scan with assign

* vmap grad needs reduce_backward

* fuse and outer

extra/sqtt/roc.py

@@ -1,4 +1,4 @@
-import ctypes, pathlib, argparse, pickle, re, functools, dataclasses, itertools
+import ctypes, pathlib, argparse, pickle, re, functools, dataclasses, itertools, threading
 from tinygrad.helpers import temp, unwrap, DEBUG
 from tinygrad.device import ProfileEvent, ProfileDeviceEvent, ProfileProgramEvent
 from tinygrad.runtime.ops_amd import ProfileSQTTEvent, ProfilePMCEvent
@@ -94,14 +94,14 @@ def decode(profile:list[ProfileEvent]) -> _ROCParseCtx:
   ROCParseCtx = _ROCParseCtx(dev_events, sqtt_events, prog_events)
 
   @rocprof.rocprof_trace_decoder_se_data_callback_t
-  def copy_cb(buf, buf_size, data_ptr):
+  def copy_cb(buf, buf_size, _):
     if (prof_info:=ROCParseCtx.next_sqtt()) is None: return 0
     buf[0] = ctypes.cast(prof_info, ctypes.POINTER(ctypes.c_ubyte))
     buf_size[0] = len(prof_info)
     return len(prof_info)
 
   @rocprof.rocprof_trace_decoder_trace_callback_t
-  def trace_cb(record_type, events_ptr, n, data_ptr):
+  def trace_cb(record_type, events_ptr, n, _):
     match record_type:
       case rocprof.ROCPROFILER_THREAD_TRACE_DECODER_RECORD_OCCUPANCY:
         for ev in (rocprof.rocprofiler_thread_trace_decoder_occupancy_t * n).from_address(events_ptr): ROCParseCtx.on_occupancy_ev(ev)
@@ -112,7 +112,7 @@ def decode(profile:list[ProfileEvent]) -> _ROCParseCtx:
     return rocprof.ROCPROFILER_THREAD_TRACE_DECODER_STATUS_SUCCESS
 
   @rocprof.rocprof_trace_decoder_isa_callback_t
-  def isa_cb(instr_ptr, mem_size_ptr, size_ptr, pc, data_ptr):
+  def isa_cb(instr_ptr, mem_size_ptr, size_ptr, pc, _):
     instr, mem_size_ptr[0] = ROCParseCtx.disasms[(unwrap(ROCParseCtx.active_kern), pc.address)]
 
     # this is the number of bytes to next instruction, set to 0 for end_pgm
@@ -126,9 +126,11 @@ def decode(profile:list[ProfileEvent]) -> _ROCParseCtx:
 
     return rocprof.ROCPROFILER_THREAD_TRACE_DECODER_STATUS_SUCCESS
 
-  try:
-    rocprof.rocprof_trace_decoder_parse_data(copy_cb, trace_cb, isa_cb, None)
-  except AttributeError as e: raise RuntimeError("Failed to find rocprof-trace-decoder. Run sudo ./extra/sqtt/install_sqtt_decoder.py to install") from e
+  def worker():
+    try: rocprof.rocprof_trace_decoder_parse_data(copy_cb, trace_cb, isa_cb, None)
+    except AttributeError as e: raise RuntimeError("Failed to find rocprof-trace-decoder. Run sudo ./extra/sqtt/install_sqtt_decoder.py to install") from e
+  (t:=threading.Thread(target=worker, daemon=True)).start()
+  t.join()
   return ROCParseCtx
 
 if __name__ == "__main__":

test/test_outerworld.py

@@ -1,4 +1,5 @@
 import unittest
+import numpy as np
 from tinygrad import Tensor, UOp
 from tinygrad.uop.ops import AxisType, Ops
 
@@ -77,7 +78,8 @@ class TestOuterScan(unittest.TestCase):
     # 3 matmuls with SCAN
     i = UOp.range(3, -100, AxisType.OUTER)
     out = Tensor.empty(3, 1, 10)
-    comp = Tensor(i.eq(0).where(vec.uop, out[(i-1).maximum(0)].uop)) @ mats[i]
+    phi = Tensor(i.eq(0).where(vec.uop, out[(i-1).maximum(0)].uop))
+    comp = phi @ mats[i]
     store = out[i].uop.store(comp.uop).end(i)
     out = Tensor(out.uop.after(store))
     out.realize()
@@ -85,6 +87,22 @@ class TestOuterScan(unittest.TestCase):
     # TODO: testing allclose
     assert Tensor.allclose(ref, out, atol=1e-6), f"{ref.numpy()=}, {out.numpy()=}"
 
+  def test_scan_with_assign(self):
+    vec, mats, ref = self._test_scan()
+
+    a = UOp.range(3, -1, AxisType.OUTER)
+    buf_out = Tensor.empty(3, 1, 10)
+    phi = Tensor(a.eq(0).where(vec.uop, buf_out[(a-1).maximum(0)].uop))
+    out = phi @ Tensor(mats.uop.reduce_backward(a, arg=Ops.ADD))[a]
+    out = out.reshape(1, 1, 10).pad(((a,(3-a)-1), None, None))
+    out = Tensor(out.uop.reduce(a, arg=Ops.ADD))
+    out = buf_out.assign(out)
+
+    out.realize()
+
+    # TODO: testing allclose
+    np.testing.assert_allclose(ref.numpy(), out.numpy())
+
 class TestOuterworld(unittest.TestCase):
   def test_range_plus_1(self):
     t = Tensor.arange(100).reshape(10,10).realize()
@@ -144,5 +162,34 @@ class TestOuterworld(unittest.TestCase):
     out = out.reshape(1, 3).expand(a, 3).contiguous().realize()
     self.assertListEqual([[0,4,8],[4,8,12],[8,12,16]], out.tolist())
 
+class TestVmap(unittest.TestCase):
+  def test_vmap_inner(self, axis_type=AxisType.LOOP, fuse=False, grad=False):
+    x = Tensor.ones(1, 10).contiguous().requires_grad_()
+    mats = Tensor.ones(3, 10, 10).contiguous().requires_grad_()
+
+    ref = x @ mats
+    if fuse: ref = ref * 2
+
+    # vmap across axis 0
+    a = UOp.range(3, -1, axis_type)
+    out = x @ Tensor(mats.uop.reduce_backward(a, arg=Ops.ADD))[a]
+    out = out.reshape(1, 10).pad(((a,(3-a)-1), None))
+    out = Tensor(out.uop.reduce(a, arg=Ops.ADD))
+    if fuse: out = out * 2
+    if grad:
+      out.mean().backward()
+      np.testing.assert_allclose(mats.grad.numpy(), (2./30) if fuse else (1./30))
+    out.realize()
+
+    # TODO: testing allclose
+    assert Tensor.allclose(ref, out, atol=1e-6), f"{ref.numpy()=}, {out.numpy()=}"
+  def test_vmap_inner_fuse(self): self.test_vmap_inner(fuse=True)
+  def test_vmap_outer(self): self.test_vmap_inner(AxisType.OUTER)
+  def test_vmap_outer_fuse(self): self.test_vmap_inner(AxisType.OUTER, fuse=True)
+
+  def test_vmap_inner_grad(self): self.test_vmap_inner(grad=True)
+  def test_vmap_inner_fuse_grad(self): self.test_vmap_inner(fuse=True, grad=True)
+  def test_vmap_outer_grad(self): self.test_vmap_inner(AxisType.OUTER, grad=True)
+
 if __name__ == '__main__':
   unittest.main()

tinygrad/engine/schedule.py

@@ -90,7 +90,7 @@ def create_schedule_with_vars(sched_sink:UOp) -> tuple[list[ScheduleItem], dict[
       if rk.op is Ops.END: schedule.append(rk)
     else:
       raise RuntimeError(f"can't schedule {k.op}")
-    for x in children[k]:
+    for x in children[rk]:
       in_degree[x] -= 1
       if in_degree[x] == 0: queues[_heuristic(x)].append(x)
 

tinygrad/gradient.py

@@ -3,14 +3,15 @@ import math, dataclasses
 from tinygrad.uop.ops import UOp, PatternMatcher, UPat, Ops, all_metadata
 from tinygrad.helpers import argsort
 
-def reduce_gradient(ctx:UOp, ret:UOp):
+def reduce_gradient(ctx:UOp, ret:UOp, op:Ops):
   def broadcast_to_input(x): return x.reshape(x.shape+(1,)*(len(ret.src[0].shape)-len(x.shape))).expand(ret.src[0].shape)
-  if ret.arg[0] == Ops.ADD: return (broadcast_to_input(ctx),)
-  if ret.arg[0] == Ops.MAX:
+  if op == Ops.ADD: return (broadcast_to_input(ctx),)
+  if op == Ops.MAX:
+    assert ret.op is Ops.REDUCE_AXIS, "only works on REDUCE_AXIS"
     mask = ret.src[0].eq(broadcast_to_input(ret)).cast(ctx.dtype)
     count = mask.r(Ops.ADD, ret.arg[1])
     return ((mask/broadcast_to_input(count)) * broadcast_to_input(ctx),)
-  if ret.arg[0] == Ops.MUL: return (broadcast_to_input(ctx * ret) / ret.src[0],)
+  if op == Ops.MUL: return (broadcast_to_input(ctx * ret) / ret.src[0],)
 
 # ctx is grad_output
 pm_gradient = PatternMatcher([
@@ -28,7 +29,9 @@ pm_gradient = PatternMatcher([
     ((x>y).where(ctx, (x.eq(y)).where(ctx * 0.5, 0)), (x<y).where(ctx, (x.eq(y)).where(ctx * 0.5, 0)))),
   (UPat(Ops.MUL, name="ret"), lambda ctx, ret: (ret.src[1]*ctx, ret.src[0]*ctx)),
   (UPat(Ops.WHERE, name="ret"), lambda ctx, ret: (None, ret.src[0].where(ctx, ctx.const_like(0)), ret.src[0].where(ctx.const_like(0), ctx))),
-  (UPat(Ops.REDUCE_AXIS, name="ret"), reduce_gradient),
+  (UPat(Ops.REDUCE_AXIS, name="ret"), lambda ctx, ret: reduce_gradient(ctx, ret, ret.arg[0])),
+  (UPat(Ops.REDUCE, name="ret"), lambda ctx, ret: reduce_gradient(ctx, ret, ret.arg) + (None,)*(len(ret.src)-1)),
+  (UPat(Ops.REDUCE_BACKWARD, name="ret"), lambda ctx, ret: (ctx.reduce(*ret.src[1:], arg=ret.arg),) + (None,)*(len(ret.src)-1)),
   (UPat(Ops.CONTIGUOUS), lambda ctx: (ctx,)),
   (UPat(Ops.CONTIGUOUS_BACKWARD), lambda ctx: (ctx.contiguous(),)),
   (UPat(Ops.RESHAPE, name="ret"), lambda ctx, ret: (ctx.reshape(ret.src[0].shape), None)),

tinygrad/runtime/autogen/am/smu_v14_0_2.py

@@ -476,7 +476,7 @@ PP_GRTAVFS_FW_SEP_FUSE_FREQUENCY_TO_COUNT_SCALER_4 = PP_GRTAVFS_FW_SEP_FUSE_e.de
 PP_GRTAVFS_FW_SEP_FUSE_COUNT = PP_GRTAVFS_FW_SEP_FUSE_e.define('PP_GRTAVFS_FW_SEP_FUSE_COUNT', 19)
 
 class SviTelemetryScale_t(Struct): pass
-int8_t = ctypes.c_char
+int8_t = ctypes.c_byte
 SviTelemetryScale_t._fields_ = [
   ('Offset', int8_t),
   ('Padding', uint8_t),

tinygrad/runtime/autogen/mesa.py

@@ -89,7 +89,7 @@ NIR_CMAT_C_SIGNED = nir_cmat_signed.define('NIR_CMAT_C_SIGNED', 4)
 NIR_CMAT_RESULT_SIGNED = nir_cmat_signed.define('NIR_CMAT_RESULT_SIGNED', 8)
 
 class nir_const_value(ctypes.Union): pass
-int8_t = ctypes.c_char
+int8_t = ctypes.c_byte
 uint8_t = ctypes.c_ubyte
 int16_t = ctypes.c_int16
 uint16_t = ctypes.c_uint16
@@ -3723,7 +3723,7 @@ struct__IO_FILE._fields_ = [
   ('_flags2', ctypes.c_int32),
   ('_old_offset', ctypes.c_int64),
   ('_cur_column', ctypes.c_uint16),
-  ('_vtable_offset', ctypes.c_char),
+  ('_vtable_offset', ctypes.c_byte),
   ('_shortbuf', (ctypes.c_char * 1)),
   ('_lock', ctypes.POINTER(_IO_lock_t)),
   ('_offset', ctypes.c_int64),

tinygrad/runtime/ops_amd.py

@@ -8,6 +8,7 @@ from tinygrad.runtime.support.hcq import MMIOInterface, BumpAllocator, hcq_filte
 from tinygrad.uop.ops import sint
 from tinygrad.device import Compiled, DMAFdRef, BufferSpec, CompilerPairT
 from tinygrad.helpers import getenv, round_up, data64_le, DEBUG, PROFILE, ProfileEvent, suppress_finalizing, lo32, hi32, colored, prod, ContextVar
+from tinygrad.helpers import VIZ
 from tinygrad.renderer.cstyle import AMDRenderer
 from tinygrad.renderer.llvmir import AMDLLVMRenderer
 from tinygrad.runtime.autogen import kfd, hsa, pci, sqtt
@@ -19,7 +20,7 @@ from tinygrad.runtime.support.amd import AMDReg, AMDIP, import_module, import_so
 from tinygrad.runtime.support.system import System, PCIIfaceBase, PCIAllocationMeta, PCIDevice, USBPCIDevice, MAP_FIXED, MAP_NORESERVE
 if getenv("IOCTL"): import extra.hip_gpu_driver.hip_ioctl  # noqa: F401 # pylint: disable=unused-import
 
-SQTT, SQTT_ITRACE_SE_MASK, PMC = ContextVar("SQTT", 0), ContextVar("SQTT_ITRACE_SE_MASK", 0b11), ContextVar("PMC", 0)
+SQTT, SQTT_ITRACE_SE_MASK, PMC = ContextVar("SQTT", VIZ.value>=2), ContextVar("SQTT_ITRACE_SE_MASK", 0b11), ContextVar("PMC", 0)
 EVENT_INDEX_PARTIAL_FLUSH = 4 # based on a comment in nvd.h
 WAIT_REG_MEM_FUNCTION_EQ  = 3 # ==
 WAIT_REG_MEM_FUNCTION_NEQ = 4 # !=

tinygrad/runtime/support/autogen.py

@@ -103,7 +103,7 @@ def gen(dll, files, args=[], prolog=[], rules=[], epilog=[], recsym=False, use_e
     suggested_name = anon_names.get(f"{loc_file(loc(decl:=clang.clang_getTypeDeclaration(t)))}:{loc_line(loc(decl))}", suggested_name)
     nonlocal lines, types, anoncnt, objc
     tmap = {clang.CXType_Void:"None", clang.CXType_Char_U:"ctypes.c_ubyte", clang.CXType_UChar:"ctypes.c_ubyte", clang.CXType_Char_S:"ctypes.c_char",
-            clang.CXType_SChar:"ctypes.c_char",
+            clang.CXType_SChar:"ctypes.c_byte",
             **{getattr(clang, f'CXType_{k}'):f"ctypes.c_{k.lower()}" for k in ["Bool", "WChar", "Float", "Double", "LongDouble"]},
             **{getattr(clang, f'CXType_{k}'):f"ctypes.c_{'u' if 'U' in k else ''}int{sz}" for sz,k in
                [(16, "UShort"), (16, "Short"), (32, "UInt"), (32, "Int"), (64, "ULong"), (64, "Long"), (64, "ULongLong"), (64, "LongLong")]}}

tinygrad/schedule/indexing.py

@@ -18,6 +18,11 @@ def realize_srcs(ctx:dict[UOp, None], rb:UOp) -> None:
 
 def realize_assign(ctx:dict[UOp, None], a:UOp) -> None:
   if a.src[1].op not in ALWAYS_CONTIGUOUS: ctx[a.src[1]] = None
+
+  # if it's an outer reduce, we *don't* realize it
+  if a.src[1].op is Ops.REDUCE and any(tr.arg[-1] == AxisType.OUTER for tr in a.src[1].src[1:]):
+    del ctx[a.src[1]]
+
   # if it's a kernel, we don't realize it
   if a.src[1].op is not Ops.KERNEL: ctx[a] = None
 
@@ -26,6 +31,8 @@ pm_generate_realize_map = PatternMatcher([
   (UPat(Ops.SINK, name="s"), lambda ctx,s: ctx.update((x.base, None) for x in s.src if x.base.op not in ALWAYS_CONTIGUOUS)),
   # always realize COPY/BUFFER_VIEW/CONTIGUOUS/STORE
   (UPat({Ops.COPY, Ops.BUFFER_VIEW, Ops.CONTIGUOUS, Ops.STORE}, name="tr"), realize),
+  # always realize REDUCE on outer ranges
+  (UPat(Ops.REDUCE, name="r"), lambda ctx,r: realize(ctx, r) if any(tr.arg[-1] == AxisType.OUTER for tr in r.src[1:]) else None),
   # realize srcs of COPY, MSELECT, MSTACK
   (UPat((Ops.COPY, Ops.MSELECT, Ops.MSTACK), name="rb"), realize_srcs),
   # realize ASSIGN and input to assign (might be optimized out)

tinygrad/schedule/rangeify.py

@@ -65,7 +65,7 @@ mop_cleanup = PatternMatcher([
 
 earliest_rewrites = mop_cleanup+PatternMatcher([
   # just removing it works...
-  (UPat((Ops.DETACH, Ops.CONTIGUOUS_BACKWARD), name="x"), lambda x: x.src[0]),
+  (UPat((Ops.DETACH, Ops.CONTIGUOUS_BACKWARD, Ops.REDUCE_BACKWARD), name="x"), lambda x: x.src[0]),
 
   # 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)),
@@ -472,6 +472,7 @@ pm_add_range_tags = PatternMatcher([
 ])
 
 def split_store(ctx:list[UOp], x:UOp) -> UOp|None:
+  # if we have any outer ranges open here, we don't split
   if len([r for r in x.ranges if r.arg[-1] != AxisType.OUTER]): return None
 
   # ends of outer range don't go in kernels
@@ -499,7 +500,12 @@ def split_store(ctx:list[UOp], x:UOp) -> UOp|None:
     raise RuntimeError(f"all buffers must be on the same device: {tuple(b.buf_uop.buffer for b in kernel.src)}")
   return kernel
 
+def split_inner_and_outer_end(x: UOp):
+  outer_ranges, inner_ranges = partition(x.src[1:], lambda r: r.arg[-1] == AxisType.OUTER)
+  if len(outer_ranges) and len(inner_ranges): return x.src[0].end(*inner_ranges).end(*outer_ranges)
+
 split_kernels = PatternMatcher([
+  (UPat(Ops.END, name="x"), split_inner_and_outer_end),
   (UPat((Ops.STORE, Ops.END), name="x"), split_store),
 ])
 
@@ -531,6 +537,24 @@ replace_contiguous = PatternMatcher([
   (UPat(GroupOp.ALU, name="alu"), lambda ctx,alu: alu.replace(src=new_src) if (new_src:=tuple(ctx.get(s, s) for s in alu.src)) != alu.src else None),
 ])
 
+pm_fix_vmap = PatternMatcher([
+  # x>=y and x<(y+1) means x==y (this can go in symbolic)
+  ((UPat.var("x", dtype=dtypes.index) >= UPat.var("y")) & (UPat.var("x") < (UPat.var("y")+1)), lambda x,y: x.eq(y)),
+  # remove the reduce if it's compare reduce w assign (keep the outer range)
+  (UPat(Ops.BUFFERIZE, name="buf", src=(
+   UPat(Ops.ASSIGN, name="assign", allow_any_len=True, src=(UPat.var("assign_buf"),
+   (UPat.var("r1", dtype=dtypes.index) != UPat.var("r2")).where(0, UPat.var("val")).reduce(UPat.var("r2"), arg=Ops.ADD),)),), allow_any_len=True),
+   lambda r1,r2,val,buf,assign_buf,assign: buf.replace(src=(UOp(Ops.ASSIGN, val.dtype,
+    src=(assign_buf,val)+assign.src[2:]),)+buf.src[1:]).substitute({r1:r2}) if r1 in buf.src[1:] and r2.arg[-1] == AxisType.OUTER else None),
+  # remove the reduce if it's compare reduce (keep the outer range)
+  (UPat(Ops.BUFFERIZE, name="buf", src=(
+   (UPat.var("r1", dtype=dtypes.index) != UPat.var("r2")).where(0, UPat.var("val")).reduce(UPat.var("r2"), arg=Ops.ADD),), allow_any_len=True),
+   lambda r1,r2,val,buf: buf.replace(src=(val,)+buf.src[1:]).substitute({r1:r2}) if r1 in buf.src[1:] and r2.arg[-1] == AxisType.OUTER else None),
+  # remove the reduce if it's compare reduce
+  ((UPat.var("r1", dtype=dtypes.index) != UPat.var("r2")).where(0, UPat.var("val")).reduce(UPat.var("r2"), arg=Ops.ADD),
+   lambda r1,r2,val: val.substitute({r2:r1}) if r2.arg[-1] != AxisType.OUTER else None),
+])
+
 @disable_gc()
 @track_rewrites(lambda _,ret: f"Schedule {pluralize('Kernel', len([u for u in UOp.sink(*ret.values()).toposort() if u.op is Ops.KERNEL]))}", True)
 def get_rangeify_map(sink:UOp) -> dict[UOp, UOp]:
@@ -543,9 +567,9 @@ def get_rangeify_map(sink:UOp) -> dict[UOp, UOp]:
   # convert movement ops to ranges
   tsink, rctx = run_rangeify(tsink, DEBUG_RANGEIFY)
 
-  tsink = graph_rewrite(tsink, symbolic+pm_reduce_simplify+pm_const_buffer_folding, name="symbolic+reduce_collapse")  # this does const folding
+  tsink = graph_rewrite(tsink, symbolic+pm_reduce_simplify+pm_const_buffer_folding+pm_fix_vmap, name="symbolic+reduce_collapse")
   tsink = graph_rewrite(tsink, pm_remove_bufferize, bottom_up=True, name="remove bufferize with cost function")
-  tsink = graph_rewrite(tsink, symbolic+pm_reduce_simplify+pm_const_buffer_folding, name="symbolic+reduce_collapse pt 2")
+  tsink = graph_rewrite(tsink, symbolic+pm_reduce_simplify+pm_const_buffer_folding+pm_fix_vmap, name="symbolic+reduce_collapse pt 2")
   tsink = graph_rewrite(tsink, pm_limit_bufs, ctx=rctx, name="limit buffers")
 
   # rebuild the sink with all the BUFFERIZEs with tags, this is what's ending up in the tensor graph

tinygrad/uop/__init__.py

@@ -91,6 +91,7 @@ class Ops(FastEnum):
 
   # reduce
   REDUCE_AXIS = auto(); REDUCE = auto(); ALLREDUCE = auto()
+  REDUCE_BACKWARD = auto()
 
   # errors/placeholders
   REWRITE_ERROR = auto(); SENTINEL = auto()

tinygrad/uop/ops.py

@@ -89,8 +89,8 @@ class UOpMetaClass(type):
     if SPEC > 1:
       from tinygrad.uop.spec import full_spec, test_pyrender
       if SPEC > 2: test_pyrender(created)
-      with Context(IGNORE_OOB=1): ret = full_spec.rewrite(created)
-      if cast(bool|None, ret) is not True: raise RuntimeError(f"SPEC ISSUE {ret}: {created}")
+      with Context(IGNORE_OOB=1): fret = cast(bool|None, full_spec.rewrite(created))
+      if fret is not True: raise RuntimeError(f"SPEC ISSUE {fret}: {created}")
     return created
 
 # some uops map to other stuff
@@ -219,7 +219,7 @@ class UOp(OpMixin, metaclass=UOpMetaClass):
       case Ops.DEFINE_GLOBAL | Ops.DEFINE_LOCAL | Ops.DEFINE_REG: return (self.ptrdtype.size,)
 
       # passthrough ops
-      case Ops.REDUCE | Ops.MSTACK | Ops.MSELECT | Ops.DETACH | Ops.CONTIGUOUS | Ops.CONTIGUOUS_BACKWARD | Ops.AFTER | Ops.END:
+      case Ops.REDUCE | Ops.REDUCE_BACKWARD | Ops.MSTACK | Ops.MSELECT | Ops.DETACH | Ops.CONTIGUOUS | Ops.CONTIGUOUS_BACKWARD | Ops.AFTER | Ops.END:
         return self.src[0]._shape
 
       # ops with custom handling
@@ -442,6 +442,7 @@ class UOp(OpMixin, metaclass=UOpMetaClass):
     assert self.dtype.scalar() is dtypes.index, "Can only call get_valid on index dtype"
     return self.src[0] if self.op is Ops.WHERE and self.src[2].arg is Invalid else UOp.const(dtypes.bool, self.arg is not Invalid)
   def reduce(self, *src:UOp, **kwargs): return UOp(Ops.REDUCE, kwargs.pop('dtype', self.dtype), src=(self,)+src, **kwargs)
+  def reduce_backward(self, *src:UOp, **kwargs): return UOp(Ops.REDUCE_BACKWARD, kwargs.pop('dtype', self.dtype), src=(self,)+src, **kwargs)
 
   def is_contiguous(self):
     # TODO: this is is_realized
@@ -583,7 +584,7 @@ class UOp(OpMixin, metaclass=UOpMetaClass):
   def new_buffer(device:str|tuple[str, ...], size:int, dtype:DType, num=None):
     return UOp(Ops.BUFFER, dtype, (UOp.unique(num), UOp(Ops.DEVICE, arg=device)), size)
   @property
-  def device(self) -> str|tuple[str, ...]: return cast(str|tuple[str, ...], unwrap(self._device))
+  def device(self) -> str|tuple[str, ...]: return unwrap(self._device)
   @recursive_property
   def _device(self) -> str|tuple[str, ...]|None:
     if self.op is Ops.DEVICE: return self.arg
@@ -1164,12 +1165,12 @@ class RewriteContext:
 
   def cached_pm_rewrite(self, x:UOp):
     if (ret:=self.pm_cache.get(x,SENTINEL)) is not SENTINEL: return ret
-    ret = self.pm_cache[x] = cast(PatternMatcher, self.pm).rewrite(x, self.ctx)
+    ret = self.pm_cache[x] = unwrap(self.pm).rewrite(x, self.ctx)
     return ret
 
   def cached_bpm_rewrite(self, x:UOp):
     if (ret:=self.bpm_cache.get(x,SENTINEL)) is not SENTINEL: return ret
-    ret = self.bpm_cache[x] = cast(PatternMatcher, self.bpm).rewrite(x, self.ctx)
+    ret = self.bpm_cache[x] = unwrap(self.bpm).rewrite(x, self.ctx)
     return ret
 
   def unified_rewrite(self, root:UOp) -> UOp:

tinygrad/uop/spec.py

@@ -108,7 +108,7 @@ _tensor_spec = PatternMatcher([
   (UPat(Ops.REDUCE_AXIS, name="x"), lambda x: isinstance(x.arg, tuple) and len(x.arg) >= 2 and x.arg[0] in {Ops.ADD, Ops.MUL, Ops.MAX}),
 
   # REDUCE with an outerworld range
-  (UPat(Ops.REDUCE, src=(UPat(),), allow_any_len=True, name="x"), lambda x: all(y.dtype == dtypes.index for y in x.src[1:])),
+  (UPat((Ops.REDUCE, Ops.REDUCE_BACKWARD), src=(UPat(),), allow_any_len=True, name="x"), lambda x: all(y.dtype == dtypes.index for y in x.src[1:])),
 
   # AFTER if things were kernelized
   (UPat(Ops.AFTER, src=(UPat((Ops.BUFFER, Ops.AFTER)),), allow_any_len=True), lambda: True),

tinygrad/uop/symbolic.py

@@ -407,14 +407,10 @@ def uop_given_valid(valid:UOp, uop:UOp, try_simplex=True) -> UOp:
     expr, is_upper, c = res
     bounds[expr][int(is_upper)] = c
 
-  # don't simplify any other gates, can lead to OOB, we substitute them back later
-  uop = uop.substitute((load_subs:={u: UOp(Ops.NOOP, dtype=u.dtype, arg=u) for u in uop.toposort() if u.op is Ops.INDEX}))
-
   # simplify uop given that valid is True
   all_candidates = []
   for i,(expr,v) in enumerate(bounds.items()):
     v0, v1 = (expr.vmin if v[0] is None else v[0], expr.vmax if v[1] is None else v[1])
-    expr = expr.substitute(load_subs)  # make sure expr appears in same form in the uop
     # try checking the whole clause
     all_candidates.append((expr, UOp.variable(f"fake{i}", v0, v1, expr.dtype)))
 
@@ -438,8 +434,6 @@ def uop_given_valid(valid:UOp, uop:UOp, try_simplex=True) -> UOp:
   # try all the valids together (but only the whole expressions)
   if (s_uop:=uop.substitute(sub_dict:=dict(all_candidates))) is not uop:
     uop = s_uop.simplify().substitute({newX:X for X,newX in sub_dict.items()}).simplify(full_symbolic=False)
-  # put the loads back in
-  uop = uop.substitute({v:k for k,v in load_subs.items()})
   return uop
 
 def _valid_priority(v: UOp, valids:list[UOp]):

tinygrad/viz/index.html

@@ -271,7 +271,7 @@
   }
   #device-list > div {
     min-height: 32px;
-    width: 132px;
+    width: 134px;
     overflow-x: auto;
     overflow-y: hidden;
     white-space: nowrap;

tinygrad/viz/js/index.js

@@ -36,7 +36,7 @@ const updateProgress = ({ start, err }) => {
   d3.select("#custom").html("");
   if (err) {
     displaySelection("#custom");
-    d3.select("#custom").append(() => d3.create("div").classed("raw-text", true).call(s => s.append(() => codeBlock(err, "txt"))).node());
+    d3.select("#custom").append("div").classed("raw-text", true).call(s => s.append(() => codeBlock(err, "txt"))).node();
   }
 }