit(v_accvgpr_read(v[26], v[189]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[161], v[25]))
+  k.emit(v_accvgpr_write(v[165], v[26]))
+  k.emit(v_accvgpr_read(v[25], v[217]))
+  k.emit(v_accvgpr_read(v[26], v[221]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[193], v[25]))
+  k.emit(v_accvgpr_write(v[197], v[26]))
+  k.emit(v_accvgpr_read(v[25], v[249]))
+  k.emit(v_accvgpr_read(v[26], v[253]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[225], v[25]))
+  k.emit(v_accvgpr_write(v[229], v[26]))
+  k.emit(v_accvgpr_read(v[25], v[26]))
+  k.emit(v_accvgpr_read(v[26], v[30]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[2], v[25]))
+  k.emit(v_accvgpr_write(v[6], v[26]))
+  k.emit(v_accvgpr_read(v[25], v[58]))
+  k.emit(v_accvgpr_read(v[26], v[62]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[34], v[25]))
+  k.emit(v_accvgpr_write(v[38], v[26]))
+  k.emit(v_accvgpr_read(v[25], v[90]))
+  k.emit(v_accvgpr_read(v[26], v[94]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[66], v[25]))
+  k.emit(v_accvgpr_write(v[70], v[26]))
+  k.emit(v_accvgpr_read(v[25], v[122]))
+  k.emit(v_accvgpr_read(v[26], v[126]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[98], v[25]))
+  k.emit(v_accvgpr_write(v[102], v[26]))
+  k.emit(v_accvgpr_read(v[25], v[154]))
+  k.emit(v_accvgpr_read(v[26], v[158]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[130], v[25]))
+  k.emit(v_accvgpr_write(v[134], v[26]))
+  k.emit(v_accvgpr_read(v[25], v[186]))
+  k.emit(v_accvgpr_read(v[26], v[190]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[162], v[25]))
+  k.emit(v_accvgpr_write(v[166], v[26]))
+  k.emit(v_accvgpr_read(v[25], v[218]))
+  k.emit(v_accvgpr_read(v[26], v[222]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[194], v[25]))
+  k.emit(v_accvgpr_write(v[198], v[26]))
+  k.emit(v_accvgpr_read(v[25], v[250]))
+  k.emit(v_accvgpr_read(v[26], v[254]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[226], v[25]))
+  k.emit(v_accvgpr_write(v[230], v[26]))
+  k.emit(v_accvgpr_read(v[25], v[27]))
+  k.emit(v_accvgpr_read(v[26], v[31]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[3], v[25]))
+  k.emit(v_accvgpr_write(v[7], v[26]))
+  k.emit(v_accvgpr_read(v[25], v[59]))
+  k.emit(v_accvgpr_read(v[26], v[63]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[35], v[25]))
+  k.emit(v_accvgpr_write(v[39], v[26]))
+  k.emit(v_accvgpr_read(v[25], v[91]))
+  k.emit(v_accvgpr_read(v[26], v[95]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[67], v[25]))
+  k.emit(v_accvgpr_write(v[71], v[26]))
+  k.emit(v_accvgpr_read(v[25], v[123]))
+  k.emit(v_accvgpr_read(v[26], v[127]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[99], v[25]))
+  k.emit(v_accvgpr_write(v[103], v[26]))
+  k.emit(v_accvgpr_read(v[25], v[155]))
+  k.emit(v_accvgpr_read(v[26], v[159]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[131], v[25]))
+  k.emit(v_accvgpr_write(v[135], v[26]))
+  k.emit(v_accvgpr_read(v[25], v[187]))
+  k.emit(v_accvgpr_read(v[26], v[191]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[163], v[25]))
+  k.emit(v_accvgpr_write(v[167], v[26]))
+  k.emit(v_accvgpr_read(v[25], v[219]))
+  k.emit(v_accvgpr_read(v[26], v[223]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[195], v[25]))
+  k.emit(v_accvgpr_write(v[199], v[26]))
+  k.emit(v_accvgpr_read(v[25], v[251]))
+  k.emit(v_accvgpr_read(v[26], v[255]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[227], v[25]))
+  k.emit(v_accvgpr_write(v[231], v[26]))
+  k.emit(s_mov_b64(s[8:9], -1))
+  k.emit(s_or_saveexec_b64(VCC, s[8:9]))
+  k.emit(s_branch(), target='ShiftVectorComponents0_GLVW0')
+  k.label('ShiftVectorComponents0_GLVW3_BM0_VW0')
+  k.emit(s_mov_b32(s[8], 0))
+  k.emit(v_cmpx_eq_u32_e64(s[8:9], v[24], s[8]))
+  k.emit(v_and_b32_e32(v[18], 63, v[180]))
+  k.emit(v_lshlrev_b32_e32(v[18], 2, v[18]))
+  k.emit(v_accvgpr_read(v[25], v[20]))
+  k.emit(v_accvgpr_read(v[26], v[24]))
+  k.emit(v_accvgpr_read(v[27], v[28]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[0], v[25]))
+  k.emit(v_accvgpr_write(v[4], v[26]))
+  k.emit(v_accvgpr_write(v[8], v[27]))
+  k.emit(v_accvgpr_read(v[25], v[52]))
+  k.emit(v_accvgpr_read(v[26], v[56]))
+  k.emit(v_accvgpr_read(v[27], v[60]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[32], v[25]))
+  k.emit(v_accvgpr_write(v[36], v[26]))
+  k.emit(v_accvgpr_write(v[40], v[27]))
+  k.emit(v_accvgpr_read(v[25], v[84]))
+  k.emit(v_accvgpr_read(v[26], v[88]))
+  k.emit(v_accvgpr_read(v[27], v[92]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[64], v[25]))
+  k.emit(v_accvgpr_write(v[68], v[26]))
+  k.emit(v_accvgpr_write(v[72], v[27]))
+  k.emit(v_accvgpr_read(v[25], v[116]))
+  k.emit(v_accvgpr_read(v[26], v[120]))
+  k.emit(v_accvgpr_read(v[27], v[124]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[96], v[25]))
+  k.emit(v_accvgpr_write(v[100], v[26]))
+  k.emit(v_accvgpr_write(v[104], v[27]))
+  k.emit(v_accvgpr_read(v[25], v[148]))
+  k.emit(v_accvgpr_read(v[26], v[152]))
+  k.emit(v_accvgpr_read(v[27], v[156]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[128], v[25]))
+  k.emit(v_accvgpr_write(v[132], v[26]))
+  k.emit(v_accvgpr_write(v[136], v[27]))
+  k.emit(v_accvgpr_read(v[25], v[180]))
+  k.emit(v_accvgpr_read(v[26], v[184]))
+  k.emit(v_accvgpr_read(v[27], v[188]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[160], v[25]))
+  k.emit(v_accvgpr_write(v[164], v[26]))
+  k.emit(v_accvgpr_write(v[168], v[27]))
+  k.emit(v_accvgpr_read(v[25], v[212]))
+  k.emit(v_accvgpr_read(v[26], v[216]))
+  k.emit(v_accvgpr_read(v[27], v[220]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[192], v[25]))
+  k.emit(v_accvgpr_write(v[196], v[26]))
+  k.emit(v_accvgpr_write(v[200], v[27]))
+  k.emit(v_accvgpr_read(v[25], v[244]))
+  k.emit(v_accvgpr_read(v[26], v[248]))
+  k.emit(v_accvgpr_read(v[27], v[252]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[224], v[25]))
+  k.emit(v_accvgpr_write(v[228], v[26]))
+  k.emit(v_accvgpr_write(v[232], v[27]))
+  k.emit(v_accvgpr_read(v[25], v[21]))
+  k.emit(v_accvgpr_read(v[26], v[25]))
+  k.emit(v_accvgpr_read(v[27], v[29]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[1], v[25]))
+  k.emit(v_accvgpr_write(v[5], v[26]))
+  k.emit(v_accvgpr_write(v[9], v[27]))
+  k.emit(v_accvgpr_read(v[25], v[53]))
+  k.emit(v_accvgpr_read(v[26], v[57]))
+  k.emit(v_accvgpr_read(v[27], v[61]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[33], v[25]))
+  k.emit(v_accvgpr_write(v[37], v[26]))
+  k.emit(v_accvgpr_write(v[41], v[27]))
+  k.emit(v_accvgpr_read(v[25], v[85]))
+  k.emit(v_accvgpr_read(v[26], v[89]))
+  k.emit(v_accvgpr_read(v[27], v[93]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[65], v[25]))
+  k.emit(v_accvgpr_write(v[69], v[26]))
+  k.emit(v_accvgpr_write(v[73], v[27]))
+  k.emit(v_accvgpr_read(v[25], v[117]))
+  k.emit(v_accvgpr_read(v[26], v[121]))
+  k.emit(v_accvgpr_read(v[27], v[125]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[97], v[25]))
+  k.emit(v_accvgpr_write(v[101], v[26]))
+  k.emit(v_accvgpr_write(v[105], v[27]))
+  k.emit(v_accvgpr_read(v[25], v[149]))
+  k.emit(v_accvgpr_read(v[26], v[153]))
+  k.emit(v_accvgpr_read(v[27], v[157]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[129], v[25]))
+  k.emit(v_accvgpr_write(v[133], v[26]))
+  k.emit(v_accvgpr_write(v[137], v[27]))
+  k.emit(v_accvgpr_read(v[25], v[181]))
+  k.emit(v_accvgpr_read(v[26], v[185]))
+  k.emit(v_accvgpr_read(v[27], v[189]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[161], v[25]))
+  k.emit(v_accvgpr_write(v[165], v[26]))
+  k.emit(v_accvgpr_write(v[169], v[27]))
+  k.emit(v_accvgpr_read(v[25], v[213]))
+  k.emit(v_accvgpr_read(v[26], v[217]))
+  k.emit(v_accvgpr_read(v[27], v[221]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[193], v[25]))
+  k.emit(v_accvgpr_write(v[197], v[26]))
+  k.emit(v_accvgpr_write(v[201], v[27]))
+  k.emit(v_accvgpr_read(v[25], v[245]))
+  k.emit(v_accvgpr_read(v[26], v[249]))
+  k.emit(v_accvgpr_read(v[27], v[253]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[225], v[25]))
+  k.emit(v_accvgpr_write(v[229], v[26]))
+  k.emit(v_accvgpr_write(v[233], v[27]))
+  k.emit(v_accvgpr_read(v[25], v[22]))
+  k.emit(v_accvgpr_read(v[26], v[26]))
+  k.emit(v_accvgpr_read(v[27], v[30]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[2], v[25]))
+  k.emit(v_accvgpr_write(v[6], v[26]))
+  k.emit(v_accvgpr_write(v[10], v[27]))
+  k.emit(v_accvgpr_read(v[25], v[54]))
+  k.emit(v_accvgpr_read(v[26], v[58]))
+  k.emit(v_accvgpr_read(v[27], v[62]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[34], v[25]))
+  k.emit(v_accvgpr_write(v[38], v[26]))
+  k.emit(v_accvgpr_write(v[42], v[27]))
+  k.emit(v_accvgpr_read(v[25], v[86]))
+  k.emit(v_accvgpr_read(v[26], v[90]))
+  k.emit(v_accvgpr_read(v[27], v[94]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[66], v[25]))
+  k.emit(v_accvgpr_write(v[70], v[26]))
+  k.emit(v_accvgpr_write(v[74], v[27]))
+  k.emit(v_accvgpr_read(v[25], v[118]))
+  k.emit(v_accvgpr_read(v[26], v[122]))
+  k.emit(v_accvgpr_read(v[27], v[126]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[98], v[25]))
+  k.emit(v_accvgpr_write(v[102], v[26]))
+  k.emit(v_accvgpr_write(v[106], v[27]))
+  k.emit(v_accvgpr_read(v[25], v[150]))
+  k.emit(v_accvgpr_read(v[26], v[154]))
+  k.emit(v_accvgpr_read(v[27], v[158]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[130], v[25]))
+  k.emit(v_accvgpr_write(v[134], v[26]))
+  k.emit(v_accvgpr_write(v[138], v[27]))
+  k.emit(v_accvgpr_read(v[25], v[182]))
+  k.emit(v_accvgpr_read(v[26], v[186]))
+  k.emit(v_accvgpr_read(v[27], v[190]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[162], v[25]))
+  k.emit(v_accvgpr_write(v[166], v[26]))
+  k.emit(v_accvgpr_write(v[170], v[27]))
+  k.emit(v_accvgpr_read(v[25], v[214]))
+  k.emit(v_accvgpr_read(v[26], v[218]))
+  k.emit(v_accvgpr_read(v[27], v[222]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[194], v[25]))
+  k.emit(v_accvgpr_write(v[198], v[26]))
+  k.emit(v_accvgpr_write(v[202], v[27]))
+  k.emit(v_accvgpr_read(v[25], v[246]))
+  k.emit(v_accvgpr_read(v[26], v[250]))
+  k.emit(v_accvgpr_read(v[27], v[254]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[226], v[25]))
+  k.emit(v_accvgpr_write(v[230], v[26]))
+  k.emit(v_accvgpr_write(v[234], v[27]))
+  k.emit(v_accvgpr_read(v[25], v[23]))
+  k.emit(v_accvgpr_read(v[26], v[27]))
+  k.emit(v_accvgpr_read(v[27], v[31]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[3], v[25]))
+  k.emit(v_accvgpr_write(v[7], v[26]))
+  k.emit(v_accvgpr_write(v[11], v[27]))
+  k.emit(v_accvgpr_read(v[25], v[55]))
+  k.emit(v_accvgpr_read(v[26], v[59]))
+  k.emit(v_accvgpr_read(v[27], v[63]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[35], v[25]))
+  k.emit(v_accvgpr_write(v[39], v[26]))
+  k.emit(v_accvgpr_write(v[43], v[27]))
+  k.emit(v_accvgpr_read(v[25], v[87]))
+  k.emit(v_accvgpr_read(v[26], v[91]))
+  k.emit(v_accvgpr_read(v[27], v[95]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[67], v[25]))
+  k.emit(v_accvgpr_write(v[71], v[26]))
+  k.emit(v_accvgpr_write(v[75], v[27]))
+  k.emit(v_accvgpr_read(v[25], v[119]))
+  k.emit(v_accvgpr_read(v[26], v[123]))
+  k.emit(v_accvgpr_read(v[27], v[127]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[99], v[25]))
+  k.emit(v_accvgpr_write(v[103], v[26]))
+  k.emit(v_accvgpr_write(v[107], v[27]))
+  k.emit(v_accvgpr_read(v[25], v[151]))
+  k.emit(v_accvgpr_read(v[26], v[155]))
+  k.emit(v_accvgpr_read(v[27], v[159]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[131], v[25]))
+  k.emit(v_accvgpr_write(v[135], v[26]))
+  k.emit(v_accvgpr_write(v[139], v[27]))
+  k.emit(v_accvgpr_read(v[25], v[183]))
+  k.emit(v_accvgpr_read(v[26], v[187]))
+  k.emit(v_accvgpr_read(v[27], v[191]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[163], v[25]))
+  k.emit(v_accvgpr_write(v[167], v[26]))
+  k.emit(v_accvgpr_write(v[171], v[27]))
+  k.emit(v_accvgpr_read(v[25], v[215]))
+  k.emit(v_accvgpr_read(v[26], v[219]))
+  k.emit(v_accvgpr_read(v[27], v[223]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[195], v[25]))
+  k.emit(v_accvgpr_write(v[199], v[26]))
+  k.emit(v_accvgpr_write(v[203], v[27]))
+  k.emit(v_accvgpr_read(v[25], v[247]))
+  k.emit(v_accvgpr_read(v[26], v[251]))
+  k.emit(v_accvgpr_read(v[27], v[255]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[227], v[25]))
+  k.emit(v_accvgpr_write(v[231], v[26]))
+  k.emit(v_accvgpr_write(v[235], v[27]))
+  k.emit(s_mov_b64(s[8:9], -1))
+  k.emit(s_or_saveexec_b64(VCC, s[8:9]))
+  k.emit(s_branch(), target='ShiftVectorComponents0_GLVW0')
+  k.label('ShiftVectorComponents0_GLVW4_BM0_VW0')
+  k.emit(s_mov_b32(s[8], 0))
+  k.emit(v_cmpx_eq_u32_e64(s[8:9], v[24], s[8]))
+  k.emit(v_and_b32_e32(v[18], 63, v[180]))
+  k.emit(v_lshlrev_b32_e32(v[18], 2, v[18]))
+  k.emit(v_accvgpr_read(v[25], v[16]))
+  k.emit(v_accvgpr_read(v[26], v[20]))
+  k.emit(v_accvgpr_read(v[27], v[24]))
+  k.emit(v_accvgpr_read(v[28], v[28]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[0], v[25]))
+  k.emit(v_accvgpr_write(v[4], v[26]))
+  k.emit(v_accvgpr_write(v[8], v[27]))
+  k.emit(v_accvgpr_write(v[12], v[28]))
+  k.emit(v_accvgpr_read(v[25], v[48]))
+  k.emit(v_accvgpr_read(v[26], v[52]))
+  k.emit(v_accvgpr_read(v[27], v[56]))
+  k.emit(v_accvgpr_read(v[28], v[60]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[32], v[25]))
+  k.emit(v_accvgpr_write(v[36], v[26]))
+  k.emit(v_accvgpr_write(v[40], v[27]))
+  k.emit(v_accvgpr_write(v[44], v[28]))
+  k.emit(v_accvgpr_read(v[25], v[80]))
+  k.emit(v_accvgpr_read(v[26], v[84]))
+  k.emit(v_accvgpr_read(v[27], v[88]))
+  k.emit(v_accvgpr_read(v[28], v[92]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[64], v[25]))
+  k.emit(v_accvgpr_write(v[68], v[26]))
+  k.emit(v_accvgpr_write(v[72], v[27]))
+  k.emit(v_accvgpr_write(v[76], v[28]))
+  k.emit(v_accvgpr_read(v[25], v[112]))
+  k.emit(v_accvgpr_read(v[26], v[116]))
+  k.emit(v_accvgpr_read(v[27], v[120]))
+  k.emit(v_accvgpr_read(v[28], v[124]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[96], v[25]))
+  k.emit(v_accvgpr_write(v[100], v[26]))
+  k.emit(v_accvgpr_write(v[104], v[27]))
+  k.emit(v_accvgpr_write(v[108], v[28]))
+  k.emit(v_accvgpr_read(v[25], v[144]))
+  k.emit(v_accvgpr_read(v[26], v[148]))
+  k.emit(v_accvgpr_read(v[27], v[152]))
+  k.emit(v_accvgpr_read(v[28], v[156]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[128], v[25]))
+  k.emit(v_accvgpr_write(v[132], v[26]))
+  k.emit(v_accvgpr_write(v[136], v[27]))
+  k.emit(v_accvgpr_write(v[140], v[28]))
+  k.emit(v_accvgpr_read(v[25], v[176]))
+  k.emit(v_accvgpr_read(v[26], v[180]))
+  k.emit(v_accvgpr_read(v[27], v[184]))
+  k.emit(v_accvgpr_read(v[28], v[188]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[160], v[25]))
+  k.emit(v_accvgpr_write(v[164], v[26]))
+  k.emit(v_accvgpr_write(v[168], v[27]))
+  k.emit(v_accvgpr_write(v[172], v[28]))
+  k.emit(v_accvgpr_read(v[25], v[208]))
+  k.emit(v_accvgpr_read(v[26], v[212]))
+  k.emit(v_accvgpr_read(v[27], v[216]))
+  k.emit(v_accvgpr_read(v[28], v[220]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[192], v[25]))
+  k.emit(v_accvgpr_write(v[196], v[26]))
+  k.emit(v_accvgpr_write(v[200], v[27]))
+  k.emit(v_accvgpr_write(v[204], v[28]))
+  k.emit(v_accvgpr_read(v[25], v[240]))
+  k.emit(v_accvgpr_read(v[26], v[244]))
+  k.emit(v_accvgpr_read(v[27], v[248]))
+  k.emit(v_accvgpr_read(v[28], v[252]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[224], v[25]))
+  k.emit(v_accvgpr_write(v[228], v[26]))
+  k.emit(v_accvgpr_write(v[232], v[27]))
+  k.emit(v_accvgpr_write(v[236], v[28]))
+  k.emit(v_accvgpr_read(v[25], v[17]))
+  k.emit(v_accvgpr_read(v[26], v[21]))
+  k.emit(v_accvgpr_read(v[27], v[25]))
+  k.emit(v_accvgpr_read(v[28], v[29]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[1], v[25]))
+  k.emit(v_accvgpr_write(v[5], v[26]))
+  k.emit(v_accvgpr_write(v[9], v[27]))
+  k.emit(v_accvgpr_write(v[13], v[28]))
+  k.emit(v_accvgpr_read(v[25], v[49]))
+  k.emit(v_accvgpr_read(v[26], v[53]))
+  k.emit(v_accvgpr_read(v[27], v[57]))
+  k.emit(v_accvgpr_read(v[28], v[61]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[33], v[25]))
+  k.emit(v_accvgpr_write(v[37], v[26]))
+  k.emit(v_accvgpr_write(v[41], v[27]))
+  k.emit(v_accvgpr_write(v[45], v[28]))
+  k.emit(v_accvgpr_read(v[25], v[81]))
+  k.emit(v_accvgpr_read(v[26], v[85]))
+  k.emit(v_accvgpr_read(v[27], v[89]))
+  k.emit(v_accvgpr_read(v[28], v[93]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[65], v[25]))
+  k.emit(v_accvgpr_write(v[69], v[26]))
+  k.emit(v_accvgpr_write(v[73], v[27]))
+  k.emit(v_accvgpr_write(v[77], v[28]))
+  k.emit(v_accvgpr_read(v[25], v[113]))
+  k.emit(v_accvgpr_read(v[26], v[117]))
+  k.emit(v_accvgpr_read(v[27], v[121]))
+  k.emit(v_accvgpr_read(v[28], v[125]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[97], v[25]))
+  k.emit(v_accvgpr_write(v[101], v[26]))
+  k.emit(v_accvgpr_write(v[105], v[27]))
+  k.emit(v_accvgpr_write(v[109], v[28]))
+  k.emit(v_accvgpr_read(v[25], v[145]))
+  k.emit(v_accvgpr_read(v[26], v[149]))
+  k.emit(v_accvgpr_read(v[27], v[153]))
+  k.emit(v_accvgpr_read(v[28], v[157]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[129], v[25]))
+  k.emit(v_accvgpr_write(v[133], v[26]))
+  k.emit(v_accvgpr_write(v[137], v[27]))
+  k.emit(v_accvgpr_write(v[141], v[28]))
+  k.emit(v_accvgpr_read(v[25], v[177]))
+  k.emit(v_accvgpr_read(v[26], v[181]))
+  k.emit(v_accvgpr_read(v[27], v[185]))
+  k.emit(v_accvgpr_read(v[28], v[189]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[161], v[25]))
+  k.emit(v_accvgpr_write(v[165], v[26]))
+  k.emit(v_accvgpr_write(v[169], v[27]))
+  k.emit(v_accvgpr_write(v[173], v[28]))
+  k.emit(v_accvgpr_read(v[25], v[209]))
+  k.emit(v_accvgpr_read(v[26], v[213]))
+  k.emit(v_accvgpr_read(v[27], v[217]))
+  k.emit(v_accvgpr_read(v[28], v[221]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[193], v[25]))
+  k.emit(v_accvgpr_write(v[197], v[26]))
+  k.emit(v_accvgpr_write(v[201], v[27]))
+  k.emit(v_accvgpr_write(v[205], v[28]))
+  k.emit(v_accvgpr_read(v[25], v[241]))
+  k.emit(v_accvgpr_read(v[26], v[245]))
+  k.emit(v_accvgpr_read(v[27], v[249]))
+  k.emit(v_accvgpr_read(v[28], v[253]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[225], v[25]))
+  k.emit(v_accvgpr_write(v[229], v[26]))
+  k.emit(v_accvgpr_write(v[233], v[27]))
+  k.emit(v_accvgpr_write(v[237], v[28]))
+  k.emit(v_accvgpr_read(v[25], v[18]))
+  k.emit(v_accvgpr_read(v[26], v[22]))
+  k.emit(v_accvgpr_read(v[27], v[26]))
+  k.emit(v_accvgpr_read(v[28], v[30]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[2], v[25]))
+  k.emit(v_accvgpr_write(v[6], v[26]))
+  k.emit(v_accvgpr_write(v[10], v[27]))
+  k.emit(v_accvgpr_write(v[14], v[28]))
+  k.emit(v_accvgpr_read(v[25], v[50]))
+  k.emit(v_accvgpr_read(v[26], v[54]))
+  k.emit(v_accvgpr_read(v[27], v[58]))
+  k.emit(v_accvgpr_read(v[28], v[62]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[34], v[25]))
+  k.emit(v_accvgpr_write(v[38], v[26]))
+  k.emit(v_accvgpr_write(v[42], v[27]))
+  k.emit(v_accvgpr_write(v[46], v[28]))
+  k.emit(v_accvgpr_read(v[25], v[82]))
+  k.emit(v_accvgpr_read(v[26], v[86]))
+  k.emit(v_accvgpr_read(v[27], v[90]))
+  k.emit(v_accvgpr_read(v[28], v[94]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[66], v[25]))
+  k.emit(v_accvgpr_write(v[70], v[26]))
+  k.emit(v_accvgpr_write(v[74], v[27]))
+  k.emit(v_accvgpr_write(v[78], v[28]))
+  k.emit(v_accvgpr_read(v[25], v[114]))
+  k.emit(v_accvgpr_read(v[26], v[118]))
+  k.emit(v_accvgpr_read(v[27], v[122]))
+  k.emit(v_accvgpr_read(v[28], v[126]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[98], v[25]))
+  k.emit(v_accvgpr_write(v[102], v[26]))
+  k.emit(v_accvgpr_write(v[106], v[27]))
+  k.emit(v_accvgpr_write(v[110], v[28]))
+  k.emit(v_accvgpr_read(v[25], v[146]))
+  k.emit(v_accvgpr_read(v[26], v[150]))
+  k.emit(v_accvgpr_read(v[27], v[154]))
+  k.emit(v_accvgpr_read(v[28], v[158]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[130], v[25]))
+  k.emit(v_accvgpr_write(v[134], v[26]))
+  k.emit(v_accvgpr_write(v[138], v[27]))
+  k.emit(v_accvgpr_write(v[142], v[28]))
+  k.emit(v_accvgpr_read(v[25], v[178]))
+  k.emit(v_accvgpr_read(v[26], v[182]))
+  k.emit(v_accvgpr_read(v[27], v[186]))
+  k.emit(v_accvgpr_read(v[28], v[190]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[162], v[25]))
+  k.emit(v_accvgpr_write(v[166], v[26]))
+  k.emit(v_accvgpr_write(v[170], v[27]))
+  k.emit(v_accvgpr_write(v[174], v[28]))
+  k.emit(v_accvgpr_read(v[25], v[210]))
+  k.emit(v_accvgpr_read(v[26], v[214]))
+  k.emit(v_accvgpr_read(v[27], v[218]))
+  k.emit(v_accvgpr_read(v[28], v[222]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[194], v[25]))
+  k.emit(v_accvgpr_write(v[198], v[26]))
+  k.emit(v_accvgpr_write(v[202], v[27]))
+  k.emit(v_accvgpr_write(v[206], v[28]))
+  k.emit(v_accvgpr_read(v[25], v[242]))
+  k.emit(v_accvgpr_read(v[26], v[246]))
+  k.emit(v_accvgpr_read(v[27], v[250]))
+  k.emit(v_accvgpr_read(v[28], v[254]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[226], v[25]))
+  k.emit(v_accvgpr_write(v[230], v[26]))
+  k.emit(v_accvgpr_write(v[234], v[27]))
+  k.emit(v_accvgpr_write(v[238], v[28]))
+  k.emit(v_accvgpr_read(v[25], v[19]))
+  k.emit(v_accvgpr_read(v[26], v[23]))
+  k.emit(v_accvgpr_read(v[27], v[27]))
+  k.emit(v_accvgpr_read(v[28], v[31]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[3], v[25]))
+  k.emit(v_accvgpr_write(v[7], v[26]))
+  k.emit(v_accvgpr_write(v[11], v[27]))
+  k.emit(v_accvgpr_write(v[15], v[28]))
+  k.emit(v_accvgpr_read(v[25], v[51]))
+  k.emit(v_accvgpr_read(v[26], v[55]))
+  k.emit(v_accvgpr_read(v[27], v[59]))
+  k.emit(v_accvgpr_read(v[28], v[63]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[35], v[25]))
+  k.emit(v_accvgpr_write(v[39], v[26]))
+  k.emit(v_accvgpr_write(v[43], v[27]))
+  k.emit(v_accvgpr_write(v[47], v[28]))
+  k.emit(v_accvgpr_read(v[25], v[83]))
+  k.emit(v_accvgpr_read(v[26], v[87]))
+  k.emit(v_accvgpr_read(v[27], v[91]))
+  k.emit(v_accvgpr_read(v[28], v[95]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[67], v[25]))
+  k.emit(v_accvgpr_write(v[71], v[26]))
+  k.emit(v_accvgpr_write(v[75], v[27]))
+  k.emit(v_accvgpr_write(v[79], v[28]))
+  k.emit(v_accvgpr_read(v[25], v[115]))
+  k.emit(v_accvgpr_read(v[26], v[119]))
+  k.emit(v_accvgpr_read(v[27], v[123]))
+  k.emit(v_accvgpr_read(v[28], v[127]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[99], v[25]))
+  k.emit(v_accvgpr_write(v[103], v[26]))
+  k.emit(v_accvgpr_write(v[107], v[27]))
+  k.emit(v_accvgpr_write(v[111], v[28]))
+  k.emit(v_accvgpr_read(v[25], v[147]))
+  k.emit(v_accvgpr_read(v[26], v[151]))
+  k.emit(v_accvgpr_read(v[27], v[155]))
+  k.emit(v_accvgpr_read(v[28], v[159]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[131], v[25]))
+  k.emit(v_accvgpr_write(v[135], v[26]))
+  k.emit(v_accvgpr_write(v[139], v[27]))
+  k.emit(v_accvgpr_write(v[143], v[28]))
+  k.emit(v_accvgpr_read(v[25], v[179]))
+  k.emit(v_accvgpr_read(v[26], v[183]))
+  k.emit(v_accvgpr_read(v[27], v[187]))
+  k.emit(v_accvgpr_read(v[28], v[191]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[163], v[25]))
+  k.emit(v_accvgpr_write(v[167], v[26]))
+  k.emit(v_accvgpr_write(v[171], v[27]))
+  k.emit(v_accvgpr_write(v[175], v[28]))
+  k.emit(v_accvgpr_read(v[25], v[211]))
+  k.emit(v_accvgpr_read(v[26], v[215]))
+  k.emit(v_accvgpr_read(v[27], v[219]))
+  k.emit(v_accvgpr_read(v[28], v[223]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[195], v[25]))
+  k.emit(v_accvgpr_write(v[199], v[26]))
+  k.emit(v_accvgpr_write(v[203], v[27]))
+  k.emit(v_accvgpr_write(v[207], v[28]))
+  k.emit(v_accvgpr_read(v[25], v[243]))
+  k.emit(v_accvgpr_read(v[26], v[247]))
+  k.emit(v_accvgpr_read(v[27], v[251]))
+  k.emit(v_accvgpr_read(v[28], v[255]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[227], v[25]))
+  k.emit(v_accvgpr_write(v[231], v[26]))
+  k.emit(v_accvgpr_write(v[235], v[27]))
+  k.emit(v_accvgpr_write(v[239], v[28]))
+  k.emit(s_mov_b64(s[8:9], -1))
+  k.emit(s_or_saveexec_b64(VCC, s[8:9]))
+  k.emit(s_branch(), target='ShiftVectorComponents0_GLVW0')
+  k.label('ShiftVectorComponents0_GLVW5_BM0_VW0')
+  k.emit(s_mov_b32(s[8], 0))
+  k.emit(v_cmpx_eq_u32_e64(s[8:9], v[24], s[8]))
+  k.emit(v_and_b32_e32(v[18], 63, v[180]))
+  k.emit(v_lshlrev_b32_e32(v[18], 2, v[18]))
+  k.emit(v_accvgpr_read(v[25], v[12]))
+  k.emit(v_accvgpr_read(v[26], v[16]))
+  k.emit(v_accvgpr_read(v[27], v[20]))
+  k.emit(v_accvgpr_read(v[28], v[24]))
+  k.emit(v_accvgpr_read(v[29], v[28]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[0], v[25]))
+  k.emit(v_accvgpr_write(v[4], v[26]))
+  k.emit(v_accvgpr_write(v[8], v[27]))
+  k.emit(v_accvgpr_write(v[12], v[28]))
+  k.emit(v_accvgpr_write(v[16], v[29]))
+  k.emit(v_accvgpr_read(v[25], v[44]))
+  k.emit(v_accvgpr_read(v[26], v[48]))
+  k.emit(v_accvgpr_read(v[27], v[52]))
+  k.emit(v_accvgpr_read(v[28], v[56]))
+  k.emit(v_accvgpr_read(v[29], v[60]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[32], v[25]))
+  k.emit(v_accvgpr_write(v[36], v[26]))
+  k.emit(v_accvgpr_write(v[40], v[27]))
+  k.emit(v_accvgpr_write(v[44], v[28]))
+  k.emit(v_accvgpr_write(v[48], v[29]))
+  k.emit(v_accvgpr_read(v[25], v[76]))
+  k.emit(v_accvgpr_read(v[26], v[80]))
+  k.emit(v_accvgpr_read(v[27], v[84]))
+  k.emit(v_accvgpr_read(v[28], v[88]))
+  k.emit(v_accvgpr_read(v[29], v[92]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[64], v[25]))
+  k.emit(v_accvgpr_write(v[68], v[26]))
+  k.emit(v_accvgpr_write(v[72], v[27]))
+  k.emit(v_accvgpr_write(v[76], v[28]))
+  k.emit(v_accvgpr_write(v[80], v[29]))
+  k.emit(v_accvgpr_read(v[25], v[108]))
+  k.emit(v_accvgpr_read(v[26], v[112]))
+  k.emit(v_accvgpr_read(v[27], v[116]))
+  k.emit(v_accvgpr_read(v[28], v[120]))
+  k.emit(v_accvgpr_read(v[29], v[124]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[96], v[25]))
+  k.emit(v_accvgpr_write(v[100], v[26]))
+  k.emit(v_accvgpr_write(v[104], v[27]))
+  k.emit(v_accvgpr_write(v[108], v[28]))
+  k.emit(v_accvgpr_write(v[112], v[29]))
+  k.emit(v_accvgpr_read(v[25], v[140]))
+  k.emit(v_accvgpr_read(v[26], v[144]))
+  k.emit(v_accvgpr_read(v[27], v[148]))
+  k.emit(v_accvgpr_read(v[28], v[152]))
+  k.emit(v_accvgpr_read(v[29], v[156]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[128], v[25]))
+  k.emit(v_accvgpr_write(v[132], v[26]))
+  k.emit(v_accvgpr_write(v[136], v[27]))
+  k.emit(v_accvgpr_write(v[140], v[28]))
+  k.emit(v_accvgpr_write(v[144], v[29]))
+  k.emit(v_accvgpr_read(v[25], v[172]))
+  k.emit(v_accvgpr_read(v[26], v[176]))
+  k.emit(v_accvgpr_read(v[27], v[180]))
+  k.emit(v_accvgpr_read(v[28], v[184]))
+  k.emit(v_accvgpr_read(v[29], v[188]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[160], v[25]))
+  k.emit(v_accvgpr_write(v[164], v[26]))
+  k.emit(v_accvgpr_write(v[168], v[27]))
+  k.emit(v_accvgpr_write(v[172], v[28]))
+  k.emit(v_accvgpr_write(v[176], v[29]))
+  k.emit(v_accvgpr_read(v[25], v[204]))
+  k.emit(v_accvgpr_read(v[26], v[208]))
+  k.emit(v_accvgpr_read(v[27], v[212]))
+  k.emit(v_accvgpr_read(v[28], v[216]))
+  k.emit(v_accvgpr_read(v[29], v[220]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[192], v[25]))
+  k.emit(v_accvgpr_write(v[196], v[26]))
+  k.emit(v_accvgpr_write(v[200], v[27]))
+  k.emit(v_accvgpr_write(v[204], v[28]))
+  k.emit(v_accvgpr_write(v[208], v[29]))
+  k.emit(v_accvgpr_read(v[25], v[236]))
+  k.emit(v_accvgpr_read(v[26], v[240]))
+  k.emit(v_accvgpr_read(v[27], v[244]))
+  k.emit(v_accvgpr_read(v[28], v[248]))
+  k.emit(v_accvgpr_read(v[29], v[252]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[224], v[25]))
+  k.emit(v_accvgpr_write(v[228], v[26]))
+  k.emit(v_accvgpr_write(v[232], v[27]))
+  k.emit(v_accvgpr_write(v[236], v[28]))
+  k.emit(v_accvgpr_write(v[240], v[29]))
+  k.emit(v_accvgpr_read(v[25], v[13]))
+  k.emit(v_accvgpr_read(v[26], v[17]))
+  k.emit(v_accvgpr_read(v[27], v[21]))
+  k.emit(v_accvgpr_read(v[28], v[25]))
+  k.emit(v_accvgpr_read(v[29], v[29]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[1], v[25]))
+  k.emit(v_accvgpr_write(v[5], v[26]))
+  k.emit(v_accvgpr_write(v[9], v[27]))
+  k.emit(v_accvgpr_write(v[13], v[28]))
+  k.emit(v_accvgpr_write(v[17], v[29]))
+  k.emit(v_accvgpr_read(v[25], v[45]))
+  k.emit(v_accvgpr_read(v[26], v[49]))
+  k.emit(v_accvgpr_read(v[27], v[53]))
+  k.emit(v_accvgpr_read(v[28], v[57]))
+  k.emit(v_accvgpr_read(v[29], v[61]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[33], v[25]))
+  k.emit(v_accvgpr_write(v[37], v[26]))
+  k.emit(v_accvgpr_write(v[41], v[27]))
+  k.emit(v_accvgpr_write(v[45], v[28]))
+  k.emit(v_accvgpr_write(v[49], v[29]))
+  k.emit(v_accvgpr_read(v[25], v[77]))
+  k.emit(v_accvgpr_read(v[26], v[81]))
+  k.emit(v_accvgpr_read(v[27], v[85]))
+  k.emit(v_accvgpr_read(v[28], v[89]))
+  k.emit(v_accvgpr_read(v[29], v[93]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[65], v[25]))
+  k.emit(v_accvgpr_write(v[69], v[26]))
+  k.emit(v_accvgpr_write(v[73], v[27]))
+  k.emit(v_accvgpr_write(v[77], v[28]))
+  k.emit(v_accvgpr_write(v[81], v[29]))
+  k.emit(v_accvgpr_read(v[25], v[109]))
+  k.emit(v_accvgpr_read(v[26], v[113]))
+  k.emit(v_accvgpr_read(v[27], v[117]))
+  k.emit(v_accvgpr_read(v[28], v[121]))
+  k.emit(v_accvgpr_read(v[29], v[125]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[97], v[25]))
+  k.emit(v_accvgpr_write(v[101], v[26]))
+  k.emit(v_accvgpr_write(v[105], v[27]))
+  k.emit(v_accvgpr_write(v[109], v[28]))
+  k.emit(v_accvgpr_write(v[113], v[29]))
+  k.emit(v_accvgpr_read(v[25], v[141]))
+  k.emit(v_accvgpr_read(v[26], v[145]))
+  k.emit(v_accvgpr_read(v[27], v[149]))
+  k.emit(v_accvgpr_read(v[28], v[153]))
+  k.emit(v_accvgpr_read(v[29], v[157]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[129], v[25]))
+  k.emit(v_accvgpr_write(v[133], v[26]))
+  k.emit(v_accvgpr_write(v[137], v[27]))
+  k.emit(v_accvgpr_write(v[141], v[28]))
+  k.emit(v_accvgpr_write(v[145], v[29]))
+  k.emit(v_accvgpr_read(v[25], v[173]))
+  k.emit(v_accvgpr_read(v[26], v[177]))
+  k.emit(v_accvgpr_read(v[27], v[181]))
+  k.emit(v_accvgpr_read(v[28], v[185]))
+  k.emit(v_accvgpr_read(v[29], v[189]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[161], v[25]))
+  k.emit(v_accvgpr_write(v[165], v[26]))
+  k.emit(v_accvgpr_write(v[169], v[27]))
+  k.emit(v_accvgpr_write(v[173], v[28]))
+  k.emit(v_accvgpr_write(v[177], v[29]))
+  k.emit(v_accvgpr_read(v[25], v[205]))
+  k.emit(v_accvgpr_read(v[26], v[209]))
+  k.emit(v_accvgpr_read(v[27], v[213]))
+  k.emit(v_accvgpr_read(v[28], v[217]))
+  k.emit(v_accvgpr_read(v[29], v[221]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[193], v[25]))
+  k.emit(v_accvgpr_write(v[197], v[26]))
+  k.emit(v_accvgpr_write(v[201], v[27]))
+  k.emit(v_accvgpr_write(v[205], v[28]))
+  k.emit(v_accvgpr_write(v[209], v[29]))
+  k.emit(v_accvgpr_read(v[25], v[237]))
+  k.emit(v_accvgpr_read(v[26], v[241]))
+  k.emit(v_accvgpr_read(v[27], v[245]))
+  k.emit(v_accvgpr_read(v[28], v[249]))
+  k.emit(v_accvgpr_read(v[29], v[253]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[225], v[25]))
+  k.emit(v_accvgpr_write(v[229], v[26]))
+  k.emit(v_accvgpr_write(v[233], v[27]))
+  k.emit(v_accvgpr_write(v[237], v[28]))
+  k.emit(v_accvgpr_write(v[241], v[29]))
+  k.emit(v_accvgpr_read(v[25], v[14]))
+  k.emit(v_accvgpr_read(v[26], v[18]))
+  k.emit(v_accvgpr_read(v[27], v[22]))
+  k.emit(v_accvgpr_read(v[28], v[26]))
+  k.emit(v_accvgpr_read(v[29], v[30]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[2], v[25]))
+  k.emit(v_accvgpr_write(v[6], v[26]))
+  k.emit(v_accvgpr_write(v[10], v[27]))
+  k.emit(v_accvgpr_write(v[14], v[28]))
+  k.emit(v_accvgpr_write(v[18], v[29]))
+  k.emit(v_accvgpr_read(v[25], v[46]))
+  k.emit(v_accvgpr_read(v[26], v[50]))
+  k.emit(v_accvgpr_read(v[27], v[54]))
+  k.emit(v_accvgpr_read(v[28], v[58]))
+  k.emit(v_accvgpr_read(v[29], v[62]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[34], v[25]))
+  k.emit(v_accvgpr_write(v[38], v[26]))
+  k.emit(v_accvgpr_write(v[42], v[27]))
+  k.emit(v_accvgpr_write(v[46], v[28]))
+  k.emit(v_accvgpr_write(v[50], v[29]))
+  k.emit(v_accvgpr_read(v[25], v[78]))
+  k.emit(v_accvgpr_read(v[26], v[82]))
+  k.emit(v_accvgpr_read(v[27], v[86]))
+  k.emit(v_accvgpr_read(v[28], v[90]))
+  k.emit(v_accvgpr_read(v[29], v[94]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[66], v[25]))
+  k.emit(v_accvgpr_write(v[70], v[26]))
+  k.emit(v_accvgpr_write(v[74], v[27]))
+  k.emit(v_accvgpr_write(v[78], v[28]))
+  k.emit(v_accvgpr_write(v[82], v[29]))
+  k.emit(v_accvgpr_read(v[25], v[110]))
+  k.emit(v_accvgpr_read(v[26], v[114]))
+  k.emit(v_accvgpr_read(v[27], v[118]))
+  k.emit(v_accvgpr_read(v[28], v[122]))
+  k.emit(v_accvgpr_read(v[29], v[126]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[98], v[25]))
+  k.emit(v_accvgpr_write(v[102], v[26]))
+  k.emit(v_accvgpr_write(v[106], v[27]))
+  k.emit(v_accvgpr_write(v[110], v[28]))
+  k.emit(v_accvgpr_write(v[114], v[29]))
+  k.emit(v_accvgpr_read(v[25], v[142]))
+  k.emit(v_accvgpr_read(v[26], v[146]))
+  k.emit(v_accvgpr_read(v[27], v[150]))
+  k.emit(v_accvgpr_read(v[28], v[154]))
+  k.emit(v_accvgpr_read(v[29], v[158]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[130], v[25]))
+  k.emit(v_accvgpr_write(v[134], v[26]))
+  k.emit(v_accvgpr_write(v[138], v[27]))
+  k.emit(v_accvgpr_write(v[142], v[28]))
+  k.emit(v_accvgpr_write(v[146], v[29]))
+  k.emit(v_accvgpr_read(v[25], v[174]))
+  k.emit(v_accvgpr_read(v[26], v[178]))
+  k.emit(v_accvgpr_read(v[27], v[182]))
+  k.emit(v_accvgpr_read(v[28], v[186]))
+  k.emit(v_accvgpr_read(v[29], v[190]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[162], v[25]))
+  k.emit(v_accvgpr_write(v[166], v[26]))
+  k.emit(v_accvgpr_write(v[170], v[27]))
+  k.emit(v_accvgpr_write(v[174], v[28]))
+  k.emit(v_accvgpr_write(v[178], v[29]))
+  k.emit(v_accvgpr_read(v[25], v[206]))
+  k.emit(v_accvgpr_read(v[26], v[210]))
+  k.emit(v_accvgpr_read(v[27], v[214]))
+  k.emit(v_accvgpr_read(v[28], v[218]))
+  k.emit(v_accvgpr_read(v[29], v[222]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[194], v[25]))
+  k.emit(v_accvgpr_write(v[198], v[26]))
+  k.emit(v_accvgpr_write(v[202], v[27]))
+  k.emit(v_accvgpr_write(v[206], v[28]))
+  k.emit(v_accvgpr_write(v[210], v[29]))
+  k.emit(v_accvgpr_read(v[25], v[238]))
+  k.emit(v_accvgpr_read(v[26], v[242]))
+  k.emit(v_accvgpr_read(v[27], v[246]))
+  k.emit(v_accvgpr_read(v[28], v[250]))
+  k.emit(v_accvgpr_read(v[29], v[254]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[226], v[25]))
+  k.emit(v_accvgpr_write(v[230], v[26]))
+  k.emit(v_accvgpr_write(v[234], v[27]))
+  k.emit(v_accvgpr_write(v[238], v[28]))
+  k.emit(v_accvgpr_write(v[242], v[29]))
+  k.emit(v_accvgpr_read(v[25], v[15]))
+  k.emit(v_accvgpr_read(v[26], v[19]))
+  k.emit(v_accvgpr_read(v[27], v[23]))
+  k.emit(v_accvgpr_read(v[28], v[27]))
+  k.emit(v_accvgpr_read(v[29], v[31]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[3], v[25]))
+  k.emit(v_accvgpr_write(v[7], v[26]))
+  k.emit(v_accvgpr_write(v[11], v[27]))
+  k.emit(v_accvgpr_write(v[15], v[28]))
+  k.emit(v_accvgpr_write(v[19], v[29]))
+  k.emit(v_accvgpr_read(v[25], v[47]))
+  k.emit(v_accvgpr_read(v[26], v[51]))
+  k.emit(v_accvgpr_read(v[27], v[55]))
+  k.emit(v_accvgpr_read(v[28], v[59]))
+  k.emit(v_accvgpr_read(v[29], v[63]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[35], v[25]))
+  k.emit(v_accvgpr_write(v[39], v[26]))
+  k.emit(v_accvgpr_write(v[43], v[27]))
+  k.emit(v_accvgpr_write(v[47], v[28]))
+  k.emit(v_accvgpr_write(v[51], v[29]))
+  k.emit(v_accvgpr_read(v[25], v[79]))
+  k.emit(v_accvgpr_read(v[26], v[83]))
+  k.emit(v_accvgpr_read(v[27], v[87]))
+  k.emit(v_accvgpr_read(v[28], v[91]))
+  k.emit(v_accvgpr_read(v[29], v[95]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[67], v[25]))
+  k.emit(v_accvgpr_write(v[71], v[26]))
+  k.emit(v_accvgpr_write(v[75], v[27]))
+  k.emit(v_accvgpr_write(v[79], v[28]))
+  k.emit(v_accvgpr_write(v[83], v[29]))
+  k.emit(v_accvgpr_read(v[25], v[111]))
+  k.emit(v_accvgpr_read(v[26], v[115]))
+  k.emit(v_accvgpr_read(v[27], v[119]))
+  k.emit(v_accvgpr_read(v[28], v[123]))
+  k.emit(v_accvgpr_read(v[29], v[127]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[99], v[25]))
+  k.emit(v_accvgpr_write(v[103], v[26]))
+  k.emit(v_accvgpr_write(v[107], v[27]))
+  k.emit(v_accvgpr_write(v[111], v[28]))
+  k.emit(v_accvgpr_write(v[115], v[29]))
+  k.emit(v_accvgpr_read(v[25], v[143]))
+  k.emit(v_accvgpr_read(v[26], v[147]))
+  k.emit(v_accvgpr_read(v[27], v[151]))
+  k.emit(v_accvgpr_read(v[28], v[155]))
+  k.emit(v_accvgpr_read(v[29], v[159]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[131], v[25]))
+  k.emit(v_accvgpr_write(v[135], v[26]))
+  k.emit(v_accvgpr_write(v[139], v[27]))
+  k.emit(v_accvgpr_write(v[143], v[28]))
+  k.emit(v_accvgpr_write(v[147], v[29]))
+  k.emit(v_accvgpr_read(v[25], v[175]))
+  k.emit(v_accvgpr_read(v[26], v[179]))
+  k.emit(v_accvgpr_read(v[27], v[183]))
+  k.emit(v_accvgpr_read(v[28], v[187]))
+  k.emit(v_accvgpr_read(v[29], v[191]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[163], v[25]))
+  k.emit(v_accvgpr_write(v[167], v[26]))
+  k.emit(v_accvgpr_write(v[171], v[27]))
+  k.emit(v_accvgpr_write(v[175], v[28]))
+  k.emit(v_accvgpr_write(v[179], v[29]))
+  k.emit(v_accvgpr_read(v[25], v[207]))
+  k.emit(v_accvgpr_read(v[26], v[211]))
+  k.emit(v_accvgpr_read(v[27], v[215]))
+  k.emit(v_accvgpr_read(v[28], v[219]))
+  k.emit(v_accvgpr_read(v[29], v[223]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[195], v[25]))
+  k.emit(v_accvgpr_write(v[199], v[26]))
+  k.emit(v_accvgpr_write(v[203], v[27]))
+  k.emit(v_accvgpr_write(v[207], v[28]))
+  k.emit(v_accvgpr_write(v[211], v[29]))
+  k.emit(v_accvgpr_read(v[25], v[239]))
+  k.emit(v_accvgpr_read(v[26], v[243]))
+  k.emit(v_accvgpr_read(v[27], v[247]))
+  k.emit(v_accvgpr_read(v[28], v[251]))
+  k.emit(v_accvgpr_read(v[29], v[255]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[227], v[25]))
+  k.emit(v_accvgpr_write(v[231], v[26]))
+  k.emit(v_accvgpr_write(v[235], v[27]))
+  k.emit(v_accvgpr_write(v[239], v[28]))
+  k.emit(v_accvgpr_write(v[243], v[29]))
+  k.emit(s_mov_b64(s[8:9], -1))
+  k.emit(s_or_saveexec_b64(VCC, s[8:9]))
+  k.emit(s_branch(), target='ShiftVectorComponents0_GLVW0')
+  k.label('ShiftVectorComponents0_GLVW6_BM0_VW0')
+  k.emit(s_mov_b32(s[8], 0))
+  k.emit(v_cmpx_eq_u32_e64(s[8:9], v[24], s[8]))
+  k.emit(v_and_b32_e32(v[18], 63, v[180]))
+  k.emit(v_lshlrev_b32_e32(v[18], 2, v[18]))
+  k.emit(v_accvgpr_read(v[25], v[8]))
+  k.emit(v_accvgpr_read(v[26], v[12]))
+  k.emit(v_accvgpr_read(v[27], v[16]))
+  k.emit(v_accvgpr_read(v[28], v[20]))
+  k.emit(v_accvgpr_read(v[29], v[24]))
+  k.emit(v_accvgpr_read(v[30], v[28]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[0], v[25]))
+  k.emit(v_accvgpr_write(v[4], v[26]))
+  k.emit(v_accvgpr_write(v[8], v[27]))
+  k.emit(v_accvgpr_write(v[12], v[28]))
+  k.emit(v_accvgpr_write(v[16], v[29]))
+  k.emit(v_accvgpr_write(v[20], v[30]))
+  k.emit(v_accvgpr_read(v[25], v[40]))
+  k.emit(v_accvgpr_read(v[26], v[44]))
+  k.emit(v_accvgpr_read(v[27], v[48]))
+  k.emit(v_accvgpr_read(v[28], v[52]))
+  k.emit(v_accvgpr_read(v[29], v[56]))
+  k.emit(v_accvgpr_read(v[30], v[60]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[32], v[25]))
+  k.emit(v_accvgpr_write(v[36], v[26]))
+  k.emit(v_accvgpr_write(v[40], v[27]))
+  k.emit(v_accvgpr_write(v[44], v[28]))
+  k.emit(v_accvgpr_write(v[48], v[29]))
+  k.emit(v_accvgpr_write(v[52], v[30]))
+  k.emit(v_accvgpr_read(v[25], v[72]))
+  k.emit(v_accvgpr_read(v[26], v[76]))
+  k.emit(v_accvgpr_read(v[27], v[80]))
+  k.emit(v_accvgpr_read(v[28], v[84]))
+  k.emit(v_accvgpr_read(v[29], v[88]))
+  k.emit(v_accvgpr_read(v[30], v[92]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[64], v[25]))
+  k.emit(v_accvgpr_write(v[68], v[26]))
+  k.emit(v_accvgpr_write(v[72], v[27]))
+  k.emit(v_accvgpr_write(v[76], v[28]))
+  k.emit(v_accvgpr_write(v[80], v[29]))
+  k.emit(v_accvgpr_write(v[84], v[30]))
+  k.emit(v_accvgpr_read(v[25], v[104]))
+  k.emit(v_accvgpr_read(v[26], v[108]))
+  k.emit(v_accvgpr_read(v[27], v[112]))
+  k.emit(v_accvgpr_read(v[28], v[116]))
+  k.emit(v_accvgpr_read(v[29], v[120]))
+  k.emit(v_accvgpr_read(v[30], v[124]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[96], v[25]))
+  k.emit(v_accvgpr_write(v[100], v[26]))
+  k.emit(v_accvgpr_write(v[104], v[27]))
+  k.emit(v_accvgpr_write(v[108], v[28]))
+  k.emit(v_accvgpr_write(v[112], v[29]))
+  k.emit(v_accvgpr_write(v[116], v[30]))
+  k.emit(v_accvgpr_read(v[25], v[136]))
+  k.emit(v_accvgpr_read(v[26], v[140]))
+  k.emit(v_accvgpr_read(v[27], v[144]))
+  k.emit(v_accvgpr_read(v[28], v[148]))
+  k.emit(v_accvgpr_read(v[29], v[152]))
+  k.emit(v_accvgpr_read(v[30], v[156]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[128], v[25]))
+  k.emit(v_accvgpr_write(v[132], v[26]))
+  k.emit(v_accvgpr_write(v[136], v[27]))
+  k.emit(v_accvgpr_write(v[140], v[28]))
+  k.emit(v_accvgpr_write(v[144], v[29]))
+  k.emit(v_accvgpr_write(v[148], v[30]))
+  k.emit(v_accvgpr_read(v[25], v[168]))
+  k.emit(v_accvgpr_read(v[26], v[172]))
+  k.emit(v_accvgpr_read(v[27], v[176]))
+  k.emit(v_accvgpr_read(v[28], v[180]))
+  k.emit(v_accvgpr_read(v[29], v[184]))
+  k.emit(v_accvgpr_read(v[30], v[188]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[160], v[25]))
+  k.emit(v_accvgpr_write(v[164], v[26]))
+  k.emit(v_accvgpr_write(v[168], v[27]))
+  k.emit(v_accvgpr_write(v[172], v[28]))
+  k.emit(v_accvgpr_write(v[176], v[29]))
+  k.emit(v_accvgpr_write(v[180], v[30]))
+  k.emit(v_accvgpr_read(v[25], v[200]))
+  k.emit(v_accvgpr_read(v[26], v[204]))
+  k.emit(v_accvgpr_read(v[27], v[208]))
+  k.emit(v_accvgpr_read(v[28], v[212]))
+  k.emit(v_accvgpr_read(v[29], v[216]))
+  k.emit(v_accvgpr_read(v[30], v[220]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[192], v[25]))
+  k.emit(v_accvgpr_write(v[196], v[26]))
+  k.emit(v_accvgpr_write(v[200], v[27]))
+  k.emit(v_accvgpr_write(v[204], v[28]))
+  k.emit(v_accvgpr_write(v[208], v[29]))
+  k.emit(v_accvgpr_write(v[212], v[30]))
+  k.emit(v_accvgpr_read(v[25], v[232]))
+  k.emit(v_accvgpr_read(v[26], v[236]))
+  k.emit(v_accvgpr_read(v[27], v[240]))
+  k.emit(v_accvgpr_read(v[28], v[244]))
+  k.emit(v_accvgpr_read(v[29], v[248]))
+  k.emit(v_accvgpr_read(v[30], v[252]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[224], v[25]))
+  k.emit(v_accvgpr_write(v[228], v[26]))
+  k.emit(v_accvgpr_write(v[232], v[27]))
+  k.emit(v_accvgpr_write(v[236], v[28]))
+  k.emit(v_accvgpr_write(v[240], v[29]))
+  k.emit(v_accvgpr_write(v[244], v[30]))
+  k.emit(v_accvgpr_read(v[25], v[9]))
+  k.emit(v_accvgpr_read(v[26], v[13]))
+  k.emit(v_accvgpr_read(v[27], v[17]))
+  k.emit(v_accvgpr_read(v[28], v[21]))
+  k.emit(v_accvgpr_read(v[29], v[25]))
+  k.emit(v_accvgpr_read(v[30], v[29]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[1], v[25]))
+  k.emit(v_accvgpr_write(v[5], v[26]))
+  k.emit(v_accvgpr_write(v[9], v[27]))
+  k.emit(v_accvgpr_write(v[13], v[28]))
+  k.emit(v_accvgpr_write(v[17], v[29]))
+  k.emit(v_accvgpr_write(v[21], v[30]))
+  k.emit(v_accvgpr_read(v[25], v[41]))
+  k.emit(v_accvgpr_read(v[26], v[45]))
+  k.emit(v_accvgpr_read(v[27], v[49]))
+  k.emit(v_accvgpr_read(v[28], v[53]))
+  k.emit(v_accvgpr_read(v[29], v[57]))
+  k.emit(v_accvgpr_read(v[30], v[61]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[33], v[25]))
+  k.emit(v_accvgpr_write(v[37], v[26]))
+  k.emit(v_accvgpr_write(v[41], v[27]))
+  k.emit(v_accvgpr_write(v[45], v[28]))
+  k.emit(v_accvgpr_write(v[49], v[29]))
+  k.emit(v_accvgpr_write(v[53], v[30]))
+  k.emit(v_accvgpr_read(v[25], v[73]))
+  k.emit(v_accvgpr_read(v[26], v[77]))
+  k.emit(v_accvgpr_read(v[27], v[81]))
+  k.emit(v_accvgpr_read(v[28], v[85]))
+  k.emit(v_accvgpr_read(v[29], v[89]))
+  k.emit(v_accvgpr_read(v[30], v[93]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[65], v[25]))
+  k.emit(v_accvgpr_write(v[69], v[26]))
+  k.emit(v_accvgpr_write(v[73], v[27]))
+  k.emit(v_accvgpr_write(v[77], v[28]))
+  k.emit(v_accvgpr_write(v[81], v[29]))
+  k.emit(v_accvgpr_write(v[85], v[30]))
+  k.emit(v_accvgpr_read(v[25], v[105]))
+  k.emit(v_accvgpr_read(v[26], v[109]))
+  k.emit(v_accvgpr_read(v[27], v[113]))
+  k.emit(v_accvgpr_read(v[28], v[117]))
+  k.emit(v_accvgpr_read(v[29], v[121]))
+  k.emit(v_accvgpr_read(v[30], v[125]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[97], v[25]))
+  k.emit(v_accvgpr_write(v[101], v[26]))
+  k.emit(v_accvgpr_write(v[105], v[27]))
+  k.emit(v_accvgpr_write(v[109], v[28]))
+  k.emit(v_accvgpr_write(v[113], v[29]))
+  k.emit(v_accvgpr_write(v[117], v[30]))
+  k.emit(v_accvgpr_read(v[25], v[137]))
+  k.emit(v_accvgpr_read(v[26], v[141]))
+  k.emit(v_accvgpr_read(v[27], v[145]))
+  k.emit(v_accvgpr_read(v[28], v[149]))
+  k.emit(v_accvgpr_read(v[29], v[153]))
+  k.emit(v_accvgpr_read(v[30], v[157]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[129], v[25]))
+  k.emit(v_accvgpr_write(v[133], v[26]))
+  k.emit(v_accvgpr_write(v[137], v[27]))
+  k.emit(v_accvgpr_write(v[141], v[28]))
+  k.emit(v_accvgpr_write(v[145], v[29]))
+  k.emit(v_accvgpr_write(v[149], v[30]))
+  k.emit(v_accvgpr_read(v[25], v[169]))
+  k.emit(v_accvgpr_read(v[26], v[173]))
+  k.emit(v_accvgpr_read(v[27], v[177]))
+  k.emit(v_accvgpr_read(v[28], v[181]))
+  k.emit(v_accvgpr_read(v[29], v[185]))
+  k.emit(v_accvgpr_read(v[30], v[189]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[161], v[25]))
+  k.emit(v_accvgpr_write(v[165], v[26]))
+  k.emit(v_accvgpr_write(v[169], v[27]))
+  k.emit(v_accvgpr_write(v[173], v[28]))
+  k.emit(v_accvgpr_write(v[177], v[29]))
+  k.emit(v_accvgpr_write(v[181], v[30]))
+  k.emit(v_accvgpr_read(v[25], v[201]))
+  k.emit(v_accvgpr_read(v[26], v[205]))
+  k.emit(v_accvgpr_read(v[27], v[209]))
+  k.emit(v_accvgpr_read(v[28], v[213]))
+  k.emit(v_accvgpr_read(v[29], v[217]))
+  k.emit(v_accvgpr_read(v[30], v[221]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[193], v[25]))
+  k.emit(v_accvgpr_write(v[197], v[26]))
+  k.emit(v_accvgpr_write(v[201], v[27]))
+  k.emit(v_accvgpr_write(v[205], v[28]))
+  k.emit(v_accvgpr_write(v[209], v[29]))
+  k.emit(v_accvgpr_write(v[213], v[30]))
+  k.emit(v_accvgpr_read(v[25], v[233]))
+  k.emit(v_accvgpr_read(v[26], v[237]))
+  k.emit(v_accvgpr_read(v[27], v[241]))
+  k.emit(v_accvgpr_read(v[28], v[245]))
+  k.emit(v_accvgpr_read(v[29], v[249]))
+  k.emit(v_accvgpr_read(v[30], v[253]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[225], v[25]))
+  k.emit(v_accvgpr_write(v[229], v[26]))
+  k.emit(v_accvgpr_write(v[233], v[27]))
+  k.emit(v_accvgpr_write(v[237], v[28]))
+  k.emit(v_accvgpr_write(v[241], v[29]))
+  k.emit(v_accvgpr_write(v[245], v[30]))
+  k.emit(v_accvgpr_read(v[25], v[10]))
+  k.emit(v_accvgpr_read(v[26], v[14]))
+  k.emit(v_accvgpr_read(v[27], v[18]))
+  k.emit(v_accvgpr_read(v[28], v[22]))
+  k.emit(v_accvgpr_read(v[29], v[26]))
+  k.emit(v_accvgpr_read(v[30], v[30]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[2], v[25]))
+  k.emit(v_accvgpr_write(v[6], v[26]))
+  k.emit(v_accvgpr_write(v[10], v[27]))
+  k.emit(v_accvgpr_write(v[14], v[28]))
+  k.emit(v_accvgpr_write(v[18], v[29]))
+  k.emit(v_accvgpr_write(v[22], v[30]))
+  k.emit(v_accvgpr_read(v[25], v[42]))
+  k.emit(v_accvgpr_read(v[26], v[46]))
+  k.emit(v_accvgpr_read(v[27], v[50]))
+  k.emit(v_accvgpr_read(v[28], v[54]))
+  k.emit(v_accvgpr_read(v[29], v[58]))
+  k.emit(v_accvgpr_read(v[30], v[62]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[34], v[25]))
+  k.emit(v_accvgpr_write(v[38], v[26]))
+  k.emit(v_accvgpr_write(v[42], v[27]))
+  k.emit(v_accvgpr_write(v[46], v[28]))
+  k.emit(v_accvgpr_write(v[50], v[29]))
+  k.emit(v_accvgpr_write(v[54], v[30]))
+  k.emit(v_accvgpr_read(v[25], v[74]))
+  k.emit(v_accvgpr_read(v[26], v[78]))
+  k.emit(v_accvgpr_read(v[27], v[82]))
+  k.emit(v_accvgpr_read(v[28], v[86]))
+  k.emit(v_accvgpr_read(v[29], v[90]))
+  k.emit(v_accvgpr_read(v[30], v[94]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[66], v[25]))
+  k.emit(v_accvgpr_write(v[70], v[26]))
+  k.emit(v_accvgpr_write(v[74], v[27]))
+  k.emit(v_accvgpr_write(v[78], v[28]))
+  k.emit(v_accvgpr_write(v[82], v[29]))
+  k.emit(v_accvgpr_write(v[86], v[30]))
+  k.emit(v_accvgpr_read(v[25], v[106]))
+  k.emit(v_accvgpr_read(v[26], v[110]))
+  k.emit(v_accvgpr_read(v[27], v[114]))
+  k.emit(v_accvgpr_read(v[28], v[118]))
+  k.emit(v_accvgpr_read(v[29], v[122]))
+  k.emit(v_accvgpr_read(v[30], v[126]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[98], v[25]))
+  k.emit(v_accvgpr_write(v[102], v[26]))
+  k.emit(v_accvgpr_write(v[106], v[27]))
+  k.emit(v_accvgpr_write(v[110], v[28]))
+  k.emit(v_accvgpr_write(v[114], v[29]))
+  k.emit(v_accvgpr_write(v[118], v[30]))
+  k.emit(v_accvgpr_read(v[25], v[138]))
+  k.emit(v_accvgpr_read(v[26], v[142]))
+  k.emit(v_accvgpr_read(v[27], v[146]))
+  k.emit(v_accvgpr_read(v[28], v[150]))
+  k.emit(v_accvgpr_read(v[29], v[154]))
+  k.emit(v_accvgpr_read(v[30], v[158]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[130], v[25]))
+  k.emit(v_accvgpr_write(v[134], v[26]))
+  k.emit(v_accvgpr_write(v[138], v[27]))
+  k.emit(v_accvgpr_write(v[142], v[28]))
+  k.emit(v_accvgpr_write(v[146], v[29]))
+  k.emit(v_accvgpr_write(v[150], v[30]))
+  k.emit(v_accvgpr_read(v[25], v[170]))
+  k.emit(v_accvgpr_read(v[26], v[174]))
+  k.emit(v_accvgpr_read(v[27], v[178]))
+  k.emit(v_accvgpr_read(v[28], v[182]))
+  k.emit(v_accvgpr_read(v[29], v[186]))
+  k.emit(v_accvgpr_read(v[30], v[190]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[162], v[25]))
+  k.emit(v_accvgpr_write(v[166], v[26]))
+  k.emit(v_accvgpr_write(v[170], v[27]))
+  k.emit(v_accvgpr_write(v[174], v[28]))
+  k.emit(v_accvgpr_write(v[178], v[29]))
+  k.emit(v_accvgpr_write(v[182], v[30]))
+  k.emit(v_accvgpr_read(v[25], v[202]))
+  k.emit(v_accvgpr_read(v[26], v[206]))
+  k.emit(v_accvgpr_read(v[27], v[210]))
+  k.emit(v_accvgpr_read(v[28], v[214]))
+  k.emit(v_accvgpr_read(v[29], v[218]))
+  k.emit(v_accvgpr_read(v[30], v[222]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[194], v[25]))
+  k.emit(v_accvgpr_write(v[198], v[26]))
+  k.emit(v_accvgpr_write(v[202], v[27]))
+  k.emit(v_accvgpr_write(v[206], v[28]))
+  k.emit(v_accvgpr_write(v[210], v[29]))
+  k.emit(v_accvgpr_write(v[214], v[30]))
+  k.emit(v_accvgpr_read(v[25], v[234]))
+  k.emit(v_accvgpr_read(v[26], v[238]))
+  k.emit(v_accvgpr_read(v[27], v[242]))
+  k.emit(v_accvgpr_read(v[28], v[246]))
+  k.emit(v_accvgpr_read(v[29], v[250]))
+  k.emit(v_accvgpr_read(v[30], v[254]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[226], v[25]))
+  k.emit(v_accvgpr_write(v[230], v[26]))
+  k.emit(v_accvgpr_write(v[234], v[27]))
+  k.emit(v_accvgpr_write(v[238], v[28]))
+  k.emit(v_accvgpr_write(v[242], v[29]))
+  k.emit(v_accvgpr_write(v[246], v[30]))
+  k.emit(v_accvgpr_read(v[25], v[11]))
+  k.emit(v_accvgpr_read(v[26], v[15]))
+  k.emit(v_accvgpr_read(v[27], v[19]))
+  k.emit(v_accvgpr_read(v[28], v[23]))
+  k.emit(v_accvgpr_read(v[29], v[27]))
+  k.emit(v_accvgpr_read(v[30], v[31]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[3], v[25]))
+  k.emit(v_accvgpr_write(v[7], v[26]))
+  k.emit(v_accvgpr_write(v[11], v[27]))
+  k.emit(v_accvgpr_write(v[15], v[28]))
+  k.emit(v_accvgpr_write(v[19], v[29]))
+  k.emit(v_accvgpr_write(v[23], v[30]))
+  k.emit(v_accvgpr_read(v[25], v[43]))
+  k.emit(v_accvgpr_read(v[26], v[47]))
+  k.emit(v_accvgpr_read(v[27], v[51]))
+  k.emit(v_accvgpr_read(v[28], v[55]))
+  k.emit(v_accvgpr_read(v[29], v[59]))
+  k.emit(v_accvgpr_read(v[30], v[63]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[35], v[25]))
+  k.emit(v_accvgpr_write(v[39], v[26]))
+  k.emit(v_accvgpr_write(v[43], v[27]))
+  k.emit(v_accvgpr_write(v[47], v[28]))
+  k.emit(v_accvgpr_write(v[51], v[29]))
+  k.emit(v_accvgpr_write(v[55], v[30]))
+  k.emit(v_accvgpr_read(v[25], v[75]))
+  k.emit(v_accvgpr_read(v[26], v[79]))
+  k.emit(v_accvgpr_read(v[27], v[83]))
+  k.emit(v_accvgpr_read(v[28], v[87]))
+  k.emit(v_accvgpr_read(v[29], v[91]))
+  k.emit(v_accvgpr_read(v[30], v[95]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[67], v[25]))
+  k.emit(v_accvgpr_write(v[71], v[26]))
+  k.emit(v_accvgpr_write(v[75], v[27]))
+  k.emit(v_accvgpr_write(v[79], v[28]))
+  k.emit(v_accvgpr_write(v[83], v[29]))
+  k.emit(v_accvgpr_write(v[87], v[30]))
+  k.emit(v_accvgpr_read(v[25], v[107]))
+  k.emit(v_accvgpr_read(v[26], v[111]))
+  k.emit(v_accvgpr_read(v[27], v[115]))
+  k.emit(v_accvgpr_read(v[28], v[119]))
+  k.emit(v_accvgpr_read(v[29], v[123]))
+  k.emit(v_accvgpr_read(v[30], v[127]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[99], v[25]))
+  k.emit(v_accvgpr_write(v[103], v[26]))
+  k.emit(v_accvgpr_write(v[107], v[27]))
+  k.emit(v_accvgpr_write(v[111], v[28]))
+  k.emit(v_accvgpr_write(v[115], v[29]))
+  k.emit(v_accvgpr_write(v[119], v[30]))
+  k.emit(v_accvgpr_read(v[25], v[139]))
+  k.emit(v_accvgpr_read(v[26], v[143]))
+  k.emit(v_accvgpr_read(v[27], v[147]))
+  k.emit(v_accvgpr_read(v[28], v[151]))
+  k.emit(v_accvgpr_read(v[29], v[155]))
+  k.emit(v_accvgpr_read(v[30], v[159]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[131], v[25]))
+  k.emit(v_accvgpr_write(v[135], v[26]))
+  k.emit(v_accvgpr_write(v[139], v[27]))
+  k.emit(v_accvgpr_write(v[143], v[28]))
+  k.emit(v_accvgpr_write(v[147], v[29]))
+  k.emit(v_accvgpr_write(v[151], v[30]))
+  k.emit(v_accvgpr_read(v[25], v[171]))
+  k.emit(v_accvgpr_read(v[26], v[175]))
+  k.emit(v_accvgpr_read(v[27], v[179]))
+  k.emit(v_accvgpr_read(v[28], v[183]))
+  k.emit(v_accvgpr_read(v[29], v[187]))
+  k.emit(v_accvgpr_read(v[30], v[191]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[163], v[25]))
+  k.emit(v_accvgpr_write(v[167], v[26]))
+  k.emit(v_accvgpr_write(v[171], v[27]))
+  k.emit(v_accvgpr_write(v[175], v[28]))
+  k.emit(v_accvgpr_write(v[179], v[29]))
+  k.emit(v_accvgpr_write(v[183], v[30]))
+  k.emit(v_accvgpr_read(v[25], v[203]))
+  k.emit(v_accvgpr_read(v[26], v[207]))
+  k.emit(v_accvgpr_read(v[27], v[211]))
+  k.emit(v_accvgpr_read(v[28], v[215]))
+  k.emit(v_accvgpr_read(v[29], v[219]))
+  k.emit(v_accvgpr_read(v[30], v[223]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[195], v[25]))
+  k.emit(v_accvgpr_write(v[199], v[26]))
+  k.emit(v_accvgpr_write(v[203], v[27]))
+  k.emit(v_accvgpr_write(v[207], v[28]))
+  k.emit(v_accvgpr_write(v[211], v[29]))
+  k.emit(v_accvgpr_write(v[215], v[30]))
+  k.emit(v_accvgpr_read(v[25], v[235]))
+  k.emit(v_accvgpr_read(v[26], v[239]))
+  k.emit(v_accvgpr_read(v[27], v[243]))
+  k.emit(v_accvgpr_read(v[28], v[247]))
+  k.emit(v_accvgpr_read(v[29], v[251]))
+  k.emit(v_accvgpr_read(v[30], v[255]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[227], v[25]))
+  k.emit(v_accvgpr_write(v[231], v[26]))
+  k.emit(v_accvgpr_write(v[235], v[27]))
+  k.emit(v_accvgpr_write(v[239], v[28]))
+  k.emit(v_accvgpr_write(v[243], v[29]))
+  k.emit(v_accvgpr_write(v[247], v[30]))
+  k.emit(s_mov_b64(s[8:9], -1))
+  k.emit(s_or_saveexec_b64(VCC, s[8:9]))
+  k.emit(s_branch(), target='ShiftVectorComponents0_GLVW0')
+  k.label('ShiftVectorComponents0_GLVW7_BM0_VW0')
+  k.emit(s_mov_b32(s[8], 0))
+  k.emit(v_cmpx_eq_u32_e64(s[8:9], v[24], s[8]))
+  k.emit(v_and_b32_e32(v[18], 63, v[180]))
+  k.emit(v_lshlrev_b32_e32(v[18], 2, v[18]))
+  k.emit(v_accvgpr_read(v[25], v[4]))
+  k.emit(v_accvgpr_read(v[26], v[8]))
+  k.emit(v_accvgpr_read(v[27], v[12]))
+  k.emit(v_accvgpr_read(v[28], v[16]))
+  k.emit(v_accvgpr_read(v[29], v[20]))
+  k.emit(v_accvgpr_read(v[30], v[24]))
+  k.emit(v_accvgpr_read(v[31], v[28]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[0], v[25]))
+  k.emit(v_accvgpr_write(v[4], v[26]))
+  k.emit(v_accvgpr_write(v[8], v[27]))
+  k.emit(v_accvgpr_write(v[12], v[28]))
+  k.emit(v_accvgpr_write(v[16], v[29]))
+  k.emit(v_accvgpr_write(v[20], v[30]))
+  k.emit(v_accvgpr_write(v[24], v[31]))
+  k.emit(v_accvgpr_read(v[25], v[36]))
+  k.emit(v_accvgpr_read(v[26], v[40]))
+  k.emit(v_accvgpr_read(v[27], v[44]))
+  k.emit(v_accvgpr_read(v[28], v[48]))
+  k.emit(v_accvgpr_read(v[29], v[52]))
+  k.emit(v_accvgpr_read(v[30], v[56]))
+  k.emit(v_accvgpr_read(v[31], v[60]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[32], v[25]))
+  k.emit(v_accvgpr_write(v[36], v[26]))
+  k.emit(v_accvgpr_write(v[40], v[27]))
+  k.emit(v_accvgpr_write(v[44], v[28]))
+  k.emit(v_accvgpr_write(v[48], v[29]))
+  k.emit(v_accvgpr_write(v[52], v[30]))
+  k.emit(v_accvgpr_write(v[56], v[31]))
+  k.emit(v_accvgpr_read(v[25], v[68]))
+  k.emit(v_accvgpr_read(v[26], v[72]))
+  k.emit(v_accvgpr_read(v[27], v[76]))
+  k.emit(v_accvgpr_read(v[28], v[80]))
+  k.emit(v_accvgpr_read(v[29], v[84]))
+  k.emit(v_accvgpr_read(v[30], v[88]))
+  k.emit(v_accvgpr_read(v[31], v[92]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[64], v[25]))
+  k.emit(v_accvgpr_write(v[68], v[26]))
+  k.emit(v_accvgpr_write(v[72], v[27]))
+  k.emit(v_accvgpr_write(v[76], v[28]))
+  k.emit(v_accvgpr_write(v[80], v[29]))
+  k.emit(v_accvgpr_write(v[84], v[30]))
+  k.emit(v_accvgpr_write(v[88], v[31]))
+  k.emit(v_accvgpr_read(v[25], v[100]))
+  k.emit(v_accvgpr_read(v[26], v[104]))
+  k.emit(v_accvgpr_read(v[27], v[108]))
+  k.emit(v_accvgpr_read(v[28], v[112]))
+  k.emit(v_accvgpr_read(v[29], v[116]))
+  k.emit(v_accvgpr_read(v[30], v[120]))
+  k.emit(v_accvgpr_read(v[31], v[124]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[96], v[25]))
+  k.emit(v_accvgpr_write(v[100], v[26]))
+  k.emit(v_accvgpr_write(v[104], v[27]))
+  k.emit(v_accvgpr_write(v[108], v[28]))
+  k.emit(v_accvgpr_write(v[112], v[29]))
+  k.emit(v_accvgpr_write(v[116], v[30]))
+  k.emit(v_accvgpr_write(v[120], v[31]))
+  k.emit(v_accvgpr_read(v[25], v[132]))
+  k.emit(v_accvgpr_read(v[26], v[136]))
+  k.emit(v_accvgpr_read(v[27], v[140]))
+  k.emit(v_accvgpr_read(v[28], v[144]))
+  k.emit(v_accvgpr_read(v[29], v[148]))
+  k.emit(v_accvgpr_read(v[30], v[152]))
+  k.emit(v_accvgpr_read(v[31], v[156]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[128], v[25]))
+  k.emit(v_accvgpr_write(v[132], v[26]))
+  k.emit(v_accvgpr_write(v[136], v[27]))
+  k.emit(v_accvgpr_write(v[140], v[28]))
+  k.emit(v_accvgpr_write(v[144], v[29]))
+  k.emit(v_accvgpr_write(v[148], v[30]))
+  k.emit(v_accvgpr_write(v[152], v[31]))
+  k.emit(v_accvgpr_read(v[25], v[164]))
+  k.emit(v_accvgpr_read(v[26], v[168]))
+  k.emit(v_accvgpr_read(v[27], v[172]))
+  k.emit(v_accvgpr_read(v[28], v[176]))
+  k.emit(v_accvgpr_read(v[29], v[180]))
+  k.emit(v_accvgpr_read(v[30], v[184]))
+  k.emit(v_accvgpr_read(v[31], v[188]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[160], v[25]))
+  k.emit(v_accvgpr_write(v[164], v[26]))
+  k.emit(v_accvgpr_write(v[168], v[27]))
+  k.emit(v_accvgpr_write(v[172], v[28]))
+  k.emit(v_accvgpr_write(v[176], v[29]))
+  k.emit(v_accvgpr_write(v[180], v[30]))
+  k.emit(v_accvgpr_write(v[184], v[31]))
+  k.emit(v_accvgpr_read(v[25], v[196]))
+  k.emit(v_accvgpr_read(v[26], v[200]))
+  k.emit(v_accvgpr_read(v[27], v[204]))
+  k.emit(v_accvgpr_read(v[28], v[208]))
+  k.emit(v_accvgpr_read(v[29], v[212]))
+  k.emit(v_accvgpr_read(v[30], v[216]))
+  k.emit(v_accvgpr_read(v[31], v[220]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[192], v[25]))
+  k.emit(v_accvgpr_write(v[196], v[26]))
+  k.emit(v_accvgpr_write(v[200], v[27]))
+  k.emit(v_accvgpr_write(v[204], v[28]))
+  k.emit(v_accvgpr_write(v[208], v[29]))
+  k.emit(v_accvgpr_write(v[212], v[30]))
+  k.emit(v_accvgpr_write(v[216], v[31]))
+  k.emit(v_accvgpr_read(v[25], v[228]))
+  k.emit(v_accvgpr_read(v[26], v[232]))
+  k.emit(v_accvgpr_read(v[27], v[236]))
+  k.emit(v_accvgpr_read(v[28], v[240]))
+  k.emit(v_accvgpr_read(v[29], v[244]))
+  k.emit(v_accvgpr_read(v[30], v[248]))
+  k.emit(v_accvgpr_read(v[31], v[252]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[224], v[25]))
+  k.emit(v_accvgpr_write(v[228], v[26]))
+  k.emit(v_accvgpr_write(v[232], v[27]))
+  k.emit(v_accvgpr_write(v[236], v[28]))
+  k.emit(v_accvgpr_write(v[240], v[29]))
+  k.emit(v_accvgpr_write(v[244], v[30]))
+  k.emit(v_accvgpr_write(v[248], v[31]))
+  k.emit(v_accvgpr_read(v[25], v[5]))
+  k.emit(v_accvgpr_read(v[26], v[9]))
+  k.emit(v_accvgpr_read(v[27], v[13]))
+  k.emit(v_accvgpr_read(v[28], v[17]))
+  k.emit(v_accvgpr_read(v[29], v[21]))
+  k.emit(v_accvgpr_read(v[30], v[25]))
+  k.emit(v_accvgpr_read(v[31], v[29]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[1], v[25]))
+  k.emit(v_accvgpr_write(v[5], v[26]))
+  k.emit(v_accvgpr_write(v[9], v[27]))
+  k.emit(v_accvgpr_write(v[13], v[28]))
+  k.emit(v_accvgpr_write(v[17], v[29]))
+  k.emit(v_accvgpr_write(v[21], v[30]))
+  k.emit(v_accvgpr_write(v[25], v[31]))
+  k.emit(v_accvgpr_read(v[25], v[37]))
+  k.emit(v_accvgpr_read(v[26], v[41]))
+  k.emit(v_accvgpr_read(v[27], v[45]))
+  k.emit(v_accvgpr_read(v[28], v[49]))
+  k.emit(v_accvgpr_read(v[29], v[53]))
+  k.emit(v_accvgpr_read(v[30], v[57]))
+  k.emit(v_accvgpr_read(v[31], v[61]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[33], v[25]))
+  k.emit(v_accvgpr_write(v[37], v[26]))
+  k.emit(v_accvgpr_write(v[41], v[27]))
+  k.emit(v_accvgpr_write(v[45], v[28]))
+  k.emit(v_accvgpr_write(v[49], v[29]))
+  k.emit(v_accvgpr_write(v[53], v[30]))
+  k.emit(v_accvgpr_write(v[57], v[31]))
+  k.emit(v_accvgpr_read(v[25], v[69]))
+  k.emit(v_accvgpr_read(v[26], v[73]))
+  k.emit(v_accvgpr_read(v[27], v[77]))
+  k.emit(v_accvgpr_read(v[28], v[81]))
+  k.emit(v_accvgpr_read(v[29], v[85]))
+  k.emit(v_accvgpr_read(v[30], v[89]))
+  k.emit(v_accvgpr_read(v[31], v[93]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[65], v[25]))
+  k.emit(v_accvgpr_write(v[69], v[26]))
+  k.emit(v_accvgpr_write(v[73], v[27]))
+  k.emit(v_accvgpr_write(v[77], v[28]))
+  k.emit(v_accvgpr_write(v[81], v[29]))
+  k.emit(v_accvgpr_write(v[85], v[30]))
+  k.emit(v_accvgpr_write(v[89], v[31]))
+  k.emit(v_accvgpr_read(v[25], v[101]))
+  k.emit(v_accvgpr_read(v[26], v[105]))
+  k.emit(v_accvgpr_read(v[27], v[109]))
+  k.emit(v_accvgpr_read(v[28], v[113]))
+  k.emit(v_accvgpr_read(v[29], v[117]))
+  k.emit(v_accvgpr_read(v[30], v[121]))
+  k.emit(v_accvgpr_read(v[31], v[125]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[97], v[25]))
+  k.emit(v_accvgpr_write(v[101], v[26]))
+  k.emit(v_accvgpr_write(v[105], v[27]))
+  k.emit(v_accvgpr_write(v[109], v[28]))
+  k.emit(v_accvgpr_write(v[113], v[29]))
+  k.emit(v_accvgpr_write(v[117], v[30]))
+  k.emit(v_accvgpr_write(v[121], v[31]))
+  k.emit(v_accvgpr_read(v[25], v[133]))
+  k.emit(v_accvgpr_read(v[26], v[137]))
+  k.emit(v_accvgpr_read(v[27], v[141]))
+  k.emit(v_accvgpr_read(v[28], v[145]))
+  k.emit(v_accvgpr_read(v[29], v[149]))
+  k.emit(v_accvgpr_read(v[30], v[153]))
+  k.emit(v_accvgpr_read(v[31], v[157]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[129], v[25]))
+  k.emit(v_accvgpr_write(v[133], v[26]))
+  k.emit(v_accvgpr_write(v[137], v[27]))
+  k.emit(v_accvgpr_write(v[141], v[28]))
+  k.emit(v_accvgpr_write(v[145], v[29]))
+  k.emit(v_accvgpr_write(v[149], v[30]))
+  k.emit(v_accvgpr_write(v[153], v[31]))
+  k.emit(v_accvgpr_read(v[25], v[165]))
+  k.emit(v_accvgpr_read(v[26], v[169]))
+  k.emit(v_accvgpr_read(v[27], v[173]))
+  k.emit(v_accvgpr_read(v[28], v[177]))
+  k.emit(v_accvgpr_read(v[29], v[181]))
+  k.emit(v_accvgpr_read(v[30], v[185]))
+  k.emit(v_accvgpr_read(v[31], v[189]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[161], v[25]))
+  k.emit(v_accvgpr_write(v[165], v[26]))
+  k.emit(v_accvgpr_write(v[169], v[27]))
+  k.emit(v_accvgpr_write(v[173], v[28]))
+  k.emit(v_accvgpr_write(v[177], v[29]))
+  k.emit(v_accvgpr_write(v[181], v[30]))
+  k.emit(v_accvgpr_write(v[185], v[31]))
+  k.emit(v_accvgpr_read(v[25], v[197]))
+  k.emit(v_accvgpr_read(v[26], v[201]))
+  k.emit(v_accvgpr_read(v[27], v[205]))
+  k.emit(v_accvgpr_read(v[28], v[209]))
+  k.emit(v_accvgpr_read(v[29], v[213]))
+  k.emit(v_accvgpr_read(v[30], v[217]))
+  k.emit(v_accvgpr_read(v[31], v[221]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[193], v[25]))
+  k.emit(v_accvgpr_write(v[197], v[26]))
+  k.emit(v_accvgpr_write(v[201], v[27]))
+  k.emit(v_accvgpr_write(v[205], v[28]))
+  k.emit(v_accvgpr_write(v[209], v[29]))
+  k.emit(v_accvgpr_write(v[213], v[30]))
+  k.emit(v_accvgpr_write(v[217], v[31]))
+  k.emit(v_accvgpr_read(v[25], v[229]))
+  k.emit(v_accvgpr_read(v[26], v[233]))
+  k.emit(v_accvgpr_read(v[27], v[237]))
+  k.emit(v_accvgpr_read(v[28], v[241]))
+  k.emit(v_accvgpr_read(v[29], v[245]))
+  k.emit(v_accvgpr_read(v[30], v[249]))
+  k.emit(v_accvgpr_read(v[31], v[253]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[225], v[25]))
+  k.emit(v_accvgpr_write(v[229], v[26]))
+  k.emit(v_accvgpr_write(v[233], v[27]))
+  k.emit(v_accvgpr_write(v[237], v[28]))
+  k.emit(v_accvgpr_write(v[241], v[29]))
+  k.emit(v_accvgpr_write(v[245], v[30]))
+  k.emit(v_accvgpr_write(v[249], v[31]))
+  k.emit(v_accvgpr_read(v[25], v[6]))
+  k.emit(v_accvgpr_read(v[26], v[10]))
+  k.emit(v_accvgpr_read(v[27], v[14]))
+  k.emit(v_accvgpr_read(v[28], v[18]))
+  k.emit(v_accvgpr_read(v[29], v[22]))
+  k.emit(v_accvgpr_read(v[30], v[26]))
+  k.emit(v_accvgpr_read(v[31], v[30]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[2], v[25]))
+  k.emit(v_accvgpr_write(v[6], v[26]))
+  k.emit(v_accvgpr_write(v[10], v[27]))
+  k.emit(v_accvgpr_write(v[14], v[28]))
+  k.emit(v_accvgpr_write(v[18], v[29]))
+  k.emit(v_accvgpr_write(v[22], v[30]))
+  k.emit(v_accvgpr_write(v[26], v[31]))
+  k.emit(v_accvgpr_read(v[25], v[38]))
+  k.emit(v_accvgpr_read(v[26], v[42]))
+  k.emit(v_accvgpr_read(v[27], v[46]))
+  k.emit(v_accvgpr_read(v[28], v[50]))
+  k.emit(v_accvgpr_read(v[29], v[54]))
+  k.emit(v_accvgpr_read(v[30], v[58]))
+  k.emit(v_accvgpr_read(v[31], v[62]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[34], v[25]))
+  k.emit(v_accvgpr_write(v[38], v[26]))
+  k.emit(v_accvgpr_write(v[42], v[27]))
+  k.emit(v_accvgpr_write(v[46], v[28]))
+  k.emit(v_accvgpr_write(v[50], v[29]))
+  k.emit(v_accvgpr_write(v[54], v[30]))
+  k.emit(v_accvgpr_write(v[58], v[31]))
+  k.emit(v_accvgpr_read(v[25], v[70]))
+  k.emit(v_accvgpr_read(v[26], v[74]))
+  k.emit(v_accvgpr_read(v[27], v[78]))
+  k.emit(v_accvgpr_read(v[28], v[82]))
+  k.emit(v_accvgpr_read(v[29], v[86]))
+  k.emit(v_accvgpr_read(v[30], v[90]))
+  k.emit(v_accvgpr_read(v[31], v[94]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[66], v[25]))
+  k.emit(v_accvgpr_write(v[70], v[26]))
+  k.emit(v_accvgpr_write(v[74], v[27]))
+  k.emit(v_accvgpr_write(v[78], v[28]))
+  k.emit(v_accvgpr_write(v[82], v[29]))
+  k.emit(v_accvgpr_write(v[86], v[30]))
+  k.emit(v_accvgpr_write(v[90], v[31]))
+  k.emit(v_accvgpr_read(v[25], v[102]))
+  k.emit(v_accvgpr_read(v[26], v[106]))
+  k.emit(v_accvgpr_read(v[27], v[110]))
+  k.emit(v_accvgpr_read(v[28], v[114]))
+  k.emit(v_accvgpr_read(v[29], v[118]))
+  k.emit(v_accvgpr_read(v[30], v[122]))
+  k.emit(v_accvgpr_read(v[31], v[126]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[98], v[25]))
+  k.emit(v_accvgpr_write(v[102], v[26]))
+  k.emit(v_accvgpr_write(v[106], v[27]))
+  k.emit(v_accvgpr_write(v[110], v[28]))
+  k.emit(v_accvgpr_write(v[114], v[29]))
+  k.emit(v_accvgpr_write(v[118], v[30]))
+  k.emit(v_accvgpr_write(v[122], v[31]))
+  k.emit(v_accvgpr_read(v[25], v[134]))
+  k.emit(v_accvgpr_read(v[26], v[138]))
+  k.emit(v_accvgpr_read(v[27], v[142]))
+  k.emit(v_accvgpr_read(v[28], v[146]))
+  k.emit(v_accvgpr_read(v[29], v[150]))
+  k.emit(v_accvgpr_read(v[30], v[154]))
+  k.emit(v_accvgpr_read(v[31], v[158]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[130], v[25]))
+  k.emit(v_accvgpr_write(v[134], v[26]))
+  k.emit(v_accvgpr_write(v[138], v[27]))
+  k.emit(v_accvgpr_write(v[142], v[28]))
+  k.emit(v_accvgpr_write(v[146], v[29]))
+  k.emit(v_accvgpr_write(v[150], v[30]))
+  k.emit(v_accvgpr_write(v[154], v[31]))
+  k.emit(v_accvgpr_read(v[25], v[166]))
+  k.emit(v_accvgpr_read(v[26], v[170]))
+  k.emit(v_accvgpr_read(v[27], v[174]))
+  k.emit(v_accvgpr_read(v[28], v[178]))
+  k.emit(v_accvgpr_read(v[29], v[182]))
+  k.emit(v_accvgpr_read(v[30], v[186]))
+  k.emit(v_accvgpr_read(v[31], v[190]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[162], v[25]))
+  k.emit(v_accvgpr_write(v[166], v[26]))
+  k.emit(v_accvgpr_write(v[170], v[27]))
+  k.emit(v_accvgpr_write(v[174], v[28]))
+  k.emit(v_accvgpr_write(v[178], v[29]))
+  k.emit(v_accvgpr_write(v[182], v[30]))
+  k.emit(v_accvgpr_write(v[186], v[31]))
+  k.emit(v_accvgpr_read(v[25], v[198]))
+  k.emit(v_accvgpr_read(v[26], v[202]))
+  k.emit(v_accvgpr_read(v[27], v[206]))
+  k.emit(v_accvgpr_read(v[28], v[210]))
+  k.emit(v_accvgpr_read(v[29], v[214]))
+  k.emit(v_accvgpr_read(v[30], v[218]))
+  k.emit(v_accvgpr_read(v[31], v[222]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[194], v[25]))
+  k.emit(v_accvgpr_write(v[198], v[26]))
+  k.emit(v_accvgpr_write(v[202], v[27]))
+  k.emit(v_accvgpr_write(v[206], v[28]))
+  k.emit(v_accvgpr_write(v[210], v[29]))
+  k.emit(v_accvgpr_write(v[214], v[30]))
+  k.emit(v_accvgpr_write(v[218], v[31]))
+  k.emit(v_accvgpr_read(v[25], v[230]))
+  k.emit(v_accvgpr_read(v[26], v[234]))
+  k.emit(v_accvgpr_read(v[27], v[238]))
+  k.emit(v_accvgpr_read(v[28], v[242]))
+  k.emit(v_accvgpr_read(v[29], v[246]))
+  k.emit(v_accvgpr_read(v[30], v[250]))
+  k.emit(v_accvgpr_read(v[31], v[254]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[226], v[25]))
+  k.emit(v_accvgpr_write(v[230], v[26]))
+  k.emit(v_accvgpr_write(v[234], v[27]))
+  k.emit(v_accvgpr_write(v[238], v[28]))
+  k.emit(v_accvgpr_write(v[242], v[29]))
+  k.emit(v_accvgpr_write(v[246], v[30]))
+  k.emit(v_accvgpr_write(v[250], v[31]))
+  k.emit(v_accvgpr_read(v[25], v[7]))
+  k.emit(v_accvgpr_read(v[26], v[11]))
+  k.emit(v_accvgpr_read(v[27], v[15]))
+  k.emit(v_accvgpr_read(v[28], v[19]))
+  k.emit(v_accvgpr_read(v[29], v[23]))
+  k.emit(v_accvgpr_read(v[30], v[27]))
+  k.emit(v_accvgpr_read(v[31], v[31]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[3], v[25]))
+  k.emit(v_accvgpr_write(v[7], v[26]))
+  k.emit(v_accvgpr_write(v[11], v[27]))
+  k.emit(v_accvgpr_write(v[15], v[28]))
+  k.emit(v_accvgpr_write(v[19], v[29]))
+  k.emit(v_accvgpr_write(v[23], v[30]))
+  k.emit(v_accvgpr_write(v[27], v[31]))
+  k.emit(v_accvgpr_read(v[25], v[39]))
+  k.emit(v_accvgpr_read(v[26], v[43]))
+  k.emit(v_accvgpr_read(v[27], v[47]))
+  k.emit(v_accvgpr_read(v[28], v[51]))
+  k.emit(v_accvgpr_read(v[29], v[55]))
+  k.emit(v_accvgpr_read(v[30], v[59]))
+  k.emit(v_accvgpr_read(v[31], v[63]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[35], v[25]))
+  k.emit(v_accvgpr_write(v[39], v[26]))
+  k.emit(v_accvgpr_write(v[43], v[27]))
+  k.emit(v_accvgpr_write(v[47], v[28]))
+  k.emit(v_accvgpr_write(v[51], v[29]))
+  k.emit(v_accvgpr_write(v[55], v[30]))
+  k.emit(v_accvgpr_write(v[59], v[31]))
+  k.emit(v_accvgpr_read(v[25], v[71]))
+  k.emit(v_accvgpr_read(v[26], v[75]))
+  k.emit(v_accvgpr_read(v[27], v[79]))
+  k.emit(v_accvgpr_read(v[28], v[83]))
+  k.emit(v_accvgpr_read(v[29], v[87]))
+  k.emit(v_accvgpr_read(v[30], v[91]))
+  k.emit(v_accvgpr_read(v[31], v[95]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[67], v[25]))
+  k.emit(v_accvgpr_write(v[71], v[26]))
+  k.emit(v_accvgpr_write(v[75], v[27]))
+  k.emit(v_accvgpr_write(v[79], v[28]))
+  k.emit(v_accvgpr_write(v[83], v[29]))
+  k.emit(v_accvgpr_write(v[87], v[30]))
+  k.emit(v_accvgpr_write(v[91], v[31]))
+  k.emit(v_accvgpr_read(v[25], v[103]))
+  k.emit(v_accvgpr_read(v[26], v[107]))
+  k.emit(v_accvgpr_read(v[27], v[111]))
+  k.emit(v_accvgpr_read(v[28], v[115]))
+  k.emit(v_accvgpr_read(v[29], v[119]))
+  k.emit(v_accvgpr_read(v[30], v[123]))
+  k.emit(v_accvgpr_read(v[31], v[127]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[99], v[25]))
+  k.emit(v_accvgpr_write(v[103], v[26]))
+  k.emit(v_accvgpr_write(v[107], v[27]))
+  k.emit(v_accvgpr_write(v[111], v[28]))
+  k.emit(v_accvgpr_write(v[115], v[29]))
+  k.emit(v_accvgpr_write(v[119], v[30]))
+  k.emit(v_accvgpr_write(v[123], v[31]))
+  k.emit(v_accvgpr_read(v[25], v[135]))
+  k.emit(v_accvgpr_read(v[26], v[139]))
+  k.emit(v_accvgpr_read(v[27], v[143]))
+  k.emit(v_accvgpr_read(v[28], v[147]))
+  k.emit(v_accvgpr_read(v[29], v[151]))
+  k.emit(v_accvgpr_read(v[30], v[155]))
+  k.emit(v_accvgpr_read(v[31], v[159]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[131], v[25]))
+  k.emit(v_accvgpr_write(v[135], v[26]))
+  k.emit(v_accvgpr_write(v[139], v[27]))
+  k.emit(v_accvgpr_write(v[143], v[28]))
+  k.emit(v_accvgpr_write(v[147], v[29]))
+  k.emit(v_accvgpr_write(v[151], v[30]))
+  k.emit(v_accvgpr_write(v[155], v[31]))
+  k.emit(v_accvgpr_read(v[25], v[167]))
+  k.emit(v_accvgpr_read(v[26], v[171]))
+  k.emit(v_accvgpr_read(v[27], v[175]))
+  k.emit(v_accvgpr_read(v[28], v[179]))
+  k.emit(v_accvgpr_read(v[29], v[183]))
+  k.emit(v_accvgpr_read(v[30], v[187]))
+  k.emit(v_accvgpr_read(v[31], v[191]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[163], v[25]))
+  k.emit(v_accvgpr_write(v[167], v[26]))
+  k.emit(v_accvgpr_write(v[171], v[27]))
+  k.emit(v_accvgpr_write(v[175], v[28]))
+  k.emit(v_accvgpr_write(v[179], v[29]))
+  k.emit(v_accvgpr_write(v[183], v[30]))
+  k.emit(v_accvgpr_write(v[187], v[31]))
+  k.emit(v_accvgpr_read(v[25], v[199]))
+  k.emit(v_accvgpr_read(v[26], v[203]))
+  k.emit(v_accvgpr_read(v[27], v[207]))
+  k.emit(v_accvgpr_read(v[28], v[211]))
+  k.emit(v_accvgpr_read(v[29], v[215]))
+  k.emit(v_accvgpr_read(v[30], v[219]))
+  k.emit(v_accvgpr_read(v[31], v[223]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[195], v[25]))
+  k.emit(v_accvgpr_write(v[199], v[26]))
+  k.emit(v_accvgpr_write(v[203], v[27]))
+  k.emit(v_accvgpr_write(v[207], v[28]))
+  k.emit(v_accvgpr_write(v[211], v[29]))
+  k.emit(v_accvgpr_write(v[215], v[30]))
+  k.emit(v_accvgpr_write(v[219], v[31]))
+  k.emit(v_accvgpr_read(v[25], v[231]))
+  k.emit(v_accvgpr_read(v[26], v[235]))
+  k.emit(v_accvgpr_read(v[27], v[239]))
+  k.emit(v_accvgpr_read(v[28], v[243]))
+  k.emit(v_accvgpr_read(v[29], v[247]))
+  k.emit(v_accvgpr_read(v[30], v[251]))
+  k.emit(v_accvgpr_read(v[31], v[255]))
+  k.emit(s_nop(1))
+  k.emit(v_accvgpr_write(v[227], v[25]))
+  k.emit(v_accvgpr_write(v[231], v[26]))
+  k.emit(v_accvgpr_write(v[235], v[27]))
+  k.emit(v_accvgpr_write(v[239], v[28]))
+  k.emit(v_accvgpr_write(v[243], v[29]))
+  k.emit(v_accvgpr_write(v[247], v[30]))
+  k.emit(v_accvgpr_write(v[251], v[31]))
+  k.emit(s_mov_b64(s[8:9], -1))
+  k.emit(s_or_saveexec_b64(VCC, s[8:9]))
+  k.label('ShiftVectorComponents0_GLVW0')
+  k.emit(v_lshrrev_b32_e32(v[22], 6, v[180]))
+  k.emit(v_lshrrev_b32_e32(v[23], 1, v[22]))
+  k.emit(v_mul_lo_u32(v[23], 16, v[23]))
+  k.emit(v_and_b32_e32(v[19], 63, v[180]))
+  k.emit(v_lshrrev_b32_e32(v[19], 4, v[19]))
+  k.emit(v_lshlrev_b32_e32(v[19], 2, v[19]))
+  k.emit(v_add_lshl_u32_e64(v[19], v[23], v[19], 3))
+  k.emit(v_mul_lo_u32(v[20], v[19], s[38]))
+  k.emit(v_mul_lo_u32(v[21], v[19], s[36]))
+  k.emit(v_and_b32_e32(v[18], 1, v[22]))
+  k.emit(v_mul_lo_u32(v[18], 16, v[18]))
+  k.emit(v_and_b32_e32(v[23], 15, v[180]))
+  k.emit(v_add_lshl_u32_e64(v[18], v[23], v[18], 3))
+  k.emit(s_mul_i32(s[8], 256, s[2]))
+  k.emit(v_add_u32_e32(v[18], s[8], v[18]))
+  k.emit(s_mul_i32(s[8], 256, s[3]))
+  k.emit(v_add_u32_e32(v[19], s[8], v[19]))
+  k.waitcnt(lgkm=0)
+  k.emit(s_add_u32(s[8], s[4], 1))
+  k.emit(s_mul_i32(s[8], s[73], s[8]))
+  k.emit(s_cmp_eq_u32(s[8], 0))
+  k.emit(s_cselect_b32(s[8], s[20], s[8]))
+  k.emit(s_mov_b32(s[91], 131072))
+  k.emit(s_mov_b32(s[90], 0))
+  k.emit(s_mul_i32(s[8], 256, s[2]))
+  k.emit(v_add_u32_e32(v[26], s[8], v[180]))
+  k.emit(s_mul_i32(s[90], 4, s[90]))
+  k.emit(s_mul_i32(s[8], s[73], s[4]))
+  k.emit(v_add_u32_e32(v[24], s[8], v[26]))
+  k.emit(v_lshlrev_b32_e32(v[24], 2, v[24]))
+  k.emit(s_mul_i32(s[8], 256, s[3]))
+  k.emit(v_add_u32_e32(v[26], s[8], v[180]))
+  k.emit(buffer_load_dword(v[22], v[24], s[88:91], 0, 0, 1))
+  k.emit(v_lshlrev_b32_e32(v[26], 2, v[180]))
+  k.emit(s_barrier())
+  k.waitcnt(vm=0)
+  k.emit(ds_write_b32(v[0], v[26], v[22]))
+  k.emit(v_mov_b32_e32(v[23], 1.0))
+  k.emit(ds_write_b32(v[0], v[26], v[23], v[0], 0, 0, 4))
+  k.emit(s_mul_i32(s[8], 256, s[2]))
+  k.emit(v_add_u32_e32(v[26], s[8], v[180]))
+  k.emit(s_mul_i32(s[90], 2, s[90]))
+  k.emit(s_mul_i32(s[8], s[73], s[4]))
+  k.emit(v_add_u32_e32(v[24], s[8], v[26]))
+  k.emit(v_lshlrev_b32_e32(v[24], 1, v[24]))
+  k.emit(s_mul_i32(s[8], 256, s[3]))
+  k.emit(v_add_u32_e32(v[26], s[8], v[180]))
+  k.emit(buffer_load_short_d16(v[22], v[24], s[88:91], 0, 0, 1))
+  k.emit(v_lshlrev_b32_e32(v[26], 2, v[180]))
+  k.emit(s_barrier())
+  k.waitcnt(vm=0)
+  k.emit(v_cvt(v[22], SDWA, v[22], 0, 0, 0, 0, 0, 0, 6, 2, 4))
+  k.emit(ds_write_b32(v[0], v[26], v[22]))
+  k.emit(v_mov_b32_e32(v[23], 1.0))
+  k.emit(ds_write_b32(v[0], v[26], v[23], v[0], 0, 0, 4))
+  k.emit(s_and_b32(s[78], 255, s[20]))
+  k.emit(s_add_u32(s[79], -1, s[10]))
+  k.emit(s_cmp_ge_u32(s[2], s[79]))
+  k.emit(s_cselect_b32(s[78], s[78], 0))
+  k.emit(s_cmpk_gt_u32(s[78]))
+  k.emit(s_cbranch_scc1(), target='GW_B0_E1_M_1')
+  k.emit(s_and_b32(s[78], 255, s[21]))
+  k.emit(s_add_u32(s[79], -1, s[11]))
+  k.emit(s_cmp_ge_u32(s[3], s[79]))
+  k.emit(s_cselect_b32(s[78], s[78], 0))
+  k.emit(s_cmpk_gt_u32(s[78]))
+  k.emit(s_cbranch_scc0(), target='GW_B0_E0_1')
+  k.emit(s_cbranch_scc1(), target='GW_B0_E1_N_1')
+  k.label('GW_B0_E0_1')
+  k.emit(s_mul_i32(s[68], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[37], v[18], s[68]))
+  k.emit(v_lshlrev_b32_e32(v[37], 2, v[37]))
+  k.waitcnt(lgkm=0)
+  k.emit(s_barrier())
+  k.emit(ds_read_b128(v[88:91], v[37]))
+  k.emit(ds_read_b128(v[92:95], v[37], v[0], v[0], 0, 16))
+  k.emit(ds_read_b128(v[96:99], v[37], v[0], v[0], 0, 0, 4))
+  k.emit(ds_read_b128(v[100:103], v[37], v[0], v[0], 0, 16, 4))
+  k.emit(v_add_lshl_u32_e64(v[35], v[21], v[18], 1))
+  k.emit(v_accvgpr_read(v[40], v[0]))
+  k.emit(v_accvgpr_read(v[41], v[4]))
+  k.emit(v_accvgpr_read(v[42], v[8]))
+  k.emit(v_accvgpr_read(v[43], v[12]))
+  k.emit(v_accvgpr_read(v[44], v[16]))
+  k.emit(v_accvgpr_read(v[45], v[20]))
+  k.emit(v_accvgpr_read(v[46], v[24]))
+  k.emit(v_accvgpr_read(v[47], v[28]))
+  k.emit(v_accvgpr_read(v[48], v[32]))
+  k.emit(v_accvgpr_read(v[49], v[36]))
+  k.emit(v_accvgpr_read(v[50], v[40]))
+  k.emit(v_accvgpr_read(v[51], v[44]))
+  k.emit(v_accvgpr_read(v[52], v[48]))
+  k.emit(v_accvgpr_read(v[53], v[52]))
+  k.emit(v_accvgpr_read(v[54], v[56]))
+  k.emit(v_accvgpr_read(v[55], v[60]))
+  k.emit(v_accvgpr_read(v[56], v[64]))
+  k.emit(v_accvgpr_read(v[57], v[68]))
+  k.emit(v_accvgpr_read(v[58], v[72]))
+  k.emit(v_accvgpr_read(v[59], v[76]))
+  k.emit(v_accvgpr_read(v[60], v[80]))
+  k.emit(v_accvgpr_read(v[61], v[84]))
+  k.emit(v_accvgpr_read(v[62], v[88]))
+  k.emit(v_accvgpr_read(v[63], v[92]))
+  k.emit(v_accvgpr_read(v[64], v[96]))
+  k.emit(v_accvgpr_read(v[65], v[100]))
+  k.emit(v_accvgpr_read(v[66], v[104]))
+  k.emit(v_accvgpr_read(v[67], v[108]))
+  k.emit(v_accvgpr_read(v[68], v[112]))
+  k.emit(v_accvgpr_read(v[69], v[116]))
+  k.emit(v_accvgpr_read(v[70], v[120]))
+  k.emit(v_accvgpr_read(v[71], v[124]))
+  k.emit(v_accvgpr_read(v[72], v[128]))
+  k.emit(v_accvgpr_read(v[73], v[132]))
+  k.emit(v_accvgpr_read(v[74], v[136]))
+  k.emit(v_accvgpr_read(v[75], v[140]))
+  k.emit(v_accvgpr_read(v[76], v[144]))
+  k.emit(v_accvgpr_read(v[77], v[148]))
+  k.emit(v_accvgpr_read(v[78], v[152]))
+  k.emit(v_accvgpr_read(v[79], v[156]))
+  k.emit(v_accvgpr_read(v[80], v[160]))
+  k.emit(v_accvgpr_read(v[81], v[164]))
+  k.emit(v_accvgpr_read(v[82], v[168]))
+  k.emit(v_accvgpr_read(v[83], v[172]))
+  k.emit(v_accvgpr_read(v[84], v[176]))
+  k.emit(v_accvgpr_read(v[85], v[180]))
+  k.emit(v_accvgpr_read(v[86], v[184]))
+  k.emit(v_accvgpr_read(v[87], v[188]))
+  k.emit(v_mov_b32_e32(v[32], 4294901760))
+  k.emit(v_mov_b32_e32(v[33], 2147418112))
+  k.emit(v_mov_b32_e32(v[34], 32767))
+  k.waitcnt(lgkm=0)
+  k.emit(v_pk_mul_f32(v[40:41], v[96:97], v[40:41]))
+  k.emit(v_pk_mul_f32(v[42:43], v[98:99], v[42:43]))
+  k.emit(v_pk_mul_f32(v[44:45], v[100:101], v[44:45]))
+  k.emit(v_pk_mul_f32(v[46:47], v[102:103], v[46:47]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[40:41]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[42:43]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[44:45]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[46:47]))
+  k.emit(v_mov_b64_e32(v[40:41], v[22:23]))
+  k.emit(v_mov_b64_e32(v[42:43], v[24:25]))
+  k.emit(v_mov_b64_e32(v[44:45], v[26:27]))
+  k.emit(v_mov_b64_e32(v[46:47], v[28:29]))
+  k.emit(v_cvt_pk(v[40], v[40], v[41]))
+  k.emit(v_cvt_pk(v[41], v[42], v[43]))
+  k.emit(v_cvt_pk(v[42], v[44], v[45]))
+  k.emit(v_cvt_pk(v[43], v[46], v[47]))
+  k.emit(buffer_store_dwordx4(v[40:43], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[48:49], v[96:97], v[48:49]))
+  k.emit(v_pk_mul_f32(v[50:51], v[98:99], v[50:51]))
+  k.emit(v_pk_mul_f32(v[52:53], v[100:101], v[52:53]))
+  k.emit(v_pk_mul_f32(v[54:55], v[102:103], v[54:55]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[48:49]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[50:51]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[52:53]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[54:55]))
+  k.emit(v_mov_b64_e32(v[48:49], v[22:23]))
+  k.emit(v_mov_b64_e32(v[50:51], v[24:25]))
+  k.emit(v_mov_b64_e32(v[52:53], v[26:27]))
+  k.emit(v_mov_b64_e32(v[54:55], v[28:29]))
+  k.emit(v_cvt_pk(v[48], v[48], v[49]))
+  k.emit(v_cvt_pk(v[49], v[50], v[51]))
+  k.emit(v_cvt_pk(v[50], v[52], v[53]))
+  k.emit(v_cvt_pk(v[51], v[54], v[55]))
+  k.emit(s_lshl_b32(s[68], s[36], 1))
+  k.emit(s_add_u32(s[12], s[12], s[68]))
+  k.emit(s_addc_u32(s[13], s[13], 0))
+  k.emit(buffer_store_dwordx4(v[48:51], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[56:57], v[96:97], v[56:57]))
+  k.emit(v_pk_mul_f32(v[58:59], v[98:99], v[58:59]))
+  k.emit(v_pk_mul_f32(v[60:61], v[100:101], v[60:61]))
+  k.emit(v_pk_mul_f32(v[62:63], v[102:103], v[62:63]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[56:57]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[58:59]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[60:61]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[62:63]))
+  k.emit(v_mov_b64_e32(v[56:57], v[22:23]))
+  k.emit(v_mov_b64_e32(v[58:59], v[24:25]))
+  k.emit(v_mov_b64_e32(v[60:61], v[26:27]))
+  k.emit(v_mov_b64_e32(v[62:63], v[28:29]))
+  k.emit(v_cvt_pk(v[56], v[56], v[57]))
+  k.emit(v_cvt_pk(v[57], v[58], v[59]))
+  k.emit(v_cvt_pk(v[58], v[60], v[61]))
+  k.emit(v_cvt_pk(v[59], v[62], v[63]))
+  k.emit(s_lshl_b32(s[68], s[36], 1))
+  k.emit(s_add_u32(s[12], s[12], s[68]))
+  k.emit(s_addc_u32(s[13], s[13], 0))
+  k.emit(buffer_store_dwordx4(v[56:59], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[64:65], v[96:97], v[64:65]))
+  k.emit(v_pk_mul_f32(v[66:67], v[98:99], v[66:67]))
+  k.emit(v_pk_mul_f32(v[68:69], v[100:101], v[68:69]))
+  k.emit(v_pk_mul_f32(v[70:71], v[102:103], v[70:71]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[64:65]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[66:67]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[68:69]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[70:71]))
+  k.emit(v_mov_b64_e32(v[64:65], v[22:23]))
+  k.emit(v_mov_b64_e32(v[66:67], v[24:25]))
+  k.emit(v_mov_b64_e32(v[68:69], v[26:27]))
+  k.emit(v_mov_b64_e32(v[70:71], v[28:29]))
+  k.emit(v_cvt_pk(v[64], v[64], v[65]))
+  k.emit(v_cvt_pk(v[65], v[66], v[67]))
+  k.emit(v_cvt_pk(v[66], v[68], v[69]))
+  k.emit(v_cvt_pk(v[67], v[70], v[71]))
+  k.emit(s_lshl_b32(s[68], s[36], 1))
+  k.emit(s_add_u32(s[12], s[12], s[68]))
+  k.emit(s_addc_u32(s[13], s[13], 0))
+  k.emit(buffer_store_dwordx4(v[64:67], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[72:73], v[96:97], v[72:73]))
+  k.emit(v_pk_mul_f32(v[74:75], v[98:99], v[74:75]))
+  k.emit(v_pk_mul_f32(v[76:77], v[100:101], v[76:77]))
+  k.emit(v_pk_mul_f32(v[78:79], v[102:103], v[78:79]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[72:73]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[74:75]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[76:77]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[78:79]))
+  k.emit(v_mov_b64_e32(v[72:73], v[22:23]))
+  k.emit(v_mov_b64_e32(v[74:75], v[24:25]))
+  k.emit(v_mov_b64_e32(v[76:77], v[26:27]))
+  k.emit(v_mov_b64_e32(v[78:79], v[28:29]))
+  k.emit(v_cvt_pk(v[72], v[72], v[73]))
+  k.emit(v_cvt_pk(v[73], v[74], v[75]))
+  k.emit(v_cvt_pk(v[74], v[76], v[77]))
+  k.emit(v_cvt_pk(v[75], v[78], v[79]))
+  k.emit(s_lshl_b32(s[68], s[36], 1))
+  k.emit(s_add_u32(s[12], s[12], s[68]))
+  k.emit(s_addc_u32(s[13], s[13], 0))
+  k.emit(buffer_store_dwordx4(v[72:75], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[80:81], v[96:97], v[80:81]))
+  k.emit(v_pk_mul_f32(v[82:83], v[98:99], v[82:83]))
+  k.emit(v_pk_mul_f32(v[84:85], v[100:101], v[84:85]))
+  k.emit(v_pk_mul_f32(v[86:87], v[102:103], v[86:87]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[80:81]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[82:83]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[84:85]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[86:87]))
+  k.emit(v_mov_b64_e32(v[80:81], v[22:23]))
+  k.emit(v_mov_b64_e32(v[82:83], v[24:25]))
+  k.emit(v_mov_b64_e32(v[84:85], v[26:27]))
+  k.emit(v_mov_b64_e32(v[86:87], v[28:29]))
+  k.emit(v_cvt_pk(v[80], v[80], v[81]))
+  k.emit(v_cvt_pk(v[81], v[82], v[83]))
+  k.emit(v_cvt_pk(v[82], v[84], v[85]))
+  k.emit(v_cvt_pk(v[83], v[86], v[87]))
+  k.emit(s_lshl_b32(s[68], s[36], 1))
+  k.emit(s_add_u32(s[12], s[12], s[68]))
+  k.emit(s_addc_u32(s[13], s[13], 0))
+  k.emit(buffer_store_dwordx4(v[80:83], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_nop())
+  k.emit(ds_read_b128(v[88:91], v[37]))
+  k.emit(ds_read_b128(v[92:95], v[37], v[0], v[0], 0, 16))
+  k.emit(ds_read_b128(v[96:99], v[37], v[0], v[0], 0, 0, 4))
+  k.emit(ds_read_b128(v[100:103], v[37], v[0], v[0], 0, 16, 4))
+  k.emit(v_accvgpr_read(v[40], v[192]))
+  k.emit(v_accvgpr_read(v[41], v[196]))
+  k.emit(v_accvgpr_read(v[42], v[200]))
+  k.emit(v_accvgpr_read(v[43], v[204]))
+  k.emit(v_accvgpr_read(v[44], v[208]))
+  k.emit(v_accvgpr_read(v[45], v[212]))
+  k.emit(v_accvgpr_read(v[46], v[216]))
+  k.emit(v_accvgpr_read(v[47], v[220]))
+  k.emit(v_accvgpr_read(v[48], v[224]))
+  k.emit(v_accvgpr_read(v[49], v[228]))
+  k.emit(v_accvgpr_read(v[50], v[232]))
+  k.emit(v_accvgpr_read(v[51], v[236]))
+  k.emit(v_accvgpr_read(v[52], v[240]))
+  k.emit(v_accvgpr_read(v[53], v[244]))
+  k.emit(v_accvgpr_read(v[54], v[248]))
+  k.emit(v_accvgpr_read(v[55], v[252]))
+  k.emit(v_accvgpr_read(v[56], v[1]))
+  k.emit(v_accvgpr_read(v[57], v[5]))
+  k.emit(v_accvgpr_read(v[58], v[9]))
+  k.emit(v_accvgpr_read(v[59], v[13]))
+  k.emit(v_accvgpr_read(v[60], v[17]))
+  k.emit(v_accvgpr_read(v[61], v[21]))
+  k.emit(v_accvgpr_read(v[62], v[25]))
+  k.emit(v_accvgpr_read(v[63], v[29]))
+  k.emit(v_accvgpr_read(v[64], v[33]))
+  k.emit(v_accvgpr_read(v[65], v[37]))
+  k.emit(v_accvgpr_read(v[66], v[41]))
+  k.emit(v_accvgpr_read(v[67], v[45]))
+  k.emit(v_accvgpr_read(v[68], v[49]))
+  k.emit(v_accvgpr_read(v[69], v[53]))
+  k.emit(v_accvgpr_read(v[70], v[57]))
+  k.emit(v_accvgpr_read(v[71], v[61]))
+  k.emit(v_accvgpr_read(v[72], v[65]))
+  k.emit(v_accvgpr_read(v[73], v[69]))
+  k.emit(v_accvgpr_read(v[74], v[73]))
+  k.emit(v_accvgpr_read(v[75], v[77]))
+  k.emit(v_accvgpr_read(v[76], v[81]))
+  k.emit(v_accvgpr_read(v[77], v[85]))
+  k.emit(v_accvgpr_read(v[78], v[89]))
+  k.emit(v_accvgpr_read(v[79], v[93]))
+  k.emit(v_accvgpr_read(v[80], v[97]))
+  k.emit(v_accvgpr_read(v[81], v[101]))
+  k.emit(v_accvgpr_read(v[82], v[105]))
+  k.emit(v_accvgpr_read(v[83], v[109]))
+  k.emit(v_accvgpr_read(v[84], v[113]))
+  k.emit(v_accvgpr_read(v[85], v[117]))
+  k.emit(v_accvgpr_read(v[86], v[121]))
+  k.emit(v_accvgpr_read(v[87], v[125]))
+  k.emit(v_mov_b32_e32(v[32], 4294901760))
+  k.emit(v_mov_b32_e32(v[33], 2147418112))
+  k.emit(v_mov_b32_e32(v[34], 32767))
+  k.waitcnt(lgkm=0)
+  k.emit(v_pk_mul_f32(v[40:41], v[96:97], v[40:41]))
+  k.emit(v_pk_mul_f32(v[42:43], v[98:99], v[42:43]))
+  k.emit(v_pk_mul_f32(v[44:45], v[100:101], v[44:45]))
+  k.emit(v_pk_mul_f32(v[46:47], v[102:103], v[46:47]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[40:41]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[42:43]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[44:45]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[46:47]))
+  k.emit(v_mov_b64_e32(v[40:41], v[22:23]))
+  k.emit(v_mov_b64_e32(v[42:43], v[24:25]))
+  k.emit(v_mov_b64_e32(v[44:45], v[26:27]))
+  k.emit(v_mov_b64_e32(v[46:47], v[28:29]))
+  k.emit(v_cvt_pk(v[40], v[40], v[41]))
+  k.emit(v_cvt_pk(v[41], v[42], v[43]))
+  k.emit(v_cvt_pk(v[42], v[44], v[45]))
+  k.emit(v_cvt_pk(v[43], v[46], v[47]))
+  k.emit(s_lshl_b32(s[68], s[36], 1))
+  k.emit(s_add_u32(s[12], s[12], s[68]))
+  k.emit(s_addc_u32(s[13], s[13], 0))
+  k.emit(buffer_store_dwordx4(v[40:43], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[48:49], v[96:97], v[48:49]))
+  k.emit(v_pk_mul_f32(v[50:51], v[98:99], v[50:51]))
+  k.emit(v_pk_mul_f32(v[52:53], v[100:101], v[52:53]))
+  k.emit(v_pk_mul_f32(v[54:55], v[102:103], v[54:55]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[48:49]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[50:51]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[52:53]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[54:55]))
+  k.emit(v_mov_b64_e32(v[48:49], v[22:23]))
+  k.emit(v_mov_b64_e32(v[50:51], v[24:25]))
+  k.emit(v_mov_b64_e32(v[52:53], v[26:27]))
+  k.emit(v_mov_b64_e32(v[54:55], v[28:29]))
+  k.emit(v_cvt_pk(v[48], v[48], v[49]))
+  k.emit(v_cvt_pk(v[49], v[50], v[51]))
+  k.emit(v_cvt_pk(v[50], v[52], v[53]))
+  k.emit(v_cvt_pk(v[51], v[54], v[55]))
+  k.emit(s_lshl_b32(s[68], s[36], 1))
+  k.emit(s_add_u32(s[12], s[12], s[68]))
+  k.emit(s_addc_u32(s[13], s[13], 0))
+  k.emit(buffer_store_dwordx4(v[48:51], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[56:57], v[96:97], v[56:57]))
+  k.emit(v_pk_mul_f32(v[58:59], v[98:99], v[58:59]))
+  k.emit(v_pk_mul_f32(v[60:61], v[100:101], v[60:61]))
+  k.emit(v_pk_mul_f32(v[62:63], v[102:103], v[62:63]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[56:57]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[58:59]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[60:61]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[62:63]))
+  k.emit(v_mov_b64_e32(v[56:57], v[22:23]))
+  k.emit(v_mov_b64_e32(v[58:59], v[24:25]))
+  k.emit(v_mov_b64_e32(v[60:61], v[26:27]))
+  k.emit(v_mov_b64_e32(v[62:63], v[28:29]))
+  k.emit(v_cvt_pk(v[56], v[56], v[57]))
+  k.emit(v_cvt_pk(v[57], v[58], v[59]))
+  k.emit(v_cvt_pk(v[58], v[60], v[61]))
+  k.emit(v_cvt_pk(v[59], v[62], v[63]))
+  k.emit(s_lshl_b32(s[68], s[36], 1))
+  k.emit(s_add_u32(s[12], s[12], s[68]))
+  k.emit(s_addc_u32(s[13], s[13], 0))
+  k.emit(buffer_store_dwordx4(v[56:59], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[64:65], v[96:97], v[64:65]))
+  k.emit(v_pk_mul_f32(v[66:67], v[98:99], v[66:67]))
+  k.emit(v_pk_mul_f32(v[68:69], v[100:101], v[68:69]))
+  k.emit(v_pk_mul_f32(v[70:71], v[102:103], v[70:71]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[64:65]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[66:67]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[68:69]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[70:71]))
+  k.emit(v_mov_b64_e32(v[64:65], v[22:23]))
+  k.emit(v_mov_b64_e32(v[66:67], v[24:25]))
+  k.emit(v_mov_b64_e32(v[68:69], v[26:27]))
+  k.emit(v_mov_b64_e32(v[70:71], v[28:29]))
+  k.emit(v_cvt_pk(v[64], v[64], v[65]))
+  k.emit(v_cvt_pk(v[65], v[66], v[67]))
+  k.emit(v_cvt_pk(v[66], v[68], v[69]))
+  k.emit(v_cvt_pk(v[67], v[70], v[71]))
+  k.emit(s_lshl_b32(s[68], s[36], 1))
+  k.emit(s_add_u32(s[12], s[12], s[68]))
+  k.emit(s_addc_u32(s[13], s[13], 0))
+  k.emit(buffer_store_dwordx4(v[64:67], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[72:73], v[96:97], v[72:73]))
+  k.emit(v_pk_mul_f32(v[74:75], v[98:99], v[74:75]))
+  k.emit(v_pk_mul_f32(v[76:77], v[100:101], v[76:77]))
+  k.emit(v_pk_mul_f32(v[78:79], v[102:103], v[78:79]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[72:73]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[74:75]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[76:77]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[78:79]))
+  k.emit(v_mov_b64_e32(v[72:73], v[22:23]))
+  k.emit(v_mov_b64_e32(v[74:75], v[24:25]))
+  k.emit(v_mov_b64_e32(v[76:77], v[26:27]))
+  k.emit(v_mov_b64_e32(v[78:79], v[28:29]))
+  k.emit(v_cvt_pk(v[72], v[72], v[73]))
+  k.emit(v_cvt_pk(v[73], v[74], v[75]))
+  k.emit(v_cvt_pk(v[74], v[76], v[77]))
+  k.emit(v_cvt_pk(v[75], v[78], v[79]))
+  k.emit(s_lshl_b32(s[68], s[36], 1))
+  k.emit(s_add_u32(s[12], s[12], s[68]))
+  k.emit(s_addc_u32(s[13], s[13], 0))
+  k.emit(buffer_store_dwordx4(v[72:75], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[80:81], v[96:97], v[80:81]))
+  k.emit(v_pk_mul_f32(v[82:83], v[98:99], v[82:83]))
+  k.emit(v_pk_mul_f32(v[84:85], v[100:101], v[84:85]))
+  k.emit(v_pk_mul_f32(v[86:87], v[102:103], v[86:87]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[80:81]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[82:83]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[84:85]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[86:87]))
+  k.emit(v_mov_b64_e32(v[80:81], v[22:23]))
+  k.emit(v_mov_b64_e32(v[82:83], v[24:25]))
+  k.emit(v_mov_b64_e32(v[84:85], v[26:27]))
+  k.emit(v_mov_b64_e32(v[86:87], v[28:29]))
+  k.emit(v_cvt_pk(v[80], v[80], v[81]))
+  k.emit(v_cvt_pk(v[81], v[82], v[83]))
+  k.emit(v_cvt_pk(v[82], v[84], v[85]))
+  k.emit(v_cvt_pk(v[83], v[86], v[87]))
+  k.emit(s_lshl_b32(s[68], s[36], 1))
+  k.emit(s_add_u32(s[12], s[12], s[68]))
+  k.emit(s_addc_u32(s[13], s[13], 0))
+  k.emit(buffer_store_dwordx4(v[80:83], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_nop())
+  k.emit(ds_read_b128(v[88:91], v[37]))
+  k.emit(ds_read_b128(v[92:95], v[37], v[0], v[0], 0, 16))
+  k.emit(ds_read_b128(v[96:99], v[37], v[0], v[0], 0, 0, 4))
+  k.emit(ds_read_b128(v[100:103], v[37], v[0], v[0], 0, 16, 4))
+  k.emit(v_accvgpr_read(v[40], v[129]))
+  k.emit(v_accvgpr_read(v[41], v[133]))
+  k.emit(v_accvgpr_read(v[42], v[137]))
+  k.emit(v_accvgpr_read(v[43], v[141]))
+  k.emit(v_accvgpr_read(v[44], v[145]))
+  k.emit(v_accvgpr_read(v[45], v[149]))
+  k.emit(v_accvgpr_read(v[46], v[153]))
+  k.emit(v_accvgpr_read(v[47], v[157]))
+  k.emit(v_accvgpr_read(v[48], v[161]))
+  k.emit(v_accvgpr_read(v[49], v[165]))
+  k.emit(v_accvgpr_read(v[50], v[169]))
+  k.emit(v_accvgpr_read(v[51], v[173]))
+  k.emit(v_accvgpr_read(v[52], v[177]))
+  k.emit(v_accvgpr_read(v[53], v[181]))
+  k.emit(v_accvgpr_read(v[54], v[185]))
+  k.emit(v_accvgpr_read(v[55], v[189]))
+  k.emit(v_accvgpr_read(v[56], v[193]))
+  k.emit(v_accvgpr_read(v[57], v[197]))
+  k.emit(v_accvgpr_read(v[58], v[201]))
+  k.emit(v_accvgpr_read(v[59], v[205]))
+  k.emit(v_accvgpr_read(v[60], v[209]))
+  k.emit(v_accvgpr_read(v[61], v[213]))
+  k.emit(v_accvgpr_read(v[62], v[217]))
+  k.emit(v_accvgpr_read(v[63], v[221]))
+  k.emit(v_accvgpr_read(v[64], v[225]))
+  k.emit(v_accvgpr_read(v[65], v[229]))
+  k.emit(v_accvgpr_read(v[66], v[233]))
+  k.emit(v_accvgpr_read(v[67], v[237]))
+  k.emit(v_accvgpr_read(v[68], v[241]))
+  k.emit(v_accvgpr_read(v[69], v[245]))
+  k.emit(v_accvgpr_read(v[70], v[249]))
+  k.emit(v_accvgpr_read(v[71], v[253]))
+  k.emit(v_accvgpr_read(v[72], v[2]))
+  k.emit(v_accvgpr_read(v[73], v[6]))
+  k.emit(v_accvgpr_read(v[74], v[10]))
+  k.emit(v_accvgpr_read(v[75], v[14]))
+  k.emit(v_accvgpr_read(v[76], v[18]))
+  k.emit(v_accvgpr_read(v[77], v[22]))
+  k.emit(v_accvgpr_read(v[78], v[26]))
+  k.emit(v_accvgpr_read(v[79], v[30]))
+  k.emit(v_accvgpr_read(v[80], v[34]))
+  k.emit(v_accvgpr_read(v[81], v[38]))
+  k.emit(v_accvgpr_read(v[82], v[42]))
+  k.emit(v_accvgpr_read(v[83], v[46]))
+  k.emit(v_accvgpr_read(v[84], v[50]))
+  k.emit(v_accvgpr_read(v[85], v[54]))
+  k.emit(v_accvgpr_read(v[86], v[58]))
+  k.emit(v_accvgpr_read(v[87], v[62]))
+  k.emit(v_mov_b32_e32(v[32], 4294901760))
+  k.emit(v_mov_b32_e32(v[33], 2147418112))
+  k.emit(v_mov_b32_e32(v[34], 32767))
+  k.waitcnt(lgkm=0)
+  k.emit(v_pk_mul_f32(v[40:41], v[96:97], v[40:41]))
+  k.emit(v_pk_mul_f32(v[42:43], v[98:99], v[42:43]))
+  k.emit(v_pk_mul_f32(v[44:45], v[100:101], v[44:45]))
+  k.emit(v_pk_mul_f32(v[46:47], v[102:103], v[46:47]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[40:41]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[42:43]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[44:45]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[46:47]))
+  k.emit(v_mov_b64_e32(v[40:41], v[22:23]))
+  k.emit(v_mov_b64_e32(v[42:43], v[24:25]))
+  k.emit(v_mov_b64_e32(v[44:45], v[26:27]))
+  k.emit(v_mov_b64_e32(v[46:47], v[28:29]))
+  k.emit(v_cvt_pk(v[40], v[40], v[41]))
+  k.emit(v_cvt_pk(v[41], v[42], v[43]))
+  k.emit(v_cvt_pk(v[42], v[44], v[45]))
+  k.emit(v_cvt_pk(v[43], v[46], v[47]))
+  k.emit(s_lshl_b32(s[68], s[36], 1))
+  k.emit(s_add_u32(s[12], s[12], s[68]))
+  k.emit(s_addc_u32(s[13], s[13], 0))
+  k.emit(buffer_store_dwordx4(v[40:43], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[48:49], v[96:97], v[48:49]))
+  k.emit(v_pk_mul_f32(v[50:51], v[98:99], v[50:51]))
+  k.emit(v_pk_mul_f32(v[52:53], v[100:101], v[52:53]))
+  k.emit(v_pk_mul_f32(v[54:55], v[102:103], v[54:55]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[48:49]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[50:51]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[52:53]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[54:55]))
+  k.emit(v_mov_b64_e32(v[48:49], v[22:23]))
+  k.emit(v_mov_b64_e32(v[50:51], v[24:25]))
+  k.emit(v_mov_b64_e32(v[52:53], v[26:27]))
+  k.emit(v_mov_b64_e32(v[54:55], v[28:29]))
+  k.emit(v_cvt_pk(v[48], v[48], v[49]))
+  k.emit(v_cvt_pk(v[49], v[50], v[51]))
+  k.emit(v_cvt_pk(v[50], v[52], v[53]))
+  k.emit(v_cvt_pk(v[51], v[54], v[55]))
+  k.emit(s_lshl_b32(s[68], s[36], 1))
+  k.emit(s_add_u32(s[12], s[12], s[68]))
+  k.emit(s_addc_u32(s[13], s[13], 0))
+  k.emit(buffer_store_dwordx4(v[48:51], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[56:57], v[96:97], v[56:57]))
+  k.emit(v_pk_mul_f32(v[58:59], v[98:99], v[58:59]))
+  k.emit(v_pk_mul_f32(v[60:61], v[100:101], v[60:61]))
+  k.emit(v_pk_mul_f32(v[62:63], v[102:103], v[62:63]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[56:57]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[58:59]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[60:61]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[62:63]))
+  k.emit(v_mov_b64_e32(v[56:57], v[22:23]))
+  k.emit(v_mov_b64_e32(v[58:59], v[24:25]))
+  k.emit(v_mov_b64_e32(v[60:61], v[26:27]))
+  k.emit(v_mov_b64_e32(v[62:63], v[28:29]))
+  k.emit(v_cvt_pk(v[56], v[56], v[57]))
+  k.emit(v_cvt_pk(v[57], v[58], v[59]))
+  k.emit(v_cvt_pk(v[58], v[60], v[61]))
+  k.emit(v_cvt_pk(v[59], v[62], v[63]))
+  k.emit(s_lshl_b32(s[68], s[36], 1))
+  k.emit(s_add_u32(s[12], s[12], s[68]))
+  k.emit(s_addc_u32(s[13], s[13], 0))
+  k.emit(buffer_store_dwordx4(v[56:59], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[64:65], v[96:97], v[64:65]))
+  k.emit(v_pk_mul_f32(v[66:67], v[98:99], v[66:67]))
+  k.emit(v_pk_mul_f32(v[68:69], v[100:101], v[68:69]))
+  k.emit(v_pk_mul_f32(v[70:71], v[102:103], v[70:71]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[64:65]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[66:67]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[68:69]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[70:71]))
+  k.emit(v_mov_b64_e32(v[64:65], v[22:23]))
+  k.emit(v_mov_b64_e32(v[66:67], v[24:25]))
+  k.emit(v_mov_b64_e32(v[68:69], v[26:27]))
+  k.emit(v_mov_b64_e32(v[70:71], v[28:29]))
+  k.emit(v_cvt_pk(v[64], v[64], v[65]))
+  k.emit(v_cvt_pk(v[65], v[66], v[67]))
+  k.emit(v_cvt_pk(v[66], v[68], v[69]))
+  k.emit(v_cvt_pk(v[67], v[70], v[71]))
+  k.emit(s_lshl_b32(s[68], s[36], 1))
+  k.emit(s_add_u32(s[12], s[12], s[68]))
+  k.emit(s_addc_u32(s[13], s[13], 0))
+  k.emit(buffer_store_dwordx4(v[64:67], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[72:73], v[96:97], v[72:73]))
+  k.emit(v_pk_mul_f32(v[74:75], v[98:99], v[74:75]))
+  k.emit(v_pk_mul_f32(v[76:77], v[100:101], v[76:77]))
+  k.emit(v_pk_mul_f32(v[78:79], v[102:103], v[78:79]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[72:73]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[74:75]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[76:77]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[78:79]))
+  k.emit(v_mov_b64_e32(v[72:73], v[22:23]))
+  k.emit(v_mov_b64_e32(v[74:75], v[24:25]))
+  k.emit(v_mov_b64_e32(v[76:77], v[26:27]))
+  k.emit(v_mov_b64_e32(v[78:79], v[28:29]))
+  k.emit(v_cvt_pk(v[72], v[72], v[73]))
+  k.emit(v_cvt_pk(v[73], v[74], v[75]))
+  k.emit(v_cvt_pk(v[74], v[76], v[77]))
+  k.emit(v_cvt_pk(v[75], v[78], v[79]))
+  k.emit(s_lshl_b32(s[68], s[36], 1))
+  k.emit(s_add_u32(s[12], s[12], s[68]))
+  k.emit(s_addc_u32(s[13], s[13], 0))
+  k.emit(buffer_store_dwordx4(v[72:75], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[80:81], v[96:97], v[80:81]))
+  k.emit(v_pk_mul_f32(v[82:83], v[98:99], v[82:83]))
+  k.emit(v_pk_mul_f32(v[84:85], v[100:101], v[84:85]))
+  k.emit(v_pk_mul_f32(v[86:87], v[102:103], v[86:87]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[80:81]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[82:83]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[84:85]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[86:87]))
+  k.emit(v_mov_b64_e32(v[80:81], v[22:23]))
+  k.emit(v_mov_b64_e32(v[82:83], v[24:25]))
+  k.emit(v_mov_b64_e32(v[84:85], v[26:27]))
+  k.emit(v_mov_b64_e32(v[86:87], v[28:29]))
+  k.emit(v_cvt_pk(v[80], v[80], v[81]))
+  k.emit(v_cvt_pk(v[81], v[82], v[83]))
+  k.emit(v_cvt_pk(v[82], v[84], v[85]))
+  k.emit(v_cvt_pk(v[83], v[86], v[87]))
+  k.emit(s_lshl_b32(s[68], s[36], 1))
+  k.emit(s_add_u32(s[12], s[12], s[68]))
+  k.emit(s_addc_u32(s[13], s[13], 0))
+  k.emit(buffer_store_dwordx4(v[80:83], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_nop())
+  k.emit(ds_read_b128(v[88:91], v[37]))
+  k.emit(ds_read_b128(v[92:95], v[37], v[0], v[0], 0, 16))
+  k.emit(ds_read_b128(v[96:99], v[37], v[0], v[0], 0, 0, 4))
+  k.emit(ds_read_b128(v[100:103], v[37], v[0], v[0], 0, 16, 4))
+  k.emit(v_accvgpr_read(v[40], v[66]))
+  k.emit(v_accvgpr_read(v[41], v[70]))
+  k.emit(v_accvgpr_read(v[42], v[74]))
+  k.emit(v_accvgpr_read(v[43], v[78]))
+  k.emit(v_accvgpr_read(v[44], v[82]))
+  k.emit(v_accvgpr_read(v[45], v[86]))
+  k.emit(v_accvgpr_read(v[46], v[90]))
+  k.emit(v_accvgpr_read(v[47], v[94]))
+  k.emit(v_accvgpr_read(v[48], v[98]))
+  k.emit(v_accvgpr_read(v[49], v[102]))
+  k.emit(v_accvgpr_read(v[50], v[106]))
+  k.emit(v_accvgpr_read(v[51], v[110]))
+  k.emit(v_accvgpr_read(v[52], v[114]))
+  k.emit(v_accvgpr_read(v[53], v[118]))
+  k.emit(v_accvgpr_read(v[54], v[122]))
+  k.emit(v_accvgpr_read(v[55], v[126]))
+  k.emit(v_accvgpr_read(v[56], v[130]))
+  k.emit(v_accvgpr_read(v[57], v[134]))
+  k.emit(v_accvgpr_read(v[58], v[138]))
+  k.emit(v_accvgpr_read(v[59], v[142]))
+  k.emit(v_accvgpr_read(v[60], v[146]))
+  k.emit(v_accvgpr_read(v[61], v[150]))
+  k.emit(v_accvgpr_read(v[62], v[154]))
+  k.emit(v_accvgpr_read(v[63], v[158]))
+  k.emit(v_accvgpr_read(v[64], v[162]))
+  k.emit(v_accvgpr_read(v[65], v[166]))
+  k.emit(v_accvgpr_read(v[66], v[170]))
+  k.emit(v_accvgpr_read(v[67], v[174]))
+  k.emit(v_accvgpr_read(v[68], v[178]))
+  k.emit(v_accvgpr_read(v[69], v[182]))
+  k.emit(v_accvgpr_read(v[70], v[186]))
+  k.emit(v_accvgpr_read(v[71], v[190]))
+  k.emit(v_accvgpr_read(v[72], v[194]))
+  k.emit(v_accvgpr_read(v[73], v[198]))
+  k.emit(v_accvgpr_read(v[74], v[202]))
+  k.emit(v_accvgpr_read(v[75], v[206]))
+  k.emit(v_accvgpr_read(v[76], v[210]))
+  k.emit(v_accvgpr_read(v[77], v[214]))
+  k.emit(v_accvgpr_read(v[78], v[218]))
+  k.emit(v_accvgpr_read(v[79], v[222]))
+  k.emit(v_accvgpr_read(v[80], v[226]))
+  k.emit(v_accvgpr_read(v[81], v[230]))
+  k.emit(v_accvgpr_read(v[82], v[234]))
+  k.emit(v_accvgpr_read(v[83], v[238]))
+  k.emit(v_accvgpr_read(v[84], v[242]))
+  k.emit(v_accvgpr_read(v[85], v[246]))
+  k.emit(v_accvgpr_read(v[86], v[250]))
+  k.emit(v_accvgpr_read(v[87], v[254]))
+  k.emit(v_mov_b32_e32(v[32], 4294901760))
+  k.emit(v_mov_b32_e32(v[33], 2147418112))
+  k.emit(v_mov_b32_e32(v[34], 32767))
+  k.waitcnt(lgkm=0)
+  k.emit(v_pk_mul_f32(v[40:41], v[96:97], v[40:41]))
+  k.emit(v_pk_mul_f32(v[42:43], v[98:99], v[42:43]))
+  k.emit(v_pk_mul_f32(v[44:45], v[100:101], v[44:45]))
+  k.emit(v_pk_mul_f32(v[46:47], v[102:103], v[46:47]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[40:41]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[42:43]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[44:45]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[46:47]))
+  k.emit(v_mov_b64_e32(v[40:41], v[22:23]))
+  k.emit(v_mov_b64_e32(v[42:43], v[24:25]))
+  k.emit(v_mov_b64_e32(v[44:45], v[26:27]))
+  k.emit(v_mov_b64_e32(v[46:47], v[28:29]))
+  k.emit(v_cvt_pk(v[40], v[40], v[41]))
+  k.emit(v_cvt_pk(v[41], v[42], v[43]))
+  k.emit(v_cvt_pk(v[42], v[44], v[45]))
+  k.emit(v_cvt_pk(v[43], v[46], v[47]))
+  k.emit(s_lshl_b32(s[68], s[36], 1))
+  k.emit(s_add_u32(s[12], s[12], s[68]))
+  k.emit(s_addc_u32(s[13], s[13], 0))
+  k.emit(buffer_store_dwordx4(v[40:43], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[48:49], v[96:97], v[48:49]))
+  k.emit(v_pk_mul_f32(v[50:51], v[98:99], v[50:51]))
+  k.emit(v_pk_mul_f32(v[52:53], v[100:101], v[52:53]))
+  k.emit(v_pk_mul_f32(v[54:55], v[102:103], v[54:55]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[48:49]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[50:51]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[52:53]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[54:55]))
+  k.emit(v_mov_b64_e32(v[48:49], v[22:23]))
+  k.emit(v_mov_b64_e32(v[50:51], v[24:25]))
+  k.emit(v_mov_b64_e32(v[52:53], v[26:27]))
+  k.emit(v_mov_b64_e32(v[54:55], v[28:29]))
+  k.emit(v_cvt_pk(v[48], v[48], v[49]))
+  k.emit(v_cvt_pk(v[49], v[50], v[51]))
+  k.emit(v_cvt_pk(v[50], v[52], v[53]))
+  k.emit(v_cvt_pk(v[51], v[54], v[55]))
+  k.emit(s_lshl_b32(s[68], s[36], 1))
+  k.emit(s_add_u32(s[12], s[12], s[68]))
+  k.emit(s_addc_u32(s[13], s[13], 0))
+  k.emit(buffer_store_dwordx4(v[48:51], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[56:57], v[96:97], v[56:57]))
+  k.emit(v_pk_mul_f32(v[58:59], v[98:99], v[58:59]))
+  k.emit(v_pk_mul_f32(v[60:61], v[100:101], v[60:61]))
+  k.emit(v_pk_mul_f32(v[62:63], v[102:103], v[62:63]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[56:57]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[58:59]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[60:61]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[62:63]))
+  k.emit(v_mov_b64_e32(v[56:57], v[22:23]))
+  k.emit(v_mov_b64_e32(v[58:59], v[24:25]))
+  k.emit(v_mov_b64_e32(v[60:61], v[26:27]))
+  k.emit(v_mov_b64_e32(v[62:63], v[28:29]))
+  k.emit(v_cvt_pk(v[56], v[56], v[57]))
+  k.emit(v_cvt_pk(v[57], v[58], v[59]))
+  k.emit(v_cvt_pk(v[58], v[60], v[61]))
+  k.emit(v_cvt_pk(v[59], v[62], v[63]))
+  k.emit(s_lshl_b32(s[68], s[36], 1))
+  k.emit(s_add_u32(s[12], s[12], s[68]))
+  k.emit(s_addc_u32(s[13], s[13], 0))
+  k.emit(buffer_store_dwordx4(v[56:59], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[64:65], v[96:97], v[64:65]))
+  k.emit(v_pk_mul_f32(v[66:67], v[98:99], v[66:67]))
+  k.emit(v_pk_mul_f32(v[68:69], v[100:101], v[68:69]))
+  k.emit(v_pk_mul_f32(v[70:71], v[102:103], v[70:71]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[64:65]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[66:67]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[68:69]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[70:71]))
+  k.emit(v_mov_b64_e32(v[64:65], v[22:23]))
+  k.emit(v_mov_b64_e32(v[66:67], v[24:25]))
+  k.emit(v_mov_b64_e32(v[68:69], v[26:27]))
+  k.emit(v_mov_b64_e32(v[70:71], v[28:29]))
+  k.emit(v_cvt_pk(v[64], v[64], v[65]))
+  k.emit(v_cvt_pk(v[65], v[66], v[67]))
+  k.emit(v_cvt_pk(v[66], v[68], v[69]))
+  k.emit(v_cvt_pk(v[67], v[70], v[71]))
+  k.emit(s_lshl_b32(s[68], s[36], 1))
+  k.emit(s_add_u32(s[12], s[12], s[68]))
+  k.emit(s_addc_u32(s[13], s[13], 0))
+  k.emit(buffer_store_dwordx4(v[64:67], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[72:73], v[96:97], v[72:73]))
+  k.emit(v_pk_mul_f32(v[74:75], v[98:99], v[74:75]))
+  k.emit(v_pk_mul_f32(v[76:77], v[100:101], v[76:77]))
+  k.emit(v_pk_mul_f32(v[78:79], v[102:103], v[78:79]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[72:73]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[74:75]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[76:77]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[78:79]))
+  k.emit(v_mov_b64_e32(v[72:73], v[22:23]))
+  k.emit(v_mov_b64_e32(v[74:75], v[24:25]))
+  k.emit(v_mov_b64_e32(v[76:77], v[26:27]))
+  k.emit(v_mov_b64_e32(v[78:79], v[28:29]))
+  k.emit(v_cvt_pk(v[72], v[72], v[73]))
+  k.emit(v_cvt_pk(v[73], v[74], v[75]))
+  k.emit(v_cvt_pk(v[74], v[76], v[77]))
+  k.emit(v_cvt_pk(v[75], v[78], v[79]))
+  k.emit(s_lshl_b32(s[68], s[36], 1))
+  k.emit(s_add_u32(s[12], s[12], s[68]))
+  k.emit(s_addc_u32(s[13], s[13], 0))
+  k.emit(buffer_store_dwordx4(v[72:75], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[80:81], v[96:97], v[80:81]))
+  k.emit(v_pk_mul_f32(v[82:83], v[98:99], v[82:83]))
+  k.emit(v_pk_mul_f32(v[84:85], v[100:101], v[84:85]))
+  k.emit(v_pk_mul_f32(v[86:87], v[102:103], v[86:87]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[80:81]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[82:83]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[84:85]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[86:87]))
+  k.emit(v_mov_b64_e32(v[80:81], v[22:23]))
+  k.emit(v_mov_b64_e32(v[82:83], v[24:25]))
+  k.emit(v_mov_b64_e32(v[84:85], v[26:27]))
+  k.emit(v_mov_b64_e32(v[86:87], v[28:29]))
+  k.emit(v_cvt_pk(v[80], v[80], v[81]))
+  k.emit(v_cvt_pk(v[81], v[82], v[83]))
+  k.emit(v_cvt_pk(v[82], v[84], v[85]))
+  k.emit(v_cvt_pk(v[83], v[86], v[87]))
+  k.emit(s_lshl_b32(s[68], s[36], 1))
+  k.emit(s_add_u32(s[12], s[12], s[68]))
+  k.emit(s_addc_u32(s[13], s[13], 0))
+  k.emit(buffer_store_dwordx4(v[80:83], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_nop())
+  k.emit(ds_read_b128(v[88:91], v[37]))
+  k.emit(ds_read_b128(v[92:95], v[37], v[0], v[0], 0, 16))
+  k.emit(ds_read_b128(v[96:99], v[37], v[0], v[0], 0, 0, 4))
+  k.emit(ds_read_b128(v[100:103], v[37], v[0], v[0], 0, 16, 4))
+  k.emit(v_accvgpr_read(v[40], v[3]))
+  k.emit(v_accvgpr_read(v[41], v[7]))
+  k.emit(v_accvgpr_read(v[42], v[11]))
+  k.emit(v_accvgpr_read(v[43], v[15]))
+  k.emit(v_accvgpr_read(v[44], v[19]))
+  k.emit(v_accvgpr_read(v[45], v[23]))
+  k.emit(v_accvgpr_read(v[46], v[27]))
+  k.emit(v_accvgpr_read(v[47], v[31]))
+  k.emit(v_accvgpr_read(v[48], v[35]))
+  k.emit(v_accvgpr_read(v[49], v[39]))
+  k.emit(v_accvgpr_read(v[50], v[43]))
+  k.emit(v_accvgpr_read(v[51], v[47]))
+  k.emit(v_accvgpr_read(v[52], v[51]))
+  k.emit(v_accvgpr_read(v[53], v[55]))
+  k.emit(v_accvgpr_read(v[54], v[59]))
+  k.emit(v_accvgpr_read(v[55], v[63]))
+  k.emit(v_accvgpr_read(v[56], v[67]))
+  k.emit(v_accvgpr_read(v[57], v[71]))
+  k.emit(v_accvgpr_read(v[58], v[75]))
+  k.emit(v_accvgpr_read(v[59], v[79]))
+  k.emit(v_accvgpr_read(v[60], v[83]))
+  k.emit(v_accvgpr_read(v[61], v[87]))
+  k.emit(v_accvgpr_read(v[62], v[91]))
+  k.emit(v_accvgpr_read(v[63], v[95]))
+  k.emit(v_accvgpr_read(v[64], v[99]))
+  k.emit(v_accvgpr_read(v[65], v[103]))
+  k.emit(v_accvgpr_read(v[66], v[107]))
+  k.emit(v_accvgpr_read(v[67], v[111]))
+  k.emit(v_accvgpr_read(v[68], v[115]))
+  k.emit(v_accvgpr_read(v[69], v[119]))
+  k.emit(v_accvgpr_read(v[70], v[123]))
+  k.emit(v_accvgpr_read(v[71], v[127]))
+  k.emit(v_accvgpr_read(v[72], v[131]))
+  k.emit(v_accvgpr_read(v[73], v[135]))
+  k.emit(v_accvgpr_read(v[74], v[139]))
+  k.emit(v_accvgpr_read(v[75], v[143]))
+  k.emit(v_accvgpr_read(v[76], v[147]))
+  k.emit(v_accvgpr_read(v[77], v[151]))
+  k.emit(v_accvgpr_read(v[78], v[155]))
+  k.emit(v_accvgpr_read(v[79], v[159]))
+  k.emit(v_accvgpr_read(v[80], v[163]))
+  k.emit(v_accvgpr_read(v[81], v[167]))
+  k.emit(v_accvgpr_read(v[82], v[171]))
+  k.emit(v_accvgpr_read(v[83], v[175]))
+  k.emit(v_accvgpr_read(v[84], v[179]))
+  k.emit(v_accvgpr_read(v[85], v[183]))
+  k.emit(v_accvgpr_read(v[86], v[187]))
+  k.emit(v_accvgpr_read(v[87], v[191]))
+  k.emit(v_mov_b32_e32(v[32], 4294901760))
+  k.emit(v_mov_b32_e32(v[33], 2147418112))
+  k.emit(v_mov_b32_e32(v[34], 32767))
+  k.waitcnt(lgkm=0)
+  k.emit(v_pk_mul_f32(v[40:41], v[96:97], v[40:41]))
+  k.emit(v_pk_mul_f32(v[42:43], v[98:99], v[42:43]))
+  k.emit(v_pk_mul_f32(v[44:45], v[100:101], v[44:45]))
+  k.emit(v_pk_mul_f32(v[46:47], v[102:103], v[46:47]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[40:41]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[42:43]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[44:45]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[46:47]))
+  k.emit(v_mov_b64_e32(v[40:41], v[22:23]))
+  k.emit(v_mov_b64_e32(v[42:43], v[24:25]))
+  k.emit(v_mov_b64_e32(v[44:45], v[26:27]))
+  k.emit(v_mov_b64_e32(v[46:47], v[28:29]))
+  k.emit(v_cvt_pk(v[40], v[40], v[41]))
+  k.emit(v_cvt_pk(v[41], v[42], v[43]))
+  k.emit(v_cvt_pk(v[42], v[44], v[45]))
+  k.emit(v_cvt_pk(v[43], v[46], v[47]))
+  k.emit(s_lshl_b32(s[68], s[36], 1))
+  k.emit(s_add_u32(s[12], s[12], s[68]))
+  k.emit(s_addc_u32(s[13], s[13], 0))
+  k.emit(buffer_store_dwordx4(v[40:43], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[48:49], v[96:97], v[48:49]))
+  k.emit(v_pk_mul_f32(v[50:51], v[98:99], v[50:51]))
+  k.emit(v_pk_mul_f32(v[52:53], v[100:101], v[52:53]))
+  k.emit(v_pk_mul_f32(v[54:55], v[102:103], v[54:55]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[48:49]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[50:51]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[52:53]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[54:55]))
+  k.emit(v_mov_b64_e32(v[48:49], v[22:23]))
+  k.emit(v_mov_b64_e32(v[50:51], v[24:25]))
+  k.emit(v_mov_b64_e32(v[52:53], v[26:27]))
+  k.emit(v_mov_b64_e32(v[54:55], v[28:29]))
+  k.emit(v_cvt_pk(v[48], v[48], v[49]))
+  k.emit(v_cvt_pk(v[49], v[50], v[51]))
+  k.emit(v_cvt_pk(v[50], v[52], v[53]))
+  k.emit(v_cvt_pk(v[51], v[54], v[55]))
+  k.emit(s_lshl_b32(s[68], s[36], 1))
+  k.emit(s_add_u32(s[12], s[12], s[68]))
+  k.emit(s_addc_u32(s[13], s[13], 0))
+  k.emit(buffer_store_dwordx4(v[48:51], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[56:57], v[96:97], v[56:57]))
+  k.emit(v_pk_mul_f32(v[58:59], v[98:99], v[58:59]))
+  k.emit(v_pk_mul_f32(v[60:61], v[100:101], v[60:61]))
+  k.emit(v_pk_mul_f32(v[62:63], v[102:103], v[62:63]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[56:57]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[58:59]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[60:61]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[62:63]))
+  k.emit(v_mov_b64_e32(v[56:57], v[22:23]))
+  k.emit(v_mov_b64_e32(v[58:59], v[24:25]))
+  k.emit(v_mov_b64_e32(v[60:61], v[26:27]))
+  k.emit(v_mov_b64_e32(v[62:63], v[28:29]))
+  k.emit(v_cvt_pk(v[56], v[56], v[57]))
+  k.emit(v_cvt_pk(v[57], v[58], v[59]))
+  k.emit(v_cvt_pk(v[58], v[60], v[61]))
+  k.emit(v_cvt_pk(v[59], v[62], v[63]))
+  k.emit(s_lshl_b32(s[68], s[36], 1))
+  k.emit(s_add_u32(s[12], s[12], s[68]))
+  k.emit(s_addc_u32(s[13], s[13], 0))
+  k.emit(buffer_store_dwordx4(v[56:59], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[64:65], v[96:97], v[64:65]))
+  k.emit(v_pk_mul_f32(v[66:67], v[98:99], v[66:67]))
+  k.emit(v_pk_mul_f32(v[68:69], v[100:101], v[68:69]))
+  k.emit(v_pk_mul_f32(v[70:71], v[102:103], v[70:71]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[64:65]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[66:67]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[68:69]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[70:71]))
+  k.emit(v_mov_b64_e32(v[64:65], v[22:23]))
+  k.emit(v_mov_b64_e32(v[66:67], v[24:25]))
+  k.emit(v_mov_b64_e32(v[68:69], v[26:27]))
+  k.emit(v_mov_b64_e32(v[70:71], v[28:29]))
+  k.emit(v_cvt_pk(v[64], v[64], v[65]))
+  k.emit(v_cvt_pk(v[65], v[66], v[67]))
+  k.emit(v_cvt_pk(v[66], v[68], v[69]))
+  k.emit(v_cvt_pk(v[67], v[70], v[71]))
+  k.emit(s_lshl_b32(s[68], s[36], 1))
+  k.emit(s_add_u32(s[12], s[12], s[68]))
+  k.emit(s_addc_u32(s[13], s[13], 0))
+  k.emit(buffer_store_dwordx4(v[64:67], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[72:73], v[96:97], v[72:73]))
+  k.emit(v_pk_mul_f32(v[74:75], v[98:99], v[74:75]))
+  k.emit(v_pk_mul_f32(v[76:77], v[100:101], v[76:77]))
+  k.emit(v_pk_mul_f32(v[78:79], v[102:103], v[78:79]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[72:73]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[74:75]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[76:77]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[78:79]))
+  k.emit(v_mov_b64_e32(v[72:73], v[22:23]))
+  k.emit(v_mov_b64_e32(v[74:75], v[24:25]))
+  k.emit(v_mov_b64_e32(v[76:77], v[26:27]))
+  k.emit(v_mov_b64_e32(v[78:79], v[28:29]))
+  k.emit(v_cvt_pk(v[72], v[72], v[73]))
+  k.emit(v_cvt_pk(v[73], v[74], v[75]))
+  k.emit(v_cvt_pk(v[74], v[76], v[77]))
+  k.emit(v_cvt_pk(v[75], v[78], v[79]))
+  k.emit(s_lshl_b32(s[68], s[36], 1))
+  k.emit(s_add_u32(s[12], s[12], s[68]))
+  k.emit(s_addc_u32(s[13], s[13], 0))
+  k.emit(buffer_store_dwordx4(v[72:75], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[80:81], v[96:97], v[80:81]))
+  k.emit(v_pk_mul_f32(v[82:83], v[98:99], v[82:83]))
+  k.emit(v_pk_mul_f32(v[84:85], v[100:101], v[84:85]))
+  k.emit(v_pk_mul_f32(v[86:87], v[102:103], v[86:87]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[80:81]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[82:83]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[84:85]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[86:87]))
+  k.emit(v_mov_b64_e32(v[80:81], v[22:23]))
+  k.emit(v_mov_b64_e32(v[82:83], v[24:25]))
+  k.emit(v_mov_b64_e32(v[84:85], v[26:27]))
+  k.emit(v_mov_b64_e32(v[86:87], v[28:29]))
+  k.emit(v_cvt_pk(v[80], v[80], v[81]))
+  k.emit(v_cvt_pk(v[81], v[82], v[83]))
+  k.emit(v_cvt_pk(v[82], v[84], v[85]))
+  k.emit(v_cvt_pk(v[83], v[86], v[87]))
+  k.emit(s_lshl_b32(s[68], s[36], 1))
+  k.emit(s_add_u32(s[12], s[12], s[68]))
+  k.emit(s_addc_u32(s[13], s[13], 0))
+  k.emit(buffer_store_dwordx4(v[80:83], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_nop())
+  k.emit(ds_read_b128(v[56:59], v[37]))
+  k.emit(ds_read_b128(v[60:63], v[37], v[0], v[0], 0, 16))
+  k.emit(ds_read_b128(v[64:67], v[37], v[0], v[0], 0, 0, 4))
+  k.emit(ds_read_b128(v[68:71], v[37], v[0], v[0], 0, 16, 4))
+  k.emit(v_accvgpr_read(v[40], v[195]))
+  k.emit(v_accvgpr_read(v[41], v[199]))
+  k.emit(v_accvgpr_read(v[42], v[203]))
+  k.emit(v_accvgpr_read(v[43], v[207]))
+  k.emit(v_accvgpr_read(v[44], v[211]))
+  k.emit(v_accvgpr_read(v[45], v[215]))
+  k.emit(v_accvgpr_read(v[46], v[219]))
+  k.emit(v_accvgpr_read(v[47], v[223]))
+  k.emit(v_accvgpr_read(v[48], v[227]))
+  k.emit(v_accvgpr_read(v[49], v[231]))
+  k.emit(v_accvgpr_read(v[50], v[235]))
+  k.emit(v_accvgpr_read(v[51], v[239]))
+  k.emit(v_accvgpr_read(v[52], v[243]))
+  k.emit(v_accvgpr_read(v[53], v[247]))
+  k.emit(v_accvgpr_read(v[54], v[251]))
+  k.emit(v_accvgpr_read(v[55], v[255]))
+  k.emit(v_mov_b32_e32(v[32], 4294901760))
+  k.emit(v_mov_b32_e32(v[33], 2147418112))
+  k.emit(v_mov_b32_e32(v[34], 32767))
+  k.waitcnt(lgkm=0)
+  k.emit(v_pk_mul_f32(v[40:41], v[64:65], v[40:41]))
+  k.emit(v_pk_mul_f32(v[42:43], v[66:67], v[42:43]))
+  k.emit(v_pk_mul_f32(v[44:45], v[68:69], v[44:45]))
+  k.emit(v_pk_mul_f32(v[46:47], v[70:71], v[46:47]))
+  k.emit(v_pk_add_f32(v[22:23], v[56:57], v[40:41]))
+  k.emit(v_pk_add_f32(v[24:25], v[58:59], v[42:43]))
+  k.emit(v_pk_add_f32(v[26:27], v[60:61], v[44:45]))
+  k.emit(v_pk_add_f32(v[28:29], v[62:63], v[46:47]))
+  k.emit(v_mov_b64_e32(v[40:41], v[22:23]))
+  k.emit(v_mov_b64_e32(v[42:43], v[24:25]))
+  k.emit(v_mov_b64_e32(v[44:45], v[26:27]))
+  k.emit(v_mov_b64_e32(v[46:47], v[28:29]))
+  k.emit(v_cvt_pk(v[40], v[40], v[41]))
+  k.emit(v_cvt_pk(v[41], v[42], v[43]))
+  k.emit(v_cvt_pk(v[42], v[44], v[45]))
+  k.emit(v_cvt_pk(v[43], v[46], v[47]))
+  k.emit(s_lshl_b32(s[68], s[36], 1))
+  k.emit(s_add_u32(s[12], s[12], s[68]))
+  k.emit(s_addc_u32(s[13], s[13], 0))
+  k.emit(buffer_store_dwordx4(v[40:43], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[48:49], v[64:65], v[48:49]))
+  k.emit(v_pk_mul_f32(v[50:51], v[66:67], v[50:51]))
+  k.emit(v_pk_mul_f32(v[52:53], v[68:69], v[52:53]))
+  k.emit(v_pk_mul_f32(v[54:55], v[70:71], v[54:55]))
+  k.emit(v_pk_add_f32(v[22:23], v[56:57], v[48:49]))
+  k.emit(v_pk_add_f32(v[24:25], v[58:59], v[50:51]))
+  k.emit(v_pk_add_f32(v[26:27], v[60:61], v[52:53]))
+  k.emit(v_pk_add_f32(v[28:29], v[62:63], v[54:55]))
+  k.emit(v_mov_b64_e32(v[48:49], v[22:23]))
+  k.emit(v_mov_b64_e32(v[50:51], v[24:25]))
+  k.emit(v_mov_b64_e32(v[52:53], v[26:27]))
+  k.emit(v_mov_b64_e32(v[54:55], v[28:29]))
+  k.emit(v_cvt_pk(v[48], v[48], v[49]))
+  k.emit(v_cvt_pk(v[49], v[50], v[51]))
+  k.emit(v_cvt_pk(v[50], v[52], v[53]))
+  k.emit(v_cvt_pk(v[51], v[54], v[55]))
+  k.emit(s_lshl_b32(s[68], s[36], 1))
+  k.emit(s_add_u32(s[12], s[12], s[68]))
+  k.emit(s_addc_u32(s[13], s[13], 0))
+  k.emit(buffer_store_dwordx4(v[48:51], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_nop())
+  k.emit(s_branch(), target='GW_End_1')
+  k.label('GW_B0_E1_N_1')
+  k.emit(v_mov_b32_e32(v[30], 2147483648))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[36], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[36], 2, v[36]))
+  k.waitcnt(lgkm=0)
+  k.emit(s_barrier())
+  k.emit(ds_read_b128(v[88:91], v[36]))
+  k.emit(ds_read_b128(v[92:95], v[36], v[0], v[0], 0, 16))
+  k.emit(ds_read_b128(v[96:99], v[36], v[0], v[0], 0, 0, 4))
+  k.emit(ds_read_b128(v[100:103], v[36], v[0], v[0], 0, 16, 4))
+  k.emit(v_add_lshl_u32_e64(v[35], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[35], v[30], v[35], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[38], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[38], 2, v[38]))
+  k.emit(v_add_lshl_u32_e64(v[37], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[37], v[30], v[37], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[104], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[104], 2, v[104]))
+  k.emit(v_add_lshl_u32_e64(v[39], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[39], v[30], v[39], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[106], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[106], 2, v[106]))
+  k.emit(v_add_lshl_u32_e64(v[105], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[105], v[30], v[105], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[108], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[108], 2, v[108]))
+  k.emit(v_add_lshl_u32_e64(v[107], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[107], v[30], v[107], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[110], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[110], 2, v[110]))
+  k.emit(v_add_lshl_u32_e64(v[109], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[109], v[30], v[109], s[82:83]))
+  k.emit(v_accvgpr_read(v[40], v[0]))
+  k.emit(v_accvgpr_read(v[41], v[4]))
+  k.emit(v_accvgpr_read(v[42], v[8]))
+  k.emit(v_accvgpr_read(v[43], v[12]))
+  k.emit(v_accvgpr_read(v[44], v[16]))
+  k.emit(v_accvgpr_read(v[45], v[20]))
+  k.emit(v_accvgpr_read(v[46], v[24]))
+  k.emit(v_accvgpr_read(v[47], v[28]))
+  k.emit(v_accvgpr_read(v[48], v[32]))
+  k.emit(v_accvgpr_read(v[49], v[36]))
+  k.emit(v_accvgpr_read(v[50], v[40]))
+  k.emit(v_accvgpr_read(v[51], v[44]))
+  k.emit(v_accvgpr_read(v[52], v[48]))
+  k.emit(v_accvgpr_read(v[53], v[52]))
+  k.emit(v_accvgpr_read(v[54], v[56]))
+  k.emit(v_accvgpr_read(v[55], v[60]))
+  k.emit(v_accvgpr_read(v[56], v[64]))
+  k.emit(v_accvgpr_read(v[57], v[68]))
+  k.emit(v_accvgpr_read(v[58], v[72]))
+  k.emit(v_accvgpr_read(v[59], v[76]))
+  k.emit(v_accvgpr_read(v[60], v[80]))
+  k.emit(v_accvgpr_read(v[61], v[84]))
+  k.emit(v_accvgpr_read(v[62], v[88]))
+  k.emit(v_accvgpr_read(v[63], v[92]))
+  k.emit(v_accvgpr_read(v[64], v[96]))
+  k.emit(v_accvgpr_read(v[65], v[100]))
+  k.emit(v_accvgpr_read(v[66], v[104]))
+  k.emit(v_accvgpr_read(v[67], v[108]))
+  k.emit(v_accvgpr_read(v[68], v[112]))
+  k.emit(v_accvgpr_read(v[69], v[116]))
+  k.emit(v_accvgpr_read(v[70], v[120]))
+  k.emit(v_accvgpr_read(v[71], v[124]))
+  k.emit(v_accvgpr_read(v[72], v[128]))
+  k.emit(v_accvgpr_read(v[73], v[132]))
+  k.emit(v_accvgpr_read(v[74], v[136]))
+  k.emit(v_accvgpr_read(v[75], v[140]))
+  k.emit(v_accvgpr_read(v[76], v[144]))
+  k.emit(v_accvgpr_read(v[77], v[148]))
+  k.emit(v_accvgpr_read(v[78], v[152]))
+  k.emit(v_accvgpr_read(v[79], v[156]))
+  k.emit(v_accvgpr_read(v[80], v[160]))
+  k.emit(v_accvgpr_read(v[81], v[164]))
+  k.emit(v_accvgpr_read(v[82], v[168]))
+  k.emit(v_accvgpr_read(v[83], v[172]))
+  k.emit(v_accvgpr_read(v[84], v[176]))
+  k.emit(v_accvgpr_read(v[85], v[180]))
+  k.emit(v_accvgpr_read(v[86], v[184]))
+  k.emit(v_accvgpr_read(v[87], v[188]))
+  k.waitcnt(lgkm=0)
+  k.emit(v_mov_b32_e32(v[32], 4294901760))
+  k.emit(v_mov_b32_e32(v[33], 2147418112))
+  k.emit(v_mov_b32_e32(v[34], 32767))
+  k.emit(v_pk_mul_f32(v[40:41], v[96:97], v[40:41]))
+  k.emit(v_pk_mul_f32(v[42:43], v[98:99], v[42:43]))
+  k.emit(v_pk_mul_f32(v[44:45], v[100:101], v[44:45]))
+  k.emit(v_pk_mul_f32(v[46:47], v[102:103], v[46:47]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[40:41]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[42:43]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[44:45]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[46:47]))
+  k.emit(v_mov_b64_e32(v[40:41], v[22:23]))
+  k.emit(v_mov_b64_e32(v[42:43], v[24:25]))
+  k.emit(v_mov_b64_e32(v[44:45], v[26:27]))
+  k.emit(v_mov_b64_e32(v[46:47], v[28:29]))
+  k.emit(v_cvt_pk(v[40], v[40], v[41]))
+  k.emit(v_cvt_pk(v[41], v[42], v[43]))
+  k.emit(v_cvt_pk(v[42], v[44], v[45]))
+  k.emit(v_cvt_pk(v[43], v[46], v[47]))
+  k.emit(buffer_store_dwordx4(v[40:43], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[48:49], v[96:97], v[48:49]))
+  k.emit(v_pk_mul_f32(v[50:51], v[98:99], v[50:51]))
+  k.emit(v_pk_mul_f32(v[52:53], v[100:101], v[52:53]))
+  k.emit(v_pk_mul_f32(v[54:55], v[102:103], v[54:55]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[48:49]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[50:51]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[52:53]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[54:55]))
+  k.emit(v_mov_b64_e32(v[48:49], v[22:23]))
+  k.emit(v_mov_b64_e32(v[50:51], v[24:25]))
+  k.emit(v_mov_b64_e32(v[52:53], v[26:27]))
+  k.emit(v_mov_b64_e32(v[54:55], v[28:29]))
+  k.emit(v_cvt_pk(v[48], v[48], v[49]))
+  k.emit(v_cvt_pk(v[49], v[50], v[51]))
+  k.emit(v_cvt_pk(v[50], v[52], v[53]))
+  k.emit(v_cvt_pk(v[51], v[54], v[55]))
+  k.emit(buffer_store_dwordx4(v[48:51], v[37], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[56:57], v[96:97], v[56:57]))
+  k.emit(v_pk_mul_f32(v[58:59], v[98:99], v[58:59]))
+  k.emit(v_pk_mul_f32(v[60:61], v[100:101], v[60:61]))
+  k.emit(v_pk_mul_f32(v[62:63], v[102:103], v[62:63]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[56:57]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[58:59]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[60:61]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[62:63]))
+  k.emit(v_mov_b64_e32(v[56:57], v[22:23]))
+  k.emit(v_mov_b64_e32(v[58:59], v[24:25]))
+  k.emit(v_mov_b64_e32(v[60:61], v[26:27]))
+  k.emit(v_mov_b64_e32(v[62:63], v[28:29]))
+  k.emit(v_cvt_pk(v[56], v[56], v[57]))
+  k.emit(v_cvt_pk(v[57], v[58], v[59]))
+  k.emit(v_cvt_pk(v[58], v[60], v[61]))
+  k.emit(v_cvt_pk(v[59], v[62], v[63]))
+  k.emit(buffer_store_dwordx4(v[56:59], v[39], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[64:65], v[96:97], v[64:65]))
+  k.emit(v_pk_mul_f32(v[66:67], v[98:99], v[66:67]))
+  k.emit(v_pk_mul_f32(v[68:69], v[100:101], v[68:69]))
+  k.emit(v_pk_mul_f32(v[70:71], v[102:103], v[70:71]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[64:65]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[66:67]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[68:69]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[70:71]))
+  k.emit(v_mov_b64_e32(v[64:65], v[22:23]))
+  k.emit(v_mov_b64_e32(v[66:67], v[24:25]))
+  k.emit(v_mov_b64_e32(v[68:69], v[26:27]))
+  k.emit(v_mov_b64_e32(v[70:71], v[28:29]))
+  k.emit(v_cvt_pk(v[64], v[64], v[65]))
+  k.emit(v_cvt_pk(v[65], v[66], v[67]))
+  k.emit(v_cvt_pk(v[66], v[68], v[69]))
+  k.emit(v_cvt_pk(v[67], v[70], v[71]))
+  k.emit(buffer_store_dwordx4(v[64:67], v[105], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[72:73], v[96:97], v[72:73]))
+  k.emit(v_pk_mul_f32(v[74:75], v[98:99], v[74:75]))
+  k.emit(v_pk_mul_f32(v[76:77], v[100:101], v[76:77]))
+  k.emit(v_pk_mul_f32(v[78:79], v[102:103], v[78:79]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[72:73]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[74:75]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[76:77]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[78:79]))
+  k.emit(v_mov_b64_e32(v[72:73], v[22:23]))
+  k.emit(v_mov_b64_e32(v[74:75], v[24:25]))
+  k.emit(v_mov_b64_e32(v[76:77], v[26:27]))
+  k.emit(v_mov_b64_e32(v[78:79], v[28:29]))
+  k.emit(v_cvt_pk(v[72], v[72], v[73]))
+  k.emit(v_cvt_pk(v[73], v[74], v[75]))
+  k.emit(v_cvt_pk(v[74], v[76], v[77]))
+  k.emit(v_cvt_pk(v[75], v[78], v[79]))
+  k.emit(buffer_store_dwordx4(v[72:75], v[107], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[80:81], v[96:97], v[80:81]))
+  k.emit(v_pk_mul_f32(v[82:83], v[98:99], v[82:83]))
+  k.emit(v_pk_mul_f32(v[84:85], v[100:101], v[84:85]))
+  k.emit(v_pk_mul_f32(v[86:87], v[102:103], v[86:87]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[80:81]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[82:83]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[84:85]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[86:87]))
+  k.emit(v_mov_b64_e32(v[80:81], v[22:23]))
+  k.emit(v_mov_b64_e32(v[82:83], v[24:25]))
+  k.emit(v_mov_b64_e32(v[84:85], v[26:27]))
+  k.emit(v_mov_b64_e32(v[86:87], v[28:29]))
+  k.emit(v_cvt_pk(v[80], v[80], v[81]))
+  k.emit(v_cvt_pk(v[81], v[82], v[83]))
+  k.emit(v_cvt_pk(v[82], v[84], v[85]))
+  k.emit(v_cvt_pk(v[83], v[86], v[87]))
+  k.emit(buffer_store_dwordx4(v[80:83], v[109], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_nop())
+  k.emit(v_mov_b32_e32(v[30], 2147483648))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[36], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[36], 2, v[36]))
+  k.emit(ds_read_b128(v[88:91], v[36]))
+  k.emit(ds_read_b128(v[92:95], v[36], v[0], v[0], 0, 16))
+  k.emit(ds_read_b128(v[96:99], v[36], v[0], v[0], 0, 0, 4))
+  k.emit(ds_read_b128(v[100:103], v[36], v[0], v[0], 0, 16, 4))
+  k.emit(v_add_lshl_u32_e64(v[35], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[35], v[30], v[35], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[38], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[38], 2, v[38]))
+  k.emit(v_add_lshl_u32_e64(v[37], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[37], v[30], v[37], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[104], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[104], 2, v[104]))
+  k.emit(v_add_lshl_u32_e64(v[39], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[39], v[30], v[39], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[106], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[106], 2, v[106]))
+  k.emit(v_add_lshl_u32_e64(v[105], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[105], v[30], v[105], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[108], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[108], 2, v[108]))
+  k.emit(v_add_lshl_u32_e64(v[107], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[107], v[30], v[107], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[110], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[110], 2, v[110]))
+  k.emit(v_add_lshl_u32_e64(v[109], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[109], v[30], v[109], s[82:83]))
+  k.emit(v_accvgpr_read(v[40], v[192]))
+  k.emit(v_accvgpr_read(v[41], v[196]))
+  k.emit(v_accvgpr_read(v[42], v[200]))
+  k.emit(v_accvgpr_read(v[43], v[204]))
+  k.emit(v_accvgpr_read(v[44], v[208]))
+  k.emit(v_accvgpr_read(v[45], v[212]))
+  k.emit(v_accvgpr_read(v[46], v[216]))
+  k.emit(v_accvgpr_read(v[47], v[220]))
+  k.emit(v_accvgpr_read(v[48], v[224]))
+  k.emit(v_accvgpr_read(v[49], v[228]))
+  k.emit(v_accvgpr_read(v[50], v[232]))
+  k.emit(v_accvgpr_read(v[51], v[236]))
+  k.emit(v_accvgpr_read(v[52], v[240]))
+  k.emit(v_accvgpr_read(v[53], v[244]))
+  k.emit(v_accvgpr_read(v[54], v[248]))
+  k.emit(v_accvgpr_read(v[55], v[252]))
+  k.emit(v_accvgpr_read(v[56], v[1]))
+  k.emit(v_accvgpr_read(v[57], v[5]))
+  k.emit(v_accvgpr_read(v[58], v[9]))
+  k.emit(v_accvgpr_read(v[59], v[13]))
+  k.emit(v_accvgpr_read(v[60], v[17]))
+  k.emit(v_accvgpr_read(v[61], v[21]))
+  k.emit(v_accvgpr_read(v[62], v[25]))
+  k.emit(v_accvgpr_read(v[63], v[29]))
+  k.emit(v_accvgpr_read(v[64], v[33]))
+  k.emit(v_accvgpr_read(v[65], v[37]))
+  k.emit(v_accvgpr_read(v[66], v[41]))
+  k.emit(v_accvgpr_read(v[67], v[45]))
+  k.emit(v_accvgpr_read(v[68], v[49]))
+  k.emit(v_accvgpr_read(v[69], v[53]))
+  k.emit(v_accvgpr_read(v[70], v[57]))
+  k.emit(v_accvgpr_read(v[71], v[61]))
+  k.emit(v_accvgpr_read(v[72], v[65]))
+  k.emit(v_accvgpr_read(v[73], v[69]))
+  k.emit(v_accvgpr_read(v[74], v[73]))
+  k.emit(v_accvgpr_read(v[75], v[77]))
+  k.emit(v_accvgpr_read(v[76], v[81]))
+  k.emit(v_accvgpr_read(v[77], v[85]))
+  k.emit(v_accvgpr_read(v[78], v[89]))
+  k.emit(v_accvgpr_read(v[79], v[93]))
+  k.emit(v_accvgpr_read(v[80], v[97]))
+  k.emit(v_accvgpr_read(v[81], v[101]))
+  k.emit(v_accvgpr_read(v[82], v[105]))
+  k.emit(v_accvgpr_read(v[83], v[109]))
+  k.emit(v_accvgpr_read(v[84], v[113]))
+  k.emit(v_accvgpr_read(v[85], v[117]))
+  k.emit(v_accvgpr_read(v[86], v[121]))
+  k.emit(v_accvgpr_read(v[87], v[125]))
+  k.waitcnt(lgkm=0)
+  k.emit(v_mov_b32_e32(v[32], 4294901760))
+  k.emit(v_mov_b32_e32(v[33], 2147418112))
+  k.emit(v_mov_b32_e32(v[34], 32767))
+  k.emit(v_pk_mul_f32(v[40:41], v[96:97], v[40:41]))
+  k.emit(v_pk_mul_f32(v[42:43], v[98:99], v[42:43]))
+  k.emit(v_pk_mul_f32(v[44:45], v[100:101], v[44:45]))
+  k.emit(v_pk_mul_f32(v[46:47], v[102:103], v[46:47]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[40:41]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[42:43]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[44:45]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[46:47]))
+  k.emit(v_mov_b64_e32(v[40:41], v[22:23]))
+  k.emit(v_mov_b64_e32(v[42:43], v[24:25]))
+  k.emit(v_mov_b64_e32(v[44:45], v[26:27]))
+  k.emit(v_mov_b64_e32(v[46:47], v[28:29]))
+  k.emit(v_cvt_pk(v[40], v[40], v[41]))
+  k.emit(v_cvt_pk(v[41], v[42], v[43]))
+  k.emit(v_cvt_pk(v[42], v[44], v[45]))
+  k.emit(v_cvt_pk(v[43], v[46], v[47]))
+  k.emit(buffer_store_dwordx4(v[40:43], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[48:49], v[96:97], v[48:49]))
+  k.emit(v_pk_mul_f32(v[50:51], v[98:99], v[50:51]))
+  k.emit(v_pk_mul_f32(v[52:53], v[100:101], v[52:53]))
+  k.emit(v_pk_mul_f32(v[54:55], v[102:103], v[54:55]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[48:49]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[50:51]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[52:53]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[54:55]))
+  k.emit(v_mov_b64_e32(v[48:49], v[22:23]))
+  k.emit(v_mov_b64_e32(v[50:51], v[24:25]))
+  k.emit(v_mov_b64_e32(v[52:53], v[26:27]))
+  k.emit(v_mov_b64_e32(v[54:55], v[28:29]))
+  k.emit(v_cvt_pk(v[48], v[48], v[49]))
+  k.emit(v_cvt_pk(v[49], v[50], v[51]))
+  k.emit(v_cvt_pk(v[50], v[52], v[53]))
+  k.emit(v_cvt_pk(v[51], v[54], v[55]))
+  k.emit(buffer_store_dwordx4(v[48:51], v[37], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[56:57], v[96:97], v[56:57]))
+  k.emit(v_pk_mul_f32(v[58:59], v[98:99], v[58:59]))
+  k.emit(v_pk_mul_f32(v[60:61], v[100:101], v[60:61]))
+  k.emit(v_pk_mul_f32(v[62:63], v[102:103], v[62:63]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[56:57]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[58:59]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[60:61]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[62:63]))
+  k.emit(v_mov_b64_e32(v[56:57], v[22:23]))
+  k.emit(v_mov_b64_e32(v[58:59], v[24:25]))
+  k.emit(v_mov_b64_e32(v[60:61], v[26:27]))
+  k.emit(v_mov_b64_e32(v[62:63], v[28:29]))
+  k.emit(v_cvt_pk(v[56], v[56], v[57]))
+  k.emit(v_cvt_pk(v[57], v[58], v[59]))
+  k.emit(v_cvt_pk(v[58], v[60], v[61]))
+  k.emit(v_cvt_pk(v[59], v[62], v[63]))
+  k.emit(buffer_store_dwordx4(v[56:59], v[39], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[64:65], v[96:97], v[64:65]))
+  k.emit(v_pk_mul_f32(v[66:67], v[98:99], v[66:67]))
+  k.emit(v_pk_mul_f32(v[68:69], v[100:101], v[68:69]))
+  k.emit(v_pk_mul_f32(v[70:71], v[102:103], v[70:71]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[64:65]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[66:67]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[68:69]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[70:71]))
+  k.emit(v_mov_b64_e32(v[64:65], v[22:23]))
+  k.emit(v_mov_b64_e32(v[66:67], v[24:25]))
+  k.emit(v_mov_b64_e32(v[68:69], v[26:27]))
+  k.emit(v_mov_b64_e32(v[70:71], v[28:29]))
+  k.emit(v_cvt_pk(v[64], v[64], v[65]))
+  k.emit(v_cvt_pk(v[65], v[66], v[67]))
+  k.emit(v_cvt_pk(v[66], v[68], v[69]))
+  k.emit(v_cvt_pk(v[67], v[70], v[71]))
+  k.emit(buffer_store_dwordx4(v[64:67], v[105], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[72:73], v[96:97], v[72:73]))
+  k.emit(v_pk_mul_f32(v[74:75], v[98:99], v[74:75]))
+  k.emit(v_pk_mul_f32(v[76:77], v[100:101], v[76:77]))
+  k.emit(v_pk_mul_f32(v[78:79], v[102:103], v[78:79]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[72:73]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[74:75]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[76:77]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[78:79]))
+  k.emit(v_mov_b64_e32(v[72:73], v[22:23]))
+  k.emit(v_mov_b64_e32(v[74:75], v[24:25]))
+  k.emit(v_mov_b64_e32(v[76:77], v[26:27]))
+  k.emit(v_mov_b64_e32(v[78:79], v[28:29]))
+  k.emit(v_cvt_pk(v[72], v[72], v[73]))
+  k.emit(v_cvt_pk(v[73], v[74], v[75]))
+  k.emit(v_cvt_pk(v[74], v[76], v[77]))
+  k.emit(v_cvt_pk(v[75], v[78], v[79]))
+  k.emit(buffer_store_dwordx4(v[72:75], v[107], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[80:81], v[96:97], v[80:81]))
+  k.emit(v_pk_mul_f32(v[82:83], v[98:99], v[82:83]))
+  k.emit(v_pk_mul_f32(v[84:85], v[100:101], v[84:85]))
+  k.emit(v_pk_mul_f32(v[86:87], v[102:103], v[86:87]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[80:81]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[82:83]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[84:85]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[86:87]))
+  k.emit(v_mov_b64_e32(v[80:81], v[22:23]))
+  k.emit(v_mov_b64_e32(v[82:83], v[24:25]))
+  k.emit(v_mov_b64_e32(v[84:85], v[26:27]))
+  k.emit(v_mov_b64_e32(v[86:87], v[28:29]))
+  k.emit(v_cvt_pk(v[80], v[80], v[81]))
+  k.emit(v_cvt_pk(v[81], v[82], v[83]))
+  k.emit(v_cvt_pk(v[82], v[84], v[85]))
+  k.emit(v_cvt_pk(v[83], v[86], v[87]))
+  k.emit(buffer_store_dwordx4(v[80:83], v[109], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_nop())
+  k.emit(v_mov_b32_e32(v[30], 2147483648))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[36], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[36], 2, v[36]))
+  k.emit(ds_read_b128(v[88:91], v[36]))
+  k.emit(ds_read_b128(v[92:95], v[36], v[0], v[0], 0, 16))
+  k.emit(ds_read_b128(v[96:99], v[36], v[0], v[0], 0, 0, 4))
+  k.emit(ds_read_b128(v[100:103], v[36], v[0], v[0], 0, 16, 4))
+  k.emit(v_add_lshl_u32_e64(v[35], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[35], v[30], v[35], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[38], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[38], 2, v[38]))
+  k.emit(v_add_lshl_u32_e64(v[37], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[37], v[30], v[37], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[104], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[104], 2, v[104]))
+  k.emit(v_add_lshl_u32_e64(v[39], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[39], v[30], v[39], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[106], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[106], 2, v[106]))
+  k.emit(v_add_lshl_u32_e64(v[105], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[105], v[30], v[105], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[108], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[108], 2, v[108]))
+  k.emit(v_add_lshl_u32_e64(v[107], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[107], v[30], v[107], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[110], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[110], 2, v[110]))
+  k.emit(v_add_lshl_u32_e64(v[109], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[109], v[30], v[109], s[82:83]))
+  k.emit(v_accvgpr_read(v[40], v[129]))
+  k.emit(v_accvgpr_read(v[41], v[133]))
+  k.emit(v_accvgpr_read(v[42], v[137]))
+  k.emit(v_accvgpr_read(v[43], v[141]))
+  k.emit(v_accvgpr_read(v[44], v[145]))
+  k.emit(v_accvgpr_read(v[45], v[149]))
+  k.emit(v_accvgpr_read(v[46], v[153]))
+  k.emit(v_accvgpr_read(v[47], v[157]))
+  k.emit(v_accvgpr_read(v[48], v[161]))
+  k.emit(v_accvgpr_read(v[49], v[165]))
+  k.emit(v_accvgpr_read(v[50], v[169]))
+  k.emit(v_accvgpr_read(v[51], v[173]))
+  k.emit(v_accvgpr_read(v[52], v[177]))
+  k.emit(v_accvgpr_read(v[53], v[181]))
+  k.emit(v_accvgpr_read(v[54], v[185]))
+  k.emit(v_accvgpr_read(v[55], v[189]))
+  k.emit(v_accvgpr_read(v[56], v[193]))
+  k.emit(v_accvgpr_read(v[57], v[197]))
+  k.emit(v_accvgpr_read(v[58], v[201]))
+  k.emit(v_accvgpr_read(v[59], v[205]))
+  k.emit(v_accvgpr_read(v[60], v[209]))
+  k.emit(v_accvgpr_read(v[61], v[213]))
+  k.emit(v_accvgpr_read(v[62], v[217]))
+  k.emit(v_accvgpr_read(v[63], v[221]))
+  k.emit(v_accvgpr_read(v[64], v[225]))
+  k.emit(v_accvgpr_read(v[65], v[229]))
+  k.emit(v_accvgpr_read(v[66], v[233]))
+  k.emit(v_accvgpr_read(v[67], v[237]))
+  k.emit(v_accvgpr_read(v[68], v[241]))
+  k.emit(v_accvgpr_read(v[69], v[245]))
+  k.emit(v_accvgpr_read(v[70], v[249]))
+  k.emit(v_accvgpr_read(v[71], v[253]))
+  k.emit(v_accvgpr_read(v[72], v[2]))
+  k.emit(v_accvgpr_read(v[73], v[6]))
+  k.emit(v_accvgpr_read(v[74], v[10]))
+  k.emit(v_accvgpr_read(v[75], v[14]))
+  k.emit(v_accvgpr_read(v[76], v[18]))
+  k.emit(v_accvgpr_read(v[77], v[22]))
+  k.emit(v_accvgpr_read(v[78], v[26]))
+  k.emit(v_accvgpr_read(v[79], v[30]))
+  k.emit(v_accvgpr_read(v[80], v[34]))
+  k.emit(v_accvgpr_read(v[81], v[38]))
+  k.emit(v_accvgpr_read(v[82], v[42]))
+  k.emit(v_accvgpr_read(v[83], v[46]))
+  k.emit(v_accvgpr_read(v[84], v[50]))
+  k.emit(v_accvgpr_read(v[85], v[54]))
+  k.emit(v_accvgpr_read(v[86], v[58]))
+  k.emit(v_accvgpr_read(v[87], v[62]))
+  k.waitcnt(lgkm=0)
+  k.emit(v_mov_b32_e32(v[32], 4294901760))
+  k.emit(v_mov_b32_e32(v[33], 2147418112))
+  k.emit(v_mov_b32_e32(v[34], 32767))
+  k.emit(v_pk_mul_f32(v[40:41], v[96:97], v[40:41]))
+  k.emit(v_pk_mul_f32(v[42:43], v[98:99], v[42:43]))
+  k.emit(v_pk_mul_f32(v[44:45], v[100:101], v[44:45]))
+  k.emit(v_pk_mul_f32(v[46:47], v[102:103], v[46:47]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[40:41]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[42:43]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[44:45]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[46:47]))
+  k.emit(v_mov_b64_e32(v[40:41], v[22:23]))
+  k.emit(v_mov_b64_e32(v[42:43], v[24:25]))
+  k.emit(v_mov_b64_e32(v[44:45], v[26:27]))
+  k.emit(v_mov_b64_e32(v[46:47], v[28:29]))
+  k.emit(v_cvt_pk(v[40], v[40], v[41]))
+  k.emit(v_cvt_pk(v[41], v[42], v[43]))
+  k.emit(v_cvt_pk(v[42], v[44], v[45]))
+  k.emit(v_cvt_pk(v[43], v[46], v[47]))
+  k.emit(buffer_store_dwordx4(v[40:43], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[48:49], v[96:97], v[48:49]))
+  k.emit(v_pk_mul_f32(v[50:51], v[98:99], v[50:51]))
+  k.emit(v_pk_mul_f32(v[52:53], v[100:101], v[52:53]))
+  k.emit(v_pk_mul_f32(v[54:55], v[102:103], v[54:55]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[48:49]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[50:51]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[52:53]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[54:55]))
+  k.emit(v_mov_b64_e32(v[48:49], v[22:23]))
+  k.emit(v_mov_b64_e32(v[50:51], v[24:25]))
+  k.emit(v_mov_b64_e32(v[52:53], v[26:27]))
+  k.emit(v_mov_b64_e32(v[54:55], v[28:29]))
+  k.emit(v_cvt_pk(v[48], v[48], v[49]))
+  k.emit(v_cvt_pk(v[49], v[50], v[51]))
+  k.emit(v_cvt_pk(v[50], v[52], v[53]))
+  k.emit(v_cvt_pk(v[51], v[54], v[55]))
+  k.emit(buffer_store_dwordx4(v[48:51], v[37], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[56:57], v[96:97], v[56:57]))
+  k.emit(v_pk_mul_f32(v[58:59], v[98:99], v[58:59]))
+  k.emit(v_pk_mul_f32(v[60:61], v[100:101], v[60:61]))
+  k.emit(v_pk_mul_f32(v[62:63], v[102:103], v[62:63]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[56:57]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[58:59]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[60:61]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[62:63]))
+  k.emit(v_mov_b64_e32(v[56:57], v[22:23]))
+  k.emit(v_mov_b64_e32(v[58:59], v[24:25]))
+  k.emit(v_mov_b64_e32(v[60:61], v[26:27]))
+  k.emit(v_mov_b64_e32(v[62:63], v[28:29]))
+  k.emit(v_cvt_pk(v[56], v[56], v[57]))
+  k.emit(v_cvt_pk(v[57], v[58], v[59]))
+  k.emit(v_cvt_pk(v[58], v[60], v[61]))
+  k.emit(v_cvt_pk(v[59], v[62], v[63]))
+  k.emit(buffer_store_dwordx4(v[56:59], v[39], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[64:65], v[96:97], v[64:65]))
+  k.emit(v_pk_mul_f32(v[66:67], v[98:99], v[66:67]))
+  k.emit(v_pk_mul_f32(v[68:69], v[100:101], v[68:69]))
+  k.emit(v_pk_mul_f32(v[70:71], v[102:103], v[70:71]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[64:65]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[66:67]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[68:69]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[70:71]))
+  k.emit(v_mov_b64_e32(v[64:65], v[22:23]))
+  k.emit(v_mov_b64_e32(v[66:67], v[24:25]))
+  k.emit(v_mov_b64_e32(v[68:69], v[26:27]))
+  k.emit(v_mov_b64_e32(v[70:71], v[28:29]))
+  k.emit(v_cvt_pk(v[64], v[64], v[65]))
+  k.emit(v_cvt_pk(v[65], v[66], v[67]))
+  k.emit(v_cvt_pk(v[66], v[68], v[69]))
+  k.emit(v_cvt_pk(v[67], v[70], v[71]))
+  k.emit(buffer_store_dwordx4(v[64:67], v[105], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[72:73], v[96:97], v[72:73]))
+  k.emit(v_pk_mul_f32(v[74:75], v[98:99], v[74:75]))
+  k.emit(v_pk_mul_f32(v[76:77], v[100:101], v[76:77]))
+  k.emit(v_pk_mul_f32(v[78:79], v[102:103], v[78:79]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[72:73]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[74:75]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[76:77]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[78:79]))
+  k.emit(v_mov_b64_e32(v[72:73], v[22:23]))
+  k.emit(v_mov_b64_e32(v[74:75], v[24:25]))
+  k.emit(v_mov_b64_e32(v[76:77], v[26:27]))
+  k.emit(v_mov_b64_e32(v[78:79], v[28:29]))
+  k.emit(v_cvt_pk(v[72], v[72], v[73]))
+  k.emit(v_cvt_pk(v[73], v[74], v[75]))
+  k.emit(v_cvt_pk(v[74], v[76], v[77]))
+  k.emit(v_cvt_pk(v[75], v[78], v[79]))
+  k.emit(buffer_store_dwordx4(v[72:75], v[107], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[80:81], v[96:97], v[80:81]))
+  k.emit(v_pk_mul_f32(v[82:83], v[98:99], v[82:83]))
+  k.emit(v_pk_mul_f32(v[84:85], v[100:101], v[84:85]))
+  k.emit(v_pk_mul_f32(v[86:87], v[102:103], v[86:87]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[80:81]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[82:83]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[84:85]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[86:87]))
+  k.emit(v_mov_b64_e32(v[80:81], v[22:23]))
+  k.emit(v_mov_b64_e32(v[82:83], v[24:25]))
+  k.emit(v_mov_b64_e32(v[84:85], v[26:27]))
+  k.emit(v_mov_b64_e32(v[86:87], v[28:29]))
+  k.emit(v_cvt_pk(v[80], v[80], v[81]))
+  k.emit(v_cvt_pk(v[81], v[82], v[83]))
+  k.emit(v_cvt_pk(v[82], v[84], v[85]))
+  k.emit(v_cvt_pk(v[83], v[86], v[87]))
+  k.emit(buffer_store_dwordx4(v[80:83], v[109], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_nop())
+  k.emit(v_mov_b32_e32(v[30], 2147483648))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[36], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[36], 2, v[36]))
+  k.emit(ds_read_b128(v[88:91], v[36]))
+  k.emit(ds_read_b128(v[92:95], v[36], v[0], v[0], 0, 16))
+  k.emit(ds_read_b128(v[96:99], v[36], v[0], v[0], 0, 0, 4))
+  k.emit(ds_read_b128(v[100:103], v[36], v[0], v[0], 0, 16, 4))
+  k.emit(v_add_lshl_u32_e64(v[35], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[35], v[30], v[35], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[38], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[38], 2, v[38]))
+  k.emit(v_add_lshl_u32_e64(v[37], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[37], v[30], v[37], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[104], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[104], 2, v[104]))
+  k.emit(v_add_lshl_u32_e64(v[39], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[39], v[30], v[39], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[106], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[106], 2, v[106]))
+  k.emit(v_add_lshl_u32_e64(v[105], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[105], v[30], v[105], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[108], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[108], 2, v[108]))
+  k.emit(v_add_lshl_u32_e64(v[107], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[107], v[30], v[107], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[110], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[110], 2, v[110]))
+  k.emit(v_add_lshl_u32_e64(v[109], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[109], v[30], v[109], s[82:83]))
+  k.emit(v_accvgpr_read(v[40], v[66]))
+  k.emit(v_accvgpr_read(v[41], v[70]))
+  k.emit(v_accvgpr_read(v[42], v[74]))
+  k.emit(v_accvgpr_read(v[43], v[78]))
+  k.emit(v_accvgpr_read(v[44], v[82]))
+  k.emit(v_accvgpr_read(v[45], v[86]))
+  k.emit(v_accvgpr_read(v[46], v[90]))
+  k.emit(v_accvgpr_read(v[47], v[94]))
+  k.emit(v_accvgpr_read(v[48], v[98]))
+  k.emit(v_accvgpr_read(v[49], v[102]))
+  k.emit(v_accvgpr_read(v[50], v[106]))
+  k.emit(v_accvgpr_read(v[51], v[110]))
+  k.emit(v_accvgpr_read(v[52], v[114]))
+  k.emit(v_accvgpr_read(v[53], v[118]))
+  k.emit(v_accvgpr_read(v[54], v[122]))
+  k.emit(v_accvgpr_read(v[55], v[126]))
+  k.emit(v_accvgpr_read(v[56], v[130]))
+  k.emit(v_accvgpr_read(v[57], v[134]))
+  k.emit(v_accvgpr_read(v[58], v[138]))
+  k.emit(v_accvgpr_read(v[59], v[142]))
+  k.emit(v_accvgpr_read(v[60], v[146]))
+  k.emit(v_accvgpr_read(v[61], v[150]))
+  k.emit(v_accvgpr_read(v[62], v[154]))
+  k.emit(v_accvgpr_read(v[63], v[158]))
+  k.emit(v_accvgpr_read(v[64], v[162]))
+  k.emit(v_accvgpr_read(v[65], v[166]))
+  k.emit(v_accvgpr_read(v[66], v[170]))
+  k.emit(v_accvgpr_read(v[67], v[174]))
+  k.emit(v_accvgpr_read(v[68], v[178]))
+  k.emit(v_accvgpr_read(v[69], v[182]))
+  k.emit(v_accvgpr_read(v[70], v[186]))
+  k.emit(v_accvgpr_read(v[71], v[190]))
+  k.emit(v_accvgpr_read(v[72], v[194]))
+  k.emit(v_accvgpr_read(v[73], v[198]))
+  k.emit(v_accvgpr_read(v[74], v[202]))
+  k.emit(v_accvgpr_read(v[75], v[206]))
+  k.emit(v_accvgpr_read(v[76], v[210]))
+  k.emit(v_accvgpr_read(v[77], v[214]))
+  k.emit(v_accvgpr_read(v[78], v[218]))
+  k.emit(v_accvgpr_read(v[79], v[222]))
+  k.emit(v_accvgpr_read(v[80], v[226]))
+  k.emit(v_accvgpr_read(v[81], v[230]))
+  k.emit(v_accvgpr_read(v[82], v[234]))
+  k.emit(v_accvgpr_read(v[83], v[238]))
+  k.emit(v_accvgpr_read(v[84], v[242]))
+  k.emit(v_accvgpr_read(v[85], v[246]))
+  k.emit(v_accvgpr_read(v[86], v[250]))
+  k.emit(v_accvgpr_read(v[87], v[254]))
+  k.waitcnt(lgkm=0)
+  k.emit(v_mov_b32_e32(v[32], 4294901760))
+  k.emit(v_mov_b32_e32(v[33], 2147418112))
+  k.emit(v_mov_b32_e32(v[34], 32767))
+  k.emit(v_pk_mul_f32(v[40:41], v[96:97], v[40:41]))
+  k.emit(v_pk_mul_f32(v[42:43], v[98:99], v[42:43]))
+  k.emit(v_pk_mul_f32(v[44:45], v[100:101], v[44:45]))
+  k.emit(v_pk_mul_f32(v[46:47], v[102:103], v[46:47]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[40:41]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[42:43]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[44:45]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[46:47]))
+  k.emit(v_mov_b64_e32(v[40:41], v[22:23]))
+  k.emit(v_mov_b64_e32(v[42:43], v[24:25]))
+  k.emit(v_mov_b64_e32(v[44:45], v[26:27]))
+  k.emit(v_mov_b64_e32(v[46:47], v[28:29]))
+  k.emit(v_cvt_pk(v[40], v[40], v[41]))
+  k.emit(v_cvt_pk(v[41], v[42], v[43]))
+  k.emit(v_cvt_pk(v[42], v[44], v[45]))
+  k.emit(v_cvt_pk(v[43], v[46], v[47]))
+  k.emit(buffer_store_dwordx4(v[40:43], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[48:49], v[96:97], v[48:49]))
+  k.emit(v_pk_mul_f32(v[50:51], v[98:99], v[50:51]))
+  k.emit(v_pk_mul_f32(v[52:53], v[100:101], v[52:53]))
+  k.emit(v_pk_mul_f32(v[54:55], v[102:103], v[54:55]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[48:49]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[50:51]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[52:53]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[54:55]))
+  k.emit(v_mov_b64_e32(v[48:49], v[22:23]))
+  k.emit(v_mov_b64_e32(v[50:51], v[24:25]))
+  k.emit(v_mov_b64_e32(v[52:53], v[26:27]))
+  k.emit(v_mov_b64_e32(v[54:55], v[28:29]))
+  k.emit(v_cvt_pk(v[48], v[48], v[49]))
+  k.emit(v_cvt_pk(v[49], v[50], v[51]))
+  k.emit(v_cvt_pk(v[50], v[52], v[53]))
+  k.emit(v_cvt_pk(v[51], v[54], v[55]))
+  k.emit(buffer_store_dwordx4(v[48:51], v[37], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[56:57], v[96:97], v[56:57]))
+  k.emit(v_pk_mul_f32(v[58:59], v[98:99], v[58:59]))
+  k.emit(v_pk_mul_f32(v[60:61], v[100:101], v[60:61]))
+  k.emit(v_pk_mul_f32(v[62:63], v[102:103], v[62:63]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[56:57]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[58:59]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[60:61]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[62:63]))
+  k.emit(v_mov_b64_e32(v[56:57], v[22:23]))
+  k.emit(v_mov_b64_e32(v[58:59], v[24:25]))
+  k.emit(v_mov_b64_e32(v[60:61], v[26:27]))
+  k.emit(v_mov_b64_e32(v[62:63], v[28:29]))
+  k.emit(v_cvt_pk(v[56], v[56], v[57]))
+  k.emit(v_cvt_pk(v[57], v[58], v[59]))
+  k.emit(v_cvt_pk(v[58], v[60], v[61]))
+  k.emit(v_cvt_pk(v[59], v[62], v[63]))
+  k.emit(buffer_store_dwordx4(v[56:59], v[39], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[64:65], v[96:97], v[64:65]))
+  k.emit(v_pk_mul_f32(v[66:67], v[98:99], v[66:67]))
+  k.emit(v_pk_mul_f32(v[68:69], v[100:101], v[68:69]))
+  k.emit(v_pk_mul_f32(v[70:71], v[102:103], v[70:71]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[64:65]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[66:67]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[68:69]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[70:71]))
+  k.emit(v_mov_b64_e32(v[64:65], v[22:23]))
+  k.emit(v_mov_b64_e32(v[66:67], v[24:25]))
+  k.emit(v_mov_b64_e32(v[68:69], v[26:27]))
+  k.emit(v_mov_b64_e32(v[70:71], v[28:29]))
+  k.emit(v_cvt_pk(v[64], v[64], v[65]))
+  k.emit(v_cvt_pk(v[65], v[66], v[67]))
+  k.emit(v_cvt_pk(v[66], v[68], v[69]))
+  k.emit(v_cvt_pk(v[67], v[70], v[71]))
+  k.emit(buffer_store_dwordx4(v[64:67], v[105], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[72:73], v[96:97], v[72:73]))
+  k.emit(v_pk_mul_f32(v[74:75], v[98:99], v[74:75]))
+  k.emit(v_pk_mul_f32(v[76:77], v[100:101], v[76:77]))
+  k.emit(v_pk_mul_f32(v[78:79], v[102:103], v[78:79]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[72:73]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[74:75]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[76:77]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[78:79]))
+  k.emit(v_mov_b64_e32(v[72:73], v[22:23]))
+  k.emit(v_mov_b64_e32(v[74:75], v[24:25]))
+  k.emit(v_mov_b64_e32(v[76:77], v[26:27]))
+  k.emit(v_mov_b64_e32(v[78:79], v[28:29]))
+  k.emit(v_cvt_pk(v[72], v[72], v[73]))
+  k.emit(v_cvt_pk(v[73], v[74], v[75]))
+  k.emit(v_cvt_pk(v[74], v[76], v[77]))
+  k.emit(v_cvt_pk(v[75], v[78], v[79]))
+  k.emit(buffer_store_dwordx4(v[72:75], v[107], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[80:81], v[96:97], v[80:81]))
+  k.emit(v_pk_mul_f32(v[82:83], v[98:99], v[82:83]))
+  k.emit(v_pk_mul_f32(v[84:85], v[100:101], v[84:85]))
+  k.emit(v_pk_mul_f32(v[86:87], v[102:103], v[86:87]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[80:81]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[82:83]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[84:85]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[86:87]))
+  k.emit(v_mov_b64_e32(v[80:81], v[22:23]))
+  k.emit(v_mov_b64_e32(v[82:83], v[24:25]))
+  k.emit(v_mov_b64_e32(v[84:85], v[26:27]))
+  k.emit(v_mov_b64_e32(v[86:87], v[28:29]))
+  k.emit(v_cvt_pk(v[80], v[80], v[81]))
+  k.emit(v_cvt_pk(v[81], v[82], v[83]))
+  k.emit(v_cvt_pk(v[82], v[84], v[85]))
+  k.emit(v_cvt_pk(v[83], v[86], v[87]))
+  k.emit(buffer_store_dwordx4(v[80:83], v[109], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_nop())
+  k.emit(v_mov_b32_e32(v[30], 2147483648))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[36], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[36], 2, v[36]))
+  k.emit(ds_read_b128(v[88:91], v[36]))
+  k.emit(ds_read_b128(v[92:95], v[36], v[0], v[0], 0, 16))
+  k.emit(ds_read_b128(v[96:99], v[36], v[0], v[0], 0, 0, 4))
+  k.emit(ds_read_b128(v[100:103], v[36], v[0], v[0], 0, 16, 4))
+  k.emit(v_add_lshl_u32_e64(v[35], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[35], v[30], v[35], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[38], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[38], 2, v[38]))
+  k.emit(v_add_lshl_u32_e64(v[37], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[37], v[30], v[37], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[104], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[104], 2, v[104]))
+  k.emit(v_add_lshl_u32_e64(v[39], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[39], v[30], v[39], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[106], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[106], 2, v[106]))
+  k.emit(v_add_lshl_u32_e64(v[105], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[105], v[30], v[105], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[108], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[108], 2, v[108]))
+  k.emit(v_add_lshl_u32_e64(v[107], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[107], v[30], v[107], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[110], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[110], 2, v[110]))
+  k.emit(v_add_lshl_u32_e64(v[109], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[109], v[30], v[109], s[82:83]))
+  k.emit(v_accvgpr_read(v[40], v[3]))
+  k.emit(v_accvgpr_read(v[41], v[7]))
+  k.emit(v_accvgpr_read(v[42], v[11]))
+  k.emit(v_accvgpr_read(v[43], v[15]))
+  k.emit(v_accvgpr_read(v[44], v[19]))
+  k.emit(v_accvgpr_read(v[45], v[23]))
+  k.emit(v_accvgpr_read(v[46], v[27]))
+  k.emit(v_accvgpr_read(v[47], v[31]))
+  k.emit(v_accvgpr_read(v[48], v[35]))
+  k.emit(v_accvgpr_read(v[49], v[39]))
+  k.emit(v_accvgpr_read(v[50], v[43]))
+  k.emit(v_accvgpr_read(v[51], v[47]))
+  k.emit(v_accvgpr_read(v[52], v[51]))
+  k.emit(v_accvgpr_read(v[53], v[55]))
+  k.emit(v_accvgpr_read(v[54], v[59]))
+  k.emit(v_accvgpr_read(v[55], v[63]))
+  k.emit(v_accvgpr_read(v[56], v[67]))
+  k.emit(v_accvgpr_read(v[57], v[71]))
+  k.emit(v_accvgpr_read(v[58], v[75]))
+  k.emit(v_accvgpr_read(v[59], v[79]))
+  k.emit(v_accvgpr_read(v[60], v[83]))
+  k.emit(v_accvgpr_read(v[61], v[87]))
+  k.emit(v_accvgpr_read(v[62], v[91]))
+  k.emit(v_accvgpr_read(v[63], v[95]))
+  k.emit(v_accvgpr_read(v[64], v[99]))
+  k.emit(v_accvgpr_read(v[65], v[103]))
+  k.emit(v_accvgpr_read(v[66], v[107]))
+  k.emit(v_accvgpr_read(v[67], v[111]))
+  k.emit(v_accvgpr_read(v[68], v[115]))
+  k.emit(v_accvgpr_read(v[69], v[119]))
+  k.emit(v_accvgpr_read(v[70], v[123]))
+  k.emit(v_accvgpr_read(v[71], v[127]))
+  k.emit(v_accvgpr_read(v[72], v[131]))
+  k.emit(v_accvgpr_read(v[73], v[135]))
+  k.emit(v_accvgpr_read(v[74], v[139]))
+  k.emit(v_accvgpr_read(v[75], v[143]))
+  k.emit(v_accvgpr_read(v[76], v[147]))
+  k.emit(v_accvgpr_read(v[77], v[151]))
+  k.emit(v_accvgpr_read(v[78], v[155]))
+  k.emit(v_accvgpr_read(v[79], v[159]))
+  k.emit(v_accvgpr_read(v[80], v[163]))
+  k.emit(v_accvgpr_read(v[81], v[167]))
+  k.emit(v_accvgpr_read(v[82], v[171]))
+  k.emit(v_accvgpr_read(v[83], v[175]))
+  k.emit(v_accvgpr_read(v[84], v[179]))
+  k.emit(v_accvgpr_read(v[85], v[183]))
+  k.emit(v_accvgpr_read(v[86], v[187]))
+  k.emit(v_accvgpr_read(v[87], v[191]))
+  k.waitcnt(lgkm=0)
+  k.emit(v_mov_b32_e32(v[32], 4294901760))
+  k.emit(v_mov_b32_e32(v[33], 2147418112))
+  k.emit(v_mov_b32_e32(v[34], 32767))
+  k.emit(v_pk_mul_f32(v[40:41], v[96:97], v[40:41]))
+  k.emit(v_pk_mul_f32(v[42:43], v[98:99], v[42:43]))
+  k.emit(v_pk_mul_f32(v[44:45], v[100:101], v[44:45]))
+  k.emit(v_pk_mul_f32(v[46:47], v[102:103], v[46:47]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[40:41]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[42:43]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[44:45]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[46:47]))
+  k.emit(v_mov_b64_e32(v[40:41], v[22:23]))
+  k.emit(v_mov_b64_e32(v[42:43], v[24:25]))
+  k.emit(v_mov_b64_e32(v[44:45], v[26:27]))
+  k.emit(v_mov_b64_e32(v[46:47], v[28:29]))
+  k.emit(v_cvt_pk(v[40], v[40], v[41]))
+  k.emit(v_cvt_pk(v[41], v[42], v[43]))
+  k.emit(v_cvt_pk(v[42], v[44], v[45]))
+  k.emit(v_cvt_pk(v[43], v[46], v[47]))
+  k.emit(buffer_store_dwordx4(v[40:43], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[48:49], v[96:97], v[48:49]))
+  k.emit(v_pk_mul_f32(v[50:51], v[98:99], v[50:51]))
+  k.emit(v_pk_mul_f32(v[52:53], v[100:101], v[52:53]))
+  k.emit(v_pk_mul_f32(v[54:55], v[102:103], v[54:55]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[48:49]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[50:51]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[52:53]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[54:55]))
+  k.emit(v_mov_b64_e32(v[48:49], v[22:23]))
+  k.emit(v_mov_b64_e32(v[50:51], v[24:25]))
+  k.emit(v_mov_b64_e32(v[52:53], v[26:27]))
+  k.emit(v_mov_b64_e32(v[54:55], v[28:29]))
+  k.emit(v_cvt_pk(v[48], v[48], v[49]))
+  k.emit(v_cvt_pk(v[49], v[50], v[51]))
+  k.emit(v_cvt_pk(v[50], v[52], v[53]))
+  k.emit(v_cvt_pk(v[51], v[54], v[55]))
+  k.emit(buffer_store_dwordx4(v[48:51], v[37], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[56:57], v[96:97], v[56:57]))
+  k.emit(v_pk_mul_f32(v[58:59], v[98:99], v[58:59]))
+  k.emit(v_pk_mul_f32(v[60:61], v[100:101], v[60:61]))
+  k.emit(v_pk_mul_f32(v[62:63], v[102:103], v[62:63]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[56:57]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[58:59]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[60:61]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[62:63]))
+  k.emit(v_mov_b64_e32(v[56:57], v[22:23]))
+  k.emit(v_mov_b64_e32(v[58:59], v[24:25]))
+  k.emit(v_mov_b64_e32(v[60:61], v[26:27]))
+  k.emit(v_mov_b64_e32(v[62:63], v[28:29]))
+  k.emit(v_cvt_pk(v[56], v[56], v[57]))
+  k.emit(v_cvt_pk(v[57], v[58], v[59]))
+  k.emit(v_cvt_pk(v[58], v[60], v[61]))
+  k.emit(v_cvt_pk(v[59], v[62], v[63]))
+  k.emit(buffer_store_dwordx4(v[56:59], v[39], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[64:65], v[96:97], v[64:65]))
+  k.emit(v_pk_mul_f32(v[66:67], v[98:99], v[66:67]))
+  k.emit(v_pk_mul_f32(v[68:69], v[100:101], v[68:69]))
+  k.emit(v_pk_mul_f32(v[70:71], v[102:103], v[70:71]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[64:65]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[66:67]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[68:69]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[70:71]))
+  k.emit(v_mov_b64_e32(v[64:65], v[22:23]))
+  k.emit(v_mov_b64_e32(v[66:67], v[24:25]))
+  k.emit(v_mov_b64_e32(v[68:69], v[26:27]))
+  k.emit(v_mov_b64_e32(v[70:71], v[28:29]))
+  k.emit(v_cvt_pk(v[64], v[64], v[65]))
+  k.emit(v_cvt_pk(v[65], v[66], v[67]))
+  k.emit(v_cvt_pk(v[66], v[68], v[69]))
+  k.emit(v_cvt_pk(v[67], v[70], v[71]))
+  k.emit(buffer_store_dwordx4(v[64:67], v[105], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[72:73], v[96:97], v[72:73]))
+  k.emit(v_pk_mul_f32(v[74:75], v[98:99], v[74:75]))
+  k.emit(v_pk_mul_f32(v[76:77], v[100:101], v[76:77]))
+  k.emit(v_pk_mul_f32(v[78:79], v[102:103], v[78:79]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[72:73]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[74:75]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[76:77]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[78:79]))
+  k.emit(v_mov_b64_e32(v[72:73], v[22:23]))
+  k.emit(v_mov_b64_e32(v[74:75], v[24:25]))
+  k.emit(v_mov_b64_e32(v[76:77], v[26:27]))
+  k.emit(v_mov_b64_e32(v[78:79], v[28:29]))
+  k.emit(v_cvt_pk(v[72], v[72], v[73]))
+  k.emit(v_cvt_pk(v[73], v[74], v[75]))
+  k.emit(v_cvt_pk(v[74], v[76], v[77]))
+  k.emit(v_cvt_pk(v[75], v[78], v[79]))
+  k.emit(buffer_store_dwordx4(v[72:75], v[107], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[80:81], v[96:97], v[80:81]))
+  k.emit(v_pk_mul_f32(v[82:83], v[98:99], v[82:83]))
+  k.emit(v_pk_mul_f32(v[84:85], v[100:101], v[84:85]))
+  k.emit(v_pk_mul_f32(v[86:87], v[102:103], v[86:87]))
+  k.emit(v_pk_add_f32(v[22:23], v[88:89], v[80:81]))
+  k.emit(v_pk_add_f32(v[24:25], v[90:91], v[82:83]))
+  k.emit(v_pk_add_f32(v[26:27], v[92:93], v[84:85]))
+  k.emit(v_pk_add_f32(v[28:29], v[94:95], v[86:87]))
+  k.emit(v_mov_b64_e32(v[80:81], v[22:23]))
+  k.emit(v_mov_b64_e32(v[82:83], v[24:25]))
+  k.emit(v_mov_b64_e32(v[84:85], v[26:27]))
+  k.emit(v_mov_b64_e32(v[86:87], v[28:29]))
+  k.emit(v_cvt_pk(v[80], v[80], v[81]))
+  k.emit(v_cvt_pk(v[81], v[82], v[83]))
+  k.emit(v_cvt_pk(v[82], v[84], v[85]))
+  k.emit(v_cvt_pk(v[83], v[86], v[87]))
+  k.emit(buffer_store_dwordx4(v[80:83], v[109], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_nop())
+  k.emit(v_mov_b32_e32(v[30], 2147483648))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[36], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[36], 2, v[36]))
+  k.emit(ds_read_b128(v[56:59], v[36]))
+  k.emit(ds_read_b128(v[60:63], v[36], v[0], v[0], 0, 16))
+  k.emit(ds_read_b128(v[64:67], v[36], v[0], v[0], 0, 0, 4))
+  k.emit(ds_read_b128(v[68:71], v[36], v[0], v[0], 0, 16, 4))
+  k.emit(v_add_lshl_u32_e64(v[35], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[35], v[30], v[35], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[38], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[38], 2, v[38]))
+  k.emit(v_add_lshl_u32_e64(v[37], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[37], v[30], v[37], s[82:83]))
+  k.emit(v_accvgpr_read(v[40], v[195]))
+  k.emit(v_accvgpr_read(v[41], v[199]))
+  k.emit(v_accvgpr_read(v[42], v[203]))
+  k.emit(v_accvgpr_read(v[43], v[207]))
+  k.emit(v_accvgpr_read(v[44], v[211]))
+  k.emit(v_accvgpr_read(v[45], v[215]))
+  k.emit(v_accvgpr_read(v[46], v[219]))
+  k.emit(v_accvgpr_read(v[47], v[223]))
+  k.emit(v_accvgpr_read(v[48], v[227]))
+  k.emit(v_accvgpr_read(v[49], v[231]))
+  k.emit(v_accvgpr_read(v[50], v[235]))
+  k.emit(v_accvgpr_read(v[51], v[239]))
+  k.emit(v_accvgpr_read(v[52], v[243]))
+  k.emit(v_accvgpr_read(v[53], v[247]))
+  k.emit(v_accvgpr_read(v[54], v[251]))
+  k.emit(v_accvgpr_read(v[55], v[255]))
+  k.waitcnt(lgkm=0)
+  k.emit(v_mov_b32_e32(v[32], 4294901760))
+  k.emit(v_mov_b32_e32(v[33], 2147418112))
+  k.emit(v_mov_b32_e32(v[34], 32767))
+  k.emit(v_pk_mul_f32(v[40:41], v[64:65], v[40:41]))
+  k.emit(v_pk_mul_f32(v[42:43], v[66:67], v[42:43]))
+  k.emit(v_pk_mul_f32(v[44:45], v[68:69], v[44:45]))
+  k.emit(v_pk_mul_f32(v[46:47], v[70:71], v[46:47]))
+  k.emit(v_pk_add_f32(v[22:23], v[56:57], v[40:41]))
+  k.emit(v_pk_add_f32(v[24:25], v[58:59], v[42:43]))
+  k.emit(v_pk_add_f32(v[26:27], v[60:61], v[44:45]))
+  k.emit(v_pk_add_f32(v[28:29], v[62:63], v[46:47]))
+  k.emit(v_mov_b64_e32(v[40:41], v[22:23]))
+  k.emit(v_mov_b64_e32(v[42:43], v[24:25]))
+  k.emit(v_mov_b64_e32(v[44:45], v[26:27]))
+  k.emit(v_mov_b64_e32(v[46:47], v[28:29]))
+  k.emit(v_cvt_pk(v[40], v[40], v[41]))
+  k.emit(v_cvt_pk(v[41], v[42], v[43]))
+  k.emit(v_cvt_pk(v[42], v[44], v[45]))
+  k.emit(v_cvt_pk(v[43], v[46], v[47]))
+  k.emit(buffer_store_dwordx4(v[40:43], v[35], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_pk_mul_f32(v[48:49], v[64:65], v[48:49]))
+  k.emit(v_pk_mul_f32(v[50:51], v[66:67], v[50:51]))
+  k.emit(v_pk_mul_f32(v[52:53], v[68:69], v[52:53]))
+  k.emit(v_pk_mul_f32(v[54:55], v[70:71], v[54:55]))
+  k.emit(v_pk_add_f32(v[22:23], v[56:57], v[48:49]))
+  k.emit(v_pk_add_f32(v[24:25], v[58:59], v[50:51]))
+  k.emit(v_pk_add_f32(v[26:27], v[60:61], v[52:53]))
+  k.emit(v_pk_add_f32(v[28:29], v[62:63], v[54:55]))
+  k.emit(v_mov_b64_e32(v[48:49], v[22:23]))
+  k.emit(v_mov_b64_e32(v[50:51], v[24:25]))
+  k.emit(v_mov_b64_e32(v[52:53], v[26:27]))
+  k.emit(v_mov_b64_e32(v[54:55], v[28:29]))
+  k.emit(v_cvt_pk(v[48], v[48], v[49]))
+  k.emit(v_cvt_pk(v[49], v[50], v[51]))
+  k.emit(v_cvt_pk(v[50], v[52], v[53]))
+  k.emit(v_cvt_pk(v[51], v[54], v[55]))
+  k.emit(buffer_store_dwordx4(v[48:51], v[37], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_nop())
+  k.emit(s_branch(), target='GW_End_1')
+  k.label('GW_B0_E1_M_1')
+  k.emit(v_mov_b32_e32(v[30], 2147483648))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[78], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[78], 2, v[78]))
+  k.waitcnt(lgkm=0)
+  k.emit(s_barrier())
+  k.emit(ds_read_b32(v[75], v[78]))
+  k.emit(ds_read_b32(v[76], v[78], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[77], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[77], v[30], v[77], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 1))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[82], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[82], 2, v[82]))
+  k.emit(ds_read_b32(v[79], v[82]))
+  k.emit(ds_read_b32(v[80], v[82], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[81], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[81], v[30], v[81], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 2))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[86], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[86], 2, v[86]))
+  k.emit(ds_read_b32(v[83], v[86]))
+  k.emit(ds_read_b32(v[84], v[86], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[85], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[85], v[30], v[85], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 3))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[90], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[90], 2, v[90]))
+  k.emit(ds_read_b32(v[87], v[90]))
+  k.emit(ds_read_b32(v[88], v[90], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[89], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[89], v[30], v[89], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 4))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[94], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[94], 2, v[94]))
+  k.emit(ds_read_b32(v[91], v[94]))
+  k.emit(ds_read_b32(v[92], v[94], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[93], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[93], v[30], v[93], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 5))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[98], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[98], 2, v[98]))
+  k.emit(ds_read_b32(v[95], v[98]))
+  k.emit(ds_read_b32(v[96], v[98], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[97], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[97], v[30], v[97], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 6))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[102], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[102], 2, v[102]))
+  k.emit(ds_read_b32(v[99], v[102]))
+  k.emit(ds_read_b32(v[100], v[102], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[101], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[101], v[30], v[101], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 7))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[106], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[106], 2, v[106]))
+  k.emit(ds_read_b32(v[103], v[106]))
+  k.emit(ds_read_b32(v[104], v[106], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[105], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[105], v[30], v[105], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[108], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[108], 2, v[108]))
+  k.emit(v_add_lshl_u32_e64(v[107], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[107], v[30], v[107], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 1))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[110], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[110], 2, v[110]))
+  k.emit(v_add_lshl_u32_e64(v[109], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[109], v[30], v[109], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 2))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[112], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[112], 2, v[112]))
+  k.emit(v_add_lshl_u32_e64(v[111], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[111], v[30], v[111], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 3))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[114], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[114], 2, v[114]))
+  k.emit(v_add_lshl_u32_e64(v[113], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[113], v[30], v[113], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 4))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[116], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[116], 2, v[116]))
+  k.emit(v_add_lshl_u32_e64(v[115], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[115], v[30], v[115], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 5))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[118], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[118], 2, v[118]))
+  k.emit(v_add_lshl_u32_e64(v[117], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[117], v[30], v[117], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 6))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[120], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[120], 2, v[120]))
+  k.emit(v_add_lshl_u32_e64(v[119], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[119], v[30], v[119], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 7))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[122], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[122], 2, v[122]))
+  k.emit(v_add_lshl_u32_e64(v[121], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[121], v[30], v[121], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[124], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[124], 2, v[124]))
+  k.emit(v_add_lshl_u32_e64(v[123], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[123], v[30], v[123], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 1))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[126], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[126], 2, v[126]))
+  k.emit(v_add_lshl_u32_e64(v[125], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[125], v[30], v[125], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 2))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[128], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[128], 2, v[128]))
+  k.emit(v_add_lshl_u32_e64(v[127], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[127], v[30], v[127], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 3))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[130], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[130], 2, v[130]))
+  k.emit(v_add_lshl_u32_e64(v[129], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[129], v[30], v[129], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 4))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[132], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[132], 2, v[132]))
+  k.emit(v_add_lshl_u32_e64(v[131], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[131], v[30], v[131], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 5))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[134], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[134], 2, v[134]))
+  k.emit(v_add_lshl_u32_e64(v[133], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[133], v[30], v[133], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 6))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[136], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[136], 2, v[136]))
+  k.emit(v_add_lshl_u32_e64(v[135], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[135], v[30], v[135], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 7))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[138], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[138], 2, v[138]))
+  k.emit(v_add_lshl_u32_e64(v[137], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[137], v[30], v[137], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[140], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[140], 2, v[140]))
+  k.emit(v_add_lshl_u32_e64(v[139], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[139], v[30], v[139], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 1))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[142], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[142], 2, v[142]))
+  k.emit(v_add_lshl_u32_e64(v[141], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[141], v[30], v[141], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 2))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[144], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[144], 2, v[144]))
+  k.emit(v_add_lshl_u32_e64(v[143], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[143], v[30], v[143], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 3))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[146], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[146], 2, v[146]))
+  k.emit(v_add_lshl_u32_e64(v[145], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[145], v[30], v[145], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 4))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[148], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[148], 2, v[148]))
+  k.emit(v_add_lshl_u32_e64(v[147], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[147], v[30], v[147], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 5))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[150], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[150], 2, v[150]))
+  k.emit(v_add_lshl_u32_e64(v[149], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[149], v[30], v[149], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 6))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[152], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[152], 2, v[152]))
+  k.emit(v_add_lshl_u32_e64(v[151], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[151], v[30], v[151], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 7))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[154], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[154], 2, v[154]))
+  k.emit(v_add_lshl_u32_e64(v[153], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[153], v[30], v[153], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[156], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[156], 2, v[156]))
+  k.emit(v_add_lshl_u32_e64(v[155], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[155], v[30], v[155], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 1))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[158], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[158], 2, v[158]))
+  k.emit(v_add_lshl_u32_e64(v[157], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[157], v[30], v[157], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 2))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[160], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[160], 2, v[160]))
+  k.emit(v_add_lshl_u32_e64(v[159], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[159], v[30], v[159], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 3))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[162], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[162], 2, v[162]))
+  k.emit(v_add_lshl_u32_e64(v[161], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[161], v[30], v[161], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 4))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[164], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[164], 2, v[164]))
+  k.emit(v_add_lshl_u32_e64(v[163], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[163], v[30], v[163], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 5))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[166], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[166], 2, v[166]))
+  k.emit(v_add_lshl_u32_e64(v[165], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[165], v[30], v[165], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 6))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[168], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[168], 2, v[168]))
+  k.emit(v_add_lshl_u32_e64(v[167], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[167], v[30], v[167], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 7))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[170], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[170], 2, v[170]))
+  k.emit(v_add_lshl_u32_e64(v[169], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[169], v[30], v[169], s[82:83]))
+  k.emit(v_accvgpr_read(v[35], v[0]))
+  k.emit(v_accvgpr_read(v[36], v[4]))
+  k.emit(v_accvgpr_read(v[37], v[8]))
+  k.emit(v_accvgpr_read(v[38], v[12]))
+  k.emit(v_accvgpr_read(v[39], v[16]))
+  k.emit(v_accvgpr_read(v[40], v[20]))
+  k.emit(v_accvgpr_read(v[41], v[24]))
+  k.emit(v_accvgpr_read(v[42], v[28]))
+  k.emit(v_accvgpr_read(v[43], v[32]))
+  k.emit(v_accvgpr_read(v[44], v[36]))
+  k.emit(v_accvgpr_read(v[45], v[40]))
+  k.emit(v_accvgpr_read(v[46], v[44]))
+  k.emit(v_accvgpr_read(v[47], v[48]))
+  k.emit(v_accvgpr_read(v[48], v[52]))
+  k.emit(v_accvgpr_read(v[49], v[56]))
+  k.emit(v_accvgpr_read(v[50], v[60]))
+  k.emit(v_accvgpr_read(v[51], v[64]))
+  k.emit(v_accvgpr_read(v[52], v[68]))
+  k.emit(v_accvgpr_read(v[53], v[72]))
+  k.emit(v_accvgpr_read(v[54], v[76]))
+  k.emit(v_accvgpr_read(v[55], v[80]))
+  k.emit(v_accvgpr_read(v[56], v[84]))
+  k.emit(v_accvgpr_read(v[57], v[88]))
+  k.emit(v_accvgpr_read(v[58], v[92]))
+  k.emit(v_accvgpr_read(v[59], v[96]))
+  k.emit(v_accvgpr_read(v[60], v[100]))
+  k.emit(v_accvgpr_read(v[61], v[104]))
+  k.emit(v_accvgpr_read(v[62], v[108]))
+  k.emit(v_accvgpr_read(v[63], v[112]))
+  k.emit(v_accvgpr_read(v[64], v[116]))
+  k.emit(v_accvgpr_read(v[65], v[120]))
+  k.emit(v_accvgpr_read(v[66], v[124]))
+  k.emit(v_accvgpr_read(v[67], v[128]))
+  k.emit(v_accvgpr_read(v[68], v[132]))
+  k.emit(v_accvgpr_read(v[69], v[136]))
+  k.emit(v_accvgpr_read(v[70], v[140]))
+  k.emit(v_accvgpr_read(v[71], v[144]))
+  k.emit(v_accvgpr_read(v[72], v[148]))
+  k.emit(v_accvgpr_read(v[73], v[152]))
+  k.emit(v_accvgpr_read(v[74], v[156]))
+  k.waitcnt(lgkm=0)
+  k.emit(v_mov_b32_e32(v[32], 4294901760))
+  k.emit(v_mov_b32_e32(v[33], 2147418112))
+  k.emit(v_mov_b32_e32(v[34], 32767))
+  k.emit(v_mul_f32_e32(v[35], v[76], v[35]))
+  k.emit(v_add_f32_e32(v[22], v[75], v[35]))
+  k.emit(v_mov_b32_e32(v[35], v[22]))
+  k.emit(v_cvt_pk(v[35], v[35], v[35]))
+  k.emit(buffer_store_short(v[35], v[77], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[36], v[80], v[36]))
+  k.emit(v_add_f32_e32(v[22], v[79], v[36]))
+  k.emit(v_mov_b32_e32(v[36], v[22]))
+  k.emit(v_cvt_pk(v[36], v[36], v[36]))
+  k.emit(buffer_store_short(v[36], v[81], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[37], v[84], v[37]))
+  k.emit(v_add_f32_e32(v[22], v[83], v[37]))
+  k.emit(v_mov_b32_e32(v[37], v[22]))
+  k.emit(v_cvt_pk(v[37], v[37], v[37]))
+  k.emit(buffer_store_short(v[37], v[85], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[38], v[88], v[38]))
+  k.emit(v_add_f32_e32(v[22], v[87], v[38]))
+  k.emit(v_mov_b32_e32(v[38], v[22]))
+  k.emit(v_cvt_pk(v[38], v[38], v[38]))
+  k.emit(buffer_store_short(v[38], v[89], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[39], v[92], v[39]))
+  k.emit(v_add_f32_e32(v[22], v[91], v[39]))
+  k.emit(v_mov_b32_e32(v[39], v[22]))
+  k.emit(v_cvt_pk(v[39], v[39], v[39]))
+  k.emit(buffer_store_short(v[39], v[93], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[40], v[96], v[40]))
+  k.emit(v_add_f32_e32(v[22], v[95], v[40]))
+  k.emit(v_mov_b32_e32(v[40], v[22]))
+  k.emit(v_cvt_pk(v[40], v[40], v[40]))
+  k.emit(buffer_store_short(v[40], v[97], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[41], v[100], v[41]))
+  k.emit(v_add_f32_e32(v[22], v[99], v[41]))
+  k.emit(v_mov_b32_e32(v[41], v[22]))
+  k.emit(v_cvt_pk(v[41], v[41], v[41]))
+  k.emit(buffer_store_short(v[41], v[101], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[42], v[104], v[42]))
+  k.emit(v_add_f32_e32(v[22], v[103], v[42]))
+  k.emit(v_mov_b32_e32(v[42], v[22]))
+  k.emit(v_cvt_pk(v[42], v[42], v[42]))
+  k.emit(buffer_store_short(v[42], v[105], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[43], v[76], v[43]))
+  k.emit(v_add_f32_e32(v[22], v[75], v[43]))
+  k.emit(v_mov_b32_e32(v[43], v[22]))
+  k.emit(v_cvt_pk(v[43], v[43], v[43]))
+  k.emit(buffer_store_short(v[43], v[107], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[44], v[80], v[44]))
+  k.emit(v_add_f32_e32(v[22], v[79], v[44]))
+  k.emit(v_mov_b32_e32(v[44], v[22]))
+  k.emit(v_cvt_pk(v[44], v[44], v[44]))
+  k.emit(buffer_store_short(v[44], v[109], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[45], v[84], v[45]))
+  k.emit(v_add_f32_e32(v[22], v[83], v[45]))
+  k.emit(v_mov_b32_e32(v[45], v[22]))
+  k.emit(v_cvt_pk(v[45], v[45], v[45]))
+  k.emit(buffer_store_short(v[45], v[111], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[46], v[88], v[46]))
+  k.emit(v_add_f32_e32(v[22], v[87], v[46]))
+  k.emit(v_mov_b32_e32(v[46], v[22]))
+  k.emit(v_cvt_pk(v[46], v[46], v[46]))
+  k.emit(buffer_store_short(v[46], v[113], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[47], v[92], v[47]))
+  k.emit(v_add_f32_e32(v[22], v[91], v[47]))
+  k.emit(v_mov_b32_e32(v[47], v[22]))
+  k.emit(v_cvt_pk(v[47], v[47], v[47]))
+  k.emit(buffer_store_short(v[47], v[115], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[48], v[96], v[48]))
+  k.emit(v_add_f32_e32(v[22], v[95], v[48]))
+  k.emit(v_mov_b32_e32(v[48], v[22]))
+  k.emit(v_cvt_pk(v[48], v[48], v[48]))
+  k.emit(buffer_store_short(v[48], v[117], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[49], v[100], v[49]))
+  k.emit(v_add_f32_e32(v[22], v[99], v[49]))
+  k.emit(v_mov_b32_e32(v[49], v[22]))
+  k.emit(v_cvt_pk(v[49], v[49], v[49]))
+  k.emit(buffer_store_short(v[49], v[119], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[50], v[104], v[50]))
+  k.emit(v_add_f32_e32(v[22], v[103], v[50]))
+  k.emit(v_mov_b32_e32(v[50], v[22]))
+  k.emit(v_cvt_pk(v[50], v[50], v[50]))
+  k.emit(buffer_store_short(v[50], v[121], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[51], v[76], v[51]))
+  k.emit(v_add_f32_e32(v[22], v[75], v[51]))
+  k.emit(v_mov_b32_e32(v[51], v[22]))
+  k.emit(v_cvt_pk(v[51], v[51], v[51]))
+  k.emit(buffer_store_short(v[51], v[123], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[52], v[80], v[52]))
+  k.emit(v_add_f32_e32(v[22], v[79], v[52]))
+  k.emit(v_mov_b32_e32(v[52], v[22]))
+  k.emit(v_cvt_pk(v[52], v[52], v[52]))
+  k.emit(buffer_store_short(v[52], v[125], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[53], v[84], v[53]))
+  k.emit(v_add_f32_e32(v[22], v[83], v[53]))
+  k.emit(v_mov_b32_e32(v[53], v[22]))
+  k.emit(v_cvt_pk(v[53], v[53], v[53]))
+  k.emit(buffer_store_short(v[53], v[127], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[54], v[88], v[54]))
+  k.emit(v_add_f32_e32(v[22], v[87], v[54]))
+  k.emit(v_mov_b32_e32(v[54], v[22]))
+  k.emit(v_cvt_pk(v[54], v[54], v[54]))
+  k.emit(buffer_store_short(v[54], v[129], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[55], v[92], v[55]))
+  k.emit(v_add_f32_e32(v[22], v[91], v[55]))
+  k.emit(v_mov_b32_e32(v[55], v[22]))
+  k.emit(v_cvt_pk(v[55], v[55], v[55]))
+  k.emit(buffer_store_short(v[55], v[131], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[56], v[96], v[56]))
+  k.emit(v_add_f32_e32(v[22], v[95], v[56]))
+  k.emit(v_mov_b32_e32(v[56], v[22]))
+  k.emit(v_cvt_pk(v[56], v[56], v[56]))
+  k.emit(buffer_store_short(v[56], v[133], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[57], v[100], v[57]))
+  k.emit(v_add_f32_e32(v[22], v[99], v[57]))
+  k.emit(v_mov_b32_e32(v[57], v[22]))
+  k.emit(v_cvt_pk(v[57], v[57], v[57]))
+  k.emit(buffer_store_short(v[57], v[135], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[58], v[104], v[58]))
+  k.emit(v_add_f32_e32(v[22], v[103], v[58]))
+  k.emit(v_mov_b32_e32(v[58], v[22]))
+  k.emit(v_cvt_pk(v[58], v[58], v[58]))
+  k.emit(buffer_store_short(v[58], v[137], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[59], v[76], v[59]))
+  k.emit(v_add_f32_e32(v[22], v[75], v[59]))
+  k.emit(v_mov_b32_e32(v[59], v[22]))
+  k.emit(v_cvt_pk(v[59], v[59], v[59]))
+  k.emit(buffer_store_short(v[59], v[139], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[60], v[80], v[60]))
+  k.emit(v_add_f32_e32(v[22], v[79], v[60]))
+  k.emit(v_mov_b32_e32(v[60], v[22]))
+  k.emit(v_cvt_pk(v[60], v[60], v[60]))
+  k.emit(buffer_store_short(v[60], v[141], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[61], v[84], v[61]))
+  k.emit(v_add_f32_e32(v[22], v[83], v[61]))
+  k.emit(v_mov_b32_e32(v[61], v[22]))
+  k.emit(v_cvt_pk(v[61], v[61], v[61]))
+  k.emit(buffer_store_short(v[61], v[143], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[62], v[88], v[62]))
+  k.emit(v_add_f32_e32(v[22], v[87], v[62]))
+  k.emit(v_mov_b32_e32(v[62], v[22]))
+  k.emit(v_cvt_pk(v[62], v[62], v[62]))
+  k.emit(buffer_store_short(v[62], v[145], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[63], v[92], v[63]))
+  k.emit(v_add_f32_e32(v[22], v[91], v[63]))
+  k.emit(v_mov_b32_e32(v[63], v[22]))
+  k.emit(v_cvt_pk(v[63], v[63], v[63]))
+  k.emit(buffer_store_short(v[63], v[147], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[64], v[96], v[64]))
+  k.emit(v_add_f32_e32(v[22], v[95], v[64]))
+  k.emit(v_mov_b32_e32(v[64], v[22]))
+  k.emit(v_cvt_pk(v[64], v[64], v[64]))
+  k.emit(buffer_store_short(v[64], v[149], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[65], v[100], v[65]))
+  k.emit(v_add_f32_e32(v[22], v[99], v[65]))
+  k.emit(v_mov_b32_e32(v[65], v[22]))
+  k.emit(v_cvt_pk(v[65], v[65], v[65]))
+  k.emit(buffer_store_short(v[65], v[151], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[66], v[104], v[66]))
+  k.emit(v_add_f32_e32(v[22], v[103], v[66]))
+  k.emit(v_mov_b32_e32(v[66], v[22]))
+  k.emit(v_cvt_pk(v[66], v[66], v[66]))
+  k.emit(buffer_store_short(v[66], v[153], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[67], v[76], v[67]))
+  k.emit(v_add_f32_e32(v[22], v[75], v[67]))
+  k.emit(v_mov_b32_e32(v[67], v[22]))
+  k.emit(v_cvt_pk(v[67], v[67], v[67]))
+  k.emit(buffer_store_short(v[67], v[155], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[68], v[80], v[68]))
+  k.emit(v_add_f32_e32(v[22], v[79], v[68]))
+  k.emit(v_mov_b32_e32(v[68], v[22]))
+  k.emit(v_cvt_pk(v[68], v[68], v[68]))
+  k.emit(buffer_store_short(v[68], v[157], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[69], v[84], v[69]))
+  k.emit(v_add_f32_e32(v[22], v[83], v[69]))
+  k.emit(v_mov_b32_e32(v[69], v[22]))
+  k.emit(v_cvt_pk(v[69], v[69], v[69]))
+  k.emit(buffer_store_short(v[69], v[159], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[70], v[88], v[70]))
+  k.emit(v_add_f32_e32(v[22], v[87], v[70]))
+  k.emit(v_mov_b32_e32(v[70], v[22]))
+  k.emit(v_cvt_pk(v[70], v[70], v[70]))
+  k.emit(buffer_store_short(v[70], v[161], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[71], v[92], v[71]))
+  k.emit(v_add_f32_e32(v[22], v[91], v[71]))
+  k.emit(v_mov_b32_e32(v[71], v[22]))
+  k.emit(v_cvt_pk(v[71], v[71], v[71]))
+  k.emit(buffer_store_short(v[71], v[163], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[72], v[96], v[72]))
+  k.emit(v_add_f32_e32(v[22], v[95], v[72]))
+  k.emit(v_mov_b32_e32(v[72], v[22]))
+  k.emit(v_cvt_pk(v[72], v[72], v[72]))
+  k.emit(buffer_store_short(v[72], v[165], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[73], v[100], v[73]))
+  k.emit(v_add_f32_e32(v[22], v[99], v[73]))
+  k.emit(v_mov_b32_e32(v[73], v[22]))
+  k.emit(v_cvt_pk(v[73], v[73], v[73]))
+  k.emit(buffer_store_short(v[73], v[167], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[74], v[104], v[74]))
+  k.emit(v_add_f32_e32(v[22], v[103], v[74]))
+  k.emit(v_mov_b32_e32(v[74], v[22]))
+  k.emit(v_cvt_pk(v[74], v[74], v[74]))
+  k.emit(buffer_store_short(v[74], v[169], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_nop())
+  k.emit(v_mov_b32_e32(v[30], 2147483648))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[78], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[78], 2, v[78]))
+  k.emit(ds_read_b32(v[75], v[78]))
+  k.emit(ds_read_b32(v[76], v[78], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[77], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[77], v[30], v[77], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 1))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[82], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[82], 2, v[82]))
+  k.emit(ds_read_b32(v[79], v[82]))
+  k.emit(ds_read_b32(v[80], v[82], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[81], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[81], v[30], v[81], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 2))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[86], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[86], 2, v[86]))
+  k.emit(ds_read_b32(v[83], v[86]))
+  k.emit(ds_read_b32(v[84], v[86], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[85], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[85], v[30], v[85], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 3))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[90], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[90], 2, v[90]))
+  k.emit(ds_read_b32(v[87], v[90]))
+  k.emit(ds_read_b32(v[88], v[90], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[89], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[89], v[30], v[89], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 4))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[94], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[94], 2, v[94]))
+  k.emit(ds_read_b32(v[91], v[94]))
+  k.emit(ds_read_b32(v[92], v[94], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[93], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[93], v[30], v[93], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 5))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[98], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[98], 2, v[98]))
+  k.emit(ds_read_b32(v[95], v[98]))
+  k.emit(ds_read_b32(v[96], v[98], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[97], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[97], v[30], v[97], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 6))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[102], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[102], 2, v[102]))
+  k.emit(ds_read_b32(v[99], v[102]))
+  k.emit(ds_read_b32(v[100], v[102], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[101], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[101], v[30], v[101], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 7))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[106], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[106], 2, v[106]))
+  k.emit(ds_read_b32(v[103], v[106]))
+  k.emit(ds_read_b32(v[104], v[106], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[105], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[105], v[30], v[105], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[108], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[108], 2, v[108]))
+  k.emit(v_add_lshl_u32_e64(v[107], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[107], v[30], v[107], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 1))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[110], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[110], 2, v[110]))
+  k.emit(v_add_lshl_u32_e64(v[109], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[109], v[30], v[109], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 2))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[112], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[112], 2, v[112]))
+  k.emit(v_add_lshl_u32_e64(v[111], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[111], v[30], v[111], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 3))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[114], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[114], 2, v[114]))
+  k.emit(v_add_lshl_u32_e64(v[113], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[113], v[30], v[113], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 4))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[116], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[116], 2, v[116]))
+  k.emit(v_add_lshl_u32_e64(v[115], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[115], v[30], v[115], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 5))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[118], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[118], 2, v[118]))
+  k.emit(v_add_lshl_u32_e64(v[117], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[117], v[30], v[117], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 6))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[120], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[120], 2, v[120]))
+  k.emit(v_add_lshl_u32_e64(v[119], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[119], v[30], v[119], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 7))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[122], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[122], 2, v[122]))
+  k.emit(v_add_lshl_u32_e64(v[121], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[121], v[30], v[121], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[124], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[124], 2, v[124]))
+  k.emit(v_add_lshl_u32_e64(v[123], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[123], v[30], v[123], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 1))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[126], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[126], 2, v[126]))
+  k.emit(v_add_lshl_u32_e64(v[125], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[125], v[30], v[125], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 2))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[128], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[128], 2, v[128]))
+  k.emit(v_add_lshl_u32_e64(v[127], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[127], v[30], v[127], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 3))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[130], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[130], 2, v[130]))
+  k.emit(v_add_lshl_u32_e64(v[129], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[129], v[30], v[129], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 4))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[132], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[132], 2, v[132]))
+  k.emit(v_add_lshl_u32_e64(v[131], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[131], v[30], v[131], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 5))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[134], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[134], 2, v[134]))
+  k.emit(v_add_lshl_u32_e64(v[133], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[133], v[30], v[133], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 6))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[136], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[136], 2, v[136]))
+  k.emit(v_add_lshl_u32_e64(v[135], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[135], v[30], v[135], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 7))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[138], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[138], 2, v[138]))
+  k.emit(v_add_lshl_u32_e64(v[137], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[137], v[30], v[137], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[140], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[140], 2, v[140]))
+  k.emit(v_add_lshl_u32_e64(v[139], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[139], v[30], v[139], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 1))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[142], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[142], 2, v[142]))
+  k.emit(v_add_lshl_u32_e64(v[141], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[141], v[30], v[141], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 2))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[144], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[144], 2, v[144]))
+  k.emit(v_add_lshl_u32_e64(v[143], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[143], v[30], v[143], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 3))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[146], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[146], 2, v[146]))
+  k.emit(v_add_lshl_u32_e64(v[145], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[145], v[30], v[145], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 4))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[148], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[148], 2, v[148]))
+  k.emit(v_add_lshl_u32_e64(v[147], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[147], v[30], v[147], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 5))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[150], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[150], 2, v[150]))
+  k.emit(v_add_lshl_u32_e64(v[149], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[149], v[30], v[149], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 6))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[152], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[152], 2, v[152]))
+  k.emit(v_add_lshl_u32_e64(v[151], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[151], v[30], v[151], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 7))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[154], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[154], 2, v[154]))
+  k.emit(v_add_lshl_u32_e64(v[153], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[153], v[30], v[153], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[156], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[156], 2, v[156]))
+  k.emit(v_add_lshl_u32_e64(v[155], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[155], v[30], v[155], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 1))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[158], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[158], 2, v[158]))
+  k.emit(v_add_lshl_u32_e64(v[157], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[157], v[30], v[157], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 2))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[160], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[160], 2, v[160]))
+  k.emit(v_add_lshl_u32_e64(v[159], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[159], v[30], v[159], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 3))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[162], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[162], 2, v[162]))
+  k.emit(v_add_lshl_u32_e64(v[161], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[161], v[30], v[161], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 4))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[164], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[164], 2, v[164]))
+  k.emit(v_add_lshl_u32_e64(v[163], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[163], v[30], v[163], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 5))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[166], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[166], 2, v[166]))
+  k.emit(v_add_lshl_u32_e64(v[165], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[165], v[30], v[165], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 6))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[168], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[168], 2, v[168]))
+  k.emit(v_add_lshl_u32_e64(v[167], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[167], v[30], v[167], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 7))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[170], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[170], 2, v[170]))
+  k.emit(v_add_lshl_u32_e64(v[169], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[169], v[30], v[169], s[82:83]))
+  k.emit(v_accvgpr_read(v[35], v[160]))
+  k.emit(v_accvgpr_read(v[36], v[164]))
+  k.emit(v_accvgpr_read(v[37], v[168]))
+  k.emit(v_accvgpr_read(v[38], v[172]))
+  k.emit(v_accvgpr_read(v[39], v[176]))
+  k.emit(v_accvgpr_read(v[40], v[180]))
+  k.emit(v_accvgpr_read(v[41], v[184]))
+  k.emit(v_accvgpr_read(v[42], v[188]))
+  k.emit(v_accvgpr_read(v[43], v[192]))
+  k.emit(v_accvgpr_read(v[44], v[196]))
+  k.emit(v_accvgpr_read(v[45], v[200]))
+  k.emit(v_accvgpr_read(v[46], v[204]))
+  k.emit(v_accvgpr_read(v[47], v[208]))
+  k.emit(v_accvgpr_read(v[48], v[212]))
+  k.emit(v_accvgpr_read(v[49], v[216]))
+  k.emit(v_accvgpr_read(v[50], v[220]))
+  k.emit(v_accvgpr_read(v[51], v[224]))
+  k.emit(v_accvgpr_read(v[52], v[228]))
+  k.emit(v_accvgpr_read(v[53], v[232]))
+  k.emit(v_accvgpr_read(v[54], v[236]))
+  k.emit(v_accvgpr_read(v[55], v[240]))
+  k.emit(v_accvgpr_read(v[56], v[244]))
+  k.emit(v_accvgpr_read(v[57], v[248]))
+  k.emit(v_accvgpr_read(v[58], v[252]))
+  k.emit(v_accvgpr_read(v[59], v[1]))
+  k.emit(v_accvgpr_read(v[60], v[5]))
+  k.emit(v_accvgpr_read(v[61], v[9]))
+  k.emit(v_accvgpr_read(v[62], v[13]))
+  k.emit(v_accvgpr_read(v[63], v[17]))
+  k.emit(v_accvgpr_read(v[64], v[21]))
+  k.emit(v_accvgpr_read(v[65], v[25]))
+  k.emit(v_accvgpr_read(v[66], v[29]))
+  k.emit(v_accvgpr_read(v[67], v[33]))
+  k.emit(v_accvgpr_read(v[68], v[37]))
+  k.emit(v_accvgpr_read(v[69], v[41]))
+  k.emit(v_accvgpr_read(v[70], v[45]))
+  k.emit(v_accvgpr_read(v[71], v[49]))
+  k.emit(v_accvgpr_read(v[72], v[53]))
+  k.emit(v_accvgpr_read(v[73], v[57]))
+  k.emit(v_accvgpr_read(v[74], v[61]))
+  k.waitcnt(lgkm=0)
+  k.emit(v_mov_b32_e32(v[32], 4294901760))
+  k.emit(v_mov_b32_e32(v[33], 2147418112))
+  k.emit(v_mov_b32_e32(v[34], 32767))
+  k.emit(v_mul_f32_e32(v[35], v[76], v[35]))
+  k.emit(v_add_f32_e32(v[22], v[75], v[35]))
+  k.emit(v_mov_b32_e32(v[35], v[22]))
+  k.emit(v_cvt_pk(v[35], v[35], v[35]))
+  k.emit(buffer_store_short(v[35], v[77], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[36], v[80], v[36]))
+  k.emit(v_add_f32_e32(v[22], v[79], v[36]))
+  k.emit(v_mov_b32_e32(v[36], v[22]))
+  k.emit(v_cvt_pk(v[36], v[36], v[36]))
+  k.emit(buffer_store_short(v[36], v[81], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[37], v[84], v[37]))
+  k.emit(v_add_f32_e32(v[22], v[83], v[37]))
+  k.emit(v_mov_b32_e32(v[37], v[22]))
+  k.emit(v_cvt_pk(v[37], v[37], v[37]))
+  k.emit(buffer_store_short(v[37], v[85], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[38], v[88], v[38]))
+  k.emit(v_add_f32_e32(v[22], v[87], v[38]))
+  k.emit(v_mov_b32_e32(v[38], v[22]))
+  k.emit(v_cvt_pk(v[38], v[38], v[38]))
+  k.emit(buffer_store_short(v[38], v[89], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[39], v[92], v[39]))
+  k.emit(v_add_f32_e32(v[22], v[91], v[39]))
+  k.emit(v_mov_b32_e32(v[39], v[22]))
+  k.emit(v_cvt_pk(v[39], v[39], v[39]))
+  k.emit(buffer_store_short(v[39], v[93], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[40], v[96], v[40]))
+  k.emit(v_add_f32_e32(v[22], v[95], v[40]))
+  k.emit(v_mov_b32_e32(v[40], v[22]))
+  k.emit(v_cvt_pk(v[40], v[40], v[40]))
+  k.emit(buffer_store_short(v[40], v[97], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[41], v[100], v[41]))
+  k.emit(v_add_f32_e32(v[22], v[99], v[41]))
+  k.emit(v_mov_b32_e32(v[41], v[22]))
+  k.emit(v_cvt_pk(v[41], v[41], v[41]))
+  k.emit(buffer_store_short(v[41], v[101], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[42], v[104], v[42]))
+  k.emit(v_add_f32_e32(v[22], v[103], v[42]))
+  k.emit(v_mov_b32_e32(v[42], v[22]))
+  k.emit(v_cvt_pk(v[42], v[42], v[42]))
+  k.emit(buffer_store_short(v[42], v[105], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[43], v[76], v[43]))
+  k.emit(v_add_f32_e32(v[22], v[75], v[43]))
+  k.emit(v_mov_b32_e32(v[43], v[22]))
+  k.emit(v_cvt_pk(v[43], v[43], v[43]))
+  k.emit(buffer_store_short(v[43], v[107], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[44], v[80], v[44]))
+  k.emit(v_add_f32_e32(v[22], v[79], v[44]))
+  k.emit(v_mov_b32_e32(v[44], v[22]))
+  k.emit(v_cvt_pk(v[44], v[44], v[44]))
+  k.emit(buffer_store_short(v[44], v[109], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[45], v[84], v[45]))
+  k.emit(v_add_f32_e32(v[22], v[83], v[45]))
+  k.emit(v_mov_b32_e32(v[45], v[22]))
+  k.emit(v_cvt_pk(v[45], v[45], v[45]))
+  k.emit(buffer_store_short(v[45], v[111], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[46], v[88], v[46]))
+  k.emit(v_add_f32_e32(v[22], v[87], v[46]))
+  k.emit(v_mov_b32_e32(v[46], v[22]))
+  k.emit(v_cvt_pk(v[46], v[46], v[46]))
+  k.emit(buffer_store_short(v[46], v[113], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[47], v[92], v[47]))
+  k.emit(v_add_f32_e32(v[22], v[91], v[47]))
+  k.emit(v_mov_b32_e32(v[47], v[22]))
+  k.emit(v_cvt_pk(v[47], v[47], v[47]))
+  k.emit(buffer_store_short(v[47], v[115], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[48], v[96], v[48]))
+  k.emit(v_add_f32_e32(v[22], v[95], v[48]))
+  k.emit(v_mov_b32_e32(v[48], v[22]))
+  k.emit(v_cvt_pk(v[48], v[48], v[48]))
+  k.emit(buffer_store_short(v[48], v[117], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[49], v[100], v[49]))
+  k.emit(v_add_f32_e32(v[22], v[99], v[49]))
+  k.emit(v_mov_b32_e32(v[49], v[22]))
+  k.emit(v_cvt_pk(v[49], v[49], v[49]))
+  k.emit(buffer_store_short(v[49], v[119], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[50], v[104], v[50]))
+  k.emit(v_add_f32_e32(v[22], v[103], v[50]))
+  k.emit(v_mov_b32_e32(v[50], v[22]))
+  k.emit(v_cvt_pk(v[50], v[50], v[50]))
+  k.emit(buffer_store_short(v[50], v[121], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[51], v[76], v[51]))
+  k.emit(v_add_f32_e32(v[22], v[75], v[51]))
+  k.emit(v_mov_b32_e32(v[51], v[22]))
+  k.emit(v_cvt_pk(v[51], v[51], v[51]))
+  k.emit(buffer_store_short(v[51], v[123], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[52], v[80], v[52]))
+  k.emit(v_add_f32_e32(v[22], v[79], v[52]))
+  k.emit(v_mov_b32_e32(v[52], v[22]))
+  k.emit(v_cvt_pk(v[52], v[52], v[52]))
+  k.emit(buffer_store_short(v[52], v[125], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[53], v[84], v[53]))
+  k.emit(v_add_f32_e32(v[22], v[83], v[53]))
+  k.emit(v_mov_b32_e32(v[53], v[22]))
+  k.emit(v_cvt_pk(v[53], v[53], v[53]))
+  k.emit(buffer_store_short(v[53], v[127], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[54], v[88], v[54]))
+  k.emit(v_add_f32_e32(v[22], v[87], v[54]))
+  k.emit(v_mov_b32_e32(v[54], v[22]))
+  k.emit(v_cvt_pk(v[54], v[54], v[54]))
+  k.emit(buffer_store_short(v[54], v[129], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[55], v[92], v[55]))
+  k.emit(v_add_f32_e32(v[22], v[91], v[55]))
+  k.emit(v_mov_b32_e32(v[55], v[22]))
+  k.emit(v_cvt_pk(v[55], v[55], v[55]))
+  k.emit(buffer_store_short(v[55], v[131], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[56], v[96], v[56]))
+  k.emit(v_add_f32_e32(v[22], v[95], v[56]))
+  k.emit(v_mov_b32_e32(v[56], v[22]))
+  k.emit(v_cvt_pk(v[56], v[56], v[56]))
+  k.emit(buffer_store_short(v[56], v[133], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[57], v[100], v[57]))
+  k.emit(v_add_f32_e32(v[22], v[99], v[57]))
+  k.emit(v_mov_b32_e32(v[57], v[22]))
+  k.emit(v_cvt_pk(v[57], v[57], v[57]))
+  k.emit(buffer_store_short(v[57], v[135], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[58], v[104], v[58]))
+  k.emit(v_add_f32_e32(v[22], v[103], v[58]))
+  k.emit(v_mov_b32_e32(v[58], v[22]))
+  k.emit(v_cvt_pk(v[58], v[58], v[58]))
+  k.emit(buffer_store_short(v[58], v[137], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[59], v[76], v[59]))
+  k.emit(v_add_f32_e32(v[22], v[75], v[59]))
+  k.emit(v_mov_b32_e32(v[59], v[22]))
+  k.emit(v_cvt_pk(v[59], v[59], v[59]))
+  k.emit(buffer_store_short(v[59], v[139], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[60], v[80], v[60]))
+  k.emit(v_add_f32_e32(v[22], v[79], v[60]))
+  k.emit(v_mov_b32_e32(v[60], v[22]))
+  k.emit(v_cvt_pk(v[60], v[60], v[60]))
+  k.emit(buffer_store_short(v[60], v[141], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[61], v[84], v[61]))
+  k.emit(v_add_f32_e32(v[22], v[83], v[61]))
+  k.emit(v_mov_b32_e32(v[61], v[22]))
+  k.emit(v_cvt_pk(v[61], v[61], v[61]))
+  k.emit(buffer_store_short(v[61], v[143], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[62], v[88], v[62]))
+  k.emit(v_add_f32_e32(v[22], v[87], v[62]))
+  k.emit(v_mov_b32_e32(v[62], v[22]))
+  k.emit(v_cvt_pk(v[62], v[62], v[62]))
+  k.emit(buffer_store_short(v[62], v[145], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[63], v[92], v[63]))
+  k.emit(v_add_f32_e32(v[22], v[91], v[63]))
+  k.emit(v_mov_b32_e32(v[63], v[22]))
+  k.emit(v_cvt_pk(v[63], v[63], v[63]))
+  k.emit(buffer_store_short(v[63], v[147], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[64], v[96], v[64]))
+  k.emit(v_add_f32_e32(v[22], v[95], v[64]))
+  k.emit(v_mov_b32_e32(v[64], v[22]))
+  k.emit(v_cvt_pk(v[64], v[64], v[64]))
+  k.emit(buffer_store_short(v[64], v[149], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[65], v[100], v[65]))
+  k.emit(v_add_f32_e32(v[22], v[99], v[65]))
+  k.emit(v_mov_b32_e32(v[65], v[22]))
+  k.emit(v_cvt_pk(v[65], v[65], v[65]))
+  k.emit(buffer_store_short(v[65], v[151], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[66], v[104], v[66]))
+  k.emit(v_add_f32_e32(v[22], v[103], v[66]))
+  k.emit(v_mov_b32_e32(v[66], v[22]))
+  k.emit(v_cvt_pk(v[66], v[66], v[66]))
+  k.emit(buffer_store_short(v[66], v[153], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[67], v[76], v[67]))
+  k.emit(v_add_f32_e32(v[22], v[75], v[67]))
+  k.emit(v_mov_b32_e32(v[67], v[22]))
+  k.emit(v_cvt_pk(v[67], v[67], v[67]))
+  k.emit(buffer_store_short(v[67], v[155], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[68], v[80], v[68]))
+  k.emit(v_add_f32_e32(v[22], v[79], v[68]))
+  k.emit(v_mov_b32_e32(v[68], v[22]))
+  k.emit(v_cvt_pk(v[68], v[68], v[68]))
+  k.emit(buffer_store_short(v[68], v[157], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[69], v[84], v[69]))
+  k.emit(v_add_f32_e32(v[22], v[83], v[69]))
+  k.emit(v_mov_b32_e32(v[69], v[22]))
+  k.emit(v_cvt_pk(v[69], v[69], v[69]))
+  k.emit(buffer_store_short(v[69], v[159], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[70], v[88], v[70]))
+  k.emit(v_add_f32_e32(v[22], v[87], v[70]))
+  k.emit(v_mov_b32_e32(v[70], v[22]))
+  k.emit(v_cvt_pk(v[70], v[70], v[70]))
+  k.emit(buffer_store_short(v[70], v[161], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[71], v[92], v[71]))
+  k.emit(v_add_f32_e32(v[22], v[91], v[71]))
+  k.emit(v_mov_b32_e32(v[71], v[22]))
+  k.emit(v_cvt_pk(v[71], v[71], v[71]))
+  k.emit(buffer_store_short(v[71], v[163], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[72], v[96], v[72]))
+  k.emit(v_add_f32_e32(v[22], v[95], v[72]))
+  k.emit(v_mov_b32_e32(v[72], v[22]))
+  k.emit(v_cvt_pk(v[72], v[72], v[72]))
+  k.emit(buffer_store_short(v[72], v[165], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[73], v[100], v[73]))
+  k.emit(v_add_f32_e32(v[22], v[99], v[73]))
+  k.emit(v_mov_b32_e32(v[73], v[22]))
+  k.emit(v_cvt_pk(v[73], v[73], v[73]))
+  k.emit(buffer_store_short(v[73], v[167], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[74], v[104], v[74]))
+  k.emit(v_add_f32_e32(v[22], v[103], v[74]))
+  k.emit(v_mov_b32_e32(v[74], v[22]))
+  k.emit(v_cvt_pk(v[74], v[74], v[74]))
+  k.emit(buffer_store_short(v[74], v[169], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_nop())
+  k.emit(v_mov_b32_e32(v[30], 2147483648))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[78], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[78], 2, v[78]))
+  k.emit(ds_read_b32(v[75], v[78]))
+  k.emit(ds_read_b32(v[76], v[78], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[77], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[77], v[30], v[77], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 1))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[82], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[82], 2, v[82]))
+  k.emit(ds_read_b32(v[79], v[82]))
+  k.emit(ds_read_b32(v[80], v[82], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[81], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[81], v[30], v[81], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 2))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[86], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[86], 2, v[86]))
+  k.emit(ds_read_b32(v[83], v[86]))
+  k.emit(ds_read_b32(v[84], v[86], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[85], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[85], v[30], v[85], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 3))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[90], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[90], 2, v[90]))
+  k.emit(ds_read_b32(v[87], v[90]))
+  k.emit(ds_read_b32(v[88], v[90], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[89], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[89], v[30], v[89], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 4))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[94], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[94], 2, v[94]))
+  k.emit(ds_read_b32(v[91], v[94]))
+  k.emit(ds_read_b32(v[92], v[94], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[93], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[93], v[30], v[93], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 5))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[98], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[98], 2, v[98]))
+  k.emit(ds_read_b32(v[95], v[98]))
+  k.emit(ds_read_b32(v[96], v[98], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[97], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[97], v[30], v[97], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 6))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[102], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[102], 2, v[102]))
+  k.emit(ds_read_b32(v[99], v[102]))
+  k.emit(ds_read_b32(v[100], v[102], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[101], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[101], v[30], v[101], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 7))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[106], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[106], 2, v[106]))
+  k.emit(ds_read_b32(v[103], v[106]))
+  k.emit(ds_read_b32(v[104], v[106], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[105], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[105], v[30], v[105], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[108], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[108], 2, v[108]))
+  k.emit(v_add_lshl_u32_e64(v[107], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[107], v[30], v[107], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 1))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[110], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[110], 2, v[110]))
+  k.emit(v_add_lshl_u32_e64(v[109], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[109], v[30], v[109], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 2))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[112], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[112], 2, v[112]))
+  k.emit(v_add_lshl_u32_e64(v[111], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[111], v[30], v[111], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 3))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[114], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[114], 2, v[114]))
+  k.emit(v_add_lshl_u32_e64(v[113], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[113], v[30], v[113], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 4))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[116], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[116], 2, v[116]))
+  k.emit(v_add_lshl_u32_e64(v[115], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[115], v[30], v[115], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 5))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[118], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[118], 2, v[118]))
+  k.emit(v_add_lshl_u32_e64(v[117], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[117], v[30], v[117], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 6))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[120], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[120], 2, v[120]))
+  k.emit(v_add_lshl_u32_e64(v[119], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[119], v[30], v[119], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 7))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[122], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[122], 2, v[122]))
+  k.emit(v_add_lshl_u32_e64(v[121], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[121], v[30], v[121], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[124], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[124], 2, v[124]))
+  k.emit(v_add_lshl_u32_e64(v[123], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[123], v[30], v[123], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 1))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[126], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[126], 2, v[126]))
+  k.emit(v_add_lshl_u32_e64(v[125], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[125], v[30], v[125], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 2))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[128], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[128], 2, v[128]))
+  k.emit(v_add_lshl_u32_e64(v[127], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[127], v[30], v[127], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 3))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[130], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[130], 2, v[130]))
+  k.emit(v_add_lshl_u32_e64(v[129], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[129], v[30], v[129], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 4))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[132], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[132], 2, v[132]))
+  k.emit(v_add_lshl_u32_e64(v[131], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[131], v[30], v[131], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 5))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[134], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[134], 2, v[134]))
+  k.emit(v_add_lshl_u32_e64(v[133], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[133], v[30], v[133], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 6))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[136], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[136], 2, v[136]))
+  k.emit(v_add_lshl_u32_e64(v[135], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[135], v[30], v[135], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 7))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[138], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[138], 2, v[138]))
+  k.emit(v_add_lshl_u32_e64(v[137], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[137], v[30], v[137], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[140], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[140], 2, v[140]))
+  k.emit(v_add_lshl_u32_e64(v[139], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[139], v[30], v[139], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 1))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[142], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[142], 2, v[142]))
+  k.emit(v_add_lshl_u32_e64(v[141], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[141], v[30], v[141], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 2))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[144], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[144], 2, v[144]))
+  k.emit(v_add_lshl_u32_e64(v[143], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[143], v[30], v[143], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 3))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[146], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[146], 2, v[146]))
+  k.emit(v_add_lshl_u32_e64(v[145], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[145], v[30], v[145], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 4))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[148], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[148], 2, v[148]))
+  k.emit(v_add_lshl_u32_e64(v[147], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[147], v[30], v[147], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 5))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[150], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[150], 2, v[150]))
+  k.emit(v_add_lshl_u32_e64(v[149], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[149], v[30], v[149], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 6))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[152], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[152], 2, v[152]))
+  k.emit(v_add_lshl_u32_e64(v[151], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[151], v[30], v[151], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 7))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[154], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[154], 2, v[154]))
+  k.emit(v_add_lshl_u32_e64(v[153], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[153], v[30], v[153], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[156], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[156], 2, v[156]))
+  k.emit(v_add_lshl_u32_e64(v[155], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[155], v[30], v[155], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 1))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[158], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[158], 2, v[158]))
+  k.emit(v_add_lshl_u32_e64(v[157], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[157], v[30], v[157], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 2))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[160], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[160], 2, v[160]))
+  k.emit(v_add_lshl_u32_e64(v[159], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[159], v[30], v[159], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 3))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[162], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[162], 2, v[162]))
+  k.emit(v_add_lshl_u32_e64(v[161], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[161], v[30], v[161], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 4))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[164], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[164], 2, v[164]))
+  k.emit(v_add_lshl_u32_e64(v[163], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[163], v[30], v[163], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 5))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[166], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[166], 2, v[166]))
+  k.emit(v_add_lshl_u32_e64(v[165], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[165], v[30], v[165], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 6))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[168], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[168], 2, v[168]))
+  k.emit(v_add_lshl_u32_e64(v[167], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[167], v[30], v[167], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 7))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[170], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[170], 2, v[170]))
+  k.emit(v_add_lshl_u32_e64(v[169], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[169], v[30], v[169], s[82:83]))
+  k.emit(v_accvgpr_read(v[35], v[65]))
+  k.emit(v_accvgpr_read(v[36], v[69]))
+  k.emit(v_accvgpr_read(v[37], v[73]))
+  k.emit(v_accvgpr_read(v[38], v[77]))
+  k.emit(v_accvgpr_read(v[39], v[81]))
+  k.emit(v_accvgpr_read(v[40], v[85]))
+  k.emit(v_accvgpr_read(v[41], v[89]))
+  k.emit(v_accvgpr_read(v[42], v[93]))
+  k.emit(v_accvgpr_read(v[43], v[97]))
+  k.emit(v_accvgpr_read(v[44], v[101]))
+  k.emit(v_accvgpr_read(v[45], v[105]))
+  k.emit(v_accvgpr_read(v[46], v[109]))
+  k.emit(v_accvgpr_read(v[47], v[113]))
+  k.emit(v_accvgpr_read(v[48], v[117]))
+  k.emit(v_accvgpr_read(v[49], v[121]))
+  k.emit(v_accvgpr_read(v[50], v[125]))
+  k.emit(v_accvgpr_read(v[51], v[129]))
+  k.emit(v_accvgpr_read(v[52], v[133]))
+  k.emit(v_accvgpr_read(v[53], v[137]))
+  k.emit(v_accvgpr_read(v[54], v[141]))
+  k.emit(v_accvgpr_read(v[55], v[145]))
+  k.emit(v_accvgpr_read(v[56], v[149]))
+  k.emit(v_accvgpr_read(v[57], v[153]))
+  k.emit(v_accvgpr_read(v[58], v[157]))
+  k.emit(v_accvgpr_read(v[59], v[161]))
+  k.emit(v_accvgpr_read(v[60], v[165]))
+  k.emit(v_accvgpr_read(v[61], v[169]))
+  k.emit(v_accvgpr_read(v[62], v[173]))
+  k.emit(v_accvgpr_read(v[63], v[177]))
+  k.emit(v_accvgpr_read(v[64], v[181]))
+  k.emit(v_accvgpr_read(v[65], v[185]))
+  k.emit(v_accvgpr_read(v[66], v[189]))
+  k.emit(v_accvgpr_read(v[67], v[193]))
+  k.emit(v_accvgpr_read(v[68], v[197]))
+  k.emit(v_accvgpr_read(v[69], v[201]))
+  k.emit(v_accvgpr_read(v[70], v[205]))
+  k.emit(v_accvgpr_read(v[71], v[209]))
+  k.emit(v_accvgpr_read(v[72], v[213]))
+  k.emit(v_accvgpr_read(v[73], v[217]))
+  k.emit(v_accvgpr_read(v[74], v[221]))
+  k.waitcnt(lgkm=0)
+  k.emit(v_mov_b32_e32(v[32], 4294901760))
+  k.emit(v_mov_b32_e32(v[33], 2147418112))
+  k.emit(v_mov_b32_e32(v[34], 32767))
+  k.emit(v_mul_f32_e32(v[35], v[76], v[35]))
+  k.emit(v_add_f32_e32(v[22], v[75], v[35]))
+  k.emit(v_mov_b32_e32(v[35], v[22]))
+  k.emit(v_cvt_pk(v[35], v[35], v[35]))
+  k.emit(buffer_store_short(v[35], v[77], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[36], v[80], v[36]))
+  k.emit(v_add_f32_e32(v[22], v[79], v[36]))
+  k.emit(v_mov_b32_e32(v[36], v[22]))
+  k.emit(v_cvt_pk(v[36], v[36], v[36]))
+  k.emit(buffer_store_short(v[36], v[81], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[37], v[84], v[37]))
+  k.emit(v_add_f32_e32(v[22], v[83], v[37]))
+  k.emit(v_mov_b32_e32(v[37], v[22]))
+  k.emit(v_cvt_pk(v[37], v[37], v[37]))
+  k.emit(buffer_store_short(v[37], v[85], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[38], v[88], v[38]))
+  k.emit(v_add_f32_e32(v[22], v[87], v[38]))
+  k.emit(v_mov_b32_e32(v[38], v[22]))
+  k.emit(v_cvt_pk(v[38], v[38], v[38]))
+  k.emit(buffer_store_short(v[38], v[89], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[39], v[92], v[39]))
+  k.emit(v_add_f32_e32(v[22], v[91], v[39]))
+  k.emit(v_mov_b32_e32(v[39], v[22]))
+  k.emit(v_cvt_pk(v[39], v[39], v[39]))
+  k.emit(buffer_store_short(v[39], v[93], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[40], v[96], v[40]))
+  k.emit(v_add_f32_e32(v[22], v[95], v[40]))
+  k.emit(v_mov_b32_e32(v[40], v[22]))
+  k.emit(v_cvt_pk(v[40], v[40], v[40]))
+  k.emit(buffer_store_short(v[40], v[97], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[41], v[100], v[41]))
+  k.emit(v_add_f32_e32(v[22], v[99], v[41]))
+  k.emit(v_mov_b32_e32(v[41], v[22]))
+  k.emit(v_cvt_pk(v[41], v[41], v[41]))
+  k.emit(buffer_store_short(v[41], v[101], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[42], v[104], v[42]))
+  k.emit(v_add_f32_e32(v[22], v[103], v[42]))
+  k.emit(v_mov_b32_e32(v[42], v[22]))
+  k.emit(v_cvt_pk(v[42], v[42], v[42]))
+  k.emit(buffer_store_short(v[42], v[105], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[43], v[76], v[43]))
+  k.emit(v_add_f32_e32(v[22], v[75], v[43]))
+  k.emit(v_mov_b32_e32(v[43], v[22]))
+  k.emit(v_cvt_pk(v[43], v[43], v[43]))
+  k.emit(buffer_store_short(v[43], v[107], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[44], v[80], v[44]))
+  k.emit(v_add_f32_e32(v[22], v[79], v[44]))
+  k.emit(v_mov_b32_e32(v[44], v[22]))
+  k.emit(v_cvt_pk(v[44], v[44], v[44]))
+  k.emit(buffer_store_short(v[44], v[109], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[45], v[84], v[45]))
+  k.emit(v_add_f32_e32(v[22], v[83], v[45]))
+  k.emit(v_mov_b32_e32(v[45], v[22]))
+  k.emit(v_cvt_pk(v[45], v[45], v[45]))
+  k.emit(buffer_store_short(v[45], v[111], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[46], v[88], v[46]))
+  k.emit(v_add_f32_e32(v[22], v[87], v[46]))
+  k.emit(v_mov_b32_e32(v[46], v[22]))
+  k.emit(v_cvt_pk(v[46], v[46], v[46]))
+  k.emit(buffer_store_short(v[46], v[113], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[47], v[92], v[47]))
+  k.emit(v_add_f32_e32(v[22], v[91], v[47]))
+  k.emit(v_mov_b32_e32(v[47], v[22]))
+  k.emit(v_cvt_pk(v[47], v[47], v[47]))
+  k.emit(buffer_store_short(v[47], v[115], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[48], v[96], v[48]))
+  k.emit(v_add_f32_e32(v[22], v[95], v[48]))
+  k.emit(v_mov_b32_e32(v[48], v[22]))
+  k.emit(v_cvt_pk(v[48], v[48], v[48]))
+  k.emit(buffer_store_short(v[48], v[117], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[49], v[100], v[49]))
+  k.emit(v_add_f32_e32(v[22], v[99], v[49]))
+  k.emit(v_mov_b32_e32(v[49], v[22]))
+  k.emit(v_cvt_pk(v[49], v[49], v[49]))
+  k.emit(buffer_store_short(v[49], v[119], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[50], v[104], v[50]))
+  k.emit(v_add_f32_e32(v[22], v[103], v[50]))
+  k.emit(v_mov_b32_e32(v[50], v[22]))
+  k.emit(v_cvt_pk(v[50], v[50], v[50]))
+  k.emit(buffer_store_short(v[50], v[121], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[51], v[76], v[51]))
+  k.emit(v_add_f32_e32(v[22], v[75], v[51]))
+  k.emit(v_mov_b32_e32(v[51], v[22]))
+  k.emit(v_cvt_pk(v[51], v[51], v[51]))
+  k.emit(buffer_store_short(v[51], v[123], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[52], v[80], v[52]))
+  k.emit(v_add_f32_e32(v[22], v[79], v[52]))
+  k.emit(v_mov_b32_e32(v[52], v[22]))
+  k.emit(v_cvt_pk(v[52], v[52], v[52]))
+  k.emit(buffer_store_short(v[52], v[125], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[53], v[84], v[53]))
+  k.emit(v_add_f32_e32(v[22], v[83], v[53]))
+  k.emit(v_mov_b32_e32(v[53], v[22]))
+  k.emit(v_cvt_pk(v[53], v[53], v[53]))
+  k.emit(buffer_store_short(v[53], v[127], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[54], v[88], v[54]))
+  k.emit(v_add_f32_e32(v[22], v[87], v[54]))
+  k.emit(v_mov_b32_e32(v[54], v[22]))
+  k.emit(v_cvt_pk(v[54], v[54], v[54]))
+  k.emit(buffer_store_short(v[54], v[129], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[55], v[92], v[55]))
+  k.emit(v_add_f32_e32(v[22], v[91], v[55]))
+  k.emit(v_mov_b32_e32(v[55], v[22]))
+  k.emit(v_cvt_pk(v[55], v[55], v[55]))
+  k.emit(buffer_store_short(v[55], v[131], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[56], v[96], v[56]))
+  k.emit(v_add_f32_e32(v[22], v[95], v[56]))
+  k.emit(v_mov_b32_e32(v[56], v[22]))
+  k.emit(v_cvt_pk(v[56], v[56], v[56]))
+  k.emit(buffer_store_short(v[56], v[133], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[57], v[100], v[57]))
+  k.emit(v_add_f32_e32(v[22], v[99], v[57]))
+  k.emit(v_mov_b32_e32(v[57], v[22]))
+  k.emit(v_cvt_pk(v[57], v[57], v[57]))
+  k.emit(buffer_store_short(v[57], v[135], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[58], v[104], v[58]))
+  k.emit(v_add_f32_e32(v[22], v[103], v[58]))
+  k.emit(v_mov_b32_e32(v[58], v[22]))
+  k.emit(v_cvt_pk(v[58], v[58], v[58]))
+  k.emit(buffer_store_short(v[58], v[137], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[59], v[76], v[59]))
+  k.emit(v_add_f32_e32(v[22], v[75], v[59]))
+  k.emit(v_mov_b32_e32(v[59], v[22]))
+  k.emit(v_cvt_pk(v[59], v[59], v[59]))
+  k.emit(buffer_store_short(v[59], v[139], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[60], v[80], v[60]))
+  k.emit(v_add_f32_e32(v[22], v[79], v[60]))
+  k.emit(v_mov_b32_e32(v[60], v[22]))
+  k.emit(v_cvt_pk(v[60], v[60], v[60]))
+  k.emit(buffer_store_short(v[60], v[141], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[61], v[84], v[61]))
+  k.emit(v_add_f32_e32(v[22], v[83], v[61]))
+  k.emit(v_mov_b32_e32(v[61], v[22]))
+  k.emit(v_cvt_pk(v[61], v[61], v[61]))
+  k.emit(buffer_store_short(v[61], v[143], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[62], v[88], v[62]))
+  k.emit(v_add_f32_e32(v[22], v[87], v[62]))
+  k.emit(v_mov_b32_e32(v[62], v[22]))
+  k.emit(v_cvt_pk(v[62], v[62], v[62]))
+  k.emit(buffer_store_short(v[62], v[145], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[63], v[92], v[63]))
+  k.emit(v_add_f32_e32(v[22], v[91], v[63]))
+  k.emit(v_mov_b32_e32(v[63], v[22]))
+  k.emit(v_cvt_pk(v[63], v[63], v[63]))
+  k.emit(buffer_store_short(v[63], v[147], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[64], v[96], v[64]))
+  k.emit(v_add_f32_e32(v[22], v[95], v[64]))
+  k.emit(v_mov_b32_e32(v[64], v[22]))
+  k.emit(v_cvt_pk(v[64], v[64], v[64]))
+  k.emit(buffer_store_short(v[64], v[149], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[65], v[100], v[65]))
+  k.emit(v_add_f32_e32(v[22], v[99], v[65]))
+  k.emit(v_mov_b32_e32(v[65], v[22]))
+  k.emit(v_cvt_pk(v[65], v[65], v[65]))
+  k.emit(buffer_store_short(v[65], v[151], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[66], v[104], v[66]))
+  k.emit(v_add_f32_e32(v[22], v[103], v[66]))
+  k.emit(v_mov_b32_e32(v[66], v[22]))
+  k.emit(v_cvt_pk(v[66], v[66], v[66]))
+  k.emit(buffer_store_short(v[66], v[153], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[67], v[76], v[67]))
+  k.emit(v_add_f32_e32(v[22], v[75], v[67]))
+  k.emit(v_mov_b32_e32(v[67], v[22]))
+  k.emit(v_cvt_pk(v[67], v[67], v[67]))
+  k.emit(buffer_store_short(v[67], v[155], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[68], v[80], v[68]))
+  k.emit(v_add_f32_e32(v[22], v[79], v[68]))
+  k.emit(v_mov_b32_e32(v[68], v[22]))
+  k.emit(v_cvt_pk(v[68], v[68], v[68]))
+  k.emit(buffer_store_short(v[68], v[157], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[69], v[84], v[69]))
+  k.emit(v_add_f32_e32(v[22], v[83], v[69]))
+  k.emit(v_mov_b32_e32(v[69], v[22]))
+  k.emit(v_cvt_pk(v[69], v[69], v[69]))
+  k.emit(buffer_store_short(v[69], v[159], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[70], v[88], v[70]))
+  k.emit(v_add_f32_e32(v[22], v[87], v[70]))
+  k.emit(v_mov_b32_e32(v[70], v[22]))
+  k.emit(v_cvt_pk(v[70], v[70], v[70]))
+  k.emit(buffer_store_short(v[70], v[161], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[71], v[92], v[71]))
+  k.emit(v_add_f32_e32(v[22], v[91], v[71]))
+  k.emit(v_mov_b32_e32(v[71], v[22]))
+  k.emit(v_cvt_pk(v[71], v[71], v[71]))
+  k.emit(buffer_store_short(v[71], v[163], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[72], v[96], v[72]))
+  k.emit(v_add_f32_e32(v[22], v[95], v[72]))
+  k.emit(v_mov_b32_e32(v[72], v[22]))
+  k.emit(v_cvt_pk(v[72], v[72], v[72]))
+  k.emit(buffer_store_short(v[72], v[165], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[73], v[100], v[73]))
+  k.emit(v_add_f32_e32(v[22], v[99], v[73]))
+  k.emit(v_mov_b32_e32(v[73], v[22]))
+  k.emit(v_cvt_pk(v[73], v[73], v[73]))
+  k.emit(buffer_store_short(v[73], v[167], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[74], v[104], v[74]))
+  k.emit(v_add_f32_e32(v[22], v[103], v[74]))
+  k.emit(v_mov_b32_e32(v[74], v[22]))
+  k.emit(v_cvt_pk(v[74], v[74], v[74]))
+  k.emit(buffer_store_short(v[74], v[169], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_nop())
+  k.emit(v_mov_b32_e32(v[30], 2147483648))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[78], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[78], 2, v[78]))
+  k.emit(ds_read_b32(v[75], v[78]))
+  k.emit(ds_read_b32(v[76], v[78], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[77], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[77], v[30], v[77], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 1))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[82], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[82], 2, v[82]))
+  k.emit(ds_read_b32(v[79], v[82]))
+  k.emit(ds_read_b32(v[80], v[82], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[81], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[81], v[30], v[81], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 2))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[86], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[86], 2, v[86]))
+  k.emit(ds_read_b32(v[83], v[86]))
+  k.emit(ds_read_b32(v[84], v[86], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[85], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[85], v[30], v[85], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 3))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[90], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[90], 2, v[90]))
+  k.emit(ds_read_b32(v[87], v[90]))
+  k.emit(ds_read_b32(v[88], v[90], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[89], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[89], v[30], v[89], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 4))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[94], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[94], 2, v[94]))
+  k.emit(ds_read_b32(v[91], v[94]))
+  k.emit(ds_read_b32(v[92], v[94], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[93], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[93], v[30], v[93], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 5))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[98], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[98], 2, v[98]))
+  k.emit(ds_read_b32(v[95], v[98]))
+  k.emit(ds_read_b32(v[96], v[98], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[97], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[97], v[30], v[97], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 6))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[102], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[102], 2, v[102]))
+  k.emit(ds_read_b32(v[99], v[102]))
+  k.emit(ds_read_b32(v[100], v[102], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[101], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[101], v[30], v[101], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 7))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[106], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[106], 2, v[106]))
+  k.emit(ds_read_b32(v[103], v[106]))
+  k.emit(ds_read_b32(v[104], v[106], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[105], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[105], v[30], v[105], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[108], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[108], 2, v[108]))
+  k.emit(v_add_lshl_u32_e64(v[107], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[107], v[30], v[107], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 1))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[110], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[110], 2, v[110]))
+  k.emit(v_add_lshl_u32_e64(v[109], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[109], v[30], v[109], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 2))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[112], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[112], 2, v[112]))
+  k.emit(v_add_lshl_u32_e64(v[111], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[111], v[30], v[111], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 3))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[114], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[114], 2, v[114]))
+  k.emit(v_add_lshl_u32_e64(v[113], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[113], v[30], v[113], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 4))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[116], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[116], 2, v[116]))
+  k.emit(v_add_lshl_u32_e64(v[115], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[115], v[30], v[115], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 5))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[118], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[118], 2, v[118]))
+  k.emit(v_add_lshl_u32_e64(v[117], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[117], v[30], v[117], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 6))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[120], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[120], 2, v[120]))
+  k.emit(v_add_lshl_u32_e64(v[119], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[119], v[30], v[119], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 7))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[122], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[122], 2, v[122]))
+  k.emit(v_add_lshl_u32_e64(v[121], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[121], v[30], v[121], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[124], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[124], 2, v[124]))
+  k.emit(v_add_lshl_u32_e64(v[123], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[123], v[30], v[123], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 1))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[126], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[126], 2, v[126]))
+  k.emit(v_add_lshl_u32_e64(v[125], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[125], v[30], v[125], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 2))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[128], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[128], 2, v[128]))
+  k.emit(v_add_lshl_u32_e64(v[127], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[127], v[30], v[127], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 3))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[130], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[130], 2, v[130]))
+  k.emit(v_add_lshl_u32_e64(v[129], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[129], v[30], v[129], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 4))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[132], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[132], 2, v[132]))
+  k.emit(v_add_lshl_u32_e64(v[131], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[131], v[30], v[131], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 5))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[134], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[134], 2, v[134]))
+  k.emit(v_add_lshl_u32_e64(v[133], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[133], v[30], v[133], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 6))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[136], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[136], 2, v[136]))
+  k.emit(v_add_lshl_u32_e64(v[135], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[135], v[30], v[135], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 7))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[138], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[138], 2, v[138]))
+  k.emit(v_add_lshl_u32_e64(v[137], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[137], v[30], v[137], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[140], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[140], 2, v[140]))
+  k.emit(v_add_lshl_u32_e64(v[139], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[139], v[30], v[139], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 1))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[142], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[142], 2, v[142]))
+  k.emit(v_add_lshl_u32_e64(v[141], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[141], v[30], v[141], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 2))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[144], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[144], 2, v[144]))
+  k.emit(v_add_lshl_u32_e64(v[143], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[143], v[30], v[143], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 3))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[146], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[146], 2, v[146]))
+  k.emit(v_add_lshl_u32_e64(v[145], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[145], v[30], v[145], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 4))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[148], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[148], 2, v[148]))
+  k.emit(v_add_lshl_u32_e64(v[147], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[147], v[30], v[147], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 5))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[150], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[150], 2, v[150]))
+  k.emit(v_add_lshl_u32_e64(v[149], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[149], v[30], v[149], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 6))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[152], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[152], 2, v[152]))
+  k.emit(v_add_lshl_u32_e64(v[151], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[151], v[30], v[151], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 7))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[154], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[154], 2, v[154]))
+  k.emit(v_add_lshl_u32_e64(v[153], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[153], v[30], v[153], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[156], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[156], 2, v[156]))
+  k.emit(v_add_lshl_u32_e64(v[155], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[155], v[30], v[155], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 1))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[158], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[158], 2, v[158]))
+  k.emit(v_add_lshl_u32_e64(v[157], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[157], v[30], v[157], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 2))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[160], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[160], 2, v[160]))
+  k.emit(v_add_lshl_u32_e64(v[159], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[159], v[30], v[159], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 3))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[162], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[162], 2, v[162]))
+  k.emit(v_add_lshl_u32_e64(v[161], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[161], v[30], v[161], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 4))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[164], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[164], 2, v[164]))
+  k.emit(v_add_lshl_u32_e64(v[163], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[163], v[30], v[163], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 5))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[166], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[166], 2, v[166]))
+  k.emit(v_add_lshl_u32_e64(v[165], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[165], v[30], v[165], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 6))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[168], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[168], 2, v[168]))
+  k.emit(v_add_lshl_u32_e64(v[167], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[167], v[30], v[167], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 7))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[170], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[170], 2, v[170]))
+  k.emit(v_add_lshl_u32_e64(v[169], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[169], v[30], v[169], s[82:83]))
+  k.emit(v_accvgpr_read(v[35], v[225]))
+  k.emit(v_accvgpr_read(v[36], v[229]))
+  k.emit(v_accvgpr_read(v[37], v[233]))
+  k.emit(v_accvgpr_read(v[38], v[237]))
+  k.emit(v_accvgpr_read(v[39], v[241]))
+  k.emit(v_accvgpr_read(v[40], v[245]))
+  k.emit(v_accvgpr_read(v[41], v[249]))
+  k.emit(v_accvgpr_read(v[42], v[253]))
+  k.emit(v_accvgpr_read(v[43], v[2]))
+  k.emit(v_accvgpr_read(v[44], v[6]))
+  k.emit(v_accvgpr_read(v[45], v[10]))
+  k.emit(v_accvgpr_read(v[46], v[14]))
+  k.emit(v_accvgpr_read(v[47], v[18]))
+  k.emit(v_accvgpr_read(v[48], v[22]))
+  k.emit(v_accvgpr_read(v[49], v[26]))
+  k.emit(v_accvgpr_read(v[50], v[30]))
+  k.emit(v_accvgpr_read(v[51], v[34]))
+  k.emit(v_accvgpr_read(v[52], v[38]))
+  k.emit(v_accvgpr_read(v[53], v[42]))
+  k.emit(v_accvgpr_read(v[54], v[46]))
+  k.emit(v_accvgpr_read(v[55], v[50]))
+  k.emit(v_accvgpr_read(v[56], v[54]))
+  k.emit(v_accvgpr_read(v[57], v[58]))
+  k.emit(v_accvgpr_read(v[58], v[62]))
+  k.emit(v_accvgpr_read(v[59], v[66]))
+  k.emit(v_accvgpr_read(v[60], v[70]))
+  k.emit(v_accvgpr_read(v[61], v[74]))
+  k.emit(v_accvgpr_read(v[62], v[78]))
+  k.emit(v_accvgpr_read(v[63], v[82]))
+  k.emit(v_accvgpr_read(v[64], v[86]))
+  k.emit(v_accvgpr_read(v[65], v[90]))
+  k.emit(v_accvgpr_read(v[66], v[94]))
+  k.emit(v_accvgpr_read(v[67], v[98]))
+  k.emit(v_accvgpr_read(v[68], v[102]))
+  k.emit(v_accvgpr_read(v[69], v[106]))
+  k.emit(v_accvgpr_read(v[70], v[110]))
+  k.emit(v_accvgpr_read(v[71], v[114]))
+  k.emit(v_accvgpr_read(v[72], v[118]))
+  k.emit(v_accvgpr_read(v[73], v[122]))
+  k.emit(v_accvgpr_read(v[74], v[126]))
+  k.waitcnt(lgkm=0)
+  k.emit(v_mov_b32_e32(v[32], 4294901760))
+  k.emit(v_mov_b32_e32(v[33], 2147418112))
+  k.emit(v_mov_b32_e32(v[34], 32767))
+  k.emit(v_mul_f32_e32(v[35], v[76], v[35]))
+  k.emit(v_add_f32_e32(v[22], v[75], v[35]))
+  k.emit(v_mov_b32_e32(v[35], v[22]))
+  k.emit(v_cvt_pk(v[35], v[35], v[35]))
+  k.emit(buffer_store_short(v[35], v[77], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[36], v[80], v[36]))
+  k.emit(v_add_f32_e32(v[22], v[79], v[36]))
+  k.emit(v_mov_b32_e32(v[36], v[22]))
+  k.emit(v_cvt_pk(v[36], v[36], v[36]))
+  k.emit(buffer_store_short(v[36], v[81], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[37], v[84], v[37]))
+  k.emit(v_add_f32_e32(v[22], v[83], v[37]))
+  k.emit(v_mov_b32_e32(v[37], v[22]))
+  k.emit(v_cvt_pk(v[37], v[37], v[37]))
+  k.emit(buffer_store_short(v[37], v[85], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[38], v[88], v[38]))
+  k.emit(v_add_f32_e32(v[22], v[87], v[38]))
+  k.emit(v_mov_b32_e32(v[38], v[22]))
+  k.emit(v_cvt_pk(v[38], v[38], v[38]))
+  k.emit(buffer_store_short(v[38], v[89], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[39], v[92], v[39]))
+  k.emit(v_add_f32_e32(v[22], v[91], v[39]))
+  k.emit(v_mov_b32_e32(v[39], v[22]))
+  k.emit(v_cvt_pk(v[39], v[39], v[39]))
+  k.emit(buffer_store_short(v[39], v[93], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[40], v[96], v[40]))
+  k.emit(v_add_f32_e32(v[22], v[95], v[40]))
+  k.emit(v_mov_b32_e32(v[40], v[22]))
+  k.emit(v_cvt_pk(v[40], v[40], v[40]))
+  k.emit(buffer_store_short(v[40], v[97], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[41], v[100], v[41]))
+  k.emit(v_add_f32_e32(v[22], v[99], v[41]))
+  k.emit(v_mov_b32_e32(v[41], v[22]))
+  k.emit(v_cvt_pk(v[41], v[41], v[41]))
+  k.emit(buffer_store_short(v[41], v[101], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[42], v[104], v[42]))
+  k.emit(v_add_f32_e32(v[22], v[103], v[42]))
+  k.emit(v_mov_b32_e32(v[42], v[22]))
+  k.emit(v_cvt_pk(v[42], v[42], v[42]))
+  k.emit(buffer_store_short(v[42], v[105], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[43], v[76], v[43]))
+  k.emit(v_add_f32_e32(v[22], v[75], v[43]))
+  k.emit(v_mov_b32_e32(v[43], v[22]))
+  k.emit(v_cvt_pk(v[43], v[43], v[43]))
+  k.emit(buffer_store_short(v[43], v[107], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[44], v[80], v[44]))
+  k.emit(v_add_f32_e32(v[22], v[79], v[44]))
+  k.emit(v_mov_b32_e32(v[44], v[22]))
+  k.emit(v_cvt_pk(v[44], v[44], v[44]))
+  k.emit(buffer_store_short(v[44], v[109], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[45], v[84], v[45]))
+  k.emit(v_add_f32_e32(v[22], v[83], v[45]))
+  k.emit(v_mov_b32_e32(v[45], v[22]))
+  k.emit(v_cvt_pk(v[45], v[45], v[45]))
+  k.emit(buffer_store_short(v[45], v[111], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[46], v[88], v[46]))
+  k.emit(v_add_f32_e32(v[22], v[87], v[46]))
+  k.emit(v_mov_b32_e32(v[46], v[22]))
+  k.emit(v_cvt_pk(v[46], v[46], v[46]))
+  k.emit(buffer_store_short(v[46], v[113], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[47], v[92], v[47]))
+  k.emit(v_add_f32_e32(v[22], v[91], v[47]))
+  k.emit(v_mov_b32_e32(v[47], v[22]))
+  k.emit(v_cvt_pk(v[47], v[47], v[47]))
+  k.emit(buffer_store_short(v[47], v[115], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[48], v[96], v[48]))
+  k.emit(v_add_f32_e32(v[22], v[95], v[48]))
+  k.emit(v_mov_b32_e32(v[48], v[22]))
+  k.emit(v_cvt_pk(v[48], v[48], v[48]))
+  k.emit(buffer_store_short(v[48], v[117], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[49], v[100], v[49]))
+  k.emit(v_add_f32_e32(v[22], v[99], v[49]))
+  k.emit(v_mov_b32_e32(v[49], v[22]))
+  k.emit(v_cvt_pk(v[49], v[49], v[49]))
+  k.emit(buffer_store_short(v[49], v[119], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[50], v[104], v[50]))
+  k.emit(v_add_f32_e32(v[22], v[103], v[50]))
+  k.emit(v_mov_b32_e32(v[50], v[22]))
+  k.emit(v_cvt_pk(v[50], v[50], v[50]))
+  k.emit(buffer_store_short(v[50], v[121], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[51], v[76], v[51]))
+  k.emit(v_add_f32_e32(v[22], v[75], v[51]))
+  k.emit(v_mov_b32_e32(v[51], v[22]))
+  k.emit(v_cvt_pk(v[51], v[51], v[51]))
+  k.emit(buffer_store_short(v[51], v[123], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[52], v[80], v[52]))
+  k.emit(v_add_f32_e32(v[22], v[79], v[52]))
+  k.emit(v_mov_b32_e32(v[52], v[22]))
+  k.emit(v_cvt_pk(v[52], v[52], v[52]))
+  k.emit(buffer_store_short(v[52], v[125], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[53], v[84], v[53]))
+  k.emit(v_add_f32_e32(v[22], v[83], v[53]))
+  k.emit(v_mov_b32_e32(v[53], v[22]))
+  k.emit(v_cvt_pk(v[53], v[53], v[53]))
+  k.emit(buffer_store_short(v[53], v[127], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[54], v[88], v[54]))
+  k.emit(v_add_f32_e32(v[22], v[87], v[54]))
+  k.emit(v_mov_b32_e32(v[54], v[22]))
+  k.emit(v_cvt_pk(v[54], v[54], v[54]))
+  k.emit(buffer_store_short(v[54], v[129], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[55], v[92], v[55]))
+  k.emit(v_add_f32_e32(v[22], v[91], v[55]))
+  k.emit(v_mov_b32_e32(v[55], v[22]))
+  k.emit(v_cvt_pk(v[55], v[55], v[55]))
+  k.emit(buffer_store_short(v[55], v[131], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[56], v[96], v[56]))
+  k.emit(v_add_f32_e32(v[22], v[95], v[56]))
+  k.emit(v_mov_b32_e32(v[56], v[22]))
+  k.emit(v_cvt_pk(v[56], v[56], v[56]))
+  k.emit(buffer_store_short(v[56], v[133], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[57], v[100], v[57]))
+  k.emit(v_add_f32_e32(v[22], v[99], v[57]))
+  k.emit(v_mov_b32_e32(v[57], v[22]))
+  k.emit(v_cvt_pk(v[57], v[57], v[57]))
+  k.emit(buffer_store_short(v[57], v[135], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[58], v[104], v[58]))
+  k.emit(v_add_f32_e32(v[22], v[103], v[58]))
+  k.emit(v_mov_b32_e32(v[58], v[22]))
+  k.emit(v_cvt_pk(v[58], v[58], v[58]))
+  k.emit(buffer_store_short(v[58], v[137], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[59], v[76], v[59]))
+  k.emit(v_add_f32_e32(v[22], v[75], v[59]))
+  k.emit(v_mov_b32_e32(v[59], v[22]))
+  k.emit(v_cvt_pk(v[59], v[59], v[59]))
+  k.emit(buffer_store_short(v[59], v[139], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[60], v[80], v[60]))
+  k.emit(v_add_f32_e32(v[22], v[79], v[60]))
+  k.emit(v_mov_b32_e32(v[60], v[22]))
+  k.emit(v_cvt_pk(v[60], v[60], v[60]))
+  k.emit(buffer_store_short(v[60], v[141], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[61], v[84], v[61]))
+  k.emit(v_add_f32_e32(v[22], v[83], v[61]))
+  k.emit(v_mov_b32_e32(v[61], v[22]))
+  k.emit(v_cvt_pk(v[61], v[61], v[61]))
+  k.emit(buffer_store_short(v[61], v[143], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[62], v[88], v[62]))
+  k.emit(v_add_f32_e32(v[22], v[87], v[62]))
+  k.emit(v_mov_b32_e32(v[62], v[22]))
+  k.emit(v_cvt_pk(v[62], v[62], v[62]))
+  k.emit(buffer_store_short(v[62], v[145], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[63], v[92], v[63]))
+  k.emit(v_add_f32_e32(v[22], v[91], v[63]))
+  k.emit(v_mov_b32_e32(v[63], v[22]))
+  k.emit(v_cvt_pk(v[63], v[63], v[63]))
+  k.emit(buffer_store_short(v[63], v[147], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[64], v[96], v[64]))
+  k.emit(v_add_f32_e32(v[22], v[95], v[64]))
+  k.emit(v_mov_b32_e32(v[64], v[22]))
+  k.emit(v_cvt_pk(v[64], v[64], v[64]))
+  k.emit(buffer_store_short(v[64], v[149], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[65], v[100], v[65]))
+  k.emit(v_add_f32_e32(v[22], v[99], v[65]))
+  k.emit(v_mov_b32_e32(v[65], v[22]))
+  k.emit(v_cvt_pk(v[65], v[65], v[65]))
+  k.emit(buffer_store_short(v[65], v[151], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[66], v[104], v[66]))
+  k.emit(v_add_f32_e32(v[22], v[103], v[66]))
+  k.emit(v_mov_b32_e32(v[66], v[22]))
+  k.emit(v_cvt_pk(v[66], v[66], v[66]))
+  k.emit(buffer_store_short(v[66], v[153], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[67], v[76], v[67]))
+  k.emit(v_add_f32_e32(v[22], v[75], v[67]))
+  k.emit(v_mov_b32_e32(v[67], v[22]))
+  k.emit(v_cvt_pk(v[67], v[67], v[67]))
+  k.emit(buffer_store_short(v[67], v[155], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[68], v[80], v[68]))
+  k.emit(v_add_f32_e32(v[22], v[79], v[68]))
+  k.emit(v_mov_b32_e32(v[68], v[22]))
+  k.emit(v_cvt_pk(v[68], v[68], v[68]))
+  k.emit(buffer_store_short(v[68], v[157], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[69], v[84], v[69]))
+  k.emit(v_add_f32_e32(v[22], v[83], v[69]))
+  k.emit(v_mov_b32_e32(v[69], v[22]))
+  k.emit(v_cvt_pk(v[69], v[69], v[69]))
+  k.emit(buffer_store_short(v[69], v[159], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[70], v[88], v[70]))
+  k.emit(v_add_f32_e32(v[22], v[87], v[70]))
+  k.emit(v_mov_b32_e32(v[70], v[22]))
+  k.emit(v_cvt_pk(v[70], v[70], v[70]))
+  k.emit(buffer_store_short(v[70], v[161], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[71], v[92], v[71]))
+  k.emit(v_add_f32_e32(v[22], v[91], v[71]))
+  k.emit(v_mov_b32_e32(v[71], v[22]))
+  k.emit(v_cvt_pk(v[71], v[71], v[71]))
+  k.emit(buffer_store_short(v[71], v[163], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[72], v[96], v[72]))
+  k.emit(v_add_f32_e32(v[22], v[95], v[72]))
+  k.emit(v_mov_b32_e32(v[72], v[22]))
+  k.emit(v_cvt_pk(v[72], v[72], v[72]))
+  k.emit(buffer_store_short(v[72], v[165], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[73], v[100], v[73]))
+  k.emit(v_add_f32_e32(v[22], v[99], v[73]))
+  k.emit(v_mov_b32_e32(v[73], v[22]))
+  k.emit(v_cvt_pk(v[73], v[73], v[73]))
+  k.emit(buffer_store_short(v[73], v[167], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[74], v[104], v[74]))
+  k.emit(v_add_f32_e32(v[22], v[103], v[74]))
+  k.emit(v_mov_b32_e32(v[74], v[22]))
+  k.emit(v_cvt_pk(v[74], v[74], v[74]))
+  k.emit(buffer_store_short(v[74], v[169], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_nop())
+  k.emit(v_mov_b32_e32(v[30], 2147483648))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[78], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[78], 2, v[78]))
+  k.emit(ds_read_b32(v[75], v[78]))
+  k.emit(ds_read_b32(v[76], v[78], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[77], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[77], v[30], v[77], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 1))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[82], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[82], 2, v[82]))
+  k.emit(ds_read_b32(v[79], v[82]))
+  k.emit(ds_read_b32(v[80], v[82], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[81], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[81], v[30], v[81], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 2))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[86], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[86], 2, v[86]))
+  k.emit(ds_read_b32(v[83], v[86]))
+  k.emit(ds_read_b32(v[84], v[86], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[85], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[85], v[30], v[85], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 3))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[90], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[90], 2, v[90]))
+  k.emit(ds_read_b32(v[87], v[90]))
+  k.emit(ds_read_b32(v[88], v[90], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[89], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[89], v[30], v[89], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 4))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[94], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[94], 2, v[94]))
+  k.emit(ds_read_b32(v[91], v[94]))
+  k.emit(ds_read_b32(v[92], v[94], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[93], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[93], v[30], v[93], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 5))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[98], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[98], 2, v[98]))
+  k.emit(ds_read_b32(v[95], v[98]))
+  k.emit(ds_read_b32(v[96], v[98], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[97], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[97], v[30], v[97], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 6))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[102], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[102], 2, v[102]))
+  k.emit(ds_read_b32(v[99], v[102]))
+  k.emit(ds_read_b32(v[100], v[102], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[101], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[101], v[30], v[101], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 7))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[106], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[106], 2, v[106]))
+  k.emit(ds_read_b32(v[103], v[106]))
+  k.emit(ds_read_b32(v[104], v[106], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[105], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[105], v[30], v[105], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[108], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[108], 2, v[108]))
+  k.emit(v_add_lshl_u32_e64(v[107], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[107], v[30], v[107], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 1))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[110], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[110], 2, v[110]))
+  k.emit(v_add_lshl_u32_e64(v[109], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[109], v[30], v[109], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 2))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[112], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[112], 2, v[112]))
+  k.emit(v_add_lshl_u32_e64(v[111], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[111], v[30], v[111], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 3))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[114], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[114], 2, v[114]))
+  k.emit(v_add_lshl_u32_e64(v[113], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[113], v[30], v[113], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 4))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[116], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[116], 2, v[116]))
+  k.emit(v_add_lshl_u32_e64(v[115], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[115], v[30], v[115], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 5))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[118], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[118], 2, v[118]))
+  k.emit(v_add_lshl_u32_e64(v[117], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[117], v[30], v[117], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 6))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[120], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[120], 2, v[120]))
+  k.emit(v_add_lshl_u32_e64(v[119], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[119], v[30], v[119], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 7))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[122], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[122], 2, v[122]))
+  k.emit(v_add_lshl_u32_e64(v[121], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[121], v[30], v[121], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[124], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[124], 2, v[124]))
+  k.emit(v_add_lshl_u32_e64(v[123], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[123], v[30], v[123], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 1))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[126], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[126], 2, v[126]))
+  k.emit(v_add_lshl_u32_e64(v[125], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[125], v[30], v[125], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 2))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[128], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[128], 2, v[128]))
+  k.emit(v_add_lshl_u32_e64(v[127], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[127], v[30], v[127], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 3))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[130], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[130], 2, v[130]))
+  k.emit(v_add_lshl_u32_e64(v[129], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[129], v[30], v[129], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 4))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[132], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[132], 2, v[132]))
+  k.emit(v_add_lshl_u32_e64(v[131], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[131], v[30], v[131], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 5))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[134], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[134], 2, v[134]))
+  k.emit(v_add_lshl_u32_e64(v[133], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[133], v[30], v[133], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 6))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[136], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[136], 2, v[136]))
+  k.emit(v_add_lshl_u32_e64(v[135], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[135], v[30], v[135], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 7))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[138], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[138], 2, v[138]))
+  k.emit(v_add_lshl_u32_e64(v[137], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[137], v[30], v[137], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[140], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[140], 2, v[140]))
+  k.emit(v_add_lshl_u32_e64(v[139], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[139], v[30], v[139], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 1))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[142], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[142], 2, v[142]))
+  k.emit(v_add_lshl_u32_e64(v[141], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[141], v[30], v[141], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 2))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[144], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[144], 2, v[144]))
+  k.emit(v_add_lshl_u32_e64(v[143], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[143], v[30], v[143], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 3))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[146], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[146], 2, v[146]))
+  k.emit(v_add_lshl_u32_e64(v[145], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[145], v[30], v[145], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 4))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[148], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[148], 2, v[148]))
+  k.emit(v_add_lshl_u32_e64(v[147], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[147], v[30], v[147], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 5))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[150], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[150], 2, v[150]))
+  k.emit(v_add_lshl_u32_e64(v[149], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[149], v[30], v[149], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 6))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[152], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[152], 2, v[152]))
+  k.emit(v_add_lshl_u32_e64(v[151], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[151], v[30], v[151], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 7))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[154], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[154], 2, v[154]))
+  k.emit(v_add_lshl_u32_e64(v[153], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[153], v[30], v[153], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[156], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[156], 2, v[156]))
+  k.emit(v_add_lshl_u32_e64(v[155], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[155], v[30], v[155], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 1))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[158], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[158], 2, v[158]))
+  k.emit(v_add_lshl_u32_e64(v[157], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[157], v[30], v[157], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 2))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[160], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[160], 2, v[160]))
+  k.emit(v_add_lshl_u32_e64(v[159], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[159], v[30], v[159], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 3))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[162], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[162], 2, v[162]))
+  k.emit(v_add_lshl_u32_e64(v[161], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[161], v[30], v[161], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 4))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[164], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[164], 2, v[164]))
+  k.emit(v_add_lshl_u32_e64(v[163], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[163], v[30], v[163], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 5))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[166], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[166], 2, v[166]))
+  k.emit(v_add_lshl_u32_e64(v[165], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[165], v[30], v[165], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 6))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[168], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[168], 2, v[168]))
+  k.emit(v_add_lshl_u32_e64(v[167], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[167], v[30], v[167], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 7))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[170], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[170], 2, v[170]))
+  k.emit(v_add_lshl_u32_e64(v[169], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[169], v[30], v[169], s[82:83]))
+  k.emit(v_accvgpr_read(v[35], v[130]))
+  k.emit(v_accvgpr_read(v[36], v[134]))
+  k.emit(v_accvgpr_read(v[37], v[138]))
+  k.emit(v_accvgpr_read(v[38], v[142]))
+  k.emit(v_accvgpr_read(v[39], v[146]))
+  k.emit(v_accvgpr_read(v[40], v[150]))
+  k.emit(v_accvgpr_read(v[41], v[154]))
+  k.emit(v_accvgpr_read(v[42], v[158]))
+  k.emit(v_accvgpr_read(v[43], v[162]))
+  k.emit(v_accvgpr_read(v[44], v[166]))
+  k.emit(v_accvgpr_read(v[45], v[170]))
+  k.emit(v_accvgpr_read(v[46], v[174]))
+  k.emit(v_accvgpr_read(v[47], v[178]))
+  k.emit(v_accvgpr_read(v[48], v[182]))
+  k.emit(v_accvgpr_read(v[49], v[186]))
+  k.emit(v_accvgpr_read(v[50], v[190]))
+  k.emit(v_accvgpr_read(v[51], v[194]))
+  k.emit(v_accvgpr_read(v[52], v[198]))
+  k.emit(v_accvgpr_read(v[53], v[202]))
+  k.emit(v_accvgpr_read(v[54], v[206]))
+  k.emit(v_accvgpr_read(v[55], v[210]))
+  k.emit(v_accvgpr_read(v[56], v[214]))
+  k.emit(v_accvgpr_read(v[57], v[218]))
+  k.emit(v_accvgpr_read(v[58], v[222]))
+  k.emit(v_accvgpr_read(v[59], v[226]))
+  k.emit(v_accvgpr_read(v[60], v[230]))
+  k.emit(v_accvgpr_read(v[61], v[234]))
+  k.emit(v_accvgpr_read(v[62], v[238]))
+  k.emit(v_accvgpr_read(v[63], v[242]))
+  k.emit(v_accvgpr_read(v[64], v[246]))
+  k.emit(v_accvgpr_read(v[65], v[250]))
+  k.emit(v_accvgpr_read(v[66], v[254]))
+  k.emit(v_accvgpr_read(v[67], v[3]))
+  k.emit(v_accvgpr_read(v[68], v[7]))
+  k.emit(v_accvgpr_read(v[69], v[11]))
+  k.emit(v_accvgpr_read(v[70], v[15]))
+  k.emit(v_accvgpr_read(v[71], v[19]))
+  k.emit(v_accvgpr_read(v[72], v[23]))
+  k.emit(v_accvgpr_read(v[73], v[27]))
+  k.emit(v_accvgpr_read(v[74], v[31]))
+  k.waitcnt(lgkm=0)
+  k.emit(v_mov_b32_e32(v[32], 4294901760))
+  k.emit(v_mov_b32_e32(v[33], 2147418112))
+  k.emit(v_mov_b32_e32(v[34], 32767))
+  k.emit(v_mul_f32_e32(v[35], v[76], v[35]))
+  k.emit(v_add_f32_e32(v[22], v[75], v[35]))
+  k.emit(v_mov_b32_e32(v[35], v[22]))
+  k.emit(v_cvt_pk(v[35], v[35], v[35]))
+  k.emit(buffer_store_short(v[35], v[77], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[36], v[80], v[36]))
+  k.emit(v_add_f32_e32(v[22], v[79], v[36]))
+  k.emit(v_mov_b32_e32(v[36], v[22]))
+  k.emit(v_cvt_pk(v[36], v[36], v[36]))
+  k.emit(buffer_store_short(v[36], v[81], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[37], v[84], v[37]))
+  k.emit(v_add_f32_e32(v[22], v[83], v[37]))
+  k.emit(v_mov_b32_e32(v[37], v[22]))
+  k.emit(v_cvt_pk(v[37], v[37], v[37]))
+  k.emit(buffer_store_short(v[37], v[85], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[38], v[88], v[38]))
+  k.emit(v_add_f32_e32(v[22], v[87], v[38]))
+  k.emit(v_mov_b32_e32(v[38], v[22]))
+  k.emit(v_cvt_pk(v[38], v[38], v[38]))
+  k.emit(buffer_store_short(v[38], v[89], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[39], v[92], v[39]))
+  k.emit(v_add_f32_e32(v[22], v[91], v[39]))
+  k.emit(v_mov_b32_e32(v[39], v[22]))
+  k.emit(v_cvt_pk(v[39], v[39], v[39]))
+  k.emit(buffer_store_short(v[39], v[93], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[40], v[96], v[40]))
+  k.emit(v_add_f32_e32(v[22], v[95], v[40]))
+  k.emit(v_mov_b32_e32(v[40], v[22]))
+  k.emit(v_cvt_pk(v[40], v[40], v[40]))
+  k.emit(buffer_store_short(v[40], v[97], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[41], v[100], v[41]))
+  k.emit(v_add_f32_e32(v[22], v[99], v[41]))
+  k.emit(v_mov_b32_e32(v[41], v[22]))
+  k.emit(v_cvt_pk(v[41], v[41], v[41]))
+  k.emit(buffer_store_short(v[41], v[101], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[42], v[104], v[42]))
+  k.emit(v_add_f32_e32(v[22], v[103], v[42]))
+  k.emit(v_mov_b32_e32(v[42], v[22]))
+  k.emit(v_cvt_pk(v[42], v[42], v[42]))
+  k.emit(buffer_store_short(v[42], v[105], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[43], v[76], v[43]))
+  k.emit(v_add_f32_e32(v[22], v[75], v[43]))
+  k.emit(v_mov_b32_e32(v[43], v[22]))
+  k.emit(v_cvt_pk(v[43], v[43], v[43]))
+  k.emit(buffer_store_short(v[43], v[107], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[44], v[80], v[44]))
+  k.emit(v_add_f32_e32(v[22], v[79], v[44]))
+  k.emit(v_mov_b32_e32(v[44], v[22]))
+  k.emit(v_cvt_pk(v[44], v[44], v[44]))
+  k.emit(buffer_store_short(v[44], v[109], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[45], v[84], v[45]))
+  k.emit(v_add_f32_e32(v[22], v[83], v[45]))
+  k.emit(v_mov_b32_e32(v[45], v[22]))
+  k.emit(v_cvt_pk(v[45], v[45], v[45]))
+  k.emit(buffer_store_short(v[45], v[111], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[46], v[88], v[46]))
+  k.emit(v_add_f32_e32(v[22], v[87], v[46]))
+  k.emit(v_mov_b32_e32(v[46], v[22]))
+  k.emit(v_cvt_pk(v[46], v[46], v[46]))
+  k.emit(buffer_store_short(v[46], v[113], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[47], v[92], v[47]))
+  k.emit(v_add_f32_e32(v[22], v[91], v[47]))
+  k.emit(v_mov_b32_e32(v[47], v[22]))
+  k.emit(v_cvt_pk(v[47], v[47], v[47]))
+  k.emit(buffer_store_short(v[47], v[115], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[48], v[96], v[48]))
+  k.emit(v_add_f32_e32(v[22], v[95], v[48]))
+  k.emit(v_mov_b32_e32(v[48], v[22]))
+  k.emit(v_cvt_pk(v[48], v[48], v[48]))
+  k.emit(buffer_store_short(v[48], v[117], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[49], v[100], v[49]))
+  k.emit(v_add_f32_e32(v[22], v[99], v[49]))
+  k.emit(v_mov_b32_e32(v[49], v[22]))
+  k.emit(v_cvt_pk(v[49], v[49], v[49]))
+  k.emit(buffer_store_short(v[49], v[119], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[50], v[104], v[50]))
+  k.emit(v_add_f32_e32(v[22], v[103], v[50]))
+  k.emit(v_mov_b32_e32(v[50], v[22]))
+  k.emit(v_cvt_pk(v[50], v[50], v[50]))
+  k.emit(buffer_store_short(v[50], v[121], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[51], v[76], v[51]))
+  k.emit(v_add_f32_e32(v[22], v[75], v[51]))
+  k.emit(v_mov_b32_e32(v[51], v[22]))
+  k.emit(v_cvt_pk(v[51], v[51], v[51]))
+  k.emit(buffer_store_short(v[51], v[123], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[52], v[80], v[52]))
+  k.emit(v_add_f32_e32(v[22], v[79], v[52]))
+  k.emit(v_mov_b32_e32(v[52], v[22]))
+  k.emit(v_cvt_pk(v[52], v[52], v[52]))
+  k.emit(buffer_store_short(v[52], v[125], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[53], v[84], v[53]))
+  k.emit(v_add_f32_e32(v[22], v[83], v[53]))
+  k.emit(v_mov_b32_e32(v[53], v[22]))
+  k.emit(v_cvt_pk(v[53], v[53], v[53]))
+  k.emit(buffer_store_short(v[53], v[127], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[54], v[88], v[54]))
+  k.emit(v_add_f32_e32(v[22], v[87], v[54]))
+  k.emit(v_mov_b32_e32(v[54], v[22]))
+  k.emit(v_cvt_pk(v[54], v[54], v[54]))
+  k.emit(buffer_store_short(v[54], v[129], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[55], v[92], v[55]))
+  k.emit(v_add_f32_e32(v[22], v[91], v[55]))
+  k.emit(v_mov_b32_e32(v[55], v[22]))
+  k.emit(v_cvt_pk(v[55], v[55], v[55]))
+  k.emit(buffer_store_short(v[55], v[131], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[56], v[96], v[56]))
+  k.emit(v_add_f32_e32(v[22], v[95], v[56]))
+  k.emit(v_mov_b32_e32(v[56], v[22]))
+  k.emit(v_cvt_pk(v[56], v[56], v[56]))
+  k.emit(buffer_store_short(v[56], v[133], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[57], v[100], v[57]))
+  k.emit(v_add_f32_e32(v[22], v[99], v[57]))
+  k.emit(v_mov_b32_e32(v[57], v[22]))
+  k.emit(v_cvt_pk(v[57], v[57], v[57]))
+  k.emit(buffer_store_short(v[57], v[135], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[58], v[104], v[58]))
+  k.emit(v_add_f32_e32(v[22], v[103], v[58]))
+  k.emit(v_mov_b32_e32(v[58], v[22]))
+  k.emit(v_cvt_pk(v[58], v[58], v[58]))
+  k.emit(buffer_store_short(v[58], v[137], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[59], v[76], v[59]))
+  k.emit(v_add_f32_e32(v[22], v[75], v[59]))
+  k.emit(v_mov_b32_e32(v[59], v[22]))
+  k.emit(v_cvt_pk(v[59], v[59], v[59]))
+  k.emit(buffer_store_short(v[59], v[139], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[60], v[80], v[60]))
+  k.emit(v_add_f32_e32(v[22], v[79], v[60]))
+  k.emit(v_mov_b32_e32(v[60], v[22]))
+  k.emit(v_cvt_pk(v[60], v[60], v[60]))
+  k.emit(buffer_store_short(v[60], v[141], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[61], v[84], v[61]))
+  k.emit(v_add_f32_e32(v[22], v[83], v[61]))
+  k.emit(v_mov_b32_e32(v[61], v[22]))
+  k.emit(v_cvt_pk(v[61], v[61], v[61]))
+  k.emit(buffer_store_short(v[61], v[143], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[62], v[88], v[62]))
+  k.emit(v_add_f32_e32(v[22], v[87], v[62]))
+  k.emit(v_mov_b32_e32(v[62], v[22]))
+  k.emit(v_cvt_pk(v[62], v[62], v[62]))
+  k.emit(buffer_store_short(v[62], v[145], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[63], v[92], v[63]))
+  k.emit(v_add_f32_e32(v[22], v[91], v[63]))
+  k.emit(v_mov_b32_e32(v[63], v[22]))
+  k.emit(v_cvt_pk(v[63], v[63], v[63]))
+  k.emit(buffer_store_short(v[63], v[147], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[64], v[96], v[64]))
+  k.emit(v_add_f32_e32(v[22], v[95], v[64]))
+  k.emit(v_mov_b32_e32(v[64], v[22]))
+  k.emit(v_cvt_pk(v[64], v[64], v[64]))
+  k.emit(buffer_store_short(v[64], v[149], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[65], v[100], v[65]))
+  k.emit(v_add_f32_e32(v[22], v[99], v[65]))
+  k.emit(v_mov_b32_e32(v[65], v[22]))
+  k.emit(v_cvt_pk(v[65], v[65], v[65]))
+  k.emit(buffer_store_short(v[65], v[151], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[66], v[104], v[66]))
+  k.emit(v_add_f32_e32(v[22], v[103], v[66]))
+  k.emit(v_mov_b32_e32(v[66], v[22]))
+  k.emit(v_cvt_pk(v[66], v[66], v[66]))
+  k.emit(buffer_store_short(v[66], v[153], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[67], v[76], v[67]))
+  k.emit(v_add_f32_e32(v[22], v[75], v[67]))
+  k.emit(v_mov_b32_e32(v[67], v[22]))
+  k.emit(v_cvt_pk(v[67], v[67], v[67]))
+  k.emit(buffer_store_short(v[67], v[155], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[68], v[80], v[68]))
+  k.emit(v_add_f32_e32(v[22], v[79], v[68]))
+  k.emit(v_mov_b32_e32(v[68], v[22]))
+  k.emit(v_cvt_pk(v[68], v[68], v[68]))
+  k.emit(buffer_store_short(v[68], v[157], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[69], v[84], v[69]))
+  k.emit(v_add_f32_e32(v[22], v[83], v[69]))
+  k.emit(v_mov_b32_e32(v[69], v[22]))
+  k.emit(v_cvt_pk(v[69], v[69], v[69]))
+  k.emit(buffer_store_short(v[69], v[159], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[70], v[88], v[70]))
+  k.emit(v_add_f32_e32(v[22], v[87], v[70]))
+  k.emit(v_mov_b32_e32(v[70], v[22]))
+  k.emit(v_cvt_pk(v[70], v[70], v[70]))
+  k.emit(buffer_store_short(v[70], v[161], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[71], v[92], v[71]))
+  k.emit(v_add_f32_e32(v[22], v[91], v[71]))
+  k.emit(v_mov_b32_e32(v[71], v[22]))
+  k.emit(v_cvt_pk(v[71], v[71], v[71]))
+  k.emit(buffer_store_short(v[71], v[163], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[72], v[96], v[72]))
+  k.emit(v_add_f32_e32(v[22], v[95], v[72]))
+  k.emit(v_mov_b32_e32(v[72], v[22]))
+  k.emit(v_cvt_pk(v[72], v[72], v[72]))
+  k.emit(buffer_store_short(v[72], v[165], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[73], v[100], v[73]))
+  k.emit(v_add_f32_e32(v[22], v[99], v[73]))
+  k.emit(v_mov_b32_e32(v[73], v[22]))
+  k.emit(v_cvt_pk(v[73], v[73], v[73]))
+  k.emit(buffer_store_short(v[73], v[167], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[74], v[104], v[74]))
+  k.emit(v_add_f32_e32(v[22], v[103], v[74]))
+  k.emit(v_mov_b32_e32(v[74], v[22]))
+  k.emit(v_cvt_pk(v[74], v[74], v[74]))
+  k.emit(buffer_store_short(v[74], v[169], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_nop())
+  k.emit(v_mov_b32_e32(v[30], 2147483648))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[78], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[78], 2, v[78]))
+  k.emit(ds_read_b32(v[75], v[78]))
+  k.emit(ds_read_b32(v[76], v[78], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[77], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[77], v[30], v[77], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 1))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[82], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[82], 2, v[82]))
+  k.emit(ds_read_b32(v[79], v[82]))
+  k.emit(ds_read_b32(v[80], v[82], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[81], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[81], v[30], v[81], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 2))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[86], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[86], 2, v[86]))
+  k.emit(ds_read_b32(v[83], v[86]))
+  k.emit(ds_read_b32(v[84], v[86], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[85], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[85], v[30], v[85], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 3))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[90], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[90], 2, v[90]))
+  k.emit(ds_read_b32(v[87], v[90]))
+  k.emit(ds_read_b32(v[88], v[90], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[89], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[89], v[30], v[89], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 4))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[94], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[94], 2, v[94]))
+  k.emit(ds_read_b32(v[91], v[94]))
+  k.emit(ds_read_b32(v[92], v[94], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[93], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[93], v[30], v[93], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 5))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[98], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[98], 2, v[98]))
+  k.emit(ds_read_b32(v[95], v[98]))
+  k.emit(ds_read_b32(v[96], v[98], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[97], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[97], v[30], v[97], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 6))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[102], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[102], 2, v[102]))
+  k.emit(ds_read_b32(v[99], v[102]))
+  k.emit(ds_read_b32(v[100], v[102], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[101], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[101], v[30], v[101], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 7))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[106], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[106], 2, v[106]))
+  k.emit(ds_read_b32(v[103], v[106]))
+  k.emit(ds_read_b32(v[104], v[106], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[105], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[105], v[30], v[105], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[108], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[108], 2, v[108]))
+  k.emit(v_add_lshl_u32_e64(v[107], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[107], v[30], v[107], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 1))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[110], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[110], 2, v[110]))
+  k.emit(v_add_lshl_u32_e64(v[109], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[109], v[30], v[109], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 2))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[112], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[112], 2, v[112]))
+  k.emit(v_add_lshl_u32_e64(v[111], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[111], v[30], v[111], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 3))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[114], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[114], 2, v[114]))
+  k.emit(v_add_lshl_u32_e64(v[113], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[113], v[30], v[113], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 4))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[116], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[116], 2, v[116]))
+  k.emit(v_add_lshl_u32_e64(v[115], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[115], v[30], v[115], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 5))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[118], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[118], 2, v[118]))
+  k.emit(v_add_lshl_u32_e64(v[117], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[117], v[30], v[117], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 6))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[120], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[120], 2, v[120]))
+  k.emit(v_add_lshl_u32_e64(v[119], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[119], v[30], v[119], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 7))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[122], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[122], 2, v[122]))
+  k.emit(v_add_lshl_u32_e64(v[121], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[121], v[30], v[121], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[124], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[124], 2, v[124]))
+  k.emit(v_add_lshl_u32_e64(v[123], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[123], v[30], v[123], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 1))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[126], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[126], 2, v[126]))
+  k.emit(v_add_lshl_u32_e64(v[125], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[125], v[30], v[125], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 2))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[128], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[128], 2, v[128]))
+  k.emit(v_add_lshl_u32_e64(v[127], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[127], v[30], v[127], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 3))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[130], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[130], 2, v[130]))
+  k.emit(v_add_lshl_u32_e64(v[129], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[129], v[30], v[129], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 4))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[132], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[132], 2, v[132]))
+  k.emit(v_add_lshl_u32_e64(v[131], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[131], v[30], v[131], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 5))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[134], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[134], 2, v[134]))
+  k.emit(v_add_lshl_u32_e64(v[133], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[133], v[30], v[133], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 6))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[136], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[136], 2, v[136]))
+  k.emit(v_add_lshl_u32_e64(v[135], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[135], v[30], v[135], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 7))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[138], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[138], 2, v[138]))
+  k.emit(v_add_lshl_u32_e64(v[137], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[137], v[30], v[137], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[140], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[140], 2, v[140]))
+  k.emit(v_add_lshl_u32_e64(v[139], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[139], v[30], v[139], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 1))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[142], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[142], 2, v[142]))
+  k.emit(v_add_lshl_u32_e64(v[141], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[141], v[30], v[141], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 2))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[144], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[144], 2, v[144]))
+  k.emit(v_add_lshl_u32_e64(v[143], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[143], v[30], v[143], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 3))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[146], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[146], 2, v[146]))
+  k.emit(v_add_lshl_u32_e64(v[145], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[145], v[30], v[145], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 4))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[148], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[148], 2, v[148]))
+  k.emit(v_add_lshl_u32_e64(v[147], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[147], v[30], v[147], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 5))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[150], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[150], 2, v[150]))
+  k.emit(v_add_lshl_u32_e64(v[149], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[149], v[30], v[149], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 6))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[152], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[152], 2, v[152]))
+  k.emit(v_add_lshl_u32_e64(v[151], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[151], v[30], v[151], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 7))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[154], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[154], 2, v[154]))
+  k.emit(v_add_lshl_u32_e64(v[153], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[153], v[30], v[153], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[156], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[156], 2, v[156]))
+  k.emit(v_add_lshl_u32_e64(v[155], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[155], v[30], v[155], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 1))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[158], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[158], 2, v[158]))
+  k.emit(v_add_lshl_u32_e64(v[157], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[157], v[30], v[157], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 2))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[160], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[160], 2, v[160]))
+  k.emit(v_add_lshl_u32_e64(v[159], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[159], v[30], v[159], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 3))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[162], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[162], 2, v[162]))
+  k.emit(v_add_lshl_u32_e64(v[161], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[161], v[30], v[161], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 4))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[164], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[164], 2, v[164]))
+  k.emit(v_add_lshl_u32_e64(v[163], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[163], v[30], v[163], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 5))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[166], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[166], 2, v[166]))
+  k.emit(v_add_lshl_u32_e64(v[165], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[165], v[30], v[165], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 6))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[168], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[168], 2, v[168]))
+  k.emit(v_add_lshl_u32_e64(v[167], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[167], v[30], v[167], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 7))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[170], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[170], 2, v[170]))
+  k.emit(v_add_lshl_u32_e64(v[169], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[169], v[30], v[169], s[82:83]))
+  k.emit(v_accvgpr_read(v[35], v[35]))
+  k.emit(v_accvgpr_read(v[36], v[39]))
+  k.emit(v_accvgpr_read(v[37], v[43]))
+  k.emit(v_accvgpr_read(v[38], v[47]))
+  k.emit(v_accvgpr_read(v[39], v[51]))
+  k.emit(v_accvgpr_read(v[40], v[55]))
+  k.emit(v_accvgpr_read(v[41], v[59]))
+  k.emit(v_accvgpr_read(v[42], v[63]))
+  k.emit(v_accvgpr_read(v[43], v[67]))
+  k.emit(v_accvgpr_read(v[44], v[71]))
+  k.emit(v_accvgpr_read(v[45], v[75]))
+  k.emit(v_accvgpr_read(v[46], v[79]))
+  k.emit(v_accvgpr_read(v[47], v[83]))
+  k.emit(v_accvgpr_read(v[48], v[87]))
+  k.emit(v_accvgpr_read(v[49], v[91]))
+  k.emit(v_accvgpr_read(v[50], v[95]))
+  k.emit(v_accvgpr_read(v[51], v[99]))
+  k.emit(v_accvgpr_read(v[52], v[103]))
+  k.emit(v_accvgpr_read(v[53], v[107]))
+  k.emit(v_accvgpr_read(v[54], v[111]))
+  k.emit(v_accvgpr_read(v[55], v[115]))
+  k.emit(v_accvgpr_read(v[56], v[119]))
+  k.emit(v_accvgpr_read(v[57], v[123]))
+  k.emit(v_accvgpr_read(v[58], v[127]))
+  k.emit(v_accvgpr_read(v[59], v[131]))
+  k.emit(v_accvgpr_read(v[60], v[135]))
+  k.emit(v_accvgpr_read(v[61], v[139]))
+  k.emit(v_accvgpr_read(v[62], v[143]))
+  k.emit(v_accvgpr_read(v[63], v[147]))
+  k.emit(v_accvgpr_read(v[64], v[151]))
+  k.emit(v_accvgpr_read(v[65], v[155]))
+  k.emit(v_accvgpr_read(v[66], v[159]))
+  k.emit(v_accvgpr_read(v[67], v[163]))
+  k.emit(v_accvgpr_read(v[68], v[167]))
+  k.emit(v_accvgpr_read(v[69], v[171]))
+  k.emit(v_accvgpr_read(v[70], v[175]))
+  k.emit(v_accvgpr_read(v[71], v[179]))
+  k.emit(v_accvgpr_read(v[72], v[183]))
+  k.emit(v_accvgpr_read(v[73], v[187]))
+  k.emit(v_accvgpr_read(v[74], v[191]))
+  k.waitcnt(lgkm=0)
+  k.emit(v_mov_b32_e32(v[32], 4294901760))
+  k.emit(v_mov_b32_e32(v[33], 2147418112))
+  k.emit(v_mov_b32_e32(v[34], 32767))
+  k.emit(v_mul_f32_e32(v[35], v[76], v[35]))
+  k.emit(v_add_f32_e32(v[22], v[75], v[35]))
+  k.emit(v_mov_b32_e32(v[35], v[22]))
+  k.emit(v_cvt_pk(v[35], v[35], v[35]))
+  k.emit(buffer_store_short(v[35], v[77], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[36], v[80], v[36]))
+  k.emit(v_add_f32_e32(v[22], v[79], v[36]))
+  k.emit(v_mov_b32_e32(v[36], v[22]))
+  k.emit(v_cvt_pk(v[36], v[36], v[36]))
+  k.emit(buffer_store_short(v[36], v[81], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[37], v[84], v[37]))
+  k.emit(v_add_f32_e32(v[22], v[83], v[37]))
+  k.emit(v_mov_b32_e32(v[37], v[22]))
+  k.emit(v_cvt_pk(v[37], v[37], v[37]))
+  k.emit(buffer_store_short(v[37], v[85], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[38], v[88], v[38]))
+  k.emit(v_add_f32_e32(v[22], v[87], v[38]))
+  k.emit(v_mov_b32_e32(v[38], v[22]))
+  k.emit(v_cvt_pk(v[38], v[38], v[38]))
+  k.emit(buffer_store_short(v[38], v[89], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[39], v[92], v[39]))
+  k.emit(v_add_f32_e32(v[22], v[91], v[39]))
+  k.emit(v_mov_b32_e32(v[39], v[22]))
+  k.emit(v_cvt_pk(v[39], v[39], v[39]))
+  k.emit(buffer_store_short(v[39], v[93], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[40], v[96], v[40]))
+  k.emit(v_add_f32_e32(v[22], v[95], v[40]))
+  k.emit(v_mov_b32_e32(v[40], v[22]))
+  k.emit(v_cvt_pk(v[40], v[40], v[40]))
+  k.emit(buffer_store_short(v[40], v[97], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[41], v[100], v[41]))
+  k.emit(v_add_f32_e32(v[22], v[99], v[41]))
+  k.emit(v_mov_b32_e32(v[41], v[22]))
+  k.emit(v_cvt_pk(v[41], v[41], v[41]))
+  k.emit(buffer_store_short(v[41], v[101], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[42], v[104], v[42]))
+  k.emit(v_add_f32_e32(v[22], v[103], v[42]))
+  k.emit(v_mov_b32_e32(v[42], v[22]))
+  k.emit(v_cvt_pk(v[42], v[42], v[42]))
+  k.emit(buffer_store_short(v[42], v[105], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[43], v[76], v[43]))
+  k.emit(v_add_f32_e32(v[22], v[75], v[43]))
+  k.emit(v_mov_b32_e32(v[43], v[22]))
+  k.emit(v_cvt_pk(v[43], v[43], v[43]))
+  k.emit(buffer_store_short(v[43], v[107], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[44], v[80], v[44]))
+  k.emit(v_add_f32_e32(v[22], v[79], v[44]))
+  k.emit(v_mov_b32_e32(v[44], v[22]))
+  k.emit(v_cvt_pk(v[44], v[44], v[44]))
+  k.emit(buffer_store_short(v[44], v[109], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[45], v[84], v[45]))
+  k.emit(v_add_f32_e32(v[22], v[83], v[45]))
+  k.emit(v_mov_b32_e32(v[45], v[22]))
+  k.emit(v_cvt_pk(v[45], v[45], v[45]))
+  k.emit(buffer_store_short(v[45], v[111], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[46], v[88], v[46]))
+  k.emit(v_add_f32_e32(v[22], v[87], v[46]))
+  k.emit(v_mov_b32_e32(v[46], v[22]))
+  k.emit(v_cvt_pk(v[46], v[46], v[46]))
+  k.emit(buffer_store_short(v[46], v[113], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[47], v[92], v[47]))
+  k.emit(v_add_f32_e32(v[22], v[91], v[47]))
+  k.emit(v_mov_b32_e32(v[47], v[22]))
+  k.emit(v_cvt_pk(v[47], v[47], v[47]))
+  k.emit(buffer_store_short(v[47], v[115], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[48], v[96], v[48]))
+  k.emit(v_add_f32_e32(v[22], v[95], v[48]))
+  k.emit(v_mov_b32_e32(v[48], v[22]))
+  k.emit(v_cvt_pk(v[48], v[48], v[48]))
+  k.emit(buffer_store_short(v[48], v[117], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[49], v[100], v[49]))
+  k.emit(v_add_f32_e32(v[22], v[99], v[49]))
+  k.emit(v_mov_b32_e32(v[49], v[22]))
+  k.emit(v_cvt_pk(v[49], v[49], v[49]))
+  k.emit(buffer_store_short(v[49], v[119], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[50], v[104], v[50]))
+  k.emit(v_add_f32_e32(v[22], v[103], v[50]))
+  k.emit(v_mov_b32_e32(v[50], v[22]))
+  k.emit(v_cvt_pk(v[50], v[50], v[50]))
+  k.emit(buffer_store_short(v[50], v[121], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[51], v[76], v[51]))
+  k.emit(v_add_f32_e32(v[22], v[75], v[51]))
+  k.emit(v_mov_b32_e32(v[51], v[22]))
+  k.emit(v_cvt_pk(v[51], v[51], v[51]))
+  k.emit(buffer_store_short(v[51], v[123], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[52], v[80], v[52]))
+  k.emit(v_add_f32_e32(v[22], v[79], v[52]))
+  k.emit(v_mov_b32_e32(v[52], v[22]))
+  k.emit(v_cvt_pk(v[52], v[52], v[52]))
+  k.emit(buffer_store_short(v[52], v[125], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[53], v[84], v[53]))
+  k.emit(v_add_f32_e32(v[22], v[83], v[53]))
+  k.emit(v_mov_b32_e32(v[53], v[22]))
+  k.emit(v_cvt_pk(v[53], v[53], v[53]))
+  k.emit(buffer_store_short(v[53], v[127], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[54], v[88], v[54]))
+  k.emit(v_add_f32_e32(v[22], v[87], v[54]))
+  k.emit(v_mov_b32_e32(v[54], v[22]))
+  k.emit(v_cvt_pk(v[54], v[54], v[54]))
+  k.emit(buffer_store_short(v[54], v[129], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[55], v[92], v[55]))
+  k.emit(v_add_f32_e32(v[22], v[91], v[55]))
+  k.emit(v_mov_b32_e32(v[55], v[22]))
+  k.emit(v_cvt_pk(v[55], v[55], v[55]))
+  k.emit(buffer_store_short(v[55], v[131], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[56], v[96], v[56]))
+  k.emit(v_add_f32_e32(v[22], v[95], v[56]))
+  k.emit(v_mov_b32_e32(v[56], v[22]))
+  k.emit(v_cvt_pk(v[56], v[56], v[56]))
+  k.emit(buffer_store_short(v[56], v[133], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[57], v[100], v[57]))
+  k.emit(v_add_f32_e32(v[22], v[99], v[57]))
+  k.emit(v_mov_b32_e32(v[57], v[22]))
+  k.emit(v_cvt_pk(v[57], v[57], v[57]))
+  k.emit(buffer_store_short(v[57], v[135], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[58], v[104], v[58]))
+  k.emit(v_add_f32_e32(v[22], v[103], v[58]))
+  k.emit(v_mov_b32_e32(v[58], v[22]))
+  k.emit(v_cvt_pk(v[58], v[58], v[58]))
+  k.emit(buffer_store_short(v[58], v[137], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[59], v[76], v[59]))
+  k.emit(v_add_f32_e32(v[22], v[75], v[59]))
+  k.emit(v_mov_b32_e32(v[59], v[22]))
+  k.emit(v_cvt_pk(v[59], v[59], v[59]))
+  k.emit(buffer_store_short(v[59], v[139], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[60], v[80], v[60]))
+  k.emit(v_add_f32_e32(v[22], v[79], v[60]))
+  k.emit(v_mov_b32_e32(v[60], v[22]))
+  k.emit(v_cvt_pk(v[60], v[60], v[60]))
+  k.emit(buffer_store_short(v[60], v[141], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[61], v[84], v[61]))
+  k.emit(v_add_f32_e32(v[22], v[83], v[61]))
+  k.emit(v_mov_b32_e32(v[61], v[22]))
+  k.emit(v_cvt_pk(v[61], v[61], v[61]))
+  k.emit(buffer_store_short(v[61], v[143], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[62], v[88], v[62]))
+  k.emit(v_add_f32_e32(v[22], v[87], v[62]))
+  k.emit(v_mov_b32_e32(v[62], v[22]))
+  k.emit(v_cvt_pk(v[62], v[62], v[62]))
+  k.emit(buffer_store_short(v[62], v[145], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[63], v[92], v[63]))
+  k.emit(v_add_f32_e32(v[22], v[91], v[63]))
+  k.emit(v_mov_b32_e32(v[63], v[22]))
+  k.emit(v_cvt_pk(v[63], v[63], v[63]))
+  k.emit(buffer_store_short(v[63], v[147], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[64], v[96], v[64]))
+  k.emit(v_add_f32_e32(v[22], v[95], v[64]))
+  k.emit(v_mov_b32_e32(v[64], v[22]))
+  k.emit(v_cvt_pk(v[64], v[64], v[64]))
+  k.emit(buffer_store_short(v[64], v[149], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[65], v[100], v[65]))
+  k.emit(v_add_f32_e32(v[22], v[99], v[65]))
+  k.emit(v_mov_b32_e32(v[65], v[22]))
+  k.emit(v_cvt_pk(v[65], v[65], v[65]))
+  k.emit(buffer_store_short(v[65], v[151], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[66], v[104], v[66]))
+  k.emit(v_add_f32_e32(v[22], v[103], v[66]))
+  k.emit(v_mov_b32_e32(v[66], v[22]))
+  k.emit(v_cvt_pk(v[66], v[66], v[66]))
+  k.emit(buffer_store_short(v[66], v[153], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[67], v[76], v[67]))
+  k.emit(v_add_f32_e32(v[22], v[75], v[67]))
+  k.emit(v_mov_b32_e32(v[67], v[22]))
+  k.emit(v_cvt_pk(v[67], v[67], v[67]))
+  k.emit(buffer_store_short(v[67], v[155], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[68], v[80], v[68]))
+  k.emit(v_add_f32_e32(v[22], v[79], v[68]))
+  k.emit(v_mov_b32_e32(v[68], v[22]))
+  k.emit(v_cvt_pk(v[68], v[68], v[68]))
+  k.emit(buffer_store_short(v[68], v[157], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[69], v[84], v[69]))
+  k.emit(v_add_f32_e32(v[22], v[83], v[69]))
+  k.emit(v_mov_b32_e32(v[69], v[22]))
+  k.emit(v_cvt_pk(v[69], v[69], v[69]))
+  k.emit(buffer_store_short(v[69], v[159], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[70], v[88], v[70]))
+  k.emit(v_add_f32_e32(v[22], v[87], v[70]))
+  k.emit(v_mov_b32_e32(v[70], v[22]))
+  k.emit(v_cvt_pk(v[70], v[70], v[70]))
+  k.emit(buffer_store_short(v[70], v[161], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[71], v[92], v[71]))
+  k.emit(v_add_f32_e32(v[22], v[91], v[71]))
+  k.emit(v_mov_b32_e32(v[71], v[22]))
+  k.emit(v_cvt_pk(v[71], v[71], v[71]))
+  k.emit(buffer_store_short(v[71], v[163], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[72], v[96], v[72]))
+  k.emit(v_add_f32_e32(v[22], v[95], v[72]))
+  k.emit(v_mov_b32_e32(v[72], v[22]))
+  k.emit(v_cvt_pk(v[72], v[72], v[72]))
+  k.emit(buffer_store_short(v[72], v[165], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[73], v[100], v[73]))
+  k.emit(v_add_f32_e32(v[22], v[99], v[73]))
+  k.emit(v_mov_b32_e32(v[73], v[22]))
+  k.emit(v_cvt_pk(v[73], v[73], v[73]))
+  k.emit(buffer_store_short(v[73], v[167], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[74], v[104], v[74]))
+  k.emit(v_add_f32_e32(v[22], v[103], v[74]))
+  k.emit(v_mov_b32_e32(v[74], v[22]))
+  k.emit(v_cvt_pk(v[74], v[74], v[74]))
+  k.emit(buffer_store_short(v[74], v[169], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_nop())
+  k.emit(v_mov_b32_e32(v[30], 2147483648))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[54], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[54], 2, v[54]))
+  k.emit(ds_read_b32(v[51], v[54]))
+  k.emit(ds_read_b32(v[52], v[54], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[53], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[53], v[30], v[53], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 1))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[58], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[58], 2, v[58]))
+  k.emit(ds_read_b32(v[55], v[58]))
+  k.emit(ds_read_b32(v[56], v[58], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[57], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[57], v[30], v[57], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 2))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[62], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[62], 2, v[62]))
+  k.emit(ds_read_b32(v[59], v[62]))
+  k.emit(ds_read_b32(v[60], v[62], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[61], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[61], v[30], v[61], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 3))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[66], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[66], 2, v[66]))
+  k.emit(ds_read_b32(v[63], v[66]))
+  k.emit(ds_read_b32(v[64], v[66], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[65], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[65], v[30], v[65], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 4))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[70], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[70], 2, v[70]))
+  k.emit(ds_read_b32(v[67], v[70]))
+  k.emit(ds_read_b32(v[68], v[70], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[69], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[69], v[30], v[69], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 5))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[74], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[74], 2, v[74]))
+  k.emit(ds_read_b32(v[71], v[74]))
+  k.emit(ds_read_b32(v[72], v[74], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[73], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[73], v[30], v[73], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 6))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[78], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[78], 2, v[78]))
+  k.emit(ds_read_b32(v[75], v[78]))
+  k.emit(ds_read_b32(v[76], v[78], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[77], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[77], v[30], v[77], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 7))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[82], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[82], 2, v[82]))
+  k.emit(ds_read_b32(v[79], v[82]))
+  k.emit(ds_read_b32(v[80], v[82], v[0], v[0], 0, 0, 4))
+  k.emit(v_add_lshl_u32_e64(v[81], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[81], v[30], v[81], s[82:83]))
+  k.emit(v_add_co_u32(v[19], VCC, v[19], 1))
+  k.emit(v_add_u32_e64(v[20], v[20], s[38]))
+  k.emit(v_add_u32_e64(v[21], v[21], s[36]))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[18], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[84], v[18], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[84], 2, v[84]))
+  k.emit(v_add_lshl_u32_e64(v[83], v[21], v[18], 1))
+  k.emit(v_cndmask_b32_e64(v[83], v[30], v[83], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 1))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[86], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[86], 2, v[86]))
+  k.emit(v_add_lshl_u32_e64(v[85], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[85], v[30], v[85], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 2))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[88], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[88], 2, v[88]))
+  k.emit(v_add_lshl_u32_e64(v[87], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[87], v[30], v[87], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 3))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[90], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[90], 2, v[90]))
+  k.emit(v_add_lshl_u32_e64(v[89], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[89], v[30], v[89], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 4))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[92], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[92], 2, v[92]))
+  k.emit(v_add_lshl_u32_e64(v[91], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[91], v[30], v[91], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 5))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[94], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[94], 2, v[94]))
+  k.emit(v_add_lshl_u32_e64(v[93], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[93], v[30], v[93], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 6))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[96], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[96], 2, v[96]))
+  k.emit(v_add_lshl_u32_e64(v[95], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[95], v[30], v[95], s[82:83]))
+  k.emit(v_add_co_u32(v[22], VCC, v[18], 7))
+  k.emit(v_cmp_lt_u32_e64(s[78:79], v[22], s[20]))
+  k.emit(v_cmp_lt_u32_e64(s[82:83], v[19], s[21]))
+  k.emit(s_and_b64(s[82:83], s[78:79], s[82:83]))
+  k.emit(s_mul_i32(s[78], 256, s[2]))
+  k.emit(v_sub_u32_e64(v[98], v[22], s[78]))
+  k.emit(v_lshlrev_b32_e32(v[98], 2, v[98]))
+  k.emit(v_add_lshl_u32_e64(v[97], v[21], v[22], 1))
+  k.emit(v_cndmask_b32_e64(v[97], v[30], v[97], s[82:83]))
+  k.emit(v_accvgpr_read(v[35], v[195]))
+  k.emit(v_accvgpr_read(v[36], v[199]))
+  k.emit(v_accvgpr_read(v[37], v[203]))
+  k.emit(v_accvgpr_read(v[38], v[207]))
+  k.emit(v_accvgpr_read(v[39], v[211]))
+  k.emit(v_accvgpr_read(v[40], v[215]))
+  k.emit(v_accvgpr_read(v[41], v[219]))
+  k.emit(v_accvgpr_read(v[42], v[223]))
+  k.emit(v_accvgpr_read(v[43], v[227]))
+  k.emit(v_accvgpr_read(v[44], v[231]))
+  k.emit(v_accvgpr_read(v[45], v[235]))
+  k.emit(v_accvgpr_read(v[46], v[239]))
+  k.emit(v_accvgpr_read(v[47], v[243]))
+  k.emit(v_accvgpr_read(v[48], v[247]))
+  k.emit(v_accvgpr_read(v[49], v[251]))
+  k.emit(v_accvgpr_read(v[50], v[255]))
+  k.waitcnt(lgkm=0)
+  k.emit(v_mov_b32_e32(v[32], 4294901760))
+  k.emit(v_mov_b32_e32(v[33], 2147418112))
+  k.emit(v_mov_b32_e32(v[34], 32767))
+  k.emit(v_mul_f32_e32(v[35], v[52], v[35]))
+  k.emit(v_add_f32_e32(v[22], v[51], v[35]))
+  k.emit(v_mov_b32_e32(v[35], v[22]))
+  k.emit(v_cvt_pk(v[35], v[35], v[35]))
+  k.emit(buffer_store_short(v[35], v[53], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[36], v[56], v[36]))
+  k.emit(v_add_f32_e32(v[22], v[55], v[36]))
+  k.emit(v_mov_b32_e32(v[36], v[22]))
+  k.emit(v_cvt_pk(v[36], v[36], v[36]))
+  k.emit(buffer_store_short(v[36], v[57], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[37], v[60], v[37]))
+  k.emit(v_add_f32_e32(v[22], v[59], v[37]))
+  k.emit(v_mov_b32_e32(v[37], v[22]))
+  k.emit(v_cvt_pk(v[37], v[37], v[37]))
+  k.emit(buffer_store_short(v[37], v[61], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[38], v[64], v[38]))
+  k.emit(v_add_f32_e32(v[22], v[63], v[38]))
+  k.emit(v_mov_b32_e32(v[38], v[22]))
+  k.emit(v_cvt_pk(v[38], v[38], v[38]))
+  k.emit(buffer_store_short(v[38], v[65], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[39], v[68], v[39]))
+  k.emit(v_add_f32_e32(v[22], v[67], v[39]))
+  k.emit(v_mov_b32_e32(v[39], v[22]))
+  k.emit(v_cvt_pk(v[39], v[39], v[39]))
+  k.emit(buffer_store_short(v[39], v[69], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[40], v[72], v[40]))
+  k.emit(v_add_f32_e32(v[22], v[71], v[40]))
+  k.emit(v_mov_b32_e32(v[40], v[22]))
+  k.emit(v_cvt_pk(v[40], v[40], v[40]))
+  k.emit(buffer_store_short(v[40], v[73], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[41], v[76], v[41]))
+  k.emit(v_add_f32_e32(v[22], v[75], v[41]))
+  k.emit(v_mov_b32_e32(v[41], v[22]))
+  k.emit(v_cvt_pk(v[41], v[41], v[41]))
+  k.emit(buffer_store_short(v[41], v[77], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[42], v[80], v[42]))
+  k.emit(v_add_f32_e32(v[22], v[79], v[42]))
+  k.emit(v_mov_b32_e32(v[42], v[22]))
+  k.emit(v_cvt_pk(v[42], v[42], v[42]))
+  k.emit(buffer_store_short(v[42], v[81], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[43], v[52], v[43]))
+  k.emit(v_add_f32_e32(v[22], v[51], v[43]))
+  k.emit(v_mov_b32_e32(v[43], v[22]))
+  k.emit(v_cvt_pk(v[43], v[43], v[43]))
+  k.emit(buffer_store_short(v[43], v[83], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[44], v[56], v[44]))
+  k.emit(v_add_f32_e32(v[22], v[55], v[44]))
+  k.emit(v_mov_b32_e32(v[44], v[22]))
+  k.emit(v_cvt_pk(v[44], v[44], v[44]))
+  k.emit(buffer_store_short(v[44], v[85], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[45], v[60], v[45]))
+  k.emit(v_add_f32_e32(v[22], v[59], v[45]))
+  k.emit(v_mov_b32_e32(v[45], v[22]))
+  k.emit(v_cvt_pk(v[45], v[45], v[45]))
+  k.emit(buffer_store_short(v[45], v[87], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[46], v[64], v[46]))
+  k.emit(v_add_f32_e32(v[22], v[63], v[46]))
+  k.emit(v_mov_b32_e32(v[46], v[22]))
+  k.emit(v_cvt_pk(v[46], v[46], v[46]))
+  k.emit(buffer_store_short(v[46], v[89], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[47], v[68], v[47]))
+  k.emit(v_add_f32_e32(v[22], v[67], v[47]))
+  k.emit(v_mov_b32_e32(v[47], v[22]))
+  k.emit(v_cvt_pk(v[47], v[47], v[47]))
+  k.emit(buffer_store_short(v[47], v[91], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[48], v[72], v[48]))
+  k.emit(v_add_f32_e32(v[22], v[71], v[48]))
+  k.emit(v_mov_b32_e32(v[48], v[22]))
+  k.emit(v_cvt_pk(v[48], v[48], v[48]))
+  k.emit(buffer_store_short(v[48], v[93], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[49], v[76], v[49]))
+  k.emit(v_add_f32_e32(v[22], v[75], v[49]))
+  k.emit(v_mov_b32_e32(v[49], v[22]))
+  k.emit(v_cvt_pk(v[49], v[49], v[49]))
+  k.emit(buffer_store_short(v[49], v[95], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(v_mul_f32_e32(v[50], v[80], v[50]))
+  k.emit(v_add_f32_e32(v[22], v[79], v[50]))
+  k.emit(v_mov_b32_e32(v[50], v[22]))
+  k.emit(v_cvt_pk(v[50], v[50], v[50]))
+  k.emit(buffer_store_short(v[50], v[97], s[12:15], 0, 0, 1, 0, 0, 0, 0, 0, 1))
+  k.emit(s_nop())
+  k.emit(s_branch(), target='GW_End_1')
+  k.label('GW_End_1')
+  k.emit(s_cmp_ge_u32(s[58], s[59]))
+  k.emit(s_cbranch_scc1(), target='KernelEnd')
+  k.emit(s_branch(), target='PersistentLoopStart')
+  k.label('KernelEnd')
+  k.emit(s_endpgm())
+  return k
diff --git a/extra/gemm/asm/cdna/gemm.py b/extra/gemm/asm/cdna/gemm.py
new file mode 100644
index 0000000000..1201966715
--- /dev/null
+++ b/extra/gemm/asm/cdna/gemm.py
@@ -0,0 +1,95 @@
+import atexit, functools
+from tinygrad.runtime.support.compiler_amd import HIPCompiler
+from tinygrad import Tensor, Device, dtypes
+from tinygrad.uop.ops import UOp, Ops, KernelInfo, AxisType
+from tinygrad.renderer import Estimates
+from tinygrad.helpers import getenv, all_same, dedup
+from extra.gemm.asm.cdna.asm import build_kernel, GEMM_ARGS
+
+# ** CDNA4 assembly gemm
+
+WORKGROUP_SIZE = 256
+
+def custom_asm_gemm(C:UOp, A:UOp, B:UOp, dname:str, arch:str, wg:int) -> UOp:
+  batch, M, K = A.shape
+  K2, N = B.shape[(1 if B.ndim == 3 else 0):]
+  assert K == K2
+  lidx = UOp.special(WORKGROUP_SIZE, "lidx0")
+  gidx = UOp.special(wg, "gidx0")
+  k = build_kernel(batch, M, N, K, A.dtype.base)
+  sink = UOp.sink(C.base, A.base, B.base, lidx, gidx,
+                  arg=KernelInfo(name=k.name, estimates=Estimates(ops=2*batch*M*N*K, mem=(batch*M*K + K*N + batch*M*N)*2)))
+  binary = HIPCompiler(arch).compile(k.to_asm())
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)),
+                               UOp(Ops.SOURCE, arg=k.to_text()), UOp(Ops.BINARY, arg=binary)))
+
+counters = {"used":0, "todos":[]}
+def todo(msg:str) -> bool: counters["todos"].append(msg); return False
+atexit.register(lambda: print(f'asm_gemm: {counters["used"]} used, {len(counters["todos"])} not used'))
+
+def can_use_asm_gemm(a:Tensor, b:Tensor) -> bool:
+  if a.dtype != b.dtype: return todo(f"dtypes must match {a.dtype} != {b.dtype}")
+  if a.dtype not in {dtypes.bfloat16, dtypes.float16}: return todo(f"only bfloat16/float16, got {a.dtype}")
+  # only sharding on the batch is tested, others might work too
+  if isinstance(a.device, tuple) and not (a.ndim == 3 and a.uop.axis == 0 and b.uop.axis is None):
+    return todo(f"sharding mismatch a.ndim={a.ndim} a.uop.axis={a.uop.axis} b.uop.axis={b.uop.axis}")
+  batch, M, K = (1, *a.shape) if a.ndim == 2 else a.shape
+  N = b.shape[1]
+  if isinstance(a.device, tuple): batch //= len(a.device)
+  if batch not in {1, 2}: return todo(f"GEMM batch size {batch}")
+  if (key:=(M, N, K)) not in GEMM_ARGS: return todo(f"GEMM shape not supported {key}")
+  return True
+
+# ** UOp gemm to test Tensor.custom_kernel multi and backward correctness on non cdna4
+# note: this can be removed after we have GEMM on mixins
+
+def custom_uop_gemm(C:UOp, A:UOp, B:UOp) -> UOp:
+  M, K = A.shape[0]*A.shape[1], A.shape[2]
+  K2, N = B.shape[(1 if B.ndim == 3 else 0):]
+  assert K == K2
+  m = UOp.range(M, 1, AxisType.LOOP)
+  n = UOp.range(N, 2, AxisType.LOOP)
+  k = UOp.range(K, 0, AxisType.REDUCE)
+  mul = (A.index((m*UOp.const(dtypes.index, K)+k))*B.index((k*UOp.const(dtypes.index, N)+n))).cast(dtypes.float32)
+  red = mul.reduce(k, arg=Ops.ADD, dtype=dtypes.float32).cast(C.dtype.base)
+  store = C.index((m*UOp.const(dtypes.index, N)+n), ptr=True).store(red).end(m, n)
+  return store.sink(arg=KernelInfo(name=f'uop_gemm_{M}_{N}_{K}'))
+
+# ** backward gemm, might use the asm gemm
+
+def custom_gemm_bw(gradient:UOp, kernel:UOp):
+  out, a, b = kernel.src
+  assert all_same([gradient.device, a.device, b.device, out.device])
+  a_t, b_t, g_t = Tensor(a, device=a.device), Tensor(b, device=a.device), Tensor(gradient, device=a.device)
+  grad_a = (g_t @ b_t.T).uop
+  a_T = a_t.transpose(-2, -1)
+  a_T = a_T.reshape(*a_T.shape[:-1], 1, a_T.shape[-1])
+  g_r = g_t.reshape(*g_t.shape[:-2], 1, *g_t.shape[-2:]).transpose(-1, -2)
+  grad_b = (a_T * g_r).sum((-1, 0)).uop
+  return (None, grad_a, grad_b)
+
+# ** main gemm function
+
+def asm_gemm(a:Tensor, b:Tensor) -> Tensor:
+  assert can_use_asm_gemm(a, b), f"{counters['todos'][-1]}"
+  counters["used"] += 1
+  squeeze = a.ndim == 2
+  if squeeze: a = a.unsqueeze(0)
+
+  batch, M, K = a.shape
+  N = b.shape[1]
+  is_multi = isinstance(a.device, tuple)
+
+  if is_multi:
+    out = Tensor(Tensor.empty(batch//len(a.device), M, N, dtype=a.dtype, device=a.device).uop.multi(0), device=a.device)
+  else:
+    out = Tensor.empty(batch, M, N, dtype=a.dtype, device=a.device)
+
+  dname = a.device[0] if is_multi else a.device
+  arch = getattr(Device[dname].renderer, "arch", None)
+  if arch.startswith("gfx950") and getenv("USE_ASM", 1):
+    numWG = GEMM_ARGS[(M, N, K)][0]
+    out = Tensor.custom_kernel(out, a, b, fxn=functools.partial(custom_asm_gemm, dname=dname, wg=numWG, arch=arch), grad_fxn=custom_gemm_bw)[0]
+  else:
+    out = Tensor.custom_kernel(out, a, b, fxn=custom_uop_gemm, grad_fxn=custom_gemm_bw)[0]
+  return out.squeeze(0) if squeeze else out
diff --git a/extra/gemm/asm/cdna/gemm.s b/extra/gemm/asm/cdna/gemm.s
deleted file mode 100644
index 81bc04e12f..0000000000
--- a/extra/gemm/asm/cdna/gemm.s
+++ /dev/null
@@ -1,1621 +0,0 @@
-	// ** global buffers
-	s_load_dwordx2  s[28:29], s[0:1], 0x0    // C
-	s_load_dwordx2  s[34:35], s[0:1], 0x08   // A
-	s_load_dwordx2  s[32:33], s[0:1], 0x10   // B
-	// ** others kernel args
-	s_load_dword    s24, s[0:1], 0x18        // N
-	s_waitcnt lgkmcnt(0)
-	// "info"
-	s_mov_b32 s51, 1             // gemm_info = 1
-	s_mov_b32 s53, 1             // kernel_info0 = 1
-	s_mov_b32 s11, 0x40010020    // kernel_info1 = 0x40010020
-	// sizes / strides
-	s_mov_b32 s25, s24           // sizesFree1 = N
-	s_mov_b32 s26, 1             // sizesFree2 = BATCH
-	s_mov_b32 s27, s24           // sizesSum0  = K (== N)
-	// Strides: major=N, minor=0 (addr = base + idx0*N + idx1*0)
-	s_mov_b32 s36, s24           // strideD0
-	s_mov_b32 s37, 0             // strideD1
-	s_mov_b32 s38, s24           // strideC0
-	s_mov_b32 s39, 0             // strideC1
-	s_mov_b32 s40, s24           // strideA0
-	s_mov_b32 s41, 0             // strideA1
-	s_mov_b32 s42, s24           // strideB0
-	s_mov_b32 s43, 0             // strideB1
-	// ** workgroup mapping
-	s_lshr_b32 s52, s51, 30                                    // 000000002924: 8F349E33
-	s_and_b32 s51, 0x3fffffff, s51                             // 000000002928: 863333FF 3FFFFFFF
-	s_cmp_eq_u32 s52, 0                                        // 000000002930: BF068034
-	s_and_b32 s10, s53, 0xffff0000                             // 000000002A70: 860AFF35 FFFF0000
-	s_lshr_b32 s10, s10, 16                                    // 000000002A78: 8F0A900A
-	s_and_b32 s50, s53, 0xffff                                 // 000000002A7C: 8632FF35 0000FFFF
-	s_mov_b32 s5, s52                                          // 000000002A84: BE850034
-	s_mov_b32 m0, 0x20800                                      // 000000002A88: BEFC00FF 00020800
-	v_mov_b32_e32 v134, v0                                     // 000000002A90: 7F0C0300
-	s_lshr_b32 s60, s11, 16                                    // 000000002A94: 8F3C900B
-	s_ff1_i32_b32 s60, s60                                     // 000000002A98: BEBC103C
-	s_lshr_b32 s61, s11, 22                                    // 000000002A9C: 8F3D960B
-	v_and_b32_e32 v5, 63, v134                                 // 000000002BB0: 260B0CBF
-	v_and_b32_e32 v4, 15, v5                                   // 000000002BB4: 26080A8F
-	v_lshlrev_b32_e32 v4, 6, v4                                // 000000002BB8: 24080886
-	v_lshlrev_b32_e32 v4, 3, v4                                // 000000002BBC: 24080883
-	v_lshrrev_b32_e32 v5, 4, v5                                // 000000002BC0: 200A0A84
-	v_lshl_add_u32 v4, v5, 3, v4                               // 000000002BC4: D1FD0004 04110705
-	v_lshrrev_b32_e32 v8, 6, v134                              // 000000002BCC: 20110C86
-	v_and_b32_e32 v8, 1, v8                                    // 000000002BD0: 26101081
-	v_lshl_add_u32 v4, v8, 13, v4                              // 000000002BD4: D1FD0004 04111B08
-	v_and_b32_e32 v6, 63, v134                                 // 000000002BDC: 260D0CBF
-	v_and_b32_e32 v5, 15, v6                                   // 000000002BE0: 260A0C8F
-	v_lshlrev_b32_e32 v5, 6, v5                                // 000000002BE4: 240A0A86
-	v_lshlrev_b32_e32 v5, 3, v5                                // 000000002BE8: 240A0A83
-	v_lshrrev_b32_e32 v6, 4, v6                                // 000000002BEC: 200C0C84
-	v_lshl_add_u32 v5, v6, 3, v5                               // 000000002BF0: D1FD0005 04150706
-	v_lshrrev_b32_e32 v7, 7, v134                              // 000000002BF8: 200F0C87
-	v_and_b32_e32 v7, 1, v7                                    // 000000002BFC: 260E0E81
-	v_lshl_add_u32 v5, v7, 13, v5                              // 000000002C00: D1FD0005 04151B07
-	v_lshrrev_b32_e32 v6, 6, v134                              // 000000002C08: 200D0C86
-	v_lshrrev_b32_e32 v6, 2, v6                                // 000000002C0C: 200C0C82
-	s_mov_b32 s53, 64                                          // 000000002C10: BEB500C0
-	v_mul_lo_u32 v6, s53, v6                                   // 000000002C14: D2850006 00020C35
-	v_add_lshl_u32 v2, v6, v4, 1                               // 000000002C1C: D1FE0002 02060906
-	v_lshrrev_b32_e32 v7, 10, v2                               // 000000002C24: 200E048A
-	v_lshl_add_u32 v2, v7, 4, v2                               // 000000002C28: D1FD0002 04090907
-	v_lshrrev_b32_e32 v4, 6, v134                              // 000000002C30: 20090C86
-	v_lshrrev_b32_e32 v4, 2, v4                                // 000000002C34: 20080882
-	v_mul_lo_u32 v4, s53, v4                                   // 000000002C38: D2850004 00020835
-	v_add_lshl_u32 v3, v4, v5, 1                               // 000000002C40: D1FE0003 02060B04
-	v_lshrrev_b32_e32 v6, 10, v3                               // 000000002C48: 200C068A
-	v_lshl_add_u32 v3, v6, 4, v3                               // 000000002C4C: D1FD0003 040D0906
-	v_add_co_u32_e32 v3, vcc, 0x8200, v3                       // 000000002C54: 320606FF 00008200
-	v_add_u32_e32 v132, 0x10400, v2                            // 000000002C5C: 690804FF 00010400
-	v_xor_b32_e32 v132, v132, v2                               // 000000002C64: 2B080584
-	v_add_u32_e32 v133, 0x10400, v3                            // 000000002C68: 690A06FF 00010400
-	v_xor_b32_e32 v133, v133, v3                               // 000000002C70: 2B0A0785
-	v_lshrrev_b32_e32 v4, 3, v134                              // 000000002C74: 20090C83
-	v_and_b32_e32 v5, 7, v134                                  // 000000002C78: 260B0C87
-	v_lshlrev_b32_e32 v5, 3, v5                                // 000000002C7C: 240A0A83
-	v_mov_b32_e32 v8, v5                                       // 000000002C80: 7E100305
-	v_lshrrev_b32_e32 v6, 3, v134                              // 000000002C84: 200D0C83
-	v_and_b32_e32 v7, 7, v134                                  // 000000002C88: 260F0C87
-	v_lshlrev_b32_e32 v7, 3, v7                                // 000000002C8C: 240E0E83
-	v_mov_b32_e32 v9, v7                                       // 000000002C90: 7E120307
-	v_mul_u32_u24_e32 v10, 64, v4                              // 000000002C94: 101408C0
-	v_add_lshl_u32 v10, v8, v10, 1                             // 000000002C98: D1FE000A 02061508
-	v_lshrrev_b32_e32 v12, 10, v10                             // 000000002CA0: 2018148A
-	v_lshl_add_u32 v10, v12, 4, v10                            // 000000002CA4: D1FD000A 0429090C
-	s_nop 0                                                    // 000000002CAC: BF800000
-	v_readfirstlane_b32 s46, v10                               // 000000002CB0: 7E5C050A
-	s_nop 0                                                    // 000000002CB4: BF800000
-	s_add_u32 s48, s46, 0x10400                                // 000000002CB8: 8030FF2E 00010400
-	s_xor_b32 s48, s48, s46                                    // 000000002CC0: 88302E30
-	v_mul_u32_u24_e32 v10, 64, v6                              // 000000002CC4: 10140CC0
-	v_add_lshl_u32 v10, v9, v10, 1                             // 000000002CC8: D1FE000A 02061509
-	v_lshrrev_b32_e32 v12, 10, v10                             // 000000002CD0: 2018148A
-	v_lshl_add_u32 v10, v12, 4, v10                            // 000000002CD4: D1FD000A 0429090C
-	v_add_co_u32_e32 v10, vcc, 0x8200, v10                     // 000000002CDC: 321414FF 00008200
-	s_nop 0                                                    // 000000002CE4: BF800000
-	v_readfirstlane_b32 s47, v10                               // 000000002CE8: 7E5E050A
-	s_nop 0                                                    // 000000002CEC: BF800000
-	s_add_u32 s49, s47, 0x10400                                // 000000002CF0: 8031FF2F 00010400
-	s_xor_b32 s49, s49, s47                                    // 000000002CF8: 88312F31
-	v_mov_b32_e32 v12, 0x100                                   // 000000002CFC: 7E1802FF 00000100
-	v_mov_b32_e32 v11, s24                                     // 000000002D04: 7E160218
-	v_cvt_f32_u32_e32 v10, v12                                 // 000000002D08: 7E140D0C
-	v_rcp_iflag_f32_e32 v10, v10                               // 000000002D0C: 7E14470A
-	v_cvt_f32_u32_e32 v13, v11                                 // 000000002D10: 7E1A0D0B
-	v_mul_f32_e32 v10, v10, v13                                // 000000002D14: 0A141B0A
-	v_cvt_u32_f32_e32 v10, v10                                 // 000000002D18: 7E140F0A
-	v_mul_u32_u24_e32 v13, v10, v12                            // 000000002D1C: 101A190A
-	v_sub_u32_e32 v13, v11, v13                                // 000000002D20: 6A1A1B0B
-	v_cmp_ne_u32_e64 vcc, v13, 0                               // 000000002D24: D0CD006A 0001010D
-	v_addc_co_u32_e64 v10, vcc, v10, 0, vcc                    // 000000002D2C: D11C6A0A 01A9010A
-	v_mov_b32_e32 v12, 0x100                                   // 000000002D34: 7E1802FF 00000100
-	v_mov_b32_e32 v11, s25                                     // 000000002D3C: 7E160219
-	v_readfirstlane_b32 s14, v10                               // 000000002D40: 7E1C050A
-	v_cvt_f32_u32_e32 v10, v12                                 // 000000002D44: 7E140D0C
-	v_rcp_iflag_f32_e32 v10, v10                               // 000000002D48: 7E14470A
-	v_cvt_f32_u32_e32 v13, v11                                 // 000000002D4C: 7E1A0D0B
-	v_mul_f32_e32 v10, v10, v13                                // 000000002D50: 0A141B0A
-	v_cvt_u32_f32_e32 v10, v10                                 // 000000002D54: 7E140F0A
-	v_mul_u32_u24_e32 v13, v10, v12                            // 000000002D58: 101A190A
-	v_sub_u32_e32 v13, v11, v13                                // 000000002D5C: 6A1A1B0B
-	v_cmp_ne_u32_e64 vcc, v13, 0                               // 000000002D60: D0CD006A 0001010D
-	v_addc_co_u32_e64 v10, vcc, v10, 0, vcc                    // 000000002D68: D11C6A0A 01A9010A
-	s_nop 0                                                    // 000000002D70: BF800000
-	v_readfirstlane_b32 s15, v10                               // 000000002D74: 7E1E050A
-	s_waitcnt lgkmcnt(0)                                       // 000000002D78: BF8CC07F
-	s_mul_i32 s52, s14, s15                                    // 000000002D7C: 92340F0E
-	s_and_b32 s53, s50, 0x3fff                                 // 000000002D80: 8635FF32 00003FFF
-	s_mul_i32 s52, s52, s53                                    // 000000002D88: 92343534
-	v_cvt_f32_u32_e32 v10, s52                                 // 000000002D8C: 7E140C34
-	v_rcp_iflag_f32_e32 v10, v10                               // 000000002D90: 7E14470A
-	v_cvt_f32_u32_e32 v11, s2                                  // 000000002D94: 7E160C02
-	v_mul_f32_e32 v10, v10, v11                                // 000000002D98: 0A14170A
-	v_cvt_u32_f32_e32 v10, v10                                 // 000000002D9C: 7E140F0A
-	v_mul_u32_u24_e64 v11, v10, s52                            // 000000002DA0: D108000B 0000690A
-	v_sub_u32_e32 v11, s2, v11                                 // 000000002DA8: 6A161602
-	v_cmpx_eq_u32_e64 exec, v11, s52                           // 000000002DAC: D0DA007E 0000690B
-	v_add_u32_e32 v10, 1, v10                                  // 000000002DB4: 68141481
-	s_mov_b64 exec, -1                                         // 000000002DB8: BEFE01C1
-	v_cmpx_gt_u32_e64 exec, v11, s52                           // 000000002DBC: D0DC007E 0000690B
-	v_sub_u32_e64 v10, v10, 1                                  // 000000002DC4: D135000A 0001030A
-	s_mov_b64 exec, -1                                         // 000000002DCC: BEFE01C1
-	v_readfirstlane_b32 s52, v10                               // 000000002DD0: 7E68050A
-	s_mov_b32 s4, s52                                          // 000000002DD4: BE840034
-	s_mul_i32 s52, s15, s14                                    // 000000002DD8: 92340E0F
-	s_mul_i32 s52, s52, s4                                     // 000000002DDC: 92340434
-	s_mul_i32 s52, s52, s53                                    // 000000002DE0: 92343534
-	s_sub_u32 s2, s2, s52                                      // 000000002DE4: 80823402
-	v_cvt_f32_u32_e32 v10, s14                                 // 000000002DE8: 7E140C0E
-	v_rcp_iflag_f32_e32 v10, v10                               // 000000002DEC: 7E14470A
-	v_cvt_f32_u32_e32 v11, s2                                  // 000000002DF0: 7E160C02
-	v_mul_f32_e32 v10, v10, v11                                // 000000002DF4: 0A14170A
-	v_cvt_u32_f32_e32 v10, v10                                 // 000000002DF8: 7E140F0A
-	v_mul_u32_u24_e64 v11, v10, s14                            // 000000002DFC: D108000B 00001D0A
-	v_sub_u32_e32 v11, s2, v11                                 // 000000002E04: 6A161602
-	v_cmpx_eq_u32_e64 exec, v11, s14                           // 000000002E08: D0DA007E 00001D0B
-	v_add_u32_e32 v10, 1, v10                                  // 000000002E10: 68141481
-	s_mov_b64 exec, -1                                         // 000000002E14: BEFE01C1
-	v_cmpx_gt_u32_e64 exec, v11, s14                           // 000000002E18: D0DC007E 00001D0B
-	v_sub_u32_e64 v10, v10, 1                                  // 000000002E20: D135000A 0001030A
-	s_mov_b64 exec, -1                                         // 000000002E28: BEFE01C1
-	v_readfirstlane_b32 s52, v10                               // 000000002E2C: 7E68050A
-	s_mov_b32 s3, s52                                          // 000000002E30: BE830034
-	s_mul_i32 s52, s3, s14                                     // 000000002E34: 92340E03
-	s_sub_u32 s2, s2, s52                                      // 000000002E38: 80823402
-	s_sub_u32 s32, s32, 16                                     // 000000002E3C: 80A09020
-	s_subb_u32 s33, s33, 0                                     // 000000002E40: 82A18021
-	s_sub_u32 s34, s34, 16                                     // 000000002E44: 80A29022
-	s_subb_u32 s35, s35, 0                                     // 000000002E48: 82A38023
-	s_and_b32 s84, s50, 0x3fff                                 // 000000002E5C: 8654FF32 00003FFF
-	s_mov_b64 s[6:7], 0                                        // 000000002F48: BE860180
-	s_mov_b32 s8, 1                                            // 000000002F4C: BE880081
-	s_mov_b32 s9, 1                                            // 000000002F50: BE890081
-
-	s_sext_i32_i16 s11, s11                                    // 000000002F54: BE8B170B
-	v_mul_lo_u32 v10, s40, v4                                  // 0000000031B8: D285000A 00020828
-	v_add_co_u32_e32 v0, vcc, v5, v10                          // 0000000031C0: 32001505
-	v_add_u32_e32 v0, 8, v0                                    // 0000000031C4: 68000088
-	v_lshlrev_b32_e32 v0, 1, v0                                // 0000000031C8: 24000081
-	s_mul_i32 s70, s40, 32                                     // 0000000031CC: 9246A028
-	s_lshl_b32 s70, s70, 1                                     // 0000000031D0: 8E468146
-	s_mul_i32 s71, s40, 64                                     // 0000000031D4: 9247C028
-	s_lshl_b32 s71, s71, 1                                     // 0000000031D8: 8E478147
-	s_mul_i32 s72, s40, 0x60                                   // 0000000031DC: 9248FF28 00000060
-	s_lshl_b32 s72, s72, 1                                     // 0000000031E4: 8E488148
-	s_mul_i32 s73, s40, 0x80                                   // 0000000031E8: 9249FF28 00000080
-	s_lshl_b32 s73, s73, 1                                     // 0000000031F0: 8E498149
-	s_mul_i32 s74, s40, 0xa0                                   // 0000000031F4: 924AFF28 000000A0
-	s_lshl_b32 s74, s74, 1                                     // 0000000031FC: 8E4A814A
-	s_mul_i32 s75, s40, 0xc0                                   // 000000003200: 924BFF28 000000C0
-	s_lshl_b32 s75, s75, 1                                     // 000000003208: 8E4B814B
-	s_mul_i32 s76, s40, 0xe0                                   // 00000000320C: 924CFF28 000000E0
-	s_lshl_b32 s76, s76, 1                                     // 000000003214: 8E4C814C
-	v_mul_lo_u32 v10, s42, v6                                  // 000000003218: D285000A 00020C2A
-	v_add_co_u32_e32 v1, vcc, v7, v10                          // 000000003220: 32021507
-	v_add_u32_e32 v1, 8, v1                                    // 000000003224: 68020288
-	v_lshlrev_b32_e32 v1, 1, v1                                // 000000003228: 24020281
-	s_mul_i32 s77, s42, 32                                     // 00000000322C: 924DA02A
-	s_lshl_b32 s77, s77, 1                                     // 000000003230: 8E4D814D
-	s_mul_i32 s78, s42, 64                                     // 000000003234: 924EC02A
-	s_lshl_b32 s78, s78, 1                                     // 000000003238: 8E4E814E
-	s_mul_i32 s79, s42, 0x60                                   // 00000000323C: 924FFF2A 00000060
-	s_lshl_b32 s79, s79, 1                                     // 000000003244: 8E4F814F
-	s_mul_i32 s80, s42, 0x80                                   // 000000003248: 9250FF2A 00000080
-	s_lshl_b32 s80, s80, 1                                     // 000000003250: 8E508150
-	s_mul_i32 s81, s42, 0xa0                                   // 000000003254: 9251FF2A 000000A0
-	s_lshl_b32 s81, s81, 1                                     // 00000000325C: 8E518151
-	s_mul_i32 s82, s42, 0xc0                                   // 000000003260: 9252FF2A 000000C0
-	s_lshl_b32 s82, s82, 1                                     // 000000003268: 8E528152
-	s_mul_i32 s83, s42, 0xe0                                   // 00000000326C: 9253FF2A 000000E0
-	s_lshl_b32 s83, s83, 1                                     // 000000003274: 8E538153
-	s_mul_hi_u32 s87, s2, 0x100                                // 000000003278: 9657FF02 00000100
-	s_mul_i32 s86, s2, 0x100                                   // 000000003280: 9256FF02 00000100
-	s_mul_hi_u32 s87, s86, s40                                 // 000000003288: 96572856
-	s_mul_i32 s86, s86, s40                                    // 00000000328C: 92562856
-	s_and_b32 s84, s50, 0x8000                                 // 000000003290: 8654FF32 00008000
-	s_cbranch_scc1 skip_offset_A                                // 000000003298: BF850003
-	s_mul_hi_u32 s85, 64, s6                                   // 00000000329C: 965506C0
-	s_mul_i32 s84, 64, s6                                      // 0000000032A0: 925406C0
-
-skip_offset_A:
-	s_add_u32 s86, s86, s84                                    // 000000003330: 80565456
-	s_addc_u32 s87, s87, s85                                   // 000000003334: 82575557
-	s_mov_b64 s[60:61], 1                                      // 000000003338: BEBC0181
-	s_sub_u32 s84, s27, 1                                      // 00000000333C: 80D4811B
-	s_mul_hi_u32 s85, 1, s84                                   // 000000003340: 96555481
-	s_mul_i32 s84, 1, s84                                      // 000000003344: 92545481
-	s_add_u32 s60, s60, s84                                    // 000000003348: 803C543C
-	s_addc_u32 s61, s61, s85                                   // 00000000334C: 823D553D
-	s_sub_u32 s84, s24, 1                                      // 000000003350: 80D48118
-	s_mul_hi_u32 s85, s40, s84                                 // 000000003354: 96555428
-	s_mul_i32 s84, s40, s84                                    // 000000003358: 92545428
-	s_add_u32 s60, s60, s84                                    // 00000000335C: 803C543C
-	s_addc_u32 s61, s61, s85                                   // 000000003360: 823D553D
-	s_sub_u32 s60, s60, s86                                    // 000000003364: 80BC563C
-	s_subb_u32 s61, s61, s87                                   // 000000003368: 82BD573D
-	s_lshl_b64 s[60:61], s[60:61], 1                           // 00000000336C: 8EBC813C
-	s_add_u32 s60, s60, 16                                     // 000000003370: 803C903C
-	s_addc_u32 s61, s61, 0                                     // 000000003374: 823D803D
-	s_cmp_eq_u32 s61, 0                                        // 000000003378: BF06803D
-	s_cselect_b32 s54, s60, -1                                 // 00000000337C: 8536C13C
-	s_mul_hi_u32 s85, s41, s4                                  // 000000003380: 96550429
-	s_mul_i32 s84, s41, s4                                     // 000000003384: 92540429
-	s_add_u32 s86, s86, s84                                    // 000000003388: 80565456
-	s_addc_u32 s87, s87, s85                                   // 00000000338C: 82575557
-	s_lshl_b64 s[86:87], s[86:87], 1                           // 000000003390: 8ED68156
-	s_add_u32 s52, s32, s86                                    // 000000003394: 80345620
-	s_addc_u32 s53, s33, s87                                   // 000000003398: 82355721
-	s_mov_b32 s55, 0x20000                                     // 00000000339C: BEB700FF 00020000
-	s_mul_hi_u32 s87, s3, 0x100                                // 0000000033A4: 9657FF03 00000100
-	s_mul_i32 s86, s3, 0x100                                   // 0000000033AC: 9256FF03 00000100
-	s_mul_hi_u32 s87, s86, s42                                 // 0000000033B4: 96572A56
-	s_mul_i32 s86, s86, s42                                    // 0000000033B8: 92562A56
-	s_and_b32 s84, s50, 0x8000                                 // 0000000033BC: 8654FF32 00008000
-	s_cbranch_scc1 skip_offset_B                                // 0000000033C4: BF850003
-	s_mul_hi_u32 s85, 64, s6                                   // 0000000033C8: 965506C0
-	s_mul_i32 s84, 64, s6                                      // 0000000033CC: 925406C0
-
-skip_offset_B:
-	s_add_u32 s86, s86, s84                                    // 00000000345C: 80565456
-	s_addc_u32 s87, s87, s85                                   // 000000003460: 82575557
-	s_mov_b64 s[62:63], 1                                      // 000000003464: BEBE0181
-	s_sub_u32 s84, s27, 1                                      // 000000003468: 80D4811B
-	s_mul_hi_u32 s85, 1, s84                                   // 00000000346C: 96555481
-	s_mul_i32 s84, 1, s84                                      // 000000003470: 92545481
-	s_add_u32 s62, s62, s84                                    // 000000003474: 803E543E
-	s_addc_u32 s63, s63, s85                                   // 000000003478: 823F553F
-	s_sub_u32 s84, s25, 1                                      // 00000000347C: 80D48119
-	s_mul_hi_u32 s85, s42, s84                                 // 000000003480: 9655542A
-	s_mul_i32 s84, s42, s84                                    // 000000003484: 9254542A
-	s_add_u32 s62, s62, s84                                    // 000000003488: 803E543E
-	s_addc_u32 s63, s63, s85                                   // 00000000348C: 823F553F
-	s_sub_u32 s62, s62, s86                                    // 000000003490: 80BE563E
-	s_subb_u32 s63, s63, s87                                   // 000000003494: 82BF573F
-	s_lshl_b64 s[62:63], s[62:63], 1                           // 000000003498: 8EBE813E
-	s_add_u32 s62, s62, 16                                     // 00000000349C: 803E903E
-	s_addc_u32 s63, s63, 0                                     // 0000000034A0: 823F803F
-	s_cmp_eq_u32 s63, 0                                        // 0000000034A4: BF06803F
-	s_cselect_b32 s58, s62, -1                                 // 0000000034A8: 853AC13E
-	s_mul_hi_u32 s85, s43, s4                                  // 0000000034AC: 9655042B
-	s_mul_i32 s84, s43, s4                                     // 0000000034B0: 9254042B
-	s_add_u32 s86, s86, s84                                    // 0000000034B4: 80565456
-	s_addc_u32 s87, s87, s85                                   // 0000000034B8: 82575557
-	s_lshl_b64 s[86:87], s[86:87], 1                           // 0000000034BC: 8ED68156
-	s_add_u32 s56, s34, s86                                    // 0000000034C0: 80385622
-	s_addc_u32 s57, s35, s87                                   // 0000000034C4: 82395723
-	s_mov_b32 s59, 0x20000                                     // 0000000034C8: BEBB00FF 00020000
-	s_and_b32 s85, s50, 0x3fff                                 // 0000000034D0: 8655FF32 00003FFF
-	s_mul_i32 s85, s85, 0x80                                   // 0000000034D8: 9255FF55 00000080
-	s_and_b32 s84, s50, 0x8000                                 // 0000000034E0: 8654FF32 00008000
-	s_cselect_b32 s68, 0x80, s85                               // 0000000034E8: 854455FF 00000080
-	s_and_b32 s85, s50, 0x3fff                                 // 0000000034F0: 8655FF32 00003FFF
-	s_mul_i32 s85, s85, 0x80                                   // 0000000034F8: 9255FF55 00000080
-	s_and_b32 s84, s50, 0x8000                                 // 000000003500: 8654FF32 00008000
-	s_cselect_b32 s69, 0x80, s85                               // 000000003508: 854555FF 00000080
-	s_lshr_b32 s12, s27, 6                                     // 000000003510: 8F0C861B
-	s_and_b32 s84, s50, 0x3fff                                 // 000000003514: 8654FF32 00003FFF
-	s_mov_b32 s13, s12                                         // 000000003594: BE8D000C
-	s_and_b32 s86, s10, 0x1f00                                 // 000000003598: 8656FF0A 00001F00
-	s_lshr_b32 s86, s86, 8                                     // 0000000035A0: 8F568856
-	s_and_b32 s87, s10, 0xe000                                 // 0000000035A4: 8657FF0A 0000E000
-	s_and_b32 s10, s10, 0xff                                   // 0000000035AC: 860AFF0A 000000FF
-	s_mov_b32 s84, s10                                         // 0000000035B4: BED4000A
-	s_lshl_b32 s85, s84, s86                                   // 0000000035B8: 8E555654
-	s_cmp_ge_u32 s13, s85                                      // 0000000035BC: BF09550D
-	s_sub_u32 s85, s84, 1                                      // 0000000035CC: 80D58154
-	s_cmp_ge_u32 s84, 1                                        // 0000000035D0: BF098154
-	s_cselect_b32 s51, s85, 0                                  // 0000000035D4: 85338055
-	s_cmp_eq_u32 s87, 0x2000                                   // 0000000035E8: BF06FF57 00002000
-	s_and_b32 s51, s51, s84                                    // 000000003640: 86335433
-	s_lshl_b32 s51, s51, s86                                   // 000000003644: 8E335633
-	s_mul_hi_i32 s85, s51, s68                                 // 000000003648: 96D54433
-	s_mul_i32 s84, s51, s68                                    // 00000000364C: 92544433
-	s_mul_hi_i32 s65, s12, s68                                 // 000000003650: 96C1440C
-	s_mul_i32 s64, s12, s68                                    // 000000003654: 9240440C
-	s_sub_u32 s64, s68, s64                                    // 000000003658: 80C04044
-	s_subb_u32 s65, 0, s65                                     // 00000000365C: 82C14180
-	s_add_u32 s52, s52, s84                                    // 000000003660: 80345434
-	s_addc_u32 s53, s53, s85                                   // 000000003664: 82355535
-	s_sub_u32 s60, s60, s84                                    // 000000003668: 80BC543C
-	s_subb_u32 s61, s61, s85                                   // 00000000366C: 82BD553D
-	s_cmp_eq_u32 s61, 0                                        // 000000003670: BF06803D
-	s_cselect_b32 s54, s60, -1                                 // 000000003674: 8536C13C
-	s_mul_hi_i32 s85, s51, s69                                 // 000000003678: 96D54533
-	s_mul_i32 s84, s51, s69                                    // 00000000367C: 92544533
-	s_mul_hi_i32 s67, s12, s69                                 // 000000003680: 96C3450C
-	s_mul_i32 s66, s12, s69                                    // 000000003684: 9242450C
-	s_sub_u32 s66, s69, s66                                    // 000000003688: 80C24245
-	s_subb_u32 s67, 0, s67                                     // 00000000368C: 82C34380
-	s_add_u32 s56, s56, s84                                    // 000000003690: 80385438
-	s_addc_u32 s57, s57, s85                                   // 000000003694: 82395539
-	s_sub_u32 s62, s62, s84                                    // 000000003698: 80BE543E
-	s_subb_u32 s63, s63, s85                                   // 00000000369C: 82BF553F
-	s_cmp_eq_u32 s63, 0                                        // 0000000036A0: BF06803F
-	s_cselect_b32 s58, s62, -1                                 // 0000000036A4: 853AC13E
-	s_add_u32 s51, s51, 2                                      // 0000000036A8: 80338233
-	s_cmp_eq_u32 s12, 0                                        // 0000000036AC: BF06800C
-	s_cbranch_scc1 init_output_buffers                       // 0000000036B0: BF850092
-	s_mov_b32 m0, s46                                          // 0000000036B4: BEFC002E
-	buffer_load_dwordx4 v0, s[52:55], 0 offen lds              // 0000000036B8: E05D1000 800D0000
-	s_add_u32 m0, m0, 0x1040                                   // 0000000036C0: 807CFF7C 00001040
-	buffer_load_dwordx4 v0, s[52:55], s70 offen lds            // 0000000036C8: E05D1000 460D0000
-	s_add_u32 m0, m0, 0x1040                                   // 0000000036D0: 807CFF7C 00001040
-	buffer_load_dwordx4 v0, s[52:55], s71 offen lds            // 0000000036D8: E05D1000 470D0000
-	s_add_u32 m0, m0, 0x1040                                   // 0000000036E0: 807CFF7C 00001040
-	buffer_load_dwordx4 v0, s[52:55], s72 offen lds            // 0000000036E8: E05D1000 480D0000
-	s_add_u32 m0, m0, 0x1040                                   // 0000000036F0: 807CFF7C 00001040
-	buffer_load_dwordx4 v0, s[52:55], s73 offen lds            // 0000000036F8: E05D1000 490D0000
-	s_add_u32 m0, m0, 0x1040                                   // 000000003700: 807CFF7C 00001040
-	buffer_load_dwordx4 v0, s[52:55], s74 offen lds            // 000000003708: E05D1000 4A0D0000
-	s_add_u32 m0, m0, 0x1040                                   // 000000003710: 807CFF7C 00001040
-	buffer_load_dwordx4 v0, s[52:55], s75 offen lds            // 000000003718: E05D1000 4B0D0000
-	s_add_u32 m0, m0, 0x1040                                   // 000000003720: 807CFF7C 00001040
-	buffer_load_dwordx4 v0, s[52:55], s76 offen lds            // 000000003728: E05D1000 4C0D0000
-	s_mov_b32 m0, s47                                          // 000000003730: BEFC002F
-	buffer_load_dwordx4 v1, s[56:59], 0 offen lds              // 000000003734: E05D1000 800E0001
-	s_add_u32 m0, m0, 0x1040                                   // 00000000373C: 807CFF7C 00001040
-	buffer_load_dwordx4 v1, s[56:59], s77 offen lds            // 000000003744: E05D1000 4D0E0001
-	s_add_u32 m0, m0, 0x1040                                   // 00000000374C: 807CFF7C 00001040
-	buffer_load_dwordx4 v1, s[56:59], s78 offen lds            // 000000003754: E05D1000 4E0E0001
-	s_add_u32 m0, m0, 0x1040                                   // 00000000375C: 807CFF7C 00001040
-	v_accvgpr_write_b32 a0, 0                                  // 000000003764: D3D94000 18000080
-	v_accvgpr_write_b32 a1, 0                                  // 00000000376C: D3D94001 18000080
-	v_accvgpr_write_b32 a2, 0                                  // 000000003774: D3D94002 18000080
-	v_accvgpr_write_b32 a3, 0                                  // 00000000377C: D3D94003 18000080
-	v_accvgpr_write_b32 a4, 0                                  // 000000003784: D3D94004 18000080
-	v_accvgpr_write_b32 a5, 0                                  // 00000000378C: D3D94005 18000080
-	v_accvgpr_write_b32 a6, 0                                  // 000000003794: D3D94006 18000080
-	v_accvgpr_write_b32 a7, 0                                  // 00000000379C: D3D94007 18000080
-	v_accvgpr_write_b32 a8, 0                                  // 0000000037A4: D3D94008 18000080
-	v_accvgpr_write_b32 a9, 0                                  // 0000000037AC: D3D94009 18000080
-	v_accvgpr_write_b32 a10, 0                                 // 0000000037B4: D3D9400A 18000080
-	v_accvgpr_write_b32 a11, 0                                 // 0000000037BC: D3D9400B 18000080
-	v_accvgpr_write_b32 a12, 0                                 // 0000000037C4: D3D9400C 18000080
-	v_accvgpr_write_b32 a13, 0                                 // 0000000037CC: D3D9400D 18000080
-	v_accvgpr_write_b32 a14, 0                                 // 0000000037D4: D3D9400E 18000080
-	v_accvgpr_write_b32 a15, 0                                 // 0000000037DC: D3D9400F 18000080
-	v_mov_b64_e32 v[6:7], 0                                    // 0000000037E4: 7E0C7080
-	v_mov_b64_e32 v[8:9], 0                                    // 0000000037E8: 7E107080
-	v_mfma_f32_32x32x16_bf16 a[16:31], v[6:9], v[6:9], a[0:15] // 0000000037EC: D3B78010 04020D06
-	v_mfma_f32_32x32x16_bf16 a[32:47], v[6:9], v[6:9], a[0:15] // 0000000037F4: D3B78020 04020D06
-	v_mfma_f32_32x32x16_bf16 a[48:63], v[6:9], v[6:9], a[0:15] // 0000000037FC: D3B78030 04020D06
-	v_mfma_f32_32x32x16_bf16 a[64:79], v[6:9], v[6:9], a[0:15] // 000000003804: D3B78040 04020D06
-	v_mfma_f32_32x32x16_bf16 a[80:95], v[6:9], v[6:9], a[0:15] // 00000000380C: D3B78050 04020D06
-	v_mfma_f32_32x32x16_bf16 a[96:111], v[6:9], v[6:9], a[0:15]// 000000003814: D3B78060 04020D06
-	v_mfma_f32_32x32x16_bf16 a[112:127], v[6:9], v[6:9], a[0:15]// 00000000381C: D3B78070 04020D06
-	v_mfma_f32_32x32x16_bf16 a[128:143], v[6:9], v[6:9], a[0:15]// 000000003824: D3B78080 04020D06
-	buffer_load_dwordx4 v1, s[56:59], s79 offen lds            // 00000000382C: E05D1000 4F0E0001
-	s_add_u32 m0, m0, 0x1040                                   // 000000003834: 807CFF7C 00001040
-	v_mfma_f32_32x32x16_bf16 a[144:159], v[6:9], v[6:9], a[0:15]// 00000000383C: D3B78090 04020D06
-	v_mfma_f32_32x32x16_bf16 a[160:175], v[6:9], v[6:9], a[0:15]// 000000003844: D3B780A0 04020D06
-	v_mfma_f32_32x32x16_bf16 a[176:191], v[6:9], v[6:9], a[0:15]// 00000000384C: D3B780B0 04020D06
-	v_mfma_f32_32x32x16_bf16 a[192:207], v[6:9], v[6:9], a[0:15]// 000000003854: D3B780C0 04020D06
-	v_mfma_f32_32x32x16_bf16 a[208:223], v[6:9], v[6:9], a[0:15]// 00000000385C: D3B780D0 04020D06
-	v_mfma_f32_32x32x16_bf16 a[224:239], v[6:9], v[6:9], a[0:15]// 000000003864: D3B780E0 04020D06
-	v_mfma_f32_32x32x16_bf16 a[240:255], v[6:9], v[6:9], a[0:15]// 00000000386C: D3B780F0 04020D06
-	buffer_load_dwordx4 v1, s[56:59], s80 offen lds            // 000000003874: E05D1000 500E0001
-	s_add_u32 m0, m0, 0x1040                                   // 00000000387C: 807CFF7C 00001040
-	buffer_load_dwordx4 v1, s[56:59], s81 offen lds            // 000000003884: E05D1000 510E0001
-	s_add_u32 m0, m0, 0x1040                                   // 00000000388C: 807CFF7C 00001040
-	buffer_load_dwordx4 v1, s[56:59], s82 offen lds            // 000000003894: E05D1000 520E0001
-	s_add_u32 m0, m0, 0x1040                                   // 00000000389C: 807CFF7C 00001040
-	buffer_load_dwordx4 v1, s[56:59], s83 offen lds            // 0000000038A4: E05D1000 530E0001
-	s_add_u32 s86, s12, 1                                      // 0000000038AC: 8056810C
-	s_cmp_eq_u32 s51, s86                                      // 0000000038B0: BF065633
-	s_cselect_b32 s84, s64, s68                                // 0000000038B4: 85544440
-	s_cselect_b32 s85, s65, 0                                  // 0000000038B8: 85558041
-	s_add_u32 s52, s52, s84                                    // 0000000038BC: 80345434
-	s_addc_u32 s53, s53, s85                                   // 0000000038C0: 82355535
-	s_sub_u32 s60, s60, s84                                    // 0000000038C4: 80BC543C
-	s_subb_u32 s61, s61, s85                                   // 0000000038C8: 82BD553D
-	s_cmp_eq_u32 s61, 0                                        // 0000000038CC: BF06803D
-	s_cselect_b32 s54, s60, -1                                 // 0000000038D0: 8536C13C
-	s_add_u32 s86, s12, 1                                      // 0000000038D4: 8056810C
-	s_cmp_eq_u32 s51, s86                                      // 0000000038D8: BF065633
-	s_cselect_b32 s84, s66, s69                                // 0000000038DC: 85544542
-	s_cselect_b32 s85, s67, 0                                  // 0000000038E0: 85558043
-	s_add_u32 s56, s56, s84                                    // 0000000038E4: 80385438
-	s_addc_u32 s57, s57, s85                                   // 0000000038E8: 82395539
-	s_sub_u32 s62, s62, s84                                    // 0000000038EC: 80BE543E
-	s_subb_u32 s63, s63, s85                                   // 0000000038F0: 82BF553F
-	s_cmp_eq_u32 s63, 0                                        // 0000000038F4: BF06803F
-	s_cselect_b32 s58, s62, -1                                 // 0000000038F8: 853AC13E
-
-init_output_buffers:
-	s_mov_b64 s[16:17], s[28:29]                               // 0000000038FC: BE90011C
-	s_mov_b32 s18, 0x80000000                                  // 000000003900: BE9200FF 80000000
-	s_mov_b32 s19, 0x20000                                     // 000000003908: BE9300FF 00020000
-	s_mov_b64 s[20:21], s[30:31]                               // 000000003910: BE94011E
-	s_mov_b32 s22, 0x80000000                                  // 000000003914: BE9600FF 80000000
-	s_mov_b32 s23, 0x20000                                     // 00000000391C: BE9700FF 00020000
-	s_mul_i32 s86, 0x100, s3                                   // 000000003924: 925603FF 00000100
-	s_mul_hi_u32 s85, s86, s38                                 // 00000000392C: 96552656
-	s_mul_i32 s84, s86, s38                                    // 000000003930: 92542656
-	s_lshl_b64 s[84:85], s[84:85], s8                          // 000000003934: 8ED40854
-	s_add_u32 s20, s30, s84                                    // 000000003938: 8014541E
-	s_addc_u32 s21, s31, s85                                   // 00000000393C: 8215551F
-	s_mul_hi_u32 s85, s86, s36                                 // 000000003940: 96552456
-	s_mul_i32 s84, s86, s36                                    // 000000003944: 92542456
-	s_lshl_b64 s[84:85], s[84:85], s9                          // 000000003948: 8ED40954
-	s_add_u32 s16, s28, s84                                    // 00000000394C: 8010541C
-	s_addc_u32 s17, s29, s85                                   // 000000003950: 8211551D
-	s_mul_hi_u32 s85, s4, s39                                  // 000000003954: 96552704
-	s_mul_i32 s84, s4, s39                                     // 000000003958: 92542704
-	s_lshl_b64 s[84:85], s[84:85], s8                          // 00000000395C: 8ED40854
-	s_add_u32 s20, s20, s84                                    // 000000003960: 80145414
-	s_addc_u32 s21, s21, s85                                   // 000000003964: 82155515
-	s_mul_hi_u32 s85, s4, s37                                  // 000000003968: 96552504
-	s_mul_i32 s84, s4, s37                                     // 00000000396C: 92542504
-	s_lshl_b64 s[84:85], s[84:85], s9                          // 000000003970: 8ED40954
-	s_add_u32 s16, s16, s84                                    // 000000003974: 80105410
-	s_addc_u32 s17, s17, s85                                   // 000000003978: 82115511
-	s_mul_hi_u32 s85, s24, s6                                  // 00000000398C: 96550618
-	s_mul_i32 s84, s24, s6                                     // 000000003990: 92540618
-	s_sub_u32 s86, s25, 1                                      // 000000003994: 80D68119
-	s_mul_i32 s86, s86, s6                                     // 000000003998: 92560656
-	s_mul_hi_u32 s87, s86, s38                                 // 00000000399C: 96572656
-	s_mul_i32 s86, s86, s38                                    // 0000000039A0: 92562656
-	s_add_u32 s84, s84, s86                                    // 0000000039A4: 80545654
-	s_addc_u32 s85, s85, s87                                   // 0000000039A8: 82555755
-	s_sub_u32 s86, s26, 1                                      // 0000000039AC: 80D6811A
-	s_mul_i32 s86, s86, s6                                     // 0000000039B0: 92560656
-	s_mul_hi_u32 s87, s86, s39                                 // 0000000039B4: 96572756
-	s_mul_i32 s86, s86, s39                                    // 0000000039B8: 92562756
-	s_add_u32 s84, s84, s86                                    // 0000000039BC: 80545654
-	s_addc_u32 s85, s85, s87                                   // 0000000039C0: 82555755
-	s_lshl_b64 s[84:85], s[84:85], 2                           // 0000000039C4: 8ED48254
-	s_add_u32 s16, s16, s84                                    // 0000000039C8: 80105410
-	s_addc_u32 s17, s17, s85                                   // 0000000039CC: 82115511
-	s_xor_b32 s46, s48, s46                                    // 0000000039F0: 882E2E30
-	s_xor_b32 s47, s49, s47                                    // 0000000039F4: 882F2F31
-	s_cmp_eq_u32 s12, 1                                        // 0000000039F8: BF06810C
-	s_cbranch_scc1 after_prefetch                              // 0000000039FC: BF850040
-	s_mov_b32 m0, s46                                          // 000000003A00: BEFC002E
-	buffer_load_dwordx4 v0, s[52:55], 0 offen lds              // 000000003A04: E05D1000 800D0000
-	s_add_u32 m0, m0, 0x1040                                   // 000000003A0C: 807CFF7C 00001040
-	buffer_load_dwordx4 v0, s[52:55], s70 offen lds            // 000000003A14: E05D1000 460D0000
-	s_add_u32 m0, m0, 0x1040                                   // 000000003A1C: 807CFF7C 00001040
-	buffer_load_dwordx4 v0, s[52:55], s71 offen lds            // 000000003A24: E05D1000 470D0000
-	s_add_u32 m0, m0, 0x1040                                   // 000000003A2C: 807CFF7C 00001040
-	buffer_load_dwordx4 v0, s[52:55], s72 offen lds            // 000000003A34: E05D1000 480D0000
-	s_add_u32 m0, m0, 0x1040                                   // 000000003A3C: 807CFF7C 00001040
-	buffer_load_dwordx4 v0, s[52:55], s73 offen lds            // 000000003A44: E05D1000 490D0000
-	s_add_u32 m0, m0, 0x1040                                   // 000000003A4C: 807CFF7C 00001040
-	buffer_load_dwordx4 v0, s[52:55], s74 offen lds            // 000000003A54: E05D1000 4A0D0000
-	s_add_u32 m0, m0, 0x1040                                   // 000000003A5C: 807CFF7C 00001040
-	buffer_load_dwordx4 v0, s[52:55], s75 offen lds            // 000000003A64: E05D1000 4B0D0000
-	s_add_u32 m0, m0, 0x1040                                   // 000000003A6C: 807CFF7C 00001040
-	buffer_load_dwordx4 v0, s[52:55], s76 offen lds            // 000000003A74: E05D1000 4C0D0000
-	s_mov_b32 m0, s47                                          // 000000003A7C: BEFC002F
-	buffer_load_dwordx4 v1, s[56:59], 0 offen lds              // 000000003A80: E05D1000 800E0001
-	s_add_u32 m0, m0, 0x1040                                   // 000000003A88: 807CFF7C 00001040
-	buffer_load_dwordx4 v1, s[56:59], s77 offen lds            // 000000003A90: E05D1000 4D0E0001
-	s_add_u32 m0, m0, 0x1040                                   // 000000003A98: 807CFF7C 00001040
-	buffer_load_dwordx4 v1, s[56:59], s78 offen lds            // 000000003AA0: E05D1000 4E0E0001
-	s_add_u32 m0, m0, 0x1040                                   // 000000003AA8: 807CFF7C 00001040
-	buffer_load_dwordx4 v1, s[56:59], s79 offen lds            // 000000003AB0: E05D1000 4F0E0001
-	s_add_u32 m0, m0, 0x1040                                   // 000000003AB8: 807CFF7C 00001040
-	buffer_load_dwordx4 v1, s[56:59], s80 offen lds            // 000000003AC0: E05D1000 500E0001
-	s_add_u32 m0, m0, 0x1040                                   // 000000003AC8: 807CFF7C 00001040
-	buffer_load_dwordx4 v1, s[56:59], s81 offen lds            // 000000003AD0: E05D1000 510E0001
-	s_add_u32 m0, m0, 0x1040                                   // 000000003AD8: 807CFF7C 00001040
-	buffer_load_dwordx4 v1, s[56:59], s82 offen lds            // 000000003AE0: E05D1000 520E0001
-	s_add_u32 m0, m0, 0x1040                                   // 000000003AE8: 807CFF7C 00001040
-	buffer_load_dwordx4 v1, s[56:59], s83 offen lds            // 000000003AF0: E05D1000 530E0001
-	s_xor_b32 s46, s48, s46                                    // 000000003AF8: 882E2E30
-	s_xor_b32 s47, s49, s47                                    // 000000003AFC: 882F2F31
-
-after_prefetch:
-	s_waitcnt vmcnt(24)                                        // 000000003B00: BF8C4F78
-	s_barrier                                                  // 000000003B04: BF8A0000
-	ds_read_b128 v[4:7], v2                                    // 000000003B08: D9FE0000 04000002
-	ds_read_b128 v[8:11], v2 offset:128                        // 000000003B10: D9FE0080 08000002
-	ds_read_b128 v[12:15], v2 offset:256                       // 000000003B18: D9FE0100 0C000002
-	ds_read_b128 v[16:19], v2 offset:384                       // 000000003B20: D9FE0180 10000002
-	ds_read_b128 v[20:23], v2 offset:512                       // 000000003B28: D9FE0200 14000002
-	ds_read_b128 v[24:27], v2 offset:640                       // 000000003B30: D9FE0280 18000002
-	ds_read_b128 v[28:31], v2 offset:768                       // 000000003B38: D9FE0300 1C000002
-	ds_read_b128 v[32:35], v2 offset:896                       // 000000003B40: D9FE0380 20000002
-	s_waitcnt vmcnt(16)                                        // 000000003B48: BF8C4F70
-	s_barrier                                                  // 000000003B4C: BF8A0000
-	ds_read_b128 v[68:71], v3                                  // 000000003B50: D9FE0000 44000003
-	ds_read_b128 v[72:75], v3 offset:128                       // 000000003B58: D9FE0080 48000003
-	ds_read_b128 v[76:79], v3 offset:256                       // 000000003B60: D9FE0100 4C000003
-	ds_read_b128 v[80:83], v3 offset:384                       // 000000003B68: D9FE0180 50000003
-	ds_read_b128 v[84:87], v3 offset:512                       // 000000003B70: D9FE0200 54000003
-	ds_read_b128 v[88:91], v3 offset:640                       // 000000003B78: D9FE0280 58000003
-	ds_read_b128 v[92:95], v3 offset:768                       // 000000003B80: D9FE0300 5C000003
-	ds_read_b128 v[96:99], v3 offset:896                       // 000000003B88: D9FE0380 60000003
-	s_waitcnt lgkmcnt(0)                                       // 000000003B90: BF8CC07F
-	s_cmp_eq_u32 s12, 1                                        // 000000003B94: BF06810C
-	s_cbranch_scc1 final_compute                                // 000000003B98: BF8502E5
-	s_cmp_le_u32 s12, 2                                        // 000000003B9C: BF0B820C
-	s_cbranch_scc1 loop_epilogue                              // 000000003BA0: BF85019E
-
-main_loop:
-	v_mfma_f32_16x16x32_bf16 a[0:3], v[68:71], v[4:7], a[0:3]  // 000000003BA4: D3B58000 04020944
-	ds_read_b128 v[36:39], v2 offset:64                        // 000000003BAC: D9FE0040 24000002
-	v_mfma_f32_16x16x32_bf16 a[4:7], v[68:71], v[8:11], a[4:7] // 000000003BB4: D3B58004 04121144
-	s_cmp_eq_u32 s12, s51                                      // 000000003BBC: BF06330C
-	s_cselect_b32 s84, s64, s68                                // 000000003BC0: 85544440
-	v_mfma_f32_16x16x32_bf16 a[8:11], v[68:71], v[12:15], a[8:11]// 000000003BC4: D3B58008 04221944
-	ds_read_b128 v[40:43], v2 offset:192                       // 000000003BCC: D9FE00C0 28000002
-	v_mfma_f32_16x16x32_bf16 a[12:15], v[68:71], v[16:19], a[12:15]// 000000003BD4: D3B5800C 04322144
-	s_cselect_b32 s85, s65, 0                                  // 000000003BDC: 85558041
-	s_add_u32 s52, s52, s84                                    // 000000003BE0: 80345434
-	v_mfma_f32_16x16x32_bf16 a[16:19], v[68:71], v[20:23], a[16:19]// 000000003BE4: D3B58010 04422944
-	ds_read_b128 v[44:47], v2 offset:320                       // 000000003BEC: D9FE0140 2C000002
-	v_mfma_f32_16x16x32_bf16 a[20:23], v[68:71], v[24:27], a[20:23]// 000000003BF4: D3B58014 04523144
-	s_addc_u32 s53, s53, s85                                   // 000000003BFC: 82355535
-	s_sub_u32 s60, s60, s84                                    // 000000003C00: 80BC543C
-	v_mfma_f32_16x16x32_bf16 a[24:27], v[68:71], v[28:31], a[24:27]// 000000003C04: D3B58018 04623944
-	ds_read_b128 v[48:51], v2 offset:448                       // 000000003C0C: D9FE01C0 30000002
-	v_mfma_f32_16x16x32_bf16 a[28:31], v[68:71], v[32:35], a[28:31]// 000000003C14: D3B5801C 04724144
-	s_subb_u32 s61, s61, s85                                   // 000000003C1C: 82BD553D
-	s_cmp_eq_u32 s61, 0                                        // 000000003C20: BF06803D
-	v_mfma_f32_16x16x32_bf16 a[32:35], v[72:75], v[4:7], a[32:35]// 000000003C24: D3B58020 04820948
-	ds_read_b128 v[52:55], v2 offset:576                       // 000000003C2C: D9FE0240 34000002
-	v_mfma_f32_16x16x32_bf16 a[36:39], v[72:75], v[8:11], a[36:39]// 000000003C34: D3B58024 04921148
-	s_cselect_b32 s54, s60, -1                                 // 000000003C3C: 8536C13C
-	s_cmp_eq_u32 s12, s51                                      // 000000003C40: BF06330C
-	v_mfma_f32_16x16x32_bf16 a[40:43], v[72:75], v[12:15], a[40:43]// 000000003C44: D3B58028 04A21948
-	ds_read_b128 v[56:59], v2 offset:704                       // 000000003C4C: D9FE02C0 38000002
-	v_mfma_f32_16x16x32_bf16 a[44:47], v[72:75], v[16:19], a[44:47]// 000000003C54: D3B5802C 04B22148
-	s_cselect_b32 s84, s66, s69                                // 000000003C5C: 85544542
-	s_cselect_b32 s85, s67, 0                                  // 000000003C60: 85558043
-	v_mfma_f32_16x16x32_bf16 a[48:51], v[72:75], v[20:23], a[48:51]// 000000003C64: D3B58030 04C22948
-	ds_read_b128 v[60:63], v2 offset:832                       // 000000003C6C: D9FE0340 3C000002
-	v_mfma_f32_16x16x32_bf16 a[52:55], v[72:75], v[24:27], a[52:55]// 000000003C74: D3B58034 04D23148
-	s_add_u32 s56, s56, s84                                    // 000000003C7C: 80385438
-	s_addc_u32 s57, s57, s85                                   // 000000003C80: 82395539
-	v_mfma_f32_16x16x32_bf16 a[56:59], v[72:75], v[28:31], a[56:59]// 000000003C84: D3B58038 04E23948
-	ds_read_b128 v[64:67], v2 offset:960                       // 000000003C8C: D9FE03C0 40000002
-	v_mfma_f32_16x16x32_bf16 a[60:63], v[72:75], v[32:35], a[60:63]// 000000003C94: D3B5803C 04F24148
-	s_mov_b32 m0, s46                                          // 000000003C9C: BEFC002E
-	s_sub_u32 s62, s62, s84                                    // 000000003CA0: 80BE543E
-	v_mfma_f32_16x16x32_bf16 a[64:67], v[76:79], v[4:7], a[64:67]// 000000003CA4: D3B58040 0502094C
-	s_subb_u32 s63, s63, s85                                   // 000000003CAC: 82BF553F
-	s_cmp_eq_u32 s63, 0                                        // 000000003CB0: BF06803F
-	v_mfma_f32_16x16x32_bf16 a[68:71], v[76:79], v[8:11], a[68:71]// 000000003CB4: D3B58044 0512114C
-	s_cselect_b32 s58, s62, -1                                 // 000000003CBC: 853AC13E
-	v_mfma_f32_16x16x32_bf16 a[72:75], v[76:79], v[12:15], a[72:75]// 000000003CC0: D3B58048 0522194C
-	v_mfma_f32_16x16x32_bf16 a[76:79], v[76:79], v[16:19], a[76:79]// 000000003CC8: D3B5804C 0532214C
-	v_mfma_f32_16x16x32_bf16 a[80:83], v[76:79], v[20:23], a[80:83]// 000000003CD0: D3B58050 0542294C
-	s_waitcnt lgkmcnt(0)                                       // 000000003CD8: BF8CC07F
-	v_mfma_f32_16x16x32_bf16 a[84:87], v[76:79], v[24:27], a[84:87]// 000000003CDC: D3B58054 0552314C
-	s_barrier                                                  // 000000003CE4: BF8A0000
-	v_mfma_f32_16x16x32_bf16 a[88:91], v[76:79], v[28:31], a[88:91]// 000000003CE8: D3B58058 0562394C
-	buffer_load_dwordx4 v0, s[52:55], 0 offen lds              // 000000003CF0: E05D1000 800D0000
-	v_mfma_f32_16x16x32_bf16 a[92:95], v[76:79], v[32:35], a[92:95]// 000000003CF8: D3B5805C 0572414C
-	s_add_u32 m0, m0, 0x1040                                   // 000000003D00: 807CFF7C 00001040
-	v_mfma_f32_16x16x32_bf16 a[96:99], v[80:83], v[4:7], a[96:99]// 000000003D08: D3B58060 05820950
-	ds_read_b128 v[100:103], v3 offset:64                      // 000000003D10: D9FE0040 64000003
-	v_mfma_f32_16x16x32_bf16 a[100:103], v[80:83], v[8:11], a[100:103]// 000000003D18: D3B58064 05921150
-	buffer_load_dwordx4 v0, s[52:55], s70 offen lds            // 000000003D20: E05D1000 460D0000
-	v_mfma_f32_16x16x32_bf16 a[104:107], v[80:83], v[12:15], a[104:107]// 000000003D28: D3B58068 05A21950
-	s_add_u32 m0, m0, 0x1040                                   // 000000003D30: 807CFF7C 00001040
-	v_mfma_f32_16x16x32_bf16 a[108:111], v[80:83], v[16:19], a[108:111]// 000000003D38: D3B5806C 05B22150
-	ds_read_b128 v[104:107], v3 offset:192                     // 000000003D40: D9FE00C0 68000003
-	v_mfma_f32_16x16x32_bf16 a[112:115], v[80:83], v[20:23], a[112:115]// 000000003D48: D3B58070 05C22950
-	buffer_load_dwordx4 v0, s[52:55], s71 offen lds            // 000000003D50: E05D1000 470D0000
-	v_mfma_f32_16x16x32_bf16 a[116:119], v[80:83], v[24:27], a[116:119]// 000000003D58: D3B58074 05D23150
-	s_add_u32 m0, m0, 0x1040                                   // 000000003D60: 807CFF7C 00001040
-	v_mfma_f32_16x16x32_bf16 a[120:123], v[80:83], v[28:31], a[120:123]// 000000003D68: D3B58078 05E23950
-	ds_read_b128 v[108:111], v3 offset:320                     // 000000003D70: D9FE0140 6C000003
-	v_mfma_f32_16x16x32_bf16 a[124:127], v[80:83], v[32:35], a[124:127]// 000000003D78: D3B5807C 05F24150
-	buffer_load_dwordx4 v0, s[52:55], s72 offen lds            // 000000003D80: E05D1000 480D0000
-	v_mfma_f32_16x16x32_bf16 a[128:131], v[84:87], v[4:7], a[128:131]// 000000003D88: D3B58080 06020954
-	s_add_u32 m0, m0, 0x1040                                   // 000000003D90: 807CFF7C 00001040
-	v_mfma_f32_16x16x32_bf16 a[132:135], v[84:87], v[8:11], a[132:135]// 000000003D98: D3B58084 06121154
-	ds_read_b128 v[112:115], v3 offset:448                     // 000000003DA0: D9FE01C0 70000003
-	v_mfma_f32_16x16x32_bf16 a[136:139], v[84:87], v[12:15], a[136:139]// 000000003DA8: D3B58088 06221954
-	buffer_load_dwordx4 v0, s[52:55], s73 offen lds            // 000000003DB0: E05D1000 490D0000
-	v_mfma_f32_16x16x32_bf16 a[140:143], v[84:87], v[16:19], a[140:143]// 000000003DB8: D3B5808C 06322154
-	s_add_u32 m0, m0, 0x1040                                   // 000000003DC0: 807CFF7C 00001040
-	v_mfma_f32_16x16x32_bf16 a[144:147], v[84:87], v[20:23], a[144:147]// 000000003DC8: D3B58090 06422954
-	ds_read_b128 v[116:119], v3 offset:576                     // 000000003DD0: D9FE0240 74000003
-	v_mfma_f32_16x16x32_bf16 a[148:151], v[84:87], v[24:27], a[148:151]// 000000003DD8: D3B58094 06523154
-	v_mfma_f32_16x16x32_bf16 a[152:155], v[84:87], v[28:31], a[152:155]// 000000003DE0: D3B58098 06623954
-	ds_read_b128 v[120:123], v3 offset:704                     // 000000003DE8: D9FE02C0 78000003
-	v_mfma_f32_16x16x32_bf16 a[156:159], v[84:87], v[32:35], a[156:159]// 000000003DF0: D3B5809C 06724154
-	v_mfma_f32_16x16x32_bf16 a[160:163], v[88:91], v[4:7], a[160:163]// 000000003DF8: D3B580A0 06820958
-	ds_read_b128 v[124:127], v3 offset:832                     // 000000003E00: D9FE0340 7C000003
-	v_mfma_f32_16x16x32_bf16 a[164:167], v[88:91], v[8:11], a[164:167]// 000000003E08: D3B580A4 06921158
-	v_mfma_f32_16x16x32_bf16 a[168:171], v[88:91], v[12:15], a[168:171]// 000000003E10: D3B580A8 06A21958
-	ds_read_b128 v[128:131], v3 offset:960                     // 000000003E18: D9FE03C0 80000003
-	v_mfma_f32_16x16x32_bf16 a[172:175], v[88:91], v[16:19], a[172:175]// 000000003E20: D3B580AC 06B22158
-	v_mfma_f32_16x16x32_bf16 a[176:179], v[88:91], v[20:23], a[176:179]// 000000003E28: D3B580B0 06C22958
-	v_mfma_f32_16x16x32_bf16 a[180:183], v[88:91], v[24:27], a[180:183]// 000000003E30: D3B580B4 06D23158
-	v_mfma_f32_16x16x32_bf16 a[184:187], v[88:91], v[28:31], a[184:187]// 000000003E38: D3B580B8 06E23958
-	v_mfma_f32_16x16x32_bf16 a[188:191], v[88:91], v[32:35], a[188:191]// 000000003E40: D3B580BC 06F24158
-	v_mfma_f32_16x16x32_bf16 a[192:195], v[92:95], v[4:7], a[192:195]// 000000003E48: D3B580C0 0702095C
-	v_mfma_f32_16x16x32_bf16 a[196:199], v[92:95], v[8:11], a[196:199]// 000000003E50: D3B580C4 0712115C
-	v_mfma_f32_16x16x32_bf16 a[200:203], v[92:95], v[12:15], a[200:203]// 000000003E58: D3B580C8 0722195C
-	s_waitcnt lgkmcnt(0)                                       // 000000003E60: BF8CC07F
-	v_mfma_f32_16x16x32_bf16 a[204:207], v[92:95], v[16:19], a[204:207]// 000000003E64: D3B580CC 0732215C
-	s_barrier                                                  // 000000003E6C: BF8A0000
-	v_mfma_f32_16x16x32_bf16 a[208:211], v[92:95], v[20:23], a[208:211]// 000000003E70: D3B580D0 0742295C
-	buffer_load_dwordx4 v0, s[52:55], s74 offen lds            // 000000003E78: E05D1000 4A0D0000
-	v_mfma_f32_16x16x32_bf16 a[212:215], v[92:95], v[24:27], a[212:215]// 000000003E80: D3B580D4 0752315C
-	s_add_u32 m0, m0, 0x1040                                   // 000000003E88: 807CFF7C 00001040
-	v_mfma_f32_16x16x32_bf16 a[216:219], v[92:95], v[28:31], a[216:219]// 000000003E90: D3B580D8 0762395C
-	v_mfma_f32_16x16x32_bf16 a[220:223], v[92:95], v[32:35], a[220:223]// 000000003E98: D3B580DC 0772415C
-	buffer_load_dwordx4 v0, s[52:55], s75 offen lds            // 000000003EA0: E05D1000 4B0D0000
-	v_mfma_f32_16x16x32_bf16 a[224:227], v[96:99], v[4:7], a[224:227]// 000000003EA8: D3B580E0 07820960
-	s_add_u32 m0, m0, 0x1040                                   // 000000003EB0: 807CFF7C 00001040
-	v_mfma_f32_16x16x32_bf16 a[228:231], v[96:99], v[8:11], a[228:231]// 000000003EB8: D3B580E4 07921160
-	v_mfma_f32_16x16x32_bf16 a[232:235], v[96:99], v[12:15], a[232:235]// 000000003EC0: D3B580E8 07A21960
-	buffer_load_dwordx4 v0, s[52:55], s76 offen lds            // 000000003EC8: E05D1000 4C0D0000
-	v_mfma_f32_16x16x32_bf16 a[236:239], v[96:99], v[16:19], a[236:239]// 000000003ED0: D3B580EC 07B22160
-	s_mov_b32 m0, s47                                          // 000000003ED8: BEFC002F
-	v_mfma_f32_16x16x32_bf16 a[240:243], v[96:99], v[20:23], a[240:243]// 000000003EDC: D3B580F0 07C22960
-	v_mfma_f32_16x16x32_bf16 a[244:247], v[96:99], v[24:27], a[244:247]// 000000003EE4: D3B580F4 07D23160
-	buffer_load_dwordx4 v1, s[56:59], 0 offen lds              // 000000003EEC: E05D1000 800E0001
-	v_mfma_f32_16x16x32_bf16 a[248:251], v[96:99], v[28:31], a[248:251]// 000000003EF4: D3B580F8 07E23960
-	s_add_u32 m0, m0, 0x1040                                   // 000000003EFC: 807CFF7C 00001040
-	v_mfma_f32_16x16x32_bf16 a[252:255], v[96:99], v[32:35], a[252:255]// 000000003F04: D3B580FC 07F24160
-	v_mfma_f32_16x16x32_bf16 a[0:3], v[100:103], v[36:39], a[0:3]// 000000003F0C: D3B58000 04024964
-	buffer_load_dwordx4 v1, s[56:59], s77 offen lds            // 000000003F14: E05D1000 4D0E0001
-	v_mfma_f32_16x16x32_bf16 a[4:7], v[100:103], v[40:43], a[4:7]// 000000003F1C: D3B58004 04125164
-	s_add_u32 m0, m0, 0x1040                                   // 000000003F24: 807CFF7C 00001040
-	s_xor_b32 s46, s48, s46                                    // 000000003F2C: 882E2E30
-	v_mfma_f32_16x16x32_bf16 a[8:11], v[100:103], v[44:47], a[8:11]// 000000003F30: D3B58008 04225964
-	v_mfma_f32_16x16x32_bf16 a[12:15], v[100:103], v[48:51], a[12:15]// 000000003F38: D3B5800C 04326164
-	v_mfma_f32_16x16x32_bf16 a[16:19], v[100:103], v[52:55], a[16:19]// 000000003F40: D3B58010 04426964
-	v_mfma_f32_16x16x32_bf16 a[20:23], v[100:103], v[56:59], a[20:23]// 000000003F48: D3B58014 04527164
-	v_mfma_f32_16x16x32_bf16 a[24:27], v[100:103], v[60:63], a[24:27]// 000000003F50: D3B58018 04627964
-	v_mfma_f32_16x16x32_bf16 a[28:31], v[100:103], v[64:67], a[28:31]// 000000003F58: D3B5801C 04728164
-	v_mfma_f32_16x16x32_bf16 a[32:35], v[104:107], v[36:39], a[32:35]// 000000003F60: D3B58020 04824968
-	v_mfma_f32_16x16x32_bf16 a[36:39], v[104:107], v[40:43], a[36:39]// 000000003F68: D3B58024 04925168
-	v_mfma_f32_16x16x32_bf16 a[40:43], v[104:107], v[44:47], a[40:43]// 000000003F70: D3B58028 04A25968
-	v_mfma_f32_16x16x32_bf16 a[44:47], v[104:107], v[48:51], a[44:47]// 000000003F78: D3B5802C 04B26168
-	v_mfma_f32_16x16x32_bf16 a[48:51], v[104:107], v[52:55], a[48:51]// 000000003F80: D3B58030 04C26968
-	v_mfma_f32_16x16x32_bf16 a[52:55], v[104:107], v[56:59], a[52:55]// 000000003F88: D3B58034 04D27168
-	v_mfma_f32_16x16x32_bf16 a[56:59], v[104:107], v[60:63], a[56:59]// 000000003F90: D3B58038 04E27968
-	v_mfma_f32_16x16x32_bf16 a[60:63], v[104:107], v[64:67], a[60:63]// 000000003F98: D3B5803C 04F28168
-	v_mfma_f32_16x16x32_bf16 a[64:67], v[108:111], v[36:39], a[64:67]// 000000003FA0: D3B58040 0502496C
-	v_mfma_f32_16x16x32_bf16 a[68:71], v[108:111], v[40:43], a[68:71]// 000000003FA8: D3B58044 0512516C
-	v_mfma_f32_16x16x32_bf16 a[72:75], v[108:111], v[44:47], a[72:75]// 000000003FB0: D3B58048 0522596C
-	v_mfma_f32_16x16x32_bf16 a[76:79], v[108:111], v[48:51], a[76:79]// 000000003FB8: D3B5804C 0532616C
-	v_mfma_f32_16x16x32_bf16 a[80:83], v[108:111], v[52:55], a[80:83]// 000000003FC0: D3B58050 0542696C
-	v_xor_b32_e32 v2, v132, v2                                 // 000000003FC8: 2A040584
-	v_xor_b32_e32 v3, v133, v3                                 // 000000003FCC: 2A060785
-	v_mfma_f32_16x16x32_bf16 a[84:87], v[108:111], v[56:59], a[84:87]// 000000003FD0: D3B58054 0552716C
-	buffer_load_dwordx4 v1, s[56:59], s78 offen lds            // 000000003FD8: E05D1000 4E0E0001
-	v_mfma_f32_16x16x32_bf16 a[88:91], v[108:111], v[60:63], a[88:91]// 000000003FE0: D3B58058 0562796C
-	s_add_u32 m0, m0, 0x1040                                   // 000000003FE8: 807CFF7C 00001040
-	v_mfma_f32_16x16x32_bf16 a[92:95], v[108:111], v[64:67], a[92:95]// 000000003FF0: D3B5805C 0572816C
-	buffer_load_dwordx4 v1, s[56:59], s79 offen lds            // 000000003FF8: E05D1000 4F0E0001
-	v_mfma_f32_16x16x32_bf16 a[96:99], v[112:115], v[36:39], a[96:99]// 000000004000: D3B58060 05824970
-	s_add_u32 m0, m0, 0x1040                                   // 000000004008: 807CFF7C 00001040
-	v_mfma_f32_16x16x32_bf16 a[100:103], v[112:115], v[40:43], a[100:103]// 000000004010: D3B58064 05925170
-	buffer_load_dwordx4 v1, s[56:59], s80 offen lds            // 000000004018: E05D1000 500E0001
-	v_mfma_f32_16x16x32_bf16 a[104:107], v[112:115], v[44:47], a[104:107]// 000000004020: D3B58068 05A25970
-	v_mfma_f32_16x16x32_bf16 a[108:111], v[112:115], v[48:51], a[108:111]// 000000004028: D3B5806C 05B26170
-	s_waitcnt vmcnt(13)                                        // 000000004030: BF8C0F7D
-	v_mfma_f32_16x16x32_bf16 a[112:115], v[112:115], v[52:55], a[112:115]// 000000004034: D3B58070 05C26970
-	s_barrier                                                  // 00000000403C: BF8A0000
-	v_mfma_f32_16x16x32_bf16 a[116:119], v[112:115], v[56:59], a[116:119]// 000000004040: D3B58074 05D27170
-	ds_read_b128 v[4:7], v2                                    // 000000004048: D9FE0000 04000002
-	v_mfma_f32_16x16x32_bf16 a[120:123], v[112:115], v[60:63], a[120:123]// 000000004050: D3B58078 05E27970
-	ds_read_b128 v[8:11], v2 offset:128                        // 000000004058: D9FE0080 08000002
-	s_add_u32 m0, m0, 0x1040                                   // 000000004060: 807CFF7C 00001040
-	v_mfma_f32_16x16x32_bf16 a[124:127], v[112:115], v[64:67], a[124:127]// 000000004068: D3B5807C 05F28170
-	ds_read_b128 v[12:15], v2 offset:256                       // 000000004070: D9FE0100 0C000002
-	v_mfma_f32_16x16x32_bf16 a[128:131], v[116:119], v[36:39], a[128:131]// 000000004078: D3B58080 06024974
-	buffer_load_dwordx4 v1, s[56:59], s81 offen lds            // 000000004080: E05D1000 510E0001
-	v_mfma_f32_16x16x32_bf16 a[132:135], v[116:119], v[40:43], a[132:135]// 000000004088: D3B58084 06125174
-	ds_read_b128 v[16:19], v2 offset:384                       // 000000004090: D9FE0180 10000002
-	v_mfma_f32_16x16x32_bf16 a[136:139], v[116:119], v[44:47], a[136:139]// 000000004098: D3B58088 06225974
-	ds_read_b128 v[20:23], v2 offset:512                       // 0000000040A0: D9FE0200 14000002
-	s_add_u32 m0, m0, 0x1040                                   // 0000000040A8: 807CFF7C 00001040
-	v_mfma_f32_16x16x32_bf16 a[140:143], v[116:119], v[48:51], a[140:143]// 0000000040B0: D3B5808C 06326174
-	v_mfma_f32_16x16x32_bf16 a[144:147], v[116:119], v[52:55], a[144:147]// 0000000040B8: D3B58090 06426974
-	buffer_load_dwordx4 v1, s[56:59], s82 offen lds            // 0000000040C0: E05D1000 520E0001
-	v_mfma_f32_16x16x32_bf16 a[148:151], v[116:119], v[56:59], a[148:151]// 0000000040C8: D3B58094 06527174
-	v_mfma_f32_16x16x32_bf16 a[152:155], v[116:119], v[60:63], a[152:155]// 0000000040D0: D3B58098 06627974
-	ds_read_b128 v[24:27], v2 offset:640                       // 0000000040D8: D9FE0280 18000002
-	s_add_u32 m0, m0, 0x1040                                   // 0000000040E0: 807CFF7C 00001040
-	v_mfma_f32_16x16x32_bf16 a[156:159], v[116:119], v[64:67], a[156:159]// 0000000040E8: D3B5809C 06728174
-	ds_read_b128 v[28:31], v2 offset:768                       // 0000000040F0: D9FE0300 1C000002
-	v_mfma_f32_16x16x32_bf16 a[160:163], v[120:123], v[36:39], a[160:163]// 0000000040F8: D3B580A0 06824978
-	ds_read_b128 v[32:35], v2 offset:896                       // 000000004100: D9FE0380 20000002
-	v_mfma_f32_16x16x32_bf16 a[164:167], v[120:123], v[40:43], a[164:167]// 000000004108: D3B580A4 06925178
-	ds_read_b128 v[68:71], v3                                  // 000000004110: D9FE0000 44000003
-	v_mfma_f32_16x16x32_bf16 a[168:171], v[120:123], v[44:47], a[168:171]// 000000004118: D3B580A8 06A25978
-	ds_read_b128 v[72:75], v3 offset:128                       // 000000004120: D9FE0080 48000003
-	v_mfma_f32_16x16x32_bf16 a[172:175], v[120:123], v[48:51], a[172:175]// 000000004128: D3B580AC 06B26178
-	v_mfma_f32_16x16x32_bf16 a[176:179], v[120:123], v[52:55], a[176:179]// 000000004130: D3B580B0 06C26978
-	v_mfma_f32_16x16x32_bf16 a[180:183], v[120:123], v[56:59], a[180:183]// 000000004138: D3B580B4 06D27178
-	ds_read_b128 v[76:79], v3 offset:256                       // 000000004140: D9FE0100 4C000003
-	v_mfma_f32_16x16x32_bf16 a[184:187], v[120:123], v[60:63], a[184:187]// 000000004148: D3B580B8 06E27978
-	v_mfma_f32_16x16x32_bf16 a[188:191], v[120:123], v[64:67], a[188:191]// 000000004150: D3B580BC 06F28178
-	v_mfma_f32_16x16x32_bf16 a[192:195], v[124:127], v[36:39], a[192:195]// 000000004158: D3B580C0 0702497C
-	ds_read_b128 v[80:83], v3 offset:384                       // 000000004160: D9FE0180 50000003
-	v_mfma_f32_16x16x32_bf16 a[196:199], v[124:127], v[40:43], a[196:199]// 000000004168: D3B580C4 0712517C
-	v_mfma_f32_16x16x32_bf16 a[200:203], v[124:127], v[44:47], a[200:203]// 000000004170: D3B580C8 0722597C
-	ds_read_b128 v[84:87], v3 offset:512                       // 000000004178: D9FE0200 54000003
-	v_mfma_f32_16x16x32_bf16 a[204:207], v[124:127], v[48:51], a[204:207]// 000000004180: D3B580CC 0732617C
-	v_mfma_f32_16x16x32_bf16 a[208:211], v[124:127], v[52:55], a[208:211]// 000000004188: D3B580D0 0742697C
-	v_mfma_f32_16x16x32_bf16 a[212:215], v[124:127], v[56:59], a[212:215]// 000000004190: D3B580D4 0752717C
-	ds_read_b128 v[88:91], v3 offset:640                       // 000000004198: D9FE0280 58000003
-	v_mfma_f32_16x16x32_bf16 a[216:219], v[124:127], v[60:63], a[216:219]// 0000000041A0: D3B580D8 0762797C
-	v_mfma_f32_16x16x32_bf16 a[220:223], v[124:127], v[64:67], a[220:223]// 0000000041A8: D3B580DC 0772817C
-	v_mfma_f32_16x16x32_bf16 a[224:227], v[128:131], v[36:39], a[224:227]// 0000000041B0: D3B580E0 07824980
-	ds_read_b128 v[92:95], v3 offset:768                       // 0000000041B8: D9FE0300 5C000003
-	v_mfma_f32_16x16x32_bf16 a[228:231], v[128:131], v[40:43], a[228:231]// 0000000041C0: D3B580E4 07925180
-	v_mfma_f32_16x16x32_bf16 a[232:235], v[128:131], v[44:47], a[232:235]// 0000000041C8: D3B580E8 07A25980
-	v_mfma_f32_16x16x32_bf16 a[236:239], v[128:131], v[48:51], a[236:239]// 0000000041D0: D3B580EC 07B26180
-	ds_read_b128 v[96:99], v3 offset:896                       // 0000000041D8: D9FE0380 60000003
-	v_mfma_f32_16x16x32_bf16 a[240:243], v[128:131], v[52:55], a[240:243]// 0000000041E0: D3B580F0 07C26980
-	buffer_load_dwordx4 v1, s[56:59], s83 offen lds            // 0000000041E8: E05D1000 530E0001
-	v_mfma_f32_16x16x32_bf16 a[244:247], v[128:131], v[56:59], a[244:247]// 0000000041F0: D3B580F4 07D27180
-	s_xor_b32 s47, s49, s47                                    // 0000000041F8: 882F2F31
-	s_sub_u32 s12, s12, 1                                      // 0000000041FC: 808C810C
-	v_mfma_f32_16x16x32_bf16 a[248:251], v[128:131], v[60:63], a[248:251]// 000000004200: D3B580F8 07E27980
-	s_cmp_eq_i32 s12, 2                                        // 000000004208: BF00820C
-	s_waitcnt lgkmcnt(0)                                       // 00000000420C: BF8CC07F
-	v_mfma_f32_16x16x32_bf16 a[252:255], v[128:131], v[64:67], a[252:255]// 000000004210: D3B580FC 07F28180
-	s_cbranch_scc0 main_loop                            // 000000004218: BF84FE62
-
-loop_epilogue:
-	v_mfma_f32_16x16x32_bf16 a[0:3], v[68:71], v[4:7], a[0:3]  // 00000000421C: D3B58000 04020944
-	ds_read_b128 v[36:39], v2 offset:64                        // 000000004224: D9FE0040 24000002
-	v_mfma_f32_16x16x32_bf16 a[4:7], v[68:71], v[8:11], a[4:7] // 00000000422C: D3B58004 04121144
-	v_mfma_f32_16x16x32_bf16 a[8:11], v[68:71], v[12:15], a[8:11]// 000000004234: D3B58008 04221944
-	ds_read_b128 v[100:103], v3 offset:64                      // 00000000423C: D9FE0040 64000003
-	v_mfma_f32_16x16x32_bf16 a[12:15], v[68:71], v[16:19], a[12:15]// 000000004244: D3B5800C 04322144
-	v_mfma_f32_16x16x32_bf16 a[16:19], v[68:71], v[20:23], a[16:19]// 00000000424C: D3B58010 04422944
-	ds_read_b128 v[40:43], v2 offset:192                       // 000000004254: D9FE00C0 28000002
-	v_mfma_f32_16x16x32_bf16 a[20:23], v[68:71], v[24:27], a[20:23]// 00000000425C: D3B58014 04523144
-	v_mfma_f32_16x16x32_bf16 a[24:27], v[68:71], v[28:31], a[24:27]// 000000004264: D3B58018 04623944
-	ds_read_b128 v[44:47], v2 offset:320                       // 00000000426C: D9FE0140 2C000002
-	v_mfma_f32_16x16x32_bf16 a[28:31], v[68:71], v[32:35], a[28:31]// 000000004274: D3B5801C 04724144
-	v_mfma_f32_16x16x32_bf16 a[32:35], v[72:75], v[4:7], a[32:35]// 00000000427C: D3B58020 04820948
-	ds_read_b128 v[48:51], v2 offset:448                       // 000000004284: D9FE01C0 30000002
-	v_mfma_f32_16x16x32_bf16 a[36:39], v[72:75], v[8:11], a[36:39]// 00000000428C: D3B58024 04921148
-	v_mfma_f32_16x16x32_bf16 a[40:43], v[72:75], v[12:15], a[40:43]// 000000004294: D3B58028 04A21948
-	ds_read_b128 v[52:55], v2 offset:576                       // 00000000429C: D9FE0240 34000002
-	v_mfma_f32_16x16x32_bf16 a[44:47], v[72:75], v[16:19], a[44:47]// 0000000042A4: D3B5802C 04B22148
-	v_mfma_f32_16x16x32_bf16 a[48:51], v[72:75], v[20:23], a[48:51]// 0000000042AC: D3B58030 04C22948
-	ds_read_b128 v[56:59], v2 offset:704                       // 0000000042B4: D9FE02C0 38000002
-	v_mfma_f32_16x16x32_bf16 a[52:55], v[72:75], v[24:27], a[52:55]// 0000000042BC: D3B58034 04D23148
-	v_mfma_f32_16x16x32_bf16 a[56:59], v[72:75], v[28:31], a[56:59]// 0000000042C4: D3B58038 04E23948
-	ds_read_b128 v[60:63], v2 offset:832                       // 0000000042CC: D9FE0340 3C000002
-	v_mfma_f32_16x16x32_bf16 a[60:63], v[72:75], v[32:35], a[60:63]// 0000000042D4: D3B5803C 04F24148
-	v_mfma_f32_16x16x32_bf16 a[64:67], v[76:79], v[4:7], a[64:67]// 0000000042DC: D3B58040 0502094C
-	ds_read_b128 v[64:67], v2 offset:960                       // 0000000042E4: D9FE03C0 40000002
-	v_mfma_f32_16x16x32_bf16 a[68:71], v[76:79], v[8:11], a[68:71]// 0000000042EC: D3B58044 0512114C
-	v_mfma_f32_16x16x32_bf16 a[72:75], v[76:79], v[12:15], a[72:75]// 0000000042F4: D3B58048 0522194C
-	ds_read_b128 v[104:107], v3 offset:192                     // 0000000042FC: D9FE00C0 68000003
-	v_mfma_f32_16x16x32_bf16 a[76:79], v[76:79], v[16:19], a[76:79]// 000000004304: D3B5804C 0532214C
-	v_mfma_f32_16x16x32_bf16 a[80:83], v[76:79], v[20:23], a[80:83]// 00000000430C: D3B58050 0542294C
-	ds_read_b128 v[108:111], v3 offset:320                     // 000000004314: D9FE0140 6C000003
-	v_mfma_f32_16x16x32_bf16 a[84:87], v[76:79], v[24:27], a[84:87]// 00000000431C: D3B58054 0552314C
-	v_mfma_f32_16x16x32_bf16 a[88:91], v[76:79], v[28:31], a[88:91]// 000000004324: D3B58058 0562394C
-	ds_read_b128 v[112:115], v3 offset:448                     // 00000000432C: D9FE01C0 70000003
-	v_mfma_f32_16x16x32_bf16 a[92:95], v[76:79], v[32:35], a[92:95]// 000000004334: D3B5805C 0572414C
-	v_mfma_f32_16x16x32_bf16 a[96:99], v[80:83], v[4:7], a[96:99]// 00000000433C: D3B58060 05820950
-	ds_read_b128 v[116:119], v3 offset:576                     // 000000004344: D9FE0240 74000003
-	v_mfma_f32_16x16x32_bf16 a[100:103], v[80:83], v[8:11], a[100:103]// 00000000434C: D3B58064 05921150
-	v_mfma_f32_16x16x32_bf16 a[104:107], v[80:83], v[12:15], a[104:107]// 000000004354: D3B58068 05A21950
-	ds_read_b128 v[120:123], v3 offset:704                     // 00000000435C: D9FE02C0 78000003
-	v_mfma_f32_16x16x32_bf16 a[108:111], v[80:83], v[16:19], a[108:111]// 000000004364: D3B5806C 05B22150
-	v_mfma_f32_16x16x32_bf16 a[112:115], v[80:83], v[20:23], a[112:115]// 00000000436C: D3B58070 05C22950
-	ds_read_b128 v[124:127], v3 offset:832                     // 000000004374: D9FE0340 7C000003
-	v_mfma_f32_16x16x32_bf16 a[116:119], v[80:83], v[24:27], a[116:119]// 00000000437C: D3B58074 05D23150
-	v_mfma_f32_16x16x32_bf16 a[120:123], v[80:83], v[28:31], a[120:123]// 000000004384: D3B58078 05E23950
-	ds_read_b128 v[128:131], v3 offset:960                     // 00000000438C: D9FE03C0 80000003
-	v_mfma_f32_16x16x32_bf16 a[124:127], v[80:83], v[32:35], a[124:127]// 000000004394: D3B5807C 05F24150
-	v_mfma_f32_16x16x32_bf16 a[128:131], v[84:87], v[4:7], a[128:131]// 00000000439C: D3B58080 06020954
-	v_mfma_f32_16x16x32_bf16 a[132:135], v[84:87], v[8:11], a[132:135]// 0000000043A4: D3B58084 06121154
-	v_mfma_f32_16x16x32_bf16 a[136:139], v[84:87], v[12:15], a[136:139]// 0000000043AC: D3B58088 06221954
-	v_mfma_f32_16x16x32_bf16 a[140:143], v[84:87], v[16:19], a[140:143]// 0000000043B4: D3B5808C 06322154
-	v_mfma_f32_16x16x32_bf16 a[144:147], v[84:87], v[20:23], a[144:147]// 0000000043BC: D3B58090 06422954
-	v_mfma_f32_16x16x32_bf16 a[148:151], v[84:87], v[24:27], a[148:151]// 0000000043C4: D3B58094 06523154
-	v_mfma_f32_16x16x32_bf16 a[152:155], v[84:87], v[28:31], a[152:155]// 0000000043CC: D3B58098 06623954
-	v_mfma_f32_16x16x32_bf16 a[156:159], v[84:87], v[32:35], a[156:159]// 0000000043D4: D3B5809C 06724154
-	v_mfma_f32_16x16x32_bf16 a[160:163], v[88:91], v[4:7], a[160:163]// 0000000043DC: D3B580A0 06820958
-	v_mfma_f32_16x16x32_bf16 a[164:167], v[88:91], v[8:11], a[164:167]// 0000000043E4: D3B580A4 06921158
-	v_mfma_f32_16x16x32_bf16 a[168:171], v[88:91], v[12:15], a[168:171]// 0000000043EC: D3B580A8 06A21958
-	v_mfma_f32_16x16x32_bf16 a[172:175], v[88:91], v[16:19], a[172:175]// 0000000043F4: D3B580AC 06B22158
-	v_mfma_f32_16x16x32_bf16 a[176:179], v[88:91], v[20:23], a[176:179]// 0000000043FC: D3B580B0 06C22958
-	v_mfma_f32_16x16x32_bf16 a[180:183], v[88:91], v[24:27], a[180:183]// 000000004404: D3B580B4 06D23158
-	v_mfma_f32_16x16x32_bf16 a[184:187], v[88:91], v[28:31], a[184:187]// 00000000440C: D3B580B8 06E23958
-	v_mfma_f32_16x16x32_bf16 a[188:191], v[88:91], v[32:35], a[188:191]// 000000004414: D3B580BC 06F24158
-	v_mfma_f32_16x16x32_bf16 a[192:195], v[92:95], v[4:7], a[192:195]// 00000000441C: D3B580C0 0702095C
-	v_mfma_f32_16x16x32_bf16 a[196:199], v[92:95], v[8:11], a[196:199]// 000000004424: D3B580C4 0712115C
-	v_mfma_f32_16x16x32_bf16 a[200:203], v[92:95], v[12:15], a[200:203]// 00000000442C: D3B580C8 0722195C
-	v_mfma_f32_16x16x32_bf16 a[204:207], v[92:95], v[16:19], a[204:207]// 000000004434: D3B580CC 0732215C
-	v_mfma_f32_16x16x32_bf16 a[208:211], v[92:95], v[20:23], a[208:211]// 00000000443C: D3B580D0 0742295C
-	v_mfma_f32_16x16x32_bf16 a[212:215], v[92:95], v[24:27], a[212:215]// 000000004444: D3B580D4 0752315C
-	v_mfma_f32_16x16x32_bf16 a[216:219], v[92:95], v[28:31], a[216:219]// 00000000444C: D3B580D8 0762395C
-	v_mfma_f32_16x16x32_bf16 a[220:223], v[92:95], v[32:35], a[220:223]// 000000004454: D3B580DC 0772415C
-	v_mfma_f32_16x16x32_bf16 a[224:227], v[96:99], v[4:7], a[224:227]// 00000000445C: D3B580E0 07820960
-	v_mfma_f32_16x16x32_bf16 a[228:231], v[96:99], v[8:11], a[228:231]// 000000004464: D3B580E4 07921160
-	v_mfma_f32_16x16x32_bf16 a[232:235], v[96:99], v[12:15], a[232:235]// 00000000446C: D3B580E8 07A21960
-	v_mfma_f32_16x16x32_bf16 a[236:239], v[96:99], v[16:19], a[236:239]// 000000004474: D3B580EC 07B22160
-	v_mfma_f32_16x16x32_bf16 a[240:243], v[96:99], v[20:23], a[240:243]// 00000000447C: D3B580F0 07C22960
-	v_mfma_f32_16x16x32_bf16 a[244:247], v[96:99], v[24:27], a[244:247]// 000000004484: D3B580F4 07D23160
-	v_mfma_f32_16x16x32_bf16 a[248:251], v[96:99], v[28:31], a[248:251]// 00000000448C: D3B580F8 07E23960
-	v_xor_b32_e32 v2, v132, v2                                 // 000000004494: 2A040584
-	v_xor_b32_e32 v3, v133, v3                                 // 000000004498: 2A060785
-	v_mfma_f32_16x16x32_bf16 a[252:255], v[96:99], v[32:35], a[252:255]// 00000000449C: D3B580FC 07F24160
-	s_waitcnt lgkmcnt(0)                                       // 0000000044A4: BF8CC07F
-	v_mfma_f32_16x16x32_bf16 a[0:3], v[100:103], v[36:39], a[0:3]// 0000000044A8: D3B58000 04024964
-	v_mfma_f32_16x16x32_bf16 a[4:7], v[100:103], v[40:43], a[4:7]// 0000000044B0: D3B58004 04125164
-	v_mfma_f32_16x16x32_bf16 a[8:11], v[100:103], v[44:47], a[8:11]// 0000000044B8: D3B58008 04225964
-	v_mfma_f32_16x16x32_bf16 a[12:15], v[100:103], v[48:51], a[12:15]// 0000000044C0: D3B5800C 04326164
-	v_mfma_f32_16x16x32_bf16 a[16:19], v[100:103], v[52:55], a[16:19]// 0000000044C8: D3B58010 04426964
-	v_mfma_f32_16x16x32_bf16 a[20:23], v[100:103], v[56:59], a[20:23]// 0000000044D0: D3B58014 04527164
-	v_mfma_f32_16x16x32_bf16 a[24:27], v[100:103], v[60:63], a[24:27]// 0000000044D8: D3B58018 04627964
-	v_mfma_f32_16x16x32_bf16 a[28:31], v[100:103], v[64:67], a[28:31]// 0000000044E0: D3B5801C 04728164
-	v_mfma_f32_16x16x32_bf16 a[32:35], v[104:107], v[36:39], a[32:35]// 0000000044E8: D3B58020 04824968
-	v_mfma_f32_16x16x32_bf16 a[36:39], v[104:107], v[40:43], a[36:39]// 0000000044F0: D3B58024 04925168
-	v_mfma_f32_16x16x32_bf16 a[40:43], v[104:107], v[44:47], a[40:43]// 0000000044F8: D3B58028 04A25968
-	v_mfma_f32_16x16x32_bf16 a[44:47], v[104:107], v[48:51], a[44:47]// 000000004500: D3B5802C 04B26168
-	v_mfma_f32_16x16x32_bf16 a[48:51], v[104:107], v[52:55], a[48:51]// 000000004508: D3B58030 04C26968
-	v_mfma_f32_16x16x32_bf16 a[52:55], v[104:107], v[56:59], a[52:55]// 000000004510: D3B58034 04D27168
-	v_mfma_f32_16x16x32_bf16 a[56:59], v[104:107], v[60:63], a[56:59]// 000000004518: D3B58038 04E27968
-	v_mfma_f32_16x16x32_bf16 a[60:63], v[104:107], v[64:67], a[60:63]// 000000004520: D3B5803C 04F28168
-	v_mfma_f32_16x16x32_bf16 a[64:67], v[108:111], v[36:39], a[64:67]// 000000004528: D3B58040 0502496C
-	v_mfma_f32_16x16x32_bf16 a[68:71], v[108:111], v[40:43], a[68:71]// 000000004530: D3B58044 0512516C
-	v_mfma_f32_16x16x32_bf16 a[72:75], v[108:111], v[44:47], a[72:75]// 000000004538: D3B58048 0522596C
-	v_mfma_f32_16x16x32_bf16 a[76:79], v[108:111], v[48:51], a[76:79]// 000000004540: D3B5804C 0532616C
-	v_mfma_f32_16x16x32_bf16 a[80:83], v[108:111], v[52:55], a[80:83]// 000000004548: D3B58050 0542696C
-	v_mfma_f32_16x16x32_bf16 a[84:87], v[108:111], v[56:59], a[84:87]// 000000004550: D3B58054 0552716C
-	v_mfma_f32_16x16x32_bf16 a[88:91], v[108:111], v[60:63], a[88:91]// 000000004558: D3B58058 0562796C
-	v_mfma_f32_16x16x32_bf16 a[92:95], v[108:111], v[64:67], a[92:95]// 000000004560: D3B5805C 0572816C
-	v_mfma_f32_16x16x32_bf16 a[96:99], v[112:115], v[36:39], a[96:99]// 000000004568: D3B58060 05824970
-	v_mfma_f32_16x16x32_bf16 a[100:103], v[112:115], v[40:43], a[100:103]// 000000004570: D3B58064 05925170
-	v_mfma_f32_16x16x32_bf16 a[104:107], v[112:115], v[44:47], a[104:107]// 000000004578: D3B58068 05A25970
-	v_mfma_f32_16x16x32_bf16 a[108:111], v[112:115], v[48:51], a[108:111]// 000000004580: D3B5806C 05B26170
-	v_mfma_f32_16x16x32_bf16 a[112:115], v[112:115], v[52:55], a[112:115]// 000000004588: D3B58070 05C26970
-	v_mfma_f32_16x16x32_bf16 a[116:119], v[112:115], v[56:59], a[116:119]// 000000004590: D3B58074 05D27170
-	v_mfma_f32_16x16x32_bf16 a[120:123], v[112:115], v[60:63], a[120:123]// 000000004598: D3B58078 05E27970
-	v_mfma_f32_16x16x32_bf16 a[124:127], v[112:115], v[64:67], a[124:127]// 0000000045A0: D3B5807C 05F28170
-	v_mfma_f32_16x16x32_bf16 a[128:131], v[116:119], v[36:39], a[128:131]// 0000000045A8: D3B58080 06024974
-	v_mfma_f32_16x16x32_bf16 a[132:135], v[116:119], v[40:43], a[132:135]// 0000000045B0: D3B58084 06125174
-	v_mfma_f32_16x16x32_bf16 a[136:139], v[116:119], v[44:47], a[136:139]// 0000000045B8: D3B58088 06225974
-	v_mfma_f32_16x16x32_bf16 a[140:143], v[116:119], v[48:51], a[140:143]// 0000000045C0: D3B5808C 06326174
-	v_mfma_f32_16x16x32_bf16 a[144:147], v[116:119], v[52:55], a[144:147]// 0000000045C8: D3B58090 06426974
-	v_mfma_f32_16x16x32_bf16 a[148:151], v[116:119], v[56:59], a[148:151]// 0000000045D0: D3B58094 06527174
-	v_mfma_f32_16x16x32_bf16 a[152:155], v[116:119], v[60:63], a[152:155]// 0000000045D8: D3B58098 06627974
-	v_mfma_f32_16x16x32_bf16 a[156:159], v[116:119], v[64:67], a[156:159]// 0000000045E0: D3B5809C 06728174
-	v_mfma_f32_16x16x32_bf16 a[160:163], v[120:123], v[36:39], a[160:163]// 0000000045E8: D3B580A0 06824978
-	v_mfma_f32_16x16x32_bf16 a[164:167], v[120:123], v[40:43], a[164:167]// 0000000045F0: D3B580A4 06925178
-	s_waitcnt vmcnt(0)                                         // 0000000045F8: BF8C0F70
-	v_mfma_f32_16x16x32_bf16 a[168:171], v[120:123], v[44:47], a[168:171]// 0000000045FC: D3B580A8 06A25978
-	s_barrier                                                  // 000000004604: BF8A0000
-	v_mfma_f32_16x16x32_bf16 a[172:175], v[120:123], v[48:51], a[172:175]// 000000004608: D3B580AC 06B26178
-	ds_read_b128 v[4:7], v2                                    // 000000004610: D9FE0000 04000002
-	v_mfma_f32_16x16x32_bf16 a[176:179], v[120:123], v[52:55], a[176:179]// 000000004618: D3B580B0 06C26978
-	ds_read_b128 v[68:71], v3                                  // 000000004620: D9FE0000 44000003
-	v_mfma_f32_16x16x32_bf16 a[180:183], v[120:123], v[56:59], a[180:183]// 000000004628: D3B580B4 06D27178
-	ds_read_b128 v[8:11], v2 offset:128                        // 000000004630: D9FE0080 08000002
-	v_mfma_f32_16x16x32_bf16 a[184:187], v[120:123], v[60:63], a[184:187]// 000000004638: D3B580B8 06E27978
-	ds_read_b128 v[12:15], v2 offset:256                       // 000000004640: D9FE0100 0C000002
-	v_mfma_f32_16x16x32_bf16 a[188:191], v[120:123], v[64:67], a[188:191]// 000000004648: D3B580BC 06F28178
-	ds_read_b128 v[16:19], v2 offset:384                       // 000000004650: D9FE0180 10000002
-	v_mfma_f32_16x16x32_bf16 a[192:195], v[124:127], v[36:39], a[192:195]// 000000004658: D3B580C0 0702497C
-	ds_read_b128 v[20:23], v2 offset:512                       // 000000004660: D9FE0200 14000002
-	v_mfma_f32_16x16x32_bf16 a[196:199], v[124:127], v[40:43], a[196:199]// 000000004668: D3B580C4 0712517C
-	ds_read_b128 v[24:27], v2 offset:640                       // 000000004670: D9FE0280 18000002
-	v_mfma_f32_16x16x32_bf16 a[200:203], v[124:127], v[44:47], a[200:203]// 000000004678: D3B580C8 0722597C
-	ds_read_b128 v[28:31], v2 offset:768                       // 000000004680: D9FE0300 1C000002
-	v_mfma_f32_16x16x32_bf16 a[204:207], v[124:127], v[48:51], a[204:207]// 000000004688: D3B580CC 0732617C
-	ds_read_b128 v[32:35], v2 offset:896                       // 000000004690: D9FE0380 20000002
-	v_mfma_f32_16x16x32_bf16 a[208:211], v[124:127], v[52:55], a[208:211]// 000000004698: D3B580D0 0742697C
-	ds_read_b128 v[72:75], v3 offset:128                       // 0000000046A0: D9FE0080 48000003
-	v_mfma_f32_16x16x32_bf16 a[212:215], v[124:127], v[56:59], a[212:215]// 0000000046A8: D3B580D4 0752717C
-	ds_read_b128 v[76:79], v3 offset:256                       // 0000000046B0: D9FE0100 4C000003
-	v_mfma_f32_16x16x32_bf16 a[216:219], v[124:127], v[60:63], a[216:219]// 0000000046B8: D3B580D8 0762797C
-	ds_read_b128 v[80:83], v3 offset:384                       // 0000000046C0: D9FE0180 50000003
-	v_mfma_f32_16x16x32_bf16 a[220:223], v[124:127], v[64:67], a[220:223]// 0000000046C8: D3B580DC 0772817C
-	ds_read_b128 v[84:87], v3 offset:512                       // 0000000046D0: D9FE0200 54000003
-	v_mfma_f32_16x16x32_bf16 a[224:227], v[128:131], v[36:39], a[224:227]// 0000000046D8: D3B580E0 07824980
-	ds_read_b128 v[88:91], v3 offset:640                       // 0000000046E0: D9FE0280 58000003
-	v_mfma_f32_16x16x32_bf16 a[228:231], v[128:131], v[40:43], a[228:231]// 0000000046E8: D3B580E4 07925180
-	ds_read_b128 v[92:95], v3 offset:768                       // 0000000046F0: D9FE0300 5C000003
-	v_mfma_f32_16x16x32_bf16 a[232:235], v[128:131], v[44:47], a[232:235]// 0000000046F8: D3B580E8 07A25980
-	ds_read_b128 v[96:99], v3 offset:896                       // 000000004700: D9FE0380 60000003
-	v_mfma_f32_16x16x32_bf16 a[236:239], v[128:131], v[48:51], a[236:239]// 000000004708: D3B580EC 07B26180
-	v_mfma_f32_16x16x32_bf16 a[240:243], v[128:131], v[52:55], a[240:243]// 000000004710: D3B580F0 07C26980
-	v_mfma_f32_16x16x32_bf16 a[244:247], v[128:131], v[56:59], a[244:247]// 000000004718: D3B580F4 07D27180
-	v_mfma_f32_16x16x32_bf16 a[248:251], v[128:131], v[60:63], a[248:251]// 000000004720: D3B580F8 07E27980
-	v_mfma_f32_16x16x32_bf16 a[252:255], v[128:131], v[64:67], a[252:255]// 000000004728: D3B580FC 07F28180
-
-final_compute:
-	s_and_b32 s8, s50, 0x3fff                                  // 000000004730: 8608FF32 00003FFF
-	s_and_b32 s84, 0xff, s24                                   // 000000004750: 865418FF 000000FF
-	s_add_u32 s85, -1, s14                                     // 000000004758: 80550EC1
-	s_cmp_ge_u32 s2, s85                                       // 00000000475C: BF095502
-	s_cselect_b32 s84, s84, 0                                  // 000000004760: 85548054
-	s_and_b32 s84, 0xff, s25                                   // 00000000476C: 865419FF 000000FF
-	s_add_u32 s85, -1, s15                                     // 000000004774: 80550FC1
-	s_cmp_ge_u32 s3, s85                                       // 000000004778: BF095503
-	s_cselect_b32 s84, s84, 0                                  // 00000000477C: 85548054
-	v_mfma_f32_16x16x32_bf16 a[0:3], v[68:71], v[4:7], a[0:3]  // 000000004788: D3B58000 04020944
-	ds_read_b128 v[36:39], v2 offset:64                        // 000000004790: D9FE0040 24000002
-	v_mfma_f32_16x16x32_bf16 a[4:7], v[68:71], v[8:11], a[4:7] // 000000004798: D3B58004 04121144
-	v_mfma_f32_16x16x32_bf16 a[8:11], v[68:71], v[12:15], a[8:11]// 0000000047A0: D3B58008 04221944
-	ds_read_b128 v[100:103], v3 offset:64                      // 0000000047A8: D9FE0040 64000003
-	v_mfma_f32_16x16x32_bf16 a[12:15], v[68:71], v[16:19], a[12:15]// 0000000047B0: D3B5800C 04322144
-	v_mfma_f32_16x16x32_bf16 a[16:19], v[68:71], v[20:23], a[16:19]// 0000000047B8: D3B58010 04422944
-	ds_read_b128 v[40:43], v2 offset:192                       // 0000000047C0: D9FE00C0 28000002
-	v_mfma_f32_16x16x32_bf16 a[20:23], v[68:71], v[24:27], a[20:23]// 0000000047C8: D3B58014 04523144
-	v_mfma_f32_16x16x32_bf16 a[24:27], v[68:71], v[28:31], a[24:27]// 0000000047D0: D3B58018 04623944
-	ds_read_b128 v[44:47], v2 offset:320                       // 0000000047D8: D9FE0140 2C000002
-	v_mfma_f32_16x16x32_bf16 a[28:31], v[68:71], v[32:35], a[28:31]// 0000000047E0: D3B5801C 04724144
-	v_mfma_f32_16x16x32_bf16 a[32:35], v[72:75], v[4:7], a[32:35]// 0000000047E8: D3B58020 04820948
-	ds_read_b128 v[48:51], v2 offset:448                       // 0000000047F0: D9FE01C0 30000002
-	v_mfma_f32_16x16x32_bf16 a[36:39], v[72:75], v[8:11], a[36:39]// 0000000047F8: D3B58024 04921148
-	v_mfma_f32_16x16x32_bf16 a[40:43], v[72:75], v[12:15], a[40:43]// 000000004800: D3B58028 04A21948
-	ds_read_b128 v[52:55], v2 offset:576                       // 000000004808: D9FE0240 34000002
-	v_mfma_f32_16x16x32_bf16 a[44:47], v[72:75], v[16:19], a[44:47]// 000000004810: D3B5802C 04B22148
-	v_mfma_f32_16x16x32_bf16 a[48:51], v[72:75], v[20:23], a[48:51]// 000000004818: D3B58030 04C22948
-	ds_read_b128 v[56:59], v2 offset:704                       // 000000004820: D9FE02C0 38000002
-	v_mfma_f32_16x16x32_bf16 a[52:55], v[72:75], v[24:27], a[52:55]// 000000004828: D3B58034 04D23148
-	v_mfma_f32_16x16x32_bf16 a[56:59], v[72:75], v[28:31], a[56:59]// 000000004830: D3B58038 04E23948
-	ds_read_b128 v[60:63], v2 offset:832                       // 000000004838: D9FE0340 3C000002
-	v_mfma_f32_16x16x32_bf16 a[60:63], v[72:75], v[32:35], a[60:63]// 000000004840: D3B5803C 04F24148
-	v_mfma_f32_16x16x32_bf16 a[64:67], v[76:79], v[4:7], a[64:67]// 000000004848: D3B58040 0502094C
-	ds_read_b128 v[64:67], v2 offset:960                       // 000000004850: D9FE03C0 40000002
-	v_mfma_f32_16x16x32_bf16 a[68:71], v[76:79], v[8:11], a[68:71]// 000000004858: D3B58044 0512114C
-	v_mfma_f32_16x16x32_bf16 a[72:75], v[76:79], v[12:15], a[72:75]// 000000004860: D3B58048 0522194C
-	ds_read_b128 v[104:107], v3 offset:192                     // 000000004868: D9FE00C0 68000003
-	v_mfma_f32_16x16x32_bf16 a[76:79], v[76:79], v[16:19], a[76:79]// 000000004870: D3B5804C 0532214C
-	v_mfma_f32_16x16x32_bf16 a[80:83], v[76:79], v[20:23], a[80:83]// 000000004878: D3B58050 0542294C
-	ds_read_b128 v[108:111], v3 offset:320                     // 000000004880: D9FE0140 6C000003
-	v_mfma_f32_16x16x32_bf16 a[84:87], v[76:79], v[24:27], a[84:87]// 000000004888: D3B58054 0552314C
-	v_mfma_f32_16x16x32_bf16 a[88:91], v[76:79], v[28:31], a[88:91]// 000000004890: D3B58058 0562394C
-	ds_read_b128 v[112:115], v3 offset:448                     // 000000004898: D9FE01C0 70000003
-	v_mfma_f32_16x16x32_bf16 a[92:95], v[76:79], v[32:35], a[92:95]// 0000000048A0: D3B5805C 0572414C
-	v_mfma_f32_16x16x32_bf16 a[96:99], v[80:83], v[4:7], a[96:99]// 0000000048A8: D3B58060 05820950
-	ds_read_b128 v[116:119], v3 offset:576                     // 0000000048B0: D9FE0240 74000003
-	v_mfma_f32_16x16x32_bf16 a[100:103], v[80:83], v[8:11], a[100:103]// 0000000048B8: D3B58064 05921150
-	v_mfma_f32_16x16x32_bf16 a[104:107], v[80:83], v[12:15], a[104:107]// 0000000048C0: D3B58068 05A21950
-	ds_read_b128 v[120:123], v3 offset:704                     // 0000000048C8: D9FE02C0 78000003
-	v_mfma_f32_16x16x32_bf16 a[108:111], v[80:83], v[16:19], a[108:111]// 0000000048D0: D3B5806C 05B22150
-	v_mfma_f32_16x16x32_bf16 a[112:115], v[80:83], v[20:23], a[112:115]// 0000000048D8: D3B58070 05C22950
-	ds_read_b128 v[124:127], v3 offset:832                     // 0000000048E0: D9FE0340 7C000003
-	v_mfma_f32_16x16x32_bf16 a[116:119], v[80:83], v[24:27], a[116:119]// 0000000048E8: D3B58074 05D23150
-	v_mfma_f32_16x16x32_bf16 a[120:123], v[80:83], v[28:31], a[120:123]// 0000000048F0: D3B58078 05E23950
-	ds_read_b128 v[128:131], v3 offset:960                     // 0000000048F8: D9FE03C0 80000003
-	v_mfma_f32_16x16x32_bf16 a[124:127], v[80:83], v[32:35], a[124:127]// 000000004900: D3B5807C 05F24150
-	v_mfma_f32_16x16x32_bf16 a[128:131], v[84:87], v[4:7], a[128:131]// 000000004908: D3B58080 06020954
-	v_mfma_f32_16x16x32_bf16 a[132:135], v[84:87], v[8:11], a[132:135]// 000000004910: D3B58084 06121154
-	v_mfma_f32_16x16x32_bf16 a[136:139], v[84:87], v[12:15], a[136:139]// 000000004918: D3B58088 06221954
-	v_mfma_f32_16x16x32_bf16 a[140:143], v[84:87], v[16:19], a[140:143]// 000000004920: D3B5808C 06322154
-	v_mfma_f32_16x16x32_bf16 a[144:147], v[84:87], v[20:23], a[144:147]// 000000004928: D3B58090 06422954
-	v_mfma_f32_16x16x32_bf16 a[148:151], v[84:87], v[24:27], a[148:151]// 000000004930: D3B58094 06523154
-	v_mfma_f32_16x16x32_bf16 a[152:155], v[84:87], v[28:31], a[152:155]// 000000004938: D3B58098 06623954
-	v_mfma_f32_16x16x32_bf16 a[156:159], v[84:87], v[32:35], a[156:159]// 000000004940: D3B5809C 06724154
-	v_mfma_f32_16x16x32_bf16 a[160:163], v[88:91], v[4:7], a[160:163]// 000000004948: D3B580A0 06820958
-	v_mfma_f32_16x16x32_bf16 a[164:167], v[88:91], v[8:11], a[164:167]// 000000004950: D3B580A4 06921158
-	v_mfma_f32_16x16x32_bf16 a[168:171], v[88:91], v[12:15], a[168:171]// 000000004958: D3B580A8 06A21958
-	v_mfma_f32_16x16x32_bf16 a[172:175], v[88:91], v[16:19], a[172:175]// 000000004960: D3B580AC 06B22158
-	v_mfma_f32_16x16x32_bf16 a[176:179], v[88:91], v[20:23], a[176:179]// 000000004968: D3B580B0 06C22958
-	v_mfma_f32_16x16x32_bf16 a[180:183], v[88:91], v[24:27], a[180:183]// 000000004970: D3B580B4 06D23158
-	v_mfma_f32_16x16x32_bf16 a[184:187], v[88:91], v[28:31], a[184:187]// 000000004978: D3B580B8 06E23958
-	v_mfma_f32_16x16x32_bf16 a[188:191], v[88:91], v[32:35], a[188:191]// 000000004980: D3B580BC 06F24158
-	v_mfma_f32_16x16x32_bf16 a[192:195], v[92:95], v[4:7], a[192:195]// 000000004988: D3B580C0 0702095C
-	v_mfma_f32_16x16x32_bf16 a[196:199], v[92:95], v[8:11], a[196:199]// 000000004990: D3B580C4 0712115C
-	v_mfma_f32_16x16x32_bf16 a[200:203], v[92:95], v[12:15], a[200:203]// 000000004998: D3B580C8 0722195C
-	v_mfma_f32_16x16x32_bf16 a[204:207], v[92:95], v[16:19], a[204:207]// 0000000049A0: D3B580CC 0732215C
-	v_mfma_f32_16x16x32_bf16 a[208:211], v[92:95], v[20:23], a[208:211]// 0000000049A8: D3B580D0 0742295C
-	v_mfma_f32_16x16x32_bf16 a[212:215], v[92:95], v[24:27], a[212:215]// 0000000049B0: D3B580D4 0752315C
-	v_mfma_f32_16x16x32_bf16 a[216:219], v[92:95], v[28:31], a[216:219]// 0000000049B8: D3B580D8 0762395C
-	v_mfma_f32_16x16x32_bf16 a[220:223], v[92:95], v[32:35], a[220:223]// 0000000049C0: D3B580DC 0772415C
-	v_mfma_f32_16x16x32_bf16 a[224:227], v[96:99], v[4:7], a[224:227]// 0000000049C8: D3B580E0 07820960
-	v_mfma_f32_16x16x32_bf16 a[228:231], v[96:99], v[8:11], a[228:231]// 0000000049D0: D3B580E4 07921160
-	v_mfma_f32_16x16x32_bf16 a[232:235], v[96:99], v[12:15], a[232:235]// 0000000049D8: D3B580E8 07A21960
-	v_mfma_f32_16x16x32_bf16 a[236:239], v[96:99], v[16:19], a[236:239]// 0000000049E0: D3B580EC 07B22160
-	v_mfma_f32_16x16x32_bf16 a[240:243], v[96:99], v[20:23], a[240:243]// 0000000049E8: D3B580F0 07C22960
-	v_mfma_f32_16x16x32_bf16 a[244:247], v[96:99], v[24:27], a[244:247]// 0000000049F0: D3B580F4 07D23160
-	v_mfma_f32_16x16x32_bf16 a[248:251], v[96:99], v[28:31], a[248:251]// 0000000049F8: D3B580F8 07E23960
-	v_mfma_f32_16x16x32_bf16 a[252:255], v[96:99], v[32:35], a[252:255]// 000000004A00: D3B580FC 07F24160
-	s_waitcnt lgkmcnt(0)                                       // 000000004A08: BF8CC07F
-	v_mfma_f32_16x16x32_bf16 a[0:3], v[100:103], v[36:39], a[0:3]// 000000004A0C: D3B58000 04024964
-	v_mfma_f32_16x16x32_bf16 a[4:7], v[100:103], v[40:43], a[4:7]// 000000004A14: D3B58004 04125164
-	v_mfma_f32_16x16x32_bf16 a[8:11], v[100:103], v[44:47], a[8:11]// 000000004A1C: D3B58008 04225964
-	v_mfma_f32_16x16x32_bf16 a[12:15], v[100:103], v[48:51], a[12:15]// 000000004A24: D3B5800C 04326164
-	v_mfma_f32_16x16x32_bf16 a[16:19], v[100:103], v[52:55], a[16:19]// 000000004A2C: D3B58010 04426964
-	v_mfma_f32_16x16x32_bf16 a[20:23], v[100:103], v[56:59], a[20:23]// 000000004A34: D3B58014 04527164
-	v_mfma_f32_16x16x32_bf16 a[24:27], v[100:103], v[60:63], a[24:27]// 000000004A3C: D3B58018 04627964
-	v_mfma_f32_16x16x32_bf16 a[28:31], v[100:103], v[64:67], a[28:31]// 000000004A44: D3B5801C 04728164
-	v_mfma_f32_16x16x32_bf16 a[32:35], v[104:107], v[36:39], a[32:35]// 000000004A4C: D3B58020 04824968
-	v_mfma_f32_16x16x32_bf16 a[36:39], v[104:107], v[40:43], a[36:39]// 000000004A54: D3B58024 04925168
-	v_mfma_f32_16x16x32_bf16 a[40:43], v[104:107], v[44:47], a[40:43]// 000000004A5C: D3B58028 04A25968
-	v_mfma_f32_16x16x32_bf16 a[44:47], v[104:107], v[48:51], a[44:47]// 000000004A64: D3B5802C 04B26168
-	v_mfma_f32_16x16x32_bf16 a[48:51], v[104:107], v[52:55], a[48:51]// 000000004A6C: D3B58030 04C26968
-	v_mfma_f32_16x16x32_bf16 a[52:55], v[104:107], v[56:59], a[52:55]// 000000004A74: D3B58034 04D27168
-	v_mfma_f32_16x16x32_bf16 a[56:59], v[104:107], v[60:63], a[56:59]// 000000004A7C: D3B58038 04E27968
-	v_mfma_f32_16x16x32_bf16 a[60:63], v[104:107], v[64:67], a[60:63]// 000000004A84: D3B5803C 04F28168
-	v_mfma_f32_16x16x32_bf16 a[64:67], v[108:111], v[36:39], a[64:67]// 000000004A8C: D3B58040 0502496C
-	v_mfma_f32_16x16x32_bf16 a[68:71], v[108:111], v[40:43], a[68:71]// 000000004A94: D3B58044 0512516C
-	v_mfma_f32_16x16x32_bf16 a[72:75], v[108:111], v[44:47], a[72:75]// 000000004A9C: D3B58048 0522596C
-	v_mfma_f32_16x16x32_bf16 a[76:79], v[108:111], v[48:51], a[76:79]// 000000004AA4: D3B5804C 0532616C
-	v_mfma_f32_16x16x32_bf16 a[80:83], v[108:111], v[52:55], a[80:83]// 000000004AAC: D3B58050 0542696C
-	v_mfma_f32_16x16x32_bf16 a[84:87], v[108:111], v[56:59], a[84:87]// 000000004AB4: D3B58054 0552716C
-	v_mfma_f32_16x16x32_bf16 a[88:91], v[108:111], v[60:63], a[88:91]// 000000004ABC: D3B58058 0562796C
-	v_mfma_f32_16x16x32_bf16 a[92:95], v[108:111], v[64:67], a[92:95]// 000000004AC4: D3B5805C 0572816C
-	v_mfma_f32_16x16x32_bf16 a[96:99], v[112:115], v[36:39], a[96:99]// 000000004ACC: D3B58060 05824970
-	v_mfma_f32_16x16x32_bf16 a[100:103], v[112:115], v[40:43], a[100:103]// 000000004AD4: D3B58064 05925170
-	v_mfma_f32_16x16x32_bf16 a[104:107], v[112:115], v[44:47], a[104:107]// 000000004ADC: D3B58068 05A25970
-	v_mfma_f32_16x16x32_bf16 a[108:111], v[112:115], v[48:51], a[108:111]// 000000004AE4: D3B5806C 05B26170
-	v_mfma_f32_16x16x32_bf16 a[112:115], v[112:115], v[52:55], a[112:115]// 000000004AEC: D3B58070 05C26970
-	v_mfma_f32_16x16x32_bf16 a[116:119], v[112:115], v[56:59], a[116:119]// 000000004AF4: D3B58074 05D27170
-	v_mfma_f32_16x16x32_bf16 a[120:123], v[112:115], v[60:63], a[120:123]// 000000004AFC: D3B58078 05E27970
-	v_mfma_f32_16x16x32_bf16 a[124:127], v[112:115], v[64:67], a[124:127]// 000000004B04: D3B5807C 05F28170
-	v_mfma_f32_16x16x32_bf16 a[128:131], v[116:119], v[36:39], a[128:131]// 000000004B0C: D3B58080 06024974
-	v_mfma_f32_16x16x32_bf16 a[132:135], v[116:119], v[40:43], a[132:135]// 000000004B14: D3B58084 06125174
-	v_mfma_f32_16x16x32_bf16 a[136:139], v[116:119], v[44:47], a[136:139]// 000000004B1C: D3B58088 06225974
-	v_mfma_f32_16x16x32_bf16 a[140:143], v[116:119], v[48:51], a[140:143]// 000000004B24: D3B5808C 06326174
-	v_mfma_f32_16x16x32_bf16 a[144:147], v[116:119], v[52:55], a[144:147]// 000000004B2C: D3B58090 06426974
-	v_mfma_f32_16x16x32_bf16 a[148:151], v[116:119], v[56:59], a[148:151]// 000000004B34: D3B58094 06527174
-	v_mfma_f32_16x16x32_bf16 a[152:155], v[116:119], v[60:63], a[152:155]// 000000004B3C: D3B58098 06627974
-	v_mfma_f32_16x16x32_bf16 a[156:159], v[116:119], v[64:67], a[156:159]// 000000004B44: D3B5809C 06728174
-	v_mfma_f32_16x16x32_bf16 a[160:163], v[120:123], v[36:39], a[160:163]// 000000004B4C: D3B580A0 06824978
-	v_mfma_f32_16x16x32_bf16 a[164:167], v[120:123], v[40:43], a[164:167]// 000000004B54: D3B580A4 06925178
-	v_mfma_f32_16x16x32_bf16 a[168:171], v[120:123], v[44:47], a[168:171]// 000000004B5C: D3B580A8 06A25978
-	v_mfma_f32_16x16x32_bf16 a[172:175], v[120:123], v[48:51], a[172:175]// 000000004B64: D3B580AC 06B26178
-	v_mfma_f32_16x16x32_bf16 a[176:179], v[120:123], v[52:55], a[176:179]// 000000004B6C: D3B580B0 06C26978
-	v_mfma_f32_16x16x32_bf16 a[180:183], v[120:123], v[56:59], a[180:183]// 000000004B74: D3B580B4 06D27178
-	v_mfma_f32_16x16x32_bf16 a[184:187], v[120:123], v[60:63], a[184:187]// 000000004B7C: D3B580B8 06E27978
-	v_mfma_f32_16x16x32_bf16 a[188:191], v[120:123], v[64:67], a[188:191]// 000000004B84: D3B580BC 06F28178
-	v_mfma_f32_16x16x32_bf16 a[192:195], v[124:127], v[36:39], a[192:195]// 000000004B8C: D3B580C0 0702497C
-	v_mfma_f32_16x16x32_bf16 a[196:199], v[124:127], v[40:43], a[196:199]// 000000004B94: D3B580C4 0712517C
-	v_mfma_f32_16x16x32_bf16 a[200:203], v[124:127], v[44:47], a[200:203]// 000000004B9C: D3B580C8 0722597C
-	v_mfma_f32_16x16x32_bf16 a[204:207], v[124:127], v[48:51], a[204:207]// 000000004BA4: D3B580CC 0732617C
-	v_mfma_f32_16x16x32_bf16 a[208:211], v[124:127], v[52:55], a[208:211]// 000000004BAC: D3B580D0 0742697C
-	v_mfma_f32_16x16x32_bf16 a[212:215], v[124:127], v[56:59], a[212:215]// 000000004BB4: D3B580D4 0752717C
-	v_mfma_f32_16x16x32_bf16 a[216:219], v[124:127], v[60:63], a[216:219]// 000000004BBC: D3B580D8 0762797C
-	v_mfma_f32_16x16x32_bf16 a[220:223], v[124:127], v[64:67], a[220:223]// 000000004BC4: D3B580DC 0772817C
-	v_mfma_f32_16x16x32_bf16 a[224:227], v[128:131], v[36:39], a[224:227]// 000000004BCC: D3B580E0 07824980
-	v_mfma_f32_16x16x32_bf16 a[228:231], v[128:131], v[40:43], a[228:231]// 000000004BD4: D3B580E4 07925180
-	v_mfma_f32_16x16x32_bf16 a[232:235], v[128:131], v[44:47], a[232:235]// 000000004BDC: D3B580E8 07A25980
-	v_mfma_f32_16x16x32_bf16 a[236:239], v[128:131], v[48:51], a[236:239]// 000000004BE4: D3B580EC 07B26180
-	v_mfma_f32_16x16x32_bf16 a[240:243], v[128:131], v[52:55], a[240:243]// 000000004BEC: D3B580F0 07C26980
-	v_mfma_f32_16x16x32_bf16 a[244:247], v[128:131], v[56:59], a[244:247]// 000000004BF4: D3B580F4 07D27180
-	v_mfma_f32_16x16x32_bf16 a[248:251], v[128:131], v[60:63], a[248:251]// 000000004BFC: D3B580F8 07E27980
-	v_mfma_f32_16x16x32_bf16 a[252:255], v[128:131], v[64:67], a[252:255]// 000000004C04: D3B580FC 07F28180
-
-	v_lshrrev_b32_e32 v4, 6, v134                              // 000000004C0C: 20090C86
-	v_lshrrev_b32_e32 v5, 1, v4                                // 000000004C10: 200A0881
-	v_mul_lo_u32 v5, 16, v5                                    // 000000004C14: D2850005 00020A90
-	v_and_b32_e32 v1, 63, v134                                 // 000000004C1C: 26030CBF
-	v_lshrrev_b32_e32 v1, 4, v1                                // 000000004C20: 20020284
-	v_lshlrev_b32_e32 v1, 2, v1                                // 000000004C24: 24020282
-	v_add_lshl_u32 v1, v5, v1, 3                               // 000000004C28: D1FE0001 020E0305
-	v_mul_lo_u32 v2, v1, s38                                   // 000000004C30: D2850002 00004D01
-	v_mul_lo_u32 v3, v1, s36                                   // 000000004C38: D2850003 00004901
-	v_and_b32_e32 v0, 1, v4                                    // 000000004C40: 26000881
-	v_mul_lo_u32 v0, 16, v0                                    // 000000004C44: D2850000 00020090
-	v_and_b32_e32 v5, 15, v134                                 // 000000004C4C: 260B0C8F
-	v_add_lshl_u32 v0, v5, v0, 3                               // 000000004C50: D1FE0000 020E0105
-	s_mul_i32 s8, 0x100, s2                                    // 000000004C58: 920802FF 00000100
-	v_add_u32_e32 v0, s8, v0                                   // 000000004C60: 68000008
-	s_mul_i32 s8, 0x100, s3                                    // 000000004C64: 920803FF 00000100
-	v_add_u32_e32 v1, s8, v1                                   // 000000004C6C: 68020208
-
-	v_add_lshl_u32 v11, v3, v0, 1                              // 000000004C70: D1FE000B 02060103
-	v_accvgpr_read_b32 v16, a0                                 // 000000004C78: D3D84010 18000100
-	v_accvgpr_read_b32 v17, a4                                 // 000000004C80: D3D84011 18000104
-	v_accvgpr_read_b32 v18, a8                                 // 000000004C88: D3D84012 18000108
-	v_accvgpr_read_b32 v19, a12                                // 000000004C90: D3D84013 1800010C
-	v_accvgpr_read_b32 v20, a16                                // 000000004C98: D3D84014 18000110
-	v_accvgpr_read_b32 v21, a20                                // 000000004CA0: D3D84015 18000114
-	v_accvgpr_read_b32 v22, a24                                // 000000004CA8: D3D84016 18000118
-	v_accvgpr_read_b32 v23, a28                                // 000000004CB0: D3D84017 1800011C
-	v_accvgpr_read_b32 v24, a32                                // 000000004CB8: D3D84018 18000120
-	v_accvgpr_read_b32 v25, a36                                // 000000004CC0: D3D84019 18000124
-	v_accvgpr_read_b32 v26, a40                                // 000000004CC8: D3D8401A 18000128
-	v_accvgpr_read_b32 v27, a44                                // 000000004CD0: D3D8401B 1800012C
-	v_accvgpr_read_b32 v28, a48                                // 000000004CD8: D3D8401C 18000130
-	v_accvgpr_read_b32 v29, a52                                // 000000004CE0: D3D8401D 18000134
-	v_accvgpr_read_b32 v30, a56                                // 000000004CE8: D3D8401E 18000138
-	v_accvgpr_read_b32 v31, a60                                // 000000004CF0: D3D8401F 1800013C
-	v_accvgpr_read_b32 v32, a64                                // 000000004CF8: D3D84020 18000140
-	v_accvgpr_read_b32 v33, a68                                // 000000004D00: D3D84021 18000144
-	v_accvgpr_read_b32 v34, a72                                // 000000004D08: D3D84022 18000148
-	v_accvgpr_read_b32 v35, a76                                // 000000004D10: D3D84023 1800014C
-	v_accvgpr_read_b32 v36, a80                                // 000000004D18: D3D84024 18000150
-	v_accvgpr_read_b32 v37, a84                                // 000000004D20: D3D84025 18000154
-	v_accvgpr_read_b32 v38, a88                                // 000000004D28: D3D84026 18000158
-	v_accvgpr_read_b32 v39, a92                                // 000000004D30: D3D84027 1800015C
-	v_accvgpr_read_b32 v40, a96                                // 000000004D38: D3D84028 18000160
-	v_accvgpr_read_b32 v41, a100                               // 000000004D40: D3D84029 18000164
-	v_accvgpr_read_b32 v42, a104                               // 000000004D48: D3D8402A 18000168
-	v_accvgpr_read_b32 v43, a108                               // 000000004D50: D3D8402B 1800016C
-	v_accvgpr_read_b32 v44, a112                               // 000000004D58: D3D8402C 18000170
-	v_accvgpr_read_b32 v45, a116                               // 000000004D60: D3D8402D 18000174
-	v_accvgpr_read_b32 v46, a120                               // 000000004D68: D3D8402E 18000178
-	v_accvgpr_read_b32 v47, a124                               // 000000004D70: D3D8402F 1800017C
-	v_accvgpr_read_b32 v48, a128                               // 000000004D78: D3D84030 18000180
-	v_accvgpr_read_b32 v49, a132                               // 000000004D80: D3D84031 18000184
-	v_accvgpr_read_b32 v50, a136                               // 000000004D88: D3D84032 18000188
-	v_accvgpr_read_b32 v51, a140                               // 000000004D90: D3D84033 1800018C
-	v_accvgpr_read_b32 v52, a144                               // 000000004D98: D3D84034 18000190
-	v_accvgpr_read_b32 v53, a148                               // 000000004DA0: D3D84035 18000194
-	v_accvgpr_read_b32 v54, a152                               // 000000004DA8: D3D84036 18000198
-	v_accvgpr_read_b32 v55, a156                               // 000000004DB0: D3D84037 1800019C
-	v_accvgpr_read_b32 v56, a160                               // 000000004DB8: D3D84038 180001A0
-	v_accvgpr_read_b32 v57, a164                               // 000000004DC0: D3D84039 180001A4
-	v_accvgpr_read_b32 v58, a168                               // 000000004DC8: D3D8403A 180001A8
-	v_accvgpr_read_b32 v59, a172                               // 000000004DD0: D3D8403B 180001AC
-	v_accvgpr_read_b32 v60, a176                               // 000000004DD8: D3D8403C 180001B0
-	v_accvgpr_read_b32 v61, a180                               // 000000004DE0: D3D8403D 180001B4
-	v_accvgpr_read_b32 v62, a184                               // 000000004DE8: D3D8403E 180001B8
-	v_accvgpr_read_b32 v63, a188                               // 000000004DF0: D3D8403F 180001BC
-	v_accvgpr_read_b32 v64, a192                               // 000000004DF8: D3D84040 180001C0
-	v_accvgpr_read_b32 v65, a196                               // 000000004E00: D3D84041 180001C4
-	v_accvgpr_read_b32 v66, a200                               // 000000004E08: D3D84042 180001C8
-	v_accvgpr_read_b32 v67, a204                               // 000000004E10: D3D84043 180001CC
-	v_accvgpr_read_b32 v68, a208                               // 000000004E18: D3D84044 180001D0
-	v_accvgpr_read_b32 v69, a212                               // 000000004E20: D3D84045 180001D4
-	v_accvgpr_read_b32 v70, a216                               // 000000004E28: D3D84046 180001D8
-	v_accvgpr_read_b32 v71, a220                               // 000000004E30: D3D84047 180001DC
-	v_accvgpr_read_b32 v72, a224                               // 000000004E38: D3D84048 180001E0
-	v_accvgpr_read_b32 v73, a228                               // 000000004E40: D3D84049 180001E4
-	v_accvgpr_read_b32 v74, a232                               // 000000004E48: D3D8404A 180001E8
-	v_accvgpr_read_b32 v75, a236                               // 000000004E50: D3D8404B 180001EC
-	v_accvgpr_read_b32 v76, a240                               // 000000004E58: D3D8404C 180001F0
-	v_accvgpr_read_b32 v77, a244                               // 000000004E60: D3D8404D 180001F4
-	v_accvgpr_read_b32 v78, a248                               // 000000004E68: D3D8404E 180001F8
-	v_accvgpr_read_b32 v79, a252                               // 000000004E70: D3D8404F 180001FC
-	v_accvgpr_read_b32 v80, a1                                 // 000000004E78: D3D84050 18000101
-	v_accvgpr_read_b32 v81, a5                                 // 000000004E80: D3D84051 18000105
-	v_accvgpr_read_b32 v82, a9                                 // 000000004E88: D3D84052 18000109
-	v_accvgpr_read_b32 v83, a13                                // 000000004E90: D3D84053 1800010D
-	v_accvgpr_read_b32 v84, a17                                // 000000004E98: D3D84054 18000111
-	v_accvgpr_read_b32 v85, a21                                // 000000004EA0: D3D84055 18000115
-	v_accvgpr_read_b32 v86, a25                                // 000000004EA8: D3D84056 18000119
-	v_accvgpr_read_b32 v87, a29                                // 000000004EB0: D3D84057 1800011D
-	v_accvgpr_read_b32 v88, a33                                // 000000004EB8: D3D84058 18000121
-	v_accvgpr_read_b32 v89, a37                                // 000000004EC0: D3D84059 18000125
-	v_accvgpr_read_b32 v90, a41                                // 000000004EC8: D3D8405A 18000129
-	v_accvgpr_read_b32 v91, a45                                // 000000004ED0: D3D8405B 1800012D
-	v_accvgpr_read_b32 v92, a49                                // 000000004ED8: D3D8405C 18000131
-	v_accvgpr_read_b32 v93, a53                                // 000000004EE0: D3D8405D 18000135
-	v_accvgpr_read_b32 v94, a57                                // 000000004EE8: D3D8405E 18000139
-	v_accvgpr_read_b32 v95, a61                                // 000000004EF0: D3D8405F 1800013D
-	v_accvgpr_read_b32 v96, a65                                // 000000004EF8: D3D84060 18000141
-	v_accvgpr_read_b32 v97, a69                                // 000000004F00: D3D84061 18000145
-	v_accvgpr_read_b32 v98, a73                                // 000000004F08: D3D84062 18000149
-	v_accvgpr_read_b32 v99, a77                                // 000000004F10: D3D84063 1800014D
-	v_accvgpr_read_b32 v100, a81                               // 000000004F18: D3D84064 18000151
-	v_accvgpr_read_b32 v101, a85                               // 000000004F20: D3D84065 18000155
-	v_accvgpr_read_b32 v102, a89                               // 000000004F28: D3D84066 18000159
-	v_accvgpr_read_b32 v103, a93                               // 000000004F30: D3D84067 1800015D
-	v_accvgpr_read_b32 v104, a97                               // 000000004F38: D3D84068 18000161
-	v_accvgpr_read_b32 v105, a101                              // 000000004F40: D3D84069 18000165
-	v_accvgpr_read_b32 v106, a105                              // 000000004F48: D3D8406A 18000169
-	v_accvgpr_read_b32 v107, a109                              // 000000004F50: D3D8406B 1800016D
-	v_accvgpr_read_b32 v108, a113                              // 000000004F58: D3D8406C 18000171
-	v_accvgpr_read_b32 v109, a117                              // 000000004F60: D3D8406D 18000175
-	v_accvgpr_read_b32 v110, a121                              // 000000004F68: D3D8406E 18000179
-	v_accvgpr_read_b32 v111, a125                              // 000000004F70: D3D8406F 1800017D
-	v_accvgpr_read_b32 v112, a129                              // 000000004F78: D3D84070 18000181
-	v_accvgpr_read_b32 v113, a133                              // 000000004F80: D3D84071 18000185
-	v_accvgpr_read_b32 v114, a137                              // 000000004F88: D3D84072 18000189
-	v_accvgpr_read_b32 v115, a141                              // 000000004F90: D3D84073 1800018D
-	v_accvgpr_read_b32 v116, a145                              // 000000004F98: D3D84074 18000191
-	v_accvgpr_read_b32 v117, a149                              // 000000004FA0: D3D84075 18000195
-	v_accvgpr_read_b32 v118, a153                              // 000000004FA8: D3D84076 18000199
-	v_accvgpr_read_b32 v119, a157                              // 000000004FB0: D3D84077 1800019D
-	v_accvgpr_read_b32 v120, a161                              // 000000004FB8: D3D84078 180001A1
-	v_accvgpr_read_b32 v121, a165                              // 000000004FC0: D3D84079 180001A5
-	v_accvgpr_read_b32 v122, a169                              // 000000004FC8: D3D8407A 180001A9
-	v_accvgpr_read_b32 v123, a173                              // 000000004FD0: D3D8407B 180001AD
-	v_accvgpr_read_b32 v124, a177                              // 000000004FD8: D3D8407C 180001B1
-	v_accvgpr_read_b32 v125, a181                              // 000000004FE0: D3D8407D 180001B5
-	v_accvgpr_read_b32 v126, a185                              // 000000004FE8: D3D8407E 180001B9
-	v_accvgpr_read_b32 v127, a189                              // 000000004FF0: D3D8407F 180001BD
-	v_accvgpr_read_b32 v136, a193                              // 000000004FF8: D3D84088 180001C1
-	v_accvgpr_read_b32 v137, a197                              // 000000005000: D3D84089 180001C5
-	v_accvgpr_read_b32 v138, a201                              // 000000005008: D3D8408A 180001C9
-	v_accvgpr_read_b32 v139, a205                              // 000000005010: D3D8408B 180001CD
-	v_accvgpr_read_b32 v140, a209                              // 000000005018: D3D8408C 180001D1
-	v_accvgpr_read_b32 v141, a213                              // 000000005020: D3D8408D 180001D5
-	v_accvgpr_read_b32 v142, a217                              // 000000005028: D3D8408E 180001D9
-	v_accvgpr_read_b32 v143, a221                              // 000000005030: D3D8408F 180001DD
-	v_accvgpr_read_b32 v144, a225                              // 000000005038: D3D84090 180001E1
-	v_accvgpr_read_b32 v145, a229                              // 000000005040: D3D84091 180001E5
-	v_accvgpr_read_b32 v146, a233                              // 000000005048: D3D84092 180001E9
-	v_accvgpr_read_b32 v147, a237                              // 000000005050: D3D84093 180001ED
-	v_accvgpr_read_b32 v148, a241                              // 000000005058: D3D84094 180001F1
-	v_accvgpr_read_b32 v149, a245                              // 000000005060: D3D84095 180001F5
-	v_accvgpr_read_b32 v150, a249                              // 000000005068: D3D84096 180001F9
-	v_accvgpr_read_b32 v151, a253                              // 000000005070: D3D84097 180001FD
-	v_accvgpr_read_b32 v152, a2                                // 000000005078: D3D84098 18000102
-	v_accvgpr_read_b32 v153, a6                                // 000000005080: D3D84099 18000106
-	v_accvgpr_read_b32 v154, a10                               // 000000005088: D3D8409A 1800010A
-	v_accvgpr_read_b32 v155, a14                               // 000000005090: D3D8409B 1800010E
-	v_accvgpr_read_b32 v156, a18                               // 000000005098: D3D8409C 18000112
-	v_accvgpr_read_b32 v157, a22                               // 0000000050A0: D3D8409D 18000116
-	v_accvgpr_read_b32 v158, a26                               // 0000000050A8: D3D8409E 1800011A
-	v_accvgpr_read_b32 v159, a30                               // 0000000050B0: D3D8409F 1800011E
-	v_accvgpr_read_b32 v160, a34                               // 0000000050B8: D3D840A0 18000122
-	v_accvgpr_read_b32 v161, a38                               // 0000000050C0: D3D840A1 18000126
-	v_accvgpr_read_b32 v162, a42                               // 0000000050C8: D3D840A2 1800012A
-	v_accvgpr_read_b32 v163, a46                               // 0000000050D0: D3D840A3 1800012E
-	v_accvgpr_read_b32 v164, a50                               // 0000000050D8: D3D840A4 18000132
-	v_accvgpr_read_b32 v165, a54                               // 0000000050E0: D3D840A5 18000136
-	v_accvgpr_read_b32 v166, a58                               // 0000000050E8: D3D840A6 1800013A
-	v_accvgpr_read_b32 v167, a62                               // 0000000050F0: D3D840A7 1800013E
-	v_accvgpr_read_b32 v168, a66                               // 0000000050F8: D3D840A8 18000142
-	v_accvgpr_read_b32 v169, a70                               // 000000005100: D3D840A9 18000146
-	v_accvgpr_read_b32 v170, a74                               // 000000005108: D3D840AA 1800014A
-	v_accvgpr_read_b32 v171, a78                               // 000000005110: D3D840AB 1800014E
-	v_accvgpr_read_b32 v172, a82                               // 000000005118: D3D840AC 18000152
-	v_accvgpr_read_b32 v173, a86                               // 000000005120: D3D840AD 18000156
-	v_accvgpr_read_b32 v174, a90                               // 000000005128: D3D840AE 1800015A
-	v_accvgpr_read_b32 v175, a94                               // 000000005130: D3D840AF 1800015E
-	v_accvgpr_read_b32 v176, a98                               // 000000005138: D3D840B0 18000162
-	v_accvgpr_read_b32 v177, a102                              // 000000005140: D3D840B1 18000166
-	v_accvgpr_read_b32 v178, a106                              // 000000005148: D3D840B2 1800016A
-	v_accvgpr_read_b32 v179, a110                              // 000000005150: D3D840B3 1800016E
-	v_accvgpr_read_b32 v180, a114                              // 000000005158: D3D840B4 18000172
-	v_accvgpr_read_b32 v181, a118                              // 000000005160: D3D840B5 18000176
-	v_accvgpr_read_b32 v182, a122                              // 000000005168: D3D840B6 1800017A
-	v_accvgpr_read_b32 v183, a126                              // 000000005170: D3D840B7 1800017E
-	v_accvgpr_read_b32 v184, a130                              // 000000005178: D3D840B8 18000182
-	v_accvgpr_read_b32 v185, a134                              // 000000005180: D3D840B9 18000186
-	v_accvgpr_read_b32 v186, a138                              // 000000005188: D3D840BA 1800018A
-	v_accvgpr_read_b32 v187, a142                              // 000000005190: D3D840BB 1800018E
-	v_accvgpr_read_b32 v188, a146                              // 000000005198: D3D840BC 18000192
-	v_accvgpr_read_b32 v189, a150                              // 0000000051A0: D3D840BD 18000196
-	v_accvgpr_read_b32 v190, a154                              // 0000000051A8: D3D840BE 1800019A
-	v_accvgpr_read_b32 v191, a158                              // 0000000051B0: D3D840BF 1800019E
-	v_accvgpr_read_b32 v192, a162                              // 0000000051B8: D3D840C0 180001A2
-	v_accvgpr_read_b32 v193, a166                              // 0000000051C0: D3D840C1 180001A6
-	v_accvgpr_read_b32 v194, a170                              // 0000000051C8: D3D840C2 180001AA
-	v_accvgpr_read_b32 v195, a174                              // 0000000051D0: D3D840C3 180001AE
-	v_accvgpr_read_b32 v196, a178                              // 0000000051D8: D3D840C4 180001B2
-	v_accvgpr_read_b32 v197, a182                              // 0000000051E0: D3D840C5 180001B6
-	v_accvgpr_read_b32 v198, a186                              // 0000000051E8: D3D840C6 180001BA
-	v_accvgpr_read_b32 v199, a190                              // 0000000051F0: D3D840C7 180001BE
-	v_accvgpr_read_b32 v200, a194                              // 0000000051F8: D3D840C8 180001C2
-	v_accvgpr_read_b32 v201, a198                              // 000000005200: D3D840C9 180001C6
-	v_accvgpr_read_b32 v202, a202                              // 000000005208: D3D840CA 180001CA
-	v_accvgpr_read_b32 v203, a206                              // 000000005210: D3D840CB 180001CE
-	v_accvgpr_read_b32 v204, a210                              // 000000005218: D3D840CC 180001D2
-	v_accvgpr_read_b32 v205, a214                              // 000000005220: D3D840CD 180001D6
-	v_accvgpr_read_b32 v206, a218                              // 000000005228: D3D840CE 180001DA
-	v_accvgpr_read_b32 v207, a222                              // 000000005230: D3D840CF 180001DE
-	v_accvgpr_read_b32 v208, a226                              // 000000005238: D3D840D0 180001E2
-	v_accvgpr_read_b32 v209, a230                              // 000000005240: D3D840D1 180001E6
-	v_accvgpr_read_b32 v210, a234                              // 000000005248: D3D840D2 180001EA
-	v_accvgpr_read_b32 v211, a238                              // 000000005250: D3D840D3 180001EE
-	v_accvgpr_read_b32 v212, a242                              // 000000005258: D3D840D4 180001F2
-	v_accvgpr_read_b32 v213, a246                              // 000000005260: D3D840D5 180001F6
-	v_accvgpr_read_b32 v214, a250                              // 000000005268: D3D840D6 180001FA
-	v_accvgpr_read_b32 v215, a254                              // 000000005270: D3D840D7 180001FE
-	v_accvgpr_read_b32 v216, a3                                // 000000005278: D3D840D8 18000103
-	v_accvgpr_read_b32 v217, a7                                // 000000005280: D3D840D9 18000107
-	v_accvgpr_read_b32 v218, a11                               // 000000005288: D3D840DA 1800010B
-	v_accvgpr_read_b32 v219, a15                               // 000000005290: D3D840DB 1800010F
-	v_accvgpr_read_b32 v220, a19                               // 000000005298: D3D840DC 18000113
-	v_accvgpr_read_b32 v221, a23                               // 0000000052A0: D3D840DD 18000117
-	v_accvgpr_read_b32 v222, a27                               // 0000000052A8: D3D840DE 1800011B
-	v_accvgpr_read_b32 v223, a31                               // 0000000052B0: D3D840DF 1800011F
-	v_accvgpr_read_b32 v224, a35                               // 0000000052B8: D3D840E0 18000123
-	v_accvgpr_read_b32 v225, a39                               // 0000000052C0: D3D840E1 18000127
-	v_accvgpr_read_b32 v226, a43                               // 0000000052C8: D3D840E2 1800012B
-	v_accvgpr_read_b32 v227, a47                               // 0000000052D0: D3D840E3 1800012F
-	v_accvgpr_read_b32 v228, a51                               // 0000000052D8: D3D840E4 18000133
-	v_accvgpr_read_b32 v229, a55                               // 0000000052E0: D3D840E5 18000137
-	v_accvgpr_read_b32 v230, a59                               // 0000000052E8: D3D840E6 1800013B
-	v_accvgpr_read_b32 v231, a63                               // 0000000052F0: D3D840E7 1800013F
-	v_accvgpr_read_b32 v232, a67                               // 0000000052F8: D3D840E8 18000143
-	v_accvgpr_read_b32 v233, a71                               // 000000005300: D3D840E9 18000147
-	v_accvgpr_read_b32 v234, a75                               // 000000005308: D3D840EA 1800014B
-	v_accvgpr_read_b32 v235, a79                               // 000000005310: D3D840EB 1800014F
-	v_accvgpr_read_b32 v236, a83                               // 000000005318: D3D840EC 18000153
-	v_accvgpr_read_b32 v237, a87                               // 000000005320: D3D840ED 18000157
-	v_accvgpr_read_b32 v238, a91                               // 000000005328: D3D840EE 1800015B
-	v_accvgpr_read_b32 v239, a95                               // 000000005330: D3D840EF 1800015F
-	v_accvgpr_read_b32 v240, a99                               // 000000005338: D3D840F0 18000163
-	v_accvgpr_read_b32 v241, a103                              // 000000005340: D3D840F1 18000167
-	v_accvgpr_read_b32 v242, a107                              // 000000005348: D3D840F2 1800016B
-	v_accvgpr_read_b32 v243, a111                              // 000000005350: D3D840F3 1800016F
-	v_accvgpr_read_b32 v244, a115                              // 000000005358: D3D840F4 18000173
-	v_accvgpr_read_b32 v245, a119                              // 000000005360: D3D840F5 18000177
-	v_accvgpr_read_b32 v246, a123                              // 000000005368: D3D840F6 1800017B
-	v_accvgpr_read_b32 v247, a127                              // 000000005370: D3D840F7 1800017F
-	v_mov_b32_e32 v8, 0xffff0000                               // 000000005378: 7E1002FF FFFF0000
-	v_mov_b32_e32 v9, 0x7fff0000                               // 000000005380: 7E1202FF 7FFF0000
-	v_mov_b32_e32 v10, 0x7fff                                  // 000000005388: 7E1402FF 00007FFF
-	v_cvt_pk_bf16_f32 v16, v16, v17                            // 000000005390: D2680010 00022310
-	v_cvt_pk_bf16_f32 v17, v18, v19                            // 000000005398: D2680011 00022712
-	v_cvt_pk_bf16_f32 v18, v20, v21                            // 0000000053A0: D2680012 00022B14
-	v_cvt_pk_bf16_f32 v19, v22, v23                            // 0000000053A8: D2680013 00022F16
-	buffer_store_dwordx4 v[16:19], v11, s[16:19], 0 offen nt   // 0000000053B0: E07E1000 8004100B
-	v_cvt_pk_bf16_f32 v24, v24, v25                            // 0000000053B8: D2680018 00023318
-	v_cvt_pk_bf16_f32 v25, v26, v27                            // 0000000053C0: D2680019 0002371A
-	v_cvt_pk_bf16_f32 v26, v28, v29                            // 0000000053C8: D268001A 00023B1C
-	v_cvt_pk_bf16_f32 v27, v30, v31                            // 0000000053D0: D268001B 00023F1E
-	s_lshl_b32 s12, s36, 1                                     // 0000000053D8: 8E0C8124
-	s_add_u32 s16, s16, s12                                    // 0000000053DC: 80100C10
-	s_addc_u32 s17, s17, 0                                     // 0000000053E0: 82118011
-	buffer_store_dwordx4 v[24:27], v11, s[16:19], 0 offen nt   // 0000000053E4: E07E1000 8004180B
-	v_cvt_pk_bf16_f32 v32, v32, v33                            // 0000000053EC: D2680020 00024320
-	v_cvt_pk_bf16_f32 v33, v34, v35                            // 0000000053F4: D2680021 00024722
-	v_cvt_pk_bf16_f32 v34, v36, v37                            // 0000000053FC: D2680022 00024B24
-	v_cvt_pk_bf16_f32 v35, v38, v39                            // 000000005404: D2680023 00024F26
-	s_lshl_b32 s12, s36, 1                                     // 00000000540C: 8E0C8124
-	s_add_u32 s16, s16, s12                                    // 000000005410: 80100C10
-	s_addc_u32 s17, s17, 0                                     // 000000005414: 82118011
-	buffer_store_dwordx4 v[32:35], v11, s[16:19], 0 offen nt   // 000000005418: E07E1000 8004200B
-	v_cvt_pk_bf16_f32 v40, v40, v41                            // 000000005420: D2680028 00025328
-	v_cvt_pk_bf16_f32 v41, v42, v43                            // 000000005428: D2680029 0002572A
-	v_cvt_pk_bf16_f32 v42, v44, v45                            // 000000005430: D268002A 00025B2C
-	v_cvt_pk_bf16_f32 v43, v46, v47                            // 000000005438: D268002B 00025F2E
-	s_lshl_b32 s12, s36, 1                                     // 000000005440: 8E0C8124
-	s_add_u32 s16, s16, s12                                    // 000000005444: 80100C10
-	s_addc_u32 s17, s17, 0                                     // 000000005448: 82118011
-	buffer_store_dwordx4 v[40:43], v11, s[16:19], 0 offen nt   // 00000000544C: E07E1000 8004280B
-	v_cvt_pk_bf16_f32 v48, v48, v49                            // 000000005454: D2680030 00026330
-	v_cvt_pk_bf16_f32 v49, v50, v51                            // 00000000545C: D2680031 00026732
-	v_cvt_pk_bf16_f32 v50, v52, v53                            // 000000005464: D2680032 00026B34
-	v_cvt_pk_bf16_f32 v51, v54, v55                            // 00000000546C: D2680033 00026F36
-	s_lshl_b32 s12, s36, 1                                     // 000000005474: 8E0C8124
-	s_add_u32 s16, s16, s12                                    // 000000005478: 80100C10
-	s_addc_u32 s17, s17, 0                                     // 00000000547C: 82118011
-	buffer_store_dwordx4 v[48:51], v11, s[16:19], 0 offen nt   // 000000005480: E07E1000 8004300B
-	v_cvt_pk_bf16_f32 v56, v56, v57                            // 000000005488: D2680038 00027338
-	v_cvt_pk_bf16_f32 v57, v58, v59                            // 000000005490: D2680039 0002773A
-	v_cvt_pk_bf16_f32 v58, v60, v61                            // 000000005498: D268003A 00027B3C
-	v_cvt_pk_bf16_f32 v59, v62, v63                            // 0000000054A0: D268003B 00027F3E
-	s_lshl_b32 s12, s36, 1                                     // 0000000054A8: 8E0C8124
-	s_add_u32 s16, s16, s12                                    // 0000000054AC: 80100C10
-	s_addc_u32 s17, s17, 0                                     // 0000000054B0: 82118011
-	buffer_store_dwordx4 v[56:59], v11, s[16:19], 0 offen nt   // 0000000054B4: E07E1000 8004380B
-	v_cvt_pk_bf16_f32 v64, v64, v65                            // 0000000054BC: D2680040 00028340
-	v_cvt_pk_bf16_f32 v65, v66, v67                            // 0000000054C4: D2680041 00028742
-	v_cvt_pk_bf16_f32 v66, v68, v69                            // 0000000054CC: D2680042 00028B44
-	v_cvt_pk_bf16_f32 v67, v70, v71                            // 0000000054D4: D2680043 00028F46
-	s_lshl_b32 s12, s36, 1                                     // 0000000054DC: 8E0C8124
-	s_add_u32 s16, s16, s12                                    // 0000000054E0: 80100C10
-	s_addc_u32 s17, s17, 0                                     // 0000000054E4: 82118011
-	buffer_store_dwordx4 v[64:67], v11, s[16:19], 0 offen nt   // 0000000054E8: E07E1000 8004400B
-	v_cvt_pk_bf16_f32 v72, v72, v73                            // 0000000054F0: D2680048 00029348
-	v_cvt_pk_bf16_f32 v73, v74, v75                            // 0000000054F8: D2680049 0002974A
-	v_cvt_pk_bf16_f32 v74, v76, v77                            // 000000005500: D268004A 00029B4C
-	v_cvt_pk_bf16_f32 v75, v78, v79                            // 000000005508: D268004B 00029F4E
-	s_lshl_b32 s12, s36, 1                                     // 000000005510: 8E0C8124
-	s_add_u32 s16, s16, s12                                    // 000000005514: 80100C10
-	s_addc_u32 s17, s17, 0                                     // 000000005518: 82118011
-	buffer_store_dwordx4 v[72:75], v11, s[16:19], 0 offen nt   // 00000000551C: E07E1000 8004480B
-	v_cvt_pk_bf16_f32 v80, v80, v81                            // 000000005524: D2680050 0002A350
-	v_cvt_pk_bf16_f32 v81, v82, v83                            // 00000000552C: D2680051 0002A752
-	v_cvt_pk_bf16_f32 v82, v84, v85                            // 000000005534: D2680052 0002AB54
-	v_cvt_pk_bf16_f32 v83, v86, v87                            // 00000000553C: D2680053 0002AF56
-	s_lshl_b32 s12, s36, 1                                     // 000000005544: 8E0C8124
-	s_add_u32 s16, s16, s12                                    // 000000005548: 80100C10
-	s_addc_u32 s17, s17, 0                                     // 00000000554C: 82118011
-	buffer_store_dwordx4 v[80:83], v11, s[16:19], 0 offen nt   // 000000005550: E07E1000 8004500B
-	v_cvt_pk_bf16_f32 v88, v88, v89                            // 000000005558: D2680058 0002B358
-	v_cvt_pk_bf16_f32 v89, v90, v91                            // 000000005560: D2680059 0002B75A
-	v_cvt_pk_bf16_f32 v90, v92, v93                            // 000000005568: D268005A 0002BB5C
-	v_cvt_pk_bf16_f32 v91, v94, v95                            // 000000005570: D268005B 0002BF5E
-	s_lshl_b32 s12, s36, 1                                     // 000000005578: 8E0C8124
-	s_add_u32 s16, s16, s12                                    // 00000000557C: 80100C10
-	s_addc_u32 s17, s17, 0                                     // 000000005580: 82118011
-	buffer_store_dwordx4 v[88:91], v11, s[16:19], 0 offen nt   // 000000005584: E07E1000 8004580B
-	v_cvt_pk_bf16_f32 v96, v96, v97                            // 00000000558C: D2680060 0002C360
-	v_cvt_pk_bf16_f32 v97, v98, v99                            // 000000005594: D2680061 0002C762
-	v_cvt_pk_bf16_f32 v98, v100, v101                          // 00000000559C: D2680062 0002CB64
-	v_cvt_pk_bf16_f32 v99, v102, v103                          // 0000000055A4: D2680063 0002CF66
-	s_lshl_b32 s12, s36, 1                                     // 0000000055AC: 8E0C8124
-	s_add_u32 s16, s16, s12                                    // 0000000055B0: 80100C10
-	s_addc_u32 s17, s17, 0                                     // 0000000055B4: 82118011
-	buffer_store_dwordx4 v[96:99], v11, s[16:19], 0 offen nt   // 0000000055B8: E07E1000 8004600B
-	v_cvt_pk_bf16_f32 v104, v104, v105                         // 0000000055C0: D2680068 0002D368
-	v_cvt_pk_bf16_f32 v105, v106, v107                         // 0000000055C8: D2680069 0002D76A
-	v_cvt_pk_bf16_f32 v106, v108, v109                         // 0000000055D0: D268006A 0002DB6C
-	v_cvt_pk_bf16_f32 v107, v110, v111                         // 0000000055D8: D268006B 0002DF6E
-	s_lshl_b32 s12, s36, 1                                     // 0000000055E0: 8E0C8124
-	s_add_u32 s16, s16, s12                                    // 0000000055E4: 80100C10
-	s_addc_u32 s17, s17, 0                                     // 0000000055E8: 82118011
-	buffer_store_dwordx4 v[104:107], v11, s[16:19], 0 offen nt // 0000000055EC: E07E1000 8004680B
-	v_cvt_pk_bf16_f32 v112, v112, v113                         // 0000000055F4: D2680070 0002E370
-	v_cvt_pk_bf16_f32 v113, v114, v115                         // 0000000055FC: D2680071 0002E772
-	v_cvt_pk_bf16_f32 v114, v116, v117                         // 000000005604: D2680072 0002EB74
-	v_cvt_pk_bf16_f32 v115, v118, v119                         // 00000000560C: D2680073 0002EF76
-	s_lshl_b32 s12, s36, 1                                     // 000000005614: 8E0C8124
-	s_add_u32 s16, s16, s12                                    // 000000005618: 80100C10
-	s_addc_u32 s17, s17, 0                                     // 00000000561C: 82118011
-	buffer_store_dwordx4 v[112:115], v11, s[16:19], 0 offen nt // 000000005620: E07E1000 8004700B
-	v_cvt_pk_bf16_f32 v120, v120, v121                         // 000000005628: D2680078 0002F378
-	v_cvt_pk_bf16_f32 v121, v122, v123                         // 000000005630: D2680079 0002F77A
-	v_cvt_pk_bf16_f32 v122, v124, v125                         // 000000005638: D268007A 0002FB7C
-	v_cvt_pk_bf16_f32 v123, v126, v127                         // 000000005640: D268007B 0002FF7E
-	s_lshl_b32 s12, s36, 1                                     // 000000005648: 8E0C8124
-	s_add_u32 s16, s16, s12                                    // 00000000564C: 80100C10
-	s_addc_u32 s17, s17, 0                                     // 000000005650: 82118011
-	buffer_store_dwordx4 v[120:123], v11, s[16:19], 0 offen nt // 000000005654: E07E1000 8004780B
-	v_cvt_pk_bf16_f32 v136, v136, v137                         // 00000000565C: D2680088 00031388
-	v_cvt_pk_bf16_f32 v137, v138, v139                         // 000000005664: D2680089 0003178A
-	v_cvt_pk_bf16_f32 v138, v140, v141                         // 00000000566C: D268008A 00031B8C
-	v_cvt_pk_bf16_f32 v139, v142, v143                         // 000000005674: D268008B 00031F8E
-	s_lshl_b32 s12, s36, 1                                     // 00000000567C: 8E0C8124
-	s_add_u32 s16, s16, s12                                    // 000000005680: 80100C10
-	s_addc_u32 s17, s17, 0                                     // 000000005684: 82118011
-	buffer_store_dwordx4 v[136:139], v11, s[16:19], 0 offen nt // 000000005688: E07E1000 8004880B
-	v_cvt_pk_bf16_f32 v144, v144, v145                         // 000000005690: D2680090 00032390
-	v_cvt_pk_bf16_f32 v145, v146, v147                         // 000000005698: D2680091 00032792
-	v_cvt_pk_bf16_f32 v146, v148, v149                         // 0000000056A0: D2680092 00032B94
-	v_cvt_pk_bf16_f32 v147, v150, v151                         // 0000000056A8: D2680093 00032F96
-	s_lshl_b32 s12, s36, 1                                     // 0000000056B0: 8E0C8124
-	s_add_u32 s16, s16, s12                                    // 0000000056B4: 80100C10
-	s_addc_u32 s17, s17, 0                                     // 0000000056B8: 82118011
-	buffer_store_dwordx4 v[144:147], v11, s[16:19], 0 offen nt // 0000000056BC: E07E1000 8004900B
-	v_cvt_pk_bf16_f32 v152, v152, v153                         // 0000000056C4: D2680098 00033398
-	v_cvt_pk_bf16_f32 v153, v154, v155                         // 0000000056CC: D2680099 0003379A
-	v_cvt_pk_bf16_f32 v154, v156, v157                         // 0000000056D4: D268009A 00033B9C
-	v_cvt_pk_bf16_f32 v155, v158, v159                         // 0000000056DC: D268009B 00033F9E
-	s_lshl_b32 s12, s36, 1                                     // 0000000056E4: 8E0C8124
-	s_add_u32 s16, s16, s12                                    // 0000000056E8: 80100C10
-	s_addc_u32 s17, s17, 0                                     // 0000000056EC: 82118011
-	buffer_store_dwordx4 v[152:155], v11, s[16:19], 0 offen nt // 0000000056F0: E07E1000 8004980B
-	v_cvt_pk_bf16_f32 v160, v160, v161                         // 0000000056F8: D26800A0 000343A0
-	v_cvt_pk_bf16_f32 v161, v162, v163                         // 000000005700: D26800A1 000347A2
-	v_cvt_pk_bf16_f32 v162, v164, v165                         // 000000005708: D26800A2 00034BA4
-	v_cvt_pk_bf16_f32 v163, v166, v167                         // 000000005710: D26800A3 00034FA6
-	s_lshl_b32 s12, s36, 1                                     // 000000005718: 8E0C8124
-	s_add_u32 s16, s16, s12                                    // 00000000571C: 80100C10
-	s_addc_u32 s17, s17, 0                                     // 000000005720: 82118011
-	buffer_store_dwordx4 v[160:163], v11, s[16:19], 0 offen nt // 000000005724: E07E1000 8004A00B
-	v_cvt_pk_bf16_f32 v168, v168, v169                         // 00000000572C: D26800A8 000353A8
-	v_cvt_pk_bf16_f32 v169, v170, v171                         // 000000005734: D26800A9 000357AA
-	v_cvt_pk_bf16_f32 v170, v172, v173                         // 00000000573C: D26800AA 00035BAC
-	v_cvt_pk_bf16_f32 v171, v174, v175                         // 000000005744: D26800AB 00035FAE
-	s_lshl_b32 s12, s36, 1                                     // 00000000574C: 8E0C8124
-	s_add_u32 s16, s16, s12                                    // 000000005750: 80100C10
-	s_addc_u32 s17, s17, 0                                     // 000000005754: 82118011
-	buffer_store_dwordx4 v[168:171], v11, s[16:19], 0 offen nt // 000000005758: E07E1000 8004A80B
-	v_cvt_pk_bf16_f32 v176, v176, v177                         // 000000005760: D26800B0 000363B0
-	v_cvt_pk_bf16_f32 v177, v178, v179                         // 000000005768: D26800B1 000367B2
-	v_cvt_pk_bf16_f32 v178, v180, v181                         // 000000005770: D26800B2 00036BB4
-	v_cvt_pk_bf16_f32 v179, v182, v183                         // 000000005778: D26800B3 00036FB6
-	s_lshl_b32 s12, s36, 1                                     // 000000005780: 8E0C8124
-	s_add_u32 s16, s16, s12                                    // 000000005784: 80100C10
-	s_addc_u32 s17, s17, 0                                     // 000000005788: 82118011
-	buffer_store_dwordx4 v[176:179], v11, s[16:19], 0 offen nt // 00000000578C: E07E1000 8004B00B
-	v_cvt_pk_bf16_f32 v184, v184, v185                         // 000000005794: D26800B8 000373B8
-	v_cvt_pk_bf16_f32 v185, v186, v187                         // 00000000579C: D26800B9 000377BA
-	v_cvt_pk_bf16_f32 v186, v188, v189                         // 0000000057A4: D26800BA 00037BBC
-	v_cvt_pk_bf16_f32 v187, v190, v191                         // 0000000057AC: D26800BB 00037FBE
-	s_lshl_b32 s12, s36, 1                                     // 0000000057B4: 8E0C8124
-	s_add_u32 s16, s16, s12                                    // 0000000057B8: 80100C10
-	s_addc_u32 s17, s17, 0                                     // 0000000057BC: 82118011
-	buffer_store_dwordx4 v[184:187], v11, s[16:19], 0 offen nt // 0000000057C0: E07E1000 8004B80B
-	v_cvt_pk_bf16_f32 v192, v192, v193                         // 0000000057C8: D26800C0 000383C0
-	v_cvt_pk_bf16_f32 v193, v194, v195                         // 0000000057D0: D26800C1 000387C2
-	v_cvt_pk_bf16_f32 v194, v196, v197                         // 0000000057D8: D26800C2 00038BC4
-	v_cvt_pk_bf16_f32 v195, v198, v199                         // 0000000057E0: D26800C3 00038FC6
-	s_lshl_b32 s12, s36, 1                                     // 0000000057E8: 8E0C8124
-	s_add_u32 s16, s16, s12                                    // 0000000057EC: 80100C10
-	s_addc_u32 s17, s17, 0                                     // 0000000057F0: 82118011
-	buffer_store_dwordx4 v[192:195], v11, s[16:19], 0 offen nt // 0000000057F4: E07E1000 8004C00B
-	v_cvt_pk_bf16_f32 v200, v200, v201                         // 0000000057FC: D26800C8 000393C8
-	v_cvt_pk_bf16_f32 v201, v202, v203                         // 000000005804: D26800C9 000397CA
-	v_cvt_pk_bf16_f32 v202, v204, v205                         // 00000000580C: D26800CA 00039BCC
-	v_cvt_pk_bf16_f32 v203, v206, v207                         // 000000005814: D26800CB 00039FCE
-	s_lshl_b32 s12, s36, 1                                     // 00000000581C: 8E0C8124
-	s_add_u32 s16, s16, s12                                    // 000000005820: 80100C10
-	s_addc_u32 s17, s17, 0                                     // 000000005824: 82118011
-	buffer_store_dwordx4 v[200:203], v11, s[16:19], 0 offen nt // 000000005828: E07E1000 8004C80B
-	v_cvt_pk_bf16_f32 v208, v208, v209                         // 000000005830: D26800D0 0003A3D0
-	v_cvt_pk_bf16_f32 v209, v210, v211                         // 000000005838: D26800D1 0003A7D2
-	v_cvt_pk_bf16_f32 v210, v212, v213                         // 000000005840: D26800D2 0003ABD4
-	v_cvt_pk_bf16_f32 v211, v214, v215                         // 000000005848: D26800D3 0003AFD6
-	s_lshl_b32 s12, s36, 1                                     // 000000005850: 8E0C8124
-	s_add_u32 s16, s16, s12                                    // 000000005854: 80100C10
-	s_addc_u32 s17, s17, 0                                     // 000000005858: 82118011
-	buffer_store_dwordx4 v[208:211], v11, s[16:19], 0 offen nt // 00000000585C: E07E1000 8004D00B
-	v_cvt_pk_bf16_f32 v216, v216, v217                         // 000000005864: D26800D8 0003B3D8
-	v_cvt_pk_bf16_f32 v217, v218, v219                         // 00000000586C: D26800D9 0003B7DA
-	v_cvt_pk_bf16_f32 v218, v220, v221                         // 000000005874: D26800DA 0003BBDC
-	v_cvt_pk_bf16_f32 v219, v222, v223                         // 00000000587C: D26800DB 0003BFDE
-	s_lshl_b32 s12, s36, 1                                     // 000000005884: 8E0C8124
-	s_add_u32 s16, s16, s12                                    // 000000005888: 80100C10
-	s_addc_u32 s17, s17, 0                                     // 00000000588C: 82118011
-	buffer_store_dwordx4 v[216:219], v11, s[16:19], 0 offen nt // 000000005890: E07E1000 8004D80B
-	v_cvt_pk_bf16_f32 v224, v224, v225                         // 000000005898: D26800E0 0003C3E0
-	v_cvt_pk_bf16_f32 v225, v226, v227                         // 0000000058A0: D26800E1 0003C7E2
-	v_cvt_pk_bf16_f32 v226, v228, v229                         // 0000000058A8: D26800E2 0003CBE4
-	v_cvt_pk_bf16_f32 v227, v230, v231                         // 0000000058B0: D26800E3 0003CFE6
-	s_lshl_b32 s12, s36, 1                                     // 0000000058B8: 8E0C8124
-	s_add_u32 s16, s16, s12                                    // 0000000058BC: 80100C10
-	s_addc_u32 s17, s17, 0                                     // 0000000058C0: 82118011
-	buffer_store_dwordx4 v[224:227], v11, s[16:19], 0 offen nt // 0000000058C4: E07E1000 8004E00B
-	v_cvt_pk_bf16_f32 v232, v232, v233                         // 0000000058CC: D26800E8 0003D3E8
-	v_cvt_pk_bf16_f32 v233, v234, v235                         // 0000000058D4: D26800E9 0003D7EA
-	v_cvt_pk_bf16_f32 v234, v236, v237                         // 0000000058DC: D26800EA 0003DBEC
-	v_cvt_pk_bf16_f32 v235, v238, v239                         // 0000000058E4: D26800EB 0003DFEE
-	s_lshl_b32 s12, s36, 1                                     // 0000000058EC: 8E0C8124
-	s_add_u32 s16, s16, s12                                    // 0000000058F0: 80100C10
-	s_addc_u32 s17, s17, 0                                     // 0000000058F4: 82118011
-	buffer_store_dwordx4 v[232:235], v11, s[16:19], 0 offen nt // 0000000058F8: E07E1000 8004E80B
-	v_cvt_pk_bf16_f32 v240, v240, v241                         // 000000005900: D26800F0 0003E3F0
-	v_cvt_pk_bf16_f32 v241, v242, v243                         // 000000005908: D26800F1 0003E7F2
-	v_cvt_pk_bf16_f32 v242, v244, v245                         // 000000005910: D26800F2 0003EBF4
-	v_cvt_pk_bf16_f32 v243, v246, v247                         // 000000005918: D26800F3 0003EFF6
-	s_lshl_b32 s12, s36, 1                                     // 000000005920: 8E0C8124
-	s_add_u32 s16, s16, s12                                    // 000000005924: 80100C10
-	s_addc_u32 s17, s17, 0                                     // 000000005928: 82118011
-	buffer_store_dwordx4 v[240:243], v11, s[16:19], 0 offen nt // 00000000592C: E07E1000 8004F00B
-	s_nop 0                                                    // 000000005934: BF800000
-	v_accvgpr_read_b32 v16, a131                               // 000000005938: D3D84010 18000183
-	v_accvgpr_read_b32 v17, a135                               // 000000005940: D3D84011 18000187
-	v_accvgpr_read_b32 v18, a139                               // 000000005948: D3D84012 1800018B
-	v_accvgpr_read_b32 v19, a143                               // 000000005950: D3D84013 1800018F
-	v_accvgpr_read_b32 v20, a147                               // 000000005958: D3D84014 18000193
-	v_accvgpr_read_b32 v21, a151                               // 000000005960: D3D84015 18000197
-	v_accvgpr_read_b32 v22, a155                               // 000000005968: D3D84016 1800019B
-	v_accvgpr_read_b32 v23, a159                               // 000000005970: D3D84017 1800019F
-	v_accvgpr_read_b32 v24, a163                               // 000000005978: D3D84018 180001A3
-	v_accvgpr_read_b32 v25, a167                               // 000000005980: D3D84019 180001A7
-	v_accvgpr_read_b32 v26, a171                               // 000000005988: D3D8401A 180001AB
-	v_accvgpr_read_b32 v27, a175                               // 000000005990: D3D8401B 180001AF
-	v_accvgpr_read_b32 v28, a179                               // 000000005998: D3D8401C 180001B3
-	v_accvgpr_read_b32 v29, a183                               // 0000000059A0: D3D8401D 180001B7
-	v_accvgpr_read_b32 v30, a187                               // 0000000059A8: D3D8401E 180001BB
-	v_accvgpr_read_b32 v31, a191                               // 0000000059B0: D3D8401F 180001BF
-	v_accvgpr_read_b32 v32, a195                               // 0000000059B8: D3D84020 180001C3
-	v_accvgpr_read_b32 v33, a199                               // 0000000059C0: D3D84021 180001C7
-	v_accvgpr_read_b32 v34, a203                               // 0000000059C8: D3D84022 180001CB
-	v_accvgpr_read_b32 v35, a207                               // 0000000059D0: D3D84023 180001CF
-	v_accvgpr_read_b32 v36, a211                               // 0000000059D8: D3D84024 180001D3
-	v_accvgpr_read_b32 v37, a215                               // 0000000059E0: D3D84025 180001D7
-	v_accvgpr_read_b32 v38, a219                               // 0000000059E8: D3D84026 180001DB
-	v_accvgpr_read_b32 v39, a223                               // 0000000059F0: D3D84027 180001DF
-	v_accvgpr_read_b32 v40, a227                               // 0000000059F8: D3D84028 180001E3
-	v_accvgpr_read_b32 v41, a231                               // 000000005A00: D3D84029 180001E7
-	v_accvgpr_read_b32 v42, a235                               // 000000005A08: D3D8402A 180001EB
-	v_accvgpr_read_b32 v43, a239                               // 000000005A10: D3D8402B 180001EF
-	v_accvgpr_read_b32 v44, a243                               // 000000005A18: D3D8402C 180001F3
-	v_accvgpr_read_b32 v45, a247                               // 000000005A20: D3D8402D 180001F7
-	v_accvgpr_read_b32 v46, a251                               // 000000005A28: D3D8402E 180001FB
-	v_accvgpr_read_b32 v47, a255                               // 000000005A30: D3D8402F 180001FF
-	v_mov_b32_e32 v8, 0xffff0000                               // 000000005A38: 7E1002FF FFFF0000
-	v_mov_b32_e32 v9, 0x7fff0000                               // 000000005A40: 7E1202FF 7FFF0000
-	v_mov_b32_e32 v10, 0x7fff                                  // 000000005A48: 7E1402FF 00007FFF
-	v_cvt_pk_bf16_f32 v16, v16, v17                            // 000000005A50: D2680010 00022310
-	v_cvt_pk_bf16_f32 v17, v18, v19                            // 000000005A58: D2680011 00022712
-	v_cvt_pk_bf16_f32 v18, v20, v21                            // 000000005A60: D2680012 00022B14
-	v_cvt_pk_bf16_f32 v19, v22, v23                            // 000000005A68: D2680013 00022F16
-	s_lshl_b32 s12, s36, 1                                     // 000000005A70: 8E0C8124
-	s_add_u32 s16, s16, s12                                    // 000000005A74: 80100C10
-	s_addc_u32 s17, s17, 0                                     // 000000005A78: 82118011
-	buffer_store_dwordx4 v[16:19], v11, s[16:19], 0 offen nt   // 000000005A7C: E07E1000 8004100B
-	v_cvt_pk_bf16_f32 v24, v24, v25                            // 000000005A84: D2680018 00023318
-	v_cvt_pk_bf16_f32 v25, v26, v27                            // 000000005A8C: D2680019 0002371A
-	v_cvt_pk_bf16_f32 v26, v28, v29                            // 000000005A94: D268001A 00023B1C
-	v_cvt_pk_bf16_f32 v27, v30, v31                            // 000000005A9C: D268001B 00023F1E
-	s_lshl_b32 s12, s36, 1                                     // 000000005AA4: 8E0C8124
-	s_add_u32 s16, s16, s12                                    // 000000005AA8: 80100C10
-	s_addc_u32 s17, s17, 0                                     // 000000005AAC: 82118011
-	buffer_store_dwordx4 v[24:27], v11, s[16:19], 0 offen nt   // 000000005AB0: E07E1000 8004180B
-	v_cvt_pk_bf16_f32 v32, v32, v33                            // 000000005AB8: D2680020 00024320
-	v_cvt_pk_bf16_f32 v33, v34, v35                            // 000000005AC0: D2680021 00024722
-	v_cvt_pk_bf16_f32 v34, v36, v37                            // 000000005AC8: D2680022 00024B24
-	v_cvt_pk_bf16_f32 v35, v38, v39                            // 000000005AD0: D2680023 00024F26
-	s_lshl_b32 s12, s36, 1                                     // 000000005AD8: 8E0C8124
-	s_add_u32 s16, s16, s12                                    // 000000005ADC: 80100C10
-	s_addc_u32 s17, s17, 0                                     // 000000005AE0: 82118011
-	buffer_store_dwordx4 v[32:35], v11, s[16:19], 0 offen nt   // 000000005AE4: E07E1000 8004200B
-	v_cvt_pk_bf16_f32 v40, v40, v41                            // 000000005AEC: D2680028 00025328
-	v_cvt_pk_bf16_f32 v41, v42, v43                            // 000000005AF4: D2680029 0002572A
-	v_cvt_pk_bf16_f32 v42, v44, v45                            // 000000005AFC: D268002A 00025B2C
-	v_cvt_pk_bf16_f32 v43, v46, v47                            // 000000005B04: D268002B 00025F2E
-	s_lshl_b32 s12, s36, 1                                     // 000000005B0C: 8E0C8124
-	s_add_u32 s16, s16, s12                                    // 000000005B10: 80100C10
-	s_addc_u32 s17, s17, 0                                     // 000000005B14: 82118011
-	buffer_store_dwordx4 v[40:43], v11, s[16:19], 0 offen nt   // 000000005B18: E07E1000 8004280B
-	s_nop 0                                                    // 000000005B20: BF800000
-	s_endpgm                                                   // 00000001F5D0: BF810000
diff --git a/extra/gemm/asm/cdna/template.s b/extra/gemm/asm/cdna/template.s
deleted file mode 100644
index 9ba66ed9a9..0000000000
--- a/extra/gemm/asm/cdna/template.s
+++ /dev/null
@@ -1,78 +0,0 @@
-.text
-.section	.text.
-.global	gemm
-.p2align	8
-.type	gemm,@function
-
-gemm:
-INSTRUCTIONS
-
-.section .rodata,"a",@progbits
-.p2align 6, 0x0
-.amdhsa_kernel gemm
-  # basic memory requirements
-  .amdhsa_group_segment_fixed_size 133120
-  .amdhsa_private_segment_fixed_size 0
-  .amdhsa_kernarg_size 28
-  # register usage (RSRC1)
-  .amdhsa_next_free_vgpr 504
-  .amdhsa_next_free_sgpr 96
-  # workgroup / workitem IDs (RSRC2)
-  .amdhsa_system_sgpr_workgroup_id_x 1
-  .amdhsa_system_sgpr_workgroup_id_y 1
-  .amdhsa_system_sgpr_workgroup_id_z 1
-  # user SGPRs, we only specify the kernel args ptr in s[0:1]
-  .amdhsa_user_sgpr_kernarg_segment_ptr 1
-  .amdhsa_user_sgpr_count 2
-  .amdhsa_user_sgpr_kernarg_preload_length 0
-  .amdhsa_user_sgpr_kernarg_preload_offset 0
-  # gfx90a / gfx940 specifics (RSRC3)
-  .amdhsa_accum_offset 248
-  .amdhsa_uses_dynamic_stack 0
-  .amdhsa_tg_split 0
-.end_amdhsa_kernel
-
-.amdgpu_metadata
----
-amdhsa.kernels:
-  - .name: gemm
-    .symbol: gemm.kd
-    .args:
-      - .name: C
-        .address_space: global
-        .offset: 0
-        .size: 8
-        .value_kind: global_buffer
-        .value_type: bf16
-      - .name: B
-        .address_space: global
-        .offset: 8
-        .size: 8
-        .value_kind: global_buffer
-        .value_type: bf16
-      - .name: A
-        .address_space: global
-        .offset: 16
-        .size: 8
-        .value_kind: global_buffer
-        .value_type: bf16
-      - .name: sz
-        .offset: 24
-        .size: 4
-        .value_kind: by_value
-        .value_type: u32
-    .group_segment_fixed_size: 133120
-    .private_segment_fixed_size: 0
-    .kernarg_segment_align: 8
-    .kernarg_segment_size: 28
-    .max_flat_workgroup_size: 256
-    .sgpr_count: 88
-    .sgpr_spill_count: 0
-    .vgpr_count: 248
-    .vgpr_spill_count: 0
-    .wavefront_size: 64
-amdhsa.version:
-  - 1
-  - 0
-...
-.end_amdgpu_metadata
diff --git a/extra/gemm/asm/cdna/test.py b/extra/gemm/asm/cdna/test.py
deleted file mode 100644
index 1370f5879b..0000000000
--- a/extra/gemm/asm/cdna/test.py
+++ /dev/null
@@ -1,73 +0,0 @@
-# Run assembly on the AMD runtime and check correctness
-# VIZ=2 to profile
-import pathlib
-from tinygrad import Tensor, Device, dtypes, Context
-from tinygrad.uop.ops import UOp, Ops, KernelInfo
-from tinygrad.engine.realize import Estimates
-from tinygrad.helpers import getenv
-
-fp = pathlib.Path(__file__).parent/"gemm.s"
-
-N = getenv("N", 8192)
-THREADS_PER_WG = 256
-NUM_WG = N//THREADS_PER_WG * N//THREADS_PER_WG
-
-assert N % THREADS_PER_WG == 0, "N must be divisible by THREADS_PER_WG"
-
-# ** generate inputs on CPU
-
-scale = 10.0
-
-import torch
-torch.manual_seed(0)
-A = (torch.randn(N, N, dtype=torch.float32, device="cpu") / scale).to(torch.bfloat16).contiguous()
-B = (torch.randn(N, N, dtype=torch.float32, device="cpu") / scale).to(torch.bfloat16).contiguous()
-Bt = B.t().contiguous() # transpose B for the asm gemm
-C_torch = A@B
-
-# ** copy buffers to AMD
-
-# input creation and validation run on the copy engine for simpler tracing
-
-def from_torch(t:torch.Tensor) -> Tensor:
-  return Tensor.from_blob(t.data_ptr(), t.shape, dtype=dtypes.bfloat16, device="cpu").to(Device.DEFAULT).realize()
-
-C_tiny = from_torch(A) @ from_torch(B)
-C_asm = Tensor.empty_like(C_tiny)
-
-# ** assembly custom kernel
-
-def custom_asm_gemm(C:UOp, A:UOp, B:UOp) -> UOp:
-  lidx = UOp.special(THREADS_PER_WG, "lidx0")
-  gidx = UOp.special(NUM_WG, "gidx0")
-
-  src = (pathlib.Path(__file__).parent/"template.s").read_text().replace("INSTRUCTIONS", fp.read_text())
-
-  sz = UOp.variable("SZ", 256, 8192)
-
-  sink = UOp.sink(C.base, A.base, B.base, sz, lidx, gidx, arg=KernelInfo(name="gemm", estimates=Estimates(ops=N*N*N*2, mem=N*N*4*3)))
-  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=Device.DEFAULT), UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=src)))
-
-C_asm = Tensor.custom_kernel(C_asm, from_torch(A), from_torch(Bt), fxn=custom_asm_gemm)[0]
-
-# ** run gemms
-
-sched = Tensor.schedule(C_tiny, C_asm)
-eis = [si.lower() for si in sched]
-
-with Context(DEBUG=2):
-  for ei in eis:
-    et = ei.run({"SZ":N}, wait=True)
-    print(f"{(N*N*N*2 / et)*1e-12:.2f} REAL TFLOPS")
-
-# ** correctness
-
-import ctypes
-
-def torch_bf16(t:Tensor) -> torch.tensor:
-  asm_out = t.to("cpu").realize().uop.buffer._buf
-  buf = (ctypes.c_uint16*C_asm.uop.size).from_address(asm_out.va_addr)
-  return torch.frombuffer(buf, dtype=torch.bfloat16, count=C_asm.uop.size).reshape(C_asm.shape)
-
-assert torch.allclose(torch_bf16(C_asm), C_torch, rtol=1e-2, atol=1e-3)
-assert torch.allclose(torch_bf16(C_tiny), C_torch, rtol=1e-2, atol=1e-3)
diff --git a/extra/gemm/asm/cdna/test_asm_gemm.py b/extra/gemm/asm/cdna/test_asm_gemm.py
new file mode 100644
index 0000000000..905492cc60
--- /dev/null
+++ b/extra/gemm/asm/cdna/test_asm_gemm.py
@@ -0,0 +1,46 @@
+import unittest
+from tinygrad import Tensor, Device, dtypes, Context
+from tinygrad.helpers import getenv
+from extra.gemm.asm.cdna.gemm import asm_gemm
+
+def verify_asm_gemm(batch:int, M:int, N:int, K:int, dtype=dtypes.bfloat16, multi=False) -> None:
+  Tensor.manual_seed(0)
+  a_rand = Tensor.randn((batch, M, K), dtype=dtypes.float).sub(0.5).cast(dtype)
+  b_rand = Tensor.randn((K, N), dtype=dtypes.float).sub(0.5).cast(dtype)
+  with Context(DEBUG=0):
+    Tensor.realize(a_rand, b_rand)
+
+  devs = tuple(f"{Device.DEFAULT}:{i}" for i in range(8)) if multi else None
+
+  a, b = Tensor(a_rand.numpy(), requires_grad=True).cast(dtype), Tensor(b_rand.numpy(), requires_grad=True).cast(dtype)
+  if multi: a, b = a.shard(devs, axis=0), b.shard(devs, axis=None)
+  tst = asm_gemm(a, b)
+  tst.sum().backward()
+  Tensor.realize(tst, a.grad, b.grad)
+
+  a_ref, b_ref = Tensor(a_rand.numpy(), requires_grad=True).cast(dtype), Tensor(b_rand.numpy(), requires_grad=True).cast(dtype)
+  if multi: a_ref, b_ref = a_ref.shard(devs, axis=0), b_ref.shard(devs, axis=None)
+  with Context(ASM_GEMM=0): ref = a_ref @ b_ref
+  ref.sum().backward()
+  Tensor.realize(ref, a_ref.grad, b_ref.grad)
+
+  with Context(DEBUG=0):
+    assert (tst - ref).square().max().float().item() < 1e-6, "forward mismatch"
+    assert (a.grad - a_ref.grad).square().max().float().item() < 1e-3, "grad_a mismatch"
+    assert (b.grad - b_ref.grad).square().max().float().item() < 1e-3, "grad_b mismatch"
+
+class TestGemm(unittest.TestCase):
+  def test_simple(self): verify_asm_gemm(1, N:=getenv("N", 4096), N, N, dtype=dtypes.half)
+
+  def test_gemm1(self): verify_asm_gemm(8, 8192, 4096, 14336, multi=True)
+  def test_gemm2(self): verify_asm_gemm(8, 8192, 128256, 4096, multi=True)
+  def test_gemm3(self): verify_asm_gemm(8, 8192, 14336, 4096, multi=True)
+  def test_gemm4(self): verify_asm_gemm(8, 4096, 14336, 4096, multi=True)
+  def test_gemm5(self): verify_asm_gemm(8, 4096, 4096, 14336, multi=True)
+  def test_gemm6(self): verify_asm_gemm(16, 4096, 4096, 14336, multi=True)
+  def test_gemm_unsupported(self):
+    with self.assertRaisesRegex(AssertionError, "shape not supported"):
+      verify_asm_gemm(8, 8192, 1024, 4096, multi=True)
+
+if __name__ == "__main__":
+  unittest.main()
diff --git a/extra/sqtt/examples/gfx1100/profile_empty_run_0.pkl b/extra/sqtt/examples/gfx1100/profile_empty_run_0.pkl
index 22fac0f23c..9ae7fbc50a 100644
Binary files a/extra/sqtt/examples/gfx1100/profile_empty_run_0.pkl and b/extra/sqtt/examples/gfx1100/profile_empty_run_0.pkl differ
diff --git a/extra/sqtt/examples/gfx1100/profile_empty_run_1.pkl b/extra/sqtt/examples/gfx1100/profile_empty_run_1.pkl
index 2c980df292..2c57cb5d8d 100644
Binary files a/extra/sqtt/examples/gfx1100/profile_empty_run_1.pkl and b/extra/sqtt/examples/gfx1100/profile_empty_run_1.pkl differ
diff --git a/extra/sqtt/examples/gfx1100/profile_gemm_run_0.pkl b/extra/sqtt/examples/gfx1100/profile_gemm_run_0.pkl
index e540cadc0f..ccf477c653 100644
Binary files a/extra/sqtt/examples/gfx1100/profile_gemm_run_0.pkl and b/extra/sqtt/examples/gfx1100/profile_gemm_run_0.pkl differ
diff --git a/extra/sqtt/examples/gfx1100/profile_gemm_run_1.pkl b/extra/sqtt/examples/gfx1100/profile_gemm_run_1.pkl
index 693f79f38e..be0cedc424 100644
Binary files a/extra/sqtt/examples/gfx1100/profile_gemm_run_1.pkl and b/extra/sqtt/examples/gfx1100/profile_gemm_run_1.pkl differ
diff --git a/extra/sqtt/examples/gfx1100/profile_plus_run_0.pkl b/extra/sqtt/examples/gfx1100/profile_plus_run_0.pkl
index 5416372220..1a0e2243a7 100644
Binary files a/extra/sqtt/examples/gfx1100/profile_plus_run_0.pkl and b/extra/sqtt/examples/gfx1100/profile_plus_run_0.pkl differ
diff --git a/extra/sqtt/examples/gfx1100/profile_plus_run_1.pkl b/extra/sqtt/examples/gfx1100/profile_plus_run_1.pkl
index 51aae2dbff..e5f6376fcc 100644
Binary files a/extra/sqtt/examples/gfx1100/profile_plus_run_1.pkl and b/extra/sqtt/examples/gfx1100/profile_plus_run_1.pkl differ
diff --git a/extra/sqtt/examples/gfx1200/profile_empty_run_0.pkl b/extra/sqtt/examples/gfx1200/profile_empty_run_0.pkl
index 99131b49e4..62940f0928 100644
Binary files a/extra/sqtt/examples/gfx1200/profile_empty_run_0.pkl and b/extra/sqtt/examples/gfx1200/profile_empty_run_0.pkl differ
diff --git a/extra/sqtt/examples/gfx1200/profile_empty_run_1.pkl b/extra/sqtt/examples/gfx1200/profile_empty_run_1.pkl
index 6a9ae97e85..01a9ddac0a 100644
Binary files a/extra/sqtt/examples/gfx1200/profile_empty_run_1.pkl and b/extra/sqtt/examples/gfx1200/profile_empty_run_1.pkl differ
diff --git a/extra/sqtt/examples/gfx1200/profile_gemm_run_0.pkl b/extra/sqtt/examples/gfx1200/profile_gemm_run_0.pkl
index c726b9581f..043c730e80 100644
Binary files a/extra/sqtt/examples/gfx1200/profile_gemm_run_0.pkl and b/extra/sqtt/examples/gfx1200/profile_gemm_run_0.pkl differ
diff --git a/extra/sqtt/examples/gfx1200/profile_gemm_run_1.pkl b/extra/sqtt/examples/gfx1200/profile_gemm_run_1.pkl
index a55bad145b..65f23de629 100644
Binary files a/extra/sqtt/examples/gfx1200/profile_gemm_run_1.pkl and b/extra/sqtt/examples/gfx1200/profile_gemm_run_1.pkl differ
diff --git a/extra/sqtt/examples/gfx1200/profile_plus_run_0.pkl b/extra/sqtt/examples/gfx1200/profile_plus_run_0.pkl
index 81c064609b..9d1204ca0b 100644
Binary files a/extra/sqtt/examples/gfx1200/profile_plus_run_0.pkl and b/extra/sqtt/examples/gfx1200/profile_plus_run_0.pkl differ
diff --git a/extra/sqtt/examples/gfx1200/profile_plus_run_1.pkl b/extra/sqtt/examples/gfx1200/profile_plus_run_1.pkl
index 032e0c54f0..4c1bc06cd9 100644
Binary files a/extra/sqtt/examples/gfx1200/profile_plus_run_1.pkl and b/extra/sqtt/examples/gfx1200/profile_plus_run_1.pkl differ
diff --git a/extra/sqtt/examples/gfx950/profile_empty_run_0.pkl b/extra/sqtt/examples/gfx950/profile_empty_run_0.pkl
index 5dfec279db..8ad5af2294 100644
Binary files a/extra/sqtt/examples/gfx950/profile_empty_run_0.pkl and b/extra/sqtt/examples/gfx950/profile_empty_run_0.pkl differ
diff --git a/extra/sqtt/examples/gfx950/profile_empty_run_1.pkl b/extra/sqtt/examples/gfx950/profile_empty_run_1.pkl
index 1c0ef9543a..3b94e90698 100644
Binary files a/extra/sqtt/examples/gfx950/profile_empty_run_1.pkl and b/extra/sqtt/examples/gfx950/profile_empty_run_1.pkl differ
diff --git a/extra/sqtt/examples/gfx950/profile_gemm_run_0.pkl b/extra/sqtt/examples/gfx950/profile_gemm_run_0.pkl
index 6431be7b5a..6ae6664043 100644
Binary files a/extra/sqtt/examples/gfx950/profile_gemm_run_0.pkl and b/extra/sqtt/examples/gfx950/profile_gemm_run_0.pkl differ
diff --git a/extra/sqtt/examples/gfx950/profile_gemm_run_1.pkl b/extra/sqtt/examples/gfx950/profile_gemm_run_1.pkl
index 1526352ff7..b0685edd2f 100644
Binary files a/extra/sqtt/examples/gfx950/profile_gemm_run_1.pkl and b/extra/sqtt/examples/gfx950/profile_gemm_run_1.pkl differ
diff --git a/extra/sqtt/examples/gfx950/profile_plus_run_0.pkl b/extra/sqtt/examples/gfx950/profile_plus_run_0.pkl
index c6e491de0d..c816dc509b 100644
Binary files a/extra/sqtt/examples/gfx950/profile_plus_run_0.pkl and b/extra/sqtt/examples/gfx950/profile_plus_run_0.pkl differ
diff --git a/extra/sqtt/examples/gfx950/profile_plus_run_1.pkl b/extra/sqtt/examples/gfx950/profile_plus_run_1.pkl
index a365766eb7..010125fb66 100644
Binary files a/extra/sqtt/examples/gfx950/profile_plus_run_1.pkl and b/extra/sqtt/examples/gfx950/profile_plus_run_1.pkl differ
diff --git a/extra/usbgpu/patch.py b/extra/usbgpu/patch.py
index 48c167dd11..61ffcb0420 100755
--- a/extra/usbgpu/patch.py
+++ b/extra/usbgpu/patch.py
@@ -1,6 +1,6 @@
 #!/usr/bin/env python3
-import sys, os, zlib, struct, hashlib
-from tinygrad.helpers import DEBUG, getenv, fetch
+import os, zlib, struct, hashlib
+from tinygrad.helpers import getenv
 from tinygrad.runtime.support.usb import USB3
 
 SUPPORTED_CONTROLLERS = [
@@ -50,7 +50,7 @@ patched_fw = patch(file_path, file_hash, patches)
 dev = None
 for vendor, device in SUPPORTED_CONTROLLERS:
   try:
-    dev = USB3(vendor, device, 0x81, 0x83, 0x02, 0x04)
+    dev = USB3(vendor, device, 0x81, 0x83, 0x02, 0x04, use_bot=True)
     break
   except RuntimeError: pass
 if dev is None:
diff --git a/test/mockgpu/amd/amddriver.py b/test/mockgpu/amd/amddriver.py
index 0e498f1d05..b5801f823c 100644
--- a/test/mockgpu/amd/amddriver.py
+++ b/test/mockgpu/amd/amddriver.py
@@ -4,7 +4,7 @@ import tinygrad.runtime.autogen.am.am as am
 import tinygrad.runtime.autogen.amdgpu_drm as amdgpu_drm
 from tinygrad.helpers import from_mv
 from test.mockgpu.driver import VirtDriver, VirtFileDesc, TextFileDesc, DirFileDesc, VirtFile
-from test.mockgpu.amd.amdgpu import AMDGPU, gpu_props
+from test.mockgpu.amd.amdgpu import AMDGPU, gpu_props, GFX_TARGET_VERSION, MOCKGPU_ARCH
 
 libc = ctypes.CDLL(ctypes.util.find_library("c"))
 libc.mmap.argtypes = [ctypes.c_void_p, ctypes.c_size_t, ctypes.c_int, ctypes.c_int, ctypes.c_int, ctypes.c_long]
@@ -90,35 +90,30 @@ class AMDDriver(VirtDriver):
   def _prepare_gpu(self, gpu_id):
     self.doorbells[gpu_id] = memoryview(bytearray(0x2000))
     self.gpus[gpu_id] = AMDGPU(gpu_id)
+    # IP versions: rdna3 = GC 11.0.0, NBIF 4.3.0; rdna4 = GC 12.0.0, NBIF 6.3.1
+    ip_versions = {"rdna3": {"gc": (11, 0, 0), "sdma": (6, 0, 0), "nbif": (4, 3, 0)},
+                   "rdna4": {"gc": (12, 0, 0), "sdma": (6, 0, 0), "nbif": (6, 3, 1)}}[MOCKGPU_ARCH]
+    def ip_discovery_files(hwid, ver, base_addr):
+      p = f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{hwid}/0'
+      return [VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{hwid}', functools.partial(DirFileDesc, child_names=['0'])),
+              VirtFile(f'{p}/major', functools.partial(TextFileDesc, text=str(ver[0]))),
+              VirtFile(f'{p}/minor', functools.partial(TextFileDesc, text=str(ver[1]))),
+              VirtFile(f'{p}/revision', functools.partial(TextFileDesc, text=str(ver[2]))),
+              VirtFile(f'{p}/base_addr', functools.partial(TextFileDesc, text=base_addr))]
     self.tracked_files += [
       VirtFile('/sys/module/amdgpu', functools.partial(TextFileDesc, text="1")),
       VirtFile('/sys/module/amdgpu/parameters/ppfeaturemask', functools.partial(TextFileDesc, text="0xffff3fff")),
       VirtFile(f'/sys/devices/virtual/kfd/kfd/topology/nodes/{gpu_id}', functools.partial(DirFileDesc, child_names=['gpu_id', 'properties'])),
       VirtFile(f'/sys/devices/virtual/kfd/kfd/topology/nodes/{gpu_id}/gpu_id', functools.partial(TextFileDesc, text=f"{gpu_id}")),
       VirtFile(f'/sys/devices/virtual/kfd/kfd/topology/nodes/{gpu_id}/properties',
-        functools.partial(TextFileDesc, text=gpu_props.format(drm_render_minor=gpu_id))),
+        functools.partial(TextFileDesc, text=gpu_props.format(drm_render_minor=gpu_id, gfx_target_version=GFX_TARGET_VERSION))),
       VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/power_dpm_force_performance_level',
                functools.partial(TextFileDesc, text='profile_standard\n')),
       VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0',
                functools.partial(DirFileDesc, child_names=[str(am.GC_HWID), str(am.SDMA0_HWID), str(am.NBIF_HWID)])),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.GC_HWID}', functools.partial(DirFileDesc, child_names=['0'])),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.GC_HWID}/0/major', functools.partial(TextFileDesc, text='11')),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.GC_HWID}/0/minor', functools.partial(TextFileDesc, text='0')),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.GC_HWID}/0/revision', functools.partial(TextFileDesc, text='0')),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.GC_HWID}/0/base_addr',
-               functools.partial(TextFileDesc, text='0x00001260\n0x0000A000\n0x0001C000\n0x02402C00')),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.SDMA0_HWID}', functools.partial(DirFileDesc, child_names=['0'])),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.SDMA0_HWID}/0/major', functools.partial(TextFileDesc, text='6')),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.SDMA0_HWID}/0/minor', functools.partial(TextFileDesc, text='0')),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.SDMA0_HWID}/0/revision', functools.partial(TextFileDesc, text='0')),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.SDMA0_HWID}/0/base_addr',
-               functools.partial(TextFileDesc, text='0x00001260\n0x0000A000\n0x0001C000\n0x02402C00')),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.NBIF_HWID}', functools.partial(DirFileDesc, child_names=['0'])),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.NBIF_HWID}/0/major', functools.partial(TextFileDesc, text='4')),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.NBIF_HWID}/0/minor', functools.partial(TextFileDesc, text='3')),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.NBIF_HWID}/0/revision', functools.partial(TextFileDesc, text='0')),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.NBIF_HWID}/0/base_addr',
-               functools.partial(TextFileDesc, text='0x00000000\n0x00000014\n0x00000D20\n0x00010400\n0x0241B000\n0x04040000')),
+      *ip_discovery_files(am.GC_HWID, ip_versions["gc"], '0x00001260\n0x0000A000\n0x0001C000\n0x02402C00'),
+      *ip_discovery_files(am.SDMA0_HWID, ip_versions["sdma"], '0x00001260\n0x0000A000\n0x0001C000\n0x02402C00'),
+      *ip_discovery_files(am.NBIF_HWID, ip_versions["nbif"], '0x00000000\n0x00000014\n0x00000D20\n0x00010400\n0x0241B000\n0x04040000'),
       VirtFile(f'/dev/dri/renderD{gpu_id}', functools.partial(DRMFileDesc, driver=self, gpu=f"{self.gpus[gpu_id]}")),
     ]
 
diff --git a/test/mockgpu/amd/amdgpu.py b/test/mockgpu/amd/amdgpu.py
index baf70087eb..f1c752c06c 100644
--- a/test/mockgpu/amd/amdgpu.py
+++ b/test/mockgpu/amd/amdgpu.py
@@ -1,8 +1,11 @@
 import ctypes, time
 from test.mockgpu.gpu import VirtGPU
 from test.mockgpu.helpers import _try_dlopen_remu
-from tinygrad.helpers import getbits, to_mv
+from tinygrad.helpers import getbits, to_mv, getenv
 from tinygrad.runtime.support import c
+
+MOCKGPU_ARCH = getenv("MOCKGPU_ARCH", "rdna3")
+GFX_TARGET_VERSION = {"rdna3": 110000, "rdna4": 120000}[MOCKGPU_ARCH]
 import tinygrad.runtime.autogen.amd_gpu as amd_gpu, tinygrad.runtime.autogen.am.pm4_nv as pm4
 
 SDMA_MAX_COPY_SIZE = 0x400000
@@ -194,10 +197,11 @@ class PM4Executor(AMDQueue):
     scratch_size = wavesize * 4  # This gives the scratch size per thread (lane)
 
     assert prg_sz > 0, "Invalid prg ptr (not found in mapped ranges)"
-    # Pass valid memory ranges, rsrc2, and scratch_size to Python emulator
+    # Pass valid memory ranges, rsrc2, scratch_size and arch to Python emulator
     if hasattr(remu, 'valid_mem_ranges'): remu.valid_mem_ranges = self.gpu.mapped_ranges
     if hasattr(remu, 'rsrc2'): remu.rsrc2 = rsrc2
     if hasattr(remu, 'scratch_size'): remu.scratch_size = scratch_size
+    if hasattr(remu, 'arch'): remu.arch = self.gpu.arch
     err = remu.run_asm(prg_addr, prg_sz, *gl, *lc, args_addr)
     if err != 0: raise RuntimeError("remu does not support the new instruction introduced in this kernel")
 
@@ -314,6 +318,7 @@ class AMDGPU(VirtGPU):
     self.regs = AMDGPURegisters()
     self.mapped_ranges = set()
     self.queues = []
+    self.arch = MOCKGPU_ARCH
 
   def map_range(self, vaddr, size): self.mapped_ranges.add((vaddr, size))
   def unmap_range(self, vaddr, size): self.mapped_ranges.remove((vaddr, size))
@@ -342,7 +347,7 @@ simd_arrays_per_engine 2
 cu_per_simd_array 8
 simd_per_cu 2
 max_slots_scratch_cu 32
-gfx_target_version 110000
+gfx_target_version {gfx_target_version}
 vendor_id 4098
 device_id 29772
 location_id 34304
diff --git a/test/mockgpu/helpers.py b/test/mockgpu/helpers.py
index a8ab73db18..d586598464 100644
--- a/test/mockgpu/helpers.py
+++ b/test/mockgpu/helpers.py
@@ -16,14 +16,15 @@ def _try_dlopen_gpuocelot():
   return None
 
 class PythonRemu:
-  """Python RDNA3 emulator wrapper that matches the libremu.so interface."""
+  """Python RDNA3/RDNA4 emulator wrapper that matches the libremu.so interface."""
   valid_mem_ranges: set[tuple[int, int]] = set()
   rsrc2: int = 0x19c  # Default: USER_SGPR_COUNT=14, enable X and Y workgroup IDs
   scratch_size: int = 0  # private_segment_fixed_size from kernel descriptor
+  arch: str = "rdna3"  # Architecture: rdna3 or rdna4
 
   def run_asm(self, lib: int, lib_sz: int, gx: int, gy: int, gz: int, lx: int, ly: int, lz: int, args_ptr: int) -> int:
     from extra.assembly.amd.emu import run_asm
-    return run_asm(lib, lib_sz, gx, gy, gz, lx, ly, lz, args_ptr, self.rsrc2, self.scratch_size)
+    return run_asm(lib, lib_sz, gx, gy, gz, lx, ly, lz, args_ptr, self.rsrc2, self.scratch_size, self.arch)
 
 def _try_dlopen_remu():
   # Use Python emulator only if PYTHON_REMU=1
diff --git a/test/test_edgecases.py b/test/test_edgecases.py
index 9f24524ae2..59358a5a0e 100644
--- a/test/test_edgecases.py
+++ b/test/test_edgecases.py
@@ -195,8 +195,10 @@ class TestAssignIssues(unittest.TestCase):
     t.shrink(((1, 3), (1, 3))).assign(Tensor.ones(2, 2))
     np.testing.assert_allclose(t.numpy(), torch_tensor.numpy())
 
+  @unittest.expectedFailure
   def test_assign_broadcast(self):
     # broadcasting during assign should behave like PyTorch
+    # NOTE: we don't want implicit dtype casting (int64 -> float32 loses precision), so this fails
     torch_tensor = torch.zeros(3, 5)
     torch_tensor[:] = torch.arange(5)
     t = Tensor.zeros(3, 5)
diff --git a/test/test_multitensor.py b/test/test_multitensor.py
index e7c8aa401c..375a6825f4 100644
--- a/test/test_multitensor.py
+++ b/test/test_multitensor.py
@@ -256,6 +256,18 @@ class TestMultiTensor(unittest.TestCase):
       a,b = _test_allreduce(Tensor.rand(256, 256))
       np.testing.assert_almost_equal(a.numpy(), b.numpy(), decimal=5)
 
+  def test_multiple_to_single_device_naive(self):
+    with Context(RING=0):
+      t = Tensor.arange(32).shard(devices_4, 0).to(Device.DEFAULT).realize()
+    self.assertEqual(t.device, Device.DEFAULT)
+    np.testing.assert_equal(t.numpy(), np.arange(32))
+
+  def test_multiple_to_single_device_ring(self):
+    with Context(RING=2):
+      t = Tensor.arange(32).shard(devices_4, 0).to(Device.DEFAULT).realize()
+    self.assertEqual(t.device, Device.DEFAULT)
+    np.testing.assert_equal(t.numpy(), np.arange(32))
+
   def test_allreduce_all2all(self):
     with Context(ALL2ALL=2):
       a,b = _test_allreduce(Tensor.rand(256, 256))
@@ -1273,19 +1285,20 @@ class TestMultiRamUsage(unittest.TestCase):
     _ = Tensor.zeros(self.N, self.N).contiguous().shard(devices_2, axis=0).contiguous().realize()
     self.assertUsed(self.N*self.N*4) # sharding should not increase total ram usage
 
-  def _test_matmul_half(self, devs):
+  def _test_matmul_half(self, dev_count:int):
     N = 32
     total_mem = {}
+    devs = tuple(f"NULL:{i}" for i in range(dev_count))
     for dtype in {dtypes.float, dtypes.half}:
       GlobalCounters.reset()
-      a = Tensor.empty((N, N), dtype=dtype).shard(devs, axis=0)
-      b = Tensor.empty((N, N), dtype=dtype).shard(devs, axis=None)
+      a = Tensor.empty((N, N), dtype=dtype, device=devs[0]).shard(devs, axis=0)
+      b = Tensor.empty((N, N), dtype=dtype, device=devs[0]).shard(devs, axis=None)
       (a @ b).realize()
       total_mem[dtype] = GlobalCounters.global_mem
     self.assertEqual(total_mem[dtypes.half], total_mem[dtypes.float] // 2)
 
-  def test_matmul_half(self): self._test_matmul_half(devices_2)
-  def test_matmul_half_alt(self): self._test_matmul_half(devices_4)
+  def test_matmul_half(self): self._test_matmul_half(dev_count=2)
+  def test_matmul_half_alt(self): self._test_matmul_half(dev_count=4)
 
 @unittest.skipIf(not_support_multi_device(), "need multi")
 class TestMultiFromUnrenderable(unittest.TestCase):
diff --git a/test/testextra/test_cfg_viz.py b/test/testextra/test_cfg_viz.py
index 45ed4a1fb6..3e9850b510 100644
--- a/test/testextra/test_cfg_viz.py
+++ b/test/testextra/test_cfg_viz.py
@@ -2,7 +2,7 @@
 # allow define from star imports
 
 import unittest
-import textwrap, functools
+import functools
 
 from tinygrad import Device, Tensor
 from tinygrad.uop.ops import UOp, Ops, KernelInfo
@@ -11,69 +11,27 @@ from tinygrad.runtime.support.compiler_amd import HIPCompiler
 from tinygrad.viz.serve import amdgpu_cfg
 
 from extra.assembly.amd.autogen.rdna3.ins import *
-from extra.assembly.amd.dsl import Inst
+from extra.assembly.amd.dsl import s
 
-template = """.text
-.globl fn_name
-.p2align 8
-.type fn_name,@function
-fn_name:
-  INSTRUCTION
-
-.rodata
-.p2align 6
-.amdhsa_kernel fn_name
-  .amdhsa_kernarg_size 8
-  .amdhsa_user_sgpr_kernarg_segment_ptr 1
-  .amdhsa_next_free_vgpr .amdgcn.next_free_vgpr
-  .amdhsa_next_free_sgpr .amdgcn.next_free_sgpr
-  .amdhsa_wavefront_size32 1
-.end_amdhsa_kernel
-
-.amdgpu_metadata
----
-amdhsa.version:
-  - 1
-  - 0
-amdhsa.kernels:
-  - .name: fn_name
-    .symbol: fn_name.kd
-    .group_segment_fixed_size: 0
-    .private_segment_fixed_size: 0
-    .wavefront_size: 32
-    .sgpr_count: 8
-    .vgpr_count: 8
-    .max_flat_workgroup_size: 1024
-    .kernarg_segment_align: 8
-    .kernarg_segment_size: 8
-    .args:
-      - .address_space:  global
-        .name:           a
-        .offset:         0
-        .size:           8
-        .type_name:      'float*'
-        .value_kind:     global_buffer
-...
-.end_amdgpu_metadata
-"""
+# TODO: this belongs to the dsl infrastructure
+from extra.gemm.amd_asm_matmul import Kernel
 
 # TODO: shouldn't need compiler once we can output ELF
 # outputs a text disassembly for humans and a machine readable binary
-def assemble(name:str, insts:list[str|Inst], compiler:Compiler) -> tuple[str, bytes]:
-  asm = "\n".join([inst if isinstance(inst, str) else inst.disasm() for inst in insts])
-  src = template.replace("fn_name", name).replace("INSTRUCTION", textwrap.dedent(asm))
+def assemble(name:str, k:Kernel, compiler:Compiler) -> tuple[str, bytes]:
+  src = k.to_asm()
   return (src, compiler.compile(src))
 
-def asm_kernel(out:UOp, insts:list[str|Inst], name:str, device:str, compiler:Compiler, n_threads:int=1, n_workgroups:int=1) -> UOp:
+def asm_kernel(out:UOp, k:Kernel, name:str, device:str, compiler:Compiler, n_threads:int=1, n_workgroups:int=1) -> UOp:
   lidx = UOp.special(n_threads, "lidx0")
   gidx = UOp.special(n_workgroups, "gidx0")
   sink = UOp.sink(out, lidx, gidx, arg=KernelInfo(name=name))
-  src, lib = assemble(name, insts, compiler)
+  src, lib = assemble(name, k, compiler)
   return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=device), UOp(Ops.LINEAR, src=(*sink.src, sink)),
                                UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=lib)))
 
-def run_asm(name:str, insts:list) -> None:
-  fxn = functools.partial(asm_kernel, insts=insts, name=name, device=Device.DEFAULT, compiler=HIPCompiler(Device[Device.DEFAULT].renderer.arch))
+def run_asm(name:str, k:Kernel) -> None:
+  fxn = functools.partial(asm_kernel, k=k, name=name, device=Device.DEFAULT, compiler=HIPCompiler(Device[Device.DEFAULT].renderer.arch))
   out = Tensor.custom_kernel(Tensor.empty(1), fxn=fxn)[0]
   out.realize()
 
@@ -85,32 +43,32 @@ class TestCfg(unittest.TestCase):
       self.skipTest(f"tests written for RDNA, got arch {arch}")
 
   def test_simple(self):
-    run_asm("simple", [
-      "entry:",
-        "s_branch bb1",
-      "bb1:",
-        s_endpgm(),
-        s_code_end(),
-    ])
+    k = Kernel(arch=Device["AMD"].arch)
+    k.label("entry")
+    k.emit(s_branch(), target="bb1")
+    k.label("bb1")
+    k.emit(s_endpgm())
+    k.emit(s_code_end())
+    run_asm("simple", k)
 
   def test_diamond(self):
-    run_asm("diamond", insts:=[
-      "entry:",
-        s_mov_b32(s[0], 0),
-        s_mov_b32(s[1], 0),
-        s_cmp_eq_u64(s[0:1], 0),
-        "s_cbranch_scc1 if",
-        "s_branch else",
-      "if:",
-        s_nop(1),
-        "s_branch end",
-      "else:",
-        s_nop(0),
-      "end:",
-        s_endpgm(),
-        s_code_end(),
-    ])
-    _, lib = assemble("diamond", insts, HIPCompiler(Device[Device.DEFAULT].arch))
+    k = Kernel(arch=Device["AMD"].arch)
+    k.label("entry")
+    k.emit(s_mov_b32(s[0], 0))
+    k.emit(s_mov_b32(s[1], 0))
+    k.emit(s_cmp_eq_u64(s[0:1], 0))
+    k.emit(s_cbranch_scc1(), target="if")
+    k.emit(s_branch(), target="else")
+    k.label("if")
+    k.emit(s_nop(1))
+    k.emit(s_branch(), target="end")
+    k.label("else")
+    k.emit(s_nop(0))
+    k.label("end")
+    k.emit(s_endpgm())
+    k.emit(s_code_end())
+    run_asm("diamond", k)
+    _, lib = assemble("diamond", k, HIPCompiler(Device[Device.DEFAULT].arch))
     cfg = amdgpu_cfg(lib, Device[Device.DEFAULT].device_props()["gfx_target_version"])["data"]
     self.assertEqual(len(cfg["blocks"]), 5)
     edge_count = sum(len(v) for v in cfg["paths"].values())
@@ -124,133 +82,138 @@ class TestCfg(unittest.TestCase):
     self.assertEqual(insts, ['s_mov_b32', 's_cmp_eq_u64'])
 
   def test_loop(self):
-    run_asm("simple_loop", [
-      "entry:",
-        s_mov_b32(s[1], 4),
-      "loop:",
-        s_add_u32(s[1], s[1], -1),
-        s_cmp_eq_i32(s[1], 0),
-        "s_cbranch_scc0 loop",
-        s_endpgm(),
-        s_code_end(),
-    ])
+    k = Kernel(arch=Device["AMD"].arch)
+    k.label("entry")
+    k.emit(s_mov_b32(s[1], 4))
+    k.label("loop")
+    k.emit(s_add_u32(s[1], s[1], -1))
+    k.emit(s_cmp_eq_i32(s[1], 0))
+    k.emit(s_cbranch_scc0(), target="loop")
+    k.emit(s_endpgm())
+    k.emit(s_code_end())
+    run_asm("simple_loop", k)
 
   def test_loop_branch(self):
-    run_asm("loop_if", [
-      "entry:",
-        s_mov_b32(s[1], 4),
-      "loop:",
-        s_add_u32(s[1], s[1], -1),
-        s_cmp_eq_i32(s[1], 2),
-        "s_cbranch_scc1 cond",
-        "s_branch cont",
-      "cond:",
-        s_add_u32(s[1], s[1], -2),
-      "cont:",
-        s_cmp_eq_i32(s[1], 0),
-        "s_cbranch_scc0 loop",
-        s_endpgm(),
-        s_code_end(),
-    ])
+    k = Kernel(arch=Device["AMD"].arch)
+    k.label("entry")
+    k.emit(s_mov_b32(s[1], 4))
+    k.label("loop")
+    k.emit(s_add_u32(s[1], s[1], -1))
+    k.emit(s_cmp_eq_i32(s[1], 2))
+    k.emit(s_cbranch_scc1(), target="cond")
+    k.emit(s_branch(), target="cont")
+    k.label("cond")
+    k.emit(s_add_u32(s[1], s[1], -2))
+    k.label("cont")
+    k.emit(s_cmp_eq_i32(s[1], 0))
+    k.emit(s_cbranch_scc0(), target="loop")
+    k.emit(s_endpgm())
+    k.emit(s_code_end())
+    run_asm("loop_if", k)
 
   def test_loop_break(self):
-    run_asm("loop_break", [
-      "entry:",
-        s_mov_b32(s[1], 8),
-      "loop:",
-        s_add_u32(s[1], s[1], -1),
-        s_cmp_eq_i32(s[1], 5),
-        "s_cbranch_scc1 break",
-        s_cmp_eq_i32(s[1], 0),
-        "s_cbranch_scc0 loop",
-      "break:",
-        s_endpgm(),
-        s_code_end(),
-    ])
+    k = Kernel(arch=Device["AMD"].arch)
+    k.label("entry")
+    k.emit(s_mov_b32(s[1], 8))
+    k.label("loop")
+    k.emit(s_add_u32(s[1], s[1], -1))
+    k.emit(s_cmp_eq_i32(s[1], 5))
+    k.emit(s_cbranch_scc1(), target="break")
+    k.emit(s_cmp_eq_i32(s[1], 0))
+    k.emit(s_cbranch_scc0(), target="loop")
+    k.label("break")
+    k.emit(s_endpgm())
+    k.emit(s_code_end())
+    run_asm("loop_break", k)
 
   def test_switch(self):
-    run_asm("switch_case", [
-      "entry:",
-        s_cmp_eq_i32(s[0], 0),
-        "s_cbranch_scc1 case0",
-        s_cmp_eq_i32(s[0], 1),
-        "s_cbranch_scc1 case1",
-        "s_branch case2",
-      "case0:",
-        s_nop(0),
-        "s_branch join",
-      "case1:",
-        s_nop(1),
-        "s_branch join",
-      "case2:",
-        s_nop(2),
-        "s_branch join",
-      "join:",
-        s_endpgm(),
-        s_code_end(),
-    ])
+    k = Kernel(arch=Device["AMD"].arch)
+    k.label("entry")
+    k.emit(s_cmp_eq_i32(s[0], 0))
+    k.emit(s_cbranch_scc1(), target="case0")
+    k.emit(s_cmp_eq_i32(s[0], 1))
+    k.emit(s_cbranch_scc1(), target="case1")
+    k.emit(s_branch(), target="case2")
+    k.label("case0")
+    k.emit(s_nop(0))
+    k.emit(s_branch(), target="join")
+    k.label("case1")
+    k.emit(s_nop(1))
+    k.emit(s_branch(), target="join")
+    k.label("case2")
+    k.emit(s_nop(2))
+    k.emit(s_branch(), target="join")
+    k.label("join")
+    k.emit(s_endpgm())
+    k.emit(s_code_end())
+    run_asm("switch_case", k)
 
   def test_ping_pong(self):
-    run_asm("ping_pong", [
-      "entry:",
-        s_cmp_eq_i32(s[0], 0),
-        "s_cbranch_scc1 ping",
-        "s_branch pong",
-      "ping:",
-        s_cmp_eq_i32(s[1], 0),
-        "s_cbranch_scc1 pong",
-        "s_branch end",
-      "pong:",
-        s_cmp_eq_i32(s[2], 0),
-        "s_cbranch_scc1 ping",
-      "end:",
-        s_endpgm(),
-        s_code_end(),
-    ])
+    k = Kernel(arch=Device["AMD"].arch)
+    k.label("entry")
+    k.emit(s_cmp_eq_i32(s[0], 0))
+    k.emit(s_cbranch_scc1(), target="ping")
+    k.emit(s_branch(), target="pong")
+    k.label("ping")
+    k.emit(s_cmp_eq_i32(s[1], 0))
+    k.emit(s_cbranch_scc1(), target="pong")
+    k.emit(s_branch(), target="end")
+    k.label("pong")
+    k.emit(s_cmp_eq_i32(s[2], 0))
+    k.emit(s_cbranch_scc1(), target="ping")
+    k.label("end")
+    k.emit(s_endpgm())
+    k.emit(s_code_end())
+    run_asm("ping_pong", k)
 
   def test_colored_blocks(self):
     N = 10
-    asm = ["entry:", "s_branch init0"]
+    k = Kernel(arch=Device["AMD"].arch)
+    k.label("entry")
+    k.emit(s_branch(), target="init0")
     for i in range(N):
-      asm += [f"init{i}:", s_mov_b32(s[1], i + 1), f"s_branch {(loop:=f'loop{i}')}"]
-      asm += [
-        f"{loop}:",
-          s_nop(i & 7),
-          s_add_u32(s[1], s[1], -1),
-          s_cmp_eq_i32(s[1], 0),
-          f"s_cbranch_scc0 {loop}",
-          f"s_branch {'init' + str(i+1) if i + 1 < N else 'end'}",
-      ]
-    asm += ["end:", s_endpgm(), s_code_end()]
-    run_asm("test_colored_blocks", asm)
+      loop = f"loop{i}"
+      k.label(f"init{i}")
+      k.emit(s_mov_b32(s[1], i + 1))
+      k.emit(s_branch(), target=loop)
+      k.label(loop)
+      k.emit(s_nop(i & 7))
+      k.emit(s_add_u32(s[1], s[1], -1))
+      k.emit(s_cmp_eq_i32(s[1], 0))
+      k.emit(s_cbranch_scc0(), target=loop)
+      k.emit(s_branch(), target=f"init{i+1}" if i + 1 < N else "end")
+    k.label("end")
+    k.emit(s_endpgm())
+    k.emit(s_code_end())
+    run_asm("test_colored_blocks", k)
 
   def test_jump_back_to_end(self):
-    run_asm("jump_back_to_end", [
-      "entry:",
-        s_mov_b32(s[1], 2),
-        "s_cbranch_execz loop",
-      "end:",
-        s_endpgm(),
-      "loop:",
-        s_add_u32(s[1], s[1], -1),
-        s_cmp_eq_i32(s[1], 0),
-        "s_branch end",
-        s_code_end(),
-    ])
+    k = Kernel(arch=Device["AMD"].arch)
+    k.label("entry")
+    k.emit(s_mov_b32(s[1], 2))
+    k.emit(s_cbranch_execz(), target="loop")
+    k.label("end")
+    k.emit(s_endpgm())
+    k.label("loop")
+    k.emit(s_add_u32(s[1], s[1], -1))
+    k.emit(s_cmp_eq_i32(s[1], 0))
+    k.emit(s_branch(), target="end")
+    k.emit(s_code_end())
+    run_asm("jump_back_to_end", k)
 
   def test_hit_count(self):
-    run_asm("test_hit_count", [
-      "entry:",
-        s_mov_b32(s[1], 1),
-        "s_branch alt",
-      "continue:",
-        s_mov_b32(s[2], 2),
-        s_add_u32(s[1], s[1], s[2]),
-      "alt:",
-        s_add_u32(s[1], s[1], -1),
-        s_endpgm(),
-        s_code_end(),
-    ])
+    k = Kernel(arch=Device["AMD"].arch)
+    k.label("entry")
+    k.emit(s_mov_b32(s[1], 1))
+    k.emit(s_branch(), target="alt")
+    k.label("continue")
+    k.emit(s_mov_b32(s[2], 2))
+    k.emit(s_add_u32(s[1], s[1], s[2]))
+    k.label("alt")
+    k.emit(s_add_u32(s[1], s[1], -1))
+    k.emit(s_endpgm())
+    k.emit(s_code_end())
+    run_asm("test_hit_count", k)
 
 if __name__ == "__main__":
   unittest.main()
diff --git a/test/unit/test_assign.py b/test/unit/test_assign.py
index 076f2b14f3..71d59a8607 100644
--- a/test/unit/test_assign.py
+++ b/test/unit/test_assign.py
@@ -461,6 +461,27 @@ class TestAssign(unittest.TestCase):
     a[2:5] = [1, 2, 3]
     np.testing.assert_allclose(a.numpy(), [0., 0., 1., 2., 3., 0., 0., 0.])
 
+  def test_assign_bitcast(self):
+    # assign to a bitcast view should modify the underlying buffer
+    a = Tensor([1.0, 2.0, 3.0, 4.0], dtype=dtypes.float32).realize()
+    # IEEE 754: 1.0f = 0x3f800000, 2.0f = 0x40000000, 3.0f = 0x40400000, 4.0f = 0x40800000
+    a.bitcast(dtypes.uint32).assign(Tensor([0x40800000, 0x40400000, 0x40000000, 0x3f800000], dtype=dtypes.uint32)).realize()
+    np.testing.assert_allclose(a.numpy(), [4.0, 3.0, 2.0, 1.0])
+    # double bitcast
+    b = Tensor([1.0, 2.0, 3.0, 4.0], dtype=dtypes.float32).realize()
+    b.bitcast(dtypes.uint32).bitcast(dtypes.int32).assign(Tensor([0x40800000, 0x40400000, 0x40000000, 0x3f800000], dtype=dtypes.int32)).realize()
+    np.testing.assert_allclose(b.numpy(), [4.0, 3.0, 2.0, 1.0])
+    # shrink then bitcast
+    c = Tensor([1.0, 2.0, 3.0, 4.0], dtype=dtypes.float32).realize()
+    c[0:2].bitcast(dtypes.uint32).assign(Tensor([0x40800000, 0x40400000], dtype=dtypes.uint32)).realize()
+    np.testing.assert_allclose(c.numpy(), [4.0, 3.0, 3.0, 4.0])
+
+  def test_assign_bitcast_different_size(self):
+    # different-size bitcast creates a new tensor, not a view, so assign doesn't modify the original
+    a = Tensor([0]*8, dtype=dtypes.uint8).realize()
+    a.bitcast(dtypes.int64).assign(Tensor([12345], dtype=dtypes.int64)).realize()
+    np.testing.assert_equal(a.numpy(), [0]*8)
+
   @unittest.skip("don't use output buffer, and mismatch dtype no longer supported")
   def test_cast_assignment(self):
     a = Tensor(np.arange(N*N, dtype=np.float32)).reshape(N,N)
@@ -472,6 +493,38 @@ class TestAssign(unittest.TestCase):
     assert oba1 is None and oba2 is None
     np.testing.assert_allclose(a.numpy(), np.arange(N*N,dtype=np.int32).reshape((N,N)))
 
+  def test_assign_dtype_mismatch(self):
+    # assign should not implicitly cast dtypes - this can lose precision
+    a = Tensor.zeros(4, dtype=dtypes.float32).contiguous().realize()
+    b = Tensor([1, 2, 3, 4], dtype=dtypes.int32)
+    with self.assertRaisesRegex(RuntimeError, "assign dtype mismatch"):
+      a.assign(b)
+
+  def test_assign_dtype_mismatch_int64_to_float32(self):
+    # int64 -> float32 loses precision for large values, should not be implicit
+    a = Tensor.zeros(1, dtype=dtypes.float32).contiguous().realize()
+    b = Tensor([16777217], dtype=dtypes.int64)  # 2^24 + 1, not exactly representable in float32
+    with self.assertRaisesRegex(RuntimeError, "assign dtype mismatch"):
+      a.assign(b)
+
+  def test_assign_shape_broadcast(self):
+    # shape broadcasting should work when dtypes match
+    a = Tensor.zeros(3, 5, dtype=dtypes.float32).contiguous().realize()
+    b = Tensor([1., 2., 3., 4., 5.], dtype=dtypes.float32)
+    a.assign(b)
+    a.realize()
+    expected = np.array([[1., 2., 3., 4., 5.]] * 3)
+    np.testing.assert_allclose(a.numpy(), expected)
+
+  def test_assign_shape_broadcast_2d(self):
+    # broadcast (1, 5) to (3, 5)
+    a = Tensor.zeros(3, 5, dtype=dtypes.float32).contiguous().realize()
+    b = Tensor([[1., 2., 3., 4., 5.]], dtype=dtypes.float32)
+    a.assign(b)
+    a.realize()
+    expected = np.array([[1., 2., 3., 4., 5.]] * 3)
+    np.testing.assert_allclose(a.numpy(), expected)
+
   def test_disk_assignment(self):
     a = Tensor.empty(5, device=f"disk:{temp('disk_assignment')}").assign(Tensor.ones(5)).numpy()
     np.testing.assert_equal(a, np.ones(5))
@@ -566,12 +619,12 @@ class TestAssignOrdering(unittest.TestCase):
   def test_slice_write_then_full_read(self):
     """Write to slice, then read full buffer."""
     # without .realize(): orphan slice assign not triggered by .numpy()
-    buf = Tensor.zeros(4).contiguous().realize()
+    buf = Tensor.zeros(4, dtype=dtypes.int32).contiguous().realize()
     buf[1:3].assign(Tensor([5, 6]))
     np.testing.assert_equal(buf.numpy(), [0, 0, 0, 0])  # TODO: wrong! should be [0, 5, 6, 0]
 
     # with .realize(): assign executes
-    buf = Tensor.zeros(4).contiguous().realize()
+    buf = Tensor.zeros(4, dtype=dtypes.int32).contiguous().realize()
     buf[1:3].assign(Tensor([5, 6])).realize()
     np.testing.assert_equal(buf.numpy(), [0, 5, 6, 0])
 
@@ -653,7 +706,7 @@ class TestAssignOrdering(unittest.TestCase):
 
   def test_three_buffer_chain(self):
     """Chain: A depends on B, B depends on C - ordering matters."""
-    a = Tensor.zeros(4).contiguous().realize()
+    a = Tensor.zeros(4, dtype=dtypes.int32).contiguous().realize()
     b = Tensor([1, 2, 3, 4]).contiguous().realize()
     c = Tensor([10, 10, 10, 10]).contiguous().realize()
     # b reads from c, a reads from b
@@ -665,8 +718,8 @@ class TestAssignOrdering(unittest.TestCase):
 
   def test_interleaved_assign_read_patterns(self):
     """Complex interleaved pattern: write A, read A into B, write B, read B."""
-    a = Tensor.zeros(4).contiguous().realize()
-    b = Tensor.zeros(4).contiguous().realize()
+    a = Tensor.zeros(4, dtype=dtypes.int32).contiguous().realize()
+    b = Tensor.zeros(4, dtype=dtypes.int32).contiguous().realize()
 
     a.assign(Tensor([1, 2, 3, 4]))
     b.assign(a.contiguous())       # b should get [1,2,3,4]
diff --git a/test/unit/test_autogen.py b/test/unit/test_autogen.py
index 0a0f41da87..55d107ce40 100644
--- a/test/unit/test_autogen.py
+++ b/test/unit/test_autogen.py
@@ -5,24 +5,13 @@ from tinygrad.runtime.support.c import DLL, record, init_records
 from tinygrad.runtime.support import c
 from tinygrad.runtime.support.autogen import gen
 
-class TestAutogen(unittest.TestCase):
+@unittest.skipIf(WIN, "doesn't compile on windows")
+class TestC(unittest.TestCase):
   def compile(self, src):
     with tempfile.NamedTemporaryFile(suffix=".so") as f:
       subprocess.check_output(('clang', '-x', 'c', '-fPIC', '-shared', '-', '-o', f.name), input=src.encode())
       return DLL("test", f.name)
 
-  def run_gen(self, contents):
-    with tempfile.NamedTemporaryFile(mode='w', suffix='.h') as f:
-      f.write(contents)
-      f.flush()
-
-      generated_code = gen(name="test_header", dll=None, files=[f.name])
-
-      namespace = {}
-      exec(generated_code, namespace)
-      return namespace
-
-  @unittest.skipIf(WIN, "doesn't compile on windows")
   def test_packed_struct(self):
     @record
     class Baz:
@@ -45,7 +34,6 @@ class TestAutogen(unittest.TestCase):
     assert b.c == 1
     assert b.d == 0
 
-  @unittest.skipIf(WIN, "doesn't compile on windows")
   def test_packed_struct_interop(self):
     @record
     class Baz:
@@ -75,7 +63,6 @@ class TestAutogen(unittest.TestCase):
     self.assertEqual(test(b), b.a + b.b + b.c + b.d)
 
   # https://github.com/python/cpython/issues/90914
-  @unittest.skipIf(WIN, "doesn't compile on windows")
   def test_bitfield_interop(self):
     @record
     class Baz:
@@ -103,7 +90,6 @@ class TestAutogen(unittest.TestCase):
     def test(x:Baz) -> ctypes.c_int: ...
     for i in range(8): self.assertEqual(test(Baz(*(j==i for j in range(8)))), i==2)
 
-  @unittest.skipIf(WIN, "doesn't compile on windows")
   def test_struct_interop(self):
     @record
     class Baz:
@@ -131,7 +117,6 @@ class TestAutogen(unittest.TestCase):
     def test(x:Baz) -> Baz: ...
     self.assertEqual(bytes(test(Baz(*range(8)))), struct.pack("8i", *range(7, -1, -1)))
 
-  @unittest.skipIf(WIN, "doesn't compile on windows")
   def test_aos_interop(self):
     @record
     class Item:
@@ -151,7 +136,6 @@ class TestAutogen(unittest.TestCase):
     def test(arr:(Item * 3)) -> ctypes.c_int: ...
     self.assertEqual(test((Item * 3)(Item(10), Item(20), Item(30))), 60)
 
-  @unittest.skipIf(WIN, "doesn't compile on windows")
   def test_soa_interop(self):
     @record
     class Row:
@@ -173,7 +157,6 @@ class TestAutogen(unittest.TestCase):
     self.assertEqual(r.data[1], 20)
     self.assertEqual(r.data[2], 10)
 
-  @unittest.skipIf(WIN, "doesn't compile on windows")
   def test_soa_ptr_interop(self):
     @record
     class Row:
@@ -191,7 +174,6 @@ class TestAutogen(unittest.TestCase):
     def test(x:Row) -> ctypes.c_int: ...
     assert test(Row((ctypes.c_int * 3)(10, 20, 30))) == 60
 
-  @unittest.skipIf(WIN, "doesn't compile on windows")
   def test_nested_struct_interop(self):
     @record
     class Inner:
@@ -217,7 +199,6 @@ class TestAutogen(unittest.TestCase):
     self.assertEqual(o.inner.a, 20)
     self.assertEqual(o.b, 10)
 
-  @unittest.skipIf(WIN, "doesn't compile on windows")
   def test_struct_pointer_interop(self):
     @record
     class Foo:
@@ -242,7 +223,88 @@ class TestAutogen(unittest.TestCase):
     self.assertEqual(out.contents.a, 20)
     self.assertEqual(out.contents.b, 10)
 
-  @unittest.skipIf(WIN, "doesn't compile on windows")
+  def test_pointer_field_roundtrip(self):
+    # This tests storing a pointer in a record struct field and passing it to C
+    # Mimics how mesa.struct_lp_build_tgsi_params.mask is used
+    from tinygrad.runtime.support.c import POINTER
+    @record
+    class Inner:
+      SIZE = 8
+      value: Annotated[ctypes.c_int, 0]
+      flag: Annotated[ctypes.c_int, 4]
+    @record
+    class Outer:
+      SIZE = 16
+      x: Annotated[ctypes.c_int, 0]
+      inner_ptr: Annotated[POINTER[Inner], 8]
+    init_records()
+
+    src = """
+      struct inner { int value; int flag; };
+      struct outer { int x; struct inner *inner_ptr; };
+      int test(struct inner *p) {
+        return p->value + p->flag;
+      }
+    """
+    dll = self.compile(src)
+    @dll.bind
+    def test(p:POINTER[Inner]) -> ctypes.c_int: ...
+
+    inner = Inner(value=42, flag=10)
+    outer = Outer(x=1, inner_ptr=ctypes.pointer(inner))
+    # Retrieve pointer from struct field and pass to C
+    self.assertEqual(test(outer.inner_ptr), 52)
+
+  def test_pointer_field_loses_reference(self):
+    # BUG: When a pointer is stored in a record struct field, only the address bytes are saved.
+    # The pointer's _objects dict (which prevents GC of the pointed-to object) is lost.
+    # This causes the pointed-to object to be garbage collected, leading to use-after-free.
+    from tinygrad.runtime.support.c import POINTER
+    @record
+    class MaskContext:
+      SIZE = 16
+      value: Annotated[ctypes.c_int, 0]
+      initialized: Annotated[ctypes.c_int, 4]
+      ptr: Annotated[ctypes.c_void_p, 8]
+    @record
+    class Params:
+      SIZE = 16
+      x: Annotated[ctypes.c_int, 0]
+      mask: Annotated[POINTER[MaskContext], 8]
+    init_records()
+
+    src = """
+      struct mask_ctx { int value; int initialized; void *ptr; };
+      void mask_begin(struct mask_ctx *m, int val) { m->value = val; m->initialized = 1; }
+      int mask_end(struct mask_ctx *m) { return m->value + m->initialized; }
+    """
+    dll = self.compile(src)
+    @dll.bind
+    def mask_begin(m:POINTER[MaskContext], val:ctypes.c_int) -> None: ...
+    @dll.bind
+    def mask_end(m:POINTER[MaskContext]) -> ctypes.c_int: ...
+
+    # When MaskContext() is created inline, it gets garbage collected after the pointer
+    # is stored because only the address bytes are saved, not the _objects reference.
+    params = Params(x=1, mask=ctypes.pointer(MaskContext()))
+    mask_begin(params.mask, 42)
+    result = mask_end(params.mask)
+    self.assertEqual(result, 43)  # 42 + 1
+
+@unittest.skipIf(OSX and ('MTLCompiler' in DLL._loaded_ or 'llvm' in DLL._loaded_), "libclang can't be loaded after MTLCompiler or llvm on OSX")
+@unittest.skipIf(WIN, "doesn't compile on windows")
+class TestAutogen(unittest.TestCase):
+  def run_gen(self, contents):
+    with tempfile.NamedTemporaryFile(mode='w', suffix='.h') as f:
+      f.write(contents)
+      f.flush()
+
+      generated_code = gen(name="test_header", dll=None, files=[f.name])
+
+      namespace = {}
+      exec(generated_code, namespace)
+      return namespace
+
   def test_packed_structs(self):
     ns = self.run_gen("""
 typedef unsigned NvU32;
@@ -292,47 +354,6 @@ typedef struct
     assert frts_cmd.readVbiosDesc.__class__ is FWSECLIC_READ_VBIOS_DESC
     assert frts_cmd.frtsRegionDesc.__class__ is FWSECLIC_FRTS_REGION_DESC
 
-  @unittest.skipIf(WIN, "doesn't compile on windows")
-  @unittest.skipIf(OSX, "can't find stdint?")
-  def test_packed_fields(self):
-    ns = self.run_gen("""#include <stdint.h>
-typedef struct die_info
- {
-	 uint16_t die_id;
-	 uint16_t die_offset; /* Points to the corresponding die_header structure */
- } die_info;
-
-typedef struct ip_discovery_header
- {
-	 uint32_t signature;    /* Table Signature */
-	 uint16_t version;      /* Table Version */
-	 uint16_t size;         /* Table Size */
-	 uint32_t id;           /* Table ID */
-	 uint16_t num_dies;     /* Number of Dies */
-	 die_info die_info[16]; /* list die information for up to 16 dies */
-	 union {
-		 uint16_t padding[1];	/* version <= 3 */
-		 struct {		/* version == 4 */
-			 uint8_t base_addr_64_bit : 1; /* ip structures are using 64 bit base address */
-			 uint8_t reserved : 7;
-			 uint8_t reserved2;
-		 };
-	 };
- } ip_discovery_header;
-""")
-
-    ip_discovery_header = ns['ip_discovery_header']
-
-    hdr = b'IPDS\x04\x00|\x1d\x80\x1a\xffd\x01\x00\x00\x00\x8c\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01\x00' # noqa: E501
-    ihdr = ip_discovery_header.from_buffer_copy(hdr)
-
-    assert ctypes.sizeof(ihdr) == 80
-    assert ihdr.signature == 0x53445049
-    assert ihdr.version == 0x0004
-    assert ihdr.num_dies == 1
-    assert ihdr.base_addr_64_bit == 1
-
-  @unittest.skipIf(WIN, "doesn't compile on windows")
   def test_gen_from_header(self):
     namespace = self.run_gen("""
     typedef struct {
@@ -378,7 +399,6 @@ typedef struct ip_discovery_header
     self.assertTrue(hasattr(rect, 'height'))
     self.assertTrue(hasattr(rect, 'color'))
 
-  @unittest.skipIf(WIN, "doesn't compile on windows")
   def test_struct_ordering(self):
     namespace = self.run_gen("""
     struct A;
@@ -408,77 +428,6 @@ typedef struct ip_discovery_header
     self.assertTrue(hasattr(b, 'c_ptr'))
     self.assertTrue(hasattr(c, 'a_ptr'))
 
-  @unittest.skipIf(WIN, "doesn't compile on windows")
-  def test_pointer_field_roundtrip(self):
-    # This tests storing a pointer in a record struct field and passing it to C
-    # Mimics how mesa.struct_lp_build_tgsi_params.mask is used
-    from tinygrad.runtime.support.c import POINTER
-    @record
-    class Inner:
-      SIZE = 8
-      value: Annotated[ctypes.c_int, 0]
-      flag: Annotated[ctypes.c_int, 4]
-    @record
-    class Outer:
-      SIZE = 16
-      x: Annotated[ctypes.c_int, 0]
-      inner_ptr: Annotated[POINTER[Inner], 8]
-    init_records()
-
-    src = """
-      struct inner { int value; int flag; };
-      struct outer { int x; struct inner *inner_ptr; };
-      int test(struct inner *p) {
-        return p->value + p->flag;
-      }
-    """
-    dll = self.compile(src)
-    @dll.bind
-    def test(p:POINTER[Inner]) -> ctypes.c_int: ...
-
-    inner = Inner(value=42, flag=10)
-    outer = Outer(x=1, inner_ptr=ctypes.pointer(inner))
-    # Retrieve pointer from struct field and pass to C
-    self.assertEqual(test(outer.inner_ptr), 52)
-
-  @unittest.skipIf(WIN, "doesn't compile on windows")
-  def test_pointer_field_loses_reference(self):
-    # BUG: When a pointer is stored in a record struct field, only the address bytes are saved.
-    # The pointer's _objects dict (which prevents GC of the pointed-to object) is lost.
-    # This causes the pointed-to object to be garbage collected, leading to use-after-free.
-    from tinygrad.runtime.support.c import POINTER
-    @record
-    class MaskContext:
-      SIZE = 16
-      value: Annotated[ctypes.c_int, 0]
-      initialized: Annotated[ctypes.c_int, 4]
-      ptr: Annotated[ctypes.c_void_p, 8]
-    @record
-    class Params:
-      SIZE = 16
-      x: Annotated[ctypes.c_int, 0]
-      mask: Annotated[POINTER[MaskContext], 8]
-    init_records()
-
-    src = """
-      struct mask_ctx { int value; int initialized; void *ptr; };
-      void mask_begin(struct mask_ctx *m, int val) { m->value = val; m->initialized = 1; }
-      int mask_end(struct mask_ctx *m) { return m->value + m->initialized; }
-    """
-    dll = self.compile(src)
-    @dll.bind
-    def mask_begin(m:POINTER[MaskContext], val:ctypes.c_int) -> None: ...
-    @dll.bind
-    def mask_end(m:POINTER[MaskContext]) -> ctypes.c_int: ...
-
-    # When MaskContext() is created inline, it gets garbage collected after the pointer
-    # is stored because only the address bytes are saved, not the _objects reference.
-    params = Params(x=1, mask=ctypes.pointer(MaskContext()))
-    mask_begin(params.mask, 42)
-    result = mask_end(params.mask)
-    self.assertEqual(result, 43)  # 42 + 1
-
-  @unittest.skipIf(WIN, "doesn't compile on windows")
   def test_anonymous_children(self):
     namespace = self.run_gen("""
       struct foo {
@@ -491,7 +440,6 @@ typedef struct ip_discovery_header
     self.assertIn('struct_foo', namespace)
     self.assertIn('struct_foo_bar', namespace)
 
-  @unittest.skipIf(WIN, "doesn't compile on windows")
   def test_enums(self):
     namespace = self.run_gen("""
       enum Foo { A, B, C };
@@ -511,4 +459,43 @@ typedef struct ip_discovery_header
     assert namespace["enum_Bar"].get(1) == "Y"
     assert namespace["enum_Bar"].get(2) == "Z"
 
+  @unittest.skipIf(OSX, "can't find stdint?")
+  def test_packed_fields(self):
+    ns = self.run_gen("""#include <stdint.h>
+typedef struct die_info
+ {
+	 uint16_t die_id;
+	 uint16_t die_offset; /* Points to the corresponding die_header structure */
+ } die_info;
+
+typedef struct ip_discovery_header
+ {
+	 uint32_t signature;    /* Table Signature */
+	 uint16_t version;      /* Table Version */
+	 uint16_t size;         /* Table Size */
+	 uint32_t id;           /* Table ID */
+	 uint16_t num_dies;     /* Number of Dies */
+	 die_info die_info[16]; /* list die information for up to 16 dies */
+	 union {
+		 uint16_t padding[1];	/* version <= 3 */
+		 struct {		/* version == 4 */
+			 uint8_t base_addr_64_bit : 1; /* ip structures are using 64 bit base address */
+			 uint8_t reserved : 7;
+			 uint8_t reserved2;
+		 };
+	 };
+ } ip_discovery_header;
+""")
+
+    ip_discovery_header = ns['ip_discovery_header']
+
+    hdr = b'IPDS\x04\x00|\x1d\x80\x1a\xffd\x01\x00\x00\x00\x8c\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01\x00' # noqa: E501
+    ihdr = ip_discovery_header.from_buffer_copy(hdr)
+
+    assert ctypes.sizeof(ihdr) == 80
+    assert ihdr.signature == 0x53445049
+    assert ihdr.version == 0x0004
+    assert ihdr.num_dies == 1
+    assert ihdr.base_addr_64_bit == 1
+
 if __name__ == "__main__": unittest.main()
diff --git a/test/unit/test_call.py b/test/unit/test_call.py
index 473b2962bc..1c5ef3829a 100644
--- a/test/unit/test_call.py
+++ b/test/unit/test_call.py
@@ -42,7 +42,7 @@ class TestCall(unittest.TestCase):
     b = Tensor.randn(K, N)
     Tensor.realize(a, b)
     c = Tensor.call(a, b, fxn=a.as_param(0) @ b.as_param(1))
-    np.testing.assert_allclose(c.numpy(), a.numpy() @ b.numpy(), rtol=1e-5)
+    np.testing.assert_allclose(c.numpy(), a.numpy() @ b.numpy(), rtol=1e-5, atol=1e-6)
 
   @unittest.skip("needs GEMM on mixins")
   def test_call_gemm_uop(self):
@@ -56,7 +56,7 @@ class TestCall(unittest.TestCase):
     y = UOp.param(1, dtypes.float, shape=(K, N))
     c = Tensor.call(a, b, fxn=x@y)
 
-    np.testing.assert_allclose(c.numpy(), a.numpy() @ b.numpy(), rtol=1e-5)
+    np.testing.assert_allclose(c.numpy(), a.numpy() @ b.numpy(), rtol=1e-5, atol=1e-6)
 
 if __name__ == '__main__':
   unittest.main()
diff --git a/test/unit/test_disk_tensor.py b/test/unit/test_disk_tensor.py
index 0ed52dafc4..fbc5b0f914 100644
--- a/test/unit/test_disk_tensor.py
+++ b/test/unit/test_disk_tensor.py
@@ -59,13 +59,13 @@ class TestRawDiskBuffer(unittest.TestCase):
     _test_bitcasted(t, dtypes.float32, 0.0)
     _test_bitcasted(t, dtypes.uint32, 0)
     # pi in float16 stored via int16
-    t.bitcast(dtypes.uint16).assign(Tensor.full((128, 64), 0x4248, dtype=dtypes.uint16)).realize()
+    t.assign(Tensor.full((128, 64), 0x4248, dtype=dtypes.uint16).bitcast(dtypes.uint8)).realize()
     _test_bitcasted(t, dtypes.float16, 3.140625)
     _test_bitcasted(t, dtypes.float32, 50.064727)
     _test_bitcasted(t, dtypes.uint16, 0x4248)
     _test_bitcasted(t, dtypes.uint32, 0x42484248)
     # pi in float32 stored via float32
-    t.bitcast(dtypes.float32).assign(Tensor.full((128, 32), 3.1415927, dtype=dtypes.float32)).realize()
+    t.assign(Tensor.full((128, 32), 3.1415927, dtype=dtypes.float32).bitcast(dtypes.uint8)).realize()
     _test_bitcasted(t, dtypes.float32, 3.1415927)
     _test_bitcasted(t, dtypes.uint32, 0x40490FDB)
     # doesn't suport normal cast
@@ -348,13 +348,25 @@ class TestDiskTensor(unittest.TestCase):
 
   def test_assign_with_bitcast(self):
     # bitcast assign is used in safe_save for writing header length
-    # this tests the synchronous disk assign hack handles bitcast correctly
+    # bitcast on source side works, bitcast on target side raises
     pathlib.Path(temp(fn:="dt_assign_bitcast")).unlink(missing_ok=True)
     t = Tensor.empty(16, device=f"disk:{temp(fn)}", dtype=dtypes.uint8)
-    t[0:8].bitcast(dtypes.int64).assign([12345])
-    # verify the data was written correctly
+    # correct way: bitcast the source to match target dtype
+    t[0:8].assign(Tensor([12345], dtype=dtypes.int64, device="CPU").bitcast(dtypes.uint8))
     val = int.from_bytes(t[0:8].data(), 'little')
     self.assertEqual(val, 12345)
+    # bitcast on target with non-broadcastable dtype raises
+    with self.assertRaises(RuntimeError):
+      t[0:4].bitcast(dtypes.int32).assign(Tensor([12345], dtype=dtypes.int64))
+
+  def test_assign_to_bitcast_view(self):
+    # assign float values to a float32 view of a uint8 disk buffer (used by safe_save)
+    pathlib.Path(temp(fn:="dt_bitcast_view_assign")).unlink(missing_ok=True)
+    t = Tensor.empty(32, device=f"disk:{temp(fn)}", dtype=dtypes.uint8)
+    # create float32 view of bytes 8-24 (4 floats)
+    float_view = t[8:24].bitcast(dtypes.float32)
+    float_view.assign(Tensor([1.0, 2.0, 3.0, 4.0], dtype=dtypes.float32, device="CPU"))
+    np.testing.assert_array_equal(float_view.numpy(), [1.0, 2.0, 3.0, 4.0])
 
   def test_assign_cross_device(self):
     # disk assign allows cross-device (source on GPU/CPU, target on disk)
diff --git a/test/test_memory_planner.py b/test/unit/test_memory_planner.py
similarity index 96%
rename from test/test_memory_planner.py
rename to test/unit/test_memory_planner.py
index 6aa03c6be3..32d3250820 100644
--- a/test/test_memory_planner.py
+++ b/test/unit/test_memory_planner.py
@@ -1,5 +1,5 @@
 import unittest
-from tinygrad import dtypes, Device
+from tinygrad import dtypes
 from tinygrad.device import Buffer
 from tinygrad.engine.memory import _internal_memory_planner
 
@@ -7,7 +7,7 @@ global_map = {}
 def b(i, base=None, offset=0, pin=False, size=16):
   global global_map
   if i in global_map: return global_map[i]
-  global_map[i] = Buffer(Device.DEFAULT, size, dtypes.int8, base=global_map[base] if base is not None else None, offset=offset)
+  global_map[i] = Buffer("NULL", size, dtypes.int8, base=global_map[base] if base is not None else None, offset=offset)
   if pin: global_map[i].ref(1)
   return global_map[i]
 
diff --git a/test/unit/test_schedule_cache.py b/test/unit/test_schedule_cache.py
index 71fb85b831..e7908228d7 100644
--- a/test/unit/test_schedule_cache.py
+++ b/test/unit/test_schedule_cache.py
@@ -35,7 +35,6 @@ class TestScheduleCache(unittest.TestCase):
     _, var_vals = t.schedule_with_vars()
     self.assertEqual(var_vals, {'pos': 42})
 
-  @Context(SPEC=0)
   def test_custom_kernel(self):
     for i in range(4):
       a = Tensor.empty(1)
@@ -43,7 +42,6 @@ class TestScheduleCache(unittest.TestCase):
       a.realize()
       self.assertEqual(a.item(), i)
 
-  @Context(SPEC=0)
   def test_same_custom_function_reuses_cache(self):
     schedule_cache.clear()
     fxn = functools.partial(custom_set0_kernel, num=10)
diff --git a/test/unit/test_simplify_valid_idx.py b/test/unit/test_simplify_valid_idx.py
index 5e016bbcef..c8818ec16d 100644
--- a/test/unit/test_simplify_valid_idx.py
+++ b/test/unit/test_simplify_valid_idx.py
@@ -470,5 +470,19 @@ class TestUnfoldableImageChannelSelection(unittest.TestCase):
     load = UOp(Ops.LOAD, dtypes.float, (UOp(Ops.DEFINE_GLOBAL, dtypes.imagef((10, 10, 4)), arg=0).index(x, ptr=True), UOp.const(dtypes.float, 0)))
     self.assertEqual(self._count_nans(load), 1)
 
+class TestDropTrueGate(unittest.TestCase):
+  def test_drop_true_gate_on_index(self):
+    # test that INDEX with a constant True gate gets simplified to drop the gate
+    from tinygrad.codegen.late.devectorizer import load_store_indexing
+    from tinygrad.uop.ops import graph_rewrite
+    buf = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(), arg=0)
+    idx = UOp.const(dtypes.index, 0)
+    true_gate = UOp.const(dtypes.bool, True)
+    index_with_gate = UOp(Ops.INDEX, dtypes.int.ptr(), (buf, idx, true_gate))
+    # apply the optimization
+    result = graph_rewrite(index_with_gate, load_store_indexing)
+    # the True gate should be dropped (INDEX should only have 2 sources)
+    self.assertEqual(len(result.src), 2, "True gate should be dropped from INDEX")
+
 if __name__ == '__main__':
   unittest.main()
diff --git a/test/test_uop_graph.py b/test/unit/test_uop_graph.py
similarity index 84%
rename from test/test_uop_graph.py
rename to test/unit/test_uop_graph.py
index f57c55ba41..76b252cec7 100644
--- a/test/test_uop_graph.py
+++ b/test/unit/test_uop_graph.py
@@ -479,6 +479,26 @@ class TestUOpGraph(unittest.TestCase):
     for u in uops:
       self.assertNotEqual(u.dtype, dtypes.long)
 
+  def test_load_idx_no_math_on_loaded(self):
+    # test the (x+y)<c pattern where x has loads - we shouldn't do math on loaded indices
+    c0 = UOp(Ops.DEFINE_GLOBAL, dtypes.uchar.ptr(128000), arg=0, src=())
+    c1 = UOp.range(UOp.const(dtypes.index, 512), 1, AxisType.LOOP)
+    c2 = UOp.range(UOp.const(dtypes.index, 250), 2, AxisType.LOOP)
+    c3 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(512), arg=1, src=())
+    c4 = c3.index(c1)  # c4 is a load
+    c5 = UOp.range(UOp.const(dtypes.index, 240), 0, AxisType.REDUCE)
+    c6 = ((c2*UOp.const(dtypes.index, 240))+c5)
+    c7 = UOp(Ops.DEFINE_GLOBAL, dtypes.uchar.ptr(60000), arg=2, src=())
+    c8 = c7.index(c6)
+    # (loaded + range) < const pattern - loaded value shouldn't be promoted to long
+    loaded_idx = c4.cast(dtypes.index)
+    comparison = (loaded_idx + c5) < UOp.const(dtypes.index, 60000)
+    c9 = comparison.where(c8.cast(dtypes.uint).cast(dtypes.uchar), 0).reduce(c5, arg=Ops.ADD)
+    c10 = c0.index(((c1*UOp.const(dtypes.index, 250))+c2)).store(c9).end(c1, c2)
+    uops = to_uops_list([c10])
+    for u in uops:
+      self.assertNotEqual(u.dtype, dtypes.long)
+
   def test_fold_gated_load(self):
     glbl0 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(), (), 0)
     glbl1 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(), (), 1)
@@ -686,6 +706,97 @@ class TestExpander(unittest.TestCase):
     sink = expander_rewrite(sink)
     print(sink)
 
+class TestReduceCollapse(unittest.TestCase):
+  def test_multi_range_reduce_add(self):
+    """Test that (x + y).reduce(r1, r2) distributes over multiple ranges"""
+    from tinygrad.codegen.simplify import pm_reduce_collapse
+    # Create two ranges
+    r1 = UOp.range(3, 0)
+    r2 = UOp.range(4, 1)
+    # Create x + y where x and y depend on different ranges
+    x = r1.cast(dtypes.float)
+    y = r2.cast(dtypes.float)
+    # (x + y).reduce(r1, r2) should be rewritten
+    red = (x + y).reduce(r1, r2, arg=Ops.ADD)
+    self.assertEqual(len(red.src), 3)  # value + 2 ranges
+    result = graph_rewrite(red, pm_reduce_collapse, name='test')
+    # Should become add of two separate reduces
+    self.assertEqual(result.op, Ops.ADD)
+
+class TestLoadStoreFolding(unittest.TestCase):
+  def test_gated_load_gep_preserves_alt(self):
+    """Test that LOAD(GEP, alt) preserves alt value after rewrite"""
+    from tinygrad.codegen.late.devectorizer import load_store_folding
+    buf = UOp(Ops.DEFINE_GLOBAL, dtypes.float.vec(4).ptr(), (), 0)
+    idx = UOp.const(dtypes.int, 0)
+    gate = UOp.const(dtypes.bool, True)
+    gated_index = buf.index(idx, gate)
+    gep = gated_index.gep(0)
+    alt = UOp.const(dtypes.float, 42.0)
+    gated_load = gep.load(alt)
+    self.assertEqual(len(gated_load.src), 2)  # GEP + alt
+    result = graph_rewrite(gated_load, load_store_folding, name='test')
+    # After rewrite, should still have alt value preserved
+    self.assertEqual(result.op, Ops.GEP)
+    inner_load = result.src[0]
+    self.assertEqual(inner_load.op, Ops.LOAD)
+    self.assertEqual(len(inner_load.src), 2)  # INDEX + alt
+
+  def test_gated_load_ptrcat_preserves_alt(self):
+    """Test that LOAD(PTRCAT, alt) preserves alt value after rewrite"""
+    from tinygrad.codegen.late.devectorizer import load_store_folding
+    buf1 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), (), 0)
+    buf2 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), (), 1)
+    idx = UOp.const(dtypes.int, 0)
+    idx1 = buf1.index(idx)
+    idx2 = buf2.index(idx)
+    ptrcat = UOp(Ops.PTRCAT, dtypes.float.ptr().vec(2), (idx1, idx2))
+    alt = UOp.const(dtypes.float.vec(2), 42.0)
+    gated_load = ptrcat.load(alt)
+    self.assertEqual(len(gated_load.src), 2)  # PTRCAT + alt
+    result = graph_rewrite(gated_load, load_store_folding, name='test')
+    # After rewrite, should be CAT of LOADs, each preserving alt
+    self.assertEqual(result.op, Ops.CAT)
+    for inner_load in result.src:
+      self.assertEqual(inner_load.op, Ops.LOAD)
+      self.assertEqual(len(inner_load.src), 2)  # INDEX + alt
+    self.assertEqual(inner_load.src[1].arg, 42.0)  # alt value preserved
+
+class TestConstBufferize(unittest.TestCase):
+  def test_const_bufferize_with_ranges(self):
+    """Test that CONST.BUFFERIZE with ranges is folded correctly.
+
+    BUFFERIZE can have ranges as additional sources beyond the value.
+    The pattern at rangeify.py uses allow_any_len=True because
+    CONST doesn't depend on ranges (constant is same value everywhere).
+    """
+    from tinygrad.schedule.rangeify import pm_const_buffer_folding, BufferizeOpts
+    c = UOp.const(dtypes.float, 42.0)
+    r1 = UOp.range(3, 0)
+    bufferize_with_range = UOp(Ops.BUFFERIZE, dtypes.float, (c, r1), arg=BufferizeOpts(device="CPU"))
+    self.assertEqual(len(bufferize_with_range.src), 2)  # const + 1 range
+
+    result = graph_rewrite(bufferize_with_range, pm_const_buffer_folding, name='test')
+    # BUFFERIZE should be removed, result is const broadcast to shape
+    self.assertNotEqual(result.op, Ops.BUFFERIZE)
+    const_vals = [u.arg for u in result.toposort() if u.op is Ops.CONST and u.dtype == dtypes.float]
+    self.assertIn(42.0, const_vals)
+
+  def test_const_bufferize_with_multiple_ranges(self):
+    """Test CONST.BUFFERIZE with multiple ranges is also folded."""
+    from tinygrad.schedule.rangeify import pm_const_buffer_folding, BufferizeOpts
+    c = UOp.const(dtypes.float, 3.14)
+    r1 = UOp.range(3, 0)
+    r2 = UOp.range(4, 1)
+    bufferize_with_ranges = UOp(Ops.BUFFERIZE, dtypes.float, (c, r1, r2), arg=BufferizeOpts(device="CPU"))
+    self.assertEqual(len(bufferize_with_ranges.src), 3)  # const + 2 ranges
+
+    result = graph_rewrite(bufferize_with_ranges, pm_const_buffer_folding, name='test')
+    # BUFFERIZE should be removed
+    self.assertNotEqual(result.op, Ops.BUFFERIZE)
+    const_vals = [u.arg for u in result.toposort() if u.op is Ops.CONST and u.dtype == dtypes.float]
+    self.assertIn(3.14, const_vals)
+
 class TestUOpTags(unittest.TestCase):
   def test_inc_by_one(self):
     g = UOp.const(dtypes.int, 1) + UOp.const(dtypes.int, 1)
diff --git a/test/unit/test_uop_symbolic.py b/test/unit/test_uop_symbolic.py
index 4968c81b87..826de11643 100644
--- a/test/unit/test_uop_symbolic.py
+++ b/test/unit/test_uop_symbolic.py
@@ -1074,6 +1074,24 @@ class TestGatedUopGivenValid(unittest.TestCase):
     expected_vec = UOp(Ops.VECTORIZE, dtypes.index.vec(2), (uconst(0), r0))
     self.assertEqual(idx, (r0 < 3).where(expected_vec, UOp.invalid()))
 
+class TestRangeSplitting(unittest.TestCase):
+  def test_range_split_on_mod(self):
+    # test that mark_range_mod splits RANGE(8) into RANGE(4)*2 + RANGE(2) when used with %2
+    from tinygrad.codegen.simplify import pm_split_ranges, pm_flatten_range
+    r0 = UOp.range(uconst(8), 0)
+    # create a simple expression using the range with mod: store range%2 to a buffer
+    buf = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(), arg=0)
+    val = (r0 % uconst(2)).cast(dtypes.int)
+    store = UOp(Ops.STORE, dtypes.void, (buf.index(uconst(0)), val))
+    sink = UOp(Ops.SINK, dtypes.void, (UOp(Ops.END, dtypes.void, (store, r0)),))
+    # count RANGEs before
+    ranges_before = len([u for u in sink.toposort() if u.op is Ops.RANGE])
+    # apply the range splitting optimization
+    sink_after = graph_rewrite(sink, pm_split_ranges+pm_flatten_range, ctx={}, name="test split ranges")
+    # count RANGEs after - should have more due to splitting
+    ranges_after = len([u for u in sink_after.toposort() if u.op is Ops.RANGE])
+    self.assertGreater(ranges_after, ranges_before, "RANGE should be split when used with mod of divisible constant")
+
 class TestBounds(unittest.TestCase):
   def test_unrolled_arange(self):
     # #include <metal_stdlib>
diff --git a/tinygrad/codegen/__init__.py b/tinygrad/codegen/__init__.py
index 63667d7a2a..b3273b481f 100644
--- a/tinygrad/codegen/__init__.py
+++ b/tinygrad/codegen/__init__.py
@@ -115,8 +115,8 @@ pm_linearize_cleanups = PatternMatcher([
   # if statements are not allowed in the graph
   (UPat((Ops.IF, Ops.ENDIF)), lambda: panic(RuntimeError("if not allowed in graph"))),
   # gated INDEX becomes IF-STORE-ENDIF. this is the only use of IF-ENDIF
-  (UPat(Ops.STORE, name="u", src=(UPat(Ops.INDEX, src=(UPat(), UPat(), UPat(name="gate", dtype=dtypes.bool))).or_casted(), UPat()),
-        allow_any_len=True), lambda u, gate: (u, [mif:=UOp(Ops.IF, src=(gate, u.src[0])), u, UOp(Ops.ENDIF, src=(mif,))]))
+  (UPat(Ops.STORE, name="u", src=(UPat(Ops.INDEX, src=(UPat(), UPat(), UPat(name="gate", dtype=dtypes.bool))).or_casted(), UPat())),
+   lambda u, gate: (u, [mif:=UOp(Ops.IF, src=(gate, u.src[0])), u, UOp(Ops.ENDIF, src=(mif,))]))
 ])
 
 # requires lst be toposorted. like graph rewrite, but for lines
diff --git a/tinygrad/codegen/late/devectorizer.py b/tinygrad/codegen/late/devectorizer.py
index d10c37e176..dbf4fd3ae6 100644
--- a/tinygrad/codegen/late/devectorizer.py
+++ b/tinygrad/codegen/late/devectorizer.py
@@ -123,12 +123,12 @@ load_store_folding = PatternMatcher([
   (UPat(Ops.LOAD, src=(UPat(Ops.GEP, name="gep"),), name="ld", allow_any_len=True),
    lambda gep, ld: ld.replace(dtype=ld.dtype.scalar().vec(gep.dtype.count), src=(gep.src[0],)+ld.src[1:]).gep(gep.arg)),
   # GEP on data of STORE
-  (UPat(Ops.STORE, src=(UPat(Ops.GEP, name="gep"), UPat.var("st")), allow_any_len=True, name="sto"), gep_on_store),
+  (UPat(Ops.STORE, src=(UPat(Ops.GEP, name="gep"), UPat.var("st")), name="sto"), gep_on_store),
   # put PTRCAT after LOAD
   (UPat(Ops.LOAD, src=(UPat(Ops.PTRCAT, name="cat"),), name="ld", allow_any_len=True),
    lambda cat,ld: UOp(Ops.CAT, cat.dtype.base.vec(cat.dtype.vcount), tuple(ld.replace(dtype=x.dtype.base, src=(x,)+ld.src[1:]) for x in cat.src))),
   # put PTRCAT after STORE
-  (UPat(Ops.STORE, src=(UPat(Ops.PTRCAT, name="cat"), UPat(name="data")), allow_any_len=True, name="sto"), cat_after_store),
+  (UPat(Ops.STORE, src=(UPat(Ops.PTRCAT, name="cat"), UPat(name="data")), name="sto"), cat_after_store),
 ])
 
 # *** correct load/store ***
@@ -319,8 +319,7 @@ def reduce_to_acc(ctx:ReduceContext, red:UOp):
     input_ranges = tuple([x for x in topo if x.op is Ops.RANGE and x not in reduce_range and x not in ended_ranges])
     identity = red.const(red.dtype, identity_element(red.arg, red.dtype.scalar()))
     acc = UOp(Ops.DEFINE_REG, red.dtype.ptr(size=1, addrspace=AddrSpace.REG), arg=ctx.acc_num)
-    acc_init = acc.after(*input_ranges).index(UOp.const(dtypes.int, 0)).store(identity) if len(input_ranges) else \
-               acc.index(UOp.const(dtypes.int, 0)).store(identity)
+    acc_init = acc.after(*input_ranges).index(UOp.const(dtypes.int, 0)).store(identity)
     lst = [acc.after(acc_init, *reduce_range).index(UOp.const(dtypes.int, 0))] + lst  # put acc as the first element
     ctx.acc_num += 1
   ret = functools.reduce(lambda x,y: x.alu(red.arg, y), lst)
@@ -342,6 +341,6 @@ pm_add_loads = PatternMatcher([
   (UPat(Ops.INDEX, name="idx"), lambda idx: None if isinstance(idx.dtype, (PtrDType, ImageDType)) else
     idx.replace(dtype=idx.src[0].dtype).load(dtype=idx.dtype.base)),
   # remove loads from stores
-  (UPat(Ops.STORE, src=(UPat(Ops.LOAD),), allow_any_len=True, name="s"), lambda s: s.replace(src=(s.src[0].src[0],)+s.src[1:])),
+  (UPat(Ops.STORE, src=(UPat(Ops.LOAD), UPat(name="val")), name="s"), lambda s,val: s.replace(src=(s.src[0].src[0], val))),
 ])
 
diff --git a/tinygrad/helpers.py b/tinygrad/helpers.py
index 319d44ec8d..85c471afb3 100644
--- a/tinygrad/helpers.py
+++ b/tinygrad/helpers.py
@@ -206,6 +206,8 @@ ALLOW_TF32 = ContextVar("ALLOW_TF32", 0)
 SCACHE = ContextVar("SCACHE", 1)
 # allow use of atomics for embedding backward
 USE_ATOMICS = ContextVar("USE_ATOMICS", 0)
+# allow use of assembly for gemm
+ASM_GEMM = ContextVar("ASM_GEMM", 0)
 
 @dataclass(frozen=True)
 class Metadata:
diff --git a/tinygrad/nn/state.py b/tinygrad/nn/state.py
index 454cb2b9b3..3d078674a7 100644
--- a/tinygrad/nn/state.py
+++ b/tinygrad/nn/state.py
@@ -78,7 +78,7 @@ def safe_save(tensors:dict[str, Tensor], fn:str, metadata:dict[str, Any]|None=No
   j += "\x20"*(round_up(len(j),8)-len(j))
   pathlib.Path(fn).unlink(missing_ok=True)
   t = Tensor.empty(8+len(j)+offset, dtype=dtypes.uint8, device=f"disk:{fn}")
-  t[0:8].bitcast(dtypes.int64).assign([len(j)])
+  t[0:8].assign(Tensor([len(j)], dtype=dtypes.int64, device="CPU").bitcast(dtypes.uint8))
   t[8:8+len(j)].assign(list(j.encode('utf-8')))
   for k,v in safe_load(t).items(): v.assign(tensors[k])
 
diff --git a/tinygrad/renderer/cstyle.py b/tinygrad/renderer/cstyle.py
index 6d76411c61..e57f7e9801 100644
--- a/tinygrad/renderer/cstyle.py
+++ b/tinygrad/renderer/cstyle.py
@@ -46,10 +46,10 @@ base_rewrite = PatternMatcher([
   # new load/store
   (UPat(Ops.INDEX, src=(UPat.var("buf"), UPat.var('idx')), allow_any_len=True),
    lambda ctx,buf,idx: f"({ctx[buf]}+{strip_parens(ctx[idx]) if idx.arg == Ops.ADD else ctx[idx]})"),
-  (UPat(Ops.LOAD, src=(UPat(Ops.INDEX, src=(UPat(), UPat(), UPat.var("gate"))).or_casted("bidx"), UPat.var("var")), allow_any_len=True),
+  (UPat(Ops.LOAD, src=(UPat(Ops.INDEX, src=(UPat(), UPat(), UPat.var("gate"))).or_casted("bidx"), UPat.var("var"))),
    lambda ctx,bidx,var,gate: f"({ctx[gate]}?*{ctx[bidx]}:{ctx[var]})"),
-  (UPat(Ops.LOAD, src=(UPat.var('bidx'),), allow_any_len=True), lambda ctx,bidx: f"(*{ctx[bidx]})"),
-  (UPat(Ops.STORE, src=(UPat.var('bidx'), UPat.var("var")), allow_any_len=True), lambda ctx,bidx,var: f"*{ctx[bidx]} = {ctx[var]};"),
+  (UPat(Ops.LOAD, src=(UPat.var('bidx'),)), lambda ctx,bidx: f"(*{ctx[bidx]})"),
+  (UPat(Ops.STORE, src=(UPat.var('bidx'), UPat.var("var"))), lambda ctx,bidx,var: f"*{ctx[bidx]} = {ctx[var]};"),
   # alu/gep
   # TODO: look for left-associative
   (UPat(GroupOp.ALU, name="x"), lambda ctx,x: ctx.code_for_op[x.op](
diff --git a/tinygrad/runtime/ops_metal.py b/tinygrad/runtime/ops_metal.py
index 1268ae705c..b956890bcc 100644
--- a/tinygrad/runtime/ops_metal.py
+++ b/tinygrad/runtime/ops_metal.py
@@ -4,12 +4,13 @@ import tinygrad.runtime.support.objc as objc
 from tinygrad.device import Compiled, Compiler, CompileError, LRUAllocator, ProfileDeviceEvent, CompilerSet, CompilerPair
 from tinygrad.renderer.cstyle import MetalRenderer
 from tinygrad.runtime.autogen import metal
+from tinygrad.runtime.support.c import DLL
 
 # 13 is requestType that metal uses to compile source code into MTLB, there aren't any docs or symbols.
 REQUEST_TYPE_COMPILE = 13
 
 # Must be loaded for default Metal Device: https://developer.apple.com/documentation/metal/1433401-mtlcreatesystemdefaultdevice?language=objc
-ctypes.CDLL("/System/Library/Frameworks/CoreGraphics.framework/CoreGraphics")
+DLL("CoreGraphics", "CoreGraphics")
 
 # FIXME: these need autogen to support objc categories
 # https://developer.apple.com/library/archive/documentation/Cocoa/Conceptual/ObjectiveC/Chapters/ocCategories.html
@@ -67,7 +68,7 @@ class MetalCompiler(Compiler):
   # doesn't seem to be anything we can do.
   with contextlib.suppress(FileNotFoundError, ModuleNotFoundError):
     import tinygrad.runtime.autogen.llvm # noqa: F401
-  support = ctypes.CDLL("/System/Library/PrivateFrameworks/MTLCompiler.framework/MTLCompiler")
+  support = DLL("MTLCompiler", "MTLCompiler")
   support.MTLCodeGenServiceCreate.restype = ctypes.c_void_p
 
   def __init__(self):
diff --git a/tinygrad/runtime/support/amd.py b/tinygrad/runtime/support/amd.py
index 7554e8dda2..dd121d0b7f 100644
--- a/tinygrad/runtime/support/amd.py
+++ b/tinygrad/runtime/support/amd.py
@@ -88,8 +88,8 @@ def import_asic_regs(prefix:str, version:tuple[int, ...], cls=AMDReg) -> dict[st
     return x
   def _download_file(ver, suff) -> str:
     dir_prefix = {"osssys": "oss"}.get(prefix, prefix)
-    fetch_name, file_name = f"{prefix}_{'_'.join(map(str, ver))}_{suff}.h", f"{prefix}_{'_'.join(map(str, version))}_{suff}.h"
-    return header_download(f"include/asic_reg/{dir_prefix}/{fetch_name}", name=file_name, subdir="asic_regs")
+    fetch_name = f"{prefix}_{'_'.join(map(str, ver))}_{suff}.h"
+    return header_download(f"include/asic_reg/{dir_prefix}/{fetch_name}", name=fetch_name, subdir="asic_regs")
 
   for ver in fixup_ip_version(prefix, version):
     try: offs, sh_masks = _extract_regs(_download_file(ver, "offset")), _extract_regs(_download_file(ver, "sh_mask"))
diff --git a/tinygrad/runtime/support/c.py b/tinygrad/runtime/support/c.py
index b02143dada..b6db3c13fd 100644
--- a/tinygrad/runtime/support/c.py
+++ b/tinygrad/runtime/support/c.py
@@ -131,13 +131,15 @@ def init_c_struct_t(sz:int, fields: tuple[tuple, ...]):
 def init_c_var(ty, creat_cb): return (creat_cb(v:=del_an(ty)()), v)[1]
 
 class DLL(ctypes.CDLL):
+  _loaded_: set[str] = set()
+
   @staticmethod
   def findlib(nm:str, paths:list[str], extra_paths=[]):
     if nm == 'libc' and OSX: return '/usr/lib/libc.dylib'
     if pathlib.Path(path:=getenv(nm.replace('-', '_').upper()+"_PATH", '')).is_file(): return path
     for p in paths:
       libpaths = {"posix": ["/usr/lib64", "/usr/lib", "/usr/local/lib"], "nt": os.environ['PATH'].split(os.pathsep),
-                  "darwin": ["/opt/homebrew/lib", f"/System/Library/Frameworks/{p}.framework"],
+                  "darwin": ["/opt/homebrew/lib", f"/System/Library/Frameworks/{p}.framework", f"/System/Library/PrivateFrameworks/{p}.framework"],
                   'linux': ['/lib', '/lib64', f"/lib/{sysconfig.get_config_var('MULTIARCH')}", "/usr/lib/wsl/lib/"]}
       if (pth:=pathlib.Path(p)).is_absolute():
         if pth.is_file(): return p
@@ -154,12 +156,12 @@ class DLL(ctypes.CDLL):
               if f.read(4) == b'\x7FELF': return str(l)
 
   def __init__(self, nm:str, paths:str|list[str], extra_paths=[], emsg="", **kwargs):
-    self.nm, self.emsg, self.loaded = nm, emsg, False
+    self.nm, self.emsg = nm, emsg
     if (path:= DLL.findlib(nm, paths if isinstance(paths, list) else [paths], extra_paths if isinstance(extra_paths, list) else [extra_paths])):
       if DEBUG >= 3: print(f"loading {nm} from {path}")
       try:
         super().__init__(path, **kwargs)
-        self.loaded = True
+        self._loaded_.add(self.nm)
       except OSError as e:
         self.emsg = str(e)
         if DEBUG >= 3: print(f"loading {nm} failed: {e}")
@@ -175,5 +177,6 @@ class DLL(ctypes.CDLL):
     return wrapper
 
   def __getattr__(self, nm):
-    if not self.loaded: raise AttributeError(f"failed to load library {self.nm}: " + (self.emsg or f"try setting {self.nm.upper()+'_PATH'}?"))
+    if self.nm not in self._loaded_:
+      raise AttributeError(f"failed to load library {self.nm}: " + (self.emsg or f"try setting {self.nm.upper()+'_PATH'}?"))
     return super().__getattr__(nm)
diff --git a/tinygrad/runtime/support/usb.py b/tinygrad/runtime/support/usb.py
index 2340c944cb..a227f5a35b 100644
--- a/tinygrad/runtime/support/usb.py
+++ b/tinygrad/runtime/support/usb.py
@@ -5,10 +5,10 @@ from tinygrad.helpers import DEBUG, to_mv, round_up, OSX
 from tinygrad.runtime.support.hcq import MMIOInterface
 
 class USB3:
-  def __init__(self, vendor:int, dev:int, ep_data_in:int, ep_stat_in:int, ep_data_out:int, ep_cmd_out:int, max_streams:int=31):
+  def __init__(self, vendor:int, dev:int, ep_data_in:int, ep_stat_in:int, ep_data_out:int, ep_cmd_out:int, max_streams:int=31, use_bot=False):
     self.vendor, self.dev = vendor, dev
     self.ep_data_in, self.ep_stat_in, self.ep_data_out, self.ep_cmd_out = ep_data_in, ep_stat_in, ep_data_out, ep_cmd_out
-    self.max_streams = max_streams
+    self.max_streams, self.use_bot = max_streams, use_bot
     self.ctx = ctypes.POINTER(libusb.struct_libusb_context)()
 
     if libusb.libusb_init(ctypes.byref(self.ctx)): raise RuntimeError("libusb_init failed")
@@ -25,30 +25,34 @@ class USB3:
     # Set configuration and claim interface
     if libusb.libusb_set_configuration(self.handle, 1): raise RuntimeError("set_configuration failed")
     if libusb.libusb_claim_interface(self.handle, 0): raise RuntimeError("claim_interface failed. sudo required?")
-    if libusb.libusb_set_interface_alt_setting(self.handle, 0, 1): raise RuntimeError("alt_setting failed")
 
-    # Clear any stalled endpoints
-    all_eps = (self.ep_data_out, self.ep_data_in, self.ep_stat_in, self.ep_cmd_out)
-    for ep in all_eps: libusb.libusb_clear_halt(self.handle, ep)
+    if use_bot:
+      self._tag = 0
+    else:
+      if libusb.libusb_set_interface_alt_setting(self.handle, 0, 1): raise RuntimeError("alt_setting failed")
 
-    # Allocate streams
-    stream_eps = (ctypes.c_uint8 * 3)(self.ep_data_out, self.ep_data_in, self.ep_stat_in)
-    if (rc:=libusb.libusb_alloc_streams(self.handle, self.max_streams * len(stream_eps), stream_eps, len(stream_eps))) < 0:
-      raise RuntimeError(f"alloc_streams failed: {rc}")
+      # Clear any stalled endpoints
+      all_eps = (self.ep_data_out, self.ep_data_in, self.ep_stat_in, self.ep_cmd_out)
+      for ep in all_eps: libusb.libusb_clear_halt(self.handle, ep)
 
-    # Base cmd
-    cmd_template = bytes([0x01, 0x00, 0x00, 0x01, *([0] * 12), 0xE4, 0x24, 0x00, 0xB2, 0x1A, 0x00, 0x00, 0x00, *([0] * 8)])
+      # Allocate streams
+      stream_eps = (ctypes.c_uint8 * 3)(self.ep_data_out, self.ep_data_in, self.ep_stat_in)
+      if (rc:=libusb.libusb_alloc_streams(self.handle, self.max_streams * len(stream_eps), stream_eps, len(stream_eps))) < 0:
+        raise RuntimeError(f"alloc_streams failed: {rc}")
 
-    # Init pools
-    self.tr = {ep: [libusb.libusb_alloc_transfer(0) for _ in range(self.max_streams)] for ep in all_eps}
+      # Base cmd
+      cmd_template = bytes([0x01, 0x00, 0x00, 0x01, *([0] * 12), 0xE4, 0x24, 0x00, 0xB2, 0x1A, 0x00, 0x00, 0x00, *([0] * 8)])
 
-    self.buf_cmd = [(ctypes.c_uint8 * len(cmd_template))(*cmd_template) for _ in range(self.max_streams)]
-    self.buf_stat = [(ctypes.c_uint8 * 64)() for _ in range(self.max_streams)]
-    self.buf_data_in = [(ctypes.c_uint8 * 0x1000)() for _ in range(self.max_streams)]
-    self.buf_data_out = [(ctypes.c_uint8 * 0x80000)() for _ in range(self.max_streams)]
-    self.buf_data_out_mvs = [to_mv(ctypes.addressof(self.buf_data_out[i]), 0x80000) for i in range(self.max_streams)]
+      # Init pools
+      self.tr = {ep: [libusb.libusb_alloc_transfer(0) for _ in range(self.max_streams)] for ep in all_eps}
 
-    for slot in range(self.max_streams): struct.pack_into(">B", self.buf_cmd[slot], 3, slot + 1)
+      self.buf_cmd = [(ctypes.c_uint8 * len(cmd_template))(*cmd_template) for _ in range(self.max_streams)]
+      self.buf_stat = [(ctypes.c_uint8 * 64)() for _ in range(self.max_streams)]
+      self.buf_data_in = [(ctypes.c_uint8 * 0x1000)() for _ in range(self.max_streams)]
+      self.buf_data_out = [(ctypes.c_uint8 * 0x80000)() for _ in range(self.max_streams)]
+      self.buf_data_out_mvs = [to_mv(ctypes.addressof(self.buf_data_out[i]), 0x80000) for i in range(self.max_streams)]
+
+      for slot in range(self.max_streams): struct.pack_into(">B", self.buf_cmd[slot], 3, slot + 1)
 
   def _prep_transfer(self, tr, ep, stream_id, buf, length):
     tr.contents.dev_handle, tr.contents.endpoint, tr.contents.length, tr.contents.buffer = self.handle, ep, length, buf
@@ -68,38 +72,90 @@ class USB3:
         if tr.contents.status == libusb.LIBUSB_TRANSFER_COMPLETED: running -= 1
         elif tr.contents.status != 0xFF: raise RuntimeError(f"EP 0x{tr.contents.endpoint:02X} error: {tr.contents.status}")
 
+  def _bulk_out(self, ep: int, payload: bytes, timeout: int = 1000):
+    transferred = ctypes.c_int(0)
+    rc = libusb.libusb_bulk_transfer(
+      self.handle,
+      ep,
+      (ctypes.c_ubyte * len(payload))(*payload),
+      len(payload),
+      ctypes.byref(transferred),
+      timeout,
+    )
+    assert rc == 0, f"bulk OUT 0x{ep:02X} failed: {rc}"
+    assert transferred.value == len(payload), f"bulk OUT short write on 0x{ep:02X}: {transferred.value}/{len(payload)} bytes"
+
+  def _bulk_in(self, ep: int, length: int, timeout: int = 1000) -> bytes:
+    buf, transferred = (ctypes.c_ubyte * length)(), ctypes.c_int(0)
+    rc = libusb.libusb_bulk_transfer(
+      self.handle,
+      ep,
+      buf,
+      length,
+      ctypes.byref(transferred),
+      timeout,
+    )
+    assert rc == 0, f"bulk IN 0x{ep:02X} failed: {rc}"
+    return bytes(buf[:transferred.value])
+
   def send_batch(self, cdbs:list[bytes], idata:list[int]|None=None, odata:list[bytes|None]|None=None) -> list[bytes|None]:
     idata, odata = idata or [0] * len(cdbs), odata or [None] * len(cdbs)
-    results, tr_window, op_window = [], [], []
+    results:list[bytes|None] = []
+    tr_window, op_window = [], []
 
     for idx, (cdb, rlen, send_data) in enumerate(zip(cdbs, idata, odata)):
-      # allocate slot and stream. stream is 1-based
-      slot, stream = idx % self.max_streams, (idx % self.max_streams) + 1
+      if self.use_bot:
+        dir_in = rlen > 0
+        data_len = rlen if dir_in else (len(send_data) if send_data is not None else 0)
+        assert (data_len == 0) if dir_in else (rlen == 0), "BOT mode only supports either read or write per command"
 
-      # build cmd packet
-      self.buf_cmd[slot][16:16+len(cdb)] = list(cdb)
+        # CBW
+        self._tag += 1
+        flags = 0x80 if dir_in else 0x00
+        cbw = struct.pack("<IIIBBB", 0x43425355, self._tag, data_len, flags, 0, len(cdb)) + cdb + b"\x00" * (16 - len(cdb))
+        self._bulk_out(self.ep_data_out, cbw)
 
-      # cmd + stat transfers
-      tr_window.append(self._prep_transfer(self.tr[self.ep_cmd_out][slot], self.ep_cmd_out, None, self.buf_cmd[slot], len(self.buf_cmd[slot])))
-      tr_window.append(self._prep_transfer(self.tr[self.ep_stat_in][slot], self.ep_stat_in, stream, self.buf_stat[slot], 64))
+        # DAT
+        if dir_in:
+          results.append(self._bulk_in(self.ep_data_in, rlen))
+        else:
+          if send_data is not None:
+            self._bulk_out(self.ep_data_out, send_data)
+          results.append(None)
 
-      if rlen:
-        if rlen > len(self.buf_data_in[slot]): self.buf_data_in[slot] = (ctypes.c_uint8 * round_up(rlen, 0x1000))()
-        tr_window.append(self._prep_transfer(self.tr[self.ep_data_in][slot], self.ep_data_in, stream, self.buf_data_in[slot], rlen))
+        # CSW
+        sig, rtag, residue, status = struct.unpack("<IIIB", self._bulk_in(self.ep_data_in, 13, timeout=2000))
+        assert sig == 0x53425355, f"Bad CSW signature 0x{sig:08X}, expected 0x53425355"
+        assert rtag == self._tag, f"CSW tag mismatch: got {rtag}, expected {self._tag}"
+        assert status == 0, f"SCSI command failed, CSW status=0x{status:02X}, residue={residue}"
+      else:
+        # allocate slot and stream. stream is 1-based
+        slot, stream = idx % self.max_streams, (idx % self.max_streams) + 1
 
-      if send_data is not None:
-        if len(send_data) > len(self.buf_data_out[slot]):
-          self.buf_data_out[slot] = (ctypes.c_uint8 * len(send_data))()
-          self.buf_data_out_mvs[slot] = to_mv(ctypes.addressof(self.buf_data_out[slot]), len(send_data))
+        # build cmd packet
+        self.buf_cmd[slot][16:16+len(cdb)] = list(cdb)
 
-        self.buf_data_out_mvs[slot][:len(send_data)] = bytes(send_data)
-        tr_window.append(self._prep_transfer(self.tr[self.ep_data_out][slot], self.ep_data_out, stream, self.buf_data_out[slot], len(send_data)))
+        # cmd + stat transfers
+        tr_window.append(self._prep_transfer(self.tr[self.ep_cmd_out][slot], self.ep_cmd_out, None, self.buf_cmd[slot], len(self.buf_cmd[slot])))
+        tr_window.append(self._prep_transfer(self.tr[self.ep_stat_in][slot], self.ep_stat_in, stream, self.buf_stat[slot], 64))
 
-      op_window.append((idx, slot, rlen))
-      if (idx + 1 == len(cdbs)) or len(op_window) >= self.max_streams:
-        self._submit_and_wait(tr_window)
-        for idx, slot, rlen in op_window: results.append(bytes(self.buf_data_in[slot][:rlen]) if rlen else None)
-        tr_window = []
+        if rlen:
+          if rlen > len(self.buf_data_in[slot]): self.buf_data_in[slot] = (ctypes.c_uint8 * round_up(rlen, 0x1000))()
+          tr_window.append(self._prep_transfer(self.tr[self.ep_data_in][slot], self.ep_data_in, stream, self.buf_data_in[slot], rlen))
+
+        if send_data is not None:
+          if len(send_data) > len(self.buf_data_out[slot]):
+            self.buf_data_out[slot] = (ctypes.c_uint8 * len(send_data))()
+            self.buf_data_out_mvs[slot] = to_mv(ctypes.addressof(self.buf_data_out[slot]), len(send_data))
+
+          self.buf_data_out_mvs[slot][:len(send_data)] = bytes(send_data)
+          tr_window.append(self._prep_transfer(self.tr[self.ep_data_out][slot], self.ep_data_out, stream, self.buf_data_out[slot], len(send_data)))
+
+        op_window.append((idx, slot, rlen))
+        if (idx + 1 == len(cdbs)) or len(op_window) >= self.max_streams:
+          self._submit_and_wait(tr_window)
+          for idx, slot, rlen in op_window: results.append(bytes(self.buf_data_in[slot][:rlen]) if rlen else None)
+          tr_window = []
 
     return results
 
diff --git a/tinygrad/schedule/multi.py b/tinygrad/schedule/multi.py
index b6a502c845..5539967eb2 100644
--- a/tinygrad/schedule/multi.py
+++ b/tinygrad/schedule/multi.py
@@ -71,7 +71,8 @@ def handle_allreduce(buf:UOp, red:UOp) -> UOp|None:
   # allgather
   copied_chunks = []
   for i,rc in enumerate(reduced_chunks):
-    if use_all2all: copied_chunks.append(UOp(Ops.MSTACK, buf.dtype, tuple(rc.copy_to_device(buf.device[j]) for j in range(n_lbs))))
+    if isinstance(red.src[1].arg, str): copied_chunks.append(rc.copy_to_device(red.src[1].arg))
+    elif use_all2all: copied_chunks.append(UOp(Ops.MSTACK, buf.dtype, tuple(rc.copy_to_device(buf.device[j]) for j in range(n_lbs))))
     else:
       this_chunk: list[UOp|None] = [None] * n_lbs
       this_chunk[(i+n_lbs-1)%n_lbs] = rc
@@ -93,12 +94,7 @@ def mstack_early_shrink(ms:UOp, shrink:UOp):
     return s.shrink(tuple(new_arg))
   for i, x in enumerate(ms.src):
     if x.op is Ops.COPY:
-      # if src device doesn't have a renderer, we have to view after the copy
-      # TODO: a way to understand this
-      if x.src[0].device in {"DISK", "NPY"}:
-        ret.append(apply_shrink(x, i))
-      else:
-        ret.append(apply_shrink(x.src[0], i).copy_to_device(x.device))
+      ret.append(apply_shrink(x.src[0], i).copy_to_device(x.device))
     else:
       ret.append(apply_shrink(x, i).contiguous())
   return ms.replace(src=tuple(ret))
diff --git a/tinygrad/schedule/rangeify.py b/tinygrad/schedule/rangeify.py
index 25d032f735..11ac054db3 100644
--- a/tinygrad/schedule/rangeify.py
+++ b/tinygrad/schedule/rangeify.py
@@ -60,7 +60,7 @@ def split_reduceop(reduce:UOp, x:UOp):
 
 mop_cleanup = PatternMatcher([
   # merge adjacent RESHAPES, safe because they are not tagged
-  (UPat(Ops.RESHAPE, name="x2").f(Ops.RESHAPE, allow_any_len=True, name="x"),
+  (UPat(Ops.RESHAPE, src=(UPat(Ops.RESHAPE, name="x2"), UPat()), name="x"),
    lambda x,x2: x.replace(src=(x2.src[0], x.src[1])) if x.tag is None and x2.tag is None else None),
 ])
 
@@ -90,7 +90,7 @@ earliest_rewrites = mop_cleanup+PatternMatcher([
   (UPat(Ops.CUSTOM_KERNEL, name="ck"), resolve_custom_kernel),
 
   # remove CONTIGUOUS if the BUFFER is already contiguous
-  (UPat(Ops.BUFFER).f(Ops.RESHAPE, allow_any_len=True, name="r").f(Ops.CONTIGUOUS, name="c"), lambda r,c: r.replace(tag=c.tag)),
+  (UPat(Ops.RESHAPE, src=(UPat(Ops.BUFFER), UPat()), name="r").f(Ops.CONTIGUOUS, name="c"), lambda r,c: r.replace(tag=c.tag)),
 
   # split_reduceop
   (UPat(Ops.REDUCE_AXIS, name="reduce", src=(UPat.var("x"),)), split_reduceop),
@@ -126,6 +126,10 @@ earliest_rewrites = mop_cleanup+PatternMatcher([
 
   # ** assign rules **
 
+  # move bitcast from assign target to source: a.bitcast(X).assign(src) -> a.assign(src.bitcast(a.dtype))
+  (UPat(Ops.ASSIGN, src=(UPat(Ops.BITCAST, src=(UPat(name="target"),)), UPat(name="src")), name="assign"),
+   lambda target, src, assign: target.assign(src.bitcast(target.dtype)).replace(tag=assign.tag)),
+
   # assign only to buffer, otherwise make it a CONTIGUOUS
   (UPat(Ops.ASSIGN, src=(UPat(GroupOp.All-{Ops.BUFFER}, name="target"), UPat(name="x")), name="assign"),
    lambda x,target,assign: x.f(Ops.CONTIGUOUS, tag=assign.tag) if ((t:=target.base).op is not Ops.BUFFER and \
@@ -247,7 +251,7 @@ pm_const_buffer_folding = pm_mops+PatternMatcher([
   # dont bufferize an arange
   (UPat.any((r:=UPat(dtype=dtypes.index).cast()).named("src"), r.eq(UPat()).named("src")).f(Ops.BUFFERIZE,
     allow_any_len=True, name="buf").f(Ops.INDEX, allow_any_len=True, name="idx"), remove_bufferize),
-  # no buffers for const
+  # no buffers for const (ranges don't matter for const - it's the same value everywhere)
   (UPat(Ops.CONST, name='c').f(Ops.BUFFERIZE, allow_any_len=True, name="b"), lambda c,b: b.const_like(c.arg).rtag(b.tag)),
   # indexing a const is a const
   (UPat(Ops.INDEX, src=(UPat(Ops.CONST, name="c"),),), lambda c: c),
@@ -271,20 +275,20 @@ def late_buffer_view(t:UOp, b:UOp):
   size = prod(shape)
 
   # walk up for the INDEX
+  # NOTE: even though we allow RESHAPE and SHRINK, they can combine to form non-contiguous access patterns (e.g. t[::2])
   x = t
-  while not any(u.op is Ops.INDEX for u in x.src):
-    assert x.op not in GroupOp.Elementwise, "can't buffer view elementwise"
+  while x.op is not Ops.INDEX:
+    assert x.op in {Ops.BITCAST, Ops.CONTIGUOUS, Ops.SHRINK, Ops.RESHAPE}, f"unexpected op {x.op} in buffer view walk"
     x = x.src[0]
-  x = next(u for u in x.src if u.op is Ops.INDEX)
 
   if len(shape) == 0: offset = x.src[1].arg
-  else: offset = max(sum(idx.vmin for idx in x.src[1:]), 0)
+  else: offset = sum(idx.vmin for idx in x.src[1:])
+  if offset < 0: raise RuntimeError(f"negative offset {offset} in buffer view")
 
-  return b.replace(src=(UOp(Ops.BUFFER_VIEW, t.dtype, (x.base,), (size, offset), tag=t.tag),) + b.src[1:])
+  return b.replace(src=(UOp(Ops.BUFFER_VIEW, t.dtype, (x.base,), (size, offset), tag=t.tag), b.src[1]))
 
 to_bufferview = PatternMatcher([
-  (UPat((Ops.BITCAST, Ops.CONTIGUOUS), name="t").f(Ops.BUFFERIZE, allow_any_len=True, name="b"), late_buffer_view),
-  (UPat((Ops.BITCAST, Ops.CONTIGUOUS)).f(Ops.BUFFER_VIEW, name="b"), lambda b: b.replace(src=b.src[0].src)),
+  (UPat(Ops.BUFFERIZE, src=(UPat((Ops.BITCAST, Ops.CONTIGUOUS), name="t"), UPat()), name="b"), late_buffer_view),
 ])
 
 DEVICE_MAX_BUFS = {"METAL": 31, "WEBGPU": 8} # TODO: get from device?
@@ -451,9 +455,6 @@ to_define_global = PatternMatcher([
   # this is only needed if you are using symbolic
   (UPat((Ops.CONST, Ops.DEFINE_VAR), name="c"), lambda c: c.replace(src=()) if len(c.src) else None),
 
-  # remove RANGE with 0 size
-  (UPat(Ops.RANGE, name="r"), lambda r: UOp.const(dtypes.index, 0) if r.vmax == 0 else None),
-
   # renumber the ranges starting with 0 so that kernel deduping works
   (UPat(Ops.RANGE, name="r"), renumber_range),
 ])
@@ -469,9 +470,6 @@ rangeify_codegen = PatternMatcher([
   # TODO: this can be moved into codegen?
   (UPat(Ops.NOOP, name="x"), lambda x: x.src[0]),
 
-  # strip the arg from store
-  (UPat(Ops.STORE, name="x"), lambda x: x.replace(arg=None) if x.arg is not None else None),
-
   # add loads to non ptr indexes
   # TODO: this can be moved into codegen?
   #(UPat.any(UPat(Ops.DEFINE_GLOBAL, name="dg"), UPat(Ops.DEFINE_LOCAL).f(Ops.AFTER, allow_any_len=True, name="dg"))
diff --git a/tinygrad/tensor.py b/tinygrad/tensor.py
index 59a7b72c74..aa45bdfa6e 100644
--- a/tinygrad/tensor.py
+++ b/tinygrad/tensor.py
@@ -7,7 +7,7 @@ if TYPE_CHECKING: import numpy
 from tinygrad.dtype import DType, DTypeLike, dtypes, ImageDType, ConstType, least_upper_float, least_upper_dtype, sum_acc_dtype, to_dtype, truncate
 from tinygrad.dtype import _from_np_dtype, _to_np_dtype, PyConst
 from tinygrad.helpers import argfix, make_tuple, flatten, prod, all_int, round_up, merge_dicts, argsort, getenv, all_same, fully_flatten
-from tinygrad.helpers import IMAGE, WINO, Metadata, TRACEMETA, ceildiv, fetch, polyN, is_numpy_ndarray, TracingKey, cpu_profile
+from tinygrad.helpers import IMAGE, WINO, Metadata, TRACEMETA, ASM_GEMM, ceildiv, fetch, polyN, is_numpy_ndarray, TracingKey, cpu_profile
 from tinygrad.helpers import suppress_finalizing, disable_gc
 from tinygrad.gradient import compute_gradient
 from tinygrad.mixin import OpMixin
@@ -286,20 +286,20 @@ class Tensor(OpMixin):
     return self
 
   def assign(self, x:Tensor|PyConst|list|tuple) -> Tensor:
+    is_disk = isinstance(self.device, str) and self.device.startswith("DISK")
+    if not isinstance(x, Tensor): x = Tensor(x, device="CPU" if is_disk else self.device, dtype=self.dtype)
+    if self.uop is x.uop: return self  # a self assign is a NOOP
+    # broadcast x (shape only, dtype must match)
+    if self.shape != x.shape: x = x._broadcast_to(self.shape)
+    if self.shape != x.shape: raise RuntimeError(f"assign shape mismatch {self.shape} != {x.shape}")
+    if not is_disk and self.device != x.device: raise RuntimeError(f"assign device mismatch {self.device} != {x.device}")
+    if self.dtype != x.dtype: raise RuntimeError(f"assign dtype mismatch {self.dtype} != {x.dtype}")
+    if isinstance(self.device, tuple) and self.uop.axis != x.uop.axis: raise RuntimeError(f"multi axis mismatch {self.uop.axis} != {x.uop.axis}")
+
     # TODO: this is a hack for writing to DISK. remove with working assign
-    if isinstance(self.device, str) and self.device.startswith("DISK"):
-      if not isinstance(x, Tensor): x = Tensor(x, device="CPU", dtype=self.dtype)
+    if is_disk:
       self._buffer().copyin(x._data())
       return self
-    if not isinstance(x, Tensor): x = Tensor(x, device=self.device, dtype=self.dtype)
-    if self.uop is x.uop: return self  # a self assign is a NOOP
-    # NOTE: we allow cross device assign
-    # broadcast x
-    if least_upper_dtype(self.dtype, x.dtype) == self.dtype: x = x._broadcast_to(self.shape).cast(self.dtype)
-    assert self.shape == x.shape, f"assign shape mismatch {self.shape} != {x.shape}"
-    assert self.device == x.device, f"assign device mismatch {self.device} != {x.device}"
-    assert self.dtype == x.dtype, f"assign dtype mismatch {self.dtype} != {x.dtype}"
-    assert not isinstance(self.device, tuple) or self.uop.axis == x.uop.axis, f"multi assign axis mismatch {self.uop.axis} != {x.uop.axis}"
     return self.replace(self._apply_uop(UOp.assign, x))
 
   def detach(self) -> Tensor:
@@ -2431,6 +2431,9 @@ class Tensor(OpMixin):
     ```
     """
     if IMAGE: return self.image_dot(w, dtype)
+    if ASM_GEMM:
+      from extra.gemm.asm.cdna.gemm import can_use_asm_gemm, asm_gemm
+      if can_use_asm_gemm(self, w): return asm_gemm(self, w)
     x, dx, dw = self, self.ndim, w.ndim
     if not (dx > 0 and dw > 0): raise RuntimeError(f"both tensors need to be at least 1D, got {dx}D and {dw}D")
     if x.shape[-1] != w.shape[axis_w:=-min(w.ndim,2)]: raise RuntimeError(f"cannot dot {x.shape} and {w.shape}")
@@ -3865,7 +3868,10 @@ class Tensor(OpMixin):
 
   def bitcast(self, dtype:DTypeLike) -> Tensor:
     """
-    Bitcasts `self` to the given `dtype` of the same itemsize.
+    Bitcasts `self` to the given `dtype`.
+
+    When the target dtype has the same itemsize, this is a view of the same memory.
+    When itemsizes differ, the last dimension is adjusted and a new Tensor is created.
 
     `self` must not require a gradient.
 
diff --git a/tinygrad/uop/__init__.py b/tinygrad/uop/__init__.py
index 9641fd6382..4f792ce9ea 100644
--- a/tinygrad/uop/__init__.py
+++ b/tinygrad/uop/__init__.py
@@ -14,7 +14,7 @@ class Ops(FastEnum):
   # ** 1 -- defines/special **
 
   # define GLOBAL/VAR are ptrs to outside the Kernel
-  DEFINE_GLOBAL = auto(); DEFINE_VAR = auto(); BIND = auto()
+  DEFINE_VAR = auto(); BIND = auto()
 
   # this is a RANGE for GPU dimensions, similar to symbolic shapes but not exactly
   SPECIAL = auto()
@@ -28,6 +28,9 @@ class Ops(FastEnum):
   NOOP = auto(); REWRITE_ERROR = auto()
   PARAM = auto(); CALL = auto()
 
+  # TODO: remove this alias, DEFINE_GLOBAL is PARAM now
+  DEFINE_GLOBAL = PARAM
+
   # renderer
   # LINEAR is a list of UOps, SOURCE has a str arg that's human readable, BINARY has bytes arg that's compiled
   PROGRAM = auto(); LINEAR = auto(); SOURCE = auto(); BINARY = auto()
diff --git a/tinygrad/uop/ops.py b/tinygrad/uop/ops.py
index 12611f8667..6c5ed761be 100644
--- a/tinygrad/uop/ops.py
+++ b/tinygrad/uop/ops.py
@@ -206,7 +206,7 @@ class UOp(OpMixin, metaclass=UOpMetaClass):
     match self.op:
       # late ops don't have shape
       case Ops.UNIQUE | Ops.LUNIQUE | Ops.DEVICE | Ops.RANGE | Ops.LOAD | Ops.IF | Ops.BARRIER | Ops.CUSTOM | Ops.CUSTOMI | \
-           Ops.VECTORIZE | Ops.VCONST | Ops.GEP | Ops.SPECIAL | Ops.UNROLL | Ops.CONTRACT | Ops.CUSTOM_KERNEL | \
+           Ops.VECTORIZE | Ops.VCONST | Ops.GEP | Ops.SPECIAL | Ops.UNROLL | Ops.CONTRACT | Ops.CUSTOM_KERNEL | Ops.SINK | \
            Ops.LINEAR | Ops.PROGRAM | Ops.SOURCE | Ops.BINARY:
         return None
 
@@ -224,8 +224,9 @@ class UOp(OpMixin, metaclass=UOpMetaClass):
       case Ops.BUFFER_VIEW: return (self.arg[0],)
       case Ops.ENCDEC: return self.arg[0]
       case Ops.BUFFERIZE: return tuple([int(r.vmax+1) for r in self.src[1:]])
-      case Ops.DEFINE_GLOBAL | Ops.DEFINE_LOCAL | Ops.DEFINE_REG: return (self.ptrdtype.size,)
+      case Ops.DEFINE_LOCAL | Ops.DEFINE_REG: return (self.ptrdtype.size,)
       case Ops.PARAM:
+        if isinstance(self.dtype, PtrDType): return (self.ptrdtype.size,)
         # NOTE: copied from marg
         if len(self.src) >= 1: return tuple(self.src[0].sgep(i) for i in range(self.src[0].dtype.count))
         return None
@@ -413,10 +414,8 @@ class UOp(OpMixin, metaclass=UOpMetaClass):
   def load(self, *src:UOp, **kwargs): return UOp(Ops.LOAD, dtype=kwargs.pop("dtype", self.dtype.base), src=(self,)+src, **kwargs)
   def store(self, src:UOp|ConstType, **kwargs):
     return UOp(Ops.STORE, kwargs.pop("dtype", dtypes.void), (self, UOp.const(self.dtype, src) if not isinstance(src, UOp) else src), **kwargs)
-  def end(self, *src:UOp):
-    if len(src) == 0: return self
-    return UOp(Ops.END, src=(self,)+src)
-  def after(self, *src:UOp, **kwargs): return UOp(Ops.AFTER, self.dtype, (self,)+src, **kwargs)
+  def end(self, *src:UOp): return UOp(Ops.END, src=(self,)+src) if len(src) else self
+  def after(self, *src:UOp, **kwargs): return UOp(Ops.AFTER, self.dtype, (self,)+src, **kwargs) if len(src) else self
   def assign(self, x:UOp): return UOp(Ops.ASSIGN, self.dtype, (self, x))
   def barrier(self, *src:UOp): return UOp(Ops.BARRIER, src=(self,)+src)
   def contract(self, *rngs:UOp):
@@ -1436,6 +1435,7 @@ def pyrender(ast:UOp) -> str:
       for s in u.src: to_render.add(s)
     if u.op is Ops.STORE: to_render.add(u.src[1])
     if u.op in {Ops.REDUCE, Ops.REDUCE_AXIS}: to_render.add(u.src[0])
+    if u.op in {Ops.CUSTOM_KERNEL, Ops.CALL}: raise NotImplementedError("custom_kernel / call can't be pyrendered")
     if u.op in not_rendered: continue
     # checking the consumers is not enough, you have to make sure it's not used twice by the one consumer
     if len(cmap[u]) == 1 and len([x for x in list(cmap[u].keys())[0].src if x is u]) == 1 and u.op not in always_rendered: continue
@@ -1444,7 +1444,13 @@ def pyrender(ast:UOp) -> str:
   kernels: dict[UOp, tuple[str, str]] = {}
   r: dict[UOp, str] = {}
   ret: dict[str, str] = {}
+  depth: dict[UOp, int] = {}
   for i,u in enumerate(lst):
+    # limit inline depth to avoid "too many nested parentheses" in Python parser
+    op_depth = 1 + max([depth[s] for s in u.src], default=0)
+    if op_depth > 100: to_render.add(u)
+    depth[u] = 0 if u in to_render else op_depth
+    # do the rendering
     if u.op is Ops.KERNEL:
       if u.arg.ast not in kernels:
         kernels[u.arg.ast] = (f"k{len(kernels)}", f"def k{len(kernels)}():\n  " + pyrender(u.arg.ast).replace('\n', '\n  ') + "\n  return ast\n\n")
diff --git a/tinygrad/uop/spec.py b/tinygrad/uop/spec.py
index 869166ed24..d5bb70959b 100644
--- a/tinygrad/uop/spec.py
+++ b/tinygrad/uop/spec.py
@@ -84,14 +84,11 @@ _tensor_spec = PatternMatcher([
   (UPat(Ops.LUNIQUE, dtypes.void, ()), lambda: True),
   (UPat(Ops.DEVICE, dtypes.void, (), name="d"), lambda d:
    isinstance(d.arg, str) or (isinstance(d.arg, tuple) and all(isinstance(s, str) for s in d.arg))),
-  (UPat(Ops.BUFFER, src=(UPat((Ops.LUNIQUE, Ops.UNIQUE)), UPat(Ops.DEVICE)), allow_any_len=True, name="buf"),
+  (UPat(Ops.BUFFER, src=(UPat((Ops.LUNIQUE, Ops.UNIQUE)), UPat(Ops.DEVICE)), name="buf"),
    lambda buf: isinstance(buf.arg, int) and isinstance(buf.dtype, (DType, ImageDType))),
-  (UPat(Ops.BUFFER_VIEW, src=(UPat(Ops.BUFFER),), name="buf_view"),
-   lambda buf_view: isinstance(buf_view.arg, tuple) and len(buf_view.arg) == 2 and all(isinstance(arg, (int, UOp)) for arg in buf_view.arg)),
-  (UPat(Ops.BUFFER_VIEW, src=(UPat(Ops.MSTACK, src=UPat(Ops.BUFFER)),)), lambda: True),
 
   # KERNEL can attach to an AFTER to describe the compute required to realize a BUFFER
-  (UPat(Ops.KERNEL, src=UPat((Ops.BUFFER, Ops.BUFFER_VIEW, Ops.AFTER, Ops.MSELECT, Ops.MSTACK, Ops.BIND, Ops.CONTIGUOUS))), lambda: True),
+  (UPat(Ops.KERNEL, src=UPat((Ops.BUFFER, Ops.AFTER, Ops.MSELECT, Ops.MSTACK, Ops.BIND))), lambda: True),
 
   # ASSIGN has a target and a value. It can also optionally depend on other assigns
   (UPat(Ops.ASSIGN, name="x"), lambda x: len(x.src) >= 2 and all(s.op is Ops.ASSIGN for s in x.src[2:])),
@@ -316,7 +313,8 @@ def eval_pyrender(code:str) -> UOp:
   return lcls['ast']
 
 def test_pyrender(test_ast:UOp, assert_parents=True):
-  code = pyrender(test_ast)
+  try: code = pyrender(test_ast)
+  except NotImplementedError: return None  # this is okay, not all ops can be pyrendered
   ast:UOp = eval_pyrender(code)
   if ast is not test_ast:
     if assert_parents:
diff --git a/tinygrad/viz/serve.py b/tinygrad/viz/serve.py
index be211c981f..7e93d190fd 100755
--- a/tinygrad/viz/serve.py
+++ b/tinygrad/viz/serve.py
@@ -306,11 +306,12 @@ def load_counters(profile:list[ProfileEvent]) -> None:
     # to decode a SQTT trace, we need the raw stream, program binary and device properties
     if (sqtt:=v.get(ProfileSQTTEvent)):
       for e in sqtt:
-        if e.itrace: steps.append(create_step(f"PKTS SE:{e.se}", (f"/prg-pkts-{e.se}", len(ctxs), len(steps)), data=(e.blob, prg_events[k].lib)))
+        if e.itrace: steps.append(create_step(f"PKTS SE:{e.se}", (f"/prg-pkts-{e.se}", len(ctxs), len(steps)),
+                                              data=(e.blob, prg_events[k].lib, device_props[e.device]["gfx_target_version"])))
       steps.append(create_step("SQTT", ("/prg-sqtt", len(ctxs), len(steps)), ((k, tag), sqtt, prg_events[k])))
     ctxs.append({"name":f"Exec {name}"+(f" n{run_number[k]}" if run_number[k] > 1 else ""), "steps":steps})
 
-def sqtt_timeline(data:bytes, lib:bytes) -> list[ProfileEvent]:
+def sqtt_timeline(data:bytes, lib:bytes, target:int) -> list[ProfileEvent]:
   from extra.assembly.amd.sqttmap import map_insts, InstructionInfo
   from extra.assembly.amd.sqtt import PacketType, INST, InstOp, VALUINST, IMMEDIATE, IMMEDIATE_MASK, VMEMEXEC, ALUEXEC
   ret:list[ProfileEvent] = []
@@ -321,7 +322,7 @@ def sqtt_timeline(data:bytes, lib:bytes) -> list[ProfileEvent]:
     rows.setdefault(r:=(f"WAVE:{wave}" if wave is not None else f"{p.__class__.__name__}:0 {name}"))
     key = TracingKey(f"{op_name if op_name is not None else name} OP:{idx}", ret=info.inst.disasm() if info is not None else None)
     ret.append(ProfileRangeEvent(r, key, Decimal(p._time), Decimal(p._time+width)))
-  for p, info in map_insts(data, lib):
+  for p, info in map_insts(data, lib, target):
     if len(ret) > getenv("MAX_SQTT_PKTS", 50_000): break
     if isinstance(p, INST):
       op_name = p.op.name if isinstance(p.op, InstOp) else f"0x{p.op:02x}"
@@ -346,7 +347,7 @@ def unpack_sqtt(key:tuple[str, int], data:list, p:ProfileProgramEvent) -> tuple[
   # * init decoder
   from extra.sqtt.roc import decode
   base = unwrap(p.base)
-  addr_table = amd_decode(unwrap(p.lib), device_props[p.device]["gfx_target_version"], )
+  addr_table = amd_decode(unwrap(p.lib), device_props[p.device]["gfx_target_version"])
   disasm:dict[int, tuple[str, int]] = {addr+base:(inst.disasm(), inst.size()) for addr, inst in addr_table.items()}
   rctx = decode(data, {p.tag:disasm})
   cu_events:dict[str, list[ProfileEvent]] = {}
@@ -422,16 +423,14 @@ def get_stdout(f: Callable) -> str:
   return buf.getvalue()
 
 def amd_readelf(lib:bytes) -> list[dict]:
+  from tinygrad.runtime.autogen import amdgpu_kd
   from tinygrad.runtime.support.elf import elf_loader
-  import msgpack
-  _, sections, __ = elf_loader(lib)
-  data = next((s for s in sections if s.name.startswith(".note"))).content
-  namesz, descsz, typ = struct.unpack_from(hdr:="<III", data, 0)
-  offset = (struct.calcsize(hdr)+namesz+3) & -4
-  notes = msgpack.unpackb(data[offset:offset+descsz])
-  keys = {".sgpr_count":"SGPRs", ".vgpr_count":"VGPRs", ".max_flat_workgroup_size":"Max WGP size",
-          ".group_segment_fixed_size":"LDS size", ".private_segment_fixed_size":"Scratch size"}
-  return [{"label":label, "value":v} for k,label in keys.items() if (v:=notes["amdhsa.kernels"][0][k]) > 0]
+  image, sections, __ = elf_loader(lib)
+  rodata = next((s for s in sections if s.name == ".rodata")).content
+  kd = amdgpu_kd.llvm_amdhsa_kernel_descriptor_t.from_buffer_copy(bytearray(rodata))
+  vgpr_gran = kd.compute_pgm_rsrc1 & amdgpu_kd.COMPUTE_PGM_RSRC1_GRANULATED_WORKITEM_VGPR_COUNT
+  return [{"label":f"{resource} Alloc", "value":val} for resource,val in [("VGPR", (vgpr_gran+1)*8-7), ("LDS",kd.group_segment_fixed_size),
+                                                                          ("Scratch", kd.private_segment_fixed_size)] if val > 0]
 
 def amd_decode(lib:bytes, target:int) -> dict[int, Any]: # Any is the Inst class from extra.assembly.amd.dsl
   from tinygrad.runtime.support.elf import elf_loader
@@ -558,7 +557,7 @@ def get_render(query:str) -> dict:
     rows:dict[int, dict] = {}
     for pc, (inst,_) in pc_to_inst.items():
       if start_pc is None: start_pc = pc
-      rows[pc] = {"pc":pc-start_pc, "inst":inst, "hit_count":0, "dur":0, "stall":0, "hits":{"cols":inst_columns, "rows":[]}, "type":""}
+      rows[pc] = {"pc":pc-start_pc, "inst":inst, "hit_count":0, "dur":0, "stall":0, "type":"", "hits":{"cols":inst_columns, "rows":[]}}
     for e in w.unpack_insts():
       if not (inst:=rows[e.pc]).get("type"): inst["type"] = str(e.typ).split("_")[-1]
       inst["hit_count"] += 1