simplify priority

* amd: add kern to sqtt event

* fix

.github/workflows/benchmark.yml

@@ -199,7 +199,7 @@ jobs:
     - name: Test speed vs torch
       run: NV=1 CAPTURE_PROCESS_REPLAY=0 HALF=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py | tee torch_speed.txt
     - name: Test speed vs theoretical
-      run: NV=1 IGNORE_BEAM_CACHE=1 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
+      run: NV=1 IGNORE_BEAM_CACHE=1 DISABLE_COMPILER_CACHE=1 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
     - name: Test benchmark allreduce
       run: NV=1 python test/external/external_benchmark_multitensor_allreduce.py
     - name: Test tensor cores
@@ -409,7 +409,7 @@ jobs:
     #    python3 -c "import torch; print(torch.__version__)"
     #    LD_PRELOAD="/opt/rocm/lib/libhsa-runtime64.so" HSA=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py | tee torch_speed.txt
     - name: Test speed vs theoretical
-      run: AMD=1 IGNORE_BEAM_CACHE=1 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
+      run: AMD=1 IGNORE_BEAM_CACHE=1 DISABLE_COMPILER_CACHE=1 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
     - name: Test tensor cores
       run: |
         AMD=1 AMD_LLVM=0 python3 test/opt/test_tensor_cores.py

.github/workflows/mlperf.yml

@@ -27,4 +27,4 @@ jobs:
       run: |
         rm "~/.cache/tinygrad/cache_mlperf.db" || true
         BENCHMARK_LOG=mlpert_train_resnet LOGMLPERF=0 CACHEDB="~/.cache/tinygrad/cache_mlperf.db" examples/mlperf/training_submission_v5.1/tinycorp/benchmarks/resnet/implementations/tinybox_red/run_and_time.sh
-        rm "~/.cache/tinygrad/cache_mlperf.db"
+        rm "~/.cache/tinygrad/cache_mlperf.db"

.pre-commit-config.yaml

@@ -28,7 +28,7 @@ repos:
         pass_filenames: false
       - id: tests
         name: subset of tests
-        entry: env OMP_NUM_THREADS=1 PYTHONPATH="." python3 -m pytest -n=8 test/test_ops.py test/test_dtype.py test/test_schedule.py test/test_assign.py
+        entry: env OMP_NUM_THREADS=1 PYTHONPATH="." python3 -m pytest -n=6 test/test_ops.py test/test_dtype.py test/test_schedule.py test/test_assign.py
         language: system
         always_run: true
         pass_filenames: false

autogen_stubs.sh

@@ -31,7 +31,9 @@ $(for p in "$@"; do echo "  $p,"; done)
 ]
 def _try_dlopen_$name():
   library = ctypes.util.find_library("$name")
-  if library: return ctypes.CDLL(library)
+  if library:
+    try: return ctypes.CDLL(library)
+    except OSError: pass
   for candidate in PATHS_TO_TRY:
     try: return ctypes.CDLL(candidate)
     except OSError: pass
@@ -432,11 +434,13 @@ generate_sqtt() {
     $ROCPROF_SRC/include/rocprof_trace_decoder.h \
     $ROCPROF_SRC/include/trace_decoder_instrument.h \
     $ROCPROF_SRC/include/trace_decoder_types.h \
-    -o extra/sqtt/rocprof/rocprof.py
-  fixup extra/sqtt/rocprof/rocprof.py
-  sed -i '1s/^/# pylint: skip-file\n/' extra/sqtt/rocprof/rocprof.py
-  sed -i "s/import ctypes/import ctypes, ctypes.util/g" extra/sqtt/rocprof/rocprof.py
-  sed -i "s|FunctionFactoryStub()|ctypes.CDLL(ctypes.util.find_library('rocprof-trace-decoder'))|g" extra/sqtt/rocprof/rocprof.py
+    -o $BASE/rocprof.py
+  fixup $BASE/rocprof.py
+  sed -i '1s/^/# pylint: skip-file\n/' $BASE/rocprof.py
+  sed -i "s/import ctypes/import ctypes, ctypes.util/g" $BASE/rocprof.py
+  patch_dlopen $BASE/rocprof.py rocprof-trace-decoder "'/usr/local/lib/librocprof-trace-decoder.so'" "'/usr/local/lib/librocprof-trace-decoder.dylib'"
+  sed -i "s/def _try_dlopen_rocprof-trace-decoder():/def _try_dlopen_rocprof_trace_decoder():/g" $BASE/rocprof.py
+  sed -i "s|FunctionFactoryStub()|_try_dlopen_rocprof_trace_decoder()|g" $BASE/rocprof.py
 }
 
 generate_webgpu() {

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/bert/implementations/tinybox_red/run_and_time.sh

@@ -15,7 +15,7 @@ export IGNORE_JIT_FIRST_BEAM=1
 export BASEDIR="/raid/datasets/wiki"
 
 # pip install -e ".[mlperf]"
-export LOGMLPERF=1
+export LOGMLPERF=${LOGMLPERF:-1}
 
 export SEED=$RANDOM
 DATETIME=$(date "+%m%d%H%M")

extra/gemm/amd_uop_matmul.py

@@ -1,10 +1,11 @@
 from tinygrad import Tensor, Device, Context, GlobalCounters, dtypes
-from tinygrad.uop.ops import UOp, KernelInfo
+from tinygrad.uop.ops import UOp, KernelInfo, sint, AxisType
 from tinygrad.engine.realize import ExecItem, get_runner
 from tinygrad.dtype import AddrSpace
 from tinygrad.helpers import getenv
 
 N = 4096
+M = K = N
 run_count = 5
 
 # ---------------------------
@@ -42,26 +43,17 @@ LANES_PER_WAVE_X = 8
 LANES_PER_WAVE_Y = 4
 ITERS_PER_WAVE_N = WAVE_TILE_N // (LANES_PER_WAVE_X * TN)
 ITERS_PER_WAVE_M = WAVE_TILE_M // (LANES_PER_WAVE_Y * TM)
-N_PER_ITER = WAVE_TILE_N // ITERS_PER_WAVE_N
-M_PER_ITER = WAVE_TILE_M // ITERS_PER_WAVE_M
 assert WAVE_TILE_N % (LANES_PER_WAVE_X * TN) == 0, "WAVE_TILE_N must be divisible by LANES_PER_WAVE_X*TN"
 assert WAVE_TILE_M % (LANES_PER_WAVE_Y * TM) == 0, "WAVE_TILE_M must be divisible by LANES_PER_WAVE_Y*TM"
 
+def rngs_for_shape(shape:tuple[sint, ...], rng:int, axis_type=AxisType.LOOP): return [UOp.range(s, rng+i, axis_type) for i,s in enumerate(shape)]
+def copy(dest:UOp, src:UOp, rng:int, set=False, upcast=False):
+  assert dest.shape == src.shape
+  rngs = rngs_for_shape(src.shape, rng, AxisType.UPCAST if upcast else AxisType.LOOP)
+  copy = dest[*rngs].store(src[*rngs]).end(*rngs)
+  return dest.after(copy) if set else copy
+
 def hand_spec_kernel3():
-  # ---------------------------
-  # per-thread read mapping
-  # ---------------------------
-  # A: read BK x BN tiles; B: read BN x BK tiles
-  tid = UOp.special(THREADS_PER_BLOCK, "lidx0")
-
-  waveIdx = (tid // WARP_SIZE) % WAVES_IN_BLOCK_X
-  waveIdy = (tid // WARP_SIZE) // WAVES_IN_BLOCK_X
-  assert waveIdy.vmax+1 == WAVES_IN_BLOCK_Y
-
-  idxInWave = (tid % WARP_SIZE) % LANES_PER_WAVE_X
-  idyInWave = (tid % WARP_SIZE) // LANES_PER_WAVE_X
-  assert idyInWave.vmax+1 == LANES_PER_WAVE_Y
-
   # ---------------------------
   # block indices & placeholders
   # ---------------------------
@@ -72,66 +64,66 @@ def hand_spec_kernel3():
   b = UOp.placeholder((N, N), dtypes.float, slot=2)
   c = UOp.placeholder((N, N), dtypes.float, slot=0)
 
-  BM_As_stride = (BLOCK_M + 4) if is_kernel5 else BLOCK_M
-  As = UOp.placeholder((BLOCK_K, BM_As_stride), dtypes.float, slot=0, addrspace=AddrSpace.LOCAL).shrink_to((BLOCK_K, BLOCK_M))
-  Bs = UOp.placeholder((BLOCK_K, BLOCK_N), dtypes.float, slot=1, addrspace=AddrSpace.LOCAL)
+  # index the output with the globals
+  c = c.reshape(M // BLOCK_M, BLOCK_M, N // BLOCK_N, BLOCK_N)[blockIdx_y, :, blockIdx_x, :]
 
-  A_col = UOp.placeholder((ITERS_PER_WAVE_M, TM), dtypes.float, slot=0, addrspace=AddrSpace.REG)
-  B_row = UOp.placeholder((ITERS_PER_WAVE_N, TN), dtypes.float, slot=1, addrspace=AddrSpace.REG)
-  c_regs = UOp.placeholder((ITERS_PER_WAVE_M, TM, ITERS_PER_WAVE_N, TN), dtypes.float, slot=2, addrspace=AddrSpace.REG)
+  # open the main reduction range
+  k_tile_range = UOp.range(N // BLOCK_K, 0, AxisType.REDUCE)
+  a = a.reshape(M // BLOCK_M, BLOCK_M, N // BLOCK_K, BLOCK_K)[blockIdx_y, :, k_tile_range, :]
+  b = b.reshape(N // BLOCK_K, BLOCK_K, N // BLOCK_N, BLOCK_N)[k_tile_range, :, blockIdx_x, :]
 
-  i = UOp.range(c_regs.size, 16)
-  c_regs = c_regs[i].set(0.0, end=i)
-
-  k_tile_range = UOp.range(N // BLOCK_K, 0)
+  # globals are no longer used, they are already in the indexes
+  del blockIdx_y, blockIdx_x
 
   # ---------------------------
   # GLOBAL -> LOCAL (As, Bs)
   # ---------------------------
-  b = b.reshape(N // BLOCK_K, BLOCK_K,
-                N // BLOCK_N, BLOCK_N)
-  i = UOp.range(BLOCK_N * BLOCK_K // THREADS_PER_BLOCK, 1)
-  index_x = tid % BLOCK_N
-  index_y = (tid // BLOCK_N) + (THREADS_PER_BLOCK // BLOCK_N) * i
-  Bs_store = Bs[index_y, index_x].store(b[k_tile_range, index_y, blockIdx_x, index_x]).end(i)
+  tid = UOp.special(THREADS_PER_BLOCK, "lidx0")
 
-  a = a.reshape(N // BLOCK_M, BLOCK_M,
-                N // BLOCK_K, BLOCK_K)
-  i = UOp.range(BLOCK_M * BLOCK_K // THREADS_PER_BLOCK, 2)
-  index_x = tid % BLOCK_K
-  index_y = (tid // BLOCK_K) + (THREADS_PER_BLOCK // BLOCK_K) * i
-  As_store = As[index_x, index_y].store(a[blockIdx_y, index_y, k_tile_range, index_x]).end(i)
+  # A: read BM x BK tiles (permute on store into locals)
+  BM_As_stride = (BLOCK_M + 4) if is_kernel5 else BLOCK_M
+  As = UOp.placeholder((BLOCK_K, BM_As_stride), dtypes.float, slot=0, addrspace=AddrSpace.LOCAL).shrink_to((BLOCK_K, BLOCK_M))
+  As_store = copy(As.permute((1,0)).reshape(-1, THREADS_PER_BLOCK)[:, tid], a.reshape(-1, THREADS_PER_BLOCK)[:, tid], rng=100)
+
+  # B: read BK x BN tiles
+  Bs = UOp.placeholder((BLOCK_K, BLOCK_N), dtypes.float, slot=1, addrspace=AddrSpace.LOCAL)
+  Bs_store = copy(Bs.reshape(-1, THREADS_PER_BLOCK)[:, tid], b.reshape(-1, THREADS_PER_BLOCK)[:, tid], rng=200)
 
   # TODO: can we automate barrier?
   barrier = UOp.barrier(As_store, Bs_store)
-  Bs = Bs.after(barrier)
-  As = As.after(barrier)
+  As, Bs = As.after(barrier), Bs.after(barrier)
 
   # open inner k range
-  k = UOp.range(BLOCK_K, 3)
+  k = UOp.range(BLOCK_K, 3, AxisType.REDUCE)
 
   # ---------------------------
   # LOCAL -> REG (per-wave tiles)
   # ---------------------------
-  Bs_view = Bs.reshape(BLOCK_K, WAVES_IN_BLOCK_X, ITERS_PER_WAVE_N, LANES_PER_WAVE_X, TN)
-  iterWaveN = UOp.range(ITERS_PER_WAVE_N, 4)
-  i = UOp.range(TN, 5)
-  B_row = B_row[iterWaveN, i].set(Bs_view[k, waveIdx, iterWaveN, idxInWave, i], end=(iterWaveN, i))
+  waveIdx = (tid // WARP_SIZE) % WAVES_IN_BLOCK_X
+  waveIdy = (tid // WARP_SIZE) // WAVES_IN_BLOCK_X
+  assert waveIdy.vmax+1 == WAVES_IN_BLOCK_Y
 
-  As_view = As.reshape(BLOCK_K, WAVES_IN_BLOCK_Y, ITERS_PER_WAVE_M, LANES_PER_WAVE_Y, TM)
-  iterWaveM = UOp.range(ITERS_PER_WAVE_M, 6)
-  i = UOp.range(TM, 7)
-  A_col = A_col[iterWaveM, i].set(As_view[k, waveIdy, iterWaveM, idyInWave, i], end=(iterWaveM, i))
+  laneIdx = (tid % WARP_SIZE) % LANES_PER_WAVE_X
+  laneIdy = (tid % WARP_SIZE) // LANES_PER_WAVE_X
+  assert laneIdy.vmax+1 == LANES_PER_WAVE_Y
+
+  A_col = UOp.placeholder((ITERS_PER_WAVE_M, TM), dtypes.float, slot=0, addrspace=AddrSpace.REG)
+  A_col = copy(A_col, As[k, :].reshape(WAVES_IN_BLOCK_Y, ITERS_PER_WAVE_M, LANES_PER_WAVE_Y, TM)[waveIdy, :, laneIdy, :], 300, set=True, upcast=True)
+
+  B_row = UOp.placeholder((ITERS_PER_WAVE_N, TN), dtypes.float, slot=1, addrspace=AddrSpace.REG)
+  B_row = copy(B_row, Bs[k, :].reshape(WAVES_IN_BLOCK_X, ITERS_PER_WAVE_N, LANES_PER_WAVE_X, TN)[waveIdx, :, laneIdx, :], 400, set=True, upcast=True)
 
   # ---------------------------
   # FMA: c_regs += A_col * B_row
   # ---------------------------
-  iterWaveM = UOp.range(ITERS_PER_WAVE_M, 8)
-  yt = UOp.range(TM, 9)
-  iterWaveN = UOp.range(ITERS_PER_WAVE_N, 10)
-  xt = UOp.range(TN, 12)
-  c_idx = c_regs.after(k, k_tile_range)[iterWaveM, yt, iterWaveN, xt]
-  sink = c_idx.store(c_idx + A_col[iterWaveM, yt] * B_row[iterWaveN, xt]).end(iterWaveM, iterWaveN, yt, xt)
+  c_regs = UOp.placeholder((ITERS_PER_WAVE_M, TM, ITERS_PER_WAVE_N, TN), dtypes.float, slot=2, addrspace=AddrSpace.REG)
+  i = UOp.range(c_regs.size, 16)
+  c_regs = c_regs.after(c_regs.flatten()[i].store(0.0).end(i))
+
+  # TODO: why don't these work as upcast?
+  # why if the ranges merge is it slow?!? (if you change the order on end, they will merge. big slowdown on METAL)
+  iterWaveM, yt, iterWaveN, xt = rngs = rngs_for_shape(c_regs.shape, 500)
+  sink = c_regs[*rngs].store(c_regs.after(k)[*rngs] + A_col[iterWaveM, yt] * B_row[iterWaveN, xt]).end(iterWaveM, iterWaveN, yt, xt)
 
   # Close k, sync, and close K tiles
   sink = sink.end(k).barrier().end(k_tile_range)
@@ -139,15 +131,11 @@ def hand_spec_kernel3():
   # ---------------------------
   # REG -> GLOBAL (epilogue)
   # ---------------------------
-  c = c.reshape(N//BLOCK_M, WAVES_IN_BLOCK_Y, ITERS_PER_WAVE_M, LANES_PER_WAVE_Y, TM,
-                N//BLOCK_N, WAVES_IN_BLOCK_X, ITERS_PER_WAVE_N, LANES_PER_WAVE_X, TN)
-  iterWaveM = UOp.range(ITERS_PER_WAVE_M, 1000)
-  yt = UOp.range(TM, 1001)
-  iterWaveN = UOp.range(ITERS_PER_WAVE_N, 1002)
-  xt = UOp.range(TN, 1003)
-  c_glbl_idx = c[blockIdx_y, waveIdy, iterWaveM, idyInWave, yt, blockIdx_x, waveIdx, iterWaveN, idxInWave, xt]
-  sink = c_glbl_idx.store(c_regs.after(sink)[iterWaveM, yt, iterWaveN, xt])
-  sink = sink.end(iterWaveM, iterWaveN, yt, xt)
+  c = c.reshape(WAVES_IN_BLOCK_Y, ITERS_PER_WAVE_M, LANES_PER_WAVE_Y, TM,
+                WAVES_IN_BLOCK_X, ITERS_PER_WAVE_N, LANES_PER_WAVE_X, TN)
+  c = c[waveIdy, :, laneIdy, :,
+        waveIdx, :, laneIdx, :]
+  sink = copy(c, c_regs.after(sink), rng=600)
 
   return sink.sink(arg=KernelInfo(opts_to_apply=())).simplify()
 

extra/gemm/max_matmul.py

@@ -1,17 +1,11 @@
 import numpy as np, os
 from tinygrad.helpers import getenv, flat_mv
 from tinygrad import dtypes
-from typing import Optional, List, Tuple, cast, Dict, Final, DefaultDict, Self
 from tinygrad.engine.realize import get_program
 
 # for copied uops
-from tinygrad.codegen.opt.kernel import Kernel, KernelOptError
-from tinygrad.uop.ops import UOp, Ops, BinaryOps, UnaryOps, TernaryOps, KernelInfo
-from tinygrad.codegen.opt.search import Opt, OptOps
-from tinygrad import Device, dtypes, Tensor
-from tinygrad.dtype import PtrDType, DType, DTYPES_DICT
-from tinygrad.shape.shapetracker import ShapeTracker
-from tinygrad.shape.view import View
+from tinygrad import dtypes
+from tinygrad.dtype import DTYPES_DICT
 
 script_dir = os.path.dirname(os.path.abspath(__file__))
 
@@ -53,12 +47,6 @@ def randoms():
     nc = nc.astype(np.bfloat16 if DTYPE_IN == dtypes.bfloat16 else np.float16)
   return na, nb, nc
 
-def ast_to_cuda_prog(compiler, ast, opts):
-  k = Kernel(ast)
-  k.apply_opts(opts)
-  p = get_program(k.ast, k.opts, k.applied_opts)
-  return CUDAProgram(device, p.function_name, compiler.compile(p.src))
-
 if __name__ == "__main__":
   print(f"gemm variation: {GEMM_VARIATION=} {M=} {N=} {K=} {DTYPE_IN=} {DTYPE_OUT=} {DTYPE_ACC=}")
   prog, global_size, local_size = None, None, None
@@ -189,11 +177,11 @@ if __name__ == "__main__":
 
     tms = []
     na, nb, nc = randoms()
-    cudaalloc.copyin(a, bytearray(na))
-    cudaalloc.copyin(b, bytearray(nb))
+    cudaalloc._copyin(a, memoryview(bytearray(na)))
+    cudaalloc._copyin(b, memoryview(bytearray(nb)))
     for i in range(CNT):
       tms.append(prog(*args, **kwargs))
-    cudaalloc.copyout(flat_mv(nc.data), c)
+    cudaalloc._copyout(flat_mv(nc.data), c)
     comp = na.astype(np.float32) @ nb.astype(np.float32)
     result = nc.reshape(M, N).astype(np.float32)
 

extra/gemm/mi350x_uop_matmul.py

@@ -37,11 +37,11 @@ WARPGROUP_SIZE = 1
 BLOCK_M = BLOCK_M * WARPGROUP_SIZE
 
 # TODO: improve the syntax of this. better syntax, faster iteration
-#  -- add working slice a[gx, :, i] -> shape of the : (aka (16,16,32) becomes (16,))
-#  -- add argfix to movement (traits shared with Tensor)
+#  -- DONE: add working slice a[gx, :, i] -> shape of the : (aka (16,16,32) becomes (16,))
+#  -- DONE(ish): add argfix to movement (traits shared with Tensor)
 #  -- fix WMMA to not require all the junk
 #  -- improve syntax for vectorized loads/stores (both with DEVECTORIZE and without)
-#  -- be able to use CONTRACT on a range
+#  -- DONE: be able to use CONTRACT on a range
 #  -- fix upcasted RANGE on an already vectorized buffer
 #  -- improve "all ranges not ended error" / fix the bug with after on ended ranges (if you are after end of range, range is closed)
 

extra/gemm/simple_matmul.py

@@ -17,7 +17,7 @@ M = getenv("M", N)
 K = getenv("K", N)
 CNT = getenv("CNT", 10)
 
-atol, rtol = {dtypes.bfloat16:(1e-3, 1e-2), dtypes.fp8e4m3:(1e-1, 1e-1), dtypes.fp8e5m2:(1.0, 5e-1)}.get(dtype_in, (1e-4, 3e-2))
+atol, rtol = {dtypes.half:{1e-3, 1e-2}, dtypes.bfloat16:(1e-3, 1e-2), dtypes.fp8e4m3:(1e-1, 1e-1), dtypes.fp8e5m2:(1.0, 5e-1)}.get(dtype_in, (1e-4, 3e-2))
 ATOL, RTOL = getenv("ATOL", atol), getenv("RTOL", rtol)
 
 INT_LOW = getenv("INT_LOW", 0)

extra/sqtt/disasm.py

@@ -1,68 +0,0 @@
-import ctypes
-from dataclasses import dataclass
-import tinygrad.runtime.autogen.comgr as comgr
-from tinygrad.runtime.support.compiler_amd import check
-
-@dataclass
-class InstrCtx:
-  pc:int=0
-  inst:str=""
-
-@comgr.amd_comgr_create_disassembly_info.argtypes[2]
-def instr_cb(text, user_data):
-  c = ctypes.cast(user_data, ctypes.POINTER(ctypes.py_object)).contents.value
-  c.inst = ctypes.string_at(text).decode("utf-8","replace").strip()
-  return comgr.AMD_COMGR_STATUS_SUCCESS
-
-# nop callback
-@comgr.amd_comgr_create_disassembly_info.argtypes[3]
-def addr_cb(*args): return comgr.AMD_COMGR_STATUS_SUCCESS
-
-def comgr_get_address_table(lib:bytes) -> dict[int, tuple[str, int]]:
-  check(comgr.amd_comgr_create_data(comgr.AMD_COMGR_DATA_KIND_EXECUTABLE, ctypes.byref(data_src:=comgr.amd_comgr_data_t())))
-  lib_buf = ctypes.create_string_buffer(lib, len(lib))
-  check(comgr.amd_comgr_set_data(data_src, len(lib), lib_buf))
-  check(comgr.amd_comgr_get_data_isa_name(data_src, isa_sz:=ctypes.c_size_t(128), isa:=(ctypes.c_char*isa_sz.value)()))
-
-  @comgr.amd_comgr_create_disassembly_info.argtypes[1]
-  def memory_cb(from_addr, to, size, _):
-    base, buf_len = ctypes.addressof(lib_buf), len(lib_buf)
-    start = int(from_addr) - base
-    if start < 0 or start >= buf_len: return 0
-    ctypes.memmove(to, base + start, n:=min(int(size), buf_len - start))
-    return n
-
-  info_src = comgr.amd_comgr_disassembly_info_t()
-  check(comgr.amd_comgr_create_disassembly_info(ctypes.cast(isa, ctypes.POINTER(ctypes.c_char)), memory_cb, instr_cb, addr_cb, info_src))
-
-  @comgr.amd_comgr_iterate_symbols.argtypes[1]
-  def sym_callback(sym, udata):
-    check(comgr.amd_comgr_symbol_get_info(sym, comgr.AMD_COMGR_SYMBOL_INFO_TYPE, ctypes.byref(sym_type:=ctypes.c_int())))
-    if sym_type.value != comgr.AMD_COMGR_SYMBOL_TYPE_FUNC: return comgr.AMD_COMGR_STATUS_SUCCESS
-    check(comgr.amd_comgr_symbol_get_info(sym, comgr.AMD_COMGR_SYMBOL_INFO_VALUE, ctypes.byref(vaddr:=ctypes.c_uint64())))
-    check(comgr.amd_comgr_symbol_get_info(sym, comgr.AMD_COMGR_SYMBOL_INFO_SIZE, ctypes.byref(size:=ctypes.c_uint64())))
-    check(comgr.amd_comgr_map_elf_virtual_address_to_code_object_offset(data_src, vaddr.value, ctypes.byref(offset:=ctypes.c_uint64()),
-                                                                        ctypes.byref(ctypes.c_uint64()), ctypes.byref(nobits:=ctypes.c_bool())))
-    check(nobits.value)
-    base = ctypes.addressof(lib_buf)
-    pc = base + offset.value
-    end = pc + size.value
-    addr_table = ctypes.cast(udata, ctypes.POINTER(ctypes.py_object)).contents.value
-    instr_ref = ctypes.py_object(ctx:=InstrCtx())
-    instr_ptr = ctypes.cast(ctypes.pointer(instr_ref), ctypes.c_void_p)
-    while pc < end:
-      size_read = ctypes.c_uint64(0)
-      ctx.pc = pc
-      st = comgr.amd_comgr_disassemble_instruction(info_src, ctypes.c_uint64(pc), instr_ptr, ctypes.byref(size_read))
-      if st == comgr.AMD_COMGR_STATUS_SUCCESS and size_read.value:
-        rel = (pc - base) - offset.value
-        addr_table[vaddr.value + rel] = (ctx.inst, int(size_read.value))
-        pc += size_read.value
-      else: # don't inf loop if comgr fails
-        b = ctypes.c_ubyte.from_buffer(lib_buf, pc - base).value
-        addr_table[vaddr.value + (pc - base - offset.value)] = (f"DISASSEMBLER ISSUE 0x{b:02x}", 1)
-        pc += 1
-    return comgr.AMD_COMGR_STATUS_SUCCESS
-  addr_table:dict[int, tuple[str, int]] = {}
-  check(comgr.amd_comgr_iterate_symbols(data_src, sym_callback, ctypes.cast(ctypes.pointer(ctypes.py_object(addr_table)), ctypes.c_void_p)))
-  return addr_table

extra/sqtt/install_sqtt_decoder.py

@@ -13,6 +13,6 @@ if __name__ == "__main__":
     os.chmod(fp, 0o755)
     os.system(f"sudo {fp} --prefix={fp.parent} --include-subdir")
   else:
-    lib = fetch("https://github.com/ROCm/rocprof-trace-decoder/raw/5420409ad0963b2d76450add067b9058493ccbd0/releases/linux_glibc_2_28_x86_64/librocprof-trace-decoder.so", name="librocprof-trace-decoder.so")
+    lib = fetch("https://github.com/ROCm/rocprof-trace-decoder/raw/43bf0fef74a83c3c25badfc5a09c0bd39ed8c6f9/releases/linux_glibc_2_28_x86_64/librocprof-trace-decoder.so", name="librocprof-trace-decoder.so")
   shutil.copy2(lib, DEST)
   print(f"Installed {lib.name} to", DEST)

extra/sqtt/rgptool.py

@@ -154,8 +154,22 @@ class RGP:
       if device not in device_events: raise RuntimeError(f"Device {device} not found in profile, devices in profile: {', '.join(device_events.keys())} ")
       device_event = device_events[device]
     sqtt_events = [x for x in profile if isinstance(x, ProfileSQTTEvent) and x.device == device_event.device]
+    device_props = device_event.props
+    # merge events per SE
+    merged_sqtt_events:dict[int, ProfileSQTTEvent] = {}
+    for ev in sqtt_events:
+      if ev.se not in merged_sqtt_events: merged_sqtt_events[ev.se] = ev
+      else:
+        merged_sqtt_events[ev.se] = ProfileSQTTEvent(
+          device=ev.device,
+          kern=ev.kern,
+          se=ev.se,
+          itrace=merged_sqtt_events[ev.se].itrace or ev.itrace,
+          blob=merged_sqtt_events[ev.se].blob + ev.blob,
+        )
+    sqtt_events = list(merged_sqtt_events.values())
+
     if len(sqtt_events) == 0: raise RuntimeError(f"Device {device_event.device} doesn't contain SQTT data")
-    device_props = sqtt_events[0].props
     gfx_ver = device_props['gfx_target_version'] // 10000
     gfx_iplvl = getattr(sqtt, f"SQTT_GFXIP_LEVEL_GFXIP_{device_props['gfx_target_version']//10000}_{(device_props['gfx_target_version']//100)%100}",
                         getattr(sqtt, f"SQTT_GFXIP_LEVEL_GFXIP_{device_props['gfx_target_version']//10000}", None))

extra/sqtt/roc.py

@@ -1,9 +1,32 @@
 import ctypes, pathlib, argparse, pickle, re, functools, dataclasses, itertools
-from extra.sqtt.rocprof import rocprof
-from extra.sqtt.disasm import comgr_get_address_table
-from tinygrad.helpers import temp, DEBUG
-from tinygrad.device import ProfileEvent, ProfileProgramEvent
+from tinygrad.helpers import temp, unwrap, DEBUG
+from tinygrad.device import ProfileEvent, ProfileDeviceEvent, ProfileProgramEvent
 from tinygrad.runtime.ops_amd import ProfileSQTTEvent, ProfilePMCEvent
+from tinygrad.runtime.autogen import llvm, rocprof
+from tinygrad.runtime.support.elf import elf_loader
+
+# to pass NULL to callbacks
+llvm.LLVMCreateDisasmCPUFeatures.argtypes = tuple(llvm.LLVMCreateDisasmCPUFeatures.argtypes[:5]) + (ctypes.c_void_p, ctypes.c_void_p)
+def llvm_disasm(arch:str, lib:bytes) -> dict[int, tuple[str, int]]:
+  llvm.LLVMInitializeAMDGPUTargetInfo()
+  llvm.LLVMInitializeAMDGPUTargetMC()
+  llvm.LLVMInitializeAMDGPUAsmParser()
+  llvm.LLVMInitializeAMDGPUDisassembler()
+  ctx = llvm.LLVMCreateDisasmCPUFeatures("amdgcn-amd-amdhsa".encode(), arch.encode(), "".encode(), None, 0, None, None)
+
+  image, sections, relocs = elf_loader(lib)
+  text = next((sh.header for sh in sections if sh.name == ".text"), None)
+  off, sz = unwrap(text).sh_addr, unwrap(text).sh_size
+
+  addr_table:dict[int, tuple[str, int]] = {}
+  out = ctypes.create_string_buffer(128)
+  cur_off = off
+  while cur_off < sz + off:
+    view = (ctypes.c_ubyte * ((sz + off) - cur_off)).from_buffer_copy(memoryview(image)[cur_off:])
+    instr_sz = llvm.LLVMDisasmInstruction(ctx, view, ctypes.c_uint64(len(view)), ctypes.c_uint64(0), out, ctypes.c_size_t(128))
+    addr_table[cur_off] = (out.value.decode("utf-8", "replace").strip(), instr_sz)
+    cur_off += instr_sz
+  return addr_table
 
 @dataclasses.dataclass
 class InstInfo:
@@ -17,53 +40,70 @@ class InstInfo:
   def on_ev(self, ev):
     self.hit, self.lat, self.stall = self.hit + 1, self.lat + ev.duration, self.stall + ev.stall
 
+@dataclasses.dataclass(frozen=True)
+class InstExec:
+  typ:str
+  inst:str
+  stall:int
+  dur:int
+  time:int
+
+@dataclasses.dataclass(frozen=True)
+class PrgExec:
+  name:str
+  wave:int
+  cu:int
+  simd:int
+  def __str__(self): return f"{self.name},{self.wave},{self.cu},{self.simd}"
+
 class _ROCParseCtx:
-  def __init__(self, sqtt_evs:list[ProfileSQTTEvent], prog_evs:list[ProfileProgramEvent]):
-    self.sqtt_evs, self.prog_evs = iter(sqtt_evs), prog_evs
-    self.wave_events, self.disasms, self.addr2prg = {}, {}, {}
+  def __init__(self, dev_evs:dict[str, ProfileDeviceEvent], sqtt_evs:list[ProfileSQTTEvent], prog_evs:list[ProfileProgramEvent]):
+    self.dev_evs, self.sqtt_evs, self.prog_evs = dev_evs, iter(sqtt_evs), prog_evs
+    self.wave_events:dict[PrgExec, dict[int, InstInfo]] = {}
+    self.disasms:dict[int, tuple[str, int]] = {}
+    self.inst_execs:dict[PrgExec, list[InstExec]] = {}
 
     for prog in prog_evs:
-      for addr, info in comgr_get_address_table(prog.lib).items():
-        self.disasms[prog.base + addr] = info
-        self.addr2prg[prog.base + addr] = prog
+      arch = "gfx%d%x%x" % ((trgt:=unwrap(dev_evs[prog.device].props)['gfx_target_version']) // 10000, (trgt // 100) % 100, trgt % 100)
+      for addr, info in llvm_disasm(arch, unwrap(prog.lib)).items():
+        self.disasms[(prog.name, unwrap(prog.base) + addr)] = info
 
   def next_sqtt(self):
     x = next(self.sqtt_evs, None)
+    self.active_kern = x.kern if x is not None else None
     self.active_se = x.se if x is not None else None
     return x
 
-  def find_program(self, addr): return self.addr2prg[addr]
-
   def on_occupancy_ev(self, ev):
-    if DEBUG >= 4: print("OCC", ev.time, self.active_se, ev.cu, ev.simd, ev.wave_id, ev.start)
+    if DEBUG >= 5: print("OCC", ev.time, self.active_se, ev.cu, ev.simd, ev.wave_id, ev.start)
 
   def on_wave_ev(self, ev):
-    if DEBUG >= 4: print("WAVE", ev.wave_id, self.active_se, ev.cu, ev.simd, ev.contexts, ev.begin_time, ev.end_time)
+    if DEBUG >= 5: print("WAVE", ev.wave_id, self.active_se, ev.cu, ev.simd, ev.contexts, ev.begin_time, ev.end_time)
 
-    asm = {}
+    asm:dict[int, InstInfo] = {}
+    inst_execs:list[InstExec] = []
     for j in range(ev.instructions_size):
       inst_ev = ev.instructions_array[j]
       inst_typ = rocprof.rocprofiler_thread_trace_decoder_inst_category_t__enumvalues[inst_ev.category]
-      asm.setdefault(inst_ev.pc.address, InstInfo(typ=inst_typ, inst=self.disasms[inst_ev.pc.address][0]))
+      inst_disasm = self.disasms[(self.active_kern, inst_ev.pc.address)][0]
+      asm.setdefault(inst_ev.pc.address, InstInfo(typ=inst_typ, inst=inst_disasm))
       asm[inst_ev.pc.address].on_ev(inst_ev)
+      inst_execs.append(InstExec(inst_typ, inst_disasm, inst_ev.stall, inst_ev.duration, inst_ev.time))
 
-    self.wave_events[(self.find_program(ev.instructions_array[0].pc.address).name, ev.wave_id, ev.cu, ev.simd)] = asm
+    if ev.instructions_size > 0:
+      self.wave_events[key:=PrgExec(self.active_kern, ev.wave_id, ev.cu, ev.simd)] = asm
+      self.inst_execs[key] = inst_execs
 
-if __name__ == "__main__":
-  parser = argparse.ArgumentParser()
-  parser.add_argument('--profile', type=pathlib.Path, help='Path to profile', default=pathlib.Path(temp("profile.pkl", append_user=True)))
-  args = parser.parse_args()
-
-  with args.profile.open("rb") as f: profile = pickle.load(f)
+def decode(profile:list[ProfileEvent]) -> _ROCParseCtx:
+  dev_events:dict[str, ProfileDeviceEvent] = {}
   sqtt_events:list[ProfileSQTTEvent] = []
-  pmc_events:list[ProfilePMCEvent] = []
   prog_events:list[ProfileProgramEvent] = []
   for e in profile:
+    if isinstance(e, ProfileDeviceEvent): dev_events[e.device] = e
     if isinstance(e, ProfileSQTTEvent): sqtt_events.append(e)
-    if isinstance(e, ProfilePMCEvent): pmc_events.append(e)
     if isinstance(e, ProfileProgramEvent) and e.device.startswith("AMD"): prog_events.append(e)
 
-  ROCParseCtx = _ROCParseCtx(sqtt_events, prog_events)
+  ROCParseCtx = _ROCParseCtx(dev_events, sqtt_events, prog_events)
 
   @rocprof.rocprof_trace_decoder_se_data_callback_t
   def copy_cb(buf, buf_size, data_ptr):
@@ -80,12 +120,12 @@ if __name__ == "__main__":
       case rocprof.ROCPROFILER_THREAD_TRACE_DECODER_RECORD_WAVE:
         for ev in (rocprof.rocprofiler_thread_trace_decoder_wave_t * n).from_address(events_ptr): ROCParseCtx.on_wave_ev(ev)
       case _:
-        if DEBUG >= 2: print(rocprof.rocprofiler_thread_trace_decoder_record_type_t__enumvalues[record_type], events_ptr, n)
+        if DEBUG >= 5: print(rocprof.rocprofiler_thread_trace_decoder_record_type_t__enumvalues[record_type], events_ptr, n)
     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):
-    instr, mem_size_ptr[0] = ROCParseCtx.disasms[pc.address]
+    instr, mem_size_ptr[0] = ROCParseCtx.disasms[(ROCParseCtx.active_kern, pc.address)]
 
     # this is the number of bytes to next instruction, set to 0 for end_pgm
     if instr == "s_endpgm": mem_size_ptr[0] = 0
@@ -100,10 +140,20 @@ if __name__ == "__main__":
 
   try:
     rocprof.rocprof_trace_decoder_parse_data(copy_cb, trace_cb, isa_cb, None)
-    print('SQTT:', ROCParseCtx.wave_events.keys())
-  except Exception as e: print("Error in sqtt decoder:", e)
+  except AttributeError as e: raise RuntimeError("Failed to find rocprof-trace-decoder. Run ./extra/sqtt/install_sqtt_decoder.py to install") from e
+  return ROCParseCtx
 
-  for ev in pmc_events:
+if __name__ == "__main__":
+  parser = argparse.ArgumentParser()
+  parser.add_argument('--profile', type=pathlib.Path, help='Path to profile', default=pathlib.Path(temp("profile.pkl", append_user=True)))
+  args = parser.parse_args()
+
+  with args.profile.open("rb") as f: profile = pickle.load(f)
+  rctx = decode(profile)
+  print('SQTT:', rctx.wave_events.keys())
+
+  for ev in profile:
+    if not isinstance(ev, ProfilePMCEvent): continue
     print(f"PMC Event: dev={ev.device} kern={ev.kern}")
     ptr = 0
     for s in ev.sched:

extra/sqtt/test_timing.py

@@ -0,0 +1,104 @@
+import os
+os.environ["PYTHONPATH"] = "."
+os.environ["SQTT"] = "1"
+os.environ["AMD"] = "1"
+os.environ["VIZ"] = "1"
+os.environ["AMD_LLVM"] = "0"
+
+import unittest
+import sys, contextlib
+from tinygrad import Tensor
+from tinygrad.dtype import dtypes
+from tinygrad.renderer import ProgramSpec
+from tinygrad.uop.ops import UOp, Ops, KernelInfo
+from tinygrad.engine.realize import CompiledRunner
+from tinygrad.device import Device, ProfileDeviceEvent
+
+from extra.sqtt.roc import decode, InstExec, PrgExec
+
+dev = Device["AMD"]
+
+def asm_kernel(instrs:list[str], l:int=1, g:int=1) -> Tensor:
+  name = sys._getframe(1).f_code.co_name
+  def fxn(_):
+    L = UOp.special(l, "lidx0")
+    G = UOp.special(g, "gidx0")
+    ops:list[str] = [UOp(Ops.CUSTOM, arg="asm volatile (")]
+    for inst in instrs: ops.append(UOp(Ops.CUSTOM, src=(ops[-1],), arg=f'  "{inst}\\n\\t"'))
+    ops.append(UOp(Ops.CUSTOM, src=(ops[-1],), arg=");"))
+    return UOp.sink(*ops, L, G, arg=KernelInfo(name=name))
+  k = Tensor.custom_kernel(Tensor.empty(1), fxn=fxn)[0]
+  return k
+
+@contextlib.contextmanager
+def save_sqtt():
+  # clear the old traces
+  dev.profile_events.clear()
+  sqtt:dict[PrgExec, list[InstExec]] = {}
+  yield sqtt
+  # decode sqtt
+  rctx = decode(dev.profile_events+[ProfileDeviceEvent("AMD", props=dev.device_props())])
+  assert len(rctx.inst_execs) > 0, "empty sqtt output"
+  sqtt.update(rctx.inst_execs)
+
+class TestTiming(unittest.TestCase):
+  def test_v_add(self):
+    with save_sqtt() as sqtt:
+      asm_kernel([f"v_add_f32 v{10+i} v{10+i+1} {10+i}" for i in range(3)]).realize()
+    wave = list(sqtt.values())[0][:-1]
+    assert all(s.dur == 1 for s in wave)
+    assert all(s.stall == 0 for s in wave)
+
+  def test_chain_v_add_1l(self):
+    with save_sqtt() as sqtt:
+      asm_kernel([
+        "v_add_f32_e32 v1 v0 v0",
+        "v_add_f32_e32 v2 v1 v1",
+      ]).realize()
+    wave = list(sqtt.values())[0][:-1]
+    assert all(s.dur == 1 for s in wave)
+    assert all(s.stall == 0 for s in wave)
+
+  def test_multi_cycle_inst(self):
+    with save_sqtt() as sqtt:
+      asm_kernel([
+        "v_mov_b32_e32 v4 0x3f800000",
+        "v_rcp_f32_e32 v5 v4",
+        "v_mul_f32_e32 v6 v5 v4",
+      ]).realize()
+    w = list(sqtt.values())[0]
+    rcp, mul = w[1], w[2]
+    self.assertGreater(rcp.dur, 1) # 4 cycles on gfx11
+    self.assertEqual(mul.dur, 1)
+    # mul depends on v5, how can it run before rcp is done?
+    self.assertGreaterEqual(mul.time, rcp.time+rcp.dur)
+
+  def test_wmma(self):
+    with save_sqtt() as sqtt:
+      asm_kernel([
+        "v_wmma_f32_16x16x16_f16 v[16:23], v[0:7], v[8:15], v[16:23]",
+        "v_add_f32_e32 v0 v16 v0",
+      ], l=32*4).realize()
+    assert len(sqtt) == 2, f"expected two waves, got {len(sqtt)} {list(sqtt.keys())}"
+    wmma = list(sqtt.values())[0][0]
+    self.assertGreater(wmma.dur, 1) # rgp says 32 clocks
+
+  def test_sleep(self):
+    n = 1
+    def sleep_kernel(data0):
+      assert data0.dtype.base == dtypes.ulong
+      ops:list[UOp] = []
+      ops.append(UOp(Ops.CUSTOM, arg="unsigned long long t0 = __builtin_readcyclecounter();"))
+      ops.append(UOp(Ops.CUSTOM, arg=f"__builtin_amdgcn_s_sleep({n});", src=(ops[-1],)))
+      ops.append(UOp(Ops.CUSTOM, arg="unsigned long long t1 = __builtin_readcyclecounter();", src=(ops[-1],)))
+      ops.append(UOp(Ops.CUSTOM, arg=f"data0_{data0.size}[0] = t1 - t0;", src=(ops[-1],)))
+      return UOp.sink(data0, *ops, arg=KernelInfo(name=f"sleep_{n}"))
+    diff_hw_reg = Tensor.empty(1, dtype=dtypes.ulong)
+    diff_hw_reg = Tensor.custom_kernel(diff_hw_reg, fxn=sleep_kernel)[0]
+    with save_sqtt() as sqtt:
+      diff_hw_reg.realize()
+    diff_sqtt = list(sqtt.values())[0][2]
+    self.assertEqual(diff_sqtt.dur, diff_hw_reg.item()-1) # 1 cycle for reading the counter register
+
+if __name__ == "__main__":
+  unittest.main()

extra/thunder/cuda/fa.cu

@@ -10,7 +10,7 @@ constexpr int ATTN_N = 1024;
 constexpr int ATTN_H = 16;
 constexpr int ATTN_D = 64;
 
-template<int D> constexpr size_t ROWS = 16*(128/D); // height of each worker tile (rows)
+template<int D> constexpr size_t ROWS = 16*(64/D); // height of each worker tile (rows)
 template<int D, typename T=bf16, typename L=row_l> using qkvo_tile = rt<T, ROWS<D>, D, L>;
 template<int D, typename T=float> using attn_tile = rt<T, ROWS<D>, ROWS<D>>;
 template<int D> using shared_tile = st_bf<ROWS<D>, D>;

extra/thunder/cuda/fa.py

@@ -23,7 +23,7 @@ if __name__ == "__main__":
   Tensor.realize(q, k, v, out)
 
   NUM_WORKERS = 4
-  ROWS = 16 * (128 // D)
+  ROWS = 16 * (64 // D)
 
   gsz = (N // (ROWS*NUM_WORKERS), H, B)
   for _ in range(5):

pytest.ini

@@ -1,5 +1,8 @@
 [pytest]
-norecursedirs = extra
+norecursedirs =
+    extra
+    .hypothesis
+    .git
 timeout = 300
 timeout_method = thread
 timeout_func_only = true

setup.py

@@ -32,6 +32,7 @@ setup(name='tinygrad',
         'tinygrad.codegen.opt',
         'tinygrad.codegen.late',
         'tinygrad.engine',
+        'tinygrad.mixin',
         'tinygrad.nn',
         'tinygrad.renderer',
         'tinygrad.runtime',

test/external/external_fuzz_am_interrupts.py

@@ -1,39 +0,0 @@
-import subprocess
-import random
-import time
-from concurrent.futures import ThreadPoolExecutor, as_completed
-
-def run_test(i, full_run=False):
-  print(f"\rRunning iteration {i}...", end=" ", flush=True)
-
-  p = subprocess.Popen(['python3', 'test/test_tiny.py', 'TestTiny.test_plus'],  stdout=subprocess.PIPE, stderr=subprocess.PIPE)
-
-  if not full_run:
-    time.sleep(random.uniform(0, 1200) / 1000)
-    p.kill()
-    _, stderr = p.communicate()
-  else:
-    _, stderr = p.communicate()
-
-  if full_run:
-    stderr_text = stderr.decode()
-    print(stderr_text)
-    assert "Ran 1 test in" in stderr_text and "OK" in stderr_text
-
-max_workers = 4
-with ThreadPoolExecutor(max_workers=max_workers) as executor:
-  futures = []
-  for i in range(1000000):
-    if i % 100 == 0:
-      for future in as_completed(futures):
-        try: future.result()
-        except Exception as e:
-          print(f"\nError in iteration: {e}")
-      futures = []
-
-      run_test(i, True)
-    else:
-      future = executor.submit(run_test, i, False)
-      futures.append(future)
-
-    if len(futures) > max_workers * 2: futures = [f for f in futures if not f.done()]

test/external/external_fuzz_hcq_mp.py

@@ -0,0 +1,44 @@
+import subprocess
+import random
+import time
+from concurrent.futures import ProcessPoolExecutor, as_completed
+from tinygrad.helpers import getenv
+
+# checks that HCQ drivers can be killed during operation without causing issues
+
+def run_test(i, full_run=False, force_ok=False):
+  print(f"\rRunning iteration {i}...", end=" ", flush=True)
+
+  p = subprocess.Popen(["python3", "test/test_tiny.py", "TestTiny.test_plus"], stdout=subprocess.PIPE, stderr=subprocess.PIPE)
+
+  if not full_run:
+    time.sleep(random.uniform(0, 1200) / 1000.0)
+    p.kill()
+    _, stderr = p.communicate()
+  else:
+    _, stderr = p.communicate()
+    stderr_text = stderr.decode()
+    assert ("Ran 1 test in" in stderr_text and "OK" in stderr_text) or (not force_ok and "Failed to take lock file" in stderr_text), stderr_text
+
+if __name__ == "__main__":
+  max_workers = getenv("MAX_WORKERS", 4)
+  with ProcessPoolExecutor(max_workers=max_workers) as executor:
+    futures = []
+    for i in range(1000000):
+      if i % 100 == 0:
+        # wait for everything we launched so far
+        for f in as_completed(futures):
+          try:
+            f.result()
+          except Exception as e:
+            print(f"\nError in iteration: {e}")
+        futures = []
+
+        # do a full run in the main proc
+        run_test(i, True, force_ok=True)
+      else:
+        futures.append(executor.submit(run_test, i, bool(getenv("FULL_RUN", 0))))
+
+      # keep list small
+      if len(futures) > max_workers * 2:
+        futures = [f for f in futures if not f.done()]

test/external/external_test_hcq_open.py

@@ -0,0 +1,20 @@
+import os
+if "DEV" not in os.environ: os.environ["DEV"] = "AMD"
+
+import unittest, time
+from tinygrad import Device
+
+class TestOpen(unittest.TestCase):
+  def generate_test_open(n):
+    def test(self):
+      dev = Device[Device.DEFAULT]
+      for i in range(10):
+        dev.allocator.alloc(10 << 20)
+        time.sleep(0.5)
+    test.__name__ = f'test_open_{n}'
+    return test
+
+  for i in range(64): locals()[f'test_open_{i}'] = generate_test_open(i)
+
+if __name__ == '__main__':
+  unittest.main()

test/external/speed_v_theoretical.py

@@ -85,12 +85,11 @@ class TestKernelSpeed(unittest.TestCase):
     gbs = mems / tm / 1e9
     self._compare(tm, tflops, gbs, nv_tflops, nv_gbs, amd_tflops, amd_gbs)
 
-  # NOTE: tiny7 was slower than tiny12
   # TODO: why are convs so slow?!?
   def test_conv_3x3_256_32_32_256_256(self): self._test_conv_3x3(256, 32, 32, 256, 256, nv_tflops=27, amd_tflops=14)
 
   # theoretical is nv_tflops=165, amd_tflops=123
-  def test_gemm_4096(self): self._test_matmul(4096, nv_tflops=115, amd_tflops=65)
+  def test_gemm_4096(self): self._test_matmul(4096, nv_tflops=110, amd_tflops=65)
   def test_gemm_8192(self): self._test_matmul(8192, nv_tflops=115, amd_tflops=60)
 
   # theoretical is nv_gbs=1008, amd_gbs=960

test/test_custom_kernel.py

@@ -1,5 +1,6 @@
 import unittest
 from tinygrad import Tensor, UOp, Context
+from tinygrad.dtype import AddrSpace
 from tinygrad.uop.ops import KernelInfo, AxisType
 
 # **** kernels ****
@@ -9,15 +10,18 @@ def custom_arange_kernel(C:UOp) -> UOp:
   return C[i].store(i.cast(C.dtype.base)).end(i).sink(arg=KernelInfo(name=f"custom_arange_{C.size}"))
 
 def custom_add_one_kernel(B:UOp, A:UOp) -> UOp:
+  A,B = A.flatten(), B.flatten()
   assert B.size == A.size
   i = UOp.range(A.size, 0)
   return B[i].store(A[i] + 1).end(i).sink(arg=KernelInfo(name=f"add_one_{A.size}"))
 
 def custom_elementwise_add_kernel(C:UOp, A:UOp, B:UOp) -> UOp:
+  C,A,B = C.flatten(), A.flatten(), B.flatten()
   i = UOp.range(C.size, 0)
   return C[i].store(A[i]+B[i]).end(i).sink(arg=KernelInfo(name=f"custom_add_kernel_{C.size}")).simplify()
 
 def custom_elementwise_addmul_kernel(C:UOp, D:UOp, A:UOp, B:UOp) -> UOp:
+  C,D,A,B = C.flatten(), D.flatten(), A.flatten(), B.flatten()
   assert C.size == D.size
   i = UOp.range(C.size, 0)
   store_c = C[i].store(A[i]+B[i])
@@ -39,13 +43,43 @@ def custom_sum(B:UOp, A:UOp) -> UOp:
   return B.sink(arg=KernelInfo(name=f"custom_sum_{A.shape[0]}", opts_to_apply=()))
 
 def flip_contract_kernel(dest:UOp, src:UOp):
-  assert dest.size%4 == 0
-  i = UOp.range(dest.size//4, 0)
-  j = UOp.range(4, 1, AxisType.UPCAST)
-  vec = src[i*4+j].contract(j)
-  store = UOp.group(*[dest[i*4+k].store(vec.gep(3-k)) for k in range(4)])
+  i = UOp.range(dest.shape[0], 0)
+  j = UOp.range(dest.shape[1], 1, AxisType.UPCAST)
+  vec = src[i, j].contract(j)
+  store = UOp.group(*[dest[i, k].store(vec.gep(3-k)) for k in range(4)])
   return store.end(i).sink(arg=KernelInfo(name=f"flip_contract_{dest.size}", opts_to_apply=()))
 
+def slice_sum_kernel(dest:UOp, src:UOp):
+  G = UOp.range(src.shape[0], 0)
+  slice_src = src[G, :]
+  reg = UOp.placeholder((1,), dest.dtype.base, 0, addrspace=AddrSpace.REG)
+  reg = reg.after(G)[0].set(0)
+  R = UOp.range(src.shape[1], 1, AxisType.REDUCE)
+  reg = reg[0].set(reg.after(R)[0] + slice_src[R], end=R)
+  ast = dest[G].set(reg[0], end=G)
+  return ast.sink(arg=KernelInfo(name=f"slice_sum_{src.shape[0]}_{src.shape[1]}", opts_to_apply=()))
+
+def simple_qkv_kernel(O:UOp, Q:UOp, K:UOp, V:UOp) -> UOp:
+  # attention without softmax
+  N, d = Q.shape[0], Q.shape[1]
+
+  i = UOp.range(N, 0)  # output row
+  d_out = UOp.range(d, 1)  # output column
+  j = UOp.range(N, 2, axis_type=AxisType.REDUCE)
+
+  k_inner = UOp.range(d, 3, axis_type=AxisType.REDUCE)
+  qk_acc = UOp.placeholder((1,), Q.dtype.base, 0, addrspace=AddrSpace.REG)
+  qk_acc = qk_acc.after(i, j)[0].set(0.0)
+  qk_acc = qk_acc[0].set(qk_acc.after(k_inner)[0] + Q[i, k_inner] * K[j, k_inner], end=k_inner)
+  qk_score = qk_acc[0] / (d ** 0.5)
+
+  out_acc = UOp.placeholder((1,), Q.dtype.base, 1, addrspace=AddrSpace.REG)
+  out_acc = out_acc.after(i, d_out)[0].set(0.0)
+  out_acc = out_acc[0].set(out_acc.after(j)[0] + qk_score * V[j, d_out], end=j)
+
+  store = O[i, d_out].store(out_acc[0])
+  return store.end(d_out).end(i).sink(arg=KernelInfo(name=f"simple_qkv_{N}_{d}", opts_to_apply=()))
+
 # **** backward callbacks ****
 
 def backward_gemm(gradient:UOp, kernel:UOp) -> tuple[UOp, UOp]:
@@ -111,6 +145,12 @@ class TestCustomKernel(unittest.TestCase):
     b = Tensor.custom_kernel(tst, a, fxn=custom_sum)[0]
     self.assertEqual(b.item(), 15)
 
+  def test_slice_sum(self):
+    A = Tensor.randn(16, 16).contiguous()
+    B = Tensor.empty(16)
+    B = Tensor.custom_kernel(B, A, fxn=slice_sum_kernel)[0]
+    self.assertTrue(B.allclose(A.sum(1)))
+
   def test_gemm(self):
     N = 16
     a = Tensor.randn(N, N)
@@ -152,5 +192,19 @@ class TestCustomKernel(unittest.TestCase):
     err = (grad_b - real_grad_b).square().max()
     self.assertLess(err.item(), 1e-6)
 
+  def test_simple_qkv(self):
+    N, d = 8, 4
+    Q = Tensor.randn(N, d)
+    K = Tensor.randn(N, d)
+    V = Tensor.randn(N, d)
+    O = Tensor.empty(N, d)
+
+    O_custom = Tensor.custom_kernel(O, Q, K, V, fxn=lambda o,q,k,v: simple_qkv_kernel(o,q,k,v))[0]
+    O_ref = ((Q @ K.T) / (d ** 0.5)) @ V
+
+    Tensor.realize(O_custom, O_ref)
+    err = (O_custom - O_ref).square().max()
+    self.assertLess(err.item(), 1e-6)
+
 if __name__ == '__main__':
   unittest.main()

test/test_multitensor.py

@@ -596,6 +596,12 @@ class TestMultiTensor(unittest.TestCase):
     # ast are the same on devices
     self.assertEqual(len(set(asts)), 1)
 
+  def test_flip(self):
+    rng = Tensor.rand((10, 10, 10))
+    t0 = rng.shard(devices_2, axis=1)
+    out = t0.flip(0) + 1
+    self.assertTrue((rng.flip(0)+1).allclose(out.to(rng.device)))
+
   def test_reshape_on_axis(self):
     t0 = Tensor.rand((26, 15, 7)).shard(devices_3, axis=1)
 

test/test_ops.py

@@ -1551,8 +1551,10 @@ class TestOps(unittest.TestCase):
                                 lambda x: Tensor.stack(*x.std_mean(axis=(1,2))))
 
   def test_std_mean_loaded_nan(self):
-    helper_test_op([(1,0,3,0,5)], lambda x: torch.stack(torch.std_mean(x, axis=(1,3))),
-                                  lambda x: Tensor.stack(*x.std_mean(axis=(1,3))))
+    with warnings.catch_warnings():
+      warnings.filterwarnings("ignore", message="std_mean\\(\\): degrees of freedom is <= 0")
+      helper_test_op([(1,0,3,0,5)], lambda x: torch.stack(torch.std_mean(x, axis=(1,3))),
+                                    lambda x: Tensor.stack(*x.std_mean(axis=(1,3))))
   def test_softmax(self):
     helper_test_op([(45,65)], torch.nn.Softmax(dim=1), Tensor.softmax, atol=1e-7, grad_atol=1e-7)
     helper_test_op([(45)], torch.nn.Softmax(dim=0), Tensor.softmax, atol=1e-7, grad_atol=1e-7)
@@ -2820,13 +2822,13 @@ class TestOps(unittest.TestCase):
   @slow_test
   def test_slice_fancy_indexing_list_indices(self):
     a,b,c,d,e,i,j,k,o,p = self._get_index_randoms()
-    helper_test_op([(2,5,6,5,3,4)], lambda x: x[[[0]]], lambda x: x[[[0]]])
-    helper_test_op([(2,5,6,5,3,4)], lambda x: x[[0],b,c,d,:], lambda x: x[[0],j,k,o,:])
+    helper_test_op([(2,5,6,5,3,4)], lambda x: x[((0,),)])
+    helper_test_op([(2,5,6,5,3,4)], lambda x: x[(0,),b,c,d,:], lambda x: x[(0,),j,k,o,:])
     helper_test_op([(2,5,6,5,3,4)], lambda x: x[[[[0]]],b,c,d,[[1]]], lambda x: x[[[[0]]],j,k,o,[[1]]])
-    helper_test_op([(2,5,6,5,3,4)], lambda x: x[[1,0,-1],b,c,d,:], lambda x: x[[1,0,-1],j,k,o,:])
-    helper_test_op([(2,5,6,5,3,4)], lambda x: x[a,b,c,[1,2,3],...], lambda x: x[i,j,k,[1,2,3],...])
+    helper_test_op([(2,5,6,5,3,4)], lambda x: x[(1,0,-1),b,c,d,:], lambda x: x[(1,0,-1),j,k,o,:])
+    helper_test_op([(2,5,6,5,3,4)], lambda x: x[a,b,c,(1,2,3),...], lambda x: x[i,j,k,(1,2,3),...])
     helper_test_op([(2,5,6,5,3,4)], lambda x: x[a,b,c,[[1],[2],[3]],...], lambda x: x[i,j,k,[[1],[2],[3]],...])
-    helper_test_op([(2,5,6,5,3,4)], lambda x: x[a,[2,1,0],c,[-2,1,0],e], lambda x: x[i,[2,1,0],k,[-2,1,0],p])
+    helper_test_op([(2,5,6,5,3,4)], lambda x: x[a,(2,1,0),c,(-2,1,0),e], lambda x: x[i,(2,1,0),k,(-2,1,0),p])
 
   @slow_test
   def test_slice_fancy_indexing_tuple_indices(self):
@@ -2841,11 +2843,10 @@ class TestOps(unittest.TestCase):
   @slow_test
   def test_slice_fancy_indexing_list_with_tensors(self):
     a,b,c,d,e,i,j,k,o,p = self._get_index_randoms()
-    helper_test_op([(2,5,6,5,3,4)], lambda x: x[[a]], lambda x: x[[i]])
-    helper_test_op([(2,5,6,5,3,4)], lambda x: x[[a,1]], lambda x: x[[i,1]])
-    helper_test_op([(2,5,6,5,3,4)], lambda x: x[[a,[1,1]]], lambda x: x[[i,[1,1]]])
-    helper_test_op([(2,5,6,5,3,4)], lambda x: x[[a,(1,1)]], lambda x: x[[i,(1,1)]])
-    helper_test_op([(2,5,6,5,3,4)], lambda x: x[[a,b,c,d,e]], lambda x: x[[i,j,k,o,p]])
+    helper_test_op([(2,5,6,5,3,4)], lambda x: x[(a,)], lambda x: x[(i,)])
+    helper_test_op([(2,5,6,5,3,4)], lambda x: x[(a,1)], lambda x: x[(i,1)])
+    helper_test_op([(2,5,6,5,3,4)], lambda x: x[(a,(1,1))], lambda x: x[(i,(1,1))])
+    helper_test_op([(2,5,6,5,3,4)], lambda x: x[(a,b,c,d,e)], lambda x: x[(i,j,k,o,p)])
 
   def test_slice_fancy_indexing_errors(self):
     a = Tensor.ones(10,11,12)

test/test_schedule.py

@@ -1042,13 +1042,12 @@ class TestSchedule(unittest.TestCase):
       compare = torch.nn.functional.scaled_dot_product_attention(torch.tensor(q.numpy()),torch.tensor(k.numpy()),torch.tensor(v.numpy()))
       np.testing.assert_allclose(out.numpy(), compare.numpy(), atol=1e-6, rtol=1e-3)
 
-    with Context(FUSE_ATTENTION=1):
-      out = Tensor.scaled_dot_product_attention(q,k,v)
-      run_schedule(check_schedule(out, 4)) # TODO: should be 1?
-      if getenv("CHECK", 1):
-        import torch
-        compare = torch.nn.functional.scaled_dot_product_attention(torch.tensor(q.numpy()),torch.tensor(k.numpy()),torch.tensor(v.numpy()))
-        np.testing.assert_allclose(out.numpy(), compare.numpy(), atol=1e-6, rtol=1e-3)
+    out = Tensor.scaled_dot_product_attention(q,k,v)
+    run_schedule(check_schedule(out, 4)) # TODO: should be 1?
+    if getenv("CHECK", 1):
+      import torch
+      compare = torch.nn.functional.scaled_dot_product_attention(torch.tensor(q.numpy()),torch.tensor(k.numpy()),torch.tensor(v.numpy()))
+      np.testing.assert_allclose(out.numpy(), compare.numpy(), atol=1e-6, rtol=1e-3)
 
   def test_ugly_reduceop_pairing(self):
     Tensor.manual_seed(0)

test/test_softmax_fusion.py

@@ -32,7 +32,7 @@ def run_one_schedule_item(out): lower_schedule_item(get_single_element(out.sched
 class TestFuse(unittest.TestCase):
   def _test_fuse(self, fxn, *args, atol=1e-6, allow_multiple=False, **kwargs):
     GlobalCounters.reset()
-    out_single = fxn(*args, **kwargs).fuse()
+    out_single = fxn(*args, **kwargs)
     if not allow_multiple: run_one_schedule_item(out_single)
     np_single = out_single.numpy()
     GlobalCounters.reset()
@@ -100,7 +100,7 @@ class TestFuse(unittest.TestCase):
     q = (x @ wq).contiguous()
     k = (x @ wk).contiguous()
     v = (x @ wv).contiguous()
-    attn = q.scaled_dot_product_attention(k, v).fuse()
+    attn = q.scaled_dot_product_attention(k, v)
     s = attn.schedule()
     self.assertEqual(len(s), 4) # 3 matmul and 1 attention
 
@@ -121,7 +121,7 @@ class TestFuse(unittest.TestCase):
   def test_mismatch_reduce(self):
     a = Tensor.ones(16, 10).contiguous().realize()
     b = Tensor.ones(16, 20).contiguous().realize()
-    c = (a.sum(axis=1) + b.sum(axis=1)).fuse()
+    c = (a.sum(axis=1) + b.sum(axis=1))
     self.assertListEqual(c.tolist(), [30]*16)
 
   @unittest.skipUnless(Device.DEFAULT == "METAL", "METAL TC")
@@ -129,7 +129,7 @@ class TestFuse(unittest.TestCase):
     A = Tensor.randn(8, 8).realize()
     B = Tensor.randn(8, 8).realize()
     C = Tensor.ones(1, 8, 8).pad(((1,1), None, None),).sum(0)
-    out = (C + (A @ B)).fuse()
+    out = (C + (A @ B))
     out.realize()
 
 class TestSoftmaxFusion(unittest.TestCase):
@@ -180,7 +180,7 @@ class TestSoftmaxFusion(unittest.TestCase):
 
     print("*** auto single kernel softmax ***")
     with Context(NOOPT=1, DEBUG=max(DEBUG.value, 2)):
-      out = self.test.contiguous().softmax(-1).fuse()
+      out = self.test.contiguous().softmax(-1)
       run_one_schedule_item(out)
 
     np.testing.assert_allclose(sout.numpy(), out.numpy(), atol=3e-7)

test/test_uops.py

@@ -517,7 +517,7 @@ class TestUOpStr(unittest.TestCase):
 
 class TestUPatHelpers(unittest.TestCase):
   def test_location(self):
-    self.assertEqual(sym.patterns[-1][0].location[0].replace("\\", "/").split("/")[-1], "symbolic.py")
+    self.assertEqual(sym.patterns[-1][0].location[0].replace("\\", "/").split("/")[-1], "math.py")
     self.assertEqual(shared_spec.patterns[0][0].location[0].replace("\\", "/").split("/")[-1], "spec.py")
     test_upat = UPat(Ops.CONST, dtypes.bool)
     self.assertEqual(test_upat.location[0].split("/")[-1], __file__.replace("\\", "/").split("/")[-1])

test/unit/test_indexing.py

@@ -894,7 +894,7 @@ class TestNumpy(unittest.TestCase):
     a = Tensor([[1, 2, 3],
                 [4, 5, 6],
                 [7, 8, 9]])
-    self.assertIsNot(a[...], a)
+    self.assertIs(a[...], a)
     numpy_testing_assert_equal_helper(a[...], a)
     # `a[...]` was `a` in numpy <1.9.
     #numpy_testing_assert_equal_helper(data_ptr(a[...]), data_ptr(a))
@@ -1037,9 +1037,9 @@ class TestNumpy(unittest.TestCase):
     # Before `...` would return a itself.
     a = Tensor([5])
 
-    self.assertIsNot(a, a[()])
-    self.assertIsNot(a, a[...])
-    self.assertIsNot(a, a[:])
+    self.assertIs(a, a[()])
+    self.assertIs(a, a[...])
+    self.assertIs(a, a[:])
 
   def test_broaderrors_indexing(self):
     a = Tensor.zeros(5, 5)

tinygrad/codegen/__init__.py

@@ -4,7 +4,7 @@ from tinygrad.helpers import DEVECTORIZE, TRANSCENDENTAL, SPEC
 from tinygrad.uop.ops import PatternMatcher, graph_rewrite, UOp, pm_lower_index_dtype, Ops, UPat
 from tinygrad.uop.spec import type_verify, program_spec, kernel_spec
 from tinygrad.renderer import Renderer
-from tinygrad.dtype import dtypes
+from tinygrad.dtype import dtypes, PtrDType
 from tinygrad.helpers import panic
 
 # import all pattern matchers here
@@ -19,13 +19,19 @@ from tinygrad.codegen.simplify import pm_simplify_ranges, pm_flatten_range, pm_s
 from tinygrad.schedule.rangeify import pm_add_buffers_local, rangeify_codegen, pm_mops
 from tinygrad.codegen.late.linearizer import CFGContext, pm_split_ends, pm_add_control_flow, linearize
 
+pm_syntactic_sugar = PatternMatcher([
+  # INDEX on ptr INDEX concats them
+  (UPat(Ops.INDEX, name="i1").f(Ops.INDEX, name="i2", allow_any_len=True),
+   lambda i1,i2: i2.replace(src=i1.src+i2.src[1:]) if isinstance(i1.dtype, PtrDType) and not isinstance(i2.dtype, PtrDType) else None),
+])
+
 def full_rewrite_to_sink(sink:UOp, ren:Renderer|None=None, optimize:bool=True) -> UOp:
   if ren is None: ren = Renderer()
 
   if SPEC: type_verify(sink, kernel_spec)
 
   # preprocess
-  sink = graph_rewrite(sink, pm_mops, name="early movement ops")
+  sink = graph_rewrite(sink, pm_mops+pm_syntactic_sugar, name="early movement ops", bottom_up=True)
 
   # first we optimize
   if optimize:

tinygrad/codegen/late/devectorizer.py

@@ -4,7 +4,7 @@ from collections import defaultdict
 from dataclasses import dataclass
 from tinygrad.dtype import dtypes, ImageDType, DType, AddrSpace, Invalid, PtrDType
 from tinygrad.uop.ops import UOp, Ops, UPat, PatternMatcher, graph_rewrite, GroupOp, identity_element
-from tinygrad.uop.symbolic import uop_given_valid, parse_valid, sym, symbolic_flat, invalid_gate
+from tinygrad.uop.symbolic import uop_given_valid, parse_valid, sym, symbolic, invalid_gate
 from tinygrad.helpers import getenv, flatten, AMX, prod
 from tinygrad.renderer import Renderer
 
@@ -61,7 +61,7 @@ def expand_index(buf:UOp, vec:UOp):
   if getenv("UNSAFE_DISABLE_MASK", 0): vec = vec.get_idx()
   # generate the individual indexes
   midx = graph_rewrite(UOp.sink(*[buf.index(vec.gep(i), ptr=True) for i in range(vec.dtype.count)]),
-                       symbolic_flat+load_store_indexing, name=f"index_buf_{buf.arg}")
+                       symbolic+load_store_indexing, name=f"index_buf_{buf.arg}")
   # extract all the relevant offsets
   offsets_rootsrc: defaultdict[Any, dict[int, list[int]]] = defaultdict(dict)
   for i in range(vec.dtype.count):
@@ -299,7 +299,7 @@ def reduce_to_acc(ctx:ReduceContext, red:UOp):
     ended_ranges = flatten([x.ended_ranges for x in topo if x.op is Ops.END])
     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 = 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)
     lst = [acc.after(acc_init, *reduce_range).index(UOp.const(dtypes.int, 0))] + lst  # put acc as the first element

tinygrad/codegen/late/linearizer.py

@@ -1,29 +1,27 @@
 import heapq
 from collections import defaultdict
 from tinygrad.uop.ops import PatternMatcher, UOp, Ops, UPat
+from tinygrad.helpers import prod
 
 def linearize(u:UOp) -> list[UOp]:
   # this is a toposort with priority
   lst = list(u.toposort())
   consumers: defaultdict[UOp, list[UOp]] = defaultdict(list)
   in_degree:dict[UOp, int] = {}
-  priorities:dict[UOp, int] = {}
+  priorities:dict[UOp, tuple[int, int]] = {}
 
   # get consumers and assign priorities
   # NOTE: this requires the lst be locally toposorted
   for u in reversed(lst):
     for s in u.src: consumers[s].append(u)
     in_degree[u] = len(u.src)
-    # put loads in the beginning of the block and prevent priority inversion. hack for BARRIER grouping too
-    priority = [0] + [priorities[x] for x in consumers[u]]
-    if u.op is Ops.LOAD: priority.append(-1000)
-    if u.op is Ops.BARRIER: priority.append(-1500)
-    # ranges are scheduled as late as possible so anything that can be outside is
-    # if u.op is Ops.RANGE: priority = [2000]
-    if u.op is Ops.END: priority = [-1000]
-    # move defines and consts to the top
-    if u.op in {Ops.DEFINE_GLOBAL, Ops.DEFINE_LOCAL, Ops.DEFINE_REG, Ops.DEFINE_VAR, Ops.SPECIAL, Ops.CONST}: priority.append(-2000)
-    priorities[u] = min(priority)
+
+    # we place UOps with higher run_counts later
+    # this will cause ranges to be placed late and ends to be placed early
+    run_count = prod([int(r.vmax)+1 for r in u.ranges])
+
+    # simple priority
+    priorities[u] = (run_count, 0)
 
   # number the uops in "ideal" order
   nkey = {u:i for i,u in enumerate(sorted(lst, key=lambda x: (priorities[x],)+x.tuplize))}

tinygrad/codegen/simplify.py

@@ -1,6 +1,6 @@
 import itertools
 from tinygrad.uop.ops import UOp, PatternMatcher, UPat, Ops, graph_rewrite, _substitute, range_start, ImageDType
-from tinygrad.uop.symbolic import symbolic_flat
+from tinygrad.uop.symbolic import symbolic
 from tinygrad.helpers import partition, dedup
 from tinygrad.dtype import dtypes
 
@@ -28,7 +28,7 @@ def simplify_merge_adjacent(u:UOp) -> UOp|None:
         s0, s1 = r0.src[0], r1.src[0]
         # do the merge
         new_range = r0.replace(src=(s0*s1,))
-        nidx = graph_rewrite(u, _substitute+symbolic_flat+pm_flatten_range, ctx={r0:new_range//s1, r1:new_range%s1},
+        nidx = graph_rewrite(u, _substitute+symbolic+pm_flatten_range, ctx={r0:new_range//s1, r1:new_range%s1},
                              name=f"check_merge_{r0.arg[0]}_{r1.arg[0]}")
 
         # check if it simplifies
@@ -109,7 +109,7 @@ pm_reduce_collapse = pm_reduce_unparented + PatternMatcher([
     lambda x,y,c,r: y.where(c, 0).reduce(*r.src[1:], arg=Ops.ADD)*x.cast(c.dtype)),
   # MUL casted bool
   ((UPat.var("x") * UPat.var("gate", dtype=dtypes.bool).cast()), lambda x,gate: gate.where(x, 0)),
-])+symbolic_flat
+])+symbolic
 
 pm_reduce_load_collapse = pm_reduce_collapse + PatternMatcher([
   # lift x+y out of reduce on ne

tinygrad/device.py

@@ -54,7 +54,7 @@ atexit.register(lambda: [Device[dn].finalize() for dn in Device._opened_devices]
 
 @dataclass(frozen=True)
 class ProfileDeviceEvent(ProfileEvent):
-  device:str; comp_tdiff:decimal.Decimal=decimal.Decimal(0); copy_tdiff:decimal.Decimal=decimal.Decimal(0) # noqa: E702
+  device:str; comp_tdiff:decimal.Decimal=decimal.Decimal(0); copy_tdiff:decimal.Decimal=decimal.Decimal(0); props:dict[str,Any]|None=None # noqa: E702
 
 @dataclass(frozen=True)
 class ProfileProgramEvent(ProfileEvent): device:str; name:str; lib:bytes|None; base:int|None # noqa: E702

tinygrad/gradient.py

@@ -29,14 +29,14 @@ pm_gradient = PatternMatcher([
   (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.CONTIGUOUS, Ops.FUSE)), lambda ctx: (ctx,)),
+  (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)),
   (UPat(Ops.EXPAND, name="ret"), lambda ctx, ret: (ctx.r(Ops.ADD,tuple(i for i,(s,n) in enumerate(zip(ret.src[0].shape, ret.shape)) if s!=n)), None)),
   (UPat(Ops.PAD, name="ret"), lambda ctx, ret: (ctx.shrink(tuple([(p[0], s+p[0]) for s,p in zip(ret.src[0].shape, ret.marg)])), None, None)),
   (UPat(Ops.SHRINK, name="ret"), lambda ctx, ret: (ctx.pad(tuple([(p[0], s-p[1]) for s,p in zip(ret.src[0].shape, ret.marg)])), None, None)),
   (UPat(Ops.PERMUTE, name="ret"), lambda ctx, ret: (ctx.permute(argsort(ret.marg)),)),
-  (UPat(Ops.FLIP, name="ret"), lambda ctx, ret: (ctx.flip(ret.marg),)),
+  (UPat(Ops.FLIP, name="ret"), lambda ctx, ret: (ctx.flip([i for i,x in enumerate(ret.marg) if x]),)),
   (UPat(Ops.MULTI, name="ret"), lambda ctx, ret: ctx.shard(ret.device, ret.axis).src),
   # NOTE: this is only correct when the KERNEL has a single output
   (UPat(Ops.AFTER), lambda ctx: (ctx, ctx)),

tinygrad/helpers.py

@@ -44,8 +44,7 @@ def fully_flatten(l):
     return flattened
   return [l]
 def fromimport(mod, frm): return getattr(__import__(mod, fromlist=[frm]), frm)
-def _is_balanced(s:str) -> bool:
-  return (acc:=list(itertools.accumulate([(1 if ch=='(' else -1 if ch==')' else 0) for ch in s])))[-1]==0 and all(x>=0 for x in acc)
+def _is_balanced(s:str) -> bool: return (d := 0, all((d := d + (c == '(') - (c == ')')) >= 0 for c in s))[1] and d == 0
 def strip_parens(fst:str) -> str: return fst[1:-1] if fst and fst[0]=='(' and fst[-1] == ')' and _is_balanced(fst[1:-1]) else fst
 def ceildiv(num, amt): return int(ret) if isinstance((ret:=-(num//-amt)), float) else ret
 def round_up(num:int, amt:int) -> int: return (num+amt-1)//amt * amt
@@ -171,7 +170,6 @@ DISABLE_COMPILER_CACHE = ContextVar("DISABLE_COMPILER_CACHE", 0)
 VALIDATE_WITH_CPU, DISABLE_FAST_IDIV = ContextVar("VALIDATE_WITH_CPU", 0), ContextVar("DISABLE_FAST_IDIV", 0)
 CORRECT_DIVMOD_FOLDING, FUSE_OPTIM = ContextVar("CORRECT_DIVMOD_FOLDING", 0), ContextVar("FUSE_OPTIM", 0)
 ALLOW_DEVICE_USAGE, MAX_BUFFER_SIZE = ContextVar("ALLOW_DEVICE_USAGE", 1), ContextVar("MAX_BUFFER_SIZE", 0)
-FUSE_ATTENTION = ContextVar("FUSE_ATTENTION", 0)
 EMULATE = ContextVar("EMULATE", "")
 CPU_COUNT = ContextVar("CPU_COUNT", max(1, len(os.sched_getaffinity(0)) if hasattr(os, "sched_getaffinity") else (os.cpu_count() or 1)))
 CPU_LLVM, CPU_LVP, AMD_LLVM = ContextVar("CPU_LLVM", 0), ContextVar("CPU_LVP", 0), ContextVar("AMD_LLVM", 1)

tinygrad/mixin/__init__.py

@@ -0,0 +1,4 @@
+from tinygrad.mixin.math import MathMixin
+from tinygrad.mixin.movement import MovementMixin
+
+class OpMixin(MathMixin, MovementMixin): pass

tinygrad/mixin/math.py

@@ -1,11 +1,6 @@
-# mixins add syntactic sugar to Tensor and UOp
-from typing import TypeAlias, TYPE_CHECKING, Self
+from typing import Self
 from tinygrad.uop import Ops
 from tinygrad.dtype import dtypes, ConstType
-from tinygrad.helpers import prod, argfix
-if TYPE_CHECKING:
-  from tinygrad.uop.ops import UOp
-  sint:TypeAlias = UOp|int
 
 class MathMixin:
   # required to implement
@@ -175,32 +170,3 @@ class MathMixin:
   def exp2(self): return self.alu(Ops.EXP2)
   def pow(self, x:Self|ConstType): return self.alu(Ops.POW, self.ufix(x))
   def __pow__(self, x:Self|ConstType): return self.pow(x)
-
-class MovementMixin:
-  # required to implement
-  def _mop(self, op:Ops, arg) -> Self: raise NotImplementedError
-  @property
-  def shape(self) -> tuple["sint", ...]: raise NotImplementedError
-
-  # great functions you get!
-  def view(self, shape, *args) -> Self:
-    """`.view` is an alias for `.reshape`."""
-    return self.reshape(shape, *args)
-
-  def reshape(self, shape, *args) -> Self:
-    """
-    Returns a tensor with the same data as the original tensor but with a different shape.
-    `shape` can be passed as a tuple or as separate arguments.
-
-    ```python exec="true" source="above" session="tensor" result="python"
-    t = Tensor.arange(6)
-    print(t.reshape(2, 3).numpy())
-    ```
-    """
-    # resolve None and args
-    new_shape = tuple([s if s is not None else self.shape[i] for i,s in enumerate(argfix(shape, *args))])
-    # resolve -1
-    if (c := new_shape.count(-1)) > 1: raise RuntimeError(f"only one dimension can be inferred using -1, getting {new_shape}")
-    if c: new_shape = tuple([-prod(self.shape) // prod(new_shape) if s == -1 else s for s in new_shape])
-    if prod(self.shape) != prod(new_shape): raise ValueError(f"size mismatch, can't reshape ({self.shape}) -> ({new_shape})")
-    return self._mop(Ops.RESHAPE, arg=new_shape) if new_shape != self.shape else self

tinygrad/mixin/movement.py

@@ -0,0 +1,259 @@
+# mixins add syntactic sugar to Tensor and UOp
+import functools
+from typing import TypeAlias, TYPE_CHECKING, Self
+from tinygrad.uop import Ops
+from tinygrad.helpers import prod, argfix, flatten, dedup
+if TYPE_CHECKING:
+  from tinygrad.uop.ops import UOp
+  sint:TypeAlias = UOp|int
+
+class MovementMixin:
+  # required to implement
+  def _mop(self, op:Ops, arg) -> Self: raise NotImplementedError
+  @property
+  def shape(self) -> tuple["sint", ...]: raise NotImplementedError
+
+  # great functions you get!
+  @property
+  def ndim(self) -> int:
+    """
+    Returns the number of dimensions in the tensor.
+
+    ```python exec="true" source="above" session="tensor" result="python"
+    t = Tensor([[1, 2], [3, 4]])
+    print(t.ndim)
+    ```
+    """
+    return len(self.shape)
+
+  def numel(self) -> "sint":
+    """
+    Returns the total number of elements in the tensor.
+
+    ```python exec="true" source="above" session="tensor" result="python"
+    t = Tensor([[[1, 2], [3, 4]], [[5, 6], [7, 8]]])
+    print(t.numel())
+    ```
+    """
+    return prod(self.shape)
+
+  def _resolve_dim(self, dim:int, *, extra:bool=False) -> int:
+    total = self.ndim + int(extra)
+    if not -max(1, total) <= dim <= max(1, total)-1: raise IndexError(f"{dim=} out of range {[-max(1, total), max(1, total)-1]}")
+    return dim + total if dim < 0 else dim
+
+  def reshape(self, shape, *args) -> Self:
+    """
+    Returns a tensor with the same data as the original tensor but with a different shape.
+    `shape` can be passed as a tuple or as separate arguments.
+
+    ```python exec="true" source="above" session="tensor" result="python"
+    t = Tensor.arange(6)
+    print(t.reshape(2, 3).numpy())
+    ```
+    """
+    # resolve None and args
+    new_shape = tuple([s if s is not None else self.shape[i] for i,s in enumerate(argfix(shape, *args))])
+    # resolve -1
+    if (c := new_shape.count(-1)) > 1: raise RuntimeError(f"only one dimension can be inferred using -1, getting {new_shape}")
+    if c: new_shape = tuple([-prod(self.shape) // prod(new_shape) if s == -1 else s for s in new_shape])
+    if prod(self.shape) != prod(new_shape): raise ValueError(f"size mismatch, can't reshape ({self.shape}) -> ({new_shape})")
+    ret = self._mop(Ops.RESHAPE, arg=new_shape)
+    return self if ret.shape == self.shape else ret
+
+  def shrink(self, arg:tuple[tuple["sint", "sint"]|None, ...]) -> Self:
+    """
+    Returns a tensor that shrinks the each axis based on input arg.
+    `arg` must have the same length as `self.ndim`.
+    For each axis, it can be `None`, which means no shrink, or a tuple `(start, end)` that works the same as Python slice.
+
+    ```python exec="true" source="above" session="tensor" result="python"
+    t = Tensor.arange(9).reshape(3, 3)
+    print(t.numpy())
+    ```
+    ```python exec="true" source="above" session="tensor" result="python"
+    print(t.shrink(((None, (1, 3)))).numpy())
+    ```
+    ```python exec="true" source="above" session="tensor" result="python"
+    print(t.shrink((((0, 2), (0, 2)))).numpy())
+    ```
+    """
+    if self.ndim != len(arg): raise ValueError(f"{self.ndim=} != {len(arg)=}")
+    ret = self._mop(Ops.SHRINK, arg=[x if x is not None else (0,s) for x,s in zip(arg, self.shape)])
+    return self if ret.shape == self.shape else ret
+
+  def permute(self, order, *args) -> Self:
+    """
+    Returns a tensor that is a permutation of the original tensor.
+    The new tensor has the same data as the original tensor but with the dimensions permuted according to the order specified.
+    `order` can be passed as a tuple or as separate arguments.
+
+    ```python exec="true" source="above" session="tensor" result="python"
+    t = Tensor.empty(2, 3, 5)
+    print(t.shape)
+    ```
+    ```python exec="true" source="above" session="tensor" result="python"
+    print(t.permute(2, 0, 1).shape)
+    ```
+    """
+    order_arg = tuple(self._resolve_dim(x) for x in argfix(order, *args))
+    if sorted(order_arg) != list(range(self.ndim)): raise RuntimeError(f"order is not a valid permutation, getting {order_arg}")
+    return self._mop(Ops.PERMUTE, arg=order_arg) if order_arg != tuple(range(self.ndim)) else self
+
+  def flip(self, axis, *args) -> Self:
+    """
+    Returns a tensor that reverses the order of the original tensor along given `axis`.
+    `axis` can be passed as a tuple or as separate arguments.
+
+    ```python exec="true" source="above" session="tensor" result="python"
+    t = Tensor.arange(6).reshape(2, 3)
+    print(t.numpy())
+    ```
+    ```python exec="true" source="above" session="tensor" result="python"
+    print(t.flip(0).numpy())
+    ```
+    ```python exec="true" source="above" session="tensor" result="python"
+    print(t.flip((0, 1)).numpy())
+    ```
+    """
+    axis_arg = tuple(self._resolve_dim(x) for x in argfix(axis, *args))
+    assert all(not isinstance(x, bool) and x >= 0 and x < self.ndim for x in axis_arg), f"flip args must be axis ints {axis_arg}"
+    if len(axis_arg) != len(dedup(axis_arg)): raise RuntimeError(f"dim can appear at most once, getting {axis_arg}")
+    flip_arg = tuple([i in axis_arg for i in range(len(self.shape))])
+    return self._mop(Ops.FLIP, arg=flip_arg) if any(flip_arg) else self
+
+  # **** high level ****
+
+  def view(self, shape, *args) -> Self:
+    """`.view` is an alias for `.reshape`."""
+    return self.reshape(shape, *args)
+
+  def squeeze(self, dim:int|None=None) -> Self:
+    """
+    Returns a tensor with specified dimensions of input of size 1 removed.
+    If `dim` is not specified, all dimensions with size 1 are removed.
+
+    ```python exec="true" source="above" session="tensor" result="python"
+    t = Tensor.zeros(2, 1, 2, 1, 2)
+    print(t.squeeze().shape)
+    ```
+    ```python exec="true" source="above" session="tensor" result="python"
+    print(t.squeeze(0).shape)
+    ```
+    ```python exec="true" source="above" session="tensor" result="python"
+    print(t.squeeze(1).shape)
+    ```
+    """
+    if dim is None: return self.reshape(tuple(dim for dim in self.shape if dim != 1))
+    dim = self._resolve_dim(dim)
+    return self if not self.ndim or self.shape[dim] != 1 else self.reshape(self.shape[:dim] + self.shape[dim+1:])
+
+  def unsqueeze(self, dim:int) -> Self:
+    """
+    Returns a tensor with a new dimension of size 1 inserted at the specified `dim`.
+
+    ```python exec="true" source="above" session="tensor" result="python"
+    t = Tensor([1, 2, 3, 4])
+    print(t.unsqueeze(0).numpy())
+    ```
+    ```python exec="true" source="above" session="tensor" result="python"
+    print(t.unsqueeze(1).numpy())
+    ```
+    """
+    dim = self._resolve_dim(dim, extra=True)
+    return self.reshape(self.shape[:dim] + (1,) + self.shape[dim:])
+
+  @property
+  def T(self) -> Self:
+    """`.T` is an alias for `.transpose()`."""
+    return self.transpose()
+
+  def transpose(self, dim0=1, dim1=0) -> Self:
+    """
+    Returns a tensor that is a transposed version of the original tensor.
+    The given dimensions `dim0` and `dim1` are swapped.
+
+    ```python exec="true" source="above" session="tensor" result="python"
+    t = Tensor.arange(6).reshape(2, 3)
+    print(t.numpy())
+    ```
+    ```python exec="true" source="above" session="tensor" result="python"
+    print(t.transpose(0, 1).numpy())
+    ```
+    """
+    order = list(range(self.ndim))
+    order[dim0], order[dim1] = order[dim1], order[dim0]
+    return self.permute(order)
+
+  def flatten(self, start_dim=0, end_dim=-1) -> Self:
+    """
+    Flattens the tensor by reshaping it into a one-dimensional tensor.
+    If `start_dim` or `end_dim` are passed, only dimensions starting with `start_dim` and ending with `end_dim` are flattened.
+
+    ```python exec="true" source="above" session="tensor" result="python"
+    t = Tensor.arange(8).reshape(2, 2, 2)
+    print(t.flatten().numpy())
+    ```
+    ```python exec="true" source="above" session="tensor" result="python"
+    print(t.flatten(start_dim=1).numpy())
+    ```
+    """
+    start_dim, end_dim = self._resolve_dim(start_dim), self._resolve_dim(end_dim)
+    return self.reshape(self.shape[:start_dim] + (prod(self.shape[start_dim:end_dim+1]), ) + self.shape[end_dim+1:])
+
+  def unflatten(self, dim:int, sizes:tuple[int,...]) -> Self:
+    """
+    Unflattens dimension `dim` of the tensor into multiple dimensions specified by `sizes`. `Tensor.flatten()` is the inverse of this function.
+
+    ```python exec="true" source="above" session="tensor" result="python"
+    print(Tensor.ones(3, 4, 1).unflatten(1, (2, 2)).shape)
+    ```
+    ```python exec="true" source="above" session="tensor" result="python"
+    print(Tensor.ones(3, 4, 1).unflatten(1, (-1, 2)).shape)
+    ```
+    ```python exec="true" source="above" session="tensor" result="python"
+    print(Tensor.ones(5, 12, 3).unflatten(-2, (2, 2, 3, 1, 1)).shape)
+    ```
+    """
+    dim = self._resolve_dim(dim)
+    return self.reshape(self.shape[:dim] + sizes + self.shape[dim+1:])
+
+  def rearrange(self, formula:str, **sizes) -> Self:
+    """
+    Rearranges input according to formula
+
+    See: https://einops.rocks/api/rearrange/
+
+    ```python exec="true" source="above" session="tensor" result="python"
+    x = Tensor([[1, 2], [3, 4]])
+    print(Tensor.rearrange(x, "batch channel -> (batch channel)").numpy())
+    ```
+    """
+    def parse_formula(formula: str):
+      tokens = f" {formula} ".replace("…", "...").replace("(", " ( ").replace(")", " ) ").replace(" ", "  ").replace(" 1 ", " ( ) ").split()
+      lparens, rparens = map(lambda x: [i for i, ch in enumerate(tokens) if ch == x], ("(", ")"))
+      pairs = list(zip(lparens, rparens))
+      assert len(lparens) == len(rparens) and sorted(flatten(pairs)) == flatten(pairs), "bracket mismatch"
+      return [name for name in tokens if name not in ("(", ")")], [(s - 2*i, e - 1 - 2*i) for i, (s, e) in enumerate(pairs)]
+
+    assert formula.count("->") == 1, 'need exactly one "->" in formula'
+
+    (lhs, unflatten_dims), (rhs, flatten_dims) = map(parse_formula, formula.split("->"))
+
+    for name in sizes: assert name in lhs, f"axis {name} is not used in transform"
+    assert sorted(lhs) == sorted(rhs) and len(lhs) == len(set(lhs)), f"name mismatch in {formula}"
+    for name in flatten((lhs, rhs)): assert name == "..." or (name.isidentifier() and "_" not in (name[0], name[-1])), f"invalid axis name {name}"
+    assert "..." not in flatten([lhs[s:e] for s, e in unflatten_dims]), f"cannot have collapsed ellipsis (...) in lhs of {formula}"
+    assert lhs.count("...") <= 1, f"too many ellipses in {formula}"
+
+    # resolve ellipsis
+    if "..." in lhs: ell_len = len(self.shape) - len(lhs) + 1 + sum(e - s - 1 for s, e in unflatten_dims)
+    lhs, rhs = map(lambda l: l[:(i:=l.index("..."))] + [f"...{j}" for j in range(ell_len)] + l[i + 1:] if "..." in l else l, (lhs, rhs))
+    unflatten_dims = [(s + (ell_len - 1 if "...0" in lhs[:s] else 0), e + (ell_len - 1 if "...0" in lhs[:e] else 0)) for s, e in unflatten_dims]
+    flatten_dims = [(s + (ell_len - 1 if "...0" in rhs[:s] else 0), e + (ell_len - 1 if "...0" in rhs[:e] else 0)) for s, e in flatten_dims]
+
+    # apply movement ops in order unflatten -> permute -> flatten/unsqueeze
+    t = functools.reduce(lambda x, dims: x.unflatten(dims[0], tuple(sizes.get(lhs[d], -1) for d in range(*dims))), unflatten_dims, self)
+    for i, name in enumerate(lhs): assert (name not in sizes) or sizes[name] == t.shape[i], f"size provided for dimension {name} incorrect"
+    t = t.permute([lhs.index(name) for name in rhs])
+    return functools.reduce(lambda x, dims: x.flatten(dims[0], dims[1] - 1) if dims[0]<dims[1] else x.unsqueeze(dims[0]), reversed(flatten_dims), t)

tinygrad/nn/__init__.py

@@ -323,7 +323,7 @@ class Embedding:
     if not dtypes.is_int(idx.dtype): raise TypeError(f"Expected integer dtype for index in embedding, got {idx.dtype}")
     big_shp = idx.shape+(self.vocab_sz, self.embed_sz)
     arange, idx, vals = self.arange.expand(big_shp), idx.reshape(idx.shape+(1, 1)).expand(big_shp), self.weight.expand(big_shp)
-    return (arange == idx).mul(vals).sum(-2, dtype=vals.dtype)
+    return (arange == idx).where(vals, 0).sum(-2, dtype=vals.dtype)
 
 class LSTMCell:
   """

tinygrad/renderer/nir.py

@@ -148,7 +148,7 @@ class NIRRenderer(Renderer):
     (UPat(Ops.CAST, name="x"), lambda ctx,x: ncast(ctx.b, ctx.r[x.src[0]], x.src[0].dtype, x.dtype)),
     (UPat(Ops.BITCAST, src=(UPat.var("a"),), allow_any_len=True), lambda ctx,a: ctx.r[a]),
     (UPat(Ops.GEP, src=(UPat.var("a"),), name="x"), lambda ctx,x,a: nchannel(ctx.b, ctx.r[a], x.arg[0])),
-    (UPat(Ops.DEFINE_REG, name="x"), lambda ctx,x:mesa.nir_local_variable_create(ctx.b.impl, glsl_type(x.dtype), f"acc{x.arg[0]}".encode()).contents),
+    (UPat(Ops.DEFINE_REG, name="x"), lambda ctx,x:mesa.nir_local_variable_create(ctx.b.impl, glsl_type(x.dtype), f"acc{x.arg}".encode()).contents),
     (UPat(Ops.BARRIER), lambda ctx: nbarrier(ctx.b)),
     (UPat(Ops.IF, name="x"), lambda ctx,x: mesa.nir_push_if(ctx.b, ctx.r[x.src[0]])),
     (UPat(Ops.ENDIF, name="x"), lambda ctx,x: (lambda _: mesa.nir_def())(mesa.nir_pop_if(ctx.b, ctx.r[x.src[0]])))

tinygrad/runtime/autogen/comgr.py

@@ -15,7 +15,9 @@ PATHS_TO_TRY = [
 ]
 def _try_dlopen_amd_comgr():
   library = ctypes.util.find_library("amd_comgr")
-  if library: return ctypes.CDLL(library)
+  if library:
+    try: return ctypes.CDLL(library)
+    except OSError: pass
   for candidate in PATHS_TO_TRY:
     try: return ctypes.CDLL(candidate)
     except OSError: pass

tinygrad/runtime/autogen/mesa.py

@@ -15,7 +15,9 @@ PATHS_TO_TRY = [
 ]
 def _try_dlopen_tinymesa_cpu():
   library = ctypes.util.find_library("tinymesa_cpu")
-  if library: return ctypes.CDLL(library)
+  if library:
+    try: return ctypes.CDLL(library)
+    except OSError: pass
   for candidate in PATHS_TO_TRY:
     try: return ctypes.CDLL(candidate)
     except OSError: pass

tinygrad/runtime/autogen/rocprof.py

@@ -8,11 +8,22 @@
 # LONGDOUBLE_SIZE is: 16
 #
 import ctypes, ctypes.util
+PATHS_TO_TRY = [
+  '/usr/local/lib/librocprof-trace-decoder.so',
+  '/usr/local/lib/librocprof-trace-decoder.dylib',
+]
+def _try_dlopen_rocprof_trace_decoder():
+  library = ctypes.util.find_library("rocprof-trace-decoder")
+  if library:
+    try: return ctypes.CDLL(library)
+    except OSError: pass
+  for candidate in PATHS_TO_TRY:
+    try: return ctypes.CDLL(candidate)
+    except OSError: pass
+  return None
 
 
 class AsDictMixin:
-    import sys
-    if sys.version_info >= (3, 14): _layout_ = 'ms'
     @classmethod
     def as_dict(cls, self):
         result = {}
@@ -157,7 +168,7 @@ class FunctionFactoryStub:
 # You can either re-run clan2py with -l /path/to/library.so
 # Or manually fix this by comment the ctypes.CDLL loading
 _libraries = {}
-_libraries['FIXME_STUB'] = ctypes.CDLL(ctypes.util.find_library('rocprof-trace-decoder')) #  ctypes.CDLL('FIXME_STUB')
+_libraries['FIXME_STUB'] = _try_dlopen_rocprof_trace_decoder() #  ctypes.CDLL('FIXME_STUB')
 
 
 

tinygrad/runtime/ops_amd.py

@@ -7,7 +7,7 @@ from tinygrad.runtime.support.hcq import HCQCompiled, HCQAllocator, HCQBuffer, H
 from tinygrad.runtime.support.hcq import MMIOInterface, BumpAllocator, hcq_filter_visible_devices
 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
+from tinygrad.helpers import getenv, round_up, data64_le, DEBUG, PROFILE, ProfileEvent, suppress_finalizing, lo32, hi32, colored, prod, ContextVar
 from tinygrad.renderer.cstyle import AMDRenderer
 from tinygrad.renderer.llvmir import AMDLLVMRenderer
 from tinygrad.runtime.autogen import kfd, hsa, pci, sqtt
@@ -19,7 +19,7 @@ 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, PMC = getenv("SQTT", 0), getenv("PMC", 0)
+SQTT, SQTT_ITRACE_SE_MASK, PMC = ContextVar("SQTT", 0), 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 # !=
@@ -28,10 +28,10 @@ AQL_HDR = (1 << hsa.HSA_PACKET_HEADER_BARRIER) | (hsa.HSA_FENCE_SCOPE_SYSTEM <<
         | (hsa.HSA_FENCE_SCOPE_SYSTEM << hsa.HSA_PACKET_HEADER_SCRELEASE_FENCE_SCOPE)
 
 @dataclass(frozen=True)
-class ProfileSQTTEvent(ProfileEvent): device:str; se:int; props:dict; blob:bytes; itrace:bool # noqa: E702
+class ProfileSQTTEvent(ProfileEvent): device:str; kern:str; se:int; blob:bytes; itrace:bool # noqa: E702
 
 @dataclass(frozen=True)
-class PMCSample: name:str; block:str; xcc:int; inst:int; se:int; sa:int; wgp:int; off:int; size:int; reg:str # noqa: E702
+class PMCSample: name:str; block:str; xcc:int; inst:int; se:int; sa:int; wgp:int; off:int; size:int; regsample:str # noqa: E702
 
 @dataclass(frozen=True)
 class ProfilePMCEvent(ProfileEvent): device:str; kern:str; sched:list[PMCSample]; blob:bytes # noqa: E702
@@ -72,14 +72,10 @@ class AMDComputeQueue(HWQueue):
     if self.dev.xccs > 1:
       self._q[prev_len-1] |= (len(self._q) - prev_len)
 
-  def set_grbm_broadcast(self):
-    self.wreg(self.gc.regGRBM_GFX_INDEX, **{f'{f}_broadcast_writes': 1 for f in ['se', 'sh' if self.dev.target[0] == 9 else 'sa', 'instance']})
-  def set_grbm_inst(self, n):
-    self.wreg(self.gc.regGRBM_GFX_INDEX, **{f'{f}_broadcast_writes': 1 for f in ['se', 'sh' if self.dev.target[0] == 9 else 'sa']}, instance_index=n)
-  def set_grbm_se_sh(self, se, sh):
-    self.wreg(self.gc.regGRBM_GFX_INDEX, se_index=se, **{f'{"sh" if self.dev.target[0] == 9 else "sa"}_index':sh}, instance_broadcast_writes=1)
-  def set_grbm_se_sh_wgp(self, se, sh, wgp): self.wreg(self.gc.regGRBM_GFX_INDEX, se_index=se, sa_index=sh, instance_index=wgp << 2)
-  def set_grbm_se(self, se): self.wreg(self.gc.regGRBM_GFX_INDEX, se_index=se, sh_broadcast_writes=1, instance_broadcast_writes=1)
+  def set_grbm(self, instance=None, se=None, sh=None, wgp=None):
+    instance_val = (wgp << 2 | (instance or 0)) if wgp is not None else instance
+    self.wreg(self.gc.regGRBM_GFX_INDEX, **{(f'{key}_broadcast_writes' if val is None else f'{key}_index'): (1 if val is None else val)
+      for key, val in [('instance', instance_val), ('se', se), ('sh' if self.dev.target[0] == 9 else 'sa', sh)]})
 
   def wait_reg_mem(self, value, mask=0xffffffff, mem=None, reg=None, reg_done=0, op=WAIT_REG_MEM_FUNCTION_GEQ):
     wrm_info_dw = self.pm4.WAIT_REG_MEM_MEM_SPACE(int(mem is not None)) | self.pm4.WAIT_REG_MEM_OPERATION(int(mem is None and reg_done > 0)) \
@@ -140,7 +136,7 @@ class AMDComputeQueue(HWQueue):
   ### PMC ###
 
   def pmc_reset_counters(self, en=True):
-    self.set_grbm_broadcast()
+    self.set_grbm()
     self.wreg(self.gc.regCP_PERFMON_CNTL if self.dev.target[0] <= 11 else self.gc.regCP_PERFMON_CNTL_1, perfmon_state=0)
     if en: self.wreg(self.gc.regCP_PERFMON_CNTL if self.dev.target[0] <= 11 else self.gc.regCP_PERFMON_CNTL_1, perfmon_state=1)
     return self
@@ -158,18 +154,20 @@ class AMDComputeQueue(HWQueue):
         "SQ": (1, self.dev.se_cnt // self.dev.xccs) + ((1, 1) if gfx9 else (2, self.dev.iface.props['cu_per_simd_array'] // 2))}[block]
       end_off += (rec_size:=prod((self.dev.xccs, inst_cnt, se_cnt, sa_cnt, wgp_cnt)) * 8)
 
-      if (regsel:=getattr(self.gc, (reg:=f'reg{block}_PERFCOUNTER{next(block2pid[block])}') + '_SELECT', None)) is None:
-        raise RuntimeError(f'{block} is out of perfcounter registers: ({reg} is not found)')
+      # gfx11+ and later require even-numbered SQ *_SELECT registers
+      regsample = f'reg{block}_PERFCOUNTER{(pcid:=next(block2pid[block]))}'
+      if (regsel:=getattr(self.gc, (f'reg{block}_PERFCOUNTER{(pcid*2) if self.dev.target[0]>=11 and block=="SQ" else pcid}_SELECT'), None)) is None:
+        raise RuntimeError(f'{block} is out of perfcounter registers: ({regsample} is not found)')
 
       self.wreg(regsel, perf_sel=idx, **({'simd_mask':0xf, 'sqc_bank_mask':0xf, 'sqc_client_mask':0xf} if gfx9 and block == "SQ" else {}))
-      self.dev.pmc_sched.append(PMCSample(name, block, self.dev.xccs, inst_cnt, se_cnt, sa_cnt, wgp_cnt, end_off-rec_size, rec_size, reg))
+      self.dev.pmc_sched.append(PMCSample(name, block, self.dev.xccs, inst_cnt, se_cnt, sa_cnt, wgp_cnt, end_off-rec_size, rec_size, regsample))
 
     if gfx9: self.wreg(self.gc.regSQ_PERFCOUNTER_MASK, sh0_mask=0xffff, sh1_mask=0xffff)
     self.wreg(self.gc.regCOMPUTE_PERFCOUNT_ENABLE, 1)
     return self.pmc_reset_counters(en=True)
 
   def pmc_read(self, buf, sched):
-    self.set_grbm_broadcast()
+    self.set_grbm()
     self.wreg(self.gc.regCP_PERFMON_CNTL if self.dev.target[0] <= 11 else self.gc.regCP_PERFMON_CNTL_1, perfmon_state=1, perfmon_sample_enable=1)
 
     for s in sched:
@@ -178,12 +176,10 @@ class AMDComputeQueue(HWQueue):
       for xcc in range(s.xcc):
         with self.pred_exec(xcc_mask=1 << xcc):
           for inst, se_idx, sa_idx, wgp_idx in itertools.product(range(s.inst), range(s.se), range(s.sa), range(s.wgp)):
-            if s.inst > 1: self.set_grbm_inst(inst)
-            elif self.dev.target[0] == 9: self.set_grbm_se(se_idx)
-            else: self.set_grbm_se_sh_wgp(se_idx, sa_idx, wgp_idx)
+            self.set_grbm(**({'instance':inst} if s.inst > 1 else ({'se':se_idx}|({'sh':sa_idx, 'wgp':wgp_idx} if self.dev.target[0] != 9 else {}))))
 
             # Copy counter to memory (src_sel = perf, dst_sel = tc_l2)
-            lo, hi = getattr(self.gc, f'{s.reg}_LO'), getattr(self.gc, f'{s.reg}_HI', None)
+            lo, hi = getattr(self.gc, f'{s.regsample}_LO'), getattr(self.gc, f'{s.regsample}_HI', None)
             self.pkt3(self.pm4.PACKET3_COPY_DATA, (2 << 8) | 4, lo.addr[0], 0, *data64_le(buf.va_addr+(loff:=next(offset))))
             if hi is not None: self.pkt3(self.pm4.PACKET3_COPY_DATA, (2 << 8) | 4, hi.addr[0], 0, *data64_le(buf.va_addr+loff+4))
 
@@ -201,11 +197,11 @@ class AMDComputeQueue(HWQueue):
       _0=sqtt.union_rgp_sqtt_marker_event_0(_0=sqtt.struct_rgp_sqtt_marker_event_0_0(has_thread_dims=1)),
       _2=sqtt.union_rgp_sqtt_marker_event_2(cmd_id=next(prg.dev.sqtt_next_cmd_id))), *global_size)
 
+    se_cap = max(prod([x if isinstance(x, int) else 1 for x in global_size]) // 4, 1) // 32
     for xcc in range(self.dev.xccs):
       with self.pred_exec(xcc_mask=1 << xcc):
         for i in range(8 if prg.dev.target >= (11,0,0) else 4):
-          self.wreg(getattr(self.gc, f'regCOMPUTE_STATIC_THREAD_MGMT_SE{i}'),
-            ((prg.dev.sqtt_itrace_se_mask >> ((self.dev.se_cnt // self.dev.xccs) * xcc + i)) & 0b1) if SQTT >= 2 else 0xffffffff)
+          self.wreg(getattr(self.gc, f'regCOMPUTE_STATIC_THREAD_MGMT_SE{i}'), min(0xffffffff, (1 << (se_cap + (1 if i == 0 else 0))) - 1))
 
   def sqtt_userdata(self, data, *extra_dwords):
     data_ints = [x[0] for x in struct.iter_unpack('<I', bytes(data))] + list(extra_dwords)
@@ -217,18 +213,18 @@ class AMDComputeQueue(HWQueue):
     self.wreg(self.gc.regSQ_THREAD_TRACE_CTRL, draw_event_en=1, spi_stall_en=1, sq_stall_en=1, reg_at_hwm=2, hiwater=1, util_timer=1,
       mode=int(tracing), **trace_ctrl)
 
-  def sqtt_start(self, buf0s:list[HCQBuffer], se_mask:int):
+  def sqtt_start(self, buf0s:list[HCQBuffer]):
     self.memory_barrier()
     if self.dev.target[0] == 9:
-      self.set_grbm_broadcast()
+      self.set_grbm()
       self.wreg(self.gc.regSQ_THREAD_TRACE_MASK, simd_en=0xf, cu_sel=0, sq_stall_en=1, spi_stall_en=1, reg_stall_en=1, vm_id_mask=0)
       for se in range(len(buf0s)):
         mask = (__SQTT_MISC:=1<<0) | (__SQTT_TIME:=1<<1) | (__SQTT_REG:=1<<2) | (__SQTT_WAVE_START:=1<<3) | (__SQTT_WAVE_END:=1<<6) \
              | (__SQTT_USERDATA:=1<<12) | (__SQTT_REG_CS:=1<<5) | (__SQTT_REG_CS_PRIV:=1<<15)
-        if (se_mask >> se) & 0b1: mask |= (__SQTTINST:=1<<10) | (__SQTT_INST_PC:=1<<11) | (__SQTT_ISSUE:=1<<13)
+        if (SQTT_ITRACE_SE_MASK.value >> se) & 0b1: mask |= (__SQTTINST:=1<<10) | (__SQTT_INST_PC:=1<<11) | (__SQTT_ISSUE:=1<<13)
 
         with self.pred_exec(xcc_mask=1<<(se // (ses_per_xcc:=(self.dev.se_cnt // self.dev.xccs)))):
-          self.set_grbm_se_sh(se % ses_per_xcc, 0)
+          self.set_grbm(se=se % ses_per_xcc, sh=0)
           self.wreg(self.gc.regSQ_THREAD_TRACE_TOKEN_MASK, reg_mask=0xf, token_mask=mask)
           self.wreg(self.gc.regSQ_THREAD_TRACE_TOKEN_MASK2, inst_mask=0xffffffff)
           self.wreg(self.gc.regSQ_THREAD_TRACE_BASE, addr=lo32(buf0s[se].va_addr >> 12))
@@ -240,7 +236,7 @@ class AMDComputeQueue(HWQueue):
       self.spi_config(tracing=True)
       # One buffer for one SE, mesa does it with a single buffer and ac_sqtt_get_data_offset, but this is simpler and should work just as well
       for se in range(len(buf0s)):
-        self.set_grbm_se_sh(se, 0)
+        self.set_grbm(se=se, sh=0)
 
         buf0_lo, buf0_hi = data64_le(buf0s[se].va_addr >> 12)
         if self.dev.target >= (12,0,0):
@@ -263,7 +259,7 @@ class AMDComputeQueue(HWQueue):
         token_exclude = (1 << self.soc.SQ_TT_TOKEN_EXCLUDE_PERF_SHIFT) if self.dev.target < (12,0,0) else 0
 
         # disable instr tracing
-        if not (se_mask >> se) & 0b1:
+        if not (SQTT_ITRACE_SE_MASK.value >> se) & 0b1:
           # gfx12 doesn't have enums with all fields, so it's hardcoded, but it's the same as gfx11.
           token_exclude |= (1 << self.soc.SQ_TT_TOKEN_EXCLUDE_VMEMEXEC_SHIFT | 1 << self.soc.SQ_TT_TOKEN_EXCLUDE_ALUEXEC_SHIFT | \
                             1 << self.soc.SQ_TT_TOKEN_EXCLUDE_VALUINST_SHIFT | 1 << self.soc.SQ_TT_TOKEN_EXCLUDE_IMMEDIATE_SHIFT | \
@@ -273,15 +269,15 @@ class AMDComputeQueue(HWQueue):
                   **({} if self.dev.target < (12,0,0) else {'exclude_barrier_wait': 1}))
         self.sqtt_config(tracing=True)
 
-    self.set_grbm_broadcast()
+    self.set_grbm()
     if self.dev.target[0] > 9: self.wreg(self.gc.regCOMPUTE_THREAD_TRACE_ENABLE, 1)
     self.memory_barrier()
     return self
 
   # Magic values from src/amd/common/ac_sqtt.c:ac_sqtt_emit_stop and src/amd/common/ac_sqtt.c:ac_sqtt_emit_wait
-  def sqtt_stop(self, ses:int, wptrs:HCQBuffer):
+  def sqtt_stop(self, wptrs:HCQBuffer):
     self.memory_barrier()
-    self.set_grbm_broadcast()
+    self.set_grbm()
 
     # Start shutting everything down
     if self.dev.target[0] == 9: self.wreg(self.gc.regSQ_THREAD_TRACE_MODE, mask_cs=1, autoflush_en=1, mode=0)
@@ -290,20 +286,21 @@ class AMDComputeQueue(HWQueue):
       self.pkt3(self.pm4.PACKET3_EVENT_WRITE, self.pm4.EVENT_TYPE(self.soc.THREAD_TRACE_FINISH) | self.pm4.EVENT_INDEX(0))
 
     # For each SE wait for finish to complete and copy regSQ_THREAD_TRACE_WPTR to know where in the buffer trace data ends
-    for se in range(ses):
-      self.set_grbm_se_sh(se, 0)
+    for se in range(self.dev.se_cnt):
+      with self.pred_exec(xcc_mask=1<<(se // (ses_per_xcc:=(self.dev.se_cnt // self.dev.xccs)))):
+        self.set_grbm(se=se % ses_per_xcc, sh=0)
 
-      status_reg = self.gc.regSQ_THREAD_TRACE_STATUS.addr[0] - (self.pm4.PACKET3_SET_UCONFIG_REG_START if self.dev.target[0] == 9 else 0)
-      if self.dev.target >= (10, 0, 0):
-        self.wait_reg_mem(reg=status_reg, mask=self.gc.regSQ_THREAD_TRACE_STATUS.fields_mask('finish_pending'), op=WAIT_REG_MEM_FUNCTION_EQ, value=0)
-        self.sqtt_config(tracing=False)
-      self.wait_reg_mem(reg=status_reg, mask=self.gc.regSQ_THREAD_TRACE_STATUS.fields_mask('busy'), op=WAIT_REG_MEM_FUNCTION_EQ, value=0)
-      self.pkt3(self.pm4.PACKET3_EVENT_WRITE, self.pm4.EVENT_TYPE(self.soc.CS_PARTIAL_FLUSH) | self.pm4.EVENT_INDEX(EVENT_INDEX_PARTIAL_FLUSH))
+        regstatus = self.gc.regSQ_THREAD_TRACE_STATUS.addr[0] - (self.pm4.PACKET3_SET_UCONFIG_REG_START if self.dev.target[0] == 9 else 0)
+        if self.dev.target >= (10,0,0):
+          self.wait_reg_mem(reg=regstatus, mask=self.gc.regSQ_THREAD_TRACE_STATUS.fields_mask('finish_pending'), op=WAIT_REG_MEM_FUNCTION_EQ, value=0)
+          self.sqtt_config(tracing=False)
+        self.wait_reg_mem(reg=regstatus, mask=self.gc.regSQ_THREAD_TRACE_STATUS.fields_mask('busy'), op=WAIT_REG_MEM_FUNCTION_EQ, value=0)
+        self.pkt3(self.pm4.PACKET3_EVENT_WRITE, self.pm4.EVENT_TYPE(self.soc.CS_PARTIAL_FLUSH) | self.pm4.EVENT_INDEX(EVENT_INDEX_PARTIAL_FLUSH))
 
-      # Copy WPTR to memory (src_sel = perf, dst_sel = tc_l2, wr_confirm = True)
-      self.pkt3(self.pm4.PACKET3_COPY_DATA, 1 << 20 | 2 << 8 | 4, self.gc.regSQ_THREAD_TRACE_WPTR.addr[0], 0, *data64_le(wptrs.va_addr+(se*4)))
+        # Copy WPTR to memory (src_sel = perf, dst_sel = tc_l2, wr_confirm = True)
+        self.pkt3(self.pm4.PACKET3_COPY_DATA, 1 << 20 | 2 << 8 | 4, self.gc.regSQ_THREAD_TRACE_WPTR.addr[0], 0, *data64_le(wptrs.va_addr+(se*4)))
 
-    self.set_grbm_broadcast()
+    self.set_grbm()
     if self.dev.target[0] > 9: self.spi_config(tracing=False)
     self.memory_barrier()
     return self
@@ -584,12 +581,31 @@ class AMDProgram(HCQProgram):
     weakref.finalize(self, self._fini, self.dev, self.lib_gpu, buf_spec)
 
   def __call__(self, *bufs, global_size:tuple[int,int,int]=(1,1,1), local_size:tuple[int,int,int]=(1,1,1), vals:tuple[int, ...]=(), wait=False):
+    if self.dev.sqtt_enabled: cast(AMDComputeQueue, self.dev.hw_compute_queue_t()).sqtt_start(self.dev.sqtt_buffers).submit(self.dev)
     res = super().__call__(*bufs, global_size=global_size, local_size=local_size, vals=vals, wait=wait)
     if self.dev.pmc_enabled:
       cast(AMDComputeQueue, self.dev.hw_compute_queue_t()).pmc_read(self.dev.pmc_buffer, self.dev.pmc_sched) \
                                                           .signal(self.dev.timeline_signal, self.dev.next_timeline()).submit(self.dev)
       self.dev.allocator._copyout(pmc_buf:=memoryview(bytearray(self.dev.pmc_buffer.size)), self.dev.pmc_buffer)
       Compiled.profile_events += [ProfilePMCEvent(self.dev.device, self.name, self.dev.pmc_sched, bytes(pmc_buf))]
+    if self.dev.sqtt_enabled:
+      cast(AMDComputeQueue, self.dev.hw_compute_queue_t()).sqtt_stop(self.dev.sqtt_wptrs) \
+                                                          .signal(self.dev.timeline_signal, self.dev.next_timeline()).submit(self.dev)
+      self.dev.synchronize()
+
+      for se, buf in enumerate(self.dev.sqtt_buffers):
+        wptr = ((self.dev.sqtt_wptrs.cpu_view().view(fmt='I')[se]&0x1FFFFFFF)-(((buf.va_addr//32)&0x1FFFFFFF) if self.dev.target[0] == 11 else 0))*32
+
+        if DEBUG >= 5: print(f'\t{self.dev.device}: SE {se} blob size {wptr:#x}')
+        assert wptr >= 0 and wptr <= buf.size, f"{wptr} > {buf.size}, should never happen"
+
+        # When sqtt buffer overflows, wptr stops at the last dword
+        if wptr >= buf.size - 32:
+          print(colored(f"{self.dev.device}: Warning: SQTT buffer is full (SE {se})! Increase SQTT buffer with SQTT_BUFFER_SIZE=X (in MB)", "yellow"))
+
+        self.dev.allocator._copyout(sqtt_mv:=memoryview(bytearray(wptr)), buf)
+        resbuf = (struct.pack('<Q', 0x11 | (4 << 13) | (0xf << 16) | (se << 24)) + bytes(sqtt_mv)) if self.dev.target[0] == 9 else bytes(sqtt_mv)
+        Compiled.profile_events += [ProfileSQTTEvent(self.dev.device, self.name, se, resbuf, bool((SQTT_ITRACE_SE_MASK.value >> se) & 1))]
     return res
 
 class AMDAllocator(HCQAllocator['AMDDevice']):
@@ -815,7 +831,7 @@ class PCIIface(PCIIfaceBase):
 
 class USBIface(PCIIface):
   def __init__(self, dev, dev_id): # pylint: disable=super-init-not-called
-    self.dev, self.pci_dev = dev, USBPCIDevice(f"usb:{dev_id}", bars=[0, 2, 5])
+    self.dev, self.pci_dev = dev, USBPCIDevice(dev.__class__.__name__[:2], f"usb:{dev_id}", bars=[0, 2, 5])
     self._setup_adev(self.pci_dev, dma_regions=[(0x200000, self.pci_dev.dma_view(0xf000, 0x80000))])
     self.pci_dev.usb._pci_cacheable += [(self.pci_dev.bar_info[2].addr, self.pci_dev.bar_info[2].size)] # doorbell region is cacheable
 
@@ -932,9 +948,8 @@ class AMDDevice(HCQCompiled):
 
       SQTT_BUFFER_SIZE = getenv("SQTT_BUFFER_SIZE", 256) # in mb, per shader engine
       self.sqtt_buffers = [self.allocator.alloc(SQTT_BUFFER_SIZE << 20, BufferSpec(nolru=True, uncached=True)) for _ in range(self.se_cnt)]
-      self.sqtt_itrace_se_mask = getenv("SQTT_ITRACE_SE_MASK", -1 if SQTT >= 2 else (1 << 1)) # se bitmask: -1 enable all, 0 disable all
+      self.sqtt_wptrs = self.allocator.alloc(round_up(self.se_cnt * 4, 0x1000), BufferSpec(cpu_access=True, nolru=True))
       self.sqtt_next_cmd_id = itertools.count(0)
-      cast(AMDComputeQueue, self.hw_compute_queue_t()).sqtt_start(self.sqtt_buffers, self.sqtt_itrace_se_mask).submit(self)
 
   def create_queue(self, queue_type, ring_size, ctx_save_restore_size=0, eop_buffer_size=0, ctl_stack_size=0, debug_memory_size=0):
     ring = self.iface.alloc(ring_size, uncached=True, cpu_access=True)
@@ -984,21 +999,4 @@ class AMDDevice(HCQCompiled):
 
   def on_device_hang(self): self.iface.on_device_hang()
 
-  def _at_profile_finalize(self):
-    if self.sqtt_enabled:
-      wptrs_buf = self.allocator.alloc(round_up(len(self.sqtt_buffers), 0x1000), BufferSpec(cpu_access=True, nolru=True))
-      cast(AMDComputeQueue, self.hw_compute_queue_t()).sqtt_stop(len(self.sqtt_buffers), wptrs_buf) \
-                                                      .signal(self.timeline_signal, self.next_timeline()).submit(self)
-      self.synchronize()
-      if DEBUG >= 2: print(f'{self.device}: Saving SQTT in profile...')
-      for i,buf0 in enumerate(self.sqtt_buffers):
-        wptr = ((wptrs_buf.cpu_view().view(fmt='I')[i] & 0x1FFFFFFF) - (((buf0.va_addr//32) & 0x1FFFFFFF) if self.target[0] == 11 else 0)) * 32
-        if DEBUG >= 2: print(f'\t{self.device}: SE {i} blob size {wptr:#x}')
-        assert wptr >= 0 and wptr <= buf0.size, f"{wptr} > {buf0.size}, should never happen"
-        # When sqtt buffer overflows, wptr stops at the last dword
-        if wptr >= buf0.size - 32:
-          print(colored(f"{self.device}: Warning: SQTT buffer is full (SE {i})! Increase SQTT buffer with SQTT_BUFFER_SIZE=X (in MB)", "yellow"))
-        self.allocator._copyout(sqtt_buf:=memoryview(bytearray(wptr)), buf0)
-        if self.target[0] == 9: sqtt_buf = memoryview(struct.pack('<Q', 0x11 | (4 << 13) | (0xf << 16) | (i << 24)) + sqtt_buf)
-        Compiled.profile_events += [ProfileSQTTEvent(self.device, i, self.iface.props, bytes(sqtt_buf), bool((self.sqtt_itrace_se_mask >> i) & 0b1))]
-    super()._at_profile_finalize()
+  def device_props(self): return self.iface.props

tinygrad/runtime/ops_qcom.py

@@ -322,7 +322,7 @@ class QCOMDevice(HCQCompiled):
     QCOMDevice.dummy_addr = cast(int, self._gpu_alloc(0x1000).va_addr)
 
     flags = kgsl.KGSL_CONTEXT_PREAMBLE | kgsl.KGSL_CONTEXT_PWR_CONSTRAINT | kgsl.KGSL_CONTEXT_NO_FAULT_TOLERANCE | kgsl.KGSL_CONTEXT_NO_GMEM_ALLOC \
-              | kgsl.KGSL_CONTEXT_PRIORITY(8) | kgsl.KGSL_CONTEXT_PREEMPT_STYLE(kgsl.KGSL_CONTEXT_PREEMPT_STYLE_FINEGRAIN)
+              | kgsl.KGSL_CONTEXT_PRIORITY(getenv("QCOM_PRIORITY", 8)) | kgsl.KGSL_CONTEXT_PREEMPT_STYLE(kgsl.KGSL_CONTEXT_PREEMPT_STYLE_FINEGRAIN)
     self.ctx = kgsl.IOCTL_KGSL_DRAWCTXT_CREATE(self.fd, flags=flags).drawctxt_id
 
     self.cmd_buf = self._gpu_alloc(16 << 20)

tinygrad/runtime/support/am/amdev.py

@@ -1,11 +1,11 @@
 from __future__ import annotations
-import ctypes, collections, dataclasses, functools, os, hashlib, array
+import ctypes, collections, dataclasses, functools, hashlib, array
 from tinygrad.helpers import mv_address, getenv, DEBUG, fetch
 from tinygrad.runtime.autogen.am import am
 from tinygrad.runtime.support.hcq import MMIOInterface
 from tinygrad.runtime.support.amd import AMDReg, import_module, import_asic_regs
 from tinygrad.runtime.support.memory import TLSFAllocator, MemoryManager
-from tinygrad.runtime.support.system import System, PCIDevice, PCIDevImplBase
+from tinygrad.runtime.support.system import PCIDevice, PCIDevImplBase
 from tinygrad.runtime.support.am.ip import AM_SOC, AM_GMC, AM_IH, AM_PSP, AM_SMU, AM_GFX, AM_SDMA
 
 AM_DEBUG = getenv("AM_DEBUG", 0)
@@ -122,8 +122,6 @@ class AMDev(PCIDevImplBase):
     self.pci_dev, self.devfmt, self.dma_regions = pci_dev, pci_dev.pcibus, dma_regions
     self.vram, self.doorbell64, self.mmio = self.pci_dev.map_bar(0), self.pci_dev.map_bar(2, fmt='Q'), self.pci_dev.map_bar(5, fmt='I')
 
-    self.lock_fd = System.flock_acquire(f"am_{self.devfmt}.lock")
-
     self._run_discovery()
     self._build_regs()
 
@@ -190,7 +188,6 @@ class AMDev(PCIDevImplBase):
     for ip in [self.sdma, self.gfx]: ip.fini_hw()
     self.smu.set_clocks(level=0)
     self.ih.interrupt_handler()
-    os.close(self.lock_fd)
 
   def paddr2mc(self, paddr:int) -> int: return self.gmc.mc_base + paddr
 

tinygrad/runtime/support/compiler_amd.py

@@ -13,8 +13,9 @@ from tinygrad.runtime.support.compiler_cpu import LLVMCompiler
 from tinygrad.helpers import OSX, to_char_p_p
 
 def amdgpu_disassemble(lib:bytes):
-  asm = subprocess.check_output(["llvm-objdump" if OSX else "/opt/rocm/llvm/bin/llvm-objdump", '-d', '-'], input=lib)
-  print('\n'.join([x for x in asm.decode('utf-8').split("\n") if 's_code_end' not in x]))
+  asm = subprocess.check_output(["llvm-objdump" if OSX else "/opt/rocm/llvm/bin/llvm-objdump", '-d', '-'], input=lib).decode("utf-8").splitlines()
+  while asm and ("s_nop 0" in asm[-1] or "s_code_end" in asm[-1]): asm.pop()
+  print("\n".join(asm))
 
 def check(status):
   if status != 0:

tinygrad/runtime/support/hcq.py

@@ -409,6 +409,8 @@ class HCQCompiled(Compiled, Generic[SignalType]):
       for dev in HCQCompiled.peer_groups[pg]: cast(HCQAllocator, dev.allocator).map(alc)
     return self.signal_t(base_buf=HCQCompiled.signal_pool[pg].pop(), owner=self, **kwargs)
 
+  def device_props(self) -> dict[str,Any]: return {} # to be overridden if needed. dict keys are backend dependent.
+
   def _at_profile_finalize(self):
     self.synchronize() # Expect device to be synchronizes
 
@@ -422,7 +424,7 @@ class HCQCompiled(Compiled, Generic[SignalType]):
     gpu2cpu_compute_time_diff = statistics.median([_sync(self, self.hw_compute_queue_t) for _ in range(40)])
     if self.hw_copy_queue_t is None: gpu2cpu_copy_time_diff = decimal.Decimal(0)
     else: gpu2cpu_copy_time_diff = statistics.median([_sync(self, self.hw_copy_queue_t) for _ in range(40)])
-    Compiled.profile_events += [ProfileDeviceEvent(self.device, gpu2cpu_compute_time_diff, gpu2cpu_copy_time_diff)]
+    Compiled.profile_events += [ProfileDeviceEvent(self.device, gpu2cpu_compute_time_diff, gpu2cpu_copy_time_diff, props=self.device_props())]
 
   def _wrap_timeline_signal(self):
     self.timeline_signal, self._shadow_timeline_signal, self.timeline_value = self._shadow_timeline_signal, self.timeline_signal, 1

tinygrad/runtime/support/nv/nvdev.py

@@ -73,8 +73,6 @@ class NVDev(PCIDevImplBase):
   def __init__(self, pci_dev:PCIDevice):
     self.pci_dev, self.devfmt, self.mmio = pci_dev, pci_dev.pcibus, pci_dev.map_bar(0, fmt='I')
 
-    self.lock_fd = System.flock_acquire(f"nv_{self.devfmt}.lock")
-
     self.smi_dev, self.is_booting = False, True
     self._early_init()
 

tinygrad/runtime/support/system.py

@@ -165,7 +165,8 @@ class _System:
 System = _System()
 
 class PCIDevice:
-  def __init__(self, pcibus:str, bars:list[int], resize_bars:list[int]|None=None):
+  def __init__(self, devpref:str, pcibus:str, bars:list[int], resize_bars:list[int]|None=None):
+    self.lock_fd = System.flock_acquire(f"{devpref.lower()}_{pcibus.lower()}.lock")
     self.pcibus, self.irq_poller = pcibus, None
 
     if FileIOInterface.exists(f"/sys/bus/pci/devices/{self.pcibus}/driver"):
@@ -215,7 +216,8 @@ class PCIDevice:
   def reset(self): os.system(f"sudo sh -c 'echo 1 > /sys/bus/pci/devices/{self.pcibus}/reset'")
 
 class APLPCIDevice(PCIDevice):
-  def __init__(self, pcibus:str, bars:list[int], resize_bars:list[int]|None=None):
+  def __init__(self, devpref:str, pcibus:str, bars:list[int], resize_bars:list[int]|None=None):
+    self.lock_fd = System.flock_acquire(f"{devpref.lower()}_{pcibus.lower()}.lock")
     self.pcibus, self.bars = pcibus, {b: System.iokit_pci_memmap(b) for b in bars}
     self.bar_info = {b:PCIBarInfo(0, self.bars[b].nbytes-1 if b in self.bars else 0) for b in range(6)} # NOTE: fake bar info for nv.
   def map_bar(self, bar:int, off:int=0, addr:int=0, size:int|None=None, fmt='B') -> MMIOInterface: return self.bars[bar].view(off, size, fmt)
@@ -224,7 +226,8 @@ class APLPCIDevice(PCIDevice):
   def reset(self): System.iokit_pci_rpc(__TinyGPURPCReset:=2)
 
 class USBPCIDevice(PCIDevice):
-  def __init__(self, pcibus:str, bars:list[int], resize_bars:list[int]|None=None):
+  def __init__(self, devpref:str, pcibus:str, bars:list[int], resize_bars:list[int]|None=None):
+    self.lock_fd = System.flock_acquire(f"{devpref.lower()}_{pcibus.lower()}.lock")
     self.usb = ASM24Controller()
     self.pcibus, self.bar_info = pcibus, System.pci_setup_usb_bars(self.usb, gpu_bus=4, mem_base=0x10000000, pref_mem_base=(32 << 30))
   def map_bar(self, bar, off=0, addr=0, size=None, fmt='B'):
@@ -247,12 +250,12 @@ class LNXPCIIfaceBase:
 
       # Acquire va range to avoid collisions.
       FileIOInterface.anon_mmap(va_start, va_size, 0, mmap.MAP_PRIVATE | mmap.MAP_ANONYMOUS | MAP_NORESERVE | MAP_FIXED, 0)
-    self.pci_dev, self.dev, self.vram_bar = PCIDevice(cls.gpus[dev_id], bars=bars, resize_bars=[vram_bar]), dev, vram_bar
+    self.pci_dev, self.dev, self.vram_bar = PCIDevice(dev.__class__.__name__[:2], cls.gpus[dev_id], bars=bars, resize_bars=[vram_bar]), dev, vram_bar
     self.p2p_base_addr = self.pci_dev.bar_info[vram_bar].addr
 
   def alloc(self, size:int, host=False, uncached=False, cpu_access=False, contiguous=False, force_devmem=False, **kwargs) -> HCQBuffer:
     # NOTE: logic on macos is different, since bar is small
-    should_use_sysmem = host or (((uncached or cpu_access) if OSX else (uncached and cpu_access)) and not force_devmem)
+    should_use_sysmem = host or ((cpu_access if OSX else (uncached and cpu_access)) and not force_devmem)
     if should_use_sysmem:
       vaddr = self.dev_impl.mm.alloc_vaddr(size:=round_up(size, mmap.PAGESIZE), align=mmap.PAGESIZE)
       memview, paddrs = System.alloc_sysmem(size, vaddr=vaddr, contiguous=contiguous)
@@ -281,7 +284,7 @@ class LNXPCIIfaceBase:
 
 class APLPCIIfaceBase(LNXPCIIfaceBase):
   def __init__(self, dev, dev_id, vendor, devices, bars, vram_bar, va_start, va_size):
-    self.pci_dev, self.dev, self.vram_bar = APLPCIDevice(pcibus=f'usb4:{dev_id}', bars=bars), dev, vram_bar
+    self.pci_dev, self.dev, self.vram_bar = APLPCIDevice(dev.__class__.__name__[:2], pcibus=f'usb4:{dev_id}', bars=bars), dev, vram_bar
   def map(self, b:HCQBuffer): raise RuntimeError(f"map failed: {b.owner} -> {self.dev}")
 
 PCIIfaceBase:type = APLPCIIfaceBase if OSX else LNXPCIIfaceBase

tinygrad/schedule/multi.py

@@ -68,7 +68,7 @@ def handle_allreduce(buf:UOp, red:UOp) -> UOp|None:
   # allgather
   copied_chunks = []
   for i,c in enumerate(reduced_chunks):
-    this_chunk = [None] * len(buf.device)
+    this_chunk: list[UOp|None] = [None] * len(buf.device)
     this_chunk[(i+len(buf.device)-1)%n_lbs] = c
     for step in range(n_lbs-1):
       dest = (i+step)%n_lbs
@@ -186,7 +186,7 @@ def shrink_multi(root:UOp, multi:UOp):
 
 def flip_multi(root:UOp, multi:UOp):
   assert multi.axis is None or not root.marg[multi.axis], "flipping not supported on sharded axis"
-  return multi.src[0].flip(root.marg).multi(multi.axis)
+  return multi.src[0].flip([i for i,x in enumerate(root.marg) if x]).multi(multi.axis)
 
 # from multiple devices -> one
 def copy_multi(multi:UOp, device:UOp):
@@ -214,7 +214,7 @@ multi_pm = PatternMatcher([
   (UPat(Ops.COPY, src=(UPat(Ops.MULTI, name="multi"), UPat(Ops.DEVICE, name="device"))), copy_multi),
   (UPat(Ops.ALLREDUCE, src=(UPat(Ops.MULTI, name="multi"), UPat(Ops.DEVICE, name="device")), name="red"),
     lambda multi,device,red: multi.src[0].allreduce(red.arg, device).multi(axis=multi.axis)),
-  (UPat((Ops.CAST, Ops.BITCAST, Ops.CONTIGUOUS, Ops.DETACH, Ops.CONTIGUOUS_BACKWARD, Ops.FUSE),
+  (UPat((Ops.CAST, Ops.BITCAST, Ops.CONTIGUOUS, Ops.DETACH, Ops.CONTIGUOUS_BACKWARD),
         src=(UPat(Ops.MULTI, name="multi"), ), name="root"), passthrough_multi),
 ])+replace_allreduce
 

tinygrad/schedule/rangeify.py

@@ -3,7 +3,7 @@ import itertools
 from tinygrad.dtype import dtypes, PtrDType, ImageDType, AddrSpace
 from tinygrad.uop.ops import PatternMatcher, UPat, Ops, UOp, resolve, GroupOp, _substitute, ssimplify, KernelInfo
 from tinygrad.uop.ops import track_rewrites, graph_rewrite, identity_element, sint, AxisType, BottomUpGate, Kernel, _remove_all_tags
-from tinygrad.uop.symbolic import symbolic_flat
+from tinygrad.uop.symbolic import symbolic
 from tinygrad.helpers import argsort, prod, all_same, pluralize, getenv, flatten, dedup, all_int, DEBUG, SPLIT_REDUCEOP, DEBUG_RANGEIFY
 from tinygrad.helpers import PCONTIG, partition, get_single_element, unwrap
 from tinygrad.codegen.simplify import pm_flatten_range, pm_reduce_simplify
@@ -17,7 +17,7 @@ sys.setrecursionlimit(10000)
 # movement op on INDEX as a PatternMatcher
 pm_mops = PatternMatcher([
   (UPat(GroupOp.Movement, name="r").f(Ops.INDEX, allow_any_len=True, name="idx"),
-   lambda r,idx: r.src[0].index(*apply_movement_op(r.op, r.src[0].shape, r.marg, idx.src[1:]), dtype=idx.dtype, arg=idx.arg)),  # type: ignore
+   lambda r,idx: r.src[0].index(*apply_movement_op(r.op, r.src[0].shape, r.marg, idx.src[1:]), dtype=idx.dtype, arg=idx.arg)),
   # move movement ops after AFTER
   (UPat(GroupOp.Movement, name="r").after(name="a", allow_any_len=True),
    lambda r,a: UOp(r.op, r.dtype, (a.replace(src=(r.src[0],)+a.src[1:], tag=None),)+r.src[1:], r.arg, tag=a.tag)),
@@ -65,7 +65,7 @@ mop_cleanup = PatternMatcher([
 
 earliest_rewrites = mop_cleanup+PatternMatcher([
   # just removing it works...
-  (UPat((Ops.DETACH, Ops.CONTIGUOUS_BACKWARD, Ops.FUSE), name="x"), lambda x: x.src[0]),
+  (UPat((Ops.DETACH, Ops.CONTIGUOUS_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)),
@@ -536,7 +536,7 @@ 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_flat+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, name="symbolic+reduce_collapse")  # this does const folding
   tsink = graph_rewrite(tsink, pm_remove_bufferize, bottom_up=True, name="remove bufferize with cost function")
   tsink = graph_rewrite(tsink, pm_limit_bufs, ctx=rctx, name="limit buffers")
 

tinygrad/tensor.py

@@ -5,11 +5,11 @@ from contextlib import ContextDecorator
 from typing import Callable, ClassVar, Sequence, cast, get_args, Literal, SupportsIndex, ParamSpec, TypeVar, Generic
 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
-from tinygrad.helpers import argfix, make_tuple, flatten, prod, all_int, round_up, merge_dicts, argsort, getenv, all_same, fully_flatten, dedup
-from tinygrad.helpers import IMAGE, WINO, Metadata, TRACEMETA, ceildiv, fetch, polyN, DEBUG, is_numpy_ndarray, FUSE_ATTENTION, SPEC
+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, DEBUG, is_numpy_ndarray, SPEC
 from tinygrad.helpers import suppress_finalizing
 from tinygrad.gradient import compute_gradient
-from tinygrad.uop.mixins import MathMixin, MovementMixin
+from tinygrad.mixin import OpMixin
 from tinygrad.uop.ops import smax, smin, resolve, UOp, Ops, sint, identity_element, all_metadata, _index_to_concrete_int, sint_to_uop
 from tinygrad.uop.spec import type_verify, tensor_spec
 from tinygrad.device import Device, Buffer
@@ -100,7 +100,7 @@ def _flat_to_grouped(padding:Sequence[sint]) -> tuple[tuple[sint, sint], ...]: r
 
 ReductionStr = Literal["mean", "sum", "none"]
 
-class Tensor(MathMixin, MovementMixin):
+class Tensor(OpMixin):
   """
   A `Tensor` is a multi-dimensional matrix containing elements of a single data type.
 
@@ -1055,65 +1055,6 @@ class Tensor(MathMixin, MovementMixin):
     new_shape = tuple(from_ if to == -1 or to is None else to for from_, to in zip(*(_align_left(self.shape, argfix(shape, *args)))))
     return self._broadcast_to(new_shape)
 
-  def permute(self, order, *args) -> Tensor:
-    """
-    Returns a tensor that is a permutation of the original tensor.
-    The new tensor has the same data as the original tensor but with the dimensions permuted according to the order specified.
-    `order` can be passed as a tuple or as separate arguments.
-
-    ```python exec="true" source="above" session="tensor" result="python"
-    t = Tensor.empty(2, 3, 5)
-    print(t.shape)
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(t.permute(2, 0, 1).shape)
-    ```
-    """
-    order_arg = tuple(self._resolve_dim(x) for x in argfix(order, *args))
-    if sorted(order_arg) != list(range(self.ndim)): raise RuntimeError(f"order is not a valid permutation, getting {order_arg}")
-    return self._apply_uop(UOp.permute, arg=order_arg)
-
-  def flip(self, axis, *args) -> Tensor:
-    """
-    Returns a tensor that reverses the order of the original tensor along given `axis`.
-    `axis` can be passed as a tuple or as separate arguments.
-
-    ```python exec="true" source="above" session="tensor" result="python"
-    t = Tensor.arange(6).reshape(2, 3)
-    print(t.numpy())
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(t.flip(0).numpy())
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(t.flip((0, 1)).numpy())
-    ```
-    """
-    axis_arg = tuple(self._resolve_dim(x) for x in argfix(axis, *args))
-    if len(axis_arg) != len(dedup(axis_arg)): raise RuntimeError(f"dim can appear at most once, getting {axis_arg}")
-    return self._apply_uop(UOp.flip, arg=tuple([i in axis_arg for i in range(len(self.shape))]))
-
-  def shrink(self, arg:tuple[tuple[sint, sint]|None, ...]) -> Tensor:
-    """
-    Returns a tensor that shrinks the each axis based on input arg.
-    `arg` must have the same length as `self.ndim`.
-    For each axis, it can be `None`, which means no shrink, or a tuple `(start, end)` that works the same as Python slice.
-
-    ```python exec="true" source="above" session="tensor" result="python"
-    t = Tensor.arange(9).reshape(3, 3)
-    print(t.numpy())
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(t.shrink(((None, (1, 3)))).numpy())
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(t.shrink((((0, 2), (0, 2)))).numpy())
-    ```
-    """
-    if self.ndim != len(arg): raise ValueError(f"{self.ndim=} != {len(arg)=}")
-    if (shrink_arg:=[x if x is not None else (0,s) for x,s in zip(arg, self.shape)]) == [(0,s) for s in self.shape]: return self
-    return self._apply_uop(UOp.shrink, arg=tuple(shrink_arg))
-
   def pad(self, padding:Sequence[sint]|Sequence[tuple[sint, sint]|None], mode:str="constant", value:float=0.0) -> Tensor:
     """
     Returns a tensor with padding applied based on the input `padding`.
@@ -1435,11 +1376,6 @@ class Tensor(MathMixin, MovementMixin):
     final_shape = [r*s for r,s in zip(repeats, base_shape)]
     return self.reshape(unsqueezed_shape).expand(expanded_shape).reshape(final_shape)
 
-  def _resolve_dim(self, dim:int, *, extra:bool=False) -> int:
-    total = self.ndim + int(extra)
-    if not -max(1, total) <= dim <= max(1, total)-1: raise IndexError(f"{dim=} out of range {[-max(1, total), max(1, total)-1]}")
-    return dim + total if dim < 0 else dim
-
   def split(self, sizes:int|Sequence[int], dim:int=0) -> tuple[Tensor, ...]:
     """
     Splits the tensor into chunks along the dimension specified by `dim`.
@@ -1540,96 +1476,6 @@ class Tensor(MathMixin, MovementMixin):
     output_shape = _broadcast_shape(*(t.shape for t in tensors))
     return tuple(t._broadcast_to(output_shape) for t in tensors)
 
-  def squeeze(self, dim:int|None=None) -> Tensor:
-    """
-    Returns a tensor with specified dimensions of input of size 1 removed.
-    If `dim` is not specified, all dimensions with size 1 are removed.
-
-    ```python exec="true" source="above" session="tensor" result="python"
-    t = Tensor.zeros(2, 1, 2, 1, 2)
-    print(t.squeeze().shape)
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(t.squeeze(0).shape)
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(t.squeeze(1).shape)
-    ```
-    """
-    if dim is None: return self.reshape(tuple(dim for dim in self.shape if dim != 1))
-    dim = self._resolve_dim(dim)
-    return self if not self.ndim or self.shape[dim] != 1 else self.reshape(self.shape[:dim] + self.shape[dim+1:])
-
-  def unsqueeze(self, dim:int) -> Tensor:
-    """
-    Returns a tensor with a new dimension of size 1 inserted at the specified `dim`.
-
-    ```python exec="true" source="above" session="tensor" result="python"
-    t = Tensor([1, 2, 3, 4])
-    print(t.unsqueeze(0).numpy())
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(t.unsqueeze(1).numpy())
-    ```
-    """
-    dim = self._resolve_dim(dim, extra=True)
-    return self.reshape(self.shape[:dim] + (1,) + self.shape[dim:])
-
-  @property
-  def T(self) -> Tensor:
-    """`.T` is an alias for `.transpose()`."""
-    return self.transpose()
-
-  def transpose(self, dim0=1, dim1=0) -> Tensor:
-    """
-    Returns a tensor that is a transposed version of the original tensor.
-    The given dimensions `dim0` and `dim1` are swapped.
-
-    ```python exec="true" source="above" session="tensor" result="python"
-    t = Tensor.arange(6).reshape(2, 3)
-    print(t.numpy())
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(t.transpose(0, 1).numpy())
-    ```
-    """
-    order = list(range(self.ndim))
-    order[dim0], order[dim1] = order[dim1], order[dim0]
-    return self.permute(order)
-
-  def flatten(self, start_dim=0, end_dim=-1) -> Tensor:
-    """
-    Flattens the tensor by reshaping it into a one-dimensional tensor.
-    If `start_dim` or `end_dim` are passed, only dimensions starting with `start_dim` and ending with `end_dim` are flattened.
-
-    ```python exec="true" source="above" session="tensor" result="python"
-    t = Tensor.arange(8).reshape(2, 2, 2)
-    print(t.flatten().numpy())
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(t.flatten(start_dim=1).numpy())
-    ```
-    """
-    start_dim, end_dim = self._resolve_dim(start_dim), self._resolve_dim(end_dim)
-    return self.reshape(self.shape[:start_dim] + (prod(self.shape[start_dim:end_dim+1]), ) + self.shape[end_dim+1:])
-
-  def unflatten(self, dim:int, sizes:tuple[int,...]) -> Tensor:
-    """
-    Unflattens dimension `dim` of the tensor into multiple dimensions specified by `sizes`. `Tensor.flatten()` is the inverse of this function.
-
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(Tensor.ones(3, 4, 1).unflatten(1, (2, 2)).shape)
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(Tensor.ones(3, 4, 1).unflatten(1, (-1, 2)).shape)
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(Tensor.ones(5, 12, 3).unflatten(-2, (2, 2, 3, 1, 1)).shape)
-    ```
-    """
-    dim = self._resolve_dim(dim)
-    return self.reshape(self.shape[:dim] + sizes + self.shape[dim+1:])
-
   def diag(self) -> Tensor:
     """
     Returns a 2-D square tensor with the elements of input as the main diagonal.
@@ -1675,46 +1521,6 @@ class Tensor(MathMixin, MovementMixin):
     for dim, shift in zip(dims, shifts): slices[dim] = slice(delta:=self.shape[dim]-shift%self.shape[dim], delta+self.shape[dim])
     return self.repeat(*tuple(2 if i in dims else 1 for i in range(self.ndim)))[slices]
 
-  def rearrange(self, formula:str, **sizes) -> Tensor:
-    """
-    Rearranges input according to formula
-
-    See: https://einops.rocks/api/rearrange/
-
-    ```python exec="true" source="above" session="tensor" result="python"
-    x = Tensor([[1, 2], [3, 4]])
-    print(Tensor.rearrange(x, "batch channel -> (batch channel)").numpy())
-    ```
-    """
-    def parse_formula(formula: str):
-      tokens = f" {formula} ".replace("…", "...").replace("(", " ( ").replace(")", " ) ").replace(" ", "  ").replace(" 1 ", " ( ) ").split()
-      lparens, rparens = map(lambda x: [i for i, ch in enumerate(tokens) if ch == x], ("(", ")"))
-      pairs = list(zip(lparens, rparens))
-      assert len(lparens) == len(rparens) and sorted(flatten(pairs)) == flatten(pairs), "bracket mismatch"
-      return [name for name in tokens if name not in ("(", ")")], [(s - 2*i, e - 1 - 2*i) for i, (s, e) in enumerate(pairs)]
-
-    assert formula.count("->") == 1, 'need exactly one "->" in formula'
-
-    (lhs, unflatten_dims), (rhs, flatten_dims) = map(parse_formula, formula.split("->"))
-
-    for name in sizes: assert name in lhs, f"axis {name} is not used in transform"
-    assert sorted(lhs) == sorted(rhs) and len(lhs) == len(set(lhs)), f"name mismatch in {formula}"
-    for name in flatten((lhs, rhs)): assert name == "..." or (name.isidentifier() and "_" not in (name[0], name[-1])), f"invalid axis name {name}"
-    assert "..." not in flatten([lhs[s:e] for s, e in unflatten_dims]), f"cannot have collapsed ellipsis (...) in lhs of {formula}"
-    assert lhs.count("...") <= 1, f"too many ellipses in {formula}"
-
-    # resolve ellipsis
-    if "..." in lhs: ell_len = len(self.shape) - len(lhs) + 1 + sum(e - s - 1 for s, e in unflatten_dims)
-    lhs, rhs = map(lambda l: l[:(i:=l.index("..."))] + [f"...{j}" for j in range(ell_len)] + l[i + 1:] if "..." in l else l, (lhs, rhs))
-    unflatten_dims = [(s + (ell_len - 1 if "...0" in lhs[:s] else 0), e + (ell_len - 1 if "...0" in lhs[:e] else 0)) for s, e in unflatten_dims]
-    flatten_dims = [(s + (ell_len - 1 if "...0" in rhs[:s] else 0), e + (ell_len - 1 if "...0" in rhs[:e] else 0)) for s, e in flatten_dims]
-
-    # apply movement ops in order unflatten -> permute -> flatten/unsqueeze
-    t = functools.reduce(lambda x, dims: x.unflatten(dims[0], tuple(sizes.get(lhs[d], -1) for d in range(*dims))), unflatten_dims, self)
-    for i, name in enumerate(lhs): assert (name not in sizes) or sizes[name] == t.shape[i], f"size provided for dimension {name} incorrect"
-    t = t.permute([lhs.index(name) for name in rhs])
-    return functools.reduce(lambda x, dims: x.flatten(dims[0], dims[1] - 1) if dims[0]<dims[1] else x.unsqueeze(dims[0]), reversed(flatten_dims), t)
-
   def masked_select(self, mask):
     """
     Selects elements from `self` based on the boolean `mask`.
@@ -1927,6 +1733,12 @@ class Tensor(MathMixin, MovementMixin):
     is_nan_close = (self.isnan() & other.isnan()) & equal_nan
     return is_finite_close | is_infinite_close | is_nan_close
 
+  def allclose(self, other:Tensor, rtol:float=1e-05, atol:float=1e-08, equal_nan=False) -> bool:
+    """
+    Check if all self and other are close. Return True or False.
+    """
+    return bool(self.isclose(other, rtol=rtol, atol=atol, equal_nan=equal_nan).all().item())
+
   def mean(self, axis:int|Sequence[int]|None=None, keepdim=False) -> Tensor:
     """
     Returns the mean value of the tensor along the specified axis or axes.
@@ -2131,7 +1943,7 @@ class Tensor(MathMixin, MovementMixin):
     e = m.exp()
     return m, e, e.sum(axis=axis, keepdim=True)
 
-  def softmax(self, axis=-1, dtype:DTypeLike|None=None, _single_kernel=getenv("SINGLE_KERNEL_SOFTMAX")) -> Tensor:
+  def softmax(self, axis=-1, dtype:DTypeLike|None=None) -> Tensor:
     """
     Applies the softmax function to the tensor along the specified axis.
 
@@ -2151,9 +1963,6 @@ class Tensor(MathMixin, MovementMixin):
     print(t.softmax(axis=0).numpy())
     ```
     """
-    if _single_kernel:
-      _, e, ss = self.contiguous()._softmax(axis, dtype)
-      return e.div(ss).fuse()
     _, e, ss = self._softmax(axis, dtype)
     return e.div(ss)
 
@@ -3014,15 +2823,6 @@ class Tensor(MathMixin, MovementMixin):
     """
     return self._apply_uop(UOp.contiguous, extra_args=args, **kwargs)
 
-  def fuse(self) -> Tensor:
-    """
-    Makes this a single kernel back to Ops.CONTIGUOUS on the inputs.
-
-    Useful for single kernel softmax and flash attention.
-    Careful, this can break codegen or make kernels really slow.
-    """
-    return self._apply_uop(UOp.fuse)
-
   def contiguous_backward(self) -> Tensor:
     """
     Inserts a contiguous operation in the backward pass.
@@ -4002,9 +3802,7 @@ class Tensor(MathMixin, MovementMixin):
       key = key.repeat_interleave(self.shape[-3] // key.shape[-3], dim=-3)
       value = value.repeat_interleave(self.shape[-3] // value.shape[-3], dim=-3)
 
-    if FUSE_ATTENTION: q, key, value = self.contiguous(), key.contiguous(), value.contiguous()
-    else: q = self
-
+    q = self
     qk = q.matmul(key.transpose(-2,-1), dtype=least_upper_dtype(q.dtype, key.dtype, dtypes.float32)) / math.sqrt(q.shape[-1])
     # handle attention mask
     if is_causal:
@@ -4013,8 +3811,7 @@ class Tensor(MathMixin, MovementMixin):
     if attn_mask is not None:
       if attn_mask.dtype == dtypes.bool: attn_mask = attn_mask.where(0, -float("inf"))
       qk = qk + attn_mask
-    attn = qk.cast(self.dtype).softmax(-1).dropout(dropout_p) @ value
-    return attn.fuse() if FUSE_ATTENTION else attn
+    return qk.cast(self.dtype).softmax(-1).dropout(dropout_p) @ value
 
   def _do_reduction(self, reduction:ReductionStr="mean") -> Tensor:
     if reduction not in get_args(ReductionStr): raise ValueError(f"{reduction=} must be one of {get_args(ReductionStr)}")
@@ -4209,29 +4006,6 @@ class Tensor(MathMixin, MovementMixin):
 
   # ***** Tensor Properties *****
 
-  @property
-  def ndim(self) -> int:
-    """
-    Returns the number of dimensions in the tensor.
-
-    ```python exec="true" source="above" session="tensor" result="python"
-    t = Tensor([[1, 2], [3, 4]])
-    print(t.ndim)
-    ```
-    """
-    return len(self.shape)
-
-  def numel(self) -> sint:
-    """
-    Returns the total number of elements in the tensor.
-
-    ```python exec="true" source="above" session="tensor" result="python"
-    t = Tensor([[[1, 2], [3, 4]], [[5, 6], [7, 8]]])
-    print(t.numel())
-    ```
-    """
-    return prod(self.shape)
-
   def element_size(self) -> int:
     """
     Returns the size in bytes of an individual element in the tensor.

tinygrad/uop/__init__.py

@@ -9,6 +9,22 @@ class FastEnum(IntEnum):
 
 # the order of these Ops controls the order of the toposort
 class Ops(FastEnum):
+  # ** 1 -- defines/consts **
+
+  # TODO: unify these ops into the levels of the memory hierarchy. depends on ASSIGN is STORE
+  DEFINE_GLOBAL = auto(); DEFINE_LOCAL = auto(); DEFINE_REG = auto() # noqa: E702
+
+  # this is for symbolic shapes
+  DEFINE_VAR = auto(); BIND = auto() # noqa: E702
+
+  # consts. VCONST is a vectorized const
+  VCONST = auto(); CONST = auto() # noqa: E702
+
+  # this is a RANGE for GPU dimensions, similar to symbolic shapes but not exactly
+  SPECIAL = auto()
+
+  # ** 2 -- non op uops **
+
   # uops that aren't rendered
   NOOP = auto(); SINK = auto(); UNIQUE = auto(); DEVICE = auto(); KERNEL = auto(); PRECAST = auto(); REWRITE_ERROR = auto()  # noqa: E702
   SENTINEL = auto()
@@ -26,39 +42,34 @@ class Ops(FastEnum):
   BUFFERIZE = auto()
 
   # ops that adjust the behavior of the scheduler
-  CONTIGUOUS = auto(); CONTIGUOUS_BACKWARD = auto(); DETACH = auto(); FUSE = auto() # noqa: E702
+  CONTIGUOUS = auto(); CONTIGUOUS_BACKWARD = auto(); DETACH = auto() # noqa: E702
 
   # movement ops! these only exist in the tensor graph
   RESHAPE = auto(); PERMUTE = auto(); EXPAND = auto(); PAD = auto(); SHRINK = auto(); FLIP = auto() # noqa: E702
   MULTI = auto()  # MULTI is really a movement op
 
-  # TODO: unify these ops into the levels of the memory hierarchy. depends on ASSIGN is STORE
-  DEFINE_GLOBAL = auto(); DEFINE_LOCAL = auto(); DEFINE_REG = auto() # noqa: E702
-
-  # this is for symbolic shapes
-  DEFINE_VAR = auto(); BIND = auto() # noqa: E702
-
-  # this is a RANGE for GPU dimensions, similar to symbolic shapes but not exactly
-  SPECIAL = auto()
-
-  # reduce
+  # reduce (movement)
   REDUCE_AXIS = auto(); REDUCE = auto(); ALLREDUCE = auto() # noqa: E702
 
   # optimization helper ops
   UNROLL = auto(); CONTRACT = auto(); GEP = auto(); VECTORIZE = auto(); CAT = auto(); PTRCAT = auto() # noqa: E702
 
-  # UnaryOps
-  CAST = auto(); BITCAST = auto(); EXP2 = auto(); LOG2 = auto(); SIN = auto(); SQRT = auto(); RECIPROCAL = auto(); NEG = auto(); TRUNC = auto() # noqa: E702
+  # ** 3 -- load/store **
+
+  # INDEX is a BinaryOp similar to ADD, but it operates on pointers
+  INDEX = auto()
 
   # load/store before math
   LOAD = auto(); STORE = auto() # noqa: E702
   ASSIGN = auto()  # TODO: ASSIGN is STORE, remove ASSIGN
 
+  # ** 4 -- math **
+
   # tensor core math op, not elementwise
   WMMA = auto()
 
-  # INDEX is a BinaryOp similar to ADD, but it operates on pointers
-  INDEX = auto()
+  # UnaryOps
+  CAST = auto(); BITCAST = auto(); EXP2 = auto(); LOG2 = auto(); SIN = auto(); SQRT = auto(); RECIPROCAL = auto(); NEG = auto(); TRUNC = auto() # noqa: E702
 
   # BinaryOps
   ADD = auto(); MUL = auto(); SHL = auto(); SHR = auto(); IDIV = auto(); MAX = auto(); MOD = auto() # noqa: E702
@@ -69,12 +80,11 @@ class Ops(FastEnum):
   # TernaryOps
   WHERE = auto(); MULACC = auto() # noqa: E702
 
+  # ** 5 -- control flow / other **
+
   # control flow ops
   BARRIER = auto(); RANGE = auto(); IF = auto(); END = auto(); ENDIF = auto() # noqa: E702
 
-  # consts. VCONST is a vectorized const
-  VCONST = auto(); CONST = auto() # noqa: E702
-
   # CUSTOM/CUSTOMI are used to output strings into codegen. the I makes the string inline
   CUSTOM = auto(); CUSTOMI = auto() # noqa: E702
 

tinygrad/uop/decompositions.py

@@ -282,7 +282,7 @@ def magicgu(vmax:int, d:int) -> tuple[int,int]:
 
 def fast_idiv(device: str, x: UOp, d: int, dont_cast=False) -> UOp|None:
   # If d is a power of two this is not valid for signed ints!
-  is_unsigned = True if x.vmin>=0 or x.dtype in dtypes.uints else False
+  is_unsigned = x.vmin>=0 or x.dtype in dtypes.uints
   assert d>0, "Sign should have been taken out of divisor"
   vmin,vmax = max(x.vmin, x.dtype.min), min(x.vmax, x.dtype.max)
   m,s = magicgu(max(vmax, abs(vmin)), d)
@@ -293,7 +293,7 @@ def fast_idiv(device: str, x: UOp, d: int, dont_cast=False) -> UOp|None:
     if (ret:=fast_idiv(device, x//largest_factor_of_two_in_d, d//largest_factor_of_two_in_d, dont_cast=True)) is not None: return ret
   if dont_cast: return None
   # promo_lattice needs to return an unsigned type if the type is unsigned
-  if dtypes.is_int(next_dtype := promo_lattice[x.dtype.scalar()][-1]) and is_dtype_supported(next_dtype, None if device=='' else device):
+  if dtypes.is_int(next_dtype := promo_lattice[x.dtype.scalar()][-1]) and is_dtype_supported(next_dtype, device):
     if m*vmin >= dtypes.min(next_dtype) and m*vmax <= dtypes.max(next_dtype):
       return ((x.cast(next_dtype)*m) >> s).cast(x.dtype) if is_unsigned else ((x.cast(next_dtype)*m) >> s).cast(x.dtype) + (x<0).where(x.ufix(1), 0)
   return None
@@ -318,7 +318,7 @@ def threefry2x32(x: UOp, key: UOp):
 
 powers_of_two = {2**i:i for i in range(64)}
 @functools.cache
-def get_late_rewrite_patterns(ops:tuple[Ops, ...], force_transcendental=False):
+def get_late_rewrite_patterns(ops:tuple[Ops, ...], force_transcendental):
   pat: list[tuple[UPat, Callable]] = []
   for op,f in ((Ops.EXP2, xexp2), (Ops.LOG2, xlog2), (Ops.SIN, xsin)):
     if op not in ops or force_transcendental:

tinygrad/uop/ops.py

@@ -4,11 +4,11 @@ import sys, time, functools, itertools, math, operator, hashlib, os, types, pick
 from dataclasses import dataclass
 from enum import Enum, auto
 from tinygrad.uop import Ops, GroupOp
-from tinygrad.uop.mixins import MathMixin, MovementMixin
+from tinygrad.mixin import OpMixin
 from tinygrad.dtype import ConstType, ImageDType, dtypes, DType, truncate, PtrDType, least_upper_dtype, Invalid, InvalidType, AddrSpace
 from tinygrad.helpers import ContextVar, all_int, prod, getenv, all_same, Context, partition, temp, unwrap, T, argfix, Metadata, flatten, TRACEMETA
 from tinygrad.helpers import PICKLE_BUFFERS, PROFILE, dedup, cdiv, cmod, diskcache_put, to_function_name, cpu_profile, TracingKey, VIZ, SPEC, CI
-from tinygrad.helpers import strip_parens, colored
+from tinygrad.helpers import strip_parens, colored, ansilen
 if TYPE_CHECKING:
   from tinygrad.device import Buffer, MultiBuffer
 
@@ -104,7 +104,7 @@ class recursive_property(property):
 
 # NOTE: this should be frozen, but frozen is slower
 @dataclass(eq=False, slots=True)
-class UOp(MathMixin, MovementMixin, metaclass=UOpMetaClass):
+class UOp(OpMixin, metaclass=UOpMetaClass):
   op:Ops
   dtype:DType = dtypes.void
   src:tuple[UOp, ...] = tuple()
@@ -187,10 +187,18 @@ class UOp(MathMixin, MovementMixin, metaclass=UOpMetaClass):
   def _shape(self) -> tuple[sint, ...]|None:
     match self.op:
       # late ops don't have shape
-      case Ops.UNIQUE | Ops.DEVICE | Ops.RANGE | Ops.INDEX | Ops.LOAD | Ops.IF | Ops.BARRIER | Ops.CUSTOM | Ops.CUSTOMI | \
+      case Ops.UNIQUE | 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.PRECAST | Ops.CONTRACT:
         return None
 
+      case Ops.INDEX:
+        # non pointer index doesn't have a shape
+        if not isinstance(self.dtype, PtrDType): return None
+        # fully indexed doesn't have a shape. TODO: remove this
+        if self.src[0]._shape is None or len(self.src[1:]) == len(self.src[0].shape): return None
+        # pointer index
+        return self.src[0].shape[len(self.src[1:]):]
+
       # some ops init the shape
       case Ops.CONST | Ops.DEFINE_VAR | Ops.BIND: return () if self._device is not None else None
       case Ops.BUFFER: return (self.arg,)
@@ -199,7 +207,7 @@ class UOp(MathMixin, MovementMixin, 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.FUSE | Ops.AFTER | Ops.END:
+      case Ops.REDUCE | Ops.MSTACK | Ops.MSELECT | Ops.DETACH | Ops.CONTIGUOUS | Ops.CONTIGUOUS_BACKWARD | Ops.AFTER | Ops.END:
         return self.src[0]._shape
 
       # ops with custom handling
@@ -287,14 +295,20 @@ class UOp(MathMixin, MovementMixin, metaclass=UOpMetaClass):
   def _ranges(self) -> dict[UOp, None]:
     ret: dict[UOp, None] = {}
     for s in self.src: ret.update(s.ranges)
-    if (er:=self.ended_ranges):
-      for s in UOp.sink(*er).ranges:
-        if s in ret: del ret[s]
+    for er in self.ended_ranges:
+      if er.op is Ops.RANGE:
+        # if it's a single RANGE, we don't flow through it.
+        if er in ret: del ret[er]
+      else:
+        # if it's not a RANGE, we include all ranges in srcs.
+        # technically we shouldn't flow through these ranges either, but this is pre pm_add_control_flow so it's the same.
+        for s in er.ranges:
+          if s in ret: del ret[s]
     return ret
 
   @property
   def ranges(self) -> dict[UOp, None]:
-    if self.op is Ops.RANGE: return {self:None}
+    if self.op is Ops.RANGE: return {self:None} | self._ranges
     return self._ranges
 
   # *** uop evaluation ***
@@ -344,7 +358,13 @@ class UOp(MathMixin, MovementMixin, metaclass=UOpMetaClass):
   def index(self, *srcs:UOp|None, ptr=False, **kwargs):
     return UOp(Ops.INDEX, kwargs.pop("dtype", self.dtype if ptr else self.dtype.base), (self,)+tuple([x for x in srcs if x is not None]), **kwargs)
   def __getitem__(self, idx):
-    return self.index(*[UOp.const(dtypes.index, x) if isinstance(x, int) else x for x in argfix(idx)])
+    idx = argfix(idx)
+    assert len(idx) == len(self.shape), f"__getitem__ shape mismatch, indexing {self.shape} with {len(idx)} args"
+    if len(slice_idx:=[i for i,x in enumerate(idx) if isinstance(x, slice)]):
+      perm = self.permute(tuple([i for i in range(self.ndim) if i not in slice_idx] + slice_idx))
+      return perm.index(*[UOp.const(dtypes.index, x) if isinstance(x, int) else x for x in idx if not isinstance(x, slice)], ptr=True)
+    else:
+      return self.index(*[UOp.const(dtypes.index, x) if isinstance(x, int) else x for x in idx])
   def const_like(self, b:ConstLike):
     # constants can optionally have a DEVICE source
     return UOp.const(self.dtype, b, device=self._device, shape=self._shape)
@@ -368,7 +388,8 @@ class UOp(MathMixin, MovementMixin, metaclass=UOpMetaClass):
       i = (i,)
     return UOp(Ops.GEP, self.dtype.scalar().vec(len(i)) if len(i) > 1 else self.dtype.scalar(), (self,), i)
   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, **kwargs): return UOp(Ops.STORE, kwargs.pop("dtype", dtypes.void), (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)
@@ -425,7 +446,6 @@ class UOp(MathMixin, MovementMixin, metaclass=UOpMetaClass):
     return UOp(Ops.CONTIGUOUS, dtype=self.dtype, src=(self,)+args, **kwargs)
   def contiguous_backward(self): return self.alu(Ops.CONTIGUOUS_BACKWARD)
   def bufferize(self, *args, **kwargs): return UOp(Ops.BUFFERIZE, dtype=self.dtype, src=(self,)+args, **kwargs)
-  def fuse(self): return self.alu(Ops.FUSE)
   def allreduce(self, op, device:str|tuple[str, ...]|UOp):
     assert isinstance(self.device, tuple), f"allreduce must be on tuple {self.device} isn't"
     return UOp(Ops.ALLREDUCE, self.dtype, (self, UOp(Ops.DEVICE, arg=device) if not isinstance(device, UOp) else device), op)
@@ -535,12 +555,12 @@ class UOp(MathMixin, MovementMixin, metaclass=UOpMetaClass):
   def forced_reshape(self, arg:tuple[sint, ...]): return self._mop(Ops.RESHAPE, arg, same_shape_noop=False)
   #def reshape(self, arg:tuple[sint, ...]): return self._mop(Ops.RESHAPE, arg, same_shape_noop=True)
   def expand(self, arg:tuple[sint, ...]): return self._mop(Ops.EXPAND, arg, same_shape_noop=True)
-  def shrink(self, arg:tuple[tuple[sint, sint], ...]): return self._mop(Ops.SHRINK, arg, same_shape_noop=True)
+  #def shrink(self, arg:tuple[tuple[sint, sint], ...]): return self._mop(Ops.SHRINK, arg, same_shape_noop=True)
   def pad(self, arg:tuple[tuple[sint, sint], ...]): return self._mop(Ops.PAD, arg, same_shape_noop=True)
 
   # in these two, we have custom logic to check if they are a no-op
-  def permute(self, arg:tuple[int, ...]): return self._mop(Ops.PERMUTE, arg, same_shape_noop=False) if arg != tuple(range(len(self.shape))) else self
-  def flip(self, arg:tuple[bool, ...]): return self._mop(Ops.FLIP, arg, same_shape_noop=False) if any(arg) and len(arg) == len(self.shape) else self
+  #def permute(self, arg:tuple[int, ...]): return self._mop(Ops.PERMUTE, arg, same_shape_noop=False) if arg != tuple(range(len(self.shape))) else self
+  #def flip(self, arg:tuple[bool, ...]): return self._mop(Ops.FLIP, arg, same_shape_noop=False) if any(arg) and len(arg) == len(self.shape) else self
 
   # *** uop UNIQUE ***
 
@@ -777,8 +797,8 @@ class UOp(MathMixin, MovementMixin, metaclass=UOpMetaClass):
     return UOp.placeholder(self.shape, self.dtype, slot)
 
   # set is store+end+after
-  def set(self:UOp, val:UOp|ConstType, end:UOp|tuple[UOp, ...]=()) -> UOp:
-    return self.src[0].after(self.store(UOp.const(self.dtype, val) if not isinstance(val, UOp) else val).end(*argfix(end)))
+  def set(self:UOp, val:UOp|ConstType, end:UOp|tuple[UOp, ...]|list[UOp]=()) -> UOp:
+    return self.src[0].after(self.store(val).end(*argfix(end)))
 
   def custom_kernel(*srcs:UOp, fxn:Callable, grad_fxn:Callable|None=None) -> list[UOp]:
     placeholders = [UOp.placeholder_like(s, slot=i) for i,s in enumerate(srcs)]
@@ -833,7 +853,8 @@ def exec_alu(op:Ops, dtype:DType, operands, truncate_output=True):
 def print_uops(uops:list[UOp]):
   for i,u in enumerate(uops):
     formatted_srcs = [(uops.index(x) if x.op is not Ops.CONST else f"{x.arg}") if x in uops else "--" for x in u.src]
-    print(f"{i:4d} {str(u.op):20s}: {str(u.dtype):40s} " f"{str(formatted_srcs):32s} {u.arg}")
+    formatted_range = ','.join([range_str(r, color=True) for r in sorted(u.ranges, key=lambda x: x.arg)])
+    print(f"{i:4d} {str(u.op):20s}: {(formatted_range)+' '*(10-ansilen(formatted_range))} {str(u.dtype):40s} " f"{str(formatted_srcs):32s} {u.arg}")
 
 # ***** pattern matcher *****
 
@@ -853,7 +874,7 @@ def printable(loc:tuple[str, int]) -> str:
   try: return lines(loc[0])[loc[1]-1].strip()
   except FileNotFoundError: return "<missing>"
 
-class UPat(MathMixin, MovementMixin):
+class UPat(OpMixin):
   __slots__ = ("op", "dtype", "arg", "name", "src")
   def __init__(self, op:Ops|tuple[Ops, ...]|set[Ops]|None=None, dtype:DType|tuple[DType, ...]|None=None,
                src:tuple[UPat, ...]|list[UPat]|UPat|None=None, arg:Any=None,
@@ -916,7 +937,6 @@ class UPat(MathMixin, MovementMixin):
   def store(self, *src:UPat, **kwargs): return UPat(Ops.STORE, self.dtype, (self,)+src, **kwargs)
   def assign(self, x:UPat, **kwargs): return UPat(Ops.ASSIGN, self.dtype, (self,x), **kwargs)
   def reduce(self, *src:UPat, **kwargs): return UPat(Ops.REDUCE, self.dtype, src=(self,)+src, **kwargs)
-  def fuse(self): return self.alu(Ops.FUSE)
   def broadcast(self, **kwargs): return UPat(Ops.VECTORIZE, self.dtype, src=self, **kwargs)
   def contiguous(self, *args, **kwargs): return UPat(Ops.CONTIGUOUS, dtype=self.dtype, src=(self,)+args, **kwargs)
   def after(self, *src:UPat, **kwargs): return UPat(Ops.AFTER, self.dtype, (self,)+src, **kwargs)
@@ -1309,7 +1329,8 @@ renderer_infer = PatternMatcher([
 
 def srcs(ctx, src): return f"({ctx[src[0]]},)" if len(src) == 1 else f"({', '.join([ctx[x] for x in src])})"
 def render_marg(ctx,x:UOp):
-  if x.op in {Ops.PERMUTE, Ops.FLIP}: return str(x.marg)
+  if x.op is Ops.PERMUTE: return str(x.marg)
+  if x.op is Ops.FLIP: return str(tuple([i for i,x in enumerate(x.marg) if x]))
   pieces = []
   if x.op in {Ops.RESHAPE, Ops.EXPAND}:
     pieces = [f"{ctx[a] if isinstance(a, UOp) else str(a)}" for a in x.marg]

tinygrad/uop/spec.py

@@ -89,7 +89,7 @@ _tensor_spec = PatternMatcher([
 
   # DETACH and CONTIGUOUS change how we interpret the source UOp
   # CONTIGUOUS ensures the source UOp realizes
-  (UPat((Ops.DETACH, Ops.CONTIGUOUS, Ops.CONTIGUOUS_BACKWARD, Ops.FUSE), name="root", src=(UPat.var("x"),), arg=None),
+  (UPat((Ops.DETACH, Ops.CONTIGUOUS, Ops.CONTIGUOUS_BACKWARD), name="root", src=(UPat.var("x"),), arg=None),
    lambda root,x: root.dtype == x.dtype),
 
   # CONTIGUOUS with a range
@@ -122,7 +122,7 @@ shared_codegen_spec = PatternMatcher([
   # DEFINEs
   (UPat(Ops.DEFINE_GLOBAL, name="x"), lambda x: isinstance(x.dtype, (PtrDType, ImageDType)) and x.dtype.addrspace == AddrSpace.GLOBAL),
   (UPat(Ops.DEFINE_LOCAL, name="x"), lambda x: isinstance(x.dtype, PtrDType) and x.dtype.addrspace == AddrSpace.LOCAL),
-  (UPat(Ops.DEFINE_REG, src=()), lambda: True),
+  (UPat(Ops.DEFINE_REG, src=(), name="x"), lambda x: isinstance(x.arg, int)),
 
   # allow AFTER on buffers, GROUP anywhere
   (UPat(Ops.AFTER, src=(UPat(GroupOp.Defines|{Ops.AFTER}),), allow_any_len=True), lambda: True),

tinygrad/uop/symbolic.py

@@ -382,13 +382,6 @@ symbolic = symbolic_simple+commutative+PatternMatcher([
   (UPat(Ops.VECTORIZE, src=UPat(Ops.CONST), name="vec"), lambda vec: UOp.const(vec.dtype, tuple(x.arg for x in vec.src))),
 ])+gep_pushing
 
-symbolic_flat = symbolic+PatternMatcher([
-  # ** combine terms (opinionated) **
-  (-1 * (UPat.var("x") + UPat.var("y")), lambda x,y: (-x)+(-y)),  # -(x+y) -> -x + -y
-  # (x+y)*c -> x*c+y*c. only for int, float has inf*0=nan issue
-  ((UPat.var("x", dtypes.index) + UPat.var("y")) * UPat.cvar("c"), lambda x,y,c: x*c+y*c),
-])
-
 # ******** we take a small aside to "simplify_valid" to rewrite valids ********
 
 def parse_valid(valid:UOp) -> tuple[UOp, bool, int]|None:
@@ -503,7 +496,7 @@ pm_simplify_valid = PatternMatcher([
 
 # this is symbolic 2.0
 REMOVE_FROM_SINK_LIKE = {Ops.UNROLL, Ops.NOOP, Ops.VECTORIZE, Ops.SINK}
-sym = symbolic_flat+pm_simplify_valid+PatternMatcher([
+sym = symbolic+pm_simplify_valid+PatternMatcher([
   # LOAD/STORE -> NOOP
   (UPat.var('x').store(UPat.var('x').load(), allow_any_len=True), lambda x: None if x.dtype.addrspace != AddrSpace.REG else x.src[0].src[0]),
   (UPat(Ops.LOAD, src=(UPat.cvar('c'))), lambda c: c),
@@ -553,4 +546,8 @@ sym = symbolic_flat+pm_simplify_valid+PatternMatcher([
       if any(x.op in REMOVE_FROM_SINK_LIKE for x in root.src) else None),
   # remove END with empty NOOP
   (UPat(Ops.END, src=(UPat(Ops.NOOP, src=(), name="noop"),), allow_any_len=True), lambda noop:noop),
+  # ** combine terms (opinionated) **
+  (-1 * (UPat.var("x") + UPat.var("y")), lambda x,y: (-x)+(-y)),  # -(x+y) -> -x + -y
+  # (x+y)*c -> x*c+y*c. only for int, float has inf*0=nan issue
+  ((UPat.var("x", dtypes.index) + UPat.var("y")) * UPat.cvar("c"), lambda x,y,c: x*c+y*c),
 ])

tinygrad/viz/index.html

@@ -284,12 +284,15 @@
     height: fit-content;
   }
   .raw-text {
-    padding: 0 8px;
+    padding-left: 15px;
     width: 100%;
     height: 100%;
     max-height: 100vh;
     overflow-x: auto;
   }
+  .raw-text > pre {
+    display: inline-block;
+  }
   .raw-text code {
     max-height: none !important;
   }

tinygrad/viz/js/index.js

@@ -538,7 +538,9 @@ document.getElementById("zoom-to-fit-btn").addEventListener("click", () => {
 
 function codeBlock(st, language, { loc, wrap }={}) {
   const code = document.createElement("code");
-  code.innerHTML = hljs.highlight(st, { language }).value;
+  // plaintext renders like a terminal print, otherwise render with syntax highlighting
+  if (language === "txt") code.appendChild(colored(st));
+  else code.innerHTML = hljs.highlight(st, { language }).value;
   code.className = "hljs";
   const ret = document.createElement("pre");
   if (wrap) ret.className = "wrap";
@@ -717,7 +719,7 @@ async function main() {
         const div = d3.create("div").style("background", cycleColors(colorScheme.CATEGORICAL, s.idx)).style("width", "24px").style("height", "100%");
         return [s.label.trim(), div.node()];
       })).node());
-    } else root.appendChild(codeBlock(ret.src, ret.lang));
+    } else root.appendChild(codeBlock(ret.src, ret.lang || "txt"));
     return document.querySelector("#custom").replaceChildren(root);
   }
   // ** UOp view (default)

tinygrad/viz/js/worker.js

@@ -29,10 +29,10 @@ onmessage = (e) => {
       const node = g.node(n);
       if (node.label.includes("dtypes.index")) g.removeNode(n);
     }
-    // After all layout changes are complete, remove the overlay node if it's empty
-    if (!g.node("addition")?.width) g.removeNode("addition");
   }
   dagre.layout(g);
+  // remove additions overlay if it's empty
+  if (!g.node("addition")?.width) g.removeNode("addition");
   postMessage(dagre.graphlib.json.write(g));
   self.close();
 }

tinygrad/viz/serve.py

@@ -1,6 +1,6 @@
 #!/usr/bin/env python3
 import multiprocessing, pickle, difflib, os, threading, json, time, sys, webbrowser, socket, argparse, socketserver, functools, codecs, io, struct
-import subprocess, ctypes, pathlib
+import subprocess, ctypes, pathlib, traceback
 from contextlib import redirect_stdout
 from decimal import Decimal
 from http.server import BaseHTTPRequestHandler
@@ -18,7 +18,7 @@ uops_colors = {Ops.LOAD: "#ffc0c0", Ops.STORE: "#87CEEB", Ops.CONST: "#e0e0e0",
                Ops.RANGE: "#c8a0e0", Ops.ASSIGN: "#909090", Ops.BARRIER: "#ff8080", Ops.IF: "#c8b0c0", Ops.SPECIAL: "#c0c0ff",
                Ops.INDEX: "#cef263", Ops.WMMA: "#efefc0", Ops.MULTI: "#f6ccff", Ops.KERNEL: "#3e7f55",
                **{x:"#D8F9E4" for x in GroupOp.Movement}, **{x:"#ffffc0" for x in GroupOp.ALU}, Ops.THREEFRY:"#ffff80",
-               Ops.BUFFER_VIEW: "#E5EAFF", Ops.BUFFER: "#B0BDFF", Ops.COPY: "#a040a0", Ops.FUSE: "#FFa500",
+               Ops.BUFFER_VIEW: "#E5EAFF", Ops.BUFFER: "#B0BDFF", Ops.COPY: "#a040a0",
                Ops.ALLREDUCE: "#ff40a0", Ops.MSELECT: "#d040a0", Ops.MSTACK: "#d040a0", Ops.CONTIGUOUS: "#FFC14D",
                Ops.BUFFERIZE: "#FF991C", Ops.REWRITE_ERROR: "#ff2e2e", Ops.AFTER: "#8A7866", Ops.END: "#524C46"}
 
@@ -193,10 +193,34 @@ def mem_layout(dev_events:list[tuple[int, int, float, DevEvent]], start_ts:int,
   peaks.append(peak)
   return struct.pack("<BIQ", 1, len(events), peak)+b"".join(events) if events else None
 
+def load_sqtt(profile:list[ProfileEvent]) -> None:
+  from tinygrad.runtime.ops_amd import ProfileSQTTEvent
+  if not (sqtt_events:=[e for e in profile if isinstance(e, ProfileSQTTEvent)]): return None
+  def err(name:str, msg:str|None=None) -> None:
+    step = {"name":name, "data":{"src":msg or traceback.format_exc()}, "depth":0, "query":f"/render?ctx={len(ctxs)}&step=0&fmt=counters"}
+    return ctxs.append({"name":"Counters", "steps":[step]})
+  try: from extra.sqtt.roc import decode
+  except Exception: return err("DECODER IMPORT ISSUE")
+  try:
+    rctx = decode(profile)
+    steps = [{"name":str(x[0]), "depth":0, "data":{"rows":[(e.inst, e.time, e.time-x[1][i-1].time if i else 0, e.dur, e.stall,
+                                                            str(e.typ).split("_")[-1]) for i,e in enumerate(x[1])],
+                                              "cols":["Instruction", "Clk", "Wait", "Duration", "Stall", "Type"], "summary":[]},
+              "query":f"/render?ctx={len(ctxs)}&step={i}&fmt=counters"} for i,x in enumerate(rctx.inst_execs.items())]
+    if not steps: return err("EMPTY SQTT OUTPUT", f"{len(sqtt_events)} SQTT events recorded, none got decoded")
+  except Exception: return err("DECODER ERROR")
+  ctxs.append({"name":"Counters", "steps":steps})
+
 def get_profile(profile:list[ProfileEvent]) -> bytes|None:
   # start by getting the time diffs
   for ev in profile:
     if isinstance(ev,ProfileDeviceEvent): device_ts_diffs[ev.device] = (ev.comp_tdiff, ev.copy_tdiff if ev.copy_tdiff is not None else ev.comp_tdiff)
+  # load device specific counters
+  device_decoders:dict[str, Callable[[list[ProfileEvent]], None]] = {}
+  for device in device_ts_diffs:
+    d = device.split(":")[0]
+    if d == "AMD": device_decoders[d] = load_sqtt
+  for fxn in device_decoders.values(): fxn(profile)
   # map events per device
   dev_events:dict[str, list[tuple[int, int, float, DevEvent]]] = {}
   markers:list[ProfilePointEvent] = []
@@ -248,9 +272,10 @@ def get_stdout(f:Callable) -> str:
   with redirect_stdout(buf:=io.StringIO()): f()
   return buf.getvalue()
 
-def get_render(i:int, fmt:str) -> dict|None:
+def get_render(i:int, j:int, fmt:str) -> dict|None:
+  if fmt == "counters": return ctxs[i]["steps"][j]["data"]
   if not isinstance(prg:=trace.keys[i].ret, ProgramSpec): return None
-  if fmt == "uops": return {"src":get_stdout(lambda: print_uops(prg.uops or [])), "lang":"python"}
+  if fmt == "uops": return {"src":get_stdout(lambda: print_uops(prg.uops or [])), "lang":"txt"}
   if fmt == "src": return {"src":prg.src, "lang":"cpp"}
   lib = (compiler:=Device[prg.device].compiler).compile(prg.src)
   disasm_str = get_stdout(lambda: compiler.disassemble(lib))
@@ -264,7 +289,7 @@ def get_render(i:int, fmt:str) -> dict|None:
 
 # ** HTTP server
 
-def get_int(query:dict[str, list[str]], k:str) -> int: return int(query[k][0])
+def get_int(query:dict[str, list[str]], k:str) -> int: return int(query.get(k,["0"])[0])
 
 class Handler(BaseHTTPRequestHandler):
   def do_GET(self):
@@ -279,7 +304,9 @@ class Handler(BaseHTTPRequestHandler):
         if url.path.endswith(".css"): content_type = "text/css"
       except FileNotFoundError: status_code = 404
     elif (query:=parse_qs(url.query)):
-      if url.path == "/render": ret, content_type = json.dumps(get_render(get_int(query, "ctx"), query["fmt"][0])).encode(), "application/json"
+      if url.path == "/render":
+        render_src = get_render(get_int(query, "ctx"), get_int(query, "step"), query["fmt"][0])
+        ret, content_type = json.dumps(render_src).encode(), "application/json"
       else:
         try: return self.stream_json(get_full_rewrite(trace.rewrites[i:=get_int(query, "ctx")][get_int(query, "idx")], i))
         except (KeyError, IndexError): status_code = 404