simplify priority

* amd: add kern to sqtt event

* fix

.github/workflows/autogen.yml

@@ -2,7 +2,7 @@ name: Autogen
 env:
   # increment this when downloads substantially change to avoid the internet
   DOWNLOAD_CACHE_VERSION: '12'
-  PYTHON_CACHE_VERSION: '3'
+  PYTHON_CACHE_VERSION: '4'
   APT_CACHE_VERSION: '1'
   BUILD_CACHE_VERSION: '1'
   CAPTURE_PROCESS_REPLAY: 1

.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
@@ -211,6 +211,7 @@ jobs:
         CUDA=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul.txt
         CUDA=1 SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_bfloat16.txt
         CUDA=1 SHOULD_USE_TC=1 ALLOW_TF32=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py | tee matmul_tf32.txt
+        CUDA=1 SHOULD_USE_TC=1 FP8E4M3=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_fp8.txt
     - name: Run Tensor Core GEMM (PTX)
       run: NV=1 NV_PTX=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_ptx.txt
     - name: Run Tensor Core GEMM (NV)
@@ -408,7 +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
@@ -526,7 +527,7 @@ jobs:
     - name: Run 10 CIFAR training steps
       run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=330 AMD=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
     - name: Run 10 CIFAR training steps w HALF
-      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=330 AMD=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_half.txt
+      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=350 AMD=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_half.txt
     # - name: Run 10 CIFAR training steps w BF16
     #   run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=288 AMD=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
     # TODO: too slow

.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"

.github/workflows/test.yml

@@ -2,7 +2,7 @@ name: Unit Tests
 env:
   # increment this when downloads substantially change to avoid the internet
   DOWNLOAD_CACHE_VERSION: '12'
-  PYTHON_CACHE_VERSION: '3'
+  PYTHON_CACHE_VERSION: '4'
   APT_CACHE_VERSION: '1'
   BUILD_CACHE_VERSION: '1'
   CAPTURE_PROCESS_REPLAY: 1
@@ -230,7 +230,7 @@ jobs:
       uses: ./.github/actions/setup-tinygrad
       with:
         key: linting-only
-        python-version: '3.10'
+        python-version: '3.11'
         deps: linting
     - name: Lint bad-indentation and trailing-whitespace with pylint
       run: python -m pylint --disable=all -e W0311 -e C0303 --jobs=0 --indent-string='  ' --recursive=y .
@@ -243,8 +243,9 @@ jobs:
       run: |
         python -m mypy --strict-equality --lineprecision-report .
         cat lineprecision.txt
-    - name: Run TYPED=1
-      run: TYPED=1 python -c "import tinygrad"
+    # broken because of UPatAny
+    #- name: Run TYPED=1
+    #  run: TYPED=1 python -c "import tinygrad"
 
   unittest:
     name: Unit Tests
@@ -294,6 +295,7 @@ jobs:
 
   spec:
     strategy:
+      fail-fast: false
       matrix:
         group: [1, 2]
     name: SPEC=2 (${{ matrix.group }})
@@ -308,7 +310,7 @@ jobs:
         key: spec-unit
         deps: testing_unit
     - name: Test SPEC=2
-      run: IGNORE_OOB=0 SPEC=2 PYTHONPATH="." pytest --maxfail=10 -n auto --durations=30 --ignore=test/models --ignore test/unit/test_hashing.py --timeout 40 -k "not test_setitem_big" --splits 2 --group ${{ matrix.group }}
+      run: IGNORE_OOB=0 SPEC=2 PYTHONPATH="." pytest --maxfail=10 -n auto --durations=30 --ignore=test/models --ignore test/unit/test_hashing.py --timeout 60 -k "not test_setitem_big" --splits 2 --group ${{ matrix.group }}
 
   fuzzing:
     name: Fuzzing
@@ -643,6 +645,7 @@ jobs:
         run: TRANSCENDENTAL=2 python -m pytest -n=auto test/test_ops.py::TestOps::test_sin test/test_ops.py::TestOps::test_cos test/test_ops.py::TestOps::test_tan test/test_ops.py::TestOps::test_exp test/test_ops.py::TestOps::test_log --durations=20
       - name: Run TestOps.test_add with SQTT
         run: |
+          VIZ=1 PMC=1 DEBUG=5 python3 test/test_ops.py TestOps.test_add
           VIZ=1 SQTT=1 DEBUG=5 python3 test/test_ops.py TestOps.test_add
           extra/sqtt/rgptool.py create "/tmp/profile.pkl.$USER" -o /tmp/gpu0.rgp
       - name: Run process replay tests

.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() {
@@ -531,7 +535,7 @@ generate_mesa() {
   sed -i "s/('fp_fast_math', ctypes.c_bool, 9)/('fp_fast_math', ctypes.c_uint32, 9)/" $BASE/mesa.py
   sed -i "s/('\(\w\+\)', pipe_shader_type, 8)/('\1', ctypes.c_ubyte)/" $BASE/mesa.py
   sed -i "s/\([0-9]\+\)()/\1/" $BASE/mesa.py
-  sed -i "s/\(struct_nir_builder._pack_\) = 1/\1 = 0/" $BASE/mesa.py
+  sed -i '/struct_nir_builder._pack_ = 1 # source:False/d' "$BASE/mesa.py"
   python3 -c "import tinygrad.runtime.autogen.mesa"
 }
 

docs/abstractions2.py

@@ -53,9 +53,7 @@ b = Buffer(DEVICE, 1, dtypes.int32).allocate().copyin(memoryview(bytearray(struc
 idx = UOp.const(dtypes.index, 0)
 buf_1 = UOp(Ops.DEFINE_GLOBAL, dtypes.int32.ptr(), (), 1)
 buf_2 = UOp(Ops.DEFINE_GLOBAL, dtypes.int32.ptr(), (), 2)
-ld_1 = UOp(Ops.LOAD, dtypes.int32, (buf_1.index(idx),))
-ld_2 = UOp(Ops.LOAD, dtypes.int32, (buf_2.index(idx),))
-alu = ld_1 + ld_2
+alu = buf_1.index(idx) + buf_2.index(idx)
 output_buf = UOp(Ops.DEFINE_GLOBAL, dtypes.int32.ptr(), (), 0)
 st_0 = UOp(Ops.STORE, dtypes.void, (output_buf.index(idx), alu))
 s = UOp(Ops.SINK, dtypes.void, (st_0,))

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

@@ -1,353 +1,168 @@
 from tinygrad import Tensor, Device, Context, GlobalCounters, dtypes
-from tinygrad.uop.ops import UOp, Ops, KernelInfo, graph_rewrite, AxisType, PatternMatcher, UPat
-from tinygrad.engine.realize import CompiledRunner, ExecItem, get_program
+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, colored, prod, unwrap
-from tinygrad.shape.shapetracker import ShapeTracker, View
-from tinygrad.shape.view import strides_for_shape
-from tinygrad.codegen.opt.kernel import axis_colors, Opt, OptOps
-from tinygrad.codegen.opt.swizzler import merge_views, view_left
-
-def to_colored(full_shape, axis_types): return '_'.join([colored(str(s), axis_colors[at]) for s,at in zip(full_shape, axis_types)])
+from tinygrad.helpers import getenv
 
 N = 4096
+M = K = N
 run_count = 5
 
-BN = 128
-BM = 128
-BK = 8
+# ---------------------------
+# launch/config constants
+# ---------------------------
 
-TN = 4
-TM = 4
+WARP_SIZE = 32
 
-# NOTE: this is from testgrad
-# change reduceop axes and input ShapeTrackers, view gets replaced with a reshape.
-# src->r->view  -->   src->view->r
-def swizzle_reduceop(src:UOp, r:UOp, view:UOp):
-  if r.tag is not None: return None
-  # confirm the input is in order
-  # TODO: replace this with a UOp that allows for nothing else then remove this
-  permute = tuple(i for i in range(len(src.shape)) if i not in r.axis_arg)+r.axis_arg
-  assert permute == tuple(range(len(permute))), f"reduce axis must already be in order, {permute} isn't"
+# Threadblock tile sizes (block-level tile of C that a block computes)
+BLOCK_N = 128   # columns of C (N-dim) per block
+BLOCK_M = 128   # rows of C (M-dim) per block
+BLOCK_K = 8     # K-slice per block iteration
 
-  # append the reduce shape to each of the views
-  prshape = prod(rshape:=src.shape[-len(r.axis_arg):])
-  rstrides = strides_for_shape(rshape)
-  nv = [View.create(v.shape+rshape, tuple(x*prshape for x in v.strides)+rstrides, v.offset*prshape,
-                    v.mask+tuple((0,s) for s in rshape) if v.mask is not None else None) for v in unwrap(view.st).views]
+# Register tile sizes (per-thread accumulator tile of C)
+TN = 4     # columns per thread
+TM = 4     # rows per thread
 
-  # no reshape required with shrinking REDUCE_AXIS
-  return UOp(Ops.REDUCE_AXIS, r.dtype, (src.view(ShapeTracker(tuple(nv))),),
-             (r.arg[0], tuple(range(len(view.shape), len(view.shape) + len(r.axis_arg)))))
+is_kernel5 = getenv("K5", 0)
+THREADS_PER_BLOCK = 128 if is_kernel5 else 256
+assert THREADS_PER_BLOCK % BLOCK_N == 0, "THREADS_PER_BLOCK must be divisible by BLOCK_N"
+assert THREADS_PER_BLOCK % BLOCK_K == 0, "THREADS_PER_BLOCK must be divisible by BLOCK_K"
+assert (BLOCK_N * BLOCK_K) % THREADS_PER_BLOCK == 0
+assert (BLOCK_M * BLOCK_K) % THREADS_PER_BLOCK == 0
 
-pm = PatternMatcher([
-  (UPat(Ops.VIEW, src=(UPat(Ops.REDUCE_AXIS, src=(UPat.var("src"),), name="r"),), name="view"), swizzle_reduceop),
-])
+WARPS_PER_BLOCK = THREADS_PER_BLOCK // WARP_SIZE
+WAVE_TILE_N = 128 if is_kernel5 else 64
+WAVE_TILE_M = BLOCK_N * BLOCK_M // WARPS_PER_BLOCK // WAVE_TILE_N
+assert BLOCK_N % WAVE_TILE_N == 0, "BN must be a multiple of WN"
+assert BLOCK_M % WAVE_TILE_M == 0, "BM must be a multiple of WM"
+WAVES_IN_BLOCK_X = BLOCK_N // WAVE_TILE_N
+WAVES_IN_BLOCK_Y = BLOCK_M // WAVE_TILE_M
+assert WAVES_IN_BLOCK_X * WAVES_IN_BLOCK_Y == WARPS_PER_BLOCK, "wave grid must match warps/block"
 
-def rangeify_kernel3():
-  a = Tensor.empty(N,N)
-  b = Tensor.empty(N,N)
-  c = a@b
-  #c = c.reshape((32,2,16,4,32,2,16,4)).contiguous()
-  sink = c.schedule()[-1].ast
-  #print(sink)
+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)
+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"
 
-  opts  = [Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.LOCAL, 0, 16), Opt(OptOps.UPCAST, 0, 2)]
-  opts += [Opt(OptOps.UPCAST, 1, 4), Opt(OptOps.LOCAL, 1, 16), Opt(OptOps.UPCAST, 1, 2)]
-  opts += [Opt(OptOps.UNROLL, 0, 8)]
+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
 
-  return sink.replace(arg=KernelInfo(opts_to_apply=tuple(opts)))
+def hand_spec_kernel3():
+  # ---------------------------
+  # block indices & placeholders
+  # ---------------------------
+  blockIdx_x = UOp.special(N // BLOCK_N, "gidx0")
+  blockIdx_y = UOp.special(N // BLOCK_M, "gidx1")
 
-def top_spec_kernel3():
-  a = Tensor.empty(N,N)
-  b = Tensor.empty(N,N)
-  c = a@b
-  sink = c.schedule()[-1].ast
-  L = 16
-  sink = sink.reshape((N//L, L, N//L, L)) #.lift({0:UOp.range(N//BM, 0), 2:UOp.range(N//BN, 1)})
-  sink = graph_rewrite(sink, view_left+pm)
-  axis_types = (AxisType.GLOBAL, AxisType.LOCAL, AxisType.GLOBAL, AxisType.LOCAL, AxisType.REDUCE)
-  return sink.replace(arg=KernelInfo(name="top_"+to_colored(sink.full_shape, axis_types), axis_types=axis_types))
+  a = UOp.placeholder((N, N), dtypes.float, slot=1)
+  b = UOp.placeholder((N, N), dtypes.float, slot=2)
+  c = UOp.placeholder((N, N), dtypes.float, slot=0)
 
-def hl_spec_kernel3():
-  nbIterWaveM = 2
-  nbIterWaveN = 2
+  # index the output with the globals
+  c = c.reshape(M // BLOCK_M, BLOCK_M, N // BLOCK_N, BLOCK_N)[blockIdx_y, :, blockIdx_x, :]
 
-  # define buffers
-  # TODO: remove these views once the defines have a shape
-  a = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(N*N), arg=1).view(ShapeTracker.from_shape((N,N)))
-  b = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(N*N), arg=2).view(ShapeTracker.from_shape((N,N))).permute((1,0))
-  c = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(N*N), arg=0).view(ShapeTracker.from_shape((N,N)))
-  As = UOp(Ops.DEFINE_LOCAL, dtypes.float.ptr(BK*BM, AddrSpace.LOCAL), arg=0).view(ShapeTracker.from_shape((BK, BM))).permute((1,0))
-  Bs = UOp(Ops.DEFINE_LOCAL, dtypes.float.ptr(BK*BN, AddrSpace.LOCAL), arg=1).view(ShapeTracker.from_shape((BK, BN))).permute((1,0))
-  A_col = UOp(Ops.DEFINE_REG, dtypes.float.ptr(nbIterWaveM * TM, AddrSpace.REG), arg=0).view(ShapeTracker.from_shape((nbIterWaveM * TM,)))
-  B_row = UOp(Ops.DEFINE_REG, dtypes.float.ptr(nbIterWaveN * TN, AddrSpace.REG), arg=1).view(ShapeTracker.from_shape((nbIterWaveN * TN,)))
+  # 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, :]
 
-  # shape buffers. TODO: permutes
-  full_shape = (N//BM, nbIterWaveM, BM//(nbIterWaveM * TM), TM, N//BN, nbIterWaveN, BN//(nbIterWaveN * TN), TN, N//BK, BK)
-  a = a.reshape((N//BM, nbIterWaveM, BM//(nbIterWaveM * TM), TM, 1, 1, 1, 1, N//BK, BK)).expand(full_shape)
-  b = b.reshape((1, 1, 1, 1, N//BN, nbIterWaveN, BN//(nbIterWaveN * TN), TN, N//BK, BK)).expand(full_shape)
-  c = c.reshape((N//BM, nbIterWaveM, BM//(nbIterWaveM * TM), TM, N//BN, nbIterWaveN, BN//(nbIterWaveN * TN), TN, 1, 1))
-  As = As.reshape((1, nbIterWaveM, BM//(nbIterWaveM * TM), TM, 1, 1, 1, 1, 1, BK)).expand(full_shape)
-  Bs = Bs.reshape((1, 1, 1, 1, 1, nbIterWaveN, BN//(nbIterWaveN * TN), TN, 1, BK)).expand(full_shape)
-  A_col = A_col.reshape((1, nbIterWaveM, 1, TM, 1, 1, 1, 1, 1, 1)).expand(full_shape)
-  B_row = B_row.reshape((1, 1, 1, 1, 1, nbIterWaveN, 1, TN, 1, 1)).expand(full_shape)
+  # globals are no longer used, they are already in the indexes
+  del blockIdx_y, blockIdx_x
 
-  #                     U1   L2 L3 L4 L5   U6 U7      U9   L10 L11 L12 L13   U14 U15      U17  U18  U19
-  expanded_shape = (32, 2,   2, 2, 2, 2,   2, 2,  32, 2,   2,  2,  2,  2,    2,  2,  512, 2,   2,   2)
-  assert len(expanded_shape) == 20
-  permute_a = list(range(len(expanded_shape)))
-  permute_b = permute_a[:]
+  # ---------------------------
+  # GLOBAL -> LOCAL (As, Bs)
+  # ---------------------------
+  tid = UOp.special(THREADS_PER_BLOCK, "lidx0")
 
-  # this makes all the global loads match
-  # this can also be more simply done by rebinding the RANGEs
-  # but sadly, rebinding the RANGEs doesn't work to change the order of the local axes
-  permute_a[17:20] = [11,12,13]
-  permute_a[11:14] = [17,18,19]
-  permute_a[7], permute_a[10] = permute_a[10], permute_a[7]
-  permute_a[2:7] = [3,4,5,6,2]
+  # 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)
 
-  permute_b[2:16] = [19,9,10,11,17,18,8,2,12,13,14,15,3,4]
-  permute_b[17:20] = [5,6,7]
+  # 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)
 
-  a_permute   = a.reshape(expanded_shape).permute(tuple(permute_a)).reshape(full_shape)
-  As_permute = As.reshape(expanded_shape).permute(tuple(permute_a)).reshape(full_shape)
+  # TODO: can we automate barrier?
+  barrier = UOp.barrier(As_store, Bs_store)
+  As, Bs = As.after(barrier), Bs.after(barrier)
 
-  b_permute   = b.reshape(expanded_shape).permute(tuple(permute_b)).reshape(full_shape)
-  Bs_permute = Bs.reshape(expanded_shape).permute(tuple(permute_b)).reshape(full_shape)
+  # open inner k range
+  k = UOp.range(BLOCK_K, 3, AxisType.REDUCE)
 
-  #out = (a.load() * b.load()).r(Ops.ADD, (8, 9))
-  out = (As.load(As_permute.store(a_permute.load())) * Bs.load(Bs_permute.store(b_permute.load()))).r(Ops.ADD, (8, 9))
-  #out = (A_col.load(A_col.store(As.load(As.store(a.load())))) * B_row.load(B_row.store(Bs.load(Bs.store(b.load()))))).r(Ops.ADD, (8, 9))
+  # ---------------------------
+  # LOCAL -> REG (per-wave tiles)
+  # ---------------------------
+  waveIdx = (tid // WARP_SIZE) % WAVES_IN_BLOCK_X
+  waveIdy = (tid // WARP_SIZE) // WAVES_IN_BLOCK_X
+  assert waveIdy.vmax+1 == WAVES_IN_BLOCK_Y
 
-  axis_types = (
-    AxisType.GLOBAL, AxisType.UPCAST, AxisType.LOCAL, AxisType.UPCAST,
-    AxisType.GLOBAL, AxisType.UPCAST, AxisType.LOCAL, AxisType.UPCAST,
-    AxisType.REDUCE, AxisType.REDUCE)
+  laneIdx = (tid % WARP_SIZE) % LANES_PER_WAVE_X
+  laneIdy = (tid % WARP_SIZE) // LANES_PER_WAVE_X
+  assert laneIdy.vmax+1 == LANES_PER_WAVE_Y
 
-  sink = c.store(out).sink(arg=KernelInfo(name="tg_"+to_colored(full_shape, axis_types), axis_types=axis_types))
-  sink = graph_rewrite(sink, merge_views)
-  return sink
+  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)
 
-def hand_spec_kernel3(kernel4=getenv("K4", 0), kernel5=getenv("K5", 0)):
-  BLOCK_SIZE = 128 if kernel5 else 256
+  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)
 
-  nbWaves = BLOCK_SIZE // 32
-  WN = 128 if kernel5 else 64
-  WM = BN * BM // nbWaves // WN
+  # ---------------------------
+  # FMA: c_regs += A_col * B_row
+  # ---------------------------
+  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))
 
-  nbWaveX = BN // WN
-  nbWaveY = BM // WM
+  # 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)
 
-  threadIdx_x = UOp(Ops.SPECIAL, dtypes.int, arg=("lidx0", BLOCK_SIZE))
-  waveIndex = threadIdx_x // 32
-  waveIdx = waveIndex % nbWaveX
-  waveIdy = waveIndex // nbWaveX
-  indexInWave = threadIdx_x % 32
+  # Close k, sync, and close K tiles
+  sink = sink.end(k).barrier().end(k_tile_range)
 
-  nbThreadXPerWave = 8
-  nbThreadYPerWave = 4
+  # ---------------------------
+  # REG -> GLOBAL (epilogue)
+  # ---------------------------
+  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)
 
-  idxInWave = indexInWave % nbThreadXPerWave
-  idyInWave = indexInWave // nbThreadXPerWave
+  return sink.sink(arg=KernelInfo(opts_to_apply=())).simplify()
 
-  nbIterWaveN = WN // (nbThreadXPerWave * TN)
-  nbIterWaveM = WM // (nbThreadYPerWave * TM)
+def test_matmul(sink:UOp, N=N):
+  with Context(DEBUG=0):
+    a = Tensor.randn(N, N)
+    b = Tensor.randn(N, N)
+    hc = Tensor.empty(N, N)
+    Tensor.realize(a, b, hc)
 
-  SUBWN = WN // nbIterWaveN
-  SUBWM = WM // nbIterWaveM
+  ei = ExecItem(get_runner(Device.DEFAULT, sink), [t.uop.buffer for t in [hc, a, b]])
 
-  # Thread mapping to read BKxBN block from A
-  rAIdx = threadIdx_x % BK
-  rAIdy = threadIdx_x // BK
-  # Thread mapping to read BNxBK block from B
-  rBIdx = threadIdx_x % BN
-  rBIdy = threadIdx_x // BN
+  GlobalCounters.reset()
+  ets = []
+  with Context(DEBUG=2):
+    for _ in range(run_count):
+      ets.append(ei.run(wait=True))
+  print(f"REAL TFLOPS {N * N * N * 2 / min(ets) * 1e-12:.2f}")
 
-  strideReadB = BLOCK_SIZE // BN
-  strideReadA = BLOCK_SIZE // BK
-  nbReadsB = BN * BK // BLOCK_SIZE
-  nbReadsA = BM * BK // BLOCK_SIZE
-
-  blockIdx_x = UOp(Ops.SPECIAL, dtypes.int, arg=("gidx0", N//BN))
-  blockIdx_y = UOp(Ops.SPECIAL, dtypes.int, arg=("gidx1", N//BM))
-
-  a = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(N*N), arg=1)
-  b = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(N*N), arg=2)
-  c = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(N*N), arg=0)
-
-  A_col = UOp(Ops.DEFINE_REG, dtypes.float.ptr(nbIterWaveM * TM, AddrSpace.REG), arg=0)
-  B_row = UOp(Ops.DEFINE_REG, dtypes.float.ptr(nbIterWaveN * TN, AddrSpace.REG), arg=1)
-
-  BM_As_stride = (BM+4) if kernel5 else BM
-  As = UOp(Ops.DEFINE_LOCAL, dtypes.float.ptr(BK*BM_As_stride, AddrSpace.LOCAL), arg=0)
-  Bs = UOp(Ops.DEFINE_LOCAL, dtypes.float.ptr(BK*BN, AddrSpace.LOCAL), arg=1)
-
-  c_regs = UOp(Ops.DEFINE_REG, dtypes.float.ptr(TM * nbIterWaveM * TN * nbIterWaveN), arg=2)
-
-  i = UOp.range(c_regs.dtype.size, 16)
-  init_store = c_regs[i].store(UOp.const(dtypes.float, 0.0), i)
-
-  if kernel4:
-    regA = UOp(Ops.DEFINE_REG, dtypes.float.ptr(nbReadsA, AddrSpace.REG), arg=3)
-    regB = UOp(Ops.DEFINE_REG, dtypes.float.ptr(nbReadsB, AddrSpace.REG), arg=4)
-
-    # initial load from globals into locals (0)
-    kId = 0
-
-    # load from globals into locals
-    i = UOp.range(nbReadsB, 0)
-    index_x = BN * blockIdx_x + rBIdx
-    index_y = rBIdy + i * strideReadB + kId
-    Bs_store = Bs[(index_y % BK) * BN + index_x % BN].store(b[N * index_y + index_x].load(), i)
-
-    i = UOp.range(nbReadsA, 1)
-    index_x = rAIdx + kId
-    index_y = BM * blockIdx_y + rAIdy + i * strideReadA
-    As_store = As[(index_x % BK) * BM_As_stride + index_y % BM].store(a[N * index_y + index_x].load(), i)
-
-    # iterate over the middle chunk
-    kId_range = UOp.range(N//BK-1, 2)
-    kId = kId_range*BK
-
-    barrier = UOp.barrier(As_store, Bs_store)
-
-    # load from globals into registers (next round)
-    i = UOp.range(nbReadsB, 3)
-    index_x = BN * blockIdx_x + rBIdx
-    index_y = rBIdy + i * strideReadB + kId + BK
-    regB_store = regB[i].store(b[N * index_y + index_x].load(), i)
-
-    i = UOp.range(nbReadsA, 4)
-    index_x = rAIdx + kId + BK
-    index_y = BM * blockIdx_y + rAIdy + i * strideReadA
-    regA_store = regA[i].store(a[N * index_y + index_x].load(), i)
-
-    def inner_loop(first_range, inp_dep=()):
-      # inner unroll
-      k = UOp.range(BK, first_range+0)
-
-      # load from locals into registers
-      iterWave = UOp.range(nbIterWaveN, first_range+1)
-      i = UOp.range(TN, first_range+2)
-      index = waveIdx * WN + iterWave * SUBWN + TN * idxInWave + i
-      B_row_store = B_row[iterWave*TN + i].store(Bs[k*BN + index].load(*inp_dep), iterWave, i)
-
-      iterWave = UOp.range(nbIterWaveM, first_range+3)
-      i = UOp.range(TM, first_range+4)
-      index = waveIdy * WM + iterWave * SUBWM + TM * idyInWave + i
-      A_col_store = A_col[iterWave*TM + i].store(As[k*BM_As_stride + index].load(*inp_dep), iterWave, i)
-
-      # do the GEMM math
-      iterWaveM = UOp.range(nbIterWaveM, first_range+5)
-      yt = UOp.range(TM, first_range+6)
-      iterWaveN = UOp.range(nbIterWaveN, first_range+7)
-      xt = UOp.range(TN, first_range+8)
-      x = iterWaveN * TN + xt
-      y = iterWaveM * TM + yt
-      c_regs_idx = c_regs[y * TN * nbIterWaveN + x]
-      # sketchy, this should end the kId_range but it doesn't
-      sink = c_regs_idx.store(c_regs_idx.load(init_store) + A_col[y].load(A_col_store) * B_row[x].load(B_row_store),
-                              iterWaveM, iterWaveN, yt, xt, k)
-      return sink
-
-    # TODO: kId_range should endrange after a barrier
-    sink = inner_loop(5, (barrier, regB_store, regA_store)).barrier()
-
-    # load from registers into locals
-    i = UOp.range(nbReadsB, 14)
-    index_x = BN * blockIdx_x + rBIdx
-    index_y = rBIdy + i * strideReadB + kId + BK
-    Bs_store = Bs[(index_y % BK) * BN + index_x % BN].store(regB[i].load(sink), i, kId_range)
-
-    i = UOp.range(nbReadsA, 15)
-    index_x = rAIdx + kId + BK
-    index_y = BM * blockIdx_y + rAIdy + i * strideReadA
-    As_store = As[(index_x % BK) * BM_As_stride + index_y % BM].store(regA[i].load(sink), i, kId_range)
-
-    # final iteration without the copy
-    sink = inner_loop(16, (UOp.barrier(Bs_store, As_store),))
-  else:
-    kId_range = UOp.range(N//BK, 0)
-    kId = kId_range*BK
-
-    # load from globals into locals
-    i = UOp.range(nbReadsB, 1)
-    index_x = BN * blockIdx_x + rBIdx
-    index_y = rBIdy + i * strideReadB + kId
-    Bs_store = Bs[(index_y % BK) * BN + index_x % BN].store(b[N * index_y + index_x].load(), i)
-
-    i = UOp.range(nbReadsA, 2)
-    index_x = rAIdx + kId
-    index_y = BM * blockIdx_y + rAIdy + i * strideReadA
-    As_store = As[(index_x % BK) * BM_As_stride + index_y % BM].store(a[N * index_y + index_x].load(), i)
-
-    barrier = UOp.barrier(As_store, Bs_store)
-
-    k = UOp.range(BK, 3)
-
-    # load from locals into registers
-    iterWave = UOp.range(nbIterWaveN, 4)
-    i = UOp.range(TN, 5)
-    index = waveIdx * WN + iterWave * SUBWN + TN * idxInWave + i
-    B_row_store = B_row[iterWave*TN + i].store(Bs[k*BN + index].load(barrier), iterWave, i)
-
-    iterWave = UOp.range(nbIterWaveM, 6)
-    i = UOp.range(TM, 7)
-    index = waveIdy * WM + iterWave * SUBWM + TM * idyInWave + i
-    A_col_store = A_col[iterWave*TM + i].store(As[k*BM_As_stride + index].load(barrier), iterWave, i)
-
-    # do the GEMM math
-    iterWaveM = UOp.range(nbIterWaveM, 8)
-    yt = UOp.range(TM, 9)
-    iterWaveN = UOp.range(nbIterWaveN, 10)
-    xt = UOp.range(TN, 12)
-    x = iterWaveN * TN + xt
-    y = iterWaveM * TM + yt
-    c_regs_idx = c_regs[y * TN * nbIterWaveN + x]
-    sink = c_regs_idx.store(c_regs_idx.load(init_store) + A_col[y].load(A_col_store) * B_row[x].load(B_row_store),
-                            iterWaveM, iterWaveN, yt, xt, k, kId_range)
-
-  # store c_regs into c
-  iterWaveM = UOp.range(nbIterWaveM, 1000)
-  yt = UOp.range(TM, 1001)
-  iterWaveN = UOp.range(nbIterWaveN, 1002)
-  xt = UOp.range(TN, 1003)
-  xOut = blockIdx_x * BN + waveIdx * WN + iterWaveN * SUBWN + TN * idxInWave
-  yOut = blockIdx_y * BM + waveIdy * WM + iterWaveM * SUBWM + TM * idyInWave
-  indexC = N * (yOut + yt) + xOut + xt
-  sink = c[indexC].store(c_regs[TN * nbIterWaveN * (iterWaveM * TM + yt) + (iterWaveN * TN + xt)].load(sink),
-                         iterWaveM, iterWaveN, yt, xt)
-
-  return sink.sink(arg=KernelInfo(name="tinygemm"))
+  GlobalCounters.reset()
+  with Context(DEBUG=2):
+    tc = (a @ b).realize()
+  with Context(DEBUG=0):
+    err = (hc - tc).square().mean().item()
+  print(f"mean squared error {err}")
+  if err > 1e-06:
+    raise RuntimeError("matmul is wrong!")
 
 if __name__ == "__main__":
-  HL = getenv("HL")
-  if HL == 3: hprg = rangeify_kernel3()
-  elif HL == 2: hprg = top_spec_kernel3()
-  elif HL == 1: hprg = hl_spec_kernel3()
-  else: hprg = hand_spec_kernel3()
-  if HL == 3:
-    prg = get_program(hprg, Device.default.renderer)
-  else:
-    prg = get_program(hprg, Device.default.renderer)
-  print(prg.src)
-  if getenv("SRC"): exit(0)
-  hrunner = CompiledRunner(prg)
-
-  a = Tensor.randn(N, N).realize()
-  b = Tensor.randn(N, N).realize()
-  hc = Tensor.zeros(N, N).contiguous().realize()
-
-  GlobalCounters.reset()
-  with Context(DEBUG=2):
-    for _ in range(run_count): tc = (a@b).realize()
-
-  GlobalCounters.reset()
-  buffers = [hc.uop.buffer, a.uop.buffer, b.uop.buffer]
-  ei = ExecItem(hrunner, buffers)
-  with Context(DEBUG=2):
-    for _ in range(run_count): ei.run(wait=True)
-  err = (hc-tc).square().mean().item()
-  print(f"hrunner {err}")
-  if err > 1e-06: raise RuntimeError("matmul is wrong!")
+  test_matmul(hand_spec_kernel3(), N=N)

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/metal_uop_matmul.py

@@ -0,0 +1,42 @@
+from tinygrad import UOp, dtypes
+from tinygrad.uop.ops import AxisType, Ops, KernelInfo, AddrSpace
+from extra.gemm.amd_uop_matmul import test_matmul
+
+N = 2048
+
+# metal has an 8x8 tensor core. this is the indexing
+def mat_idx(buf, g0, g1, warp, u):
+  l = [(warp//2**i)%2 for i in range(5)]
+  return buf[g0, l[4]*4 + l[2]*2 + l[1], g1, l[3]*4 + l[0]*2 + u]
+
+def hand_spec_tc_cores():
+  gx = UOp.special(N // 8, "gidx0")
+  gy = UOp.special(N // 8, "gidx1")
+  warp = UOp.special(32, "lidx0")
+
+  c = UOp.placeholder((N, N), dtypes.float, slot=0).reshape((N//8, 8, N//8, 8))
+  a = UOp.placeholder((N, N), dtypes.float, slot=1).reshape((N//8, 8, N//8, 8))
+  b = UOp.placeholder((N, N), dtypes.float, slot=2).reshape((N//8, 8, N//8, 8))
+
+  gk = UOp.range(N // 8, 0, AxisType.REDUCE)
+
+  a_tc = UOp.vectorize(*[mat_idx(a, gx, gk, warp, i) for i in range(2)])
+  b_tc = UOp.vectorize(*[mat_idx(b, gk, gy, warp, i) for i in range(2)])
+
+  acc = UOp.placeholder((2,), dtypes.float, slot=0, addrspace=AddrSpace.REG)
+  acc = acc[0].set(0.0)
+  acc = acc[1].set(0.0)
+
+  # TODO: make this simple
+  wmma_arg = ('WMMA_8_8_8_float_float', (8, 8, 8), dtypes.float, dtypes.float, 'METAL', 32, (((3, 2),), ((3, 2),), ((3, 2),)), ())
+
+  acc_load = UOp.vectorize(acc.after(gk)[0], acc.after(gk)[1])
+  out = UOp(Ops.WMMA, dtypes.float.vec(2), (a_tc, b_tc, acc_load), arg=wmma_arg)
+
+  end_loop = UOp.group(*[acc[i].store(out.gep(i)) for i in range(2)]).end(gk)
+
+  sink = UOp.group(*[mat_idx(c.after(end_loop), gx, gy, warp, i).store(acc[i]) for i in range(2)])
+  return sink.sink(arg=KernelInfo(name="custom_metal_matmul", opts_to_apply=())).simplify()
+
+if __name__ == "__main__":
+  test_matmul(hand_spec_tc_cores(), N=N)

extra/gemm/mi350x_uop_matmul.py

@@ -0,0 +1,229 @@
+import os
+import numpy as np
+np.set_printoptions(linewidth=1000000)
+os.environ["AMD_LLVM"] = "0"
+
+from tinygrad import Tensor, Context, dtypes, UOp, GlobalCounters
+from tinygrad.helpers import DEBUG, getenv
+from tinygrad.dtype import AddrSpace
+from tinygrad.uop.ops import AxisType, KernelInfo, Ops
+
+WARP_SIZE = 64
+
+# Reg tile sizes (tensor cores)
+TC_M = 16
+TC_N = 16
+TC_K = 32
+
+# 1024 matrix cores
+# 16 cycle mfma
+# 2.2 GHz
+# 16x16x32x2 FLOPS/mma = 16384
+# 2.2*1e9*16384*1024/16*1e-12 TFLOPS = 2306 TFLOPS
+
+#N,M,K = 256,256,64
+N,M,K = 4096,4096,4096
+
+# Threadblock tile sizes (block-level tile of C that a block computes)
+#BLOCK_M = 128   # rows of C (M-dim) per block
+#BLOCK_N = 128   # columns of C (N-dim) per block
+#BLOCK_K = 128   # K-slice per block iteration
+
+BLOCK_M = 64
+BLOCK_N = 64
+BLOCK_K = 128
+
+WARPGROUP_SIZE = 1
+BLOCK_M = BLOCK_M * WARPGROUP_SIZE
+
+# TODO: improve the syntax of this. better syntax, faster iteration
+#  -- 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)
+#  -- 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)
+
+CUS_PER_GPU = 256
+assert ((M//BLOCK_M) * (N//BLOCK_N)) >= CUS_PER_GPU, "not enough globals"
+
+def custom_gemm(C:UOp, A:UOp, B:UOp) -> UOp:
+  # A = (M x K)
+  # B = (K x N)
+  # C = (M x N)
+
+  # check it's proper matmul
+  assert C.shape[0] == A.shape[0]
+  assert C.shape[1] == B.shape[1]
+  assert A.shape[1] == B.shape[0]
+
+  gx, gy = UOp.special(M//BLOCK_M, "gidx0"), UOp.special(N//BLOCK_N, "gidx1")
+  warp = UOp.special(WARP_SIZE, "lidx0")
+  warpgroup = UOp.special(WARPGROUP_SIZE, "lidx1")
+
+  # generic copy logic (not good)
+  def generic_copy(glbl, gargs, lcl, rng):
+    # Fully coalesced 128-bit loads/stores.
+    INNER_SIZE = 8
+    cp_i = UOp.range(lcl.size//(WARPGROUP_SIZE*WARP_SIZE*INNER_SIZE), rng)
+    cp_inner = UOp.range(INNER_SIZE, rng+1, AxisType.UPCAST)
+    idx_i = cp_i*WARPGROUP_SIZE*WARP_SIZE*INNER_SIZE + warpgroup*WARP_SIZE*INNER_SIZE + warp*INNER_SIZE + cp_inner
+    return lcl[idx_i].store(glbl[*gargs, idx_i]).end(cp_i, cp_inner)
+
+  # split out the globals into blocks
+  C = C.reshape((M//BLOCK_M, BLOCK_M, N//BLOCK_N, BLOCK_N))
+  A = A.reshape((M//BLOCK_M, BLOCK_M, K//BLOCK_K, BLOCK_K))
+  B = B.reshape((K//BLOCK_K, BLOCK_K, N//BLOCK_N, BLOCK_N))
+
+  # this is the big accumulator
+  acc = UOp.placeholder((BLOCK_N//TC_N, BLOCK_M//TC_M//WARPGROUP_SIZE), dtypes.float.vec(4), 0, AddrSpace.REG)
+  assert acc.size*WARP_SIZE*WARPGROUP_SIZE*4 == BLOCK_M*BLOCK_N
+  acc = acc[init_l:=UOp.range(acc.size, 500)].set(UOp.const(dtypes.float.vec(4), 0.0), end=init_l)
+
+  # create locals (note A is permuted, and the stride is changed to avoid bank conflicts)
+  def make_locals(slot) -> tuple[UOp, UOp]:
+    BM_As_stride = (BLOCK_M + 1)
+    BN_Bs_stride = (BLOCK_N + 0)
+    INNER_SLICE = 8
+    As = UOp.placeholder((BLOCK_K//INNER_SLICE, BM_As_stride, INNER_SLICE), dtypes.half, slot=slot, addrspace=AddrSpace.LOCAL)
+    INNER_SLICE = 1
+    Bs = UOp.placeholder((BLOCK_K//INNER_SLICE, BN_Bs_stride, INNER_SLICE), dtypes.half, slot=slot+1, addrspace=AddrSpace.LOCAL)
+    As = As.permute((0,2,1)).reshape((BLOCK_K, BM_As_stride)).shrink_to((BLOCK_K, BLOCK_M))
+    Bs = Bs.permute((0,2,1)).reshape((BLOCK_K, BN_Bs_stride)).shrink_to((BLOCK_K, BLOCK_N))
+    return As, Bs
+
+  # load from globals into locals (TODO: use the warpgroup)
+
+  def load_to_locals(l_K_outer_loop:UOp, Asl:UOp, Bsl:UOp, rng:int, barrier=True) -> tuple[UOp, UOp]:
+    if getenv("FAKE"):
+      return Asl[0].set(0), Bsl[0].set(0)
+    else:
+      pA = A.permute((0,2,1,3)).reshape((M//BLOCK_M, K//BLOCK_K, BLOCK_M*BLOCK_K))
+      pas = Asl.permute((1,0)).reshape((BLOCK_M*BLOCK_K,))
+      As_store = generic_copy(pA, (gx, l_K_outer_loop), pas, rng)
+
+      pB = B.permute((0,2,1,3)).reshape((K//BLOCK_K, N//BLOCK_N, BLOCK_K*BLOCK_N))
+      pbs = Bsl.reshape((BLOCK_K*BLOCK_N,))
+      Bs_store = generic_copy(pB, (l_K_outer_loop, gy), pbs, rng+2)
+
+      barrier = UOp.barrier(As_store, Bs_store) if barrier else UOp.group(As_store, Bs_store)
+      return Asl.after(barrier), Bsl.after(barrier)
+
+  def compute_on_locals(acc:UOp, Asl:UOp, Bsl:UOp, rng:int, afters:tuple[UOp, ...]=()) -> UOp:
+    K_inner_loop = UOp.range(BLOCK_K//TC_K, rng, AxisType.REDUCE)
+
+    # load from locals into registers
+    Ar = UOp.placeholder((BLOCK_M//TC_M//WARPGROUP_SIZE,), dtypes.half.vec(8), slot=1, addrspace=AddrSpace.REG)
+    Br = UOp.placeholder((BLOCK_N//TC_N,), dtypes.half.vec(8), slot=2, addrspace=AddrSpace.REG)
+
+    M_load_loop = UOp.range(BLOCK_M//TC_M//WARPGROUP_SIZE, rng+10)
+    Asl = Asl.reshape((BLOCK_K//TC_K, TC_K, BLOCK_M//TC_M//WARPGROUP_SIZE, WARPGROUP_SIZE, TC_M))
+    load_rng = UOp.range(8, rng+11, axis_type=AxisType.UPCAST)
+    A_in = Asl[K_inner_loop, (warp//16)*8+load_rng, M_load_loop, warpgroup, warp%16].contract(load_rng)
+    Ar = Ar[M_load_loop].set(A_in, end=M_load_loop)
+
+    N_load_loop = UOp.range(BLOCK_N//TC_N, rng+20)
+    Bsl = Bsl.reshape((BLOCK_K//TC_K, TC_K, BLOCK_N//TC_N, TC_N))
+    load_rng = UOp.range(8, rng+21, axis_type=AxisType.UPCAST)
+    B_in = Bsl[K_inner_loop, (warp//16)*8+load_rng, N_load_loop, warp%16].contract(load_rng)
+    Br = Br[N_load_loop].set(B_in, end=N_load_loop)
+
+    M_inner_loop = UOp.range(BLOCK_M//TC_M//WARPGROUP_SIZE, rng+30)
+    N_inner_loop = UOp.range(BLOCK_N//TC_N, rng+31)
+
+    # load values
+    acc_after = acc.after(*afters, M_inner_loop, N_inner_loop, K_inner_loop)
+    acc_load = acc_after[N_inner_loop, M_inner_loop]
+
+    # do WMMA
+    wmma_arg = ('WMMA_16_16_32_half_float', (16, 16, 32), dtypes.half, dtypes.float, 'AMD', 64, ((), (), ((3, 2), (2, 2))), ())
+    out = UOp(Ops.WMMA, dtypes.float.vec(4), (Ar[M_inner_loop], Br[N_inner_loop], acc_load), arg=wmma_arg)
+
+    # store back the acc
+    acc_store = acc[N_inner_loop, M_inner_loop].store(out)
+    return acc_store.end(M_inner_loop, N_inner_loop, K_inner_loop)
+
+  # **** START INNER LOOP *****
+  # inner loop -- locals -> regs
+
+  # no pipeline
+  if not getenv("PIPELINE"):
+    As, Bs = make_locals(slot=0)
+
+    K_outer_loop = UOp.range(K//BLOCK_K, 0, AxisType.REDUCE)
+    As, Bs = load_to_locals(K_outer_loop, As, Bs, 1000, barrier=True)
+    acc_store = compute_on_locals(acc, As, Bs, 1500, afters=(K_outer_loop,))
+    acc = acc.after(acc_store.barrier().end(K_outer_loop))
+  else:
+    # this doesn't work
+    As0, Bs0 = make_locals(slot=0)
+    As1, Bs1 = make_locals(slot=2)
+    As0, Bs0 = load_to_locals(0, As0, Bs0, 1000)
+
+    K_outer_loop = UOp.range((K//BLOCK_K-2)//2, 0, AxisType.REDUCE)
+    As1, Bs1 = load_to_locals(K_outer_loop+1, As1, Bs1, 2000, barrier=False)
+    acc_store = compute_on_locals(acc, As0, Bs0, 1500, afters=(K_outer_loop,))
+    As0, Bs0 = load_to_locals(K_outer_loop+2, As0, Bs0, 3000, barrier=False)
+    acc_store = compute_on_locals(acc, As1, Bs1, 2500, afters=(acc_store, As0, Bs0))
+    acc = acc.after(acc_store.barrier().end(K_outer_loop))
+
+    #acc_store = compute_on_locals(acc, As0, Bs0, 3500, afters=(acc_store.barrier().end(K_outer_loop)))
+    """
+    As1, Bs1 = load_to_locals(K//BLOCK_K-1, As1, Bs1, 4000)
+    acc_store = compute_on_locals(acc, As1, Bs1, 4500, afters=(acc_store))
+    """
+    #acc = acc.after(acc_store)
+
+  # **** END LOOPS *****
+
+  # store the acc into gmem
+  cp_i, cp_j = UOp.range(BLOCK_M//TC_M//WARPGROUP_SIZE, 10004), UOp.range(BLOCK_N//TC_N, 10005)
+  c_load = lambda i: C[gx, cp_i*TC_M*WARPGROUP_SIZE + warpgroup*TC_M + (warp//16)*4+i, gy, cp_j*TC_N + warp%16]
+  store = UOp.group(*[c_load(i).store(acc[cp_j, cp_i].gep(i)) for i in range(4)])
+  store = store.end(cp_i, cp_j)
+
+  return store.sink(arg=KernelInfo(name="custom_gemm", opts_to_apply=())).simplify()
+
+# simplest WMMA
+"""
+# init the acc
+acc = UOp.placeholder((4,), dtypes.float, 0, AddrSpace.REG)
+acc = acc[init_l:=UOp.range(4, 1)].set(0.0, end=init_l)
+
+# do the wmma
+acc_load = UOp.vectorize(*[acc.after(K_loop)[i] for i in range(4)])
+wmma_arg = ('WMMA_16_16_32_half_float', (16, 16, 32), dtypes.half, dtypes.float, 'AMD', 64, ((), (), ((3, 2), (2, 2))), ())
+out = UOp(Ops.WMMA, dtypes.float.vec(4), (A_in, B_in, acc_load), arg=wmma_arg)
+
+# store back the acc
+acc = acc.after(UOp.group(*[acc[i].store(out.gep(i)) for i in range(4)]).end(K_loop))
+
+# store the acc into gmem
+store = UOp.group(*[C[gx, (warp//16)*4+i, gy, warp%16].store(acc[i]) for i in range(4)])
+"""
+
+if __name__ == "__main__":
+  a = Tensor.randn(M, K, dtype=dtypes.half)
+  b = Tensor.randn(K, N, dtype=dtypes.half)
+
+  #a = Tensor.zeros(M, K, dtype=dtypes.half).contiguous()
+  #a[0,16] = 1
+  #b = Tensor.ones(K, N, dtype=dtypes.half).contiguous()
+
+  c = Tensor.empty(M, N, dtype=dtypes.float)
+  with Context(DEBUG=0): Tensor.realize(a,b)
+
+  ref = a.dot(b, dtype=dtypes.float)
+  ref.realize()
+
+  GlobalCounters.reset()
+  with Context(DEBUG=max(2, DEBUG.value), DEVECTORIZE=2):
+    tst = Tensor.custom_kernel(c, a, b, fxn=custom_gemm)[0]
+    tst.realize()
+  print(f"{(N*M*K*2 / GlobalCounters.time_sum_s)*1e-12:.2f} REAL TFLOPS")
+
+  with Context(DEBUG=0):
+    #print(ref.numpy())
+    #print(tst.numpy())
+    assert Tensor.isclose(ref, tst, atol=1e-2).all().item(), "matrix not close"

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/gemm/torch_gemm.py

@@ -9,6 +9,7 @@ torch.set_num_threads(1)
 from tinygrad.helpers import getenv
 CUDA = getenv("CUDA", 1)
 MPS = getenv("MPS", 0)
+if getenv("FP16_ACC"): torch.backends.cuda.matmul.allow_fp16_accumulation = True
 
 for dtype in [torch.float32, torch.float16, torch.bfloat16]:
   for N in [256, 512, 1024, 2048, 4096]:

extra/mmapeak/mmapeak.py

@@ -84,12 +84,14 @@ if __name__=="__main__":
     NUM_WORKGROUPS = 256
     WAVE_SIZE = 64
     NUM_WAVES = 4
+    launchBenchmark("v_mfma_f32_16x16x16_f16", (3,0,1), accum=True)
     launchBenchmark("v_mfma_f32_16x16x16_bf16", (3,0,1), accum=True)
     FLOPS_PER_MATMUL = 16*16*32*2
+    launchBenchmark("v_mfma_f32_16x16x32_f16", (3,0,3), accum=True)
     launchBenchmark("v_mfma_f32_16x16x32_bf16", (3,0,3), accum=True)
     FLOPS_PER_MATMUL = 16*16*128*2
     launchBenchmark("v_mfma_f32_16x16x128_f8f6f4", (3,0,7), accum=True) # fp8
     launchBenchmark("v_mfma_f32_16x16x128_f8f6f4", (3,0,5), accum=True, extra=", cbsz:2 blgp:2") # fp6
     launchBenchmark("v_mfma_f32_16x16x128_f8f6f4", (3,0,3), accum=True, extra=", cbsz:4 blgp:4") # fp4
   else:
-    raise RuntimeError(f"arch {DEV.arch} not supported.")
+    raise RuntimeError(f"arch {DEV.arch} not supported.")

extra/qcom_gpu_driver/msm_kgsl.py

@@ -10,6 +10,8 @@ import ctypes
 
 
 class AsDictMixin:
+    import sys
+    if sys.version_info >= (3, 14): _layout_ = 'ms'
     @classmethod
     def as_dict(cls, self):
         result = {}

extra/sqtt/README.md

@@ -2,8 +2,6 @@
 
 ## Getting SQ Thread Trace
 
-Only supported on 7900XTX, requires either AM (`rmmod amdgpu`) or disabling power gating on AMD (`ppfeaturemask=0xffff3fff`, don't forget to rebuild initramfs)
-
 SQTT is implemented on top of normal tinygrad profiling, `VIZ=1 SQTT=1` to get profile pickle with sqtt data embedded in it.
 
 `SQTT_BUFFER_SIZE=X` to change size of SQTT buffer (per shader engine, 6 SEs on 7900xtx) in megabytes, default 256.

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
-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.runtime.ops_amd import ProfileSQTTEvent
+import ctypes, pathlib, argparse, pickle, re, functools, dataclasses, itertools
+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:
@@ -17,51 +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] = []
   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, 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):
@@ -78,12 +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
@@ -96,5 +138,27 @@ if __name__ == "__main__":
 
     return rocprof.ROCPROFILER_THREAD_TRACE_DECODER_STATUS_SUCCESS
 
-  rocprof.rocprof_trace_decoder_parse_data(copy_cb, trace_cb, isa_cb, None)
-  print(ROCParseCtx.wave_events.keys())
+  try:
+    rocprof.rocprof_trace_decoder_parse_data(copy_cb, trace_cb, isa_cb, None)
+  except AttributeError as e: raise RuntimeError("Failed to find rocprof-trace-decoder. Run ./extra/sqtt/install_sqtt_decoder.py to install") from e
+  return ROCParseCtx
+
+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:
+      view = memoryview(ev.blob).cast('Q')
+      print(f"\t{s.name}")
+      for xcc, inst, se_idx, sa_idx, wgp_idx in itertools.product(range(s.xcc), range(s.inst), range(s.se), range(s.sa), range(s.wgp)):
+        print(f"\t\tXCC {xcc} Inst {inst} SE {se_idx} SA {sa_idx} WGP {wgp_idx}: {view[ptr]:#x}")
+        ptr += 1

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

@@ -0,0 +1,106 @@
+#include "kittens.cuh"
+
+using namespace kittens;
+
+constexpr int NUM_WORKERS = 4;
+constexpr int PIPE_STAGES = 3;
+
+constexpr int ATTN_B = 16;
+constexpr int ATTN_N = 1024;
+constexpr int ATTN_H = 16;
+constexpr int ATTN_D = 64;
+
+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>;
+template<int D> using global_layout = gl<bf16, -1, -1, -1, D>; // B, N, H, specified at runtime, D known at compile time for this kernel
+template<int D> struct globals { global_layout<D> Qg, Kg, Vg, Og; };
+
+__launch_bounds__(NUM_WORKERS*WARP_THREADS, 1)
+__global__ void attend_ker(bf16 *O_ptr, bf16 *Q_ptr, bf16 *K_ptr, bf16 *V_ptr) {
+    constexpr int D = ATTN_D;
+    global_layout<D> Qg{Q_ptr, ATTN_B, ATTN_N, ATTN_H, nullptr};
+    global_layout<D> Kg{K_ptr, ATTN_B, ATTN_N, ATTN_H, nullptr};
+    global_layout<D> Vg{V_ptr, ATTN_B, ATTN_N, ATTN_H, nullptr};
+    global_layout<D> Og{O_ptr, ATTN_B, ATTN_N, ATTN_H, nullptr};
+    globals<D> g(Qg, Kg, Vg, Og);
+
+    using load_group = kittens::group<2>; // pairs of workers collaboratively load k, v tiles
+    int loadid = load_group::groupid(), workerid = kittens::warpid(); // which worker am I?
+    constexpr int LOAD_BLOCKS = NUM_WORKERS / load_group::GROUP_WARPS;
+    const int batch = blockIdx.z, head = blockIdx.y, q_seq = blockIdx.x * NUM_WORKERS + workerid;
+
+    extern __shared__ alignment_dummy __shm[];
+    shared_allocator al((int*)&__shm[0]);
+
+    shared_tile<D> (&k_smem)[LOAD_BLOCKS][PIPE_STAGES] = al.allocate<shared_tile<D>, LOAD_BLOCKS, PIPE_STAGES>();
+    shared_tile<D> (&v_smem)[LOAD_BLOCKS][PIPE_STAGES] = al.allocate<shared_tile<D>, LOAD_BLOCKS, PIPE_STAGES>();
+
+    shared_tile<D> (&qo_smem)[NUM_WORKERS] = reinterpret_cast<shared_tile<D>(&)[NUM_WORKERS]>(k_smem);
+    // Initialize all of the register tiles.
+    qkvo_tile<D, bf16> q_reg, k_reg; // Q and K are both row layout, as we use mma_ABt.
+    qkvo_tile<D, bf16, col_l> v_reg; // V is column layout, as we use mma_AB.
+    qkvo_tile<D, float> o_reg; // Output tile.
+    attn_tile<D, float> att_block; // attention tile, in float. (We want to use float wherever possible.)
+    attn_tile<D, bf16> att_block_mma; // bf16 attention tile for the second mma_AB. We cast right before that op.
+    typename attn_tile<D, float>::col_vec max_vec_last, max_vec, norm_vec; // these are column vectors for the in-place softmax.
+    // each warp loads its own Q tile of 16x64
+    if (q_seq*ROWS<D> < g.Qg.depth()) {
+      warp::load<1, false>(qo_smem[workerid], g.Qg, {batch, q_seq, head, 0});  // going through shared memory improves coalescing of dram reads.
+      __syncwarp();
+      warp::load(q_reg, qo_smem[workerid]);
+    }
+    __syncthreads();
+
+    if constexpr(D == 64) q_reg *= __float2bfloat16(0.125f * 1.44269504089f);
+    else if constexpr(D == 128) q_reg *= __float2bfloat16(0.08838834764f * 1.44269504089f);
+
+    max_vec = base_types::constants<float>::neg_infty();
+    norm_vec = 0.f;
+    o_reg = 0.f;
+    // launch the load of the first k, v tiles
+    int kv_blocks = (g.Kg.depth() + LOAD_BLOCKS*ROWS<D>-1) / (LOAD_BLOCKS*ROWS<D>), tic = 0;
+    load_group::load_async<1, false>(k_smem[loadid][0], g.Kg, {batch, loadid, head, 0});
+    load_group::load_async<1, false>(v_smem[loadid][0], g.Vg, {batch, loadid, head, 0});
+    // iterate over k, v for these q's that have been loaded
+    for(auto kv_idx = 0; kv_idx < kv_blocks; kv_idx++, tic=(tic+1)%3) {
+        int next_load_idx = (kv_idx+1)*LOAD_BLOCKS + loadid;
+        if(next_load_idx*ROWS<D> < g.Kg.depth()) {
+            int next_tic = (tic+1)%3;
+            load_group::load_async<1, false>(k_smem[loadid][next_tic], g.Kg, {batch, next_load_idx, head, 0});
+            load_group::load_async<1, false>(v_smem[loadid][next_tic], g.Vg, {batch, next_load_idx, head, 0});
+            load_async_wait<1>(); // next k, v can stay in flight.
+        }
+        else load_async_wait();
+        __syncthreads();
+
+        #pragma unroll LOAD_BLOCKS
+        for(int subtile = 0; subtile < LOAD_BLOCKS && (kv_idx*LOAD_BLOCKS + subtile)*ROWS<D> < g.Kg.depth(); subtile++) {
+          warp::load(k_reg, k_smem[subtile][tic]); // load k from shared into registers
+          att_block = 0.f; // zero 16x16 attention tile
+          warp::mma<transpose::N, transpose::T>(att_block, q_reg, k_reg, att_block); // [email protected]
+          // int first_index = (kv_idx*LOAD_BLOCKS + subtile)*ROWS<D>; // one past the last KV index of this tile
+          // int start_fill = g.Kg.depth()-first_index < ROWS<D> ? g.Kg.depth()-first_index : ROWS<D>;
+          // right_fill(att_block, att_block, start_fill, base_types::constants<float>::neg_infty());
+          max_vec_last = max_vec;
+          max_vec = warp::max<axis::COL>(att_block, max_vec); 
+          att_block = warp::exp2(att_block - max_vec); 
+          max_vec_last = warp::exp2(max_vec_last - max_vec); 
+          norm_vec *= max_vec_last; 
+          norm_vec = warp::sum<axis::COL>(att_block, norm_vec); 
+          att_block_mma = att_block; // copy to bf16 tile
+          warp::load(v_reg, v_smem[subtile][tic]);
+          o_reg *= max_vec_last;
+          warp::mma<transpose::N, transpose::N>(o_reg, att_block_mma, v_reg, o_reg);
+        }
+    }
+
+    o_reg /= norm_vec;
+    __syncthreads();
+    if (q_seq*ROWS<D> < g.Og.depth()) { // write out o.
+      warp::store(qo_smem[workerid], o_reg); // going through shared memory improves coalescing of dram writes.
+      __syncwarp();
+      warp::store<1, false>(g.Og, qo_smem[workerid], {batch, q_seq, head, 0});
+    }
+}

extra/thunder/cuda/fa.py

@@ -0,0 +1,43 @@
+import pathlib
+from tinygrad import Device, Tensor
+from tinygrad.helpers import Context
+from tinygrad.runtime.support.compiler_cuda import pretty_ptx, NVCCCompiler
+
+if __name__ == "__main__":
+  code = (pathlib.Path(__file__).parent / "fa.cu").read_text()
+  device = Device["CUDA"]
+  kitten_args = [f"-I{(pathlib.Path(__file__).parent / 'include').as_posix()}", "-std=c++20", "--expt-relaxed-constexpr", "-DKITTENS_4090"]
+  lib = NVCCCompiler(device.compiler.arch, kitten_args).compile(code)
+  kernel_name = lib.decode().split(".globl\t")[1].split("\n")[0]
+  print("kernel name", kernel_name)
+  print(pretty_ptx(lib.decode()))
+
+  prg = device.runtime(kernel_name, lib)
+  prg.smem = 16384 * 3
+
+  B, N, H, D = 16, 1024, 16, 64
+  q = Tensor.randn(B, N, H, D, device='CUDA', dtype="bfloat16")
+  k = Tensor.randn(B, N, H, D, device='CUDA', dtype="bfloat16")
+  v = Tensor.randn(B, N, H, D, device='CUDA', dtype="bfloat16")
+  out = Tensor.empty(B, N, H, D, device='CUDA', dtype="bfloat16")
+  Tensor.realize(q, k, v, out)
+
+  NUM_WORKERS = 4
+  ROWS = 16 * (64 // D)
+
+  gsz = (N // (ROWS*NUM_WORKERS), H, B)
+  for _ in range(5):
+    et = prg(out.uop.buffer.ensure_allocated()._buf, q.uop.buffer._buf, k.uop.buffer._buf, v.uop.buffer._buf,
+             global_size=gsz, local_size=(ROWS*NUM_WORKERS,1,1), wait=True)
+
+    attn_flops = 2 * B * H * N * N * D + \
+                 4 * B * H * N * N + \
+                 2 * B * H * N * N * D
+    print(f"{attn_flops/(et*1e9):2f} GFLOPS")
+
+  for _ in range(5):
+    with Context(DEBUG=2):
+      ref = q.scaled_dot_product_attention(k, v)
+
+  ref, out = ref.float(), out.float()
+  print((ref-out).mean().item(), (ref-out).max().item())

extra/thunder/cuda/matmul.cu

@@ -5,11 +5,11 @@ using namespace kittens;
 constexpr int g_N = 8192;
 constexpr int BLOCK_SIZE = 32;
 #define NUM_WORKERS (1)
-#define NUM_THREADS (NUM_WORKERS*kittens::WARP_THREADS)
 
 using sub_tile = st_bf<BLOCK_SIZE,BLOCK_SIZE>;
-using tile_gl =  gl<bf16,  1, 1, g_N, g_N, sub_tile>;
+using tile_gl =  gl<bf16, 1, 1, g_N, g_N>;
 
+__launch_bounds__(NUM_WORKERS*WARP_THREADS, 1)
 __global__ void kernel(bf16 *c_ptr, bf16 *a_ptr, bf16 *b_ptr) {
   tile_gl g_C{c_ptr, nullptr, nullptr, nullptr, nullptr};
   tile_gl g_A{a_ptr, nullptr, nullptr, nullptr, nullptr};

extra/thunder/cuda/matmul.py

@@ -1,10 +1,14 @@
 import pathlib
 from tinygrad import Device, Tensor
-from tinygrad.helpers import Context
+from tinygrad.helpers import Context, getenv
 from tinygrad.runtime.support.compiler_cuda import pretty_ptx, NVCCCompiler
 
 if __name__ == "__main__":
-  code = (pathlib.Path(__file__).parent / "matmul.cu").read_text()
+  if getenv("MATMUL2"):
+    code = (pathlib.Path(__file__).parent / "matmul2.cu").read_text()
+  else:
+    code = (pathlib.Path(__file__).parent / "matmul.cu").read_text()
+
   device = Device["CUDA"]
   kitten_args = [f"-I{(pathlib.Path(__file__).parent / 'include').as_posix()}", "-std=c++20", "--expt-relaxed-constexpr"]
   lib = NVCCCompiler(device.compiler.arch, kitten_args).compile(code)
@@ -13,7 +17,10 @@ if __name__ == "__main__":
   print(pretty_ptx(lib.decode()))
 
   prg = device.runtime(kernel_name, lib)
-  prg.smem = 10000
+  if getenv("MATMUL2"):
+    prg.smem = 16384 * 2
+  else:
+    prg.smem = 10000
 
   N = 8192
   a = Tensor.randn(N, N, device='CUDA', dtype="bfloat16")
@@ -21,14 +28,25 @@ if __name__ == "__main__":
   c = Tensor.empty(N, N, device='CUDA', dtype="bfloat16")
   Tensor.realize(a, b, c)
 
-  BLOCK_SIZE = 32
+  WARP_THREADS = 32
+  if getenv("MATMUL2"):
+    SUPER_N = 2
+    SUPER_M = 2
+    NUM_WORKERS = SUPER_N * SUPER_M
+    BLOCK_SIZE = 32
+    gsz = (N // (BLOCK_SIZE * SUPER_N), N // (BLOCK_SIZE * SUPER_M), 1)
+  else:
+    NUM_WORKERS = 1
+    BLOCK_SIZE = 32
+    gsz = (N // (BLOCK_SIZE), N // (BLOCK_SIZE), 1)
 
-  gsz = (N // BLOCK_SIZE, N // BLOCK_SIZE, 1)
   for _ in range(5):
     et = prg(c.uop.buffer.ensure_allocated()._buf, a.uop.buffer._buf, b.uop.buffer._buf,
-             global_size=gsz, local_size=(32,1,1), wait=True)
+             global_size=gsz, local_size=(NUM_WORKERS*WARP_THREADS,1,1), wait=True)
     print(f"{N*N*N*2/(et*1e9):2f} GFLOPS")
 
+  # print(c.tolist())
+
   for _ in range(5):
     with Context(DEBUG=2):
       ref = (a@b).realize()

extra/thunder/cuda/matmul2.cu

@@ -0,0 +1,105 @@
+#include "kittens.cuh"
+using namespace kittens;
+
+constexpr int g_N = 8192;
+
+constexpr int SUPER_N = 2;
+constexpr int SUPER_M = 2;
+constexpr int NUM_WORKERS = SUPER_N * SUPER_M;
+constexpr int LOAD_TASKS = SUPER_N + SUPER_M;
+
+constexpr int WORKER_M = 32;
+constexpr int WORKER_N = 32;
+
+constexpr int BLOCK_K = 32;
+constexpr int BLOCK_M = WORKER_M * SUPER_M;
+constexpr int BLOCK_N = WORKER_N * SUPER_N;
+
+constexpr int PIPE_STAGES = 2;
+
+using reg_tile_A = rt_bf<WORKER_M, BLOCK_K>;
+using reg_tile_B_col = rt_bf<BLOCK_K, WORKER_N, ducks::rt_layout::col>;
+using reg_tile_C = rt_fl<WORKER_M, WORKER_N>;
+
+using shared_tile_A = st_bf<WORKER_M, BLOCK_K>;
+using shared_tile_B = st_bf<BLOCK_K, WORKER_N>;
+using shared_tile_C = st_bf<WORKER_M, WORKER_N>;
+
+using gl_tile_A = gl<bf16, 1, 1, g_N, g_N, shared_tile_A>;
+using gl_tile_B = gl<bf16, 1, 1, g_N, g_N, shared_tile_B>;
+using gl_tile_C = gl<bf16, 1, 1, g_N, g_N, shared_tile_C>;
+
+__launch_bounds__(NUM_WORKERS *WARP_THREADS, 1) __global__
+    void kernel(bf16 *c_ptr, bf16 *a_ptr, bf16 *b_ptr) {
+  gl_tile_C g_C{c_ptr, nullptr, nullptr, nullptr, nullptr};
+  gl_tile_A g_A{a_ptr, nullptr, nullptr, nullptr, nullptr};
+  gl_tile_B g_B{b_ptr, nullptr, nullptr, nullptr, nullptr};
+
+  extern __shared__ alignment_dummy __shm[];
+  shared_allocator al((int *)&__shm[0]);
+
+  shared_tile_A(&As)[SUPER_M][PIPE_STAGES] =
+      al.allocate<shared_tile_A, SUPER_M, PIPE_STAGES>();
+  shared_tile_B(&Bs)[SUPER_N][PIPE_STAGES] =
+      al.allocate<shared_tile_B, SUPER_N, PIPE_STAGES>();
+
+  reg_tile_A A_reg;
+  reg_tile_B_col B_reg_col;
+  reg_tile_C C_accum;
+
+  int warpid = kittens::warpid();
+  int warp_m = warpid % SUPER_M;
+  int warp_n = warpid / SUPER_M;
+
+  int load_group_id = warpgroup::groupid();
+
+  int block_row = blockIdx.y * SUPER_M;
+  int block_col = blockIdx.x * SUPER_N;
+
+  warp::zero(C_accum);
+  int num_tiles = (g_N + BLOCK_K - 1) / BLOCK_K;
+
+  for (int load_tile = 0; load_tile < (PIPE_STAGES - 1); load_tile++) {
+    if (load_tile < num_tiles) {
+      int load_smem_idx = load_tile % PIPE_STAGES;
+      for (int task_id = warpid; task_id < LOAD_TASKS; task_id += NUM_WORKERS) {
+        if (task_id < SUPER_M) {
+          warp::load_async(As[task_id][load_smem_idx], g_A, {0, 0, block_row + task_id, load_tile});
+        } else {
+          int n_index = task_id - SUPER_M;
+          warp::load_async(Bs[n_index][load_smem_idx], g_B, {0, 0, load_tile, block_col + n_index});
+        }
+      }
+    }
+  }
+
+  for (int tile = 0; tile < num_tiles; tile++) {
+    int compute_smem_idx = tile % PIPE_STAGES;
+
+    int load_tile = tile + PIPE_STAGES - 1;
+    int load_smem_idx = load_tile % PIPE_STAGES;
+
+    if (load_tile < num_tiles) {
+      for (int task_id = warpid; task_id < LOAD_TASKS; task_id += NUM_WORKERS) {
+        if (task_id < SUPER_M) {
+          warp::load_async(As[task_id][load_smem_idx], g_A,
+                           {0, 0, block_row + task_id, load_tile});
+        } else {
+          int n_index = task_id - SUPER_M;
+          warp::load_async(Bs[n_index][load_smem_idx], g_B,
+                           {0, 0, load_tile, block_col + n_index});
+        }
+      }
+      load_async_wait<1>();
+    } else
+      load_async_wait();
+    __syncthreads();
+
+    warp::load(A_reg, As[warp_m][compute_smem_idx]);
+    warp::load(B_reg_col, Bs[warp_n][compute_smem_idx]);
+
+    warp::mma_AB(C_accum, A_reg, B_reg_col, C_accum);
+    __syncthreads();
+  }
+  warp::store(g_C, C_accum, {0, 0, block_row + warp_m, block_col + warp_n});
+}

extra/tinyfs/fetch_file.py

@@ -3,9 +3,9 @@ import argparse
 
 if __name__ == "__main__":
   parser = argparse.ArgumentParser()
-  parser.add_argument("hash", type=str, required=True, help="file hash to fetch")
-  parser.add_argument("len", type=int, required=True, help="file length to fetch")
-  parser.add_argument("dest", type=str, required=True, help="destination path to save the file")
+  parser.add_argument("--hash", type=str, required=True, help="file hash to fetch")
+  parser.add_argument("--len", type=int, required=True, help="file length to fetch")
+  parser.add_argument("--dest", type=str, required=True, help="destination path to save the file")
   args = parser.parse_args()
 
   Tensor(bytes.fromhex(args.hash), device="CPU").load(args.len).to(f"disk:{args.dest}").realize()

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',
@@ -52,7 +53,7 @@ setup(name='tinygrad',
         "License :: OSI Approved :: MIT License"
       ],
       install_requires=[],
-      python_requires='>=3.10',
+      python_requires='>=3.11',
       extras_require={
         'arm': ["unicorn"],
         'triton': ["triton-nightly>=2.1.0.dev20231014192330"],

test/external/external_benchmark_op_conv.py

@@ -0,0 +1,260 @@
+# ruff: noqa: E501
+from tinygrad import dtypes, Device
+from tinygrad.uop.ops import UOp, AxisType, Ops
+from tinygrad.codegen import full_rewrite
+from tinygrad.renderer import ProgramSpec
+from tinygrad.engine.realize import CompiledRunner
+from tinygrad.helpers import dedup
+from tinygrad.device import Buffer
+from tinygrad.dtype import ImageDType
+
+# PYTHONPATH="." DEBUG=5 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
+# kernel 672
+# faster on d59d4cd, 50% slower with the new linearizer
+
+""" d59d4cd
+c0 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((32, 1024, 4)), arg=0, src=())
+c1 = UOp.range(UOp.const(dtypes.index, 64), 3, AxisType.LOOP)
+c2 = UOp.range(UOp.const(dtypes.index, 64), 4, AxisType.LOOP)
+c3 = UOp.range(UOp.const(dtypes.index, 32), 2, AxisType.LOOP)
+c4 = (((c1*UOp.const(dtypes.index, 64))+c2)+(c3*UOp.const(dtypes.index, 4096)))
+c5 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((32, 1024, 4)), arg=1, src=())
+c6 = c5.index(c4).load()
+c7 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((32, 3072, 4)), arg=2, src=())
+c8 = UOp.range(UOp.const(dtypes.index, 48), 0, AxisType.REDUCE)
+c9 = UOp.range(UOp.const(dtypes.index, 4), 1, AxisType.REDUCE)
+c10 = c7.index(((((c8*UOp.const(dtypes.index, 4))+c9)+(c1*UOp.const(dtypes.index, 192)))+(c3*UOp.const(dtypes.index, 12288)))).load()
+c11 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((16, 192, 4)), arg=3, src=())
+c12 = c11.index(((((c9*UOp.const(dtypes.index, 4))+(c2%UOp.const(dtypes.index, 4)))+(c8*UOp.const(dtypes.index, 16)))+((c2//UOp.const(dtypes.index, 4))*UOp.const(dtypes.index, 768)))).load()
+c13 = UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(64), arg=4, src=())
+c14 = c13.index(c2).load()
+c15 = UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(64), arg=5, src=())
+c16 = c15.index(c2).load()
+c17 = (c6+(((c10*c12.cast(dtypes.float)).cast(dtypes.float).reduce(c8, c9, arg=Ops.ADD)+c14.cast(dtypes.float))*c16.cast(dtypes.float)))
+c18 = c0.index(c4).store(c17, c3, c1, c2)
+ast = c18.sink()
+more upcast axis : [(3, 320, 0, 4)]
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+__kernel void r_512_16_4_4_48_4(write_only image2d_t data0_131072, read_only image2d_t data1_131072, read_only image2d_t data2_393216, read_only image2d_t data3_12288, __global half* data4_64, __global half* data5_64) {
+const sampler_t smp = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAMP | CLK_FILTER_NEAREST;
+  float acc0[16];
+  int idx0 = get_global_id(0); /* 16 */
+  int idx1 = get_global_id(1); /* 512 */
+  int alu0 = (idx1>>4);
+  *(acc0+0) = 0.0f;
+  *(acc0+1) = 0.0f;
+  *(acc0+2) = 0.0f;
+  *(acc0+3) = 0.0f;
+  *(acc0+4) = 0.0f;
+  *(acc0+5) = 0.0f;
+  *(acc0+6) = 0.0f;
+  *(acc0+7) = 0.0f;
+  *(acc0+8) = 0.0f;
+  *(acc0+9) = 0.0f;
+  *(acc0+10) = 0.0f;
+  *(acc0+11) = 0.0f;
+  *(acc0+12) = 0.0f;
+  *(acc0+13) = 0.0f;
+  *(acc0+14) = 0.0f;
+  *(acc0+15) = 0.0f;
+  for (int Ridx0 = 0; Ridx0 < 48; Ridx0++) {
+    int alu17 = ((idx1*192)+Ridx0);
+    int alu18 = (alu17+48);
+    int alu19 = (alu17+96);
+    int alu20 = (alu17+144);
+    int alu21 = (Ridx0<<2);
+    float4 val0 = read_imagef(data3_12288, smp, (int2)(alu21,idx0));
+    float4 val1 = read_imagef(data3_12288, smp, (int2)((alu21+1),idx0));
+    float4 val2 = read_imagef(data3_12288, smp, (int2)((alu21+2),idx0));
+    float4 val3 = read_imagef(data3_12288, smp, (int2)((alu21+3),idx0));
+    float4 val4 = read_imagef(data2_393216, smp, (int2)((alu18-(3072*(((alu18>>10)*43)>>7))),alu0));
+    float4 val5 = read_imagef(data2_393216, smp, (int2)((alu19-(3072*(((alu19>>10)*43)>>7))),alu0));
+    float4 val6 = read_imagef(data2_393216, smp, (int2)((alu20-(3072*(((alu20>>10)*43)>>7))),alu0));
+    float4 val7 = read_imagef(data2_393216, smp, (int2)((alu17-(3072*(((alu17>>10)*43)>>7))),alu0));
+    *(acc0+1) = ((*(acc0+1))+(val4.x*val0.x)+(val4.y*val1.x)+(val4.z*val2.x)+(val4.w*val3.x));
+    *(acc0+5) = ((*(acc0+5))+(val4.x*val0.y)+(val4.y*val1.y)+(val4.z*val2.y)+(val4.w*val3.y));
+    *(acc0+9) = ((*(acc0+9))+(val4.x*val0.z)+(val4.y*val1.z)+(val4.z*val2.z)+(val4.w*val3.z));
+    *(acc0+13) = ((*(acc0+13))+(val4.x*val0.w)+(val4.y*val1.w)+(val4.z*val2.w)+(val4.w*val3.w));
+    *(acc0+2) = ((*(acc0+2))+(val5.x*val0.x)+(val5.y*val1.x)+(val5.z*val2.x)+(val5.w*val3.x));
+    *(acc0+6) = ((*(acc0+6))+(val5.x*val0.y)+(val5.y*val1.y)+(val5.z*val2.y)+(val5.w*val3.y));
+    *(acc0+10) = ((*(acc0+10))+(val5.x*val0.z)+(val5.y*val1.z)+(val5.z*val2.z)+(val5.w*val3.z));
+    *(acc0+14) = ((*(acc0+14))+(val5.x*val0.w)+(val5.y*val1.w)+(val5.z*val2.w)+(val5.w*val3.w));
+    *(acc0+3) = ((*(acc0+3))+(val6.x*val0.x)+(val6.y*val1.x)+(val6.z*val2.x)+(val6.w*val3.x));
+    *(acc0+7) = ((*(acc0+7))+(val6.x*val0.y)+(val6.y*val1.y)+(val6.z*val2.y)+(val6.w*val3.y));
+    *(acc0+11) = ((*(acc0+11))+(val6.x*val0.z)+(val6.y*val1.z)+(val6.z*val2.z)+(val6.w*val3.z));
+    *(acc0+15) = ((*(acc0+15))+(val6.x*val0.w)+(val6.y*val1.w)+(val6.z*val2.w)+(val6.w*val3.w));
+    *(acc0+0) = ((*(acc0+0))+(val7.x*val0.x)+(val7.y*val1.x)+(val7.z*val2.x)+(val7.w*val3.x));
+    *(acc0+4) = ((*(acc0+4))+(val7.x*val0.y)+(val7.y*val1.y)+(val7.z*val2.y)+(val7.w*val3.y));
+    *(acc0+8) = ((*(acc0+8))+(val7.x*val0.z)+(val7.y*val1.z)+(val7.z*val2.z)+(val7.w*val3.z));
+    *(acc0+12) = ((*(acc0+12))+(val7.x*val0.w)+(val7.y*val1.w)+(val7.z*val2.w)+(val7.w*val3.w));
+  }
+  int alu39 = (idx0<<2);
+  half4 val8 = (*((__global half4*)((data4_64+alu39))));
+  half4 val9 = (*((__global half4*)((data5_64+alu39))));
+  int alu40 = (idx0+(idx1<<6));
+  int2 cast0 = (int2)((alu40&1023),alu0);
+  float4 val10 = read_imagef(data1_131072, smp, cast0);
+  int2 cast1 = (int2)(((alu40+16)&1023),alu0);
+  float4 val11 = read_imagef(data1_131072, smp, cast1);
+  int2 cast2 = (int2)(((alu40+32)&1023),alu0);
+  float4 val12 = read_imagef(data1_131072, smp, cast2);
+  int2 cast3 = (int2)(((alu40+48)&1023),alu0);
+  float4 val13 = read_imagef(data1_131072, smp, cast3);
+  float cast4 = ((float)(val8.x));
+  float cast5 = ((float)(val9.x));
+  float cast6 = ((float)(val8.y));
+  float cast7 = ((float)(val9.y));
+  float cast8 = ((float)(val8.z));
+  float cast9 = ((float)(val9.z));
+  float cast10 = ((float)(val8.w));
+  float cast11 = ((float)(val9.w));
+  write_imagef(data0_131072, cast0, (float4)((val10.x+(((*(acc0+0))+cast4)*cast5)),(val10.y+(((*(acc0+4))+cast6)*cast7)),(val10.z+(((*(acc0+8))+cast8)*cast9)),(val10.w+(((*(acc0+12))+cast10)*cast11))));
+  write_imagef(data0_131072, cast1, (float4)((val11.x+(((*(acc0+1))+cast4)*cast5)),(val11.y+(((*(acc0+5))+cast6)*cast7)),(val11.z+(((*(acc0+9))+cast8)*cast9)),(val11.w+(((*(acc0+13))+cast10)*cast11))));
+  write_imagef(data0_131072, cast2, (float4)((val12.x+(((*(acc0+2))+cast4)*cast5)),(val12.y+(((*(acc0+6))+cast6)*cast7)),(val12.z+(((*(acc0+10))+cast8)*cast9)),(val12.w+(((*(acc0+14))+cast10)*cast11))));
+  write_imagef(data0_131072, cast3, (float4)((val13.x+(((*(acc0+3))+cast4)*cast5)),(val13.y+(((*(acc0+7))+cast6)*cast7)),(val13.z+(((*(acc0+11))+cast8)*cast9)),(val13.w+(((*(acc0+15))+cast10)*cast11))));
+}
+*** QCOM     672 r_512_16_4_4_48_4                              arg  6 mem   0.10 GB tm    322.55us/    77.83ms (    157 GFLOPS    4|160    GB/s) ['mul', '__add__', 'conv2d']
+"""
+
+""" master 99e76f33a0f4ec84c79c1271dbc955fe6b5a7778
+c0 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((32, 1024, 4)), (), 0)
+c2 = UOp.range(64, 3, AxisType.LOOP)
+c4 = UOp.range(64, 4, AxisType.LOOP)
+c7 = UOp.range(32, 2, AxisType.LOOP)
+c10 = (((c2*64)+c4)+(c7*4096))
+c12 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((32, 1024, 4)), (), 1)
+c14 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((32, 3072, 4)), (), 2)
+c16 = UOp.range(48, 0, AxisType.REDUCE)
+c19 = UOp.range(4, 1, AxisType.REDUCE)
+c28 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((16, 192, 4)), (), 3)
+c40 = (c14.index(((((c16*4)+c19)+(c2*192))+(c7*12288)))*c28.index(((((c19*4)+(c4%4))+(c16*16))+((c4//4)*768))))
+c42 = UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(64), (), 4)
+c46 = UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(64), (), 5)
+c50 = (c12.index(c10)+((c40.reduce(c16, c19, arg=Ops.ADD)+c42.index(c4).cast(dtypes.float))*c46.index(c4).cast(dtypes.float)))
+c52 = c0.index(c10, ptr=True).store(c50).end(c7, c2, c4)
+ast = c52.sink()
+more upcast axis : [(3, 320, 0, 4)]
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+__kernel void r_512_16_4_4_48_4(write_only image2d_t data0_131072, read_only image2d_t data1_131072, read_only image2d_t data2_393216, read_only image2d_t data3_12288, __global half* data4_64, __global half* data5_64) {
+const sampler_t smp = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAMP | CLK_FILTER_NEAREST;
+  float acc0[16];
+  int idx0 = get_global_id(0); /* 16 */
+  int idx1 = get_global_id(1); /* 512 */
+  *(acc0+0) = 0.0f;
+  *(acc0+1) = 0.0f;
+  *(acc0+2) = 0.0f;
+  *(acc0+3) = 0.0f;
+  *(acc0+4) = 0.0f;
+  *(acc0+5) = 0.0f;
+  *(acc0+6) = 0.0f;
+  *(acc0+7) = 0.0f;
+  *(acc0+8) = 0.0f;
+  *(acc0+9) = 0.0f;
+  *(acc0+10) = 0.0f;
+  *(acc0+11) = 0.0f;
+  *(acc0+12) = 0.0f;
+  *(acc0+13) = 0.0f;
+  *(acc0+14) = 0.0f;
+  *(acc0+15) = 0.0f;
+  int alu16 = (idx0<<2);
+  half4 val0 = (*((__global half4*)((data4_64+alu16))));
+  half4 val1 = (*((__global half4*)((data5_64+alu16))));
+  int alu17 = (idx0+(idx1<<6));
+  int alu18 = (idx1>>4);
+  int2 cast0 = (int2)((alu17&1023),alu18);
+  float4 val2 = read_imagef(data1_131072, smp, cast0);
+  int2 cast1 = (int2)(((alu17+16)&1023),alu18);
+  float4 val3 = read_imagef(data1_131072, smp, cast1);
+  int2 cast2 = (int2)(((alu17+32)&1023),alu18);
+  float4 val4 = read_imagef(data1_131072, smp, cast2);
+  int2 cast3 = (int2)(((alu17+48)&1023),alu18);
+  float4 val5 = read_imagef(data1_131072, smp, cast3);
+  for (int Ridx0 = 0; Ridx0 < 48; Ridx0++) {
+    int alu19 = ((idx1*192)+Ridx0);
+    int alu20 = (alu19+48);
+    int alu21 = (alu19+96);
+    int alu22 = (alu19+144);
+    int alu23 = (Ridx0<<2);
+    float4 val6 = read_imagef(data3_12288, smp, (int2)(alu23,idx0));
+    float4 val7 = read_imagef(data3_12288, smp, (int2)((alu23+1),idx0));
+    float4 val8 = read_imagef(data3_12288, smp, (int2)((alu23+2),idx0));
+    float4 val9 = read_imagef(data3_12288, smp, (int2)((alu23+3),idx0));
+    float4 val10 = read_imagef(data2_393216, smp, (int2)((alu20-(3072*(((alu20>>10)*43)>>7))),alu18));
+    *(acc0+1) = ((*(acc0+1))+(val10.x*val6.x)+(val10.y*val7.x)+(val10.z*val8.x)+(val10.w*val9.x));
+    *(acc0+5) = ((*(acc0+5))+(val10.x*val6.y)+(val10.y*val7.y)+(val10.z*val8.y)+(val10.w*val9.y));
+    *(acc0+9) = ((*(acc0+9))+(val10.x*val6.z)+(val10.y*val7.z)+(val10.z*val8.z)+(val10.w*val9.z));
+    *(acc0+13) = ((*(acc0+13))+(val10.x*val6.w)+(val10.y*val7.w)+(val10.z*val8.w)+(val10.w*val9.w));
+    float4 val11 = read_imagef(data2_393216, smp, (int2)((alu21-(3072*(((alu21>>10)*43)>>7))),alu18));
+    *(acc0+2) = ((*(acc0+2))+(val11.x*val6.x)+(val11.y*val7.x)+(val11.z*val8.x)+(val11.w*val9.x));
+    *(acc0+6) = ((*(acc0+6))+(val11.x*val6.y)+(val11.y*val7.y)+(val11.z*val8.y)+(val11.w*val9.y));
+    *(acc0+10) = ((*(acc0+10))+(val11.x*val6.z)+(val11.y*val7.z)+(val11.z*val8.z)+(val11.w*val9.z));
+    *(acc0+14) = ((*(acc0+14))+(val11.x*val6.w)+(val11.y*val7.w)+(val11.z*val8.w)+(val11.w*val9.w));
+    float4 val12 = read_imagef(data2_393216, smp, (int2)((alu22-(3072*(((alu22>>10)*43)>>7))),alu18));
+    *(acc0+3) = ((*(acc0+3))+(val12.x*val6.x)+(val12.y*val7.x)+(val12.z*val8.x)+(val12.w*val9.x));
+    *(acc0+7) = ((*(acc0+7))+(val12.x*val6.y)+(val12.y*val7.y)+(val12.z*val8.y)+(val12.w*val9.y));
+    *(acc0+11) = ((*(acc0+11))+(val12.x*val6.z)+(val12.y*val7.z)+(val12.z*val8.z)+(val12.w*val9.z));
+    *(acc0+15) = ((*(acc0+15))+(val12.x*val6.w)+(val12.y*val7.w)+(val12.z*val8.w)+(val12.w*val9.w));
+    float4 val13 = read_imagef(data2_393216, smp, (int2)((alu19-(3072*(((alu19>>10)*43)>>7))),alu18));
+    *(acc0+0) = ((*(acc0+0))+(val13.x*val6.x)+(val13.y*val7.x)+(val13.z*val8.x)+(val13.w*val9.x));
+    *(acc0+4) = ((*(acc0+4))+(val13.x*val6.y)+(val13.y*val7.y)+(val13.z*val8.y)+(val13.w*val9.y));
+    *(acc0+8) = ((*(acc0+8))+(val13.x*val6.z)+(val13.y*val7.z)+(val13.z*val8.z)+(val13.w*val9.z));
+    *(acc0+12) = ((*(acc0+12))+(val13.x*val6.w)+(val13.y*val7.w)+(val13.z*val8.w)+(val13.w*val9.w));
+  }
+  float cast4 = ((float)(val0.x));
+  float cast5 = ((float)(val1.x));
+  float cast6 = ((float)(val0.y));
+  float cast7 = ((float)(val1.y));
+  float cast8 = ((float)(val0.z));
+  float cast9 = ((float)(val1.z));
+  float cast10 = ((float)(val0.w));
+  float cast11 = ((float)(val1.w));
+  write_imagef(data0_131072, cast0, (float4)((val2.x+(((*(acc0+0))+cast4)*cast5)),(val2.y+(((*(acc0+4))+cast6)*cast7)),(val2.z+(((*(acc0+8))+cast8)*cast9)),(val2.w+(((*(acc0+12))+cast10)*cast11))));
+  write_imagef(data0_131072, cast1, (float4)((val3.x+(((*(acc0+1))+cast4)*cast5)),(val3.y+(((*(acc0+5))+cast6)*cast7)),(val3.z+(((*(acc0+9))+cast8)*cast9)),(val3.w+(((*(acc0+13))+cast10)*cast11))));
+  write_imagef(data0_131072, cast2, (float4)((val4.x+(((*(acc0+2))+cast4)*cast5)),(val4.y+(((*(acc0+6))+cast6)*cast7)),(val4.z+(((*(acc0+10))+cast8)*cast9)),(val4.w+(((*(acc0+14))+cast10)*cast11))));
+  write_imagef(data0_131072, cast3, (float4)((val5.x+(((*(acc0+3))+cast4)*cast5)),(val5.y+(((*(acc0+7))+cast6)*cast7)),(val5.z+(((*(acc0+11))+cast8)*cast9)),(val5.w+(((*(acc0+15))+cast10)*cast11))));
+}
+*** QCOM     672 r_512_16_4_4_48_4                              arg  6 mem   0.10 GB tm    527.97us/    78.94ms (     96 GFLOPS    3|98     GB/s) ['conv2d', 'mul', '__add__']
+"""
+
+c0 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((32, 1024, 4)), (), 0)
+c2 = UOp.range(64, 3, AxisType.LOOP)
+c4 = UOp.range(64, 4, AxisType.LOOP)
+c7 = UOp.range(32, 2, AxisType.LOOP)
+c10 = (((c2*64)+c4)+(c7*4096))
+c12 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((32, 1024, 4)), (), 1)
+c14 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((32, 3072, 4)), (), 2)
+c16 = UOp.range(48, 0, AxisType.REDUCE)
+c19 = UOp.range(4, 1, AxisType.REDUCE)
+c28 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((16, 192, 4)), (), 3)
+c40 = (c14.index(((((c16*4)+c19)+(c2*192))+(c7*12288)))*c28.index(((((c19*4)+(c4%4))+(c16*16))+((c4//4)*768))))
+c42 = UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(64), (), 4)
+c46 = UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(64), (), 5)
+c50 = (c12.index(c10)+((c40.reduce(c16, c19, arg=Ops.ADD)+c42.index(c4).cast(dtypes.float))*c46.index(c4).cast(dtypes.float)))
+c52 = c0.index(c10, ptr=True).store(c50).end(c7, c2, c4)
+ast = c52.sink()
+
+compiler = Device.default.compiler
+renderer = Device.default.renderer
+allocator = Device.default.allocator
+
+uops = full_rewrite(ast, renderer)
+src = renderer.render(uops)
+
+# NOLOCALS=1 IMAGE=2 DEV=CL
+lib = compiler.compile(src)
+# r_64_8_16_4_4_48_4
+# NOLOCALS: r_512_16_4_4_48_4
+ps = ProgramSpec("r_512_16_4_4_48_4", src, Device.DEFAULT, ast, uops)
+print(ps.src)
+print(ps.applied_opts)
+# (Opt(op=OptOps.UNROLL, axis=0, arg=4), Opt(op=OptOps.UPCAST, axis=1, arg=4), Opt(op=OptOps.UPCAST, axis=0, arg=4), Opt(op=OptOps.NOLOCALS, axis=None, arg=None))
+cr = CompiledRunner(ps, precompiled=lib)
+
+gs = sorted(dedup([u for u in ast.toposort() if u.op is Ops.DEFINE_GLOBAL]), key=lambda u: u.arg)
+print(len(gs))
+print([g.dtype for g in gs])
+
+bufs = [Buffer(ps.device, g.size, g.dtype if isinstance(g.dtype, ImageDType) else g.dtype._base).ensure_allocated() for g in gs]
+
+t = cr(bufs, wait=True)
+print(f"{t*1e6:.2f} us")

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_nhwc_conv.py

@@ -0,0 +1,38 @@
+from tinygrad import Tensor, nn, Context, GlobalCounters
+
+if __name__ == "__main__":
+  conv = nn.Conv2d(64, 128, 3)
+  img = Tensor.randn((1,64,128,128))
+  with Context(DEBUG=0, BEAM=0):
+    Tensor.realize(img, conv.weight, conv.bias)
+
+  tst = conv(img).permute(0,2,3,1).realize()
+  print(tst.shape)
+
+  print("NEW")
+  img_perm = img.permute(0,2,3,1).contiguous()
+  print(img_perm.shape)
+  pp = img_perm.permute(0,3,1,2)._pool((3,3)).permute(0,2,3,4,5,1)
+
+  def hwio(pp, conv):
+    pp = pp.unsqueeze(-1)
+    weight = conv.weight.permute(2,3,1,0).contiguous()
+    print(pp.shape, weight.shape, (pp*weight).shape)
+    return (pp * weight).sum([-4,-3, -2])
+
+  def ohwi(pp, conv):
+    pp = pp.unsqueeze(-4)
+    weight = conv.weight.permute(0,2,3,1).contiguous()
+    print(pp.shape, weight.shape, (pp*weight).shape)
+    return (pp * weight).sum([-3,-2,-1])
+
+  for f in [hwio, ohwi]:
+    GlobalCounters.reset()
+    print("\n**************", f.__name__, "**************")
+    out = f(pp, conv)
+    out.realize()
+    print(out.shape)
+
+    with Context(DEBUG=0, BEAM=0):
+      err = (tst-out).square()
+      print(err.mean().item(), err.max().item())

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/process_replay/process_replay.py

@@ -13,7 +13,7 @@ try:
   from tinygrad.engine.realize import get_program
   from tinygrad.uop.ops import UOp, Ops, KernelInfo
   from tinygrad.codegen.opt import Opt
-  from tinygrad.helpers import VERSION, Context, ContextVar, colored, db_connection, getenv, tqdm
+  from tinygrad.helpers import VERSION, Context, ContextVar, colored, db_connection, getenv, tqdm, BEAM
   from tinygrad.device import Device
 except ImportError as e:
   print(repr(e))
@@ -51,8 +51,8 @@ def replay_get_rangeify_map(ret:dict[UOp, UOp], big_sink:UOp) -> tuple[str, str,
   return to_str(new_sink), to_str(big_sink.substitute(ret)), (big_sink,)
 
 def replay_get_program(p:ProgramSpec, ast:UOp, renderer:Renderer|None=None, opts:list[Opt]|None=None) -> tuple[str, str, tuple[Any, ...]]:
-  # NOTE: this always uses the opts_to_apply path
-  sink_arg = ast.arg or KernelInfo(opts_to_apply=p.applied_opts)
+  # the ast.arg is non None if we are inside of search.py
+  sink_arg = ast.arg or KernelInfo(opts_to_apply=tuple(opts) if opts is not None else p.applied_opts if BEAM>=1 else None)
   input_ast = ast.replace(arg=replace(sink_arg, name=p.name))
   # if no renderer was provided, open the device to get it
   if renderer is None: renderer = Device[p.device].renderer

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/mockgpu/amd/amddriver.py

@@ -85,6 +85,8 @@ class AMDDriver(VirtDriver):
       VirtFile(f'/sys/devices/virtual/kfd/kfd/topology/nodes/{gpu_id}/gpu_id', functools.partial(TextFileDesc, text=f"{gpu_id}")),
       VirtFile(f'/sys/devices/virtual/kfd/kfd/topology/nodes/{gpu_id}/properties',
         functools.partial(TextFileDesc, text=gpu_props.format(drm_render_minor=gpu_id))),
+      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/power_dpm_force_performance_level',
+               functools.partial(TextFileDesc, text='profile_standard\n')),
       VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0',
                functools.partial(DirFileDesc, child_names=[str(am.GC_HWID), str(am.SDMA0_HWID), str(am.NBIF_HWID)])),
       VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.GC_HWID}', functools.partial(DirFileDesc, child_names=['0'])),

test/mockgpu/amd/amdgpu.py

@@ -14,6 +14,9 @@ regSQ_THREAD_TRACE_BUF0_BASE = 0x39e8 + amd_gpu.GC_BASE__INST0_SEG1
 regSQ_THREAD_TRACE_BUF0_SIZE = 0x39e9 + amd_gpu.GC_BASE__INST0_SEG1
 regSQ_THREAD_TRACE_WPTR = 0x39ef + amd_gpu.GC_BASE__INST0_SEG1
 regSQ_THREAD_TRACE_STATUS = 0x39f4 + amd_gpu.GC_BASE__INST0_SEG1
+regCP_PERFMON_CNTL = 0x3808 + amd_gpu.GC_BASE__INST0_SEG1
+regCPG_PERFCOUNTER1_LO = 0x3000 + amd_gpu.GC_BASE__INST0_SEG1
+regGUS_PERFCOUNTER_HI = 0x3643 + amd_gpu.GC_BASE__INST0_SEG1
 
 class SQTT_EVENTS:
   THREAD_TRACE_FINISH = 0x00000037
@@ -130,7 +133,7 @@ class PM4Executor(AMDQueue):
     _src_addr_hi = self._next_dword()
     dst_addr_lo = self._next_dword()
     dst_addr_hi = self._next_dword()
-    assert copy_data_flags == 0x100204, hex(copy_data_flags) # better fail than silently do the wrong thing
+    assert copy_data_flags in {0x100204, 0x000204}, hex(copy_data_flags) # better fail than silently do the wrong thing
     to_mv(dst_addr_hi<<32|dst_addr_lo, 4).cast('I')[0] = self.gpu.regs[src_addr_lo]
 
   def _exec_wait_reg_mem(self, n):
@@ -280,6 +283,9 @@ class AMDGPURegisters:
     self.regs: dict[tuple[int, int], int] = {}
   def __getitem__(self, addr:int) -> int:
     if addr == regGRBM_GFX_INDEX: return self.grbm_index
+    if regCPG_PERFCOUNTER1_LO < addr < regGUS_PERFCOUNTER_HI:
+      assert self.regs[(regCP_PERFMON_CNTL, 0)] == 0x401, "read mode should be enabled"
+      return addr << 16 | self.grbm_index
     return self.regs[(addr, getbits(self.grbm_index, 16, 23))]
   def __setitem__(self, addr:int, val:int):
     if addr == regGRBM_GFX_INDEX: self.grbm_index = val

test/models/test_real_world.py

@@ -112,7 +112,7 @@ class TestRealWorld(unittest.TestCase):
         loss.backward()
         optimizer.step()
 
-      helper_test("train_mnist", lambda: (Tensor.randn(BS, 1, 28, 28),), train, 0.07, 102)
+      helper_test("train_mnist", lambda: (Tensor.randn(BS, 1, 28, 28),), train, 0.07, 103)
 
   @unittest.skipIf(CI and Device.DEFAULT in {"CPU", "CL"}, "slow")
   def test_forward_cifar(self):
@@ -176,7 +176,7 @@ class TestRealWorld(unittest.TestCase):
       for v in data.values(): v.to_(Device.DEFAULT)
 
       helper_test("train_bert", lambda: (data["input_ids"], data["segment_ids"], data["input_mask"], data["masked_lm_positions"], \
-          data["masked_lm_ids"], data["masked_lm_weights"], data["next_sentence_labels"]), train, 0.31, 358)
+          data["masked_lm_ids"], data["masked_lm_weights"], data["next_sentence_labels"]), train, 0.31, 427)
 
 if __name__ == '__main__':
   unittest.main()

test/opt/test_tensor_cores.py

@@ -14,6 +14,8 @@ from tinygrad.codegen.opt import Opt, OptOps, KernelOptError
 # TODO: write a clean version of this
 from test.test_linearizer import helper_realized_ast, helper_linearizer_opt
 
+# NOTE: get_program always passes in Device[Device.DEFAULT].renderer explicitly for process_replay!!!
+
 def helper_tc_ensure_uops_and_opts_count(N: int, M:int, K:int, dtype_in:DType, dtype_out:DType, axis:int=0, tc_select:int=-1, tc_opt:int=0,
                                          ensure_triggered:bool=True):
   a, b = Tensor.rand(M, K, dtype=dtype_in), Tensor.rand(K, N, dtype=dtype_in)
@@ -41,7 +43,7 @@ def helper_tc_allclose(N:int, M:int, K:int, dtype_in:DType, dtype_out:DType, axi
   if dtype_in == dtypes.bfloat16: r = r.float()
   realized_ast, bufs = helper_realized_ast(r)
   opts = [Opt(op=OptOps.TC, axis=axis, arg=(tc_select, tc_opt, use_tensor_cores))]
-  prg = CompiledRunner(replace(get_program(realized_ast, opts=opts), device=Device.DEFAULT))
+  prg = CompiledRunner(replace(get_program(realized_ast, Device[Device.DEFAULT].renderer, opts=opts), device=Device.DEFAULT))
   if use_tensor_cores == 1: assert len([uop for uop in prg.p.uops if uop.op is Ops.WMMA]) > 0, "wmma not triggered"
   assert len([x for x in prg.p.uops[-1].arg.applied_opts if x.op is OptOps.TC]) == 1, "tensor core opt not included"
   prg.exec(bufs)
@@ -68,7 +70,7 @@ class TestTensorCores(unittest.TestCase):
       n, m, k = tc.dims[0], tc.dims[1], 2 if AMX else tc.dims[2]
       a, b = Tensor.rand(m, k, dtype=tc.dtype_in), Tensor.rand(k, n, dtype=tc.dtype_in)
       r = a.matmul(b, dtype=tc.dtype_out)
-      prg = get_program(r.schedule()[-1].ast, opts=[Opt(op=OptOps.TC, axis=0, arg=(-1, 2, 1))])
+      prg = get_program(r.schedule()[-1].ast, Device[Device.DEFAULT].renderer, opts=[Opt(op=OptOps.TC, axis=0, arg=(-1, 2, 1))])
       if Device.DEFAULT == "CPU" and CPU_LLVM:
         assert "0x201000" in prg.src
       elif Device.DEFAULT == "AMD" and AMD_LLVM:
@@ -154,7 +156,7 @@ class TestTensorCores(unittest.TestCase):
     r = x.matmul(y, dtype=tc.dtype_out)
     opts = [Opt(OptOps.UNROLL, 0, 4)]
     ast = helper_linearizer_opt(r, [opts], apply_tc=True, atol=3e-2, rtol=1e-3)
-    for u in get_program(ast, opts=opts).uops:
+    for u in get_program(ast, Device[Device.DEFAULT].renderer, opts=opts).uops:
       if u.op is Ops.WMMA:
         assert u.src[-1].src[0].op != Ops.STORE
 
@@ -167,7 +169,7 @@ class TestTensorCores(unittest.TestCase):
     r = x.matmul(y, dtype=tc.dtype_out)
     opts = [Opt(OptOps.UNROLL, 0, 4)]
     ast = helper_linearizer_opt(r, [opts], apply_tc=True, atol=3e-2, rtol=1e-3)
-    for u in get_program(ast, opts=opts).uops:
+    for u in get_program(ast, Device[Device.DEFAULT].renderer, opts=opts).uops:
       if u.op is Ops.WMMA:
         #assert u.src[-1].dtype == dtypes.float.vec(prod(tc.thread_local_sizes[2]))
         assert u.src[-1].src[0].op != Ops.STORE
@@ -182,7 +184,7 @@ class TestTensorCores(unittest.TestCase):
     r = x.matmul(y, dtype=tc.dtype_out).relu()
     opts = [Opt(OptOps.UNROLL, 0, 4)]
     ast = helper_linearizer_opt(r, [opts], apply_tc=True, atol=3e-2, rtol=1e-3)
-    for u in get_program(ast, opts=opts).uops:
+    for u in get_program(ast, Device[Device.DEFAULT].renderer, opts=opts).uops:
       if u.op is Ops.WMMA:
         #assert u.src[-1].dtype == dtypes.float.vec(prod(tc.thread_local_sizes[2]))
         assert u.src[-1].src[0].op != Ops.STORE

test/test_const_folding.py

@@ -1,5 +1,5 @@
 import unittest, itertools, math
-from tinygrad import Tensor, Device, dtypes
+from tinygrad import Tensor, Device, dtypes, Context
 from tinygrad.dtype import DType, ConstType
 from tinygrad.uop.ops import Ops, UOp
 from tinygrad.codegen import full_rewrite_to_sink
@@ -126,7 +126,8 @@ class TestBitcastConstFolding(unittest.TestCase):
     t({dtypes.int64: 4598983288165178391, dtypes.uint64: 4598983288165178391, dtypes.float64: 0.29485681936461233})
 
   def test_vec_bitcast(self):
-    r = full_rewrite_to_sink(UOp.const(dtypes.int32.vec(3), (-1, -2**31, 75)).bitcast(dtypes.uint32.vec(3)).sink()).src[0]
+    with Context(SPEC=0):
+      r = full_rewrite_to_sink(UOp.const(dtypes.int32.vec(3), (-1, -2**31, 75)).bitcast(dtypes.uint32.vec(3)).sink()).src[0]
     self.assertEqual(r.op, Ops.VECTORIZE)
     self.assertEqual(r.dtype, dtypes.uint32.vec(3))
     self.assertEqual(tuple(x.arg for x in r.src), (2**32-1, 2**31, 75))

test/test_custom_kernel.py

@@ -0,0 +1,210 @@
+import unittest
+from tinygrad import Tensor, UOp, Context
+from tinygrad.dtype import AddrSpace
+from tinygrad.uop.ops import KernelInfo, AxisType
+
+# **** kernels ****
+
+def custom_arange_kernel(C:UOp) -> UOp:
+  i = UOp.range(C.size, 0)
+  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])
+  store_d = D[i].store(A[i]*B[i])
+  return UOp.group(store_c, store_d).end(i).sink(arg=KernelInfo(name=f"custom_addmul_kernel_{C.size}")).simplify()
+
+def custom_gemm(C:UOp, A:UOp, B:UOp) -> UOp:
+  assert A.shape[1] == B.shape[0]
+  i, j, k = UOp.range(C.shape[0], 0), UOp.range(C.shape[1], 1), UOp.range(A.shape[1], 2, axis_type=AxisType.REDUCE)
+  C = C[i, j].set(0.0)
+  C = C[i, j].set(C.after(k)[i, j] + A[i, k] * B[k, j], end=k)
+  prog = C.end(i, j)
+  return prog.sink(arg=KernelInfo(name=f"custom_gemm_{C.shape[0]}_{C.shape[1]}_{A.shape[1]}", opts_to_apply=()))
+
+def custom_sum(B:UOp, A:UOp) -> UOp:
+  i = UOp.range(A.shape[0], 0, axis_type=AxisType.REDUCE)
+  B = B[0].set(0.0)
+  B = B[0].set(B.after(i)[0] + A[i], end=i)
+  return B.sink(arg=KernelInfo(name=f"custom_sum_{A.shape[0]}", opts_to_apply=()))
+
+def flip_contract_kernel(dest:UOp, src:UOp):
+  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]:
+  out, a, b = kernel.src
+  grad_a = (Tensor(gradient) @ Tensor(b).T).uop
+  grad_b = (Tensor(a).T @ Tensor(gradient)).uop
+  return (None, grad_a, grad_b)
+
+def backward_gemm_custom(gradient:UOp, kernel:UOp) -> tuple[UOp, UOp]:
+  out, a, b = kernel.src
+  grad_a = Tensor.empty_like(Tensor(a)).custom_kernel(Tensor(gradient), Tensor(b).T, fxn=custom_gemm)[0].uop
+  grad_b = Tensor.empty_like(Tensor(b)).custom_kernel(Tensor(a).T, Tensor(gradient), fxn=custom_gemm)[0].uop
+  return (None, grad_a, grad_b)
+
+# **** tests ****
+
+class TestCustomKernel(unittest.TestCase):
+  def test_simple(self):
+    a = Tensor.ones(16, 16).contiguous()
+    b = Tensor.ones(16, 16).contiguous()
+    c = Tensor.empty(16, 16)
+
+    c = Tensor.custom_kernel(c,a,b, fxn=custom_elementwise_add_kernel)[0]
+
+    out = c.flatten().tolist()
+    assert all(x == 2 for x in out), "all 2"
+
+  def test_multioutput(self):
+    a = Tensor.full((16, 16), 3.).contiguous()
+    b = Tensor.full((16, 16), 3.).contiguous()
+    c = Tensor.empty(16, 16)
+    d = Tensor.empty(16, 16)
+
+    c,d = Tensor.custom_kernel(c,d,a,b, fxn=custom_elementwise_addmul_kernel)[:2]
+    Tensor.realize(c,d)
+
+    assert all(x == 6 for x in c.flatten().tolist()), "all 6"
+    assert all(x == 9 for x in d.flatten().tolist()), "all 9"
+
+  def test_arange(self):
+    ref = Tensor.arange(100)
+    tst = Tensor.empty_like(ref)
+    tst = tst.custom_kernel(fxn=custom_arange_kernel)[0]
+    self.assertTrue((ref == tst).all().item())
+
+  def test_flip_contract(self):
+    a = Tensor.randn(10,4)
+    b = Tensor.empty_like(a)
+    b = b.custom_kernel(a, fxn=flip_contract_kernel)[0]
+    self.assertTrue((a.flip(1) == b).all().item())
+
+  def test_noncontig(self):
+    a = Tensor.ones(16, 16).contiguous()
+    tst = Tensor.empty_like(a)
+    b = a+1
+    b_p1 = Tensor.custom_kernel(tst, b, fxn=custom_add_one_kernel)[0]
+    self.assertTrue((b_p1 == 3).all().item())
+
+  def test_sum(self):
+    # TODO: this only works for float, and silently fails with int
+    a = Tensor([1.0, 2, 3, 4, 5])
+    tst = Tensor.empty(1)
+    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)
+    b = Tensor.randn(N, N)
+    c = Tensor.empty(N, N)
+
+    tst = Tensor.custom_kernel(c, a, b, fxn=custom_gemm)[0]
+    err = (tst - (a@b)).square().max()
+    self.assertLess(err.item(), 1e-6)
+
+  def test_gemm_backward_custom(self): self.test_gemm_backward(True)
+  # NOTE: grad_fxn doesn't work with pyrender
+  @Context(SPEC=1)
+  def test_gemm_backward(self, custom_backward_gemm=False):
+    N = 4
+    a_rand = Tensor.randn(N, 8)
+    b_rand = Tensor.randn(8, N)
+    Tensor.realize(a_rand, b_rand)
+
+    a, b = Tensor(a_rand.numpy(), requires_grad=True), Tensor(b_rand.numpy(), requires_grad=True)
+    c = Tensor.empty(N, N)
+    tst = Tensor.custom_kernel(c, a, b, fxn=custom_gemm, grad_fxn=backward_gemm_custom if custom_backward_gemm else backward_gemm)[0]
+    tst.sum().backward()
+    grad_a, grad_b = a.grad, b.grad
+    Tensor.realize(tst, grad_a, grad_b)
+
+    a, b = Tensor(a_rand.numpy(), requires_grad=True), Tensor(b_rand.numpy(), requires_grad=True)
+    ref = (a@b)
+    ref.sum().backward()
+    real_grad_a, real_grad_b = a.grad, b.grad
+    Tensor.realize(ref, real_grad_a, real_grad_b)
+
+    err = (tst - ref).square().max()
+    self.assertLess(err.item(), 1e-6)
+
+    err = (grad_a - real_grad_a).square().max()
+    self.assertLess(err.item(), 1e-6)
+
+    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_dtype_alu.py

@@ -194,6 +194,7 @@ class TestDTypeALU(unittest.TestCase):
          strat.floats(width=32, min_value=0, max_value=10.0) if skip_overflow else ht.float32,
          ht.int32, strat.sampled_from(binary_operations), strat.sampled_from(integer_binary_operations))
   @unittest.skipIf(Device.DEFAULT == "PYTHON", "TODO: fix cast inf to int32 in PYTHON")
+  @unittest.skip("broken on Mac")
   def test_float_midcast_int32(self, a, b, c, op1, op2): universal_test_midcast(a, b, c, op1, op2, dtypes.float32, dtypes.int32)
 
   @unittest.skip("broken. TODO: fix it")

test/test_linearizer_dumb.py

@@ -16,12 +16,12 @@ class TestLinearizerFailure(unittest.TestCase):
     c2 = UOp.range(UOp.const(dtypes.index, 784), 1, AxisType.GLOBAL)
     c3 = UOp.range(UOp.const(dtypes.index, 10), 3, AxisType.GLOBAL)
     c4 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(512), arg=1, src=())
-    c5 = c4.index(c1.valid(UOp.const(dtypes.bool, True))).load()
+    c5 = c4.index(c1.valid(UOp.const(dtypes.bool, True)))
     c6 = UOp.range(UOp.const(dtypes.index, 6000), 1004, AxisType.REDUCE)
     c7 = UOp.range(UOp.const(dtypes.index, 3750), 2006, AxisType.REDUCE)
     c8 = UOp.range(UOp.const(dtypes.index, 16), 2007, AxisType.GROUP_REDUCE)
     c9 = UOp(Ops.DEFINE_GLOBAL, dtypes.uchar.ptr(47040000), arg=2, src=())
-    c10 = c9.index((((c3*UOp.const(dtypes.index, 4704000))+c2)+(c6*UOp.const(dtypes.index, 784))).valid(UOp.const(dtypes.bool, True))).load()
+    c10 = c9.index((((c3*UOp.const(dtypes.index, 4704000))+c2)+(c6*UOp.const(dtypes.index, 784))).valid(UOp.const(dtypes.bool, True)))
     c11 = c5.alu(Ops.CMPNE, ((((c3*UOp.const(dtypes.index, 6000))+c6)+((c7*UOp.const(dtypes.index, 16))+c8)).alu(Ops.CMPLT, UOp.const(dtypes.index, 59999)).where(UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 1)).reduce(c7, c8, arg=Ops.ADD)+UOp.const(dtypes.int, -1))).where(UOp.const(dtypes.uchar, 0), c10).reduce(c6, arg=Ops.ADD)
     c12 = c0.index((((c1*UOp.const(dtypes.index, 7840))+(c2*UOp.const(dtypes.index, 10)))+c3).valid(UOp.const(dtypes.bool, True))).store(c11).end(c1, c2, c3)
     ast = c12.sink(arg=KernelInfo(name='test', axis_types=(), dont_use_locals=False, applied_opts=(Opt(op=OptOps.GROUP, axis=1, arg=16),), opts_to_apply=None))

test/test_linearizer_failures.py

@@ -12,9 +12,9 @@ class TestLinearizerFailures(unittest.TestCase):
     c3 = ((c1*UOp.const(dtypes.index, 32))+c2)
     c4 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(163840), arg=1, src=())
     c5 = UOp.range(UOp.const(dtypes.index, 2560), 0, AxisType.REDUCE)
-    c6 = c4.index(((((((c5//UOp.const(dtypes.index, 8))%UOp.const(dtypes.index, 8))*UOp.const(dtypes.index, 8))+(c5%UOp.const(dtypes.index, 8)))+(((c2*UOp.const(dtypes.index, 40))+(c5//UOp.const(dtypes.index, 64)))*UOp.const(dtypes.index, 64)))+(c1*UOp.const(dtypes.index, 81920)))).load()
+    c6 = c4.index(((((((c5//UOp.const(dtypes.index, 8))%UOp.const(dtypes.index, 8))*UOp.const(dtypes.index, 8))+(c5%UOp.const(dtypes.index, 8)))+(((c2*UOp.const(dtypes.index, 40))+(c5//UOp.const(dtypes.index, 64)))*UOp.const(dtypes.index, 64)))+(c1*UOp.const(dtypes.index, 81920))))
     c7 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(64), arg=2, src=())
-    c8 = c7.index(c3).load()
+    c8 = c7.index(c3)
     c9 = ((((c6+(c8*UOp.const(dtypes.float, -1.0)))*(c6+(c8*UOp.const(dtypes.float, -1.0)))).reduce(c5, arg=Ops.ADD)*UOp.const(dtypes.float, 0.000390625))+UOp.const(dtypes.float, 1e-05)).sqrt().reciprocal()
     c10 = c0.index(c3).store(c9).end(c1, c2)
     ast = c10.sink()

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_profiler.py

@@ -199,6 +199,7 @@ class TestProfiler(unittest.TestCase):
     #self.assertLess(e1.st, e2.st)
     #self.assertGreater(e1.en-e1.st, e2.en-e2.st)
 
+  @unittest.skipIf(not CI, "this test is flaky locally")
   @unittest.skipUnless(Device[Device.DEFAULT].graph is not None, "graph support required")
   def test_graph(self):
     from test.test_graph import helper_alloc_rawbuffer, helper_exec_op, helper_test_graphs

test/test_rangeify.py

@@ -1,10 +1,56 @@
 import unittest
 from tinygrad import Tensor, nn, Device
-from tinygrad.helpers import Context, GlobalCounters, CI, getenv, PCONTIG
+from tinygrad.helpers import Context, GlobalCounters, CI, getenv, PCONTIG, DEBUG
 from tinygrad.uop.ops import graph_rewrite, PatternMatcher, UPat, Ops
+from tinygrad.codegen.opt import OptOps, Opt
 from tinygrad.renderer.ptx import PTXRenderer
 from tinygrad.renderer.nir import NIRRenderer
 
+@unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, (NIRRenderer, PTXRenderer)), "broken in LVP and PTX")
+class TestDoubleMatmul(unittest.TestCase):
+  def setUp(self):
+    with Context(DEBUG=0):
+      self.a, self.b, self.c = [Tensor.randn(16, 16).contiguous().realize() for _ in range(3)]
+      self.ref = (self.a @ self.b @ self.c).realize()
+
+  def _test(self, opts):
+    with Context(PCONTIG=2, DEBUG=max(2, DEBUG.value)):
+      out = (self.a @ self.b @ self.c).contiguous(arg=opts).realize()
+
+    with Context(DEBUG=0):
+      err = (out-self.ref).square()
+      self.assertLess(err.max().item(), 1e-4)
+      self.assertLess(err.mean().item(), 1e-6)
+
+  def test_baseline(self): self._test(())
+  def test_upcast_0(self): self._test((Opt(OptOps.UPCAST, 0, 4),))
+  def test_upcast_1(self): self._test((Opt(OptOps.UPCAST, 1, 4),))
+  def test_upcast_2(self): self._test((Opt(OptOps.UPCAST, 2, 4),))
+  def test_upcast_01(self): self._test((Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.UPCAST, 1, 4)))
+  def test_upcast_01_mismatch(self): self._test((Opt(OptOps.UPCAST, 0, 2), Opt(OptOps.UPCAST, 1, 4)))
+  def test_upcast_02(self): self._test((Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.UPCAST, 2, 4)))
+  def test_upcast_12(self): self._test((Opt(OptOps.UPCAST, 1, 4), Opt(OptOps.UPCAST, 2, 4)))
+
+  def test_unroll_0(self): self._test((Opt(OptOps.UNROLL, 0, 4),))
+  def test_unroll_1(self): self._test((Opt(OptOps.UNROLL, 1, 4),))
+  def test_unroll_01(self): self._test((Opt(OptOps.UNROLL, 0, 4), Opt(OptOps.UNROLL, 1, 4)))
+
+  def test_upcast_0_unroll_0(self): self._test((Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.UNROLL, 0, 4)))
+  def test_upcast_1_unroll_0(self): self._test((Opt(OptOps.UPCAST, 1, 4), Opt(OptOps.UNROLL, 0, 4)))
+  def test_upcast_2_unroll_0(self): self._test((Opt(OptOps.UPCAST, 2, 4), Opt(OptOps.UNROLL, 0, 4)))
+
+  def test_upcast_0_unroll_1(self): self._test((Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.UNROLL, 1, 4)))
+  def test_upcast_1_unroll_1(self): self._test((Opt(OptOps.UPCAST, 1, 4), Opt(OptOps.UNROLL, 1, 4)))
+  def test_upcast_2_unroll_1(self): self._test((Opt(OptOps.UPCAST, 2, 4), Opt(OptOps.UNROLL, 1, 4)))
+
+  def test_upcast_1_unroll_1_small(self): self._test((Opt(OptOps.UPCAST, 1, 2), Opt(OptOps.UNROLL, 1, 2)))
+  def test_upcast_1_unroll_1_rev(self): self._test((Opt(OptOps.UNROLL, 1, 2), Opt(OptOps.UPCAST, 1, 2)))
+
+  def test_upcast_01_unroll_01(self):
+    self._test((Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.UPCAST, 1, 4), Opt(OptOps.UNROLL, 0, 4), Opt(OptOps.UNROLL, 1, 4)))
+  def test_upcast_12_unroll_01(self):
+    self._test((Opt(OptOps.UPCAST, 1, 4), Opt(OptOps.UPCAST, 2, 4), Opt(OptOps.UNROLL, 0, 4), Opt(OptOps.UNROLL, 1, 4)))
+
 class TestRangeifyAssign(unittest.TestCase):
   def test_assign_permuted(self):
     A = Tensor.empty(4, 4, dtype='int')
@@ -38,7 +84,7 @@ elif getenv("BIG") > 1:
   BS, HEADS, SEQLEN, EMB = 4, 32, 2048, 128
 elif getenv("BIG") > 0:
   # bigger
-  BS, HEADS, SEQLEN, EMB = 4, 32, 1024, 64
+  BS, HEADS, SEQLEN, EMB = 4, 32, 128, 128
 else:
   BS, HEADS, SEQLEN, EMB = 4, 2, 16, 8
 
@@ -85,9 +131,9 @@ class TestPcontig(unittest.TestCase):
     print(f"mse: {mse}")
     self.assertLessEqual(mse, 1e-6)
 
-  def test_flash_attention(self):
-    with Context(PCONTIG=2, DEBUG=2):
-      ret = fa().realize()
+  def test_flash_attention(self, opts=None):
+    with Context(PCONTIG=2, DEBUG=max(2, DEBUG.value)):
+      ret = fa().realize() if opts is None else fa().contiguous(arg=opts).realize()
       print(f"{GlobalCounters.global_ops/1e9:.2f} GFLOPS")
     with Context(DEBUG=2):
       cmp = fa().realize()
@@ -97,6 +143,15 @@ class TestPcontig(unittest.TestCase):
     print(f"mse: {mse}")
     self.assertLessEqual(mse, 1e-6)
 
+  def test_flash_attention_opt(self):
+    opts = ()
+    # columns in top matrix
+    opts += (Opt(OptOps.UPCAST, 0, 4),)
+    # columns in bottom matrix
+    opts += (Opt(OptOps.UPCAST, 3, 4),)
+    # rows in all the matrix
+    opts += (Opt(OptOps.UPCAST, 4, 4),)
+    self.test_flash_attention(opts)
 
 # *** non CI rangeify tests below this line ***
 

test/test_renderer_failures.py

@@ -1,5 +1,4 @@
 import unittest
-from typing import List, cast
 import numpy as np
 from tinygrad.device import Buffer, Device, is_dtype_supported
 from tinygrad.dtype import dtypes, ConstType
@@ -15,15 +14,15 @@ from tinygrad.tensor import Tensor, _to_np_dtype
 from tinygrad.codegen import full_rewrite
 from tinygrad.engine.realize import lower_schedule_item
 
-def _test_uop_result(inputs:List[Tensor], stores:List[UOp], local_size=None):
+def _test_uop_result(inputs:list[Tensor], stores:list[UOp], local_size=None):
   for x in inputs: x.realize()
   # NOTE: we only toposort the stores
-  uops: List[UOp] = []
-  def _recursive_add(uop:UOp) -> List[UOp]: return flatten([_recursive_add(x) for x in uop.src])+[uop]
+  uops: list[UOp] = []
+  def _recursive_add(uop:UOp) -> list[UOp]: return flatten([_recursive_add(x) for x in uop.src])+[uop]
   uops = dedup(flatten(_recursive_add(st) for st in stores))
   outbufs = [Buffer(Device.DEFAULT, sz:=(1 if local_size is None else prod(local_size)), (dtype:=u.src[1].dtype), \
       initial_value=np.zeros(sz, dtype=_to_np_dtype(dtype)).data) for u in uops if u.op is Ops.STORE]
-  inbufs = [cast(UOp,x.uop).base.buffer for x in inputs]
+  inbufs = [x.uop.base.buffer for x in inputs]
   src = Device[Device.DEFAULT].renderer.render(uops)
   ei = CompiledRunner(ProgramSpec(uops[-1].arg.name if uops[-1].arg is not None else "test",
                                   src, Device.DEFAULT, uops[-1], uops=uops, local_size=local_size))
@@ -35,7 +34,7 @@ def _setup_and_test_alu(alu_op:Ops, input_val:ConstType, *alu_src_uops:UOp):
   a = UOp(Ops.DEFINE_GLOBAL, dtype.ptr(), (), 0)
   b = UOp(Ops.DEFINE_GLOBAL, dtype.ptr(), (), 1)
   idx = UOp.const(dtypes.int, 0)
-  ld = UOp(Ops.LOAD, dtype, (b.index(idx),))
+  ld = b.index(idx)
   alu = ld.alu(alu_op, *alu_src_uops)
   store = UOp.store(a.index(idx), alu)
   sink = UOp(Ops.SINK, dtypes.void, (store,))
@@ -47,7 +46,7 @@ class TestRendererFailures(unittest.TestCase):
   def test_gated_store_with_alu(self):
     a = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(), (), 0)
     gate_alu = (lidx0:=UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 4),), 'lidx0')).ne(0)
-    gated_alu_store = UOp(Ops.STORE, dtypes.void, (a.index(lidx0, gate_alu), UOp.const(dtypes.int, 1)))
+    gated_alu_store = UOp(Ops.STORE, dtypes.void, (a.index(lidx0.valid(gate_alu)), UOp.const(dtypes.int, 1)))
     sink = UOp(Ops.SINK, dtypes.void, (gated_alu_store,))
     uops = full_rewrite(sink, Device[Device.DEFAULT].renderer)
     ret = _test_uop_result([], uops, local_size=[4, 1, 1])[0]
@@ -58,7 +57,7 @@ class TestRendererFailures(unittest.TestCase):
     a = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(), (), 0)
     gate_alu_0 = (lidx0:=UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 4),), 'lidx0')).ne(0)
     gate_alu_1 = (lidx1:=UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 2),), 'lidx1')).ne(0)
-    gated_alu_store = UOp(Ops.STORE, dtypes.void, (a.index(lidx0+lidx1*4, gate_alu_0&gate_alu_1), UOp.const(dtypes.int, 1)))
+    gated_alu_store = UOp(Ops.STORE, dtypes.void, (a.index((lidx0+lidx1*4).valid(gate_alu_0&gate_alu_1)), UOp.const(dtypes.int, 1)))
     sink = UOp(Ops.SINK, dtypes.void, (gated_alu_store,))
     uops = full_rewrite(sink, Device[Device.DEFAULT].renderer)
     ret = _test_uop_result([], uops, local_size=[4, 2, 1])[0]

test/test_schedule.py

@@ -370,6 +370,7 @@ class TestSchedule(unittest.TestCase):
 
   # NOTE: this is causing "LAZYCACHE=1 incorrectly reuses contiguous const" #4562
   # should contiguous dedup?
+  @unittest.skip("we do the exact opposite now")
   def test_dedup_contiguous(self):
     a = Tensor.ones(4).contiguous()
     b = Tensor.ones(4).contiguous()
@@ -446,7 +447,7 @@ class TestSchedule(unittest.TestCase):
   @unittest.skipUnless(is_dtype_supported(dtypes.ulong), "Needs ulong")
   def test_fold_conv_batchnorm_optim(self):
     # this is too high
-    for optim, cnt in [(nn.optim.Adam, 21), (nn.optim.SGD, 8)]:
+    for optim, cnt in [(nn.optim.Adam, 27), (nn.optim.SGD, 7)]:
       with self.subTest(optim=optim.__name__):
         with Tensor.train():
           img = Tensor.ones(1,3,4,4)
@@ -710,7 +711,7 @@ class TestSchedule(unittest.TestCase):
     self.assertEqual(b.buffer.numpy(), [12])
 
   # unlike schedule, kernelize can be called multiple times on a Tensor
-  def test_double_kerenlize(self):
+  def test_double_kernelize(self):
     a = Tensor.empty(10)
     b = Tensor.empty(10)
     c = (a+b)
@@ -1041,13 +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)
@@ -1220,7 +1220,7 @@ class TestSchedule(unittest.TestCase):
       _realize_weights(layer)
       opt = nn.optim.Adam(nn.state.get_parameters(layer), lr=1e-4)
       layer(x).relu().sum().backward()
-      check_schedule(opt.schedule_step(), 16)
+      check_schedule(opt.schedule_step(), 19)
 
   def test_adam_conv_fuse(self):
     with Tensor.train():
@@ -1230,7 +1230,7 @@ class TestSchedule(unittest.TestCase):
       opt = nn.optim.Adam(nn.state.get_parameters(c1), lr=1e-4)
       opt.zero_grad()
       c1(img).relu().sum().backward()
-      check_schedule(opt.schedule_step(), 16)
+      check_schedule(opt.schedule_step(), 19)
 
   def test_adam_2convs_fuse(self):
     with Tensor.train():
@@ -1241,7 +1241,7 @@ class TestSchedule(unittest.TestCase):
       opt = nn.optim.Adam(nn.state.get_parameters([c1, c2]), lr=1e-4)
       opt.zero_grad()
       c2(c1(img).relu()).relu().sum().backward()
-      check_schedule(opt.schedule_step(), 18)
+      check_schedule(opt.schedule_step(), 21)
 
   def test_sgd_conv_fuse(self):
     with Tensor.train():
@@ -1502,6 +1502,18 @@ class TestSchedule(unittest.TestCase):
     run_schedule(sched)
     np.testing.assert_allclose(dx.numpy(), [[[[0.,3.,9.],[0,1.,3.],[0.,0.,0.]]]*3]*3)
 
+  def test_fuse_arange_avg_pool2d_ceil_mode(self):
+    x = Tensor.avg_pool2d(Tensor.empty(1,1,6,6), kernel_size=(3,3), padding=1, stride=3, ceil_mode=True)
+    sched = check_schedule(x, 1)
+    self.assertEqual(len([x for x in sched[0].ast.backward_slice_with_self if x.op is Ops.REDUCE]), 1)
+
+  def test_fuse_arange_pad_circular_mode_bw(self):
+    x = Tensor.empty(1,1,5,5,5)
+    out = x.pad((1,2,3,5,1,2), mode="circular")
+    g = out.sum().gradient(x)[0]
+    sched = check_schedule(g, 1)
+    self.assertEqual(len([x for x in sched[0].ast.backward_slice_with_self if x.op is Ops.REDUCE]), 0)
+
   # TODO like openpilot with imagef
   @unittest.skipUnless(is_dtype_supported(dtypes.half), "need half")
   def test_base_change_expand_expand(self):
@@ -2162,8 +2174,8 @@ class TestCopyFolding(unittest.TestCase):
     self.assertListEqual(b.tolist(), [[0, 2], [1, 3]])
 
   def test_permute_on_disk_contiguous(self):
-    with open(temp('dt_arange_4_permute'), "wb") as f: f.write(Tensor.arange(4).realize().uop.base.buffer.as_buffer())
-    a = Tensor.empty(4, dtype=dtypes.int32, device=f"disk:{temp('dt_arange_4_permute')}")
+    with open(temp('dt_arange_4_permute_contig'), "wb") as f: f.write(Tensor.arange(4).realize().uop.base.buffer.as_buffer())
+    a = Tensor.empty(4, dtype=dtypes.int32, device=f"disk:{temp('dt_arange_4_permute_contig')}")
     b = a.reshape(2, 2).permute(1, 0).contiguous().to("CPU")
     b.realize()
     self.assertListEqual(b.tolist(), [[0, 2], [1, 3]])
@@ -2266,7 +2278,7 @@ class TestContiguous(unittest.TestCase):
   def test_double_contiguous_realizes_once(self):
     a = Tensor.empty(4, 1)
     b = a.expand((4, 4)).contiguous().contiguous()
-    check_schedule(b, 2) # TODO: should be 1?
+    check_schedule(b, 1)
 
   def test_view_does_not_realize(self):
     a = Tensor.empty(4)

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_tensor.py

@@ -810,6 +810,7 @@ class TestTensorMetadata(unittest.TestCase):
     self.assertEqual(len(si.metadata), 1)
     self.assertEqual(si.metadata[0].name, "relu")
 
+  @unittest.skip("this no longer works")
   def test_assign(self):
     x = Tensor.empty(10, 10).realize()
     x.assign(Tensor.ones(10, 10).contiguous())
@@ -839,11 +840,11 @@ class TestTensorMetadata(unittest.TestCase):
     self.assertEqual(y.grad.uop.metadata[0].name, "sigmoid")
     self.assertTrue(y.grad.uop.metadata[0].backward)
     si = Tensor.schedule(out, x.grad, y.grad)[-1]
-    self.assertEqual(len(si.metadata), 3, f"failed with {si.metadata}")
+    #self.assertEqual(len(si.metadata), 3, f"failed with {si.metadata}")
     self.assertSetEqual(set(m.name for m in si.metadata), {"sigmoid", "relu"})
-    bw = [m for m in si.metadata if m.backward]
-    self.assertEqual(len(bw), 1)
-    self.assertEqual(bw[0].name, "sigmoid")
+    #bw = [m for m in si.metadata if m.backward]
+    #self.assertEqual(len(bw), 1)
+    #self.assertEqual(bw[0].name, "sigmoid")
 
 class TestIdxUpcast(unittest.TestCase):
   def _find_op(self, ast: UOp, op: Ops):
@@ -919,5 +920,38 @@ class TestIdxUpcast(unittest.TestCase):
     a = Tensor.empty(2**11, 2**11, 1, dtype=dtypes.int8).permute((2, 0, 1)).expand((2**9+10, -1, -1)).contiguous()
     a.realize()
 
+class TestTensorUnique(unittest.TestCase):
+  def test_empty_bufs_unique(self):
+    a = Tensor.empty(10, 10).contiguous()
+    b = Tensor.empty(10, 10).contiguous()
+    Tensor.realize(a,b)
+    self.assertIsNot(a.uop.buffer, b.uop.buffer)
+
+  def test_zeros_bufs_unique_sep(self):
+    a = Tensor.zeros(10, 10).contiguous()
+    Tensor.realize(a)
+    b = Tensor.zeros(10, 10).contiguous()
+    Tensor.realize(b)
+    self.assertIsNot(a.uop.buffer, b.uop.buffer)
+
+  def test_zeros_bufs_unique(self):
+    a = Tensor.zeros(10, 10).contiguous()
+    b = Tensor.zeros(10, 10).contiguous()
+    Tensor.realize(a,b)
+    self.assertIsNot(a.uop.buffer, b.uop.buffer)
+
+  def test_eye_bufs_unique(self):
+    a = Tensor.eye(10).contiguous()
+    b = Tensor.eye(10).contiguous()
+    Tensor.realize(a,b)
+    self.assertIsNot(a.uop.buffer, b.uop.buffer)
+
+  def test_times_2_not_unique(self):
+    a = Tensor.zeros(10, 10).contiguous()
+    b = a * 2
+    c = a * 2
+    Tensor.realize(b,c)
+    self.assertIs(b.uop.buffer, c.uop.buffer)
+
 if __name__ == '__main__':
   unittest.main()

test/test_uop_graph.py

@@ -307,9 +307,10 @@ class TestUOpGraph(unittest.TestCase):
     for vec_size in [2, 4, 8]:
       consts = [UOp.const(dtypes.float, float(i)) for i in range(vec_size)]
       vec = UOp(Ops.VECTORIZE, dtypes.float.vec(vec_size), tuple(consts))
-      uops = to_uops_list([UOp(Ops.GEP, dtypes.float, (vec,), (i,)) for i in range(vec_size)])
-      for uop, const in zip(uops, consts):
-        self.assertEqual(uop, const)
+      with Context(SPEC=0):
+        uops = to_uops_list([UOp(Ops.GEP, dtypes.float, (vec,), (i,)) for i in range(vec_size)])
+        for uop, const in zip(uops, consts):
+          self.assertEqual(uop, const)
 
   @unittest.skip("no longer testable standalone")
   def test_wmma_vectorize_fold(self):
@@ -375,7 +376,7 @@ class TestUOpGraph(unittest.TestCase):
     d0 = UOp(Ops.DEFINE_GLOBAL, dtypes.bool.ptr(), arg=0)
     d1 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(), arg=1)
     idx = UOp.const(dtypes.int, 0)
-    ld = UOp(Ops.LOAD, dtypes.int, (d1.index(idx),))
+    ld = d1.index(idx)
     alu = (ld<1).cast(dtypes.bool)
     out = UOp(Ops.STORE, dtypes.void, (d0.index(idx), alu))
     uops = to_uops_list([out])
@@ -385,7 +386,7 @@ class TestUOpGraph(unittest.TestCase):
     d0 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=0)
     d1 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(), arg=1)
     idx = UOp.const(dtypes.int, 0)
-    ld = UOp(Ops.LOAD, dtypes.int, (d1.index(idx),))
+    ld = d1.index(idx)
     alu = ld.cast(dtypes.float).cast(dtypes.float)
     out = UOp(Ops.STORE, dtypes.void, (d0.index(idx), alu))
     uops = to_uops_list([out])
@@ -407,7 +408,7 @@ class TestUOpGraph(unittest.TestCase):
   def test_bitcast_to_same_dtype_fold(self):
     for dt in dtypes.ints + dtypes.floats + (dtypes.bool,):
       d0 = UOp(Ops.DEFINE_GLOBAL, dt.ptr(), arg=0)
-      v = UOp(Ops.LOAD, dt, (d0.index(UOp.const(dtypes.int, 0)),))
+      v = d0.index(UOp.const(dtypes.int, 0))
       uops = to_uops_list([v.bitcast(dt)])
       self.assertEqual(len([x for x in uops if x.op is Ops.BITCAST]), 0, f"dtype = {dt}")
 
@@ -419,7 +420,7 @@ class TestUOpGraph(unittest.TestCase):
   def test_where_on_gated_load_fold(self):
     ridx0 = UOp.range(100, 0)
     d0 = UOp(Ops.DEFINE_GLOBAL, dtypes.long.ptr(), (), 0)
-    ld = d0.index(ridx0.valid(ridx0<50)).load()
+    ld = d0.index(ridx0.valid(ridx0<50))
     w = (ridx0<50).where(ld, 5)
     uops = to_uops_list([w])
     for u in uops:
@@ -429,7 +430,7 @@ class TestUOpGraph(unittest.TestCase):
   def test_where_on_gated_load_folds_swapped_branches(self):
     ridx0 = UOp.range(100, 0)
     d0 = UOp(Ops.DEFINE_GLOBAL, dtypes.long.ptr(), (), 0)
-    ld = d0.index(ridx0.valid((ridx0<50).logical_not())).load()
+    ld = d0.index(ridx0.valid((ridx0<50).logical_not()))
     w = (ridx0<50).where(5, ld)
     uops = to_uops_list([w])
     for u in uops:
@@ -440,7 +441,7 @@ class TestUOpGraph(unittest.TestCase):
     ridx0 = UOp.range(100, 0)
     d0 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(), (), 0)
     gate_idx = ridx0.valid((ridx0<50))
-    ld = d0.index(gate_idx).load().cast(dtypes.float)
+    ld = d0.index(gate_idx).cast(dtypes.float)
     w = (ridx0<50).where(ld, 5.0)
     uops = to_uops_list([w])
     for u in uops:
@@ -466,50 +467,49 @@ class TestUOpGraph(unittest.TestCase):
     c1 = UOp.range(UOp.const(dtypes.index, 512), 1, AxisType.LOOP)
     c2 = UOp.range(UOp.const(dtypes.index, 250), 2, AxisType.LOOP)
     c3 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(512), arg=1, src=())
-    c4 = c3.index(c1).load()
+    c4 = c3.index(c1)
     c5 = UOp.range(UOp.const(dtypes.index, 240), 0, AxisType.REDUCE)
     c6 = ((c2*UOp.const(dtypes.index, 240))+c5)
     c7 = UOp(Ops.DEFINE_GLOBAL, dtypes.uchar.ptr(60000), arg=2, src=())
-    c8 = c7.index(c6).load()
+    c8 = c7.index(c6)
     c9 = ((c4<0).where((c4+60000), c4)!=c6.cast(dtypes.int)).where(0, c8.cast(dtypes.uint).cast(dtypes.uchar)).reduce(c5, arg=Ops.ADD)
     c10 = c0.index(((c1*UOp.const(dtypes.index, 250))+c2)).store(c9).end(c1, c2)
-    ast = c10.sink()
-    uops = to_uops_list([ast])
+    uops = to_uops_list([c10])
     for u in uops:
       self.assertNotEqual(u.dtype, dtypes.long)
 
   def test_in_out_of_bounds_access(self):
     with Context(IGNORE_OOB=0):
       glbl0 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(16), (), 0)
-      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(UOp.const(dtypes.int, 0)),))
+      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(UOp.const(dtypes.int, 0), ptr=True),))
       to_uops_list([ld0])
-      ld1 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(UOp.const(dtypes.int, 15)),))
+      ld1 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(UOp.const(dtypes.int, 15), ptr=True),))
       to_uops_list([ld1])
-      ld1 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(UOp.const(dtypes.int, 7)),))
+      ld1 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(UOp.const(dtypes.int, 7), ptr=True),))
       to_uops_list([ld1])
 
-      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(UOp.const(dtypes.int, 42)),))
+      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(UOp.const(dtypes.int, 42), ptr=True),))
       with self.assertRaises(RuntimeError): to_uops_list([ld0])
 
   def test_in_out_of_bounds_access_symbolic(self):
     with Context(IGNORE_OOB=0):
       glbl0 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(16), (), 0)
-      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(Variable("i", 1, 10)),))
+      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(Variable("i", 1, 10), ptr=True),))
       to_uops_list([ld0])
-      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(Variable("i", 0, 15)),))
+      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(Variable("i", 0, 15), ptr=True),))
       to_uops_list([ld0])
 
-      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(Variable("i", 0, 20)),))
+      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(Variable("i", 0, 20), ptr=True),))
       with self.assertRaises(RuntimeError): to_uops_list([ld0])
 
   def test_in_out_of_bounds_access_gated_store(self):
     with Context(IGNORE_OOB=0):
       glbl0 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(16), src=(), arg=0)
       v = Variable("v", 0, 20)
-      st0 = UOp(Ops.STORE, dtypes.void, src=(glbl0.index(v, v<16), UOp.const(dtypes.int, 0)))
+      st0 = UOp(Ops.STORE, dtypes.void, src=(glbl0.index(v.valid(v<16)), UOp.const(dtypes.int, 0)))
       to_uops_list([st0])
 
-      st1 = UOp(Ops.STORE, dtypes.void, (glbl0.index(v), v, v<20))
+      st1 = UOp(Ops.STORE, dtypes.void, (glbl0.index(v.valid(v<20)), v))
       with self.assertRaises(RuntimeError): to_uops_list([st1])
 
   @unittest.skip("if not allowed in graph")
@@ -531,7 +531,7 @@ class TestUOpGraph(unittest.TestCase):
       if_barrier = UOp(Ops.IF, dtypes.void, (gate, barrier))
 
       # Load from local memory (after the IF/barrier)
-      local_load = UOp(Ops.LOAD, dtypes.uint, (sbuf.index(lidx), if_barrier))
+      local_load = UOp(Ops.LOAD, dtypes.uint, (sbuf.index(lidx, ptr=True), if_barrier))
 
       # Store to global memory
       global_store = UOp(Ops.STORE, dtypes.void, (gbuf.index(gidx), local_load))
@@ -542,18 +542,18 @@ class TestUOpGraph(unittest.TestCase):
       ridx = UOp.range(20, 0)
       glbl0 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(16), (), 0)
       i = (ridx.cast(dtypes.float)*0.68).trunc().cast(dtypes.int)
-      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(i, ((0<=i)&(i<16))),))
+      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(i.valid((0<=i)&(i<16)), ptr=True),))
       to_uops_list([ld0])
       glblfloat = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(20), (), 0)
       ldfloat = UOp(Ops.LOAD, dtypes.float, (glblfloat.index(ridx),))
       i = (ldfloat+3.14).cast(dtypes.int)
-      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(i, ((0<=i)&(i<16))),))
+      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(i, ((0<=i)&(i<16)), ptr=True),))
 
   def test_load_cast_to_bool(self):
     with Context(IGNORE_OOB=0):
       glbl0 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(1), (), 0)
       ridx = UOp.range(20, 0)
-      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(ridx, ridx.cast(dtypes.bool).logical_not()),))
+      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(ridx.valid(ridx.cast(dtypes.bool).logical_not()), ptr=True),))
       to_uops_list([ld0])
 
   @unittest.skip("Bool load is not supported yet")
@@ -562,36 +562,36 @@ class TestUOpGraph(unittest.TestCase):
       glbl0 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(16), (), 0)
       mask = UOp(Ops.DEFINE_GLOBAL, dtypes.bool.ptr(16), (), 0)
       ridx = UOp.range(20, 0)
-      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(UOp.const(ridx, ridx<16&mask),)))
+      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(UOp.const(ridx, ridx<16&mask), ptr=True)))
       to_uops_list([ld0])
 
   def test_out_of_bounds_off_by_one_access(self):
     with Context(IGNORE_OOB=0):
       glbl0 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(16), (), 0)
-      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(UOp.const(dtypes.int, 16)),))
+      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(UOp.const(dtypes.int, 16), ptr=True),))
       with self.assertRaises(RuntimeError): to_uops_list([ld0])
 
   def test_in_out_bounds_access_with_mask(self):
     with Context(IGNORE_OOB=0):
       glbl0 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(16), (), 0)
       gidx0 = UOp.range(42, 0, AxisType.GLOBAL)
-      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(gidx0, (5<gidx0)&(gidx0<16)),))
-      ld1 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(gidx0, gidx0<16),))
+      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(gidx0.valid((5<gidx0)&(gidx0<16)), ptr=True),))
+      ld1 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(gidx0.valid(gidx0<16), ptr=True),))
       to_uops_list([ld0, ld1])
 
-      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(gidx0, gidx0<17),))
+      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(gidx0.valid(gidx0<17), ptr=True),))
       with self.assertRaises(RuntimeError): to_uops_list([ld0])
 
   def test_in_out_of_bounds_access_symbolic_mask(self):
     with Context(IGNORE_OOB=0):
       glbl0 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(16), (), 0)
       i = Variable("i", 1, 80)
-      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(i, i<10),))
+      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(i.valid(i<10), ptr=True),))
       to_uops_list([ld0])
-      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(i, i<15),))
+      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(i.valid(i<15), ptr=True),))
       to_uops_list([ld0])
 
-      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(i, i<20),))
+      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(i.valid(i<20), ptr=True),))
       with self.assertRaises(RuntimeError): to_uops_list([ld0])
 
   def test_in_out_of_bounds_access_index_load(self):
@@ -599,11 +599,11 @@ class TestUOpGraph(unittest.TestCase):
       glbl0 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(16), (), 0)
       glbl1 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(64), (), 0)
       gidx0 = UOp.range(42, 0, AxisType.GLOBAL)
-      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(gidx0, gidx0<8),)).cast(dtypes.index)
-      ld1 = UOp(Ops.LOAD, dtypes.int, (glbl1.index(ld0*2, (ld0>=0)&(ld0<32)),))
+      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(gidx0.valid(gidx0<8), ptr=True),)).cast(dtypes.index)
+      ld1 = UOp(Ops.LOAD, dtypes.int, (glbl1.index((ld0*2).valid((ld0>=0)&(ld0<32)), ptr=True),))
       to_uops_list([ld1])
 
-      ld1 = UOp(Ops.LOAD, dtypes.int, (glbl1.index(ld0*2, (ld0>=0)&(ld0<64)),))
+      ld1 = UOp(Ops.LOAD, dtypes.int, (glbl1.index((ld0*2).valid((ld0>=0)&(ld0<64)), ptr=True),))
       with self.assertRaises(RuntimeError): to_uops_list([ld1])
 
   def test_bounds_with_loaded_bool(self):
@@ -611,8 +611,8 @@ class TestUOpGraph(unittest.TestCase):
       glbl0 = UOp(Ops.DEFINE_GLOBAL, dtypes.bool.ptr(16), (), 0)
       glbl1 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(8), (), 0)
       gidx0 = UOp(Ops.SPECIAL, dtypes.index, (UOp.const(dtypes.index, 16),), "gidx0")
-      ld0 = glbl0.index(gidx0).load()
-      ld1 = glbl1.index(gidx0.valid(ld0)).load()
+      ld0 = glbl0.index(gidx0, ptr=True).load()
+      ld1 = glbl1.index(gidx0.valid(ld0), ptr=True).load()
       with self.assertRaises(RuntimeError): to_uops_list([ld1])
 
   def test_fold_gated_load(self):
@@ -620,38 +620,38 @@ class TestUOpGraph(unittest.TestCase):
     glbl1 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(), (), 1)
     glbl2 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(), (), 2)
     idx = UOp.const(dtypes.int, 0)
-    ld0 = UOp(Ops.LOAD, dtypes.int, (glbl1.index(UOp.invalid()),))
-    ld1 = UOp(Ops.LOAD, dtypes.int, (glbl2.index(idx, UOp.const(dtypes.bool, True)),))
+    ld0 = glbl1.index(UOp.invalid())
+    ld1 = glbl2.index(idx.valid(UOp.const(dtypes.bool, True)))
     uops = to_uops_list([UOp(Ops.STORE, dtypes.void, (glbl0.index(idx), ld1+ld0))])
     ld0 = uops[-1].src[-1]
     # the gate and invalid value are deleted from ld1
-    self.assertEqual(ld0, UOp.load(glbl2.index(idx), dtype=dtypes.int))
+    self.assertEqual(ld0, UOp.load(glbl2.index(idx, ptr=True), dtype=dtypes.int))
 
   def test_fold_gated_load_local(self):
     glbl0 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(), (), 0)
     smem = UOp(Ops.DEFINE_LOCAL, dtypes.int.ptr(size=18, addrspace=AddrSpace.LOCAL), (), "temp")
     lidx = UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 16),), "lidx0")
-    st = UOp(Ops.STORE, dtypes.void, (smem.index(lidx), UOp.load(glbl0.index(lidx), dtype=dtypes.int)))
+    st = UOp(Ops.STORE, dtypes.void, (smem.index(lidx, ptr=True), glbl0.index(lidx, ptr=True).load()))
     barrier = UOp(Ops.BARRIER, dtypes.void, (st, ))
-    ld0 = UOp(Ops.LOAD, dtypes.int, (smem.after(barrier).index(UOp.invalid()),))
-    ld1 = UOp(Ops.LOAD, dtypes.int, (smem.after(barrier).index(lidx+2, UOp.const(dtypes.bool, True)),))
+    ld0 = smem.after(barrier).index(UOp.invalid())
+    ld1 = smem.after(barrier).index((lidx+2).valid(UOp.const(dtypes.bool, True)))
     uops = to_uops_list([UOp(Ops.STORE, dtypes.void, (glbl0.index(lidx), ld1+ld0))])
 
     ld0 = uops[-1].src[-1]
     # the gate and invalid value are deleted from ld1
-    self.assertEqual(ld0.src[0], smem.after(barrier).index(lidx+2))
+    self.assertEqual(ld0.src[0], smem.after(barrier).index(lidx+2, ptr=True))
 
   def test_fold_gated_store(self):
     glbl = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(), (), 0)
     idx0 = UOp.const(dtypes.int, 0)
     idx1 = UOp.const(dtypes.int, 0)
     val = UOp.const(dtypes.int, 42)
-    st0 = glbl.index(UOp.invalid()).store(val)
-    st1 = glbl.index(idx0, UOp.const(dtypes.bool, True)).store(val)
+    st0 = glbl.index(UOp.invalid(), ptr=True).store(val)
+    st1 = glbl.index(idx0.valid(UOp.const(dtypes.bool, True)), ptr=True).store(val)
     uops = to_uops_list([st0, st1])
     # only the second store happens
     self.assertEqual(len(uops), 5)
-    self.assertEqual(uops[-1], glbl.index(idx1).store(val))
+    self.assertEqual(uops[-1], glbl.index(idx1, ptr=True).store(val))
 
   @unittest.skip("this is a uop type error")
   def test_asserts_bad_gate(self):

test/test_uops.py

@@ -2,13 +2,13 @@ from typing import Optional, Any
 import unittest, math
 import numpy as np
 from tinygrad.tensor import Tensor, _to_np_dtype
-from tinygrad.helpers import CI, DEBUG, getenv, Timing
+from tinygrad.helpers import CI, DEBUG, getenv, Timing, Context
 from tinygrad.dtype import dtypes, DType, AddrSpace
 from tinygrad.device import Buffer, Device
-from tinygrad.uop.ops import Ops, UOp, UPat, KernelInfo, exec_alu # noqa F401
+from tinygrad.uop.ops import Ops, UOp, UPat, KernelInfo, exec_alu, AxisType
 from tinygrad.uop.spec import shared_spec
 from tinygrad.renderer import ProgramSpec
-from tinygrad.engine.realize import CompiledRunner, get_program
+from tinygrad.engine.realize import CompiledRunner, get_program, get_runner, ExecItem
 from tinygrad.codegen import full_rewrite
 from tinygrad.uop.symbolic import sym
 from tinygrad.device import is_dtype_supported
@@ -39,9 +39,9 @@ def _test_single_value(vals, op, dts):
   output_dtype = dtypes.bool if op in (Ops.CMPLT, Ops.CMPNE) else dts[-1]
   buf_store = uop(uops, Ops.DEFINE_GLOBAL, output_dtype.ptr(), (), 0)
   buf_loads = [uop(uops, Ops.DEFINE_GLOBAL, dtype.ptr(), (), i+1) for i,dtype in enumerate(dts)]
-  loads = (uop(uops, Ops.LOAD, dtype, [buf_loads[i].index(uop(uops, Ops.CONST, dtypes.int32, (), 0))]) for i, dtype in enumerate(dts))
+  loads = (buf_loads[i].index(uop(uops, Ops.CONST, dtypes.int32, (), 0)) for i, dtype in enumerate(dts))
   alu = uop(uops, op, output_dtype, loads)
-  out = uop(uops, Ops.STORE, dtypes.void, (buf_store.index(uop(uops, Ops.CONST, dtypes.int32, (), 0)), alu))
+  out = uop(uops, Ops.STORE, dtypes.void, (buf_store.index(uop(uops, Ops.CONST, dtypes.int32, (), 0), ptr=True), alu))
   buf = Buffer(Device.DEFAULT, 1, output_dtype).allocate()
   buf2 = [Buffer(Device.DEFAULT, 1, dtype).allocate().copyin(np.array([a], dtype=_to_np_dtype(dtype)).data) for a,dtype in zip(vals, dts)]
   prg = _uops_to_prg([out])
@@ -56,7 +56,7 @@ def _test_single_value_const(vals, op, dts):
   buf_store = uop(uops, Ops.DEFINE_GLOBAL, output_dtype.ptr(), (), 0)
   loads = (uop(uops, Ops.CONST, dtype, [], a) for a,dtype in zip(vals, dts))
   alu = uop(uops, op, output_dtype, loads)
-  out = uop(uops, Ops.STORE, dtypes.void, (buf_store.index(uop(uops, Ops.CONST, dtypes.int32, (), 0)), alu))
+  out = buf_store[UOp.const(dtypes.int32, 0)].store(alu)
   buf = Buffer(Device.DEFAULT, 1, output_dtype).allocate()
   prg = _uops_to_prg([out])
   prg.exec([buf])
@@ -277,7 +277,7 @@ class TestGatedStoreRewrite(unittest.TestCase):
     gmem = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), (), 0)
     gidx0 = UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 4),), 'gidx0')
     gate = gidx0<UOp.const(dtypes.int, 1)
-    idx = UOp(Ops.INDEX, dtypes.float.ptr(), (gmem, gidx0 * UOp.const(dtypes.int, 2), gate))
+    idx = UOp(Ops.INDEX, dtypes.float.ptr(), (gmem, (gidx0 * UOp.const(dtypes.int, 2)).valid(gate)))
     val = UOp.const(dtypes.float, 42.0)
     store = UOp(Ops.STORE, dtypes.void, (idx, val))
     uops = to_uops_list([store])
@@ -294,7 +294,7 @@ class TestGatedStoreRewrite(unittest.TestCase):
     gmem1 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), (), 1)
     gidx0 = UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 4),), 'gidx0')
     idx = gidx0 * UOp.const(dtypes.int, 2)
-    idx0 = UOp(Ops.INDEX, dtypes.float.ptr(), (gmem0, idx, gidx0<UOp.const(dtypes.int, 1)))
+    idx0 = UOp(Ops.INDEX, dtypes.float.ptr(), (gmem0, idx.valid(gidx0<UOp.const(dtypes.int, 1))))
     idx1 = UOp(Ops.INDEX, dtypes.float.ptr(), (gmem1, idx))
     val = UOp.const(dtypes.float, 42.0)
     stores = [UOp.store(idx0, val), UOp.store(idx1, val)]
@@ -338,7 +338,7 @@ class TestLocalAccess(unittest.TestCase):
     smem = uop(uops, Ops.DEFINE_LOCAL, dtypes.float32.ptr(size=16, addrspace=AddrSpace.LOCAL), (), 'smem')
     st = uop(uops, Ops.STORE, dtypes.void, (smem.index(uop(uops, Ops.CONST, dtypes.int32, (), 0)), uop(uops, Ops.CONST, dtypes.float32, (), 42.0)))
     barr = uop(uops, Ops.BARRIER, dtypes.void, (st,))
-    sres = uop(uops, Ops.LOAD, dtypes.float32, (smem.after(barr).index(uop(uops, Ops.CONST, dtypes.int32, (), 0)),))
+    sres = uop(uops, Ops.LOAD, dtypes.float32, (smem.after(barr).index(uop(uops, Ops.CONST, dtypes.int32, (), 0), ptr=True),))
     self.assertEqual(_test_uops_result(dtypes.float32, uops, sres), 42)
 
   # NOTE: webgpu specific, since only webgpu performs bitpacking
@@ -348,7 +348,7 @@ class TestLocalAccess(unittest.TestCase):
     smem = uop(uops, Ops.DEFINE_LOCAL, dtypes.uint8.ptr(size=16, addrspace=AddrSpace.LOCAL), (), 'smem')
     st = uop(uops, Ops.STORE, dtypes.void, (smem.index(uop(uops, Ops.CONST, dtypes.int32, (), 0)), uop(uops, Ops.CONST, dtypes.uint8, (), 42)))
     barr = uop(uops, Ops.BARRIER, dtypes.void, (st,))
-    sres = uop(uops, Ops.LOAD, dtypes.uint8, (smem.after(barr).index(uop(uops, Ops.CONST, dtypes.int32, (), 0)),))
+    sres = smem.after(barr).index(uop(uops, Ops.CONST, dtypes.int32, (), 0))
     self.assertEqual(_test_uops_result(dtypes.uint8, uops, sres), 42)
 
   # NOTE: webgpu specific, since only webgpu performs bitpacking
@@ -382,7 +382,7 @@ class TestAssembly(unittest.TestCase):
     g1 = UOp(Ops.DEFINE_GLOBAL, dtypes.int32.ptr(), (), 0)
     c1 = UOp(Ops.CONST, dtypes.int, (), 2)
     c2 = UOp(Ops.CONST, dtypes.int, (), 3)
-    l1 = UOp(Ops.LOAD, dtypes.int, (g1.index(c1),))
+    l1 = g1.index(c1)
     a1 = UOp(Ops.MUL, dtypes.int, (l1, c1))
     a2 = UOp(Ops.MUL, dtypes.int, (l1, c2))
     uops = to_uops_list([a1,a2], ren=Device[Device.DEFAULT].renderer)
@@ -395,7 +395,7 @@ class TestAssembly(unittest.TestCase):
     for dt in (dtypes.int32, dtypes.uint32):
       g = UOp(Ops.DEFINE_GLOBAL, dt.ptr(), (), 0)
       c = UOp(Ops.CONST, dt, (), 2)
-      l = UOp(Ops.LOAD, dt, (g.index(c),))
+      l = g.index(c)
       a = UOp(Ops.IDIV, dt, (l, c))
       uops = to_uops_list([a], ren=Device[Device.DEFAULT].renderer)
       Device[Device.DEFAULT].renderer.render(uops)
@@ -406,7 +406,7 @@ class TestAssembly(unittest.TestCase):
   def test_fast_idiv_and_mod(self):
     g = UOp(Ops.DEFINE_GLOBAL, dtypes.uint32.ptr(), (), 0)
     c = UOp(Ops.CONST, dtypes.uint, (), 3)
-    l = UOp(Ops.LOAD, dtypes.uint, (g.index(c),))
+    l = g.index(c)
     a = UOp(Ops.IDIV, dtypes.uint, (l, c))
     uops = to_uops_list([a], ren=Device[Device.DEFAULT].renderer)
     Device[Device.DEFAULT].renderer.render(uops)
@@ -458,8 +458,7 @@ class TestAssembly(unittest.TestCase):
   def test_use_cmpeq(self):
     g = UOp(Ops.DEFINE_GLOBAL, dtypes.uint32.ptr(), (), 0)
     c = UOp(Ops.CONST, dtypes.uint, (), 7)
-    l = UOp(Ops.LOAD, dtypes.uint, (g.index(c),))
-    comp = l.ne(c).ne(True)
+    comp = g.index(c).ne(c).ne(True)
     uops = to_uops_list([comp], ren=Device[Device.DEFAULT].renderer)
     Device[Device.DEFAULT].renderer.render(uops)
     ops = [x.op for x in uops]
@@ -518,7 +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])
@@ -559,5 +558,80 @@ class TestUOpRender(unittest.TestCase):
     u = UOp(Ops.VECTORIZE, dtype=dtypes.int.vec(3), src=(UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 1), UOp.const(dtypes.int, 2)))
     self.assertEqual(u.render(), "(0, 1, 2)")
 
+class TestZeroRange(unittest.TestCase):
+  def test_reduce_variable(self):
+    for i in range(3,-1,-1):
+      v = UOp.variable("i", 0, 5).bind(i)
+      out = Tensor.ones(10, dtype=dtypes.int).contiguous().shrink(((0,v),)).sum()
+      self.assertEqual(out.item(), i)
+
+class TestUOpPrograms(unittest.TestCase):
+  def _run(self, prog:UOp, *tensors:Tensor):
+    ExecItem(get_runner(Device.DEFAULT, prog), [t.uop.buffer for t in tensors]).run(wait=True)
+
+  def test_simple(self):
+    out = Tensor.empty(10,10,dtype=dtypes.int)
+
+    ptr = UOp.placeholder(out.shape, out.dtype, slot=0)
+    i, j = UOp.range(10, axis_id=0), UOp.range(10, axis_id=1)
+    prog = ptr[i,j].set(42).end(i,j)
+    self._run(prog.sink(), out)
+
+    with Context(DEBUG=0): self.assertTrue((out == 42).all().item())
+
+  def test_matmul(self):
+    a = Tensor.randn(10,10)
+    b = Tensor.randn(10,10)
+    c = Tensor.empty(10,10)
+    ref = (a@b)
+    with Context(DEBUG=0): Tensor.realize(a, b, c, ref)
+
+    # C[i,j] = sum_k A[i,k] * B[k,j]
+    # Shapes: A[M,K], B[K,N], C[M,N]
+    M = N = K = 10
+    DT = dtypes.float32
+
+    # Placeholders (bind slots explicitly)
+    A = UOp.placeholder((M, K), DT, slot=0)
+    B = UOp.placeholder((K, N), DT, slot=1)
+    C = UOp.placeholder((M, N), DT, slot=2)
+
+    # Axes: i,j are spatial; k is a reduction axis over the shared dim K
+    i = UOp.range(M, axis_id=0)                             # rows of A/C
+    j = UOp.range(N, axis_id=1)                             # cols of B/C
+    k = UOp.range(K, axis_id=2, axis_type=AxisType.REDUCE)  # reduction over K
+
+    # Zero-init: write a scalar 0 to each (i,j).
+    C = C[i, j].set(0.0)
+
+    # Accumulate: C_after(k) enforces the dependency along the reduction axis
+    C = C[i, j].set(C.after(k)[i, j] + A[i, k] * B[k, j])
+
+    # Finalize the loop nest / schedule in (i, j, k) order
+    prog = C.end(i, j, k)
+
+    # run program
+    # TODO: make this work with opts_to_apply
+    self._run(prog.sink(arg=KernelInfo(opts_to_apply=())), a, b, c)
+
+    with Context(DEBUG=0): self.assertLessEqual((c-ref).square().mean().item(), 1e-6)
+
+  def test_matmul_relu(self):
+    a, b, c = Tensor.randn(10,10), Tensor.randn(10,10), Tensor.empty(10,10)
+    ref = (a@b).relu()
+    with Context(DEBUG=0): Tensor.realize(a, b, c, ref)
+
+    A, B, C = a.uop.placeholder_like(0), b.uop.placeholder_like(1), c.uop.placeholder_like(2)
+    i, j, k = UOp.range(10, 0), UOp.range(10, 1), UOp.range(10, 2, axis_type=AxisType.REDUCE)
+
+    C = C[i, j].set(0.0)
+    C = C[i, j].set(C.after(k)[i, j] + A[i, k] * B[k, j], end=k)
+    C = C[i, j].set(C[i, j].maximum(0.0))
+
+    prog = C.end(i, j)
+
+    self._run(prog.sink(arg=KernelInfo(opts_to_apply=())), a, b, c)
+    with Context(DEBUG=0): self.assertLessEqual((c-ref).square().mean().item(), 1e-6)
+
 if __name__ == '__main__':
   unittest.main(verbosity=2)

test/test_uops_stats.py

@@ -141,8 +141,8 @@ class TestUOpsStats(unittest.TestCase):
     globl = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(), tuple())
     o1 = UOp(Ops.CONST, dtypes.int, tuple(), 1)
     o2 = UOp(Ops.CONST, dtypes.int, tuple(), 2)
-    u1 = UOp(Ops.LOAD, dtypes.int, (globl.index(o1),))
-    u2 = UOp(Ops.LOAD, dtypes.int, (globl.index(o2),))
+    u1 = globl.index(o1)
+    u2 = globl.index(o2)
     u3 = UOp(Ops.CONST, dtypes.int, tuple(), 3)
     u4 = UOp(Ops.MUL, dtypes.int, (u1,u2))
     u5 = UOp(Ops.ADD, dtypes.int, (u4,u3))
@@ -151,8 +151,8 @@ class TestUOpsStats(unittest.TestCase):
     globl = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(), tuple())
     o1 = UOp(Ops.CONST, dtypes.int, tuple(), 1)
     o2 = UOp(Ops.CONST, dtypes.int, tuple(), 2)
-    u1 = UOp(Ops.LOAD, dtypes.int, (globl.index(o1),))
-    u2 = UOp(Ops.LOAD, dtypes.int, (globl.index(o2),))
+    u1 = globl.index(o1)
+    u2 = globl.index(o2)
     u3 = UOp(Ops.CONST, dtypes.int, tuple(), 3)
     u4 = UOp(Ops.MULACC, dtypes.int, (u1,u2,u3))
     uops_fma = full_rewrite(u4.sink())

test/unit/test_helpers.py

@@ -99,6 +99,11 @@ class TestStripParens(unittest.TestCase):
   def test_simple(self): self.assertEqual("1+2", strip_parens("(1+2)"))
   def test_nested(self): self.assertEqual("1+(2+3)", strip_parens("(1+(2+3))"))
   def test_casted_no_strip(self): self.assertEqual("(int)(1+2)", strip_parens("(int)(1+2)"))
+  def test_unmatched_parens(self): self.assertEqual("((c35+c39>>23&255)+-127).cast(dtypes.float)",
+    strip_parens("((c35+c39>>23&255)+-127).cast(dtypes.float)"))
+  def test_single_paren_left(self): self.assertEqual("(abc", strip_parens("(abc"))
+  def test_single_paren_right(self): self.assertEqual("abc)", strip_parens("abc)"))
+  def test_parens_at_different_depths(self): self.assertEqual("(a+(b))*(c)", strip_parens("(a+(b))*(c)"))
 
 class TestProd(unittest.TestCase):
   def test_empty(self): self.assertEqual(1, prod(tuple()))

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)

test/unit/test_linalg.py

@@ -1,5 +1,5 @@
 import unittest, functools
-from tinygrad import Tensor
+from tinygrad import Tensor, Context
 import numpy as np
 
 def orthogonality_helper(A:Tensor, tolerance=1e-5):
@@ -27,15 +27,16 @@ class TestLinAlg(unittest.TestCase):
       reconstruction_helper([U,s_diag,V],a)
 
   def _test_svd_nonfull(self, size):
-    a = Tensor.randn(size).realize()
-    U,S,V = a.svd(full_matrices=False)
-    b_shape,m,n = size[0:-2],size[-2],size[-1]
-    k = min(m,n)
-    s_diag = (S.unsqueeze(-2) * Tensor.eye(k).reshape((1,) * len(b_shape) + (k,k)).expand(b_shape + (k,k)))
-    #reduced U,V is only orthogonal along smaller dim
-    if (m < n): orthogonality_helper(U),orthogonality_helper(V)
-    else: orthogonality_helper(U.transpose(-2,-1)),orthogonality_helper(V.transpose(-2,-1))
-    reconstruction_helper([U,s_diag,V],a)
+    with Context(IGNORE_OOB=1):  # sometimes this is slow in CI
+      a = Tensor.randn(size).realize()
+      U,S,V = a.svd(full_matrices=False)
+      b_shape,m,n = size[0:-2],size[-2],size[-1]
+      k = min(m,n)
+      s_diag = (S.unsqueeze(-2) * Tensor.eye(k).reshape((1,) * len(b_shape) + (k,k)).expand(b_shape + (k,k)))
+      #reduced U,V is only orthogonal along smaller dim
+      if (m < n): orthogonality_helper(U),orthogonality_helper(V)
+      else: orthogonality_helper(U.transpose(-2,-1)),orthogonality_helper(V.transpose(-2,-1))
+      reconstruction_helper([U,s_diag,V],a)
 
   # faster for parallel pytest
   def test_svd_nonfull_2_2(self): self._test_svd_nonfull((2,2))
@@ -75,4 +76,4 @@ class TestLinAlg(unittest.TestCase):
         orthogonality_helper(b if size[-1] > size[-2] else b.transpose(-2, -1), tolerance=1e-3)
 
 if __name__ == "__main__":
-  unittest.main()
+  unittest.main()

test/unit/test_simplify_valid_idx.py

@@ -9,13 +9,13 @@ from test.unit.test_uop_symbolic import check_uop_against_string
 
 def get_gated_load_uop(valid:UOp, idx:UOp):
   return UOp(Ops.LOAD, dtypes.float, (
-    UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=0).index(idx.valid(valid)),
+    UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), arg=0).index(idx.valid(valid), ptr=True),
     UOp.const(dtypes.float, 0.0)
   ))
 
 def get_load_image_uop(image_shape:tuple[int, ...], valid:UOp, idx:tuple[UOp, UOp]):
   return UOp(Ops.LOAD, dtypes.float.vec(4), (
-    UOp(Ops.DEFINE_GLOBAL, dtypes.imagef(image_shape), arg=0).index(UOp(Ops.VECTORIZE, dtypes.index.vec(2), idx).valid(valid)),
+    UOp(Ops.DEFINE_GLOBAL, dtypes.imagef(image_shape), arg=0).index(UOp(Ops.VECTORIZE, dtypes.index.vec(2), idx).valid(valid), ptr=True),
     UOp(Ops.VECTORIZE, dtypes.float.vec(4), src=(UOp.const(dtypes.float, 0.0),) * 4)
   ))
 

test/unit/test_transcendental_helpers.py

@@ -11,7 +11,7 @@ class TestTranscendentalFunctions(unittest.TestCase):
     # TODO: Test constant input when constant folding is fixed (or maybe test both variants)
     # Load input value from a buffer to prevent constant folding
     input_buf = UOp(Ops.DEFINE_GLOBAL, dtypes.double.ptr(), arg=1, src=())
-    loaded_value = UOp.load(input_buf.index(UOp.const(dtypes.int, 0)), dtype=dtypes.double)
+    loaded_value = input_buf.index(UOp.const(dtypes.int, 0))
     def eval_payne_hanek_reduction(v:float) -> tuple[float, int]:
       return tuple(eval_uop(u, [(dtypes.float64, [v])]) for u in payne_hanek_reduction(loaded_value))
 

test/unit/test_uop_symbolic.py

@@ -769,6 +769,10 @@ class TestSymbolic(unittest.TestCase):
     self.helper_test_variable(numerator, 3, 390, "(a*((a*4)+-1))")
     self.helper_test_variable((numerator//denominator)<=0, 1, 1, "True")
 
+  def test_symbolic_range_doesnt_collapse(self):
+    r0 = UOp.range((Variable("a", 1, 10)<5).cast(dtypes.index), 0)
+    self.helper_test_variable(r0, 0, 0, "r0")
+
   def test_const_reciprocal(self):
     a = Variable("a", 1, 10, dtypes.float)
     # TODO: bounds for reciprocal
@@ -1015,10 +1019,7 @@ class TestSymbolicRealWorld(unittest.TestCase):
     #print(idx.render())
     # NOTE: this used to have 13,151,129,600 in the output which is out of int32 range.
     self.assertIn(idx.render(),
-      ("((((((((((lidx5+1)//16)*802816)+(((lidx5+1)%16)*49))+(gidx0*3211264))+(gidx1*784))+(gidx2*8))+(lidx4*100352))+lidx3)+2207744)",
-       '((lidx3+((((((((lidx5+1)//16)*802816)+(((lidx5+1)%16)*49))+(gidx0*3211264))+(gidx1*784))+(gidx2*8))+(lidx4*100352)))+2207744)',
-       '((lidx3+((lidx4*100352)+((gidx2*8)+((gidx1*784)+((gidx0*3211264)+((((lidx5+1)//16)*802816)+(((lidx5+1)%16)*49)))))))+2207744)',
-      ))
+      ("(lidx3+((lidx5+1)//16*802816+(lidx5+1)%16*49+gidx0*3211264+gidx1*784+gidx2*8+lidx4*100352)+2207744)",))
 
 class TestBounds(unittest.TestCase):
   def test_unrolled_arange(self):

tinygrad/codegen/__init__.py

@@ -1,36 +1,40 @@
 from typing import cast
-from tinygrad.helpers import QUANTIZE, DEVECTORIZE, TRANSCENDENTAL, SPEC
+import itertools
+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
-from tinygrad.codegen.quantize import pm_quant
 from tinygrad.codegen.gpudims import pm_add_gpudims
 from tinygrad.uop.symbolic import sym, symbolic_simple, gep_pushing, symbolic, pm_move_where_on_load
 from tinygrad.uop.decompositions import get_late_rewrite_patterns
-from tinygrad.codegen.late.expander import migrate_indexing, expander, pm_pre_expander, pm_group_for_reduce
+from tinygrad.codegen.late.expander import expander, pm_pre_expander, pm_group_for_reduce
 from tinygrad.codegen.late.devectorizer import load_store_folding, load_store_indexing, devectorize, pm_reduce, \
-  ReduceContext, correct_load_store, pm_render
+  ReduceContext, correct_load_store, pm_render, pm_add_loads
 from tinygrad.codegen.opt.postrange import apply_opts
 from tinygrad.codegen.simplify import pm_simplify_ranges, pm_flatten_range, pm_split_ranges, pm_load_collapse, pm_split_store
-from tinygrad.schedule.rangeify import pm_add_buffers_local, rangeify_codegen
+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+pm_syntactic_sugar, name="early movement ops", bottom_up=True)
+
   # first we optimize
   if optimize:
-    if QUANTIZE and ren.device in {"CPU", "DSP"}: sink = graph_rewrite(sink, pm_quant, name="quantize")
-
-    # TODO: fix expander and remove this
-    sink = graph_rewrite(sink, pm_add_buffers_local, name="add locals early")
-
     # collapse loads reduce (indexing by a tensor)
     sink = graph_rewrite(sink, pm_load_collapse, name="load collapse")
 
@@ -50,13 +54,13 @@ def full_rewrite_to_sink(sink:UOp, ren:Renderer|None=None, optimize:bool=True) -
     sink = apply_opts(sink, ren)
 
   # ** expander (expand_rewrite) **
-  sink = graph_rewrite(sink, sym+migrate_indexing+pm_move_where_on_load, name="postopt symbolic")
+  sink = graph_rewrite(sink, sym+pm_move_where_on_load, name="postopt symbolic")
 
   # expand
   sink = graph_rewrite(sink, sym+pm_pre_expander+pm_group_for_reduce+expander, name="expander")
 
   # add locals
-  sink = graph_rewrite(sink, pm_add_buffers_local+rangeify_codegen, name="add local buffers")
+  sink = graph_rewrite(sink, pm_add_buffers_local+rangeify_codegen, ctx=itertools.count(0), name="add local buffers")
 
   # ** devectorizer (full_graph_rewrite) **
   # remove reduce
@@ -65,6 +69,11 @@ def full_rewrite_to_sink(sink:UOp, ren:Renderer|None=None, optimize:bool=True) -
   # add gpu dims (late). this works after devectorize, but it's faster here
   sink = graph_rewrite(sink, pm_add_gpudims, ctx=ren, name="add gpudims")
 
+  # **** optimizations are done, now we lower to actual code ****
+
+  # add loads
+  sink = graph_rewrite(sink, pm_add_loads, name="** add loads (code)")
+
   # devectorize (TODO: does this need opts?)
   if DEVECTORIZE >= 2: pm_devectorize = sym+load_store_folding+load_store_indexing
   elif DEVECTORIZE: pm_devectorize = sym+devectorize+load_store_folding+correct_load_store+load_store_indexing

tinygrad/codegen/gpudims.py

@@ -80,7 +80,7 @@ def add_gpudims(ctx:Renderer, s:UOp):
   subs = {}
   for r in s_topo:
     # look for local INDEXes that are not used in the GLOBAL store, then add them as an INVALID
-    if r.op is Ops.STORE and r.src[0].ptrdtype.addrspace == AddrSpace.GLOBAL:
+    if r.op is Ops.STORE and r.buf_target().ptrdtype.addrspace == AddrSpace.GLOBAL:
       idx = r.src[0]
       missing_locals = [all_ranges[rng] for rng in local_dims if all_ranges[rng] not in idx.ranges]
       if len(missing_locals):

tinygrad/codegen/late/devectorizer.py

@@ -2,9 +2,9 @@ from typing import Any, cast
 import functools, operator, itertools
 from collections import defaultdict
 from dataclasses import dataclass
-from tinygrad.dtype import dtypes, ImageDType, DType, AddrSpace, Invalid
+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
 
@@ -12,7 +12,7 @@ from tinygrad.renderer import Renderer
 
 def simplify_valid_load(buf:UOp, start_idx:UOp, valid:UOp) -> UOp|None:
   idx = uop_given_valid(valid, start_idx)
-  if not isinstance(buf.dtype, ImageDType): return None if idx is start_idx else buf.index(idx.valid(valid))
+  if not isinstance(buf.dtype, ImageDType): return None if idx is start_idx else buf.index(idx.valid(valid), ptr=True)
 
   # wait for it to be image indexed before running simplification
   if start_idx.dtype.count != 2: return None
@@ -43,12 +43,8 @@ def simplify_valid_load(buf:UOp, start_idx:UOp, valid:UOp) -> UOp|None:
 
   if not drop_stmt and idx is start_idx: return None
   new_valid = functools.reduce(operator.and_, ss) if (ss:=[s for s in valid.split_uop(Ops.AND) if s not in drop_stmt]) else None
-  return buf.index(idx.valid(new_valid) if new_valid is not None else idx)
+  return buf.index(idx.valid(new_valid) if new_valid is not None else idx, ptr=True)
 
-def delete_redundant_gates(store:UOp, buf:UOp, idx:UOp, val:UOp, store_gate:UOp, cast:UOp|None=None) -> UOp|None:
-  if store_gate not in [gate.src[0] for gate in val.toposort() if gate.op is Ops.IF]: return None
-  # remove the gate from the index
-  return UOp.store(buf.index(idx).cast(cast.dtype) if cast is not None else buf.index(idx), val, *store.src[2:])
 
 load_store_indexing = PatternMatcher([
   # image load valid idx simplification
@@ -56,10 +52,7 @@ load_store_indexing = PatternMatcher([
   # simplify away long after index has been lowered
   (UPat(Ops.INDEX, src=(UPat.var("buf"), UPat.var("x", dtypes.long), UPat.var("c", dtypes.bool))), lambda buf,x,c: simplify_valid_load(buf, x, c)),
   # drop true gate
-  (UPat(Ops.INDEX, src=(UPat.var("buf"), UPat.var("x"), UPat.const(dtypes.bool, True)),), lambda buf,x: buf.index(x)),
-  # delete_redundant_gates (after expand)
-  (UPat(Ops.STORE, src=(UPat.any(stidx:=UPat.var("buf").index(UPat.var("idx"), UPat.var("store_gate")), stidx.cast().named("cast")),
-                                  UPat.var("val")), name="store", allow_any_len=True), delete_redundant_gates),
+  (UPat(Ops.INDEX, src=(UPat.var("buf"), UPat.var("x"), UPat.const(dtypes.bool, True)),), lambda buf,x: buf.index(x, ptr=True)),
 ])
 
 # ***** load/store grouping *****
@@ -67,8 +60,8 @@ load_store_indexing = PatternMatcher([
 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)) for i in range(vec.dtype.count)]),
-                       symbolic_flat+load_store_indexing, name=f"index_buf_{buf.arg}")
+  midx = graph_rewrite(UOp.sink(*[buf.index(vec.gep(i), ptr=True) for i in range(vec.dtype.count)]),
+                       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):
@@ -148,7 +141,7 @@ def split_load_store(ctx:Renderer|None, ls:UOp, idx:UOp):
   if ctx is not None and ctx.device == "DSP":
     lengths = [128,64,32,16,8,4]
     must_divide = False
-  elif buf.dtype.base != dtypes.float and buf.dtype.base != dtypes.half and not isinstance(buf.dtype, ImageDType):
+  elif buf.dtype.base not in (dtypes.float, dtypes.half, *dtypes.fp8s) and not isinstance(buf.dtype, ImageDType):
     pass
   elif buf.ptrdtype.addrspace == AddrSpace.REG:
     pass
@@ -170,7 +163,7 @@ def split_load_store(ctx:Renderer|None, ls:UOp, idx:UOp):
     # with 1 at the end of the lengths list, this will always hit
     for fold_length in lengths:
       if global_offset+fold_length > sz: continue
-      lidx = buf.index((offset + global_offset).valid(mask))
+      lidx = buf.index((offset + global_offset).valid(mask), ptr=True)
       if fold_length > 1: lidx = lidx.cast(buf.ptrdtype.base.vec(fold_length).ptr(size=buf.ptrdtype.size, addrspace=buf.ptrdtype.addrspace))
       if ls.op is Ops.STORE: ret.append(ls.replace(src=(lidx,ls.src[1].gep(tuple(range(global_offset, global_offset+fold_length))))+ls.src[2:]))
       else: ret.append(ls.replace(src=(lidx,)+ls.src[1:], dtype=ls.dtype.scalar().vec(fold_length)))
@@ -236,7 +229,24 @@ def no_vectorized_buf(buf:UOp):
 def no_vectorized_index(buf:UOp, cast:UOp, idx:UOp):
   cnt = cast.dtype.count
   assert idx.dtype.count == 1, f"idx dtype must be 1 {idx.dtype}"
-  return buf.broadcast(cnt).index(idx.broadcast(cnt)*cnt+UOp.const(dtypes.index.vec(cnt), tuple(range(cnt))))
+  return buf.broadcast(cnt).index(idx.broadcast(cnt)*cnt+UOp.const(dtypes.index.vec(cnt), tuple(range(cnt))), ptr=True)
+
+def no_vectorized_index_broadcast(buf:UOp, cast:UOp, bcast:UOp, idx:UOp):
+  cnt = cast.dtype.count
+  precnt = bcast.dtype.vcount
+  input_gep = bcast.arg if bcast.op is Ops.GEP else ([0]*precnt)
+  gep_arg = tuple(flatten([range(precnt) for _ in range(cnt)]))
+  sum_arg = tuple(flatten([[i+y for y in input_gep] for i in range(cnt)]))
+  return buf.broadcast(cnt*precnt).index(idx.gep(gep_arg)*cnt+UOp.const(dtypes.index.vec(cnt*precnt), sum_arg), ptr=True)
+
+devectorize_buf_and_index = PatternMatcher([
+  (UPat((Ops.DEFINE_LOCAL, Ops.DEFINE_REG), name="buf"), no_vectorized_buf),
+  (UPat((Ops.DEFINE_LOCAL, Ops.DEFINE_REG)).or_after(name="buf").cast(name="cast").index(UPat.var("idx")), no_vectorized_index),
+  (UPat((Ops.DEFINE_LOCAL, Ops.DEFINE_REG)).or_after(name="buf").cast(name="cast").broadcast(name="bcast").index(UPat.var("idx")),
+   no_vectorized_index_broadcast),
+  (UPat((Ops.DEFINE_LOCAL, Ops.DEFINE_REG)).or_after(name="buf").cast(name="cast").gep(name="bcast").index(UPat.var("idx")),
+   no_vectorized_index_broadcast),
+])
 
 devectorize = PatternMatcher([
   # CAST after AFTER
@@ -244,9 +254,7 @@ devectorize = PatternMatcher([
   # no ALU on vectorized dtypes
   (UPat((*GroupOp.ALU, Ops.CAST, Ops.BITCAST), name="alu"), no_vectorized_alu),
   (UPat(Ops.WMMA, name="wmma"), no_vectorized_wmma),
-  (UPat((Ops.DEFINE_LOCAL, Ops.DEFINE_REG), name="buf"), no_vectorized_buf),
-  (UPat((Ops.DEFINE_LOCAL, Ops.DEFINE_REG)).or_after(name="buf").cast(name="cast").index(UPat.var("idx")), no_vectorized_index),
-])
+])+devectorize_buf_and_index
 
 pm_render = PatternMatcher([
   # for rendering, we use explicit VECTORIZE
@@ -291,14 +299,14 @@ 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)).load()] + lst  # put acc as the first element
+    lst = [acc.after(acc_init, *reduce_range).index(UOp.const(dtypes.int, 0))] + lst  # put acc as the first element
     ctx.acc_num += 1
   ret = functools.reduce(lambda x,y: x.alu(red.arg, y), lst)
   if len(reduce_range) == 0: return ret
-  return acc.after(acc.index(UOp.const(dtypes.int, 0)).store(ret).end(*reduce_range)).index(UOp.const(dtypes.int, 0)).load()
+  return acc.after(acc.index(UOp.const(dtypes.int, 0)).store(ret).end(*reduce_range)).index(UOp.const(dtypes.int, 0))
 
 pm_reduce = PatternMatcher([
   # REDUCE -> DEFINE_ACC+ASSIGN
@@ -307,3 +315,14 @@ pm_reduce = PatternMatcher([
   (UPat(Ops.WMMA, name="wmma") + UPat.var("add"),
     lambda add, wmma: UOp(wmma.op, wmma.dtype, (wmma.src[0], wmma.src[1], wmma.src[2]+add), wmma.arg)),
 ])+sym
+
+# add loads
+
+pm_add_loads = PatternMatcher([
+  # add loads to non ptr index
+  (UPat(Ops.INDEX, name="idx"), lambda idx: None if isinstance(idx.dtype, (PtrDType, ImageDType)) else
+    idx.replace(dtype=idx.src[0].dtype).load(dtype=idx.dtype.base)),
+  # remove loads from stores
+  (UPat(Ops.STORE, src=(UPat(Ops.LOAD),), allow_any_len=True, name="s"), lambda s: s.replace(src=(s.src[0].src[0],)+s.src[1:])),
+])
+

tinygrad/codegen/late/expander.py

@@ -1,5 +1,5 @@
 # this converts a lowerer program into a vectorized program
-import functools, itertools, operator
+import functools, itertools
 from tinygrad.dtype import dtypes, PtrDType, AddrSpace
 from tinygrad.helpers import AMX, dedup, flatten, all_same, prod, partition
 from tinygrad.uop.ops import UOp, Ops, UPat, PatternMatcher, GroupOp, AxisType, range_start
@@ -34,10 +34,7 @@ def do_expand(root:UOp):
   new_srcs = []
   for i,src in enumerate(root.src):
     if src.op is Ops.UNROLL:
-      if root.op is Ops.IF and i == 0:
-        # IF means OR on first arg to IF
-        new_srcs.append(functools.reduce(operator.__or__, [src.src[0].gep(i) for i in range(expand_sz)]))
-      elif expand_args == src.arg:
+      if expand_args == src.arg:
         # just remove the expand
         new_srcs.append(src.src[0])
       else:
@@ -47,10 +44,7 @@ def do_expand(root:UOp):
         new_srcs.append(src.src[0].gep(tuple(lst)))
     else:
       # non-UNROLL input
-      if root.op is Ops.IF or src.op is Ops.IF:
-        # for the first arg of IF, just pass them through ignoring UNROLLS
-        new_srcs.append(src)
-      elif root.op in range_start and i >= range_start[root.op]:
+      if root.op in range_start and i >= range_start[root.op]:
         # for any range args of STORE/REDUCE, pass them through
         new_srcs.append(src)
       elif root.op is Ops.INDEX and i >= 1 and not isinstance(root.dtype, PtrDType):
@@ -81,13 +75,27 @@ def do_contract(con:UOp):
     idxs += [_expand_arg_to_idx(ex.arg, {**rpk, **lrpk}) for lrpk in _choices_from_args(con.arg)]
   return UOp(Ops.UNROLL, con.dtype, (ex.src[0].gep(tuple(idxs)),), new_ex_args)
 
+def end_unrolls(u:UOp):
+  unrolls, src = partition(u.src[1:], lambda x: x.op is Ops.UNROLL)
+  if not len(unrolls): return None
+  ret = UOp(Ops.CONTRACT, dtypes.void, (u.src[0],), sum([x.arg for x in unrolls], start=()))
+  return u.replace(src=(ret,)+tuple(src))
+
 expander = PatternMatcher([
+  # push broadcast through AFTER
+  (UPat.var("x").broadcast(name="b").after(name="a", allow_any_len=True), lambda x,b,a: x.after(*a.src[1:]).broadcast(len(b.src))),
+  (UPat.var("x").broadcast(name="b").end(name="a", allow_any_len=True), lambda x,b,a: x.end(*a.src[1:]).broadcast(len(b.src))),
+  # END on UNROLL ends the UNROLL
+  (UPat(Ops.END, name="u"), end_unrolls),
+  # BUFFERIZE puts UNROLLs for ranges as contract
+  (UPat(Ops.BUFFERIZE, src=(UPat(Ops.UNROLL), UPat(Ops.UNROLL)), name="x"),
+    lambda x: x.replace(src=tuple(UOp(Ops.CONTRACT, dtype=s.dtype.vec(x.src[1].src[0].dtype.count), src=(s,), arg=x.src[1].arg) for s in x.src))),
   # double expand
   (UPat(Ops.UNROLL, name="outer", src=(UPat(Ops.UNROLL, name="inner"),)),
    lambda outer, inner: UOp(Ops.UNROLL, outer.dtype, (inner.src[0],), inner.arg+outer.arg)),
   # do expansion
   (UPat((*GroupOp.ALU, Ops.CAST, Ops.BITCAST, Ops.GEP, Ops.WMMA, Ops.LOAD, Ops.STORE, Ops.INDEX, Ops.BUFFERIZE,
-         Ops.VECTORIZE, Ops.IF, Ops.REDUCE, Ops.END), name="root", custom_early_reject=set([Ops.UNROLL])), do_expand),
+         Ops.VECTORIZE, Ops.REDUCE, Ops.END, Ops.AFTER), name="root", custom_early_reject=set([Ops.UNROLL])), do_expand),
   (UPat(Ops.CONTRACT, name="con"), do_contract),
   # BARRIERs aren't actually expanded
   (UPat(Ops.BARRIER, src=(UPat(Ops.UNROLL, name="ex"),)),
@@ -99,22 +107,6 @@ expander = PatternMatcher([
     lambda ex,x,y: UOp(Ops.UNROLL, ex.dtype, tuple((x+y).gep(i) for i in range(256 if AMX else 8)), ex.arg)),
 ])
 
-def create_gate(root:UOp) -> UOp|None:
-  @functools.cache
-  def _gate_srcs(u:UOp, gate:UOp) -> UOp:
-    if u.op is Ops.BARRIER: return u
-    if u.op is Ops.LOAD and u.src[-1].op is Ops.BARRIER:
-      return UOp(u.op, u.dtype, u.src[:-1]+(UOp(Ops.IF, src=(gate, u.src[-1])),), arg=u.arg)
-    return u if (replace_source:=tuple(_gate_srcs(x, gate) for x in u.src)) == u.src else UOp(u.op, u.dtype, replace_source, u.arg)
-  idx = root.src[0]
-  if idx.op is Ops.CAST: idx = idx.src[0]
-  return None if idx.op is not Ops.INDEX or len(idx.src) == 2 or (ret:=_gate_srcs(root, idx.src[2])) is root else ret
-
-migrate_indexing = PatternMatcher([
-  # create gate MUST BE BEFORE expander
-  (UPat(Ops.STORE, name="root"), create_gate),
-])
-
 # ****
 
 def fix_reduce_unroll(x:UOp):

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))}
@@ -64,7 +62,10 @@ class CFGContext:
       # ranges that have dependencies on other siblings need to be scheduled after them
       order = sorted(v, key=lambda x: len([u for u in v if u in deps[x]]))
       zipped = zip(order, order[1:]) if k.op is Ops.SINK else zip([k.src[1]] + order, order)
-      for x,y in zipped: self.edges[y.src[1]] = x
+      for x,y in zipped:
+        # TODO: this can happen! it causes infinite loop in shufflenet
+        assert y.src[1] not in x.backward_slice_with_self
+        self.edges[y.src[1]] = x
 
 pm_add_control_flow = PatternMatcher([
   (UPat(Ops.RANGE, name="x"), lambda ctx,x: x.replace(src=x.src+(y,)) if (y:=ctx.edges.get(x)) is not None else None),
@@ -72,7 +73,7 @@ pm_add_control_flow = PatternMatcher([
 
 def do_split_ends(e:UOp):
   ret = e.src[0]
-  for r in list(UOp.sink(*e.src[1:]).ranges)[::-1]: ret = ret.end(r)
+  for r in sorted(UOp.sink(*e.src[1:]).ranges, key=lambda x: x.arg, reverse=True): ret = ret.end(r)
   return ret
 
 pm_split_ends = PatternMatcher([

tinygrad/codegen/opt/__init__.py

@@ -2,7 +2,6 @@
 from __future__ import annotations
 from enum import Enum, auto
 from dataclasses import dataclass
-from tinygrad.uop.ops import AxisType
 
 class OptOps(Enum):
   TC = auto(); UPCAST = auto(); UNROLL = auto(); LOCAL = auto(); THREAD = auto() # noqa: E702
@@ -16,11 +15,6 @@ class Opt:
   arg: int|tuple|None = None
   def __repr__(self): return f"Opt(op={self.op}, axis={self.axis}, arg={self.arg})"
 
-axis_letters = {AxisType.GLOBAL: "g", AxisType.THREAD: "t", AxisType.LOCAL: "l", AxisType.WARP: "w", AxisType.LOOP: "L", AxisType.UPCAST: "u",
-                AxisType.GROUP_REDUCE: "G", AxisType.REDUCE: "R", AxisType.UNROLL: "r"}
-axis_colors = {AxisType.GLOBAL: "blue", AxisType.THREAD: "BLUE", AxisType.LOCAL: "cyan", AxisType.WARP: "CYAN", AxisType.LOOP: "WHITE",
-               AxisType.UPCAST: "yellow", AxisType.GROUP_REDUCE: "RED", AxisType.REDUCE: "red", AxisType.UNROLL: "magenta"}
-
 class KernelOptError(Exception): pass
 def check(cond:bool, msg:str=""):
   if not cond: raise KernelOptError(msg)

tinygrad/codegen/opt/heuristic.py

@@ -64,8 +64,8 @@ def hand_coded_optimizations(k:Scheduler) -> Scheduler:
   MV_BLOCKSIZE, MV_THREADS_PER_ROW, MV_ROWS_PER_THREAD = getenv("MV_BLOCKSIZE", 4), getenv("MV_THREADS_PER_ROW", 8), getenv("MV_ROWS_PER_THREAD", 4)
   if k.ren.has_local and getenv("MV",1) != 0 and (MV_BLOCKSIZE > 1 or MV_THREADS_PER_ROW > 1 or MV_ROWS_PER_THREAD > 1) and  \
     k.reduceop is not None and k.reduceop.arg[0] is Ops.ADD and len(k.full_shape) >= 2 and k.ren.has_shared and \
-    (mulop:=k.reduceop.src[0]).op is Ops.MUL and mulop.src[0].op is Ops.LOAD and mulop.src[1].op is Ops.LOAD:
-    idx0, idx1 = mulop.src[0].src[0].src[1].get_idx(), mulop.src[1].src[0].src[1].get_idx()
+    (mulop:=k.reduceop.src[0]).op is Ops.MUL and mulop.src[0].op is Ops.INDEX and mulop.src[1].op is Ops.INDEX:
+    idx0, idx1 = mulop.src[0].src[1].get_idx(), mulop.src[1].src[1].get_idx()
     if k.ranges_of(AxisType.REDUCE):
       first_reduce_rng = k.ranges_of(AxisType.REDUCE)[0]
       if any(u is first_reduce_rng for u in idx0.split_uop(Ops.ADD)) and all(r in idx1.ranges for r in idx0.ranges):
@@ -73,7 +73,9 @@ def hand_coded_optimizations(k:Scheduler) -> Scheduler:
           if first_reduce_rng.src[0].divides(MV_THREADS_PER_ROW) is not None and k.full_shape[global_idx]%(MV_BLOCKSIZE*MV_ROWS_PER_THREAD) == 0:
             if DEBUG >= 3:
               print(f"MATVEC: {k.full_shape=} {first_reduce_rng.render()} {MV_BLOCKSIZE=} {MV_THREADS_PER_ROW=} {MV_ROWS_PER_THREAD=}")
-            if MV_THREADS_PER_ROW > 1: k.apply_opt(Opt(OptOps.GROUP, 0, MV_THREADS_PER_ROW))
+            try:
+              if MV_THREADS_PER_ROW > 1: k.apply_opt(Opt(OptOps.GROUP, 0, MV_THREADS_PER_ROW))
+            except KernelOptError: pass
             if MV_BLOCKSIZE > 1: k.apply_opt(Opt(OptOps.LOCAL, global_idx, MV_BLOCKSIZE))
             if MV_ROWS_PER_THREAD > 1: k.apply_opt(Opt(OptOps.UPCAST, global_idx, MV_ROWS_PER_THREAD))
             return k

tinygrad/codegen/opt/postrange.py

@@ -2,11 +2,11 @@ from __future__ import annotations
 import math, itertools
 from collections import defaultdict
 from typing import cast, Final
-from tinygrad.uop.ops import PatternMatcher, UPat, Ops, UOp, KernelInfo, graph_rewrite, AxisType, ssimplify, GroupOp
+from tinygrad.uop.ops import PatternMatcher, UPat, Ops, UOp, KernelInfo, graph_rewrite, AxisType, ssimplify, GroupOp, axis_letters, axis_colors
 from tinygrad.device import Buffer
 from tinygrad.dtype import dtypes, ImageDType
 from tinygrad.helpers import colored, BEAM, getenv, DEBUG, to_function_name, NOOPT, argsort, round_up, prod, merge_dicts, get_single_element, flatten
-from tinygrad.codegen.opt import axis_colors, Opt, OptOps, KernelOptError, check, axis_letters
+from tinygrad.codegen.opt import Opt, OptOps, KernelOptError, check
 from tinygrad.codegen.simplify import pm_flatten_range
 from tinygrad.renderer import Renderer
 
@@ -63,8 +63,15 @@ class Scheduler:
     self.ast = graph_rewrite(self.ast, pm_flatten_range, name="flatten range")
     return self.ast.replace(arg=KernelInfo(name=name, applied_opts=tuple(self.applied_opts), dont_use_locals=self.dont_use_locals), tag=1)
 
+  def _output_rngs(self) -> list[UOp]:
+    return flatten([[r for r in UOp.sink(*s.src[1:]).ranges if r.arg[-1] != AxisType.REDUCE] for s in self.ast.src if s.op is Ops.END])
   def _globalizable_rngs(self) -> list[UOp]:
-    return flatten([list(UOp.sink(*s.src[1:]).ranges) for s in self.ast.src if s.op is Ops.END])
+    ret = [r for r in self._output_rngs() if r.arg[-1] == AxisType.LOOP]
+    # exclude any output ranges from global that don't appear in all BUFFERIZE
+    for x in self.ast.toposort():
+      if x.op is Ops.BUFFERIZE:
+        ret = [r for r in ret if r in x.ranges]
+    return ret
 
   def convert_loop_to_global(self):
     if not self.ren.has_local: return None
@@ -75,11 +82,13 @@ class Scheduler:
     self.ast = self.ast.substitute(dict(zip(self.rngs, rng)))
 
   def colors(self) -> list[str]:
-    output_rngs = self._globalizable_rngs()
+    output_rngs = self._output_rngs()
+    globalizible_rngs = self._globalizable_rngs()
     ret = []
     for x,r in zip(self.axis_types, self.rngs):
       if self.dont_use_locals and x == AxisType.GLOBAL: ret.append("BLUE")
       elif r not in output_rngs and x == AxisType.LOOP: ret.append("BLACK")
+      elif r not in globalizible_rngs and x == AxisType.LOOP: ret.append("white")
       else: ret.append(axis_colors[x])
     return ret
   def colored_shape(self) -> str: return ' '.join([colored(f'{x.src[0].render():>4s}', color) for x,color in zip(self.rngs, self.colors())])
@@ -334,6 +343,6 @@ def apply_opts(ast:UOp, ren:Renderer) -> UOp:
   elif not NOOPT and (ast.arg is None or ast.arg.applied_opts == ()):
     from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
     # NOTE: hand_coded_optimizations doesn't support multiblock opts yet
-    if not any(u.op is Ops.AFTER and u.src[0].op is Ops.DEFINE_LOCAL for u in ast.backward_slice):
+    if not any(u.op is Ops.BUFFERIZE for u in ast.backward_slice):
       k = hand_coded_optimizations(k)
   return k.get_optimized_ast(name_override=ast.arg.name if ast.arg is not None and ast.arg.name != "test" else None)

tinygrad/codegen/opt/search.py

@@ -1,9 +1,8 @@
-from typing import cast
 import functools, math, time, multiprocessing, traceback, signal, atexit
 from dataclasses import replace
 from tinygrad.uop.ops import sym_infer, AxisType, pyrender
 from tinygrad.device import Device, Buffer, Compiler
-from tinygrad.helpers import prod, flatten, DEBUG, CACHELEVEL, diskcache_get, diskcache_put, getenv, Context, colored, time_to_str
+from tinygrad.helpers import prod, flatten, DEBUG, CACHELEVEL, diskcache_get, diskcache_put, getenv, Context, colored, time_to_str, unwrap
 from tinygrad.helpers import IGNORE_BEAM_CACHE
 from tinygrad.codegen.opt import Opt, OptOps, KernelOptError
 from tinygrad.tensor import Tensor
@@ -50,7 +49,7 @@ def _time_program(p:ProgramSpec, lib:bytes, var_vals:dict[str, int], rawbufs:lis
       if hasattr(dev:=Device[p.device], 'invalidate_caches'): dev.invalidate_caches()
       else:
         with Context(DEBUG=0, BEAM=0, CAPTURING=0, TRACK_MATCH_STATS=0): Tensor.ones(1024,1024).contiguous().realize(do_update_stats=False)
-    tms.append(cast(float, car(input_bufs, var_vals, wait=True))*factor)
+    tms.append(unwrap(car(input_bufs, var_vals, wait=True))*factor)
     if early_stop is not None and early_stop < min(tms): break
   return tms
 
@@ -59,7 +58,7 @@ def timeout_handler(signum, frame):
   if DEBUG >= 2: print("*** BEAM COMPILE TIMEOUT")
   raise TimeoutException()
 
-def _try_compile_linearized_w_idx(x:tuple[int,Scheduler], compiler:Compiler) -> tuple[int, tuple[ProgramSpec, bytes, float]|None]:
+def _try_compile(x:tuple[int,Scheduler], compiler:Compiler) -> tuple[int, tuple[ProgramSpec, bytes, float]|None]:
   if hasattr(signal, "alarm"):
     signal.signal(getattr(signal, 'SIGALRM'), timeout_handler)
     # set timeout
@@ -93,42 +92,42 @@ def _ensure_buffer_alloc(bufs:list[Buffer]) -> list[Buffer]: return [buf.ensure_
 # *** external API ***
 
 # get dictionary of all possible actions
-def get_kernel_actions(lin:Scheduler, include_0=True, candidates:list[Opt]|None=None) -> dict[int, Scheduler]:
-  acted_lins, max_up, max_lcl = {0:lin} if include_0 else {}, getenv("BEAM_UPCAST_MAX", 256), getenv("BEAM_LOCAL_MAX", 1024)
+def get_kernel_actions(s:Scheduler, include_0=True, candidates:list[Opt]|None=None) -> dict[int, Scheduler]:
+  acted, max_up, max_lcl = {0:s} if include_0 else {}, getenv("BEAM_UPCAST_MAX", 256), getenv("BEAM_LOCAL_MAX", 1024)
   kernel_actions = (actions if candidates is None else candidates).copy()
 
   for i,a in enumerate(kernel_actions):
     if a.axis is not None and a.op is not OptOps.TC:
-      try: ax = lin.real_axis(a.op, a.axis)
+      try: ax = s.real_axis(a.op, a.axis)
       except KernelOptError: continue
-      if (ax >= lin.shape_len) or (lin.full_shape[ax] == a.arg and Opt(a.op, a.axis, 0) in kernel_actions): continue
-    lin2 = lin.copy()
+      if (ax >= s.shape_len) or (s.full_shape[ax] == a.arg and Opt(a.op, a.axis, 0) in kernel_actions): continue
+    s2 = s.copy()
     try:
-      lin2.apply_opt(a)
-      up, lcl, tc_up = 1, 1, prod(tc.dims)//tc.threads if hasattr(lin2, 'tensor_core') and (tc:=lin2.tensor_core) else 1
-      for s,c in zip(lin2.full_shape, lin2.axis_types):
-        if c in (AxisType.UPCAST, AxisType.UNROLL): up *= s
-        elif c in (AxisType.WARP, AxisType.LOCAL, AxisType.GROUP_REDUCE): lcl *= s
+      s2.apply_opt(a)
+      up, lcl, tc_up = 1, 1, prod(tc.dims)//tc.threads if hasattr(s2, 'tensor_core') and (tc:=s2.tensor_core) else 1
+      for x,t in zip(s2.full_shape, s2.axis_types):
+        if t in (AxisType.UPCAST, AxisType.UNROLL): up *= x
+        elif t in (AxisType.WARP, AxisType.LOCAL, AxisType.GROUP_REDUCE): lcl *= x
       if up//tc_up > max_up or lcl > max_lcl:
         if getenv("BEAM_LOG_SURPASS_MAX"): print(f"too many upcast/local. {up//tc_up=}, {max_up=}, {lcl=}, {max_lcl=}")
         continue
-      acted_lins[i+1] = lin2
+      acted[i+1] = s2
     except KernelOptError: pass
-  return acted_lins
+  return acted
 
 beam_pool, BEAM_DEBUG = None, getenv("BEAM_DEBUG")
-def beam_search(lin:Scheduler, rawbufs:list[Buffer], amt:int, allow_test_size=True, disable_cache=IGNORE_BEAM_CACHE.value):
+def beam_search(s:Scheduler, rawbufs:list[Buffer], amt:int, allow_test_size=True, disable_cache=IGNORE_BEAM_CACHE.value):
   global beam_pool
-  key = {"ast": lin.ast.key, "amt": amt, "allow_test_size": allow_test_size, "device": lin.ren.device, "suffix": lin.ren.suffix}
+  key = {"ast": s.ast.key, "amt": amt, "allow_test_size": allow_test_size, "device": s.ren.device, "suffix": s.ren.suffix}
   if not disable_cache and CACHELEVEL >= 1 and (val:=diskcache_get("beam_search", key)) is not None:
-    ret = lin.copy()
-    for o in val[len(lin.applied_opts):]: ret.apply_opt(o)
+    ret = s.copy()
+    for o in val[len(s.applied_opts):]: ret.apply_opt(o)
     return ret
 
-  beam: list[tuple[Scheduler, float]] = [(lin, float("inf"))]
+  beam: list[tuple[Scheduler, float]] = [(s, float("inf"))]
   seen_libs = set()
 
-  default_parallel = multiprocessing.cpu_count() if lin.ren.device in {"CUDA", "AMD", "NV", "METAL", "HIP"} else 0
+  default_parallel = multiprocessing.cpu_count() if s.ren.device in {"CUDA", "AMD", "NV", "METAL", "HIP"} else 0
   if beam_pool is None and (workers := getenv("PARALLEL", default_parallel)):
     beam_pool = multiprocessing.get_context("spawn").Pool(workers, _init_worker, (), getenv("BEAM_MAX_TASKS_PER_CHILD", 16))
     @atexit.register
@@ -137,20 +136,20 @@ def beam_search(lin:Scheduler, rawbufs:list[Buffer], amt:int, allow_test_size=Tr
   min_progress = getenv("BEAM_MIN_PROGRESS", 0.01)/1e6
   if BEAM_DEBUG:
     print("BEAM_SEARCH:")
-    print(pyrender(lin.ast.replace(arg=None)))
-  if DEBUG >= 2: print(f"   0.00s:                from   1 ->   1 actions {lin.colored_shape()}")
+    print(pyrender(s.ast.replace(arg=None)))
+  if DEBUG >= 2: print(f"   0.00s:                from   1 ->   1 actions {s.colored_shape()}")
 
   try:
     rawbufs = _ensure_buffer_alloc(rawbufs)
-    var_vals: dict[str, int] = {k.expr:int(k.vmax+k.vmin)//2 for k in lin.ast.variables()}
+    var_vals: dict[str, int] = {k.expr:int(k.vmax+k.vmin)//2 for k in s.ast.variables()}
     exiting, st = False, time.perf_counter()
-    dev = Device[lin.ren.device]
+    dev = Device[s.ren.device]
     while not exiting:
-      acted_lins: list[Scheduler] = flatten([get_kernel_actions(lin, include_0=False).values() for lin,_ in beam])
-      timed_lins: list[tuple[Scheduler, float]] = []
-      _compile_fn = functools.partial(_try_compile_linearized_w_idx, compiler=dev.compiler)
+      candidates: list[Scheduler] = flatten([get_kernel_actions(si, include_0=False).values() for si,_ in beam])
+      timed: list[tuple[Scheduler, float]] = []
+      _compile_fn = functools.partial(_try_compile, compiler=dev.compiler)
       least_compute_ops = math.inf
-      for i,proc in (map(_compile_fn, enumerate(acted_lins)) if beam_pool is None else beam_pool.imap_unordered(_compile_fn, enumerate(acted_lins))):
+      for i,proc in (map(_compile_fn, enumerate(candidates)) if beam_pool is None else beam_pool.imap_unordered(_compile_fn, enumerate(candidates))):
         if proc is None: continue
         p, lib, compile_et = proc
         if lib in seen_libs: continue
@@ -163,26 +162,26 @@ def beam_search(lin:Scheduler, rawbufs:list[Buffer], amt:int, allow_test_size=Tr
         try: tms = _time_program(p, lib, var_vals, rawbufs, early_stop=beam[0][1]*3 if len(beam) else 1.0,
                                  allow_test_size=allow_test_size, clear_l2=hasattr(dev, 'invalidate_caches'))
         except Exception as e:
-          if BEAM_DEBUG: print(f"BEAM failed for opts: {acted_lins[i].applied_opts}\n{e}")
+          if BEAM_DEBUG: print(f"BEAM failed for opts: {candidates[i].applied_opts}\n{e}")
           if isinstance(e, RuntimeError): continue
           raise
-        timed_lins.append((acted_lins[i], min(tms)))
+        timed.append((candidates[i], min(tms)))
         if BEAM_DEBUG > 1:
-          print(f"{time.perf_counter() - st:7.2f}s: {i:5d} {len(cast(list, p.uops)):5d} uops",
-                f"{time_to_str(compile_et, w=12)} compile/{time_to_str(timed_lins[-1][1], w=12)} run",
-                f"      {len(timed_lins):4d}/{len(acted_lins):4d}         {timed_lins[-1][0].colored_shape()}")
+          print(f"{time.perf_counter() - st:7.2f}s: {i:5d} {len(unwrap(p.uops)):5d} uops",
+                f"{time_to_str(compile_et, w=12)} compile/{time_to_str(timed[-1][1], w=12)} run",
+                f"      {len(timed):4d}/{len(candidates):4d}         {timed[-1][0].colored_shape()}")
         elif DEBUG >= 2:
-          print(f"\r{time.perf_counter() - st:7.2f}s: {time_to_str(timed_lins[-1][1], w=12)}",
-                f"      {len(timed_lins):4d}/{len(acted_lins):4d}         {timed_lins[-1][0].colored_shape()}\033[K", end="")
+          print(f"\r{time.perf_counter() - st:7.2f}s: {time_to_str(timed[-1][1], w=12)}",
+                f"      {len(timed):4d}/{len(candidates):4d}         {timed[-1][0].colored_shape()}\033[K", end="")
 
       # done
-      opts = sorted(timed_lins, key=lambda x: x[1])
+      opts = sorted(timed, key=lambda x: x[1])
       exiting = len(opts) == 0 or (opts[0][1] < min_progress) or (len(beam) > 0 and ((beam[0][1]-opts[0][1]) < min_progress))
       if not exiting: beam = opts[:amt]
       elif len(opts) > 0 and opts[0][1] < beam[0][1]: beam = opts[:1]
       if DEBUG >= 2:
         print(f"\r{time.perf_counter() - st:7.2f}s:", colored(time_to_str(beam[0][1], w=12), "green" if exiting else None),
-              f"from {len(acted_lins):3d} -> {len(opts):3d} actions\033[K", beam[0][0].colored_shape())
+              f"from {len(candidates):3d} -> {len(opts):3d} actions\033[K", beam[0][0].colored_shape())
   except KeyboardInterrupt as e:
     if beam_pool is not None: beam_pool.terminate()
     raise e

tinygrad/codegen/opt/tc.py

@@ -111,7 +111,7 @@ amd_rdna4 = [TensorCore(dims=(16,16,16), threads=32, elements_per_thread=(8,8,8)
   for di,do in [(dtypes.half,dtypes.float),(dtypes.half,dtypes.half),(dtypes.bfloat16,dtypes.float),(dtypes.bfloat16,dtypes.bfloat16)]]
 
 # https://gpuopen.com/learn/amd-lab-notes/amd-lab-notes-matrix-cores-readme
-amd_cdna = [TensorCore(dims=(16,16,16), threads=64, elements_per_thread=(4,4,4), dtype_in=di, dtype_out=do,
+amd_cdna_161616 = [TensorCore(dims=(16,16,16), threads=64, elements_per_thread=(4,4,4), dtype_in=di, dtype_out=do,
   opts=("l0","l0","l0","l0","u1","u1","l1","l1"),
   swizzle=((('u0', 'u1', 'l4', 'l5', 'r2', 'r3'), ('r0', 'r1'), ('l0', 'l1', 'l2', 'l3')),
            (('l0', 'l1', 'l2', 'l3', 'r2', 'r3'), ('r0', 'r1'), ('l4', 'l5', 'u0', 'u1'))))
@@ -119,11 +119,13 @@ amd_cdna = [TensorCore(dims=(16,16,16), threads=64, elements_per_thread=(4,4,4),
 
 amd_cdna_161632 = [TensorCore(dims=(16,16,32), threads=64, elements_per_thread=(8,8,4), dtype_in=di, dtype_out=do,
   opts=("l0","l0","l0","l0","u1","u1","l1","l1"),
-  swizzle=((('u0','u1','l4','l5','r3','r4'), ('r0','r1'), ('l0','l1','l2','l3','r2')),
-           (('l0','l1','l2','l3','r3','r4'), ('r0','r1'), ('l4','l5','u0','u1','r2'))))
-  for di,do in [(dtypes.half,dtypes.float),(dtypes.bfloat16,dtypes.float)]]
+  swizzle=((('u0', 'u1', 'l4', 'l5', 'r3', 'r4'), ('r0', 'r1'), ('l0', 'l1', 'l2', 'l3', 'r2')),
+           (('l0', 'l1', 'l2', 'l3', 'r3', 'r4'), ('r0', 'r1'), ('l4', 'l5', 'u0', 'u1', 'r2'))))
+  for di,do in [(dtypes.fp8e5m2,dtypes.float),(dtypes.fp8e4m3,dtypes.float),(dtypes.half,dtypes.float),(dtypes.bfloat16,dtypes.float)]]
 
-amd_cdna4 = amd_cdna_161632 + amd_cdna
+amd_cdna3 = amd_cdna_161632[:2] + amd_cdna_161616
+
+amd_cdna4 = amd_cdna_161632 + amd_cdna_161616
 
 # ***** Apple Metal *****
 

tinygrad/codegen/quantize.py

@@ -1,59 +0,0 @@
-from tinygrad.dtype import dtypes, least_upper_dtype
-from tinygrad.uop.ops import UOp, Ops, PatternMatcher, UPat
-from tinygrad.uop.symbolic import symbolic
-
-# **** this is the "quantization preprocessor", it makes ONNX quantized models, and probably also others, actually use ints ****
-# this is badly tested and low quality. remove it?
-
-FP = (1 << 15)
-pm_quant = symbolic+PatternMatcher([
-  # cast after add/mul
-  (UPat.var("x").cast(dtypes.float32) + UPat.var("y").cast(dtypes.float32),
-   lambda x,y: (x.cast(least_upper_dtype(x.dtype, y.dtype))+y.cast(least_upper_dtype(x.dtype, y.dtype))).cast(dtypes.float32)),
-  (UPat.var("x").cast(dtypes.float32) * UPat.var("y").cast(dtypes.float32),
-   lambda x,y: (x.cast(least_upper_dtype(x.dtype, y.dtype))*y.cast(least_upper_dtype(x.dtype, y.dtype))).cast(dtypes.float32)),
-
-  # masked MUL after masked ADD
-  ((UPat.var("x") + UPat.var("v").where(UPat.var('cadd'), UPat(Ops.CONST, arg=0))) * UPat.var("v").where(UPat.var('cmul'), UPat(Ops.CONST, arg=0)),
-   lambda x,v,cadd,cmul: x*v.where(cmul, 0)+v.where(cadd*cmul, 0)),
-
-  # MUL after reduce
-  (UPat(Ops.REDUCE_AXIS, src=(UPat.var("x") * UPat.cvar("c"),), name="r"), lambda x,c,r: r.replace(src=(x,))*c.arg),
-  # CAST after reduce (doesn't work if it's a size change)
-  (UPat(Ops.REDUCE_AXIS, src=(UPat(Ops.CAST, src=(UPat.var("x"),)),), name="r"),
-    lambda x,r: r.replace(dtype=x.dtype, src=(x,)).cast(r.dtype) if dtypes.is_float(r.dtype) else None),
-
-  # x*c1 + y*c2 -> (x+y)*c1 (if c1 and c2 are close floats)
-  (UPat.var("x")*UPat.cvar("c1", dtype=dtypes.floats) + UPat.var("y")*UPat.cvar("c2", dtype=dtypes.floats),
-   lambda x,y,c1,c2: (x+y)*c1 if abs(c1.arg-c2.arg) < 1e-9 else None),
-
-  # const push through add
-  ((UPat.var("x")*UPat.cvar("c1") + UPat.var("y")*UPat.cvar("c2")) * UPat.cvar("c3"), lambda x,y,c1,c2,c3: (x*c1*c3) + (y*c2*c3)),
-
-  # fixed point mult, replace (x.float()*c1+c2).int() with an int expression
-  ((UPat.var("x").cast(dtypes.float)*UPat.var("c1")+UPat.var("cc")).cast(dtypes.int),
-   lambda x,c1,cc: ((x*(c1*FP).cast(x.dtype) + (cc*FP).cast(x.dtype)) // FP).cast(dtypes.int)),
-  # fixed point mult, replace (x.float()*c1 + y.float()*c2)*cc.int() with an int expression
-  ((UPat.var("x").cast(dtypes.float)*UPat.var("c1")+UPat.var("y").cast(dtypes.float)*UPat.var("c2")+UPat.var("cc")).cast(dtypes.int),
-   lambda x,c1,y,c2,cc: ((x*(c1*FP).cast(x.dtype) + y.cast(x.dtype)*(c2*FP).cast(x.dtype) + (cc*FP).cast(x.dtype)) // FP).cast(dtypes.int)),
-
-  # where move
-  (UPat.var("valid").where(UPat.var("yes"), UPat(Ops.CONST, arg=0))*UPat.var("mul"), lambda valid, yes, mul:
-    (yes*mul*valid.where(UOp.const(mul.dtype, 1), UOp.const(mul.dtype, 0))) if yes.op is not Ops.CONST or yes.arg != 1 else None),
-  ((UPat.var("x")*UPat.cvar("c"))*(UPat.var().where(UPat(Ops.CONST, arg=1), UPat(Ops.CONST, arg=0)).named("v")), lambda x,c,v: (x*v)*c),
-  (UPat.var("x").cast().named('c') * UPat.var('valid').where(UPat(Ops.CONST, arg=1), UPat(Ops.CONST, arg=0)), lambda x,c,valid:
-    (x*valid.where(UOp.const(x.dtype, 1), UOp.const(x.dtype, 0))).cast(c.dtype)),
-  ((UPat.var('x') * UPat.var('v1').where(UPat(Ops.CONST, arg=1), UPat(Ops.CONST, arg=0)) *
-    UPat.var('v2').where(UPat(Ops.CONST, arg=1), UPat(Ops.CONST, arg=0))).named("mul"), lambda x, mul, v1, v2:
-    x * (v1&v2).where(UOp.const(mul.dtype, 1), UOp.const(mul.dtype, 0))),
-
-  # where on two adds
-  (UPat.var("x") + UPat.var("v").where(UPat.var("a0"), UPat.var("a1")) + UPat.var("v").where(UPat.var("b0"), UPat.var("b1")),
-    lambda x,v,a0,a1,b0,b1: x + v.where(a0+b0, a1+b1)),
-
-  # split REDUCE into multiple reduces (who remembers FOIL?)
-  (UPat(Ops.REDUCE_AXIS, src=((UPat(Ops.CAST, name="v1")+UPat.var("c1")) * UPat(Ops.CAST, name="v2"),), name="r"),
-    lambda v1,v2,c1,r: r.replace(src=(v1*v2,)) + r.replace(src=(c1*v2,))),
-  (UPat(Ops.REDUCE_AXIS, src=((UPat(Ops.CAST, name="v1")+UPat.var("c1")) * (UPat(Ops.CAST, name="v2",)+UPat.var("c2")),), name="r"),
-    lambda v1,v2,c1,c2,r: r.replace(src=(v1*v2,)) + r.replace(src=(c2*v1,)) + r.replace(src=(c1*v2,)) + r.replace(src=(c1*c2,))),
-])

tinygrad/codegen/simplify.py

@@ -1,5 +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
 
@@ -18,9 +19,8 @@ pm_flatten_range = PatternMatcher([
 def count_divmod(x:UOp): return len([u for u in x.toposort() if u.op in {Ops.IDIV, Ops.MOD}])
 def simplify_merge_adjacent(u:UOp) -> UOp|None:
   reduce_ranges = [x.ranges for x in u.backward_slice_with_self if x.op is Ops.REDUCE]
-  i = 0
-  while i < len(u.ended_ranges)-1:
-    r0, r1 = u.ended_ranges[i], u.ended_ranges[i+1]
+  # on END we only want to merge adjacent ranges, on REDUCE we want to try all combinations
+  for r0, r1 in (zip(u.ended_ranges, u.ended_ranges[1:]) if u.op is Ops.END else itertools.permutations(u.ended_ranges, 2)):
     # check same type
     if r0.arg[-1] == r1.arg[-1]:
       # check if the ranges to merge are in the same reduces
@@ -28,14 +28,13 @@ 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
         if count_divmod(nidx) <= count_divmod(u):
           u = nidx
           continue
-    i += 1
   return u
 
 pm_simplify_ranges = PatternMatcher([
@@ -56,7 +55,7 @@ def do_substitute(ctx, x: UOp):
   return ret
 
 def dont_sub_ranges_for_image(ctx, x:UOp):
-  if isinstance(x.src[0].dtype, ImageDType):
+  if isinstance(x.src[0].src[0].dtype, ImageDType):
     for s in x.src[0].ranges: ctx[s] = None
 
 pm_split_ranges = PatternMatcher([
@@ -92,47 +91,59 @@ pm_reduce_collapse = pm_reduce_unparented + PatternMatcher([
   # lift x*y out of reduce
   ((UPat.var("x")*UPat.var("y")) < UPat.var("c"), lambda x,y,c: (x < ((c+y-1) // y)) if no_range(y) and no_range(c) and y.vmin > 0 else None),
   # fold the range
-  ((UPat(Ops.RANGE, name="r") < UPat.var("cut")).where(0, UPat.cvar("val")).reduce(UPat.var("r"), arg=Ops.ADD),
-   lambda r,cut,val: (r.src[0]-cut).maximum(0).minimum(r.src[0]).cast(val.dtype) * val),
-  (((UPat.var("r")<UPat.var("lower")).logical_not()&(UPat(Ops.RANGE, name="r")<UPat.var("upper"))).where(UPat.cvar("val"), 0).reduce(UPat.var("r"),
-    arg=Ops.ADD), lambda r,lower,upper,val: (upper.minimum(r.src[0])-lower.maximum(0)).maximum(0).minimum(r.src[0]).cast(val.dtype) * val),
-  ((UPat(Ops.RANGE, name="r") < UPat.var("cut")).where(UPat.cvar("val"), 0).reduce(UPat.var("r"), arg=Ops.ADD),
-   lambda r,cut,val: cut.maximum(0).minimum(r.src[0]).cast(val.dtype) * val),
+  # bound from below
+  ((UPat(Ops.RANGE, name="r") < UPat.var("cut")).where(0, UPat.var("val")).reduce(UPat.var("r"), arg=Ops.ADD),
+   lambda r,cut,val: (r.src[0]-cut).maximum(0).minimum(r.src[0]).cast(val.dtype) * val if no_range(val) else None),
+  # bound from two sides
+  (((UPat.var("r")<UPat.var("lower")).logical_not()&(UPat(Ops.RANGE, name="r")<UPat.var("upper"))).where(UPat.var("val"), 0).reduce(UPat.var("r"),
+    arg=Ops.ADD), lambda r,lower,upper,val:
+      (upper.minimum(r.src[0])-lower.maximum(0)).maximum(0).minimum(r.src[0]).cast(val.dtype) * val if no_range(val) else None),
+  # bound from above
+  ((UPat(Ops.RANGE, name="r") < UPat.var("cut")).where(UPat.var("val"), 0).reduce(UPat.var("r"), arg=Ops.ADD),
+   lambda r,cut,val: cut.maximum(0).minimum(r.src[0]).cast(val.dtype) * val if no_range(val) else None),
   # REDUCE on ADD
   ((UPat.var("x")+UPat.var("y")).reduce(arg=Ops.ADD, allow_any_len=True, name="r"),
    lambda x,y,r: x.reduce(*r.src[1:], arg=Ops.ADD) + y.reduce(*r.src[1:],arg=Ops.ADD)),
-])+symbolic_flat
-
-pm_reduce_load_collapse = PatternMatcher([
+  # AND on WHERE
+  ((UPat(Ops.DEFINE_VAR, name="x") & UPat.var("y")).where(UPat.var("c"), 0).reduce(arg=Ops.ADD, allow_any_len=True, name="r"),
+    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
+
+pm_reduce_load_collapse = pm_reduce_collapse + PatternMatcher([
   # lift x+y out of reduce on ne
   ((UPat.var("x")+UPat.var("y")).or_casted() != UPat.var("c"), lambda x,y,c: (x != (c.cast(y.dtype)-y)) if no_range(y) and no_range(c) else None),
   # reduce on gated load becomes can substitute the range and remove the reduce
   ((UPat.var("idx")!=(UPat(Ops.RANGE, name="r").or_casted())).where(0, UPat.var("expr")).reduce(UPat.var("r"), arg=Ops.ADD),
    lambda r,idx,expr: (v:=(idx.cast(r.dtype) >= 0) & (idx.cast(r.dtype) < r.src[0])).where(expr.substitute({r:idx.cast(r.dtype).valid(v)}),0)),
-])+symbolic_flat
+])
 
-def reduce_collapse(red:UOp, pm=pm_reduce_collapse):
-  included = red.src[0].toposort(gate=lambda x: any(y in x.ranges for y in red.src[1:]))
-  if any(x.op in {Ops.STORE, Ops.REDUCE} for x in included): return None
-  replaces: dict[UOp, UOp] = {}
-  for u in included:
-    for s in u.src:
-      if s in included or s in replaces or s.op in {Ops.CONST, Ops.VCONST, Ops.DEFINE_GLOBAL, Ops.DEFINE_LOCAL, Ops.DEFINE_VAR}: continue
-      replaces[s] = UOp(Ops.DEFINE_VAR, dtype=s.dtype, arg=(f'in{len(replaces)}', s.vmin, s.vmax))
-  collapse_fxn = red.substitute(replaces)
-  sink = graph_rewrite(collapse_fxn, pm, name="reduce_collapse")
-  return sink.substitute({v:k for k,v in replaces.items()}) if no_range(sink) else None
+def reduce_collapse(red:UOp, u:UOp, pm=pm_reduce_collapse):
+  for r in red.src[1:]:
+    included = u.toposort(gate=lambda x: r in x.ranges)
+    if any(x.op in {Ops.STORE, Ops.REDUCE} for x in included): return None
+    replaces: dict[UOp, UOp] = {}
+    for u in included:
+      for s in u.src:
+        if s in included or s in replaces or s.op in {Ops.CONST, Ops.VCONST, Ops.DEFINE_GLOBAL, Ops.DEFINE_LOCAL, Ops.DEFINE_VAR}: continue
+        replaces[s] = UOp(Ops.DEFINE_VAR, dtype=s.dtype, arg=(f'in{len(replaces)}', s.vmin, s.vmax))
+    collapse_fxn = u.substitute(replaces).reduce(r, arg=Ops.ADD)
+    sink = graph_rewrite(collapse_fxn, pm, name="reduce_collapse")
+    if not no_range(sink): return None
+    u = sink.substitute({v:k for k,v in replaces.items()})
+  return u
 
-def reduce_load_collapse(red:UOp): return reduce_collapse(red, pm=pm_reduce_load_collapse)
+def reduce_load_collapse(red:UOp, u:UOp): return reduce_collapse(red, u, pm=pm_reduce_load_collapse)
 
-# remove REDUCE without loads (generic arange opt / indexing). TODO: support multi range
-pm_reduce_simplify = pm_reduce_unparented + PatternMatcher([(UPat(Ops.REDUCE, src=(UPat(), UPat()), name="red"), reduce_collapse),])
+# remove REDUCE without loads (generic arange opt / indexing).
+pm_reduce_simplify = pm_reduce_unparented + PatternMatcher([
+  (UPat(Ops.REDUCE, src=(UPat.var("u"),), allow_any_len=True, arg=Ops.ADD, name="red"), reduce_collapse),
+])
 # remove REDUCE on load, comes from indexing a tensor with another tensor
-def no_load(u:UOp) -> bool: return not any(x.op is Ops.LOAD for x in u.backward_slice_with_self)
+def no_load(u:UOp) -> bool: return not any(x.op is Ops.INDEX for x in u.backward_slice_with_self)
 pm_load_collapse = PatternMatcher([
-  (UPat(Ops.REDUCE, src=(UPat(), UPat()), name="red"), reduce_load_collapse),
+  (UPat(Ops.REDUCE, src=(UPat.var("u"), UPat()), name="red"), reduce_load_collapse),
   # we want to make sure we dont do math on a loaded index since that can cause overflow, this undoes the rule in pm_reduce_load_collapse
   ((UPat.var("x", dtypes.index)+UPat.var("y"))<UPat.var("c"), lambda x,y,c: x < c-y if no_load(y) and no_load(c) and not no_load(x) else None),
 ])

tinygrad/device.py

@@ -5,7 +5,7 @@ from typing import Any, Generic, TypeVar, Iterator, Sequence, cast, Generator
 import importlib, inspect, functools, pathlib, os, platform, contextlib, sys, re, atexit, pickle, decimal
 from tinygrad.helpers import CI, OSX, LRU, getenv, diskcache_get, diskcache_put, DEBUG, GlobalCounters, flat_mv, PROFILE, temp, colored, CPU_LLVM
 from tinygrad.helpers import Context, DISABLE_COMPILER_CACHE, ALLOW_DEVICE_USAGE, MAX_BUFFER_SIZE, cpu_events, ProfileEvent, ProfilePointEvent, dedup
-from tinygrad.helpers import unwrap_class_type, suppress_finalizing
+from tinygrad.helpers import unwrap_class_type, suppress_finalizing, AMD_LLVM
 from tinygrad.dtype import DType, ImageDType, PtrDType, dtypes, _to_np_dtype
 from tinygrad.renderer import Renderer
 
@@ -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
@@ -333,6 +333,7 @@ def is_dtype_supported(dtype:DType, device:str|None=None) -> bool:
     return device in {"AMD", "PYTHON", "NULL"}
   if dtype in dtypes.fp8s:
     if device in {"CUDA", "NV"}: return not CI and not getenv(f"{device}_PTX") and not getenv("NV_NAK")
+    if device == "AMD": return not CI and not AMD_LLVM and getattr(Device["AMD"], "target") in {(9,4,2), (9,5,0)}
     return device in {"PYTHON", "NULL"}
   if device == "WEBGPU": return dtype in [dtypes.bool, dtypes.char, dtypes.uchar, dtypes.short,
                                           dtypes.ushort, dtypes.float, dtypes.int32, dtypes.uint32, dtypes.half]

tinygrad/dtype.py

@@ -221,9 +221,9 @@ def can_safe_cast(dt0:DType, dt1:DType) -> bool:
   if dt0 == dt1 or dt0 == dtypes.bool: return True
   match dt1:
     case dtypes.index: return dt0 in dtypes.ints
-    case dtypes.double: return dt0 in (dtypes.float, dtypes.half, dtypes.bfloat16,
+    case dtypes.double: return dt0 in (dtypes.float, dtypes.half, dtypes.bfloat16, *dtypes.fp8s,
       dtypes.uint32, dtypes.uint16, dtypes.uint8, dtypes.int32, dtypes.int16, dtypes.int8)
-    case dtypes.float: return dt0 in (dtypes.half, dtypes.bfloat16, dtypes.uint16, dtypes.uint8, dtypes.int16, dtypes.int8)
+    case dtypes.float: return dt0 in (dtypes.half, dtypes.bfloat16, *dtypes.fp8s, dtypes.uint16, dtypes.uint8, dtypes.int16, dtypes.int8)
     case dtypes.uint64: return dt0 in (dtypes.uint32, dtypes.uint16, dtypes.uint8)
     case dtypes.uint32: return dt0 in (dtypes.uint16, dtypes.uint8)
     case dtypes.int64: return dt0 in (dtypes.uint32, dtypes.uint16, dtypes.uint8, dtypes.int32, dtypes.int16, dtypes.int8)

tinygrad/engine/realize.py

@@ -3,6 +3,7 @@ import time, pprint, random, itertools, math
 from dataclasses import dataclass, replace, field
 from tinygrad.helpers import all_same, colored, DEBUG, GlobalCounters, ansilen, BEAM, NOOPT, all_int, CAPTURING, Metadata, TRACEMETA, TracingKey
 from tinygrad.helpers import DEVECTORIZE, time_to_str, VALIDATE_WITH_CPU, getenv, cpu_profile, PROFILE, ProfilePointEvent, cpu_events, prod, Context
+from tinygrad.helpers import unwrap
 from tinygrad.uop.ops import Ops, PatternMatcher, UOp, UPat, sym_infer, graph_rewrite, print_uops, track_rewrites, KernelInfo, pyrender
 from tinygrad.device import Device, Buffer
 from tinygrad.renderer import Renderer, ProgramSpec, Estimates
@@ -78,6 +79,7 @@ def optimize_local_size(_prg:Callable, global_size:list[int], rawbufs:list[Buffe
 
 class CompiledRunner(Runner):
   def __init__(self, p:ProgramSpec, precompiled:bytes|None=None, prg=None):
+    if DEBUG >= 3: print(p.applied_opts)
     if DEBUG >= 4: print(p.src)
     self.p:ProgramSpec = p
     if precompiled is not None: self.lib = precompiled
@@ -90,7 +92,8 @@ class CompiledRunner(Runner):
 
   def __reduce__(self): return self.__class__, (self.p, self.lib)
 
-  def __call__(self, rawbufs:list[Buffer], var_vals:dict[str, int], wait=False) -> float|None:
+  def __call__(self, rawbufs:list[Buffer], var_vals:dict[str, int]|None=None, wait=False) -> float|None:
+    if var_vals is None: var_vals = {}
     has_local = Device[self.p.device].renderer.has_local
     global_size, local_size = self.p.launch_dims(var_vals)
     if has_local and global_size is not None and local_size is None and all_int(self.p.global_size): # type: ignore[arg-type]
@@ -164,7 +167,7 @@ class ExecItem:
   fixedvars: dict[str, int] = field(default_factory=dict)
   def run(self, _var_vals:dict[str, int]|None=None, wait=False, jit=False, do_update_stats=True) -> float|None:
     var_vals = self.fixedvars if _var_vals is None else (_var_vals|self.fixedvars)
-    bufs = [cast(Buffer, x) for x in self.bufs] if jit else [cast(Buffer, x).ensure_allocated() for x in self.bufs]
+    bufs = [unwrap(x) for x in self.bufs] if jit else [unwrap(x).ensure_allocated() for x in self.bufs]
     if PROFILE:
       payload = {"metadata":self.metadata, "var_vals":var_vals, "bufs":[b.trace_num for b in bufs], "name":self.prg.display_name}
       payload["outputs"], payload["inputs"] = (self.prg.p.outs, self.prg.p.ins) if isinstance(self.prg, CompiledRunner) else ([0], [1])

tinygrad/gradient.py

@@ -29,15 +29,18 @@ 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)),
+  (UPat(Ops.KERNEL, name="k"), lambda ctx, k: k.arg.grad_fxn(ctx, k)),
   # there's no gradient for bitcast
   (UPat(Ops.BITCAST), lambda: (None,)),
 ])

tinygrad/helpers.py

@@ -44,7 +44,8 @@ def fully_flatten(l):
     return flattened
   return [l]
 def fromimport(mod, frm): return getattr(__import__(mod, fromlist=[frm]), frm)
-def strip_parens(fst:str): return fst[1:-1] if fst[0] == '(' and fst[-1] == ')' and fst[1:-1].find('(') <= fst[1:-1].find(')') else fst
+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
 def round_down(num:int, amt:int) -> int: return -round_up(-num, amt)
@@ -85,7 +86,9 @@ def word_wrap(x, wrap=80):
   while len(ansistrip(x[:i])) < wrap and i < len(x): i += 1
   return x[:i] + "\n" + word_wrap(x[i:], wrap)
 def pad_bytes(b:bytes, align:int) -> bytes: return b + b'\x00' * ((align - (len(b) % align)) % align)
-def panic(e:Exception): raise e
+def panic(e:Exception|None=None):
+  if e is None: raise RuntimeError("PANIC!")
+  raise e
 
 @functools.cache
 def canonicalize_strides(shape:tuple[T, ...], strides:tuple[T, ...]) -> tuple[T, ...]:
@@ -164,10 +167,9 @@ SPLIT_REDUCEOP, NO_MEMORY_PLANNER, RING = ContextVar("SPLIT_REDUCEOP", 1), Conte
 PICKLE_BUFFERS, LRU = ContextVar("PICKLE_BUFFERS", 1), ContextVar("LRU", 1)
 CACHELEVEL, IGNORE_BEAM_CACHE, DEVECTORIZE = ContextVar("CACHELEVEL", 2), ContextVar("IGNORE_BEAM_CACHE", 0), ContextVar("DEVECTORIZE", 1)
 DISABLE_COMPILER_CACHE = ContextVar("DISABLE_COMPILER_CACHE", 0)
-QUANTIZE, VALIDATE_WITH_CPU, DISABLE_FAST_IDIV = ContextVar("QUANTIZE", 0), ContextVar("VALIDATE_WITH_CPU", 0), ContextVar("DISABLE_FAST_IDIV", 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

@@ -1,16 +1,15 @@
-from typing import TypeVar
+from typing import Self
 from tinygrad.uop import Ops
 from tinygrad.dtype import dtypes, ConstType
 
-TMT = TypeVar("TMT", bound="MathTrait")
-class MathTrait:
+class MathMixin:
   # required to implement
-  def alu(self:TMT, op:Ops, *src:TMT) -> TMT: raise NotImplementedError
-  def const_like(self:TMT, b:ConstType) -> TMT: raise NotImplementedError
+  def alu(self, op:Ops, *src:Self) -> Self: raise NotImplementedError
+  def const_like(self, b:ConstType) -> Self: raise NotImplementedError
 
   # great functions you get!
-  def ufix(self:TMT, x:TMT|ConstType) -> TMT: return self.const_like(x) if not isinstance(x, MathTrait) else x
-  def _binop(self:TMT, op:Ops, x:TMT|ConstType, reverse:bool) -> TMT:
+  def ufix(self, x:Self|ConstType) -> Self: return self.const_like(x) if not isinstance(x, MathMixin) else x
+  def _binop(self, op:Ops, x:Self|ConstType, reverse:bool) -> Self:
     return self.ufix(x).alu(op, self) if reverse else self.alu(op, self.ufix(x))
   def logical_not(self): return self.ne(True)
   def neg(self):
@@ -20,7 +19,7 @@ class MathTrait:
     if (dtype:=getattr(self, 'dtype')) is not None:
       if isinstance(dtype, tuple): dtype = dtype[0]
       if not (dtypes.is_bool(dtype) or dtypes.is_int(dtype)): raise RuntimeError(f"{dtype} is not supported")
-  def add(self:TMT, x:TMT|ConstType, reverse:bool=False):
+  def add(self, x:Self|ConstType, reverse:bool=False):
     """
     Adds `self` and `x`.
     Equivalent to `self + x`.
@@ -38,7 +37,7 @@ class MathTrait:
     ```
     """
     return self._binop(Ops.ADD, x, reverse)
-  def mul(self:TMT, x:TMT|ConstType, reverse:bool=False):
+  def mul(self, x:Self|ConstType, reverse:bool=False):
     """
     Multiplies `self` and `x`.
     Equivalent to `self * x`.
@@ -57,7 +56,7 @@ class MathTrait:
     ```
     """
     return self._binop(Ops.MUL, x, reverse)
-  def bitwise_and(self:TMT, x:TMT|ConstType, reverse:bool=False):
+  def bitwise_and(self, x:Self|ConstType, reverse:bool=False):
     """
     Computes the bitwise AND of `self` and `x`.
     Equivalent to `self & x`.
@@ -71,7 +70,7 @@ class MathTrait:
     """
     self._check_dtype()
     return self._binop(Ops.AND, x, reverse)
-  def bitwise_or(self:TMT, x:TMT|ConstType, reverse:bool=False):
+  def bitwise_or(self, x:Self|ConstType, reverse:bool=False):
     """
     Computes the bitwise OR of `self` and `x`.
     Equivalent to `self | x`.
@@ -85,7 +84,7 @@ class MathTrait:
     """
     self._check_dtype()
     return self._binop(Ops.OR, x, reverse)
-  def bitwise_xor(self:TMT, x:TMT|ConstType, reverse:bool=False):
+  def bitwise_xor(self, x:Self|ConstType, reverse:bool=False):
     """
     Computes bitwise xor of `self` and `x`.
     Equivalent to `self ^ x`.
@@ -100,7 +99,7 @@ class MathTrait:
     """
     self._check_dtype()
     return self._binop(Ops.XOR, x, reverse)
-  def idiv(self:TMT, x:TMT|ConstType, reverse:bool=False):
+  def idiv(self, x:Self|ConstType, reverse:bool=False):
     """
     Divides `self` by `x`.
     Equivalent to `self // x`.
@@ -112,62 +111,62 @@ class MathTrait:
     ```
     """
     return self._binop(Ops.IDIV, x, reverse)
-  def mod(self:TMT, x:TMT|ConstType, reverse:bool=False): return self._binop(Ops.MOD, x, reverse)
-  def sub(self:TMT, x:TMT|ConstType, reverse:bool=False): return self.ufix(x).alu(Ops.ADD, -self) if reverse else self.alu(Ops.ADD, self.ufix(-x))
-  def div(self:TMT, x:TMT|ConstType, reverse:bool=False):
+  def mod(self, x:Self|ConstType, reverse:bool=False): return self._binop(Ops.MOD, x, reverse)
+  def sub(self, x:Self|ConstType, reverse:bool=False): return self.ufix(x).alu(Ops.ADD, -self) if reverse else self.alu(Ops.ADD, self.ufix(-x))
+  def div(self, x:Self|ConstType, reverse:bool=False):
     return (self.ufix(x)*self.alu(Ops.RECIPROCAL)) if reverse else (self*self.ufix(x).alu(Ops.RECIPROCAL))
 
   def __neg__(self): return self.neg()
 
-  def __add__(self:TMT, x:TMT|ConstType): return self.add(x)
-  def __sub__(self:TMT, x:TMT|ConstType): return self.sub(x)
-  def __mul__(self:TMT, x:TMT|ConstType): return self.mul(x)
-  def __truediv__(self:TMT, x:TMT|ConstType): return self.div(x)
-  def __floordiv__(self:TMT, x:TMT|ConstType): return self.idiv(x)  # TODO: idiv is trunc div, not floordiv
-  def __mod__(self:TMT, x:TMT|ConstType): return self.mod(x)
-  def __and__(self:TMT, x:TMT|ConstType): return self.bitwise_and(x)
-  def __or__(self:TMT, x:TMT|ConstType): return self.bitwise_or(x)
-  def __xor__(self:TMT, x:TMT|ConstType): return self.bitwise_xor(x)
+  def __add__(self, x:Self|ConstType): return self.add(x)
+  def __sub__(self, x:Self|ConstType): return self.sub(x)
+  def __mul__(self, x:Self|ConstType): return self.mul(x)
+  def __truediv__(self, x:Self|ConstType): return self.div(x)
+  def __floordiv__(self, x:Self|ConstType): return self.idiv(x)  # TODO: idiv is trunc div, not floordiv
+  def __mod__(self, x:Self|ConstType): return self.mod(x)
+  def __and__(self, x:Self|ConstType): return self.bitwise_and(x)
+  def __or__(self, x:Self|ConstType): return self.bitwise_or(x)
+  def __xor__(self, x:Self|ConstType): return self.bitwise_xor(x)
 
-  def __radd__(self:TMT, x:TMT|ConstType): return self.add(x, True)
-  def __rsub__(self:TMT, x:TMT|ConstType): return self.sub(x, True)
-  def __rmul__(self:TMT, x:TMT|ConstType): return self.mul(x, True)
-  def __rtruediv__(self:TMT, x:TMT|ConstType): return self.div(x, True)
-  def __rfloordiv__(self:TMT, x:TMT|ConstType): return self.idiv(x, True)
-  def __rand__(self:TMT, x:TMT|ConstType): return self.bitwise_and(x, True)
-  def __ror__(self:TMT, x:TMT|ConstType): return self.bitwise_or(x, True)
-  def __rxor__(self:TMT, x:TMT|ConstType): return self.bitwise_xor(x, True)
-  def __rmod__(self:TMT, x:TMT|ConstType): return self.mod(x, True)
+  def __radd__(self, x:Self|ConstType): return self.add(x, True)
+  def __rsub__(self, x:Self|ConstType): return self.sub(x, True)
+  def __rmul__(self, x:Self|ConstType): return self.mul(x, True)
+  def __rtruediv__(self, x:Self|ConstType): return self.div(x, True)
+  def __rfloordiv__(self, x:Self|ConstType): return self.idiv(x, True)
+  def __rand__(self, x:Self|ConstType): return self.bitwise_and(x, True)
+  def __ror__(self, x:Self|ConstType): return self.bitwise_or(x, True)
+  def __rxor__(self, x:Self|ConstType): return self.bitwise_xor(x, True)
+  def __rmod__(self, x:Self|ConstType): return self.mod(x, True)
 
-  def __lt__(self:TMT, x:TMT|ConstType): return self.alu(Ops.CMPLT, self.ufix(x))
-  def __gt__(self:TMT, x:TMT|ConstType): return self.ufix(x).alu(Ops.CMPLT, self)
-  def __ge__(self:TMT, x:TMT|ConstType): return (self < x).logical_not()
-  def __le__(self:TMT, x:TMT|ConstType): return (self > x).logical_not()
+  def __lt__(self, x:Self|ConstType): return self.alu(Ops.CMPLT, self.ufix(x))
+  def __gt__(self, x:Self|ConstType): return self.ufix(x).alu(Ops.CMPLT, self)
+  def __ge__(self, x:Self|ConstType): return (self < x).logical_not()
+  def __le__(self, x:Self|ConstType): return (self > x).logical_not()
 
-  def ne(self:TMT, x:TMT|ConstType): return self.alu(Ops.CMPNE, self.ufix(x))
-  def eq(self:TMT, x:TMT|ConstType): return self.ne(x).logical_not()
-  def __ne__(self:TMT, x:TMT|ConstType): return self.ne(x)  # type: ignore[override]
+  def ne(self, x:Self|ConstType): return self.alu(Ops.CMPNE, self.ufix(x))
+  def eq(self, x:Self|ConstType): return self.ne(x).logical_not()
+  def __ne__(self, x:Self|ConstType): return self.ne(x)  # type: ignore[override]
   # NOTE: __eq__ isn't overridden, and means the same thing as is by default
 
-  def lshift(self:TMT, x:TMT|int, reverse:bool=False): return self._binop(Ops.SHL, x, reverse)
-  def rshift(self:TMT, x:TMT|int, reverse:bool=False): return self._binop(Ops.SHR, x, reverse)
-  def __lshift__(self:TMT, x:TMT|int): return self.lshift(x)
-  def __rshift__(self:TMT, x:TMT|int): return self.rshift(x)
-  def __rlshift__(self:TMT, x:TMT|int): return self.lshift(x, True)
-  def __rrshift__(self:TMT, x:TMT|int): return self.rshift(x, True)
+  def lshift(self, x:Self|int, reverse:bool=False): return self._binop(Ops.SHL, x, reverse)
+  def rshift(self, x:Self|int, reverse:bool=False): return self._binop(Ops.SHR, x, reverse)
+  def __lshift__(self, x:Self|int): return self.lshift(x)
+  def __rshift__(self, x:Self|int): return self.rshift(x)
+  def __rlshift__(self, x:Self|int): return self.lshift(x, True)
+  def __rrshift__(self, x:Self|int): return self.rshift(x, True)
 
-  def maximum(self:TMT, x:TMT|ConstType): return self.alu(Ops.MAX, self.ufix(x))
-  def minimum(self:TMT, x:TMT|ConstType): return -(-self).maximum(-x)
-  def where(self:TMT, x:TMT|ConstType, y:TMT|ConstType):
+  def maximum(self, x:Self|ConstType): return self.alu(Ops.MAX, self.ufix(x))
+  def minimum(self, x:Self|ConstType): return -(-self).maximum(-x)
+  def where(self, x:Self|ConstType, y:Self|ConstType):
     if isinstance(x, type(self)): return self.alu(Ops.WHERE, x, x.ufix(y))
     if isinstance(y, type(self)): return self.alu(Ops.WHERE, y.ufix(x), y)
     raise RuntimeError("where needs at least one UOp arg")
-  def threefry(self:TMT, seed:TMT): return self.alu(Ops.THREEFRY, seed)
+  def threefry(self, seed:Self): return self.alu(Ops.THREEFRY, seed)
   def reciprocal(self): return self.alu(Ops.RECIPROCAL)
   def trunc(self): return self.alu(Ops.TRUNC)
   def sqrt(self): return self.alu(Ops.SQRT)
   def sin(self): return self.alu(Ops.SIN)
   def log2(self): return self.alu(Ops.LOG2)
   def exp2(self): return self.alu(Ops.EXP2)
-  def pow(self:TMT, x:TMT|ConstType): return self.alu(Ops.POW, self.ufix(x))
-  def __pow__(self:TMT, x:TMT|ConstType): return self.pow(x)
+  def pow(self, x:Self|ConstType): return self.alu(Ops.POW, self.ufix(x))
+  def __pow__(self, x:Self|ConstType): return self.pow(x)

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/__init__.py

@@ -1,13 +1,12 @@
 from __future__ import annotations
-from typing import Callable, cast, TYPE_CHECKING
+from typing import Callable, cast
 import functools
 from dataclasses import dataclass, field
 from tinygrad.helpers import to_function_name, dedup, prod
 from tinygrad.uop.ops import Ops, UOp, sym_infer, sint, Variable, ssimplify, GroupOp, PatternMatcher
 from tinygrad.dtype import AddrSpace, PtrDType
-if TYPE_CHECKING:
-  from tinygrad.codegen.opt.tc import TensorCore
-  from tinygrad.codegen.opt import Opt
+from tinygrad.codegen.opt.tc import TensorCore
+from tinygrad.codegen.opt import Opt
 
 @dataclass(frozen=True)
 class Estimates:
@@ -30,7 +29,7 @@ class Estimates:
     if ignore_indexing:
       def range_gate(x): return x.op is not Ops.RANGE
       for u in uops:
-        if u.op in {Ops.LOAD, Ops.STORE} and (not isinstance(u.src[0].dtype, PtrDType) or u.src[0].dtype.addrspace != AddrSpace.REG):
+        if u.op in {Ops.LOAD, Ops.STORE}:
           # if u.src[0] is INDEX, we have to include the buffer since it might be an AFTER
           dont_count = dont_count.union((UOp.sink(*u.src[0].src[1:]) if u.src[0].op is Ops.INDEX else u.src[0]).toposort(range_gate))
           # TODO: is this correct? this all needs to be cleaned up
@@ -81,16 +80,15 @@ class ProgramSpec:
       for u in self.uops:
         if u.op is Ops.DEFINE_VAR: self.vars.append(u)
         if u.op is Ops.DEFINE_GLOBAL: self.globals.append(u.arg)
-        if u.op is Ops.STORE and (u.src[0].op is Ops.INDEX or (u.src[0].op is Ops.CAST and u.src[0].src[0].op is Ops.INDEX)):
-          idx = u.src[0] if u.src[0].op is Ops.INDEX else u.src[0].src[0]
-          if (buf:=idx.src[0]).op is Ops.DEFINE_GLOBAL: self.outs.append(buf.arg)
-        if u.op is Ops.LOAD and (u.src[0].op is Ops.INDEX or (u.src[0].op is Ops.CAST and u.src[0].src[0].op is Ops.INDEX)):
-          idx = u.src[0] if u.src[0].op is Ops.INDEX else u.src[0].src[0]
-          if (buf:=idx.src[0]).op is Ops.DEFINE_GLOBAL: self.ins.append(buf.arg)
+        if u.op in (Ops.STORE, Ops.LOAD):
+          if (idx:=u.src[0]).op is Ops.INDEX or (u.src[0].op is Ops.CAST and (idx:=u.src[0].src[0]).op is Ops.INDEX):
+            if (buf:=idx.src[0]).op is Ops.DEFINE_GLOBAL: (self.outs if u.op is Ops.STORE else self.ins).append(buf.arg)
+          # TODO: can else happen?
         if u.op is Ops.SPECIAL:
           # NOTE: you have to set local_size and global_size to the base [1,1,1] outside this
           if u.arg[0] == 'i': self.local_size = None
           special_size = self.local_size if u.arg[0] == 'l' else self.global_size
+          # TODO: this cast is wrong, u.src[0].ssimplify() can be sint
           if special_size is not None: special_size[int(u.arg[-1])] = cast(int, u.src[0].ssimplify())
       self.vars = sorted(self.vars, key=lambda v: v.arg)
       self.outs = sorted(dedup(self.outs))
@@ -105,8 +103,10 @@ class ProgramSpec:
   def function_name(self) -> str: return to_function_name(self.name)
 
   @property
-  def applied_opts(self) -> tuple[Opt, ...]|None: return self.uops[-1].arg.applied_opts if \
-    self.uops is not None and self.uops[-1].op is Ops.SINK and self.uops[-1].arg is not None else None
+  def applied_opts(self) -> tuple[Opt, ...]|None:
+    if self.uops is None: return None
+    assert self.uops[-1].op is Ops.SINK, self.uops[-1].op
+    return self.uops[-1].arg.applied_opts
 
   def launch_dims(self, var_vals:dict[str, int]):
     global_size = [sym_infer(sz, var_vals) for sz in self.global_size] if self.global_size is not None else None

tinygrad/renderer/cstyle.py

@@ -1,8 +1,8 @@
 from typing import Literal, Callable, cast
 import os, math, sys
 from collections import defaultdict, Counter
-from tinygrad.codegen.opt import tc, axis_letters
-from tinygrad.uop.ops import GroupOp, Ops, UOp, PatternMatcher, UPat, range_str
+from tinygrad.codegen.opt import tc
+from tinygrad.uop.ops import GroupOp, Ops, UOp, PatternMatcher, UPat, range_str, axis_letters
 from tinygrad.helpers import strip_parens, getenv, prod, dedup, AMX, CPU_COUNT
 from tinygrad.dtype import ImageDType, dtypes, DType, PtrDType, AddrSpace, truncate
 from tinygrad.renderer import Renderer
@@ -71,11 +71,26 @@ extra_pm = PatternMatcher([
   (UPat(Ops.WHERE, name="alu"), no_vectorized_alu),
 ])
 
+def create_non_native_float_pats(dts:tuple[DType, ...], casting:bool=True):
+  patterns = PatternMatcher([
+    (UPat(Ops.WHERE, src=(UPat.var("b"), UPat.var("x", dtype=dts), UPat.var("y", dtype=dts))),
+     lambda b,x,y: UOp(Ops.WHERE, dtype=dtypes.float, src=(b,x.cast(dtypes.float),y.cast(dtypes.float))).cast(x.dtype)),
+    (UPat(GroupOp.ALU, dtype=dts, name="x"),
+     lambda x: UOp(x.op, dtypes.float, tuple(vv.cast(dtypes.float) for vv in x.src), x.arg).cast(x.dtype)),
+    (UPat(GroupOp.ALU, dtypes.bool, name="alu", src=(UPat.var("x", dtype=dts), UPat.var("y", dtype=dts))),
+     lambda alu,x,y: UOp(alu.op, dtypes.bool, (x.cast(dtypes.float), y.cast(dtypes.float)), alu.arg))])
+  if casting:
+    # add float intermediate casting
+    patterns += PatternMatcher([
+      (UPat(Ops.CAST, dts, (UPat.var("x"),), name="y"), lambda x,y: x.cast(dtypes.float).cast(y.dtype) if x.dtype!=dtypes.float else None),
+      (UPat(Ops.CAST, name="x", src=(UPat.var("y", dts),)), lambda x,y: y.cast(dtypes.float).cast(x.dtype) if x.dtype!=dtypes.float else None)])
+  return patterns
+
 def uops_to_dtypes(uops:list[UOp]) -> list[DType]: return dedup(u.dtype for u in uops if not isinstance(u.dtype, (ImageDType, PtrDType)))
 
 # (name, dims, dtype_in, dtype_out, device, threads, upcast_axes, reduce_axes)
 def wmma_args(uops:list[UOp]):
-  return dedup((uop.arg[0], uop.arg[1], uop.src[0].dtype.scalar(), uop.dtype.scalar(), *(uop.arg[4:8])) for uop in uops if uop.op is Ops.WMMA)
+  return dedup((uop.arg[0], uop.arg[1], uop.arg[2], uop.dtype.scalar(), *(uop.arg[4:8])) for uop in uops if uop.op is Ops.WMMA)
 
 class CStyleLanguage(Renderer):
   kernel_typedef: str = "void"
@@ -367,12 +382,8 @@ class CUDARenderer(CStyleLanguage):
     Ops.SQRT: lambda x,dtype: f"hsqrt({x})" if dtype in (dtypes.half, dtypes.bfloat16) else f"sqrt({x})",
     Ops.RECIPROCAL: lambda x,dtype: f"hrcp({x})" if dtype in (dtypes.half, dtypes.bfloat16) else f"(1/{x})" }
   type_map = {dtypes.bfloat16: "nv_bfloat16", dtypes.fp8e4m3: "__nv_fp8_e4m3", dtypes.fp8e5m2: "__nv_fp8_e5m2"}
-  extra_matcher = PatternMatcher([
+  extra_matcher = create_non_native_float_pats(dtypes.fp8s, casting=False) + PatternMatcher([
     (UPat(Ops.CAST, dtypes.fp8s, UPat.var("x", dtypes.fp8s), name='y'), lambda x,y: x.cast(dtypes.float).cast(y.dtype) if x.dtype!=y.dtype else None),
-    (UPat(GroupOp.ALU, dtype=dtypes.fp8s, name="x"),
-     lambda x: UOp(x.op, dtypes.float, tuple(vv.cast(dtypes.float) for vv in x.src), x.arg).cast(x.dtype)),
-    (UPat(GroupOp.ALU, dtypes.bool, name="alu", src=(UPat.var("x", dtype=dtypes.fp8s), UPat.var("y", dtype=dtypes.fp8s))),
-     lambda alu,x,y: UOp(alu.op, dtypes.bool, (x.cast(dtypes.float), y.cast(dtypes.float)), alu.arg)),
   ]) + extra_pm
   def render_vector_prefix(self, dt:DType) -> str:
     vec, scal = self.render_dtype(dt), self.render_dtype(dt.scalar()),
@@ -388,7 +399,7 @@ class CUDARenderer(CStyleLanguage):
     if any(dt.scalar() == dtypes.half for dt in used_dtypes): prefix.append("#include <cuda_fp16.h>")
     if any(dt.scalar() == dtypes.bfloat16 for dt in used_dtypes): prefix.append("#include <cuda_bf16.h>")
     prefix += [self.render_vector_prefix(dt) for dt in used_dtypes if (dt.count in (4,8) and dt.scalar() in {dtypes.half, dtypes.bfloat16})
-      or (dt.count in (8,16) and dt.scalar() in dtypes.fp8s)]
+      or (dt.count in (2,4,8,16) and dt.scalar() in dtypes.fp8s)]
     dt_map_in = { dtypes.float: "tf32", dtypes.half: "f16", dtypes.bfloat16: "bf16", dtypes.fp8e4m3: "e4m3", dtypes.fp8e5m2: "e5m2" }
     dt_map_out = { dtypes.float: "f32", dtypes.half: "f16" }
     for name, (N, M, K), dtype_in, dtype_out, _, _, upcast_axes, _ in wmma_args(uops):
@@ -423,13 +434,20 @@ class AMDRenderer(CStyleLanguage):
 
   @staticmethod
   def get_tensor_cores(arch):
-    return {"gfx942": tc.amd_cdna, "gfx950": tc.amd_cdna4, "gfx1200": tc.amd_rdna4, "gfx1201": tc.amd_rdna4}.get(arch.split(":")[0], tc.amd_rdna3)
+    return {"gfx942": tc.amd_cdna3, "gfx950": tc.amd_cdna4, "gfx1200": tc.amd_rdna4, "gfx1201": tc.amd_rdna4}.get(arch.split(":")[0], tc.amd_rdna3)
+  @staticmethod
+  def is_cdna(arch): return arch.split(":")[0] in {"gfx942", "gfx950"}
   def __init__(self, arch:str): # gfx942 => MI300, gfx1100 => RX 7900, gfx1201 => RX 9700
     self.arch = arch
     self.tensor_cores = self.get_tensor_cores(arch)
-    if self.tensor_cores == tc.amd_cdna:
+    if self.is_cdna(self.arch):
       self.string_rewrite = PatternMatcher([
-        (UPat(Ops.WMMA, name="x"), lambda ctx,x: f"__{x.arg[0]}({ctx[x.src[0]]}, {ctx[x.src[1]]}, {ctx[x.src[2]]}, 0, 0, 0)")]) + base_rewrite
+        (UPat(Ops.WMMA, name="x"), lambda ctx,x: f"__{x.arg[0]}({ctx[x.src[0]]}, {ctx[x.src[1]]}, {ctx[x.src[2]]}, 0, 0, 0)"),
+        (UPat(Ops.CAST, dtypes.fp8s, (UPat.var("y", dtypes.float),), name="x",),
+          lambda ctx,x, y: f"f32_to_fp8({ctx[x.src[0]]}, {'1' if x.dtype == dtypes.fp8e5m2 else '0'})"),
+        (UPat(Ops.CAST, dtypes.float, (UPat.var("y", dtypes.fp8s),), name="x",),
+          lambda ctx,x, y: f"__builtin_amdgcn_cvt_f32_{'bf8' if y.dtype == dtypes.fp8e5m2 else 'fp8'}((unsigned int){ctx[x.src[0]]}, 0)"),
+      ]) + base_rewrite
   def __reduce__(self): return self.__class__, (self.arch,)
 
   # language options
@@ -455,20 +473,11 @@ class AMDRenderer(CStyleLanguage):
   barrier = '__builtin_amdgcn_fence(__ATOMIC_RELEASE, "workgroup");' + '__builtin_amdgcn_s_barrier();' + \
             '__builtin_amdgcn_fence(__ATOMIC_ACQUIRE, "workgroup");'
   float4 = "make_float4"
-  type_map = {dtypes.bfloat16: "hip_bfloat16"}
-  extra_matcher = PatternMatcher([
-    # cast bfloat16 alus to float
-    (UPat(Ops.WHERE, src=(UPat.var("b"), UPat.var("x", dtype=dtypes.bfloat16), UPat.var("y", dtype=dtypes.bfloat16))),
-      lambda b,x,y: UOp(Ops.WHERE, dtype=dtypes.float, src=(b,x.cast(dtypes.float),y.cast(dtypes.float))).cast(dtypes.bfloat16)),
-    (UPat(GroupOp.ALU, dtype=dtypes.bfloat16, name="x"),
-      lambda x: UOp(x.op, dtypes.float, tuple(vv.cast(dtypes.float) for vv in x.src), x.arg).cast(dtypes.bfloat16)),
-    (UPat(GroupOp.ALU, dtypes.bool, name="alu", src=(UPat.var("x", dtype=dtypes.bfloat16), UPat.var("y", dtype=dtypes.bfloat16))),
-      lambda alu,x,y: UOp(alu.op, dtypes.bool, (x.cast(dtypes.float), y.cast(dtypes.float)), alu.arg)),
-    # add float intermediate casting for bfloat16
-    (UPat(Ops.CAST, name="x", src=(UPat.var("y", dtypes.bfloat16),)),
-      lambda x,y: y.cast(dtypes.float).cast(x.dtype) if x.dtype!=dtypes.float else None),
-    (UPat(Ops.CAST, dtypes.bfloat16, (UPat.var("x"),)),
-      lambda x: x.cast(dtypes.float).cast(dtypes.bfloat16) if x.dtype!=dtypes.float else None),
+  type_map = {dtypes.bfloat16: "hip_bfloat16", dtypes.fp8e4m3: "hip_fp8", dtypes.fp8e5m2: "hip_bf8"}
+  extra_matcher = create_non_native_float_pats((dtypes.bfloat16, *dtypes.fp8s)) + PatternMatcher([
+    (UPat(Ops.WMMA, name="x", dtype=dtypes.float.vec(4)),
+      lambda x: UOp(Ops.WMMA, x.dtype, (x.src[0].bitcast(dtypes.uint64), x.src[1].bitcast(dtypes.uint64),
+        x.src[2]), (*x.arg,)) if x.src[0].dtype in (dtypes.fp8e4m3.vec(8), dtypes.fp8e5m2.vec(8)) else None),
     # bfloat16 casting
     (UPat.cvar('x', dtypes.bfloat16), lambda x: cast_float_to_bf16(UOp.const(dtypes.float, x.arg))),
     (UPat(Ops.CAST, dtypes.float, (UPat.var("x", dtypes.bfloat16),)),
@@ -482,14 +491,21 @@ class AMDRenderer(CStyleLanguage):
 
   def render_kernel(self, function_name, kernel, bufs, uops, prefix=None) -> str:
     prefix = ["#define INFINITY (__builtin_inff())","#define NAN (__builtin_nanf(\"\"))","typedef long unsigned int size_t;","#define half _Float16"]
-    type_map = { dtypes.bfloat16: "bf16", dtypes.float: "f32", dtypes.half: "f16" }
+    type_map = { dtypes.bfloat16: "bf16", dtypes.float: "f32", dtypes.half: "f16", dtypes.fp8e4m3: "_fp8_fp8", dtypes.fp8e5m2: "_bf8_bf8" }
     used_dtypes = uops_to_dtypes(uops)
     if any(dt.scalar() == dtypes.bfloat16 for dt in used_dtypes): prefix.append("typedef unsigned short hip_bfloat16;")
+    if any(dt.scalar() in dtypes.fp8s for dt in used_dtypes):
+      prefix += ["typedef unsigned char hip_bf8;", "typedef unsigned char hip_fp8;"]
+      prefix.append("""static inline __attribute__((device)) unsigned char f32_to_fp8(float v, int is_bf8) {
+  v = (((*(unsigned*)&v)&0x7F800000)!=0x7F800000)?__builtin_amdgcn_fmed3f(v,is_bf8?57344.0f:448.0f,is_bf8?-57344.0f:-448.0f) : v;
+  return (unsigned char)(is_bf8?__builtin_amdgcn_cvt_pk_bf8_f32(v,v,0,false):__builtin_amdgcn_cvt_pk_fp8_f32(v,v,0,false));\n}""")
     prefix += [self.render_vector_prefix(dt) for dt in used_dtypes if dt.count > 1]
 
-    for name, _, dtype_in, dtype_out, _, _, _, _ in wmma_args(uops): # TODO: handle TCs f32_bf16 and bf16_bf16 w/ wrapper
-      if self.tensor_cores == tc.amd_cdna:
-        prefix.append(f"#define __{name} __builtin_amdgcn_mfma_f32_16x16x16{'f16' if dtype_in == dtypes.half else 'bf16_1k'}")
+    for name, (N, M, K), dtype_in, dtype_out, _, _, _, _ in wmma_args(uops): # TODO: handle TCs f32_bf16 and bf16_bf16 w/ wrapper
+      if self.is_cdna(self.arch):
+        if (N, M, K) == (16, 16, 16): type_map[dtypes.bfloat16] = 'bf16_1k'
+        elif (N, M, K) == (16, 16, 32): type_map = {**type_map, dtypes.bfloat16: "_bf16", dtypes.half: "_f16"}
+        prefix.append(f"#define __{name} __builtin_amdgcn_mfma_f32_{N}x{M}x{K}{type_map[dtype_in]}")
       # #define __WMMA_16_16_16_half_half __builtin_amdgcn_wmma_f16_16x16x16_f16_w32_gfx12
       elif self.tensor_cores == tc.amd_rdna4:
         prefix.append(f"#define __{name} __builtin_amdgcn_wmma_{type_map[dtype_out]}_16x16x16_{type_map[dtype_in]}_w32_gfx12")

tinygrad/renderer/llvmir.py

@@ -107,14 +107,20 @@ base_rewrite = PatternMatcher([
 
   # range
   (UPat(Ops.RANGE, name="r"), lambda ctx,r:
-   f"  br label %loop_entry_{range_str(r)}\nloop_entry_{range_str(r)}:\n"
-   f"  br label %loop_body_{range_str(r)}\nloop_body_{range_str(r)}:\n"
-   f"  {ctx[r]} = phi {ldt(r.dtype)} [ 0, %loop_entry_{range_str(r)} ], [ {ctx[r]}phi, %loop_latch_{range_str(r)} ]"),
-  (UPat(Ops.END, src=(UPat(), UPat(Ops.RANGE, name="r")), name="x"), lambda ctx,x,r:
-   f"  br label %loop_latch_{range_str(r)}\nloop_latch_{range_str(r)}:\n"
+   f"  br label %loop_entry_{range_str(r)}\n"
+   f"loop_entry_{range_str(r)}:\n"
+   f"  br label %loop_latch_{range_str(r)}\n"
+   f"loop_latch_{range_str(r)}:\n"
+   f"  {ctx[r]} = phi {ldt(r.dtype)} [ 0, %loop_entry_{range_str(r)} ], [ {ctx[r]}phi, %loop_footer_{range_str(r)} ]\n"
    f"  {ctx[r]}phi = add {ldt(r.dtype)} {ctx[r]}, 1\n"
-   f"  {ctx[x]} = icmp ult {ldt(r.dtype)} {ctx[r]}phi, {ctx[r.src[0]]}\n"
-   f"  br i1 {ctx[x]}, label %loop_body_{range_str(r)}, label %loop_exit_{range_str(r)}\nloop_exit_{range_str(r)}:"),
+   f"  {ctx[r]}cmp = icmp ult {ldt(r.dtype)} {ctx[r]}, {ctx[r.src[0]]}\n"
+   f"  br i1 {ctx[r]}cmp, label %loop_body_{range_str(r)}, label %loop_exit_{range_str(r)}\n"
+   f"loop_body_{range_str(r)}:"),
+  (UPat(Ops.END, src=(UPat(), UPat(Ops.RANGE, name="r"))), lambda r:
+   f"  br label %loop_footer_{range_str(r)}\n"
+   f"loop_footer_{range_str(r)}:\n"
+   f"  br label %loop_latch_{range_str(r)}\n"
+   f"loop_exit_{range_str(r)}:"),
 
   # if
   (UPat(Ops.IF, name="x"), lambda ctx,x: f"  br i1 {ctx[x.src[0]]}, label %ifbody_{ctx[x][1:]}, label %ifskip_{ctx[x][1:]}\nifbody_{ctx[x][1:]}:"),
@@ -181,8 +187,10 @@ class LLVMRenderer(Renderer):
       elif u.op in (Ops.DEFINE_LOCAL, Ops.DEFINE_REG):
         r[u] = f"%{'local' if u.op is Ops.DEFINE_LOCAL else 'reg'}_{str(u.arg).replace('(', '').replace(')', '').replace(',', '_').replace(' ', '')}"
         assert isinstance(u.dtype, PtrDType)
-        if self.device == "CPU" or u.op is Ops.DEFINE_REG:
+        if u.op is Ops.DEFINE_REG:
           kernel.append(f"  {r[u]} = alloca [{u.dtype.size} x {ldt(u.dtype.base)}]")
+        elif self.device == "CPU" and u.op is Ops.DEFINE_LOCAL:
+          kernel.append(f"  {r[u]} = alloca [{u.dtype.size} x {ldt(u.dtype.base)}], align 16")
         else:
           local_args.append(f"@{r[u][1:]} = internal unnamed_addr addrspace(3) global [{u.dtype.size} x {ldt(u.dtype)}] undef, align 16")
           kernel.append(f"  {r[u]} = addrspacecast [{u.dtype.size} x {ldt(u.dtype)}] addrspace(3)* @{r[u][1:]} to [{u.dtype.size} x {ldt(u.dtype)}]*")
@@ -238,7 +246,7 @@ class AMDLLVMRenderer(LLVMRenderer):
   def __init__(self, arch:str):
     self.arch = arch
     self.tensor_cores = AMDRenderer.get_tensor_cores(arch)
-    self.is_cdna = arch.split(":")[0] in {"gfx942", "gfx950"}
+    self.is_cdna = AMDRenderer.is_cdna(arch)
     self.string_rewrite += PatternMatcher([(UPat(Ops.WMMA, name="wmma"), lambda ctx, wmma, cdna=self.is_cdna: render_wmma_amd(ctx, wmma, cdna))])
     if self.is_cdna:
       self.extra_matcher += PatternMatcher([

tinygrad/renderer/nir.py

@@ -3,7 +3,7 @@ from tinygrad.dtype import AddrSpace, DType, PtrDType, dtypes
 from tinygrad.helpers import DEBUG, OSX, unwrap
 from tinygrad.renderer import Renderer
 from tinygrad.renderer.cstyle import CUDARenderer
-from tinygrad.uop.ops import GroupOp, Ops, UOp, PatternMatcher, UPat
+from tinygrad.uop.ops import GroupOp, Ops, UOp, PatternMatcher, UPat, range_str
 import tinygrad.runtime.autogen.mesa as mesa
 import base64, ctypes, ctypes.util, struct, functools, inspect
 
@@ -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]])))
@@ -182,14 +182,17 @@ class NIRRenderer(Renderer):
         self.r[u] = nimm(self.b, self.b.shader.contents.info.shared_size, dtypes.long)
         self.b.shader.contents.info.shared_size += u.dtype.nbytes()
       elif u.op == Ops.RANGE:
-        ranges.append(i:=deref_var(self.b, mesa.nir_local_variable_create(self.b.impl, glsl_type(u.dtype), f"idx{u.arg[0]}".encode()).contents))
+        ranges.append(i:=deref_var(self.b, mesa.nir_local_variable_create(self.b.impl, glsl_type(u.dtype), f"idx{range_str(u)}".encode()).contents))
         nstore(self.b, AddrSpace.REG, i, nimm(self.b, 0, u.dtype), u.dtype)
         mesa.nir_push_loop(self.b)
         self.r[u] = nload(self.b, AddrSpace.REG, i, u.dtype)
+        nif(self.b, nalu(self.b, "ilt", self.r[u], self.r[u.src[0]]), lambda: None, lambda: njump(self.b, mesa.nir_jump_break))
       elif u.op == Ops.END:
         r = u.src[1]
-        nif(self.b, nalu(self.b, "ilt", x:=nalu(self.b, "iadd", self.r[r], nimm(self.b, 1, r.dtype)), self.r[r.src[0]]),
-            functools.partial(nstore, self.b, AddrSpace.REG, ranges.pop(), x, r.dtype), lambda: njump(self.b, mesa.nir_jump_break))
+        next_i = nalu(self.b, "iadd", self.r[r], nimm(self.b, 1, r.dtype))
+        # TODO: this nif should be removable ... but TestMultiTensor.test_double_matmul_shard_W_0 segfaults with it gone
+        nif(self.b, nalu(self.b, "ilt", next_i, self.r[r.src[0]]), lambda: None, lambda: njump(self.b, mesa.nir_jump_break))
+        nstore(self.b, AddrSpace.REG, ranges.pop(), next_i, r.dtype),
         mesa.nir_pop_loop(self.b, None)
       else:
         if (d:=self.def_rewrite.rewrite(u, ctx=self)) is None: raise RuntimeError(f"failed to render {u.op} srcs {[x.dtype for x in u.src]}")

tinygrad/renderer/ptx.py

@@ -6,7 +6,7 @@ from tinygrad.uop.ops import Ops, UOp, PatternMatcher, UPat, GroupOp
 from tinygrad.dtype import dtypes, DType, PtrDType, AddrSpace
 from tinygrad.renderer import Renderer
 from tinygrad.renderer.cstyle import CUDARenderer
-from tinygrad.helpers import flatten, get_single_element, prod
+from tinygrad.helpers import flatten, get_single_element, prod, unwrap
 
 def render_val(x, dtype):
   if dtypes.is_float(dtype):
@@ -119,8 +119,12 @@ string_rewrite = PatternMatcher([
      if x.dtype.count > 1 else f"ld.{mem_type(buf)}.{ctx.mem_types[x.dtype]} {ctx.r[x]}, [{ctx.r[loc]}+0];"),
   # simple
   (UPat(Ops.DEFINE_REG, src=()), lambda ctx: []),
-  (UPat(Ops.RANGE, name="r"), lambda ctx, r: [f"mov.u32 {ctx.r[r]}, 0;", "LOOP_" + f"{ctx.r[r][1:]}:"]),
+  (UPat(Ops.RANGE, name="r"), lambda ctx, r: [
+    f"mov.u32 {ctx.r[r]}, -1;",
+    f"bra END_{ctx.r[r][1:]};",
+    "LOOP_" + f"{ctx.r[r][1:]}:"]),
   (UPat(Ops.END, name="x", src=(UPat(), UPat(Ops.RANGE, name="r"))), lambda ctx, x, r: [
+    "END_" + f"{ctx.r[r][1:]}:",
     ctx.code_for_op[Ops.ADD](ctx.r[r], ctx.r[r], "1", dtypes.int, ctx.types[dtypes.int]),
     ctx.code_for_op[Ops.CMPLT](ctx.r[x], ctx.r[r], ctx.r[r.src[0]], dtypes.int, ctx.types[dtypes.int]),
     f"@{ctx.r[x]} bra LOOP_{ctx.r[r][1:]};"]),
@@ -177,7 +181,7 @@ class PTXRenderer(Renderer):
 
     def ssa(prefix:str, u:UOp|None=None, dtype:str|None=None) -> str:
       nonlocal c, r
-      prefix += f"_{dtype if dtype is not None else self.types[cast(UOp, u).dtype.base]}_"
+      prefix += f"_{dtype if dtype is not None else self.types[unwrap(u).dtype.base]}_"
       c[prefix] += 1
       return f"%{prefix}{c[prefix]-1}"
 

tinygrad/runtime/autogen/am/am.py

@@ -10,6 +10,8 @@ import ctypes
 
 
 class AsDictMixin:
+    import sys
+    if sys.version_info >= (3, 14): _layout_ = 'ms'
     @classmethod
     def as_dict(cls, self):
         result = {}

tinygrad/runtime/autogen/am/pm4_nv.py

@@ -10,6 +10,8 @@ import ctypes
 
 
 class AsDictMixin:
+    import sys
+    if sys.version_info >= (3, 14): _layout_ = 'ms'
     @classmethod
     def as_dict(cls, self):
         result = {}

tinygrad/runtime/autogen/am/pm4_soc15.py

@@ -10,6 +10,8 @@ import ctypes
 
 
 class AsDictMixin:
+    import sys
+    if sys.version_info >= (3, 14): _layout_ = 'ms'
     @classmethod
     def as_dict(cls, self):
         result = {}

tinygrad/runtime/autogen/am/sdma_4_0_0.py

@@ -10,6 +10,8 @@ import ctypes
 
 
 class AsDictMixin:
+    import sys
+    if sys.version_info >= (3, 14): _layout_ = 'ms'
     @classmethod
     def as_dict(cls, self):
         result = {}

tinygrad/runtime/autogen/am/sdma_4_4_2.py

@@ -10,6 +10,8 @@ import ctypes
 
 
 class AsDictMixin:
+    import sys
+    if sys.version_info >= (3, 14): _layout_ = 'ms'
     @classmethod
     def as_dict(cls, self):
         result = {}

tinygrad/runtime/autogen/am/sdma_5_0_0.py

@@ -10,6 +10,8 @@ import ctypes
 
 
 class AsDictMixin:
+    import sys
+    if sys.version_info >= (3, 14): _layout_ = 'ms'
     @classmethod
     def as_dict(cls, self):
         result = {}

tinygrad/runtime/autogen/am/sdma_6_0_0.py

@@ -10,6 +10,8 @@ import ctypes
 
 
 class AsDictMixin:
+    import sys
+    if sys.version_info >= (3, 14): _layout_ = 'ms'
     @classmethod
     def as_dict(cls, self):
         result = {}

tinygrad/runtime/autogen/am/smu_v13_0_0.py

@@ -10,6 +10,8 @@ import ctypes
 
 
 class AsDictMixin:
+    import sys
+    if sys.version_info >= (3, 14): _layout_ = 'ms'
     @classmethod
     def as_dict(cls, self):
         result = {}

tinygrad/runtime/autogen/am/smu_v14_0_2.py

@@ -10,6 +10,8 @@ import ctypes
 
 
 class AsDictMixin:
+    import sys
+    if sys.version_info >= (3, 14): _layout_ = 'ms'
     @classmethod
     def as_dict(cls, self):
         result = {}

tinygrad/runtime/autogen/amd_gpu.py

@@ -10,6 +10,8 @@ import ctypes, os
 
 
 class AsDictMixin:
+    import sys
+    if sys.version_info >= (3, 14): _layout_ = 'ms'
     @classmethod
     def as_dict(cls, self):
         result = {}

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
@@ -52,6 +54,8 @@ else:
     c_long_double_t = ctypes.c_ubyte*16
 
 class AsDictMixin:
+    import sys
+    if sys.version_info >= (3, 14): _layout_ = 'ms'
     @classmethod
     def as_dict(cls, self):
         result = {}