remove on_stack from graph_rewrite

* remove restrictions on range ending in indexing

* early simplify

* Revert "early simplify"

This reverts commit 657d9972c2.

* disable const folding tests

.github/workflows/benchmark.yml

@@ -52,12 +52,12 @@ jobs:
     - name: reset process replay
       run: python3.11 test/external/process_replay/reset.py
     - name: Run Stable Diffusion
-      run: BENCHMARK_LOG=stable_diffusion JIT=1 ASSERT_MIN_STEP_TIME=1000 python3.11 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
+      run: BENCHMARK_LOG=stable_diffusion JIT=1 ASSERT_MIN_STEP_TIME=800 python3.11 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
     - name: Run Stable Diffusion without fp16
-      run: BENCHMARK_LOG=stable_diffusion_fp32 JIT=1 ASSERT_MIN_STEP_TIME=1000 python3.11 examples/stable_diffusion.py --seed 0 --noshow --timing | tee sd_no_fp16.txt
+      run: BENCHMARK_LOG=stable_diffusion_fp32 JIT=1 ASSERT_MIN_STEP_TIME=900 python3.11 examples/stable_diffusion.py --seed 0 --noshow --timing | tee sd_no_fp16.txt
     - name: Run Stable Diffusion v2
       # TODO: very slow step time
-      run: BENCHMARK_LOG=stable_diffusion_v2 JIT=1 ASSERT_MIN_STEP_TIME=100000 python3.11 examples/sdv2.py --fp16 --seed 0 --noshow --timing | tee sdv2.txt
+      run: BENCHMARK_LOG=stable_diffusion_v2 JIT=1 ASSERT_MIN_STEP_TIME=10000 python3.11 examples/sdv2.py --fp16 --seed 0 --noshow --timing | tee sdv2.txt
     # process replay can't capture this, the graph is too large
     # TODO: too slow
     # - name: Run SDXL
@@ -101,7 +101,7 @@ jobs:
     - name: Run GPT2
       run: |
         BENCHMARK_LOG=gpt2_nojit JIT=0 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
-        BENCHMARK_LOG=gpt2 JIT=1 ASSERT_MIN_STEP_TIME=16 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
+        BENCHMARK_LOG=gpt2 JIT=1 ASSERT_MIN_STEP_TIME=13 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
     - name: Run GPT2 w HALF
       run: BENCHMARK_LOG=gpt2_half HALF=1 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
     - name: Run GPT2 w HALF/BEAM
@@ -246,9 +246,9 @@ jobs:
     - name: Run GPT2
       run: |
         BENCHMARK_LOG=gpt2_nojit NV=1 JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
-        BENCHMARK_LOG=gpt2 NV=1 JIT=1 ASSERT_MIN_STEP_TIME=10 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
+        BENCHMARK_LOG=gpt2 NV=1 JIT=1 ASSERT_MIN_STEP_TIME=4 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
     - name: Run GPT2 w HALF
-      run: BENCHMARK_LOG=gpt2_half NV=1 HALF=1 ASSERT_MIN_STEP_TIME=10 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
+      run: BENCHMARK_LOG=gpt2_half NV=1 HALF=1 ASSERT_MIN_STEP_TIME=6 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
     - name: Run GPT2 w HALF/BEAM
       run: BENCHMARK_LOG=gpt2_half_beam NV=1 HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
     - uses: actions/upload-artifact@v4
@@ -316,11 +316,11 @@ jobs:
     - name: Train MNIST
       run: time PYTHONPATH=. NV=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py | tee beautiful_mnist.txt
     - name: Run 10 CIFAR training steps
-      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=850 NV=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
+      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=270 NV=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=680 NV=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=310 NV=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=750 NV=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
+      run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=310 NV=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
     # TODO: too slow
     # - name: Run 10 CIFAR training steps w winograd
     #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=350 NV=1 CAPTURE_PROCESS_REPLAY=0 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_wino.txt
@@ -426,7 +426,7 @@ jobs:
     - name: Test AM warm start time
       run: time AMD=1 python3 test/test_tiny.py TestTiny.test_plus
     - name: Run Stable Diffusion
-      run: BENCHMARK_LOG=stable_diffusion ASSERT_MIN_STEP_TIME=900 AMD=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
+      run: BENCHMARK_LOG=stable_diffusion ASSERT_MIN_STEP_TIME=550 AMD=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
     # TODO: too slow
     # - name: Run SDXL
     #   run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=3200 CAPTURE_PROCESS_REPLAY=0 AMD=1 python3 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
@@ -520,9 +520,9 @@ jobs:
     - name: Train MNIST
       run: time PYTHONPATH=. AMD=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py | tee beautiful_mnist.txt
     - name: Run 10 CIFAR training steps
-      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=400 AMD=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
+      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=500 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=330 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
@@ -625,15 +625,15 @@ jobs:
     - name: benchmark openpilot 0.9.9 dmonitoring
       run: BENCHMARK_LOG=openpilot_0_9_9_dmonitoring PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/dmonitoring_model.onnx
     - name: openpilot compile3 0.9.9 driving_vision
-      run: PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_vision.onnx
+      run: PYTHONPATH="." ASSERT_MIN_STEP_TIME=22 QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_vision.onnx
     - name: openpilot compile3 0.9.9 driving_policy
-      run: PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_policy.onnx
+      run: PYTHONPATH="." ASSERT_MIN_STEP_TIME=7 QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_policy.onnx
     - name: openpilot compile3 0.9.9 dmonitoring
-      run: PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/dmonitoring_model.onnx
+      run: PYTHONPATH="." ASSERT_MIN_STEP_TIME=15 QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/dmonitoring_model.onnx
     - name: openpilot compile3 Space Lab policy + vision
       run: |
-        PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/22aec22a10ce09384d4a4af2a0bbff08d54af7e0c888503508f356fae4ff0e29
-        PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/c824f68646a3b94f117f01c70dc8316fb466e05fbd42ccdba440b8a8dc86914b
+        PYTHONPATH="." ASSERT_MIN_STEP_TIME=4 QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/22aec22a10ce09384d4a4af2a0bbff08d54af7e0c888503508f356fae4ff0e29
+        PYTHONPATH="." ASSERT_MIN_STEP_TIME=26 QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/c824f68646a3b94f117f01c70dc8316fb466e05fbd42ccdba440b8a8dc86914b
     - name: benchmark MobileNetV2 on DSP
       run: |
         # generate quantized weights

examples/mlperf/model_train.py

@@ -3,7 +3,7 @@ from pathlib import Path
 import multiprocessing
 
 from tinygrad import Device, GlobalCounters, Tensor, TinyJit, dtypes
-from tinygrad.helpers import getenv, BEAM, WINO, round_up, diskcache_clear, FUSE_CONV_BW, Profiling
+from tinygrad.helpers import getenv, BEAM, WINO, round_up, diskcache_clear, Profiling
 from tinygrad.nn.state import get_parameters, get_state_dict, load_state_dict, safe_load, safe_save
 from tinygrad.nn.optim import LAMB, LARS, SGD, OptimizerGroup, Adam, AdamW
 
@@ -707,7 +707,7 @@ def train_unet3d():
   ```BASEDIR=<folder_path> ./examples/mlperf/scripts/setup_kits19_dataset.sh```
 
   2) To start training the model, run the following:
-  ```time PYTHONPATH=. WANDB=1 TRAIN_BEAM=3 FUSE_CONV_BW=1 GPUS=6 BS=6 MODEL=unet3d python3 examples/mlperf/model_train.py```
+  ```time PYTHONPATH=. WANDB=1 TRAIN_BEAM=3 GPUS=6 BS=6 MODEL=unet3d python3 examples/mlperf/model_train.py```
   """
   from examples.mlperf.losses import dice_ce_loss
   from examples.mlperf.metrics import dice_score
@@ -749,7 +749,6 @@ def train_unet3d():
     "train_beam": TRAIN_BEAM,
     "eval_beam": EVAL_BEAM,
     "wino": WINO.value,
-    "fuse_conv_bw": FUSE_CONV_BW.value,
     "gpus": GPUS,
     "default_float": dtypes.default_float.name
   }

examples/openpilot/compile3.py

@@ -77,13 +77,20 @@ def test_vs_compile(run, new_inputs, test_val=None):
             **{k:Tensor(v, device="NPY").realize() for k,v in new_inputs_numpy.items() if 'img' not in k}}
 
   # run 20 times
+  step_times = []
   for _ in range(20):
     st = time.perf_counter()
     out = run(**inputs)
     mt = time.perf_counter()
     val = out.numpy()
     et = time.perf_counter()
-    print(f"enqueue {(mt-st)*1e3:6.2f} ms -- total run {(et-st)*1e3:6.2f} ms")
+    step_times.append((et-st)*1e3)
+    print(f"enqueue {(mt-st)*1e3:6.2f} ms -- total run {step_times[-1]:6.2f} ms")
+
+  if (assert_time:=getenv("ASSERT_MIN_STEP_TIME")):
+    min_time = min(step_times)
+    assert min_time < assert_time, f"Speed regression, expected min step time of < {assert_time} ms but took: {min_time} ms"
+
   print(out, val.shape, val.dtype)
   if test_val is not None: np.testing.assert_equal(test_val, val)
   print("**** test done ****")

examples/stable_diffusion.py

@@ -269,12 +269,14 @@ if __name__ == "__main__":
 
   # load in weights
   with WallTimeEvent(BenchEvent.LOAD_WEIGHTS):
-    load_state_dict(model, torch_load(fetch('https://huggingface.co/CompVis/stable-diffusion-v-1-4-original/resolve/main/sd-v1-4.ckpt', 'sd-v1-4.ckpt'))['state_dict'], strict=False)
+    load_state_dict(model, torch_load(fetch('https://huggingface.co/CompVis/stable-diffusion-v-1-4-original/resolve/main/sd-v1-4.ckpt', 'sd-v1-4.ckpt'))['state_dict'], verbose=False, strict=False, realize=False)
 
     if args.fp16:
       for k,v in get_state_dict(model).items():
         if k.startswith("model"):
-          v.replace(v.cast(dtypes.float16).realize())
+          v.replace(v.cast(dtypes.float16))
+
+    Tensor.realize(*get_state_dict(model).values())
 
   # run through CLIP to get context
   tokenizer = Tokenizer.ClipTokenizer()

examples/train_resnet.py

@@ -32,7 +32,7 @@ if __name__ == "__main__":
 
   lr = 5e-3
   transform = ComposeTransforms([
-    lambda x: [Image.fromarray(xx, mode='L').resize((64, 64)) for xx in x],
+    lambda x: [Image.fromarray(xx).resize((64, 64)) for xx in x],
     lambda x: np.stack([np.asarray(xx) for xx in x], 0),
     lambda x: x / 255.0,
     lambda x: np.tile(np.expand_dims(x, 1), (1, 3, 1, 1)).astype(np.float32),

extra/test_hcopt.py

@@ -1,40 +0,0 @@
-import time
-from extra.optimization.helpers import load_worlds, ast_str_to_ast
-from tinygrad import Device
-from tinygrad.codegen.lowerer import pm_lowerer, get_index
-from tinygrad.uop.ops import graph_rewrite
-from tinygrad.codegen.opt.kernel import Kernel
-from tinygrad.codegen.opt.postrange import Scheduler
-from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
-from tinygrad.helpers import getenv
-
-if __name__ == "__main__":
-  renderer = Device.default.renderer
-  ast_strs = load_worlds()
-  if (n:=getenv("N", -1)) != -1: ast_strs = ast_strs[n:n+1]
-  good = 0
-  for i, ast_str in enumerate(ast_strs):
-    ast = ast_str_to_ast(ast_str)
-
-    st = time.perf_counter()
-    lin = Kernel(ast, renderer)
-    opt1 = hand_coded_optimizations(lin)
-    et_lin = time.perf_counter() - st
-
-    lowered = graph_rewrite(ast, pm_lowerer, ctx=get_index(ast), bottom_up=True)
-    st = time.perf_counter()
-    sch = Scheduler(lowered, renderer)
-    sch.convert_loop_to_global()
-    sch.simplify_merge_adjacent()
-    opt2 = hand_coded_optimizations(sch)
-    et_sch = time.perf_counter() - st
-
-    if opt1 != opt2:
-      print(f"******* {i:6d}")
-      print("Kernel:    ", lin.colored_shape(), "->", lin.apply_opts(opt1).colored_shape())
-      print("Scheduler: ", sch.colored_shape(), "->", sch.apply_opts(opt2).colored_shape())
-      print(opt1)
-      print(opt2)
-    else:
-      good += 1
-      print(f"******* {i:6d} MATCH {good/(i+1)*100:.2f}% -- {et_lin/et_sch:4.2f}x speedup")

test/external/external_benchmark_openpilot.py

@@ -39,10 +39,6 @@ if __name__ == "__main__":
       step_times.append(t:=(time.perf_counter_ns() - st)*1e-6)
     print(f"jitted:  {t:7.4f} ms")
 
-  if (assert_time:=getenv("ASSERT_MIN_STEP_TIME")):
-    min_time = min(step_times)
-    assert min_time < assert_time, f"Speed regression, expected min step time of < {assert_time} ms but took: {min_time} ms"
-
   suffix = ""
   if IMAGE.value < 2: suffix += f"_image{IMAGE.value}" # image=2 has no suffix for compatibility
   if getenv("FLOAT16") == 1: suffix += "_float16"

test/external/external_test_opt.py

@@ -4,7 +4,7 @@ import numpy as np
 import torch
 
 from tinygrad import GlobalCounters, Tensor, Device
-from tinygrad.helpers import getenv, Context, RANGEIFY
+from tinygrad.helpers import getenv, RANGEIFY
 from tinygrad.nn.state import get_parameters
 from tinygrad.engine.realize import capturing
 from tinygrad.tensor import _to_np_dtype
@@ -217,24 +217,22 @@ class TestOpt(unittest.TestCase):
     assert cache_len == 1, "reduceop was rerun!"
 
   def test_expand_reduce_is_folded_on_same_axis(self):
-    with Context(FUSE_CONV_BW=1):
-      for axis in [0, 1]:
-        for n in [4, 8, 16]:
-          b = torch.ones(n, n).sum(axis).reshape(n, 1).expand(n, n).sum(axis)
-          with CLCache(allowed=3 if RANGEIFY else 2):
-            a = Tensor.ones(n, n).contiguous().sum(axis).reshape(n, 1).expand(n, n).sum(axis)
-            a.realize()
-          np.testing.assert_allclose(a.numpy(), b.numpy(), rtol=1e-3, atol=1e-5)
-
-  def test_expand_reduce_is_folded_on_different_axes(self):
-    with Context(FUSE_CONV_BW=1):
-      axis1, axis2 = 0, 1
+    for axis in [0, 1]:
       for n in [4, 8, 16]:
-        b = torch.ones(n, n).sum(axis1).reshape(n, 1).expand(n, n).sum(axis2)
+        b = torch.ones(n, n).sum(axis).reshape(n, 1).expand(n, n).sum(axis)
         with CLCache(allowed=3 if RANGEIFY else 2):
-          a = Tensor.ones(n, n).contiguous().sum(axis1).reshape(n, 1).expand(n, n).sum(axis2)
+          a = Tensor.ones(n, n).contiguous().sum(axis).reshape(n, 1).expand(n, n).sum(axis)
           a.realize()
         np.testing.assert_allclose(a.numpy(), b.numpy(), rtol=1e-3, atol=1e-5)
 
+  def test_expand_reduce_is_folded_on_different_axes(self):
+    axis1, axis2 = 0, 1
+    for n in [4, 8, 16]:
+      b = torch.ones(n, n).sum(axis1).reshape(n, 1).expand(n, n).sum(axis2)
+      with CLCache(allowed=3 if RANGEIFY else 2):
+        a = Tensor.ones(n, n).contiguous().sum(axis1).reshape(n, 1).expand(n, n).sum(axis2)
+        a.realize()
+      np.testing.assert_allclose(a.numpy(), b.numpy(), rtol=1e-3, atol=1e-5)
+
 if __name__ == '__main__':
   unittest.main()

test/mockgpu/nv/nvgpu.py

@@ -1,4 +1,4 @@
-import ctypes, ctypes.util, time
+import ctypes, time
 import tinygrad.runtime.autogen.nv_gpu as nv_gpu
 from enum import Enum, auto
 from test.mockgpu.gpu import VirtGPU

test/test_const_folding.py

@@ -69,9 +69,12 @@ class TestBinaryOpsConstFolding(unittest.TestCase):
   def test_tensor_one_mul(self):
     _check_ast_count(0, Tensor.ones(4) * Tensor([1.0, 2, 3, 4]))
 
+  # TODO: these will be fixed with better folding
+  @unittest.expectedFailure
   def test_bool_tensor_mul_bool(self):
     _check_ast_count(0, Tensor([True, False]) * True)
     _check_ast_count(0, Tensor([True, False]) * False)
+  @unittest.expectedFailure
   def test_bool_mul_bool_tensor(self):
     _check_ast_count(0, True * Tensor([True, False]))
     _check_ast_count(0, False * Tensor([True, False]))

test/test_linearizer.py

@@ -432,7 +432,7 @@ def helper_realized_ast(r:Tensor|list[Tensor]) -> tuple[UOp, list[Buffer]]:
 def helper_linearizer_ast(ast:UOp, inputs:list[Tensor], *args, **kwargs):
   assert isinstance(ast, UOp), "ast must be UOp"
   inbufs = [x.uop.base.buffer for x in inputs]
-  outbufs = [Buffer(inbufs[-1].device if inbufs else Device.DEFAULT, out.st_arg.size, out.src[1].dtype).allocate() for out in ast.src]
+  outbufs = [Buffer(inbufs[-1].device if inbufs else Device.DEFAULT, out.size, out.src[1].dtype).allocate() for out in ast.src]
   _helper_linearizer_opt_ast(ast, outbufs+inbufs, *args, **kwargs)
 
 def helper_linearizer_opt(r:Tensor|list[Tensor], *args, **kwargs):

test/test_rangeify.py

@@ -300,11 +300,12 @@ class TestOuterworld(unittest.TestCase):
     o.contiguous(i).realize()
     self.assertTrue((t==o).all().item())
 
-from tinygrad.schedule.rangeify import pm_rangeify, RangeifyContext
+@unittest.skip("pm_rangeify no longer exists. test this in a different way")
 class TestRangeifyPM(unittest.TestCase):
   def setUp(self): self.base = Tensor.empty(10*10).reshape(10, 10).contiguous()
   def assert_same(self, a, b):
     def run_pm_rangeify(t:Tensor):
+      from tinygrad.schedule.rangeify import pm_rangeify, RangeifyContext
       sink = t.uop.sink()
       pm_realize = PatternMatcher([(UPat(Ops.CONTIGUOUS, name="x"), lambda x: x.replace(op=Ops.REALIZE))])
       sink = graph_rewrite(sink, pm_realize)

test/test_schedule.py

@@ -45,7 +45,7 @@ def check_schedule(t:Tensor|list[Tensor]|UOp, allowed:int, to_prerealize:list[Te
 def _realize_weights(m):
   for p in nn.state.get_parameters(m): p.realize()
 
-def _test_conv2d(allowed:int, dtype:DType=dtypes.float, **kwargs):
+def _test_conv2d(allowed:int, dtype:DType=dtypes.float):
   old_default_float, dtypes.default_float = dtypes.default_float, dtype
   dtypes.default_float = dtype
   Tensor.manual_seed(0)
@@ -54,7 +54,7 @@ def _test_conv2d(allowed:int, dtype:DType=dtypes.float, **kwargs):
   w = Tensor.uniform(16, CIN, 3, 3, requires_grad=True).realize()
   ret = Tensor.conv2d(img, w).relu().mean().backward()
   dtypes.default_float = old_default_float
-  with Context(**kwargs): s = Tensor.schedule(ret, img.grad, w.grad)
+  s = Tensor.schedule(ret, img.grad, w.grad)
   run_schedule(s.copy())
   cnt = len([si for si in s if si.ast.op is Ops.SINK])
   assert cnt == allowed, f"expected {allowed} kernels, got {cnt}"
@@ -470,15 +470,14 @@ class TestSchedule(unittest.TestCase):
           check_schedule(opt.schedule_step(), cnt)
 
   def test_fold_batchnorm_backward(self):
-    with Context(FUSE_CONV_BW=1):
-      with Tensor.train():
-        x = Tensor.empty((2, 16, 8, 8)).contiguous()
-        bn = nn.BatchNorm2d(16)
-        bn.weight.requires_grad = bn.bias.requires_grad = x.requires_grad = True
-        fw = bn(x).contiguous_backward().relu().contiguous()
-        fw.sum().backward()
-        # TODO: this is too many
-        check_schedule([x.grad, bn.weight.grad, bn.bias.grad, fw], 10)
+    with Tensor.train():
+      x = Tensor.empty((2, 16, 8, 8)).contiguous()
+      bn = nn.BatchNorm2d(16)
+      bn.weight.requires_grad = bn.bias.requires_grad = x.requires_grad = True
+      fw = bn(x).contiguous_backward().relu().contiguous()
+      fw.sum().backward()
+      # TODO: this is too many
+      check_schedule([x.grad, bn.weight.grad, bn.bias.grad, fw], 10)
 
   def test_fold_conv_relu(self):
     c1 = nn.Conv2d(3,16,3)
@@ -1321,7 +1320,7 @@ class TestSchedule(unittest.TestCase):
       opt = nn.optim.SGD(nn.state.get_parameters([c1, c2, c3, c4]))
       opt.zero_grad()
       c4(c3(c2(c1(img).relu()).relu()).relu()).relu().sum().backward()
-      with Context(FUSE_CONV_BW=1): check_schedule(opt.schedule_step(), 14)
+      check_schedule(opt.schedule_step(), 14)
 
   @unittest.skipUnless(is_dtype_supported(dtypes.half), "need half")
   @expect_rangeify_fails
@@ -1625,15 +1624,14 @@ class TestSchedule(unittest.TestCase):
     out = x.argmax(1)
     run_schedule(check_schedule(out, 2))
 
-  def test_conv2d(self): _test_conv2d(5 if RANGEIFY else 7)
-  def test_conv2d_fused(self): _test_conv2d(5 if RANGEIFY else 5, FUSE_CONV_BW=1)
+  def test_conv2d(self): _test_conv2d(5 if SPLIT_REDUCEOP else 4)
+  def test_conv2d_fused(self): _test_conv2d(5 if SPLIT_REDUCEOP else 4)
 
   @unittest.skipUnless(is_dtype_supported(dtypes.half) and is_dtype_supported(dtypes.ulong), "need half and ulong")
-  def test_conv2d_half(self): _test_conv2d(5 if RANGEIFY else 7, dtype=dtypes.half)
+  def test_conv2d_half(self): _test_conv2d(5 if SPLIT_REDUCEOP else 4, dtype=dtypes.half)
   @unittest.skipUnless(is_dtype_supported(dtypes.half), "need half")
   @unittest.skipIf(Device.DEFAULT == "WEBGPU", "Causes other tests to fail")
-  @unittest.skipIf(not RANGEIFY, "passes on RANGEIFY")
-  def test_conv2d_fused_half(self): _test_conv2d(5, dtype=dtypes.half)
+  def test_conv2d_fused_half(self): _test_conv2d(5 if SPLIT_REDUCEOP else 4, dtype=dtypes.half)
 
   def test_schedule_mem_used(self):
     base = GlobalCounters.mem_used
@@ -2291,7 +2289,7 @@ class TestBufferUOp(unittest.TestCase):
 
   def test_buffer_view_not_allowed(self):
     permuted_view = Tensor.empty(1, 2, 3).permute(0, 2, 1)
-    with self.assertRaisesRegex(AssertionError, "VIEW only works here if it's contiguous"):
+    with self.assertRaisesRegex(AssertionError, "can only be RESHAPE"):
       permuted_view.uop.buffer # cannot access Buffer of a non contiguous VIEW
 
   def test_buffer_only_after_realize(self):

test/test_tensor_uop.py

@@ -4,6 +4,7 @@ import unittest
 from tinygrad import Tensor, Device, dtypes
 from tinygrad.engine.realize import run_schedule
 from tinygrad.uop.ops import Ops, UOp, UPat
+from tinygrad.helpers import SPLIT_REDUCEOP
 
 class TestTensorUOp(unittest.TestCase):
   def test_fromcpu_shape_tracker(self):
@@ -94,6 +95,7 @@ class TestTensorUOp(unittest.TestCase):
     self.assertEqual(out.tolist(), Tensor.zeros(4, 8).tolist())
 
 reduce_kernel = UPat(Ops.SINK, src=(UPat(Ops.STORE, allow_any_len=True, src=(UPat(), UPat((Ops.REDUCE_AXIS, Ops.REDUCE))))))
+@unittest.skipUnless(SPLIT_REDUCEOP, "only for SPLIT_REDUCEOP")
 class TestReduceOp(unittest.TestCase):
   def test_no_split_reduce_kernel(self):
     a = Tensor.rand(4, 4).realize()

test/unit/test_helpers.py

@@ -3,7 +3,7 @@ from tinygrad import Variable
 from tinygrad.helpers import Context, ContextVar, argfix, colored, word_wrap, is_numpy_ndarray, CI, mv_address
 from tinygrad.helpers import merge_dicts, strip_parens, prod, round_up, fetch, fully_flatten, from_mv, to_mv, polyN, time_to_str, cdiv, cmod, getbits
 from tinygrad.tensor import Tensor, get_shape
-from tinygrad.shape.view import get_contraction, get_contraction_with_reduce
+from tinygrad.shape.view import get_contraction
 import numpy as np
 
 VARIABLE = ContextVar("VARIABLE", 0)
@@ -219,20 +219,6 @@ class TestMemoryview(unittest.TestCase):
     print(f"from_mv vs mv_address: {fmv_us:8.3f} µs vs {mva_us:8.3f} µs")
 
 class TestGetContraction(unittest.TestCase):
-  def test_contraction_with_reduce(self):
-    r = get_contraction((16, 1, 1, 1), (16, 1, 1))
-    self.assertEqual(r, [[0], [], [1, 2, 3]])
-    r = get_contraction_with_reduce((16, 1, 1, 1), (16, 1, 1), (1,))
-    self.assertEqual(r, [[0], [1, 2], [3]])
-
-    r = get_contraction((16, 1, 1, 1, 1), (16, 1, 1, 1))
-    self.assertEqual(r, [[0], [], [], [1, 2, 3, 4]])
-    r = get_contraction_with_reduce((16, 1, 1, 1, 1), (16, 1, 1, 1), (1,))
-    self.assertEqual(r, [[0], [1, 2], [3], [4]])
-
-    r = get_contraction_with_reduce((2, 512, 1, 1), (2, 1, 512), (1,))
-    self.assertIsNone(r)
-
   def test_contraction(self):
     r = get_contraction((1,2,3,4), (2,3,4))
     self.assertEqual(r, [[0, 1], [2], [3]])

test/unit/test_rewrite_not_ready.py

@@ -1,110 +0,0 @@
-import unittest
-from dataclasses import dataclass, field
-from tinygrad.uop.ops import PatternMatcher, UOp, graph_rewrite, Ops, UPat, GroupOp, RewriteNotReady
-
-# we could insert CHILDREN node
-
-@dataclass
-class ChildrenContext:
-  children: dict[UOp, list[UOp]]|None = None
-
-# this is a generic child labeller
-def extract_children(ctx:ChildrenContext, x:UOp):
-  if ctx.children is not None: return
-  ctx.children = {k:list(v.keys()) for k,v in x.get_consumer_map().items() if len(v) > 1}
-
-def mark_children(ctx:ChildrenContext, x:UOp):
-  new_srcs = [(UOp(Ops.CHILD, s.dtype, src=(s,), arg=(ctx.children[s].index(x), len(ctx.children[s]))) if s in ctx.children else s) for s in x.src]
-  return x.replace(src=tuple(new_srcs))
-
-pm_children = PatternMatcher([
-  (UPat(Ops.SINK, name="x"), extract_children),
-  (UPat(GroupOp.All-{Ops.CHILD}, name="x"), mark_children),
-])
-
-@dataclass
-class TestContext:
-  seen_children: dict[UOp, set[int]] = field(default_factory=dict)
-  ready_children: dict[UOp, set[int]] = field(default_factory=dict)
-  seen_consts:int = 0
-  saved_seen_consts:int = 0
-  exp2_visit_count:int = 0
-
-# this is a generic pattern
-def visit_child(ctx:ChildrenContext, x:UOp):
-  if x.src[0] not in ctx.seen_children:
-    ctx.seen_children[x.src[0]] = set()
-    ctx.ready_children[x.src[0]] = set()
-  ctx.seen_children[x.src[0]].add(x.arg[0])
-  if len(ctx.seen_children[x.src[0]]) != x.arg[1]:
-    print(f"visit CHILD {x.arg} bottom up -- not ready {ctx.seen_children[x.src[0]]}")
-    raise RewriteNotReady
-  print(f"visit CHILD {x.arg} bottom up -- READY {ctx.seen_children[x.src[0]]}")
-  ctx.ready_children[x.src[0]].add(x.arg[0])
-
-pm_child_visitor = PatternMatcher([
-  (UPat(Ops.CHILD, name="x"), visit_child),
-])
-
-# this is for the test
-def see_const(ctx:ChildrenContext, c:UOp): ctx.seen_consts += c.arg
-def see_exp2(ctx:ChildrenContext): ctx.exp2_visit_count += 1
-def save_seen_consts(ctx:ChildrenContext, x:UOp): ctx.saved_seen_consts = ctx.seen_consts
-pm_consts = PatternMatcher([
-  (UPat(Ops.DEFINE_VAR, name="x"), save_seen_consts),
-  (UPat()+UPat.cvar("c"), see_const),
-  (UPat(Ops.EXP2), see_exp2),
-])
-
-class TestChildrenRewrite(unittest.TestCase):
-  def test_not_ready_double_simple(self):
-    global_a = UOp.variable("a", 0, 10).exp2()
-    inter = (global_a+global_a).exp2()
-    global_sink = (inter+inter).sink()
-
-    sink = graph_rewrite(global_sink, pm_children, ctx=ChildrenContext(), bottom_up=True)
-    ctx = TestContext()
-    graph_rewrite(sink, pm_consts, ctx=ctx, bottom_up=True)
-    self.assertEqual(ctx.exp2_visit_count, 2)
-
-  def test_not_ready_double(self):
-    global_a = UOp.variable("a", 0, 10).exp2()
-    inter = ((global_a+1000)+(global_a+100)).exp2()
-    global_sink = ((inter+10)+(inter+1)).sink()
-
-    sink = graph_rewrite(global_sink, pm_children, ctx=ChildrenContext(), bottom_up=True)
-    print("test_not_ready_double")
-    ctx = TestContext()
-    graph_rewrite(sink, pm_child_visitor+pm_consts, ctx=ctx, bottom_up=True)
-    self.assertEqual(ctx.exp2_visit_count, 2)
-    self.assertEqual(ctx.seen_consts, ctx.saved_seen_consts)
-    self.assertEqual(ctx.seen_consts, 1111)
-
-  def test_in_srcs_twice(self):
-    global_a = UOp.variable("a", 0, 10).exp2()
-    global_sink = (global_a+global_a).sink()
-
-    ctx = TestContext()
-    graph_rewrite(global_sink, pm_consts, ctx=ctx, bottom_up=True)
-    self.assertEqual(ctx.exp2_visit_count, 1)
-
-  def test_not_ready(self):
-    global_a = UOp.variable("a", 0, 10).exp2()
-    global_sink = ((global_a+2)+(global_a+3)).sink()
-
-    # without children and not ready, we don't see both adds before the DEFINE_VAR
-    ctx = TestContext()
-    graph_rewrite(global_sink, pm_consts, ctx=ctx, bottom_up=True)
-    self.assertNotEqual(ctx.seen_consts, ctx.saved_seen_consts)
-    self.assertEqual(ctx.exp2_visit_count, 1)
-
-    # with children and not ready we do
-    sink = graph_rewrite(global_sink, pm_children, ctx=ChildrenContext(), bottom_up=True)
-    ctx = TestContext()
-    graph_rewrite(sink, pm_child_visitor+pm_consts, ctx=ctx, bottom_up=True)
-    self.assertEqual(ctx.seen_consts, ctx.saved_seen_consts)
-    self.assertEqual(ctx.exp2_visit_count, 1)
-    self.assertSetEqual(list(ctx.ready_children.values())[0], {0,1})
-
-if __name__ == '__main__':
-  unittest.main()

test/unit/test_shapetracker.py

@@ -175,12 +175,6 @@ class TestRealSimplifies(unittest.TestCase):
       View.create((8, 3, 3, 11, 2, 28), (924, 308, 0, 28, 0, 1), 0, None),
       View.create((8, 1, 6, 10, 28, 3, 2, 1), (5544, 0, 0, 56, 1, 1848, 672, 0), 0, None)))
 
-class TestViewMinify(unittest.TestCase):
-  def test_minifies(self):
-    assert len(View.create((10,10)).minify().shape) == 1
-    assert len(View.create((10,10)).permute((1,0)).minify().shape) == 2
-    assert len(View.create((10,10,10,10)).permute((1,0,2,3)).minify().shape) == 3
-
 class TestIndexExpressions2d(unittest.TestCase):
   def setUp(self):
     shapes = [(30, 5), (15, 10), (15, 1), (5, 10), (5, 1)] # Make sure dim0 is a multiple of 5, one of the tests divides this dimension by 5

test/unit/test_view.py

@@ -10,25 +10,6 @@ class TestView(unittest.TestCase):
     v = View.create(shape=(4,3,2), strides=(1,4,10), mask=((0,4),(0,3),(0,2)))
     self.assertIsNone(v.mask)
 
-  def test_minify_zero_strided_dims(self):
-    target = View.create(shape=(2,2), strides=(30,2), offset=7, mask=None)
-    v = View.create(shape=(2,1,2), strides=(30,0,2), offset=7, mask=None)
-    self.assertEqual(v.minify(), target)
-    v = View.create(shape=(1,2,2), strides=(0,30,2), offset=7, mask=None)
-    self.assertEqual(v.minify(), target)
-    v = View.create(shape=(2,2,1), strides=(30,2,0), offset=7, mask=None)
-    self.assertEqual(v.minify(), target)
-    v = View.create(shape=(2,1,1,2), strides=(30,0,0,2), offset=7, mask=None)
-    self.assertEqual(v.minify(), target)
-    v = View.create(shape=(1,1,2,2), strides=(0,0,30,2), offset=7, mask=None)
-    self.assertEqual(v.minify(), target)
-    v = View.create(shape=(2,2,1,1), strides=(30,2,0,0), offset=7, mask=None)
-    self.assertEqual(v.minify(), target)
-    v = View.create(shape=(1,2,2,1), strides=(0,30,2,0), offset=7, mask=None)
-    self.assertEqual(v.minify(), target)
-    v = View.create(shape=(1,2,1,2), strides=(0,30,0,2), offset=7, mask=None)
-    self.assertEqual(v.minify(), target)
-
   def test_empty_mask_contiguous(self):
     v1 = View.create(shape=(2,2,2), strides=(4,2,1), mask=None)
     v2 = View.create(shape=(2,2,2), strides=(4,2,1), mask=((0,2),(0,2),(0,2)))

tinygrad/codegen/__init__.py

@@ -7,7 +7,6 @@ from tinygrad.uop.spec import type_verify
 from tinygrad.renderer import Renderer
 
 # import all pattern matchers here
-from tinygrad.codegen.lowerer import pm_lowerer, get_index
 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
@@ -51,7 +50,6 @@ def _get_rewrites_for_renderer(opts:Renderer, optimize:bool, linearizer:bool, _Q
 
     # lowerer first
     if _QUANTIZE and opts.device in {"CPU", "DSP"}: ret.append(RewriteStep(pm_quant, name="quantize"))
-    ret.append(RewriteStep(pm_lowerer, get_index, name="lowerer", bottom_up=True))
 
     # split ranges
     if _RANGEIFY:

tinygrad/codegen/lowerer.py

@@ -1,48 +0,0 @@
-# the job of the lowerer is to do indexing
-from dataclasses import dataclass
-from tinygrad.uop.ops import KernelInfo, UOp, Ops, PatternMatcher, UPat, sint_to_uop, AxisType, graph_rewrite
-
-# ***** indexing *****
-
-@dataclass
-class IndexContext:
-  axis_types: tuple[AxisType, ...]
-  idxs: list[UOp]
-  start: int = 0
-
-def shape_to_idx(s, axis_types, start=0):
-  return [UOp.range(sint_to_uop(s), start+i, at) for i, (s, at) in enumerate(zip(s, axis_types))]
-
-def get_index(ast:UOp) -> IndexContext:
-  axis_types = ast.arg.axis_types if isinstance(ast.arg, KernelInfo) else ()
-  #if len(ast.full_shape) != len(axis_types) and ast.st is not None:
-  #  axis_types = tuple([AxisType.REDUCE if resolve(s != fs) else AxisType.LOOP for s,fs in zip(ast.shape, ast.full_shape)])
-  return IndexContext(axis_types, [], 0)
-
-# ***** lowering (given index) *****
-
-def subblock(ctx: IndexContext, full_new_idx: list[UOp], src: UOp):
-  lc = IndexContext(ctx.axis_types, full_new_idx, ctx.start+1000)
-  ctx.start = lc.start
-  return graph_rewrite(src, pm_lowerer, lc, name="subblock", bottom_up=True)
-
-def fixup_wmma(ctx:IndexContext, x:UOp):
-  if x.tag is not None: return None
-  new_idxs = shape_to_idx(x.src[0].shape, ctx.axis_types, ctx.start)
-  full_new_idx = list(ctx.idxs)
-  for a in x.arg[-1]: full_new_idx[a] = new_idxs[a]
-
-  srcs = subblock(ctx, full_new_idx, UOp.sink(*x.src)).src
-
-  # NOTE: this assumes these are expanded. which now shouldn't change anything
-  new_x_arg_m2 = tuple([tuple([(full_new_idx[a].arg[0], sz) for a,sz in v]) for v in x.arg[-2]])
-  new_x_arg_m1 = tuple([full_new_idx[a].arg[0] for a in x.arg[-1]])
-  return x.replace(src=srcs, arg=x.arg[:-2]+(new_x_arg_m2, new_x_arg_m1), tag=1)
-
-pm_lowerer = PatternMatcher([
-  (UPat(Ops.WMMA, name="x"), fixup_wmma),
-
-  # axis fixups for WMMA
-  (UPat((Ops.CONTRACT, Ops.UNROLL), name="x"),
-   lambda ctx,x: x.replace(tag=1, arg=tuple([(ctx.idxs[a].arg[0], sz) for a,sz in x.arg])) if x.tag is None else None),
-])

tinygrad/codegen/opt/postrange.py

@@ -2,7 +2,7 @@ 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, can_pad, GroupOp
+from tinygrad.uop.ops import PatternMatcher, UPat, Ops, UOp, KernelInfo, graph_rewrite, AxisType, ssimplify, GroupOp
 from tinygrad.device import Buffer
 from tinygrad.dtype import AddrSpace, dtypes, ImageDType
 from tinygrad.helpers import colored, BEAM, getenv, DEBUG, to_function_name, NOOPT, argsort, round_up, prod, merge_dicts, get_single_element
@@ -188,7 +188,7 @@ class Scheduler:
       check(rng.arg[-1] is not AxisType.THREAD, "cannot pad thread")
       # ok to pad SUM if all parent ALU ops have f(0) = 0
       if (r:=self.reduceop) is not None and rng.arg[-1] in (AxisType.GROUP_REDUCE, AxisType.REDUCE):
-        check(r.arg[0] is Ops.ADD and can_pad(r, {}), f"cannot pad {r}")
+        check(r.arg[0] is Ops.ADD and not r.op_in_backward_slice_with_self(*GroupOp.UnsafePad), f"cannot pad {r}")
       new_sz = round_up(int(rng.vmax+1), cast(int, opt.arg))
       check(rng.vmax+1 > new_sz//4, "pad adds more than quadruple the work")
       replaced_rng = UOp.range(new_sz, *rng.arg)

tinygrad/dtype.py

@@ -183,7 +183,7 @@ class dtypes:
   uints = (uint8, uint16, uint32, uint64)
   sints = (int8, int16, int32, int64)
   ints = uints + sints
-  all = floats + ints + (bool, index)
+  all = floats + ints + (bool, index) # noqa: A003
 
 if (env_default_float := getenv("DEFAULT_FLOAT", "")):
   dtypes.default_float = getattr(dtypes, env_default_float.lower())

tinygrad/helpers.py

@@ -133,7 +133,6 @@ JIT, JIT_BATCH_SIZE = ContextVar("JIT", 2 if OSX and ARCH_X86 else 1), ContextVa
 WINO, CAPTURING, TRACEMETA = ContextVar("WINO", 0), ContextVar("CAPTURING", 1), ContextVar("TRACEMETA", 1)
 USE_TC, TC_SELECT, TC_OPT, AMX = ContextVar("TC", 1), ContextVar("TC_SELECT", -1), ContextVar("TC_OPT", 0), ContextVar("AMX", 0)
 TRANSCENDENTAL, NOLOCALS = ContextVar("TRANSCENDENTAL", 1), ContextVar("NOLOCALS", 0)
-FUSE_CONV_BW = ContextVar("FUSE_CONV_BW", 0)
 SPLIT_REDUCEOP, NO_MEMORY_PLANNER, RING = ContextVar("SPLIT_REDUCEOP", 1), ContextVar("NO_MEMORY_PLANNER", 0), ContextVar("RING", 1)
 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)

tinygrad/runtime/ops_amd.py

@@ -1,6 +1,6 @@
 from __future__ import annotations
 from typing import cast, ClassVar
-import os, ctypes, ctypes.util, struct, hashlib, functools, importlib, mmap, errno, array, contextlib, sys, weakref, itertools
+import os, ctypes, struct, hashlib, functools, importlib, mmap, errno, array, contextlib, sys, weakref, itertools
 assert sys.platform != 'win32'
 from dataclasses import dataclass
 from tinygrad.runtime.support.hcq import HCQCompiled, HCQAllocator, HCQBuffer, HWQueue, CLikeArgsState, HCQSignal, HCQProgram, FileIOInterface

tinygrad/runtime/ops_cuda.py

@@ -1,5 +1,5 @@
 from __future__ import annotations
-import ctypes, ctypes.util, functools
+import ctypes, functools
 from tinygrad.helpers import DEBUG, getenv, mv_address, init_c_var, init_c_struct_t, suppress_finalizing
 from tinygrad.device import Compiled, BufferSpec, LRUAllocator, CompilerPairT
 from tinygrad.renderer.cstyle import CUDARenderer

tinygrad/runtime/support/compiler_cuda.py

@@ -1,4 +1,4 @@
-import subprocess, hashlib, tempfile, ctypes, ctypes.util, re, pathlib
+import subprocess, hashlib, tempfile, ctypes, re, pathlib
 from typing import Callable
 from tinygrad.helpers import to_char_p_p, colored, init_c_var, getenv
 import tinygrad.runtime.autogen.nvrtc as nvrtc

tinygrad/runtime/support/hcq.py

@@ -3,10 +3,11 @@ from typing import cast, Callable, Type, TypeVar, Generic, Any, Sequence
 import contextlib, decimal, statistics, time, ctypes, array, os, struct, traceback, collections
 try: import fcntl # windows misses that
 except ImportError: fcntl = None #type:ignore[assignment]
-from tinygrad.helpers import PROFILE, getenv, to_mv, round_up, ProfileRangeEvent
+from tinygrad.helpers import PROFILE, getenv, to_mv, ProfileRangeEvent
 from tinygrad.device import BufferSpec, Compiled, LRUAllocator, ProfileDeviceEvent, ProfileProgramEvent, CompilerPairT
 from tinygrad.uop.ops import sym_infer, sint, UOp
 from tinygrad.runtime.autogen import libc
+from tinygrad.runtime.support.memory import BumpAllocator
 
 class MMIOInterface:
   def __init__(self, addr:int, nbytes:int, fmt='B'): self.mv, self.addr, self.nbytes, self.fmt = to_mv(addr, nbytes).cast(fmt), addr, nbytes, fmt
@@ -62,15 +63,6 @@ ProgramType = TypeVar('ProgramType', bound='HCQProgram')
 ArgsStateType = TypeVar('ArgsStateType', bound='HCQArgsState')
 QueueType = TypeVar('QueueType', bound='HWQueue')
 
-class BumpAllocator:
-  def __init__(self, size:int, base:int=0, wrap:bool=True): self.size, self.ptr, self.base, self.wrap = size, 0, base, wrap
-  def alloc(self, size:int, alignment:int=1) -> int:
-    if round_up(self.ptr, alignment) + size > self.size:
-      if not self.wrap: raise RuntimeError("Out of memory")
-      self.ptr = 0
-    self.ptr = (res:=round_up(self.ptr, alignment)) + size
-    return res + self.base
-
 class HWQueue(Generic[SignalType, HCQDeviceType, ProgramType, ArgsStateType]):
   """
   A base class for hardware command queues in the HCQ (Hardware Command Queue) API.

tinygrad/runtime/support/llvm.py

@@ -1,4 +1,4 @@
-import ctypes, ctypes.util, os, sys, subprocess
+import ctypes.util, os, sys, subprocess
 from tinygrad.helpers import DEBUG, OSX, getenv
 
 if sys.platform == 'win32':

tinygrad/runtime/support/memory.py

@@ -2,6 +2,15 @@ import collections, functools, dataclasses
 from typing import Any, ClassVar
 from tinygrad.helpers import round_up, getenv
 
+class BumpAllocator:
+  def __init__(self, size:int, base:int=0, wrap:bool=True): self.size, self.ptr, self.base, self.wrap = size, 0, base, wrap
+  def alloc(self, size:int, alignment:int=1) -> int:
+    if round_up(self.ptr, alignment) + size > self.size:
+      if not self.wrap: raise RuntimeError("Out of memory")
+      self.ptr = 0
+    self.ptr = (res:=round_up(self.ptr, alignment)) + size
+    return res + self.base
+
 class TLSFAllocator:
   """
   The allocator is based on the Two-Level Segregated Fit (TLSF) algorithm. The allocator maintains 2 level of buckets:

tinygrad/runtime/support/webgpu.py

@@ -1,4 +1,4 @@
-import ctypes, ctypes.util, os, subprocess, platform, sysconfig
+import ctypes.util, os, subprocess, platform, sysconfig
 from tinygrad.helpers import OSX
 
 WEBGPU_PATH: str | None

tinygrad/schedule/indexing.py

@@ -1,11 +1,36 @@
-from typing import Iterator
+from typing import Iterator, Sequence
 import functools, operator, itertools
 from dataclasses import dataclass, field
 from tinygrad.dtype import dtypes, AddrSpace
-from tinygrad.uop.ops import PatternMatcher, UPat, Ops, UOp, resolve, GroupOp
-from tinygrad.uop.symbolic import sym
-from tinygrad.helpers import argsort, all_same, Context
-from tinygrad.uop.ops import graph_rewrite, sint, AxisType
+from tinygrad.uop.ops import PatternMatcher, UPat, Ops, UOp, resolve, GroupOp, graph_rewrite, sint, AxisType
+from tinygrad.uop.symbolic import sym, symbolic
+from tinygrad.helpers import argsort, all_same, cpu_profile, TracingKey
+
+ALWAYS_CONTIGUOUS: set[Ops] = {Ops.CONTIGUOUS, Ops.ASSIGN, Ops.COPY, Ops.BUFFER, Ops.BUFFER_VIEW,
+                     Ops.CONST, Ops.BIND, Ops.DEVICE, Ops.MSELECT, Ops.MSTACK, Ops.DEFINE_GLOBAL,
+                     Ops.DEFINE_LOCAL, Ops.DEFINE_REG, Ops.LOAD, Ops.KERNEL}
+
+def realize(ctx:dict[UOp, None], tr:UOp) -> None: ctx[tr] = None
+
+def realize_srcs(ctx:dict[UOp, None], rb:UOp) -> None:
+  for s in rb.src:
+    if s.base.op not in ALWAYS_CONTIGUOUS: ctx[s] = None
+
+def realize_assign(ctx:dict[UOp, None], a:UOp) -> None:
+  if a.src[1].op not in ALWAYS_CONTIGUOUS: ctx[a.src[1]] = None
+  # if it's a kernel, we don't realize it
+  if a.src[1].op is not Ops.KERNEL: ctx[a] = None
+
+pm_generate_realize_map = PatternMatcher([
+  # always realize SINK src
+  (UPat(Ops.SINK, name="s"), lambda ctx,s: ctx.update((x.base, None) for x in s.src if x.base.op not in ALWAYS_CONTIGUOUS)),
+  # always realize COPY/BUFFER_VIEW/CONTIGUOUS
+  (UPat({Ops.COPY, Ops.BUFFER_VIEW, Ops.CONTIGUOUS}, name="tr"), realize),
+  # realize srcs of COPY, MSELECT, MSTACK
+  (UPat((Ops.COPY, Ops.MSELECT, Ops.MSTACK), name="rb"), realize_srcs),
+  # realize ASSIGN and input to assign (might be optimized out)
+  (UPat(Ops.ASSIGN, name="a"), realize_assign),
+])
 
 @dataclass(frozen=True)
 class BufferizeOpts:
@@ -17,7 +42,6 @@ class BufferizeOpts:
 class IndexingContext:
   realize_map: dict[UOp, None] = field(default_factory=dict)
   range_map: dict[UOp, tuple[list[UOp], list[UOp]]] = field(default_factory=dict)
-  pads_gate: dict[UOp, UOp] = field(default_factory=dict)
 
   # create ranges
   range_idx: Iterator[int] = field(default_factory=itertools.count)
@@ -30,10 +54,10 @@ def create_bufferize_and_index_based_on_ranges(ctx:IndexingContext, x:UOp):
   new_srcs = []
   for s in x.src:
     new_src = s
-    if s.op in {Ops.BUFFER, Ops.MSTACK, Ops.MSELECT} or (s.op is Ops.ASSIGN and s.src[1].op is Ops.KERNEL):
+    if s.op in {Ops.BUFFER, Ops.BUFFER_VIEW, Ops.MSTACK, Ops.MSELECT} or (s.op is Ops.ASSIGN and s.src[1].op is Ops.KERNEL):
       if x in ctx.range_map: new_src = new_src.index(*ctx.range_map[x][0])
     elif s in ctx.realize_map:
-      new_src = UOp(Ops.BUFFERIZE, s.dtype, src=(s,)+tuple(ctx.range_map[s][1]), arg=BufferizeOpts(device=s.device), tag=s.tag)
+      new_src = UOp(Ops.BUFFERIZE, s.dtype, src=(new_src,)+tuple(ctx.range_map[s][1]), arg=BufferizeOpts(device=s.device), tag=s.tag)
       if x in ctx.range_map: new_src = new_src.index(*ctx.range_map[x][0])
     new_srcs.append(new_src)
   # NOTE: do we need this?
@@ -41,7 +65,8 @@ def create_bufferize_and_index_based_on_ranges(ctx:IndexingContext, x:UOp):
 
 def convert_pad_to_where_to_keep_behavior_local(ctx:IndexingContext, x:UOp):
   if x not in ctx.range_map: return None
-  ret = ctx.pads_gate[x].where(x.src[0], UOp.const(x.dtype, 0))
+  valid: UOp = functools.reduce(operator.and_, [r.get_valid() for r in ctx.range_map[x][0]], UOp.const(dtypes.bool, True))
+  ret = valid.where(x.src[0], UOp.const(x.dtype, 0))
   ctx.range_map[ret] = ctx.range_map[x]
   return ret
 
@@ -77,21 +102,52 @@ pm_apply_rangeify = PatternMatcher([
   (UPat((Ops.CONST, Ops.DEFINE_VAR), name="c"), lambda ctx,c: c.replace(src=()) if c in ctx.range_map else None),
 ])
 
-def run_rangeify(tsink:UOp, realize_map:dict[UOp, None], debug) -> tuple[UOp, IndexingContext]:
-  tsink_base = UOp.sink(*[x.base for x in tsink.src])
+# this is the definition of the movement ops
+def apply_movement_op(x:UOp, rngs:Sequence[UOp]) -> list[UOp]:
+  match x.op:
+    case Ops.SHRINK:  rngs = [a if ss == 0 else a+ss for a,(ss,_) in zip(rngs, x.arg)]
+    case Ops.PERMUTE: rngs = [rngs[p] for p in argsort(x.arg)]
+    case Ops.FLIP:    rngs = [((s-1)-a) if f else a for a,s,f in zip(rngs, x.shape, x.arg)]
+    case Ops.EXPAND:  rngs = [a if in_sh == out_sh else a.const_like(0) for a,in_sh,out_sh in zip(rngs, x.src[0].shape, x.shape)]
+    case Ops.PAD:
+      # TODO: why is multiple graph_rewrites faster than one here?
+      rngs = [r if (s == 0 and e == 0) else graph_rewrite(((r >= s) & (r < (sh-e))).where(r-s, UOp.invalid()), sym, name="pad")
+              for r,sh,(s,e) in zip(rngs, x.shape, x.arg)]
+    case Ops.RESHAPE:
+      acc = 1
+      axes_in:list[UOp] = []
+      for s,src in list(zip(x.shape, rngs))[::-1]:
+        axes_in.append(acc*src)
+        acc *= s
+      combined_axes = sum(axes_in, start=UOp.const(dtypes.index, 0))
+      axes_out:list[UOp] = []
+      for s in x.src[0].shape[::-1]:
+        axes_out.append(combined_axes % s)
+        combined_axes //= s
+      # this simplify is doing a lot of heavy lifting. this is the replacement for the reshape view merging code
+      rngs = list(graph_rewrite(UOp.sink(*axes_out[::-1]), symbolic, name="reshape").src)
+    case _: raise RuntimeError(f"{x.op} is not a MovementOp")
+  return rngs
+
+@cpu_profile(TracingKey("run_rangeify"), "TINY")
+def run_rangeify(tsink:UOp, debug:bool=False) -> tuple[UOp, IndexingContext]:
+  rctx = IndexingContext()
+
+  # get ops to realize
+  graph_rewrite(tsink, pm_generate_realize_map, ctx=rctx.realize_map, name="Input Graph")
+
+  # get the traversal order
+  with cpu_profile(TracingKey("reverse toposort"), "TINY"):
+    tsink_reverse_toposort = tsink.reverse_toposort(consumer_map:=tsink.get_consumer_map())
 
   # explicit rangeify
-  rctx = IndexingContext()
   ending_ranges: dict[UOp, bool] = {}
-  for x in tsink_base.reverse_toposort(consumer_map:=tsink_base.get_consumer_map()):
+  for x in tsink_reverse_toposort:
     if x.op in {Ops.DEVICE, Ops.UNIQUE}: continue
     ending_ranges[x] = any(ending_ranges[u] for u in consumer_map[x])
 
     # if this element has weight and it's ending a range, we (force) realize it
-    if ending_ranges[x] and x.op in GroupOp.Elementwise.union({Ops.REDUCE_AXIS}):
-      if x.op_in_backward_slice_with_self(Ops.BUFFER, Ops.REALIZE, Ops.BUFFERIZE, Ops.CONTIGUOUS):
-        if x.op_in_backward_slice_with_self(Ops.REDUCE_AXIS):
-          realize_map[x] = None
+    if ending_ranges[x] and x.op in GroupOp.Elementwise.union({Ops.REDUCE_AXIS}): rctx.realize_map[x] = None
 
     # *** the ranges on the output are
     #  1. new if this op is realized
@@ -99,7 +155,7 @@ def run_rangeify(tsink:UOp, realize_map:dict[UOp, None], debug) -> tuple[UOp, In
     #  3. potentially new if this op has 2+ consumers
 
     consumer_rngs = [rctx.range_map[c][0] for c in consumer_map[x] if c in rctx.range_map]
-    if x in realize_map:
+    if x in rctx.realize_map:
       # if this is in the realize_map, we create new ranges (at the output)
       out_rngs = [rctx.new_range(s) for s in x.shape]
       # all ranges are ended now
@@ -131,12 +187,12 @@ def run_rangeify(tsink:UOp, realize_map:dict[UOp, None], debug) -> tuple[UOp, In
         if all_all_same:
           # the new valid is the OR of all the children valids
           minimum_valid = functools.reduce(operator.or_, valids, UOp.const(dtypes.bool, False))
-          out_rngs.append(minimum_valid.where(local_rngs[0], UOp.invalid()).simplify())
+          out_rngs.append(graph_rewrite(minimum_valid.where(local_rngs[0], UOp.invalid()), symbolic, name="minimum_valid"))
         else:
           out_rngs.append(rctx.new_range(x.shape[i]))
 
       # we have to realize here if there's new ranges
-      if not all_all_same: realize_map[x] = None
+      if not all_all_same: rctx.realize_map[x] = None
 
     # TODO: some ops don't have shape, enable this after the `.st` property is removed
     #assert len(out_rngs) == len(x.shape), \
@@ -149,39 +205,9 @@ def run_rangeify(tsink:UOp, realize_map:dict[UOp, None], debug) -> tuple[UOp, In
 
     rngs = out_rngs  # rngs is the input ranges
 
-    # apply movement ops. this is the definition of them
-    if x.op is Ops.SHRINK:  rngs = [a+ss if resolve(ss != 0) else a for a,(ss,_) in zip(rngs, x.arg)]
-    if x.op is Ops.PERMUTE: rngs = [rngs[p] for p in argsort(x.arg)]
-    if x.op is Ops.FLIP:    rngs = [((s-1)-a) if f else a for a,s,f in zip(rngs, x.shape, x.arg)]
-    if x.op is Ops.EXPAND:
-      rngs = [a if resolve(x==y, False) else a.const_like(0) for a,x,y in zip(rngs, x.src[0].shape, x.shape)]
-      ending_ranges[x] = True
-    if x.op is Ops.PAD:
-      rngs = rngs[:]
-      bigwhere = UOp.const(dtypes.bool, True)
-      for i,(sh,(s,e)) in enumerate(zip(x.shape, x.arg)):
-        if s == 0 and e == 0: continue
-        where = UOp.const(dtypes.bool, True)
-        if resolve(e > 0): where = where & (rngs[i] < (sh-e))
-        if resolve(s > 0): where = where & (rngs[i] >= s)
-        bigwhere = bigwhere & where
-        with Context(TRACK_MATCH_STATS=0):
-          rngs[i] = graph_rewrite(where.where(rngs[i]-s, UOp.invalid()), sym)
-      # PAD is replaced with a WHERE in the big graph to inject the 0s at the right place
-      rctx.pads_gate[x] = bigwhere.simplify()
-    if x.op is Ops.RESHAPE:
-      acc = 1
-      to_sum = []
-      for s,src in list(zip(x.shape, rngs))[::-1]:
-        to_sum.append(acc*src)
-        acc *= s
-      mish = sum(to_sum, start=UOp.const(dtypes.index, 0))
-      ret:list[UOp] = []
-      for s in x.src[0].shape[::-1]:
-        ret.append(mish % s) # NOTE: simplify will turn this to CONST
-        mish //= s
-      # this simplify is doing a lot of heavy lifting. this is the replacement for the view merger in RESHAPE
-      rngs = list(UOp.sink(*ret[::-1]).simplify().src)
+    # apply movement ops
+    if x.op in GroupOp.Movement: rngs = apply_movement_op(x, rngs)
+    if x.op is Ops.EXPAND: ending_ranges[x] = True
 
     # REDUCE_AXIS creates ranges for the axes it is reducing
     if x.op is Ops.REDUCE_AXIS:
@@ -190,12 +216,11 @@ def run_rangeify(tsink:UOp, realize_map:dict[UOp, None], debug) -> tuple[UOp, In
         if i in x.arg[1]: rngs[i] = rctx.new_range(s, axistype=AxisType.REDUCE)
 
     if debug:
-      print("***" if x in realize_map else "   ", len(consumer_map[x]), f"{str(x.op):20s}",
+      print("***" if x in rctx.realize_map else "   ", len(consumer_map[x]), f"{str(x.op):20s}",
             UOp.sink().index(*rngs).render(), " -> ", UOp.sink().index(*out_rngs).render())
 
     # assign to the range map. rngs are the input ranges, out_rngs are the output ranges, from the x op.
     rctx.range_map[x] = (rngs, out_rngs)
 
-  rctx.realize_map = realize_map
   tsink = graph_rewrite(tsink, pm_apply_rangeify, ctx=rctx, bottom_up=True, name="apply rangeify")
   return tsink, rctx

tinygrad/schedule/rangeify.py

@@ -1,16 +1,13 @@
-from typing import Any, cast, Iterator
-
-import functools, operator, itertools
+from typing import cast
 from dataclasses import dataclass, field
 from tinygrad.dtype import dtypes, PtrDType, ImageDType, AddrSpace
-from tinygrad.uop.ops import PatternMatcher, UPat, Ops, UOp, resolve, GroupOp, RewriteNotReady, _substitute, ssimplify, KernelInfo
-from tinygrad.uop.symbolic import sym, symbolic_simple
-from tinygrad.helpers import argsort, prod, all_same, pluralize, getenv, RANGEIFY, Context, flatten, dedup, unwrap, all_int, DEBUG, SPLIT_REDUCEOP
-from tinygrad.helpers import Metadata
+from tinygrad.uop.ops import PatternMatcher, UPat, Ops, UOp, resolve, GroupOp, _substitute, ssimplify, KernelInfo
 from tinygrad.uop.ops import track_rewrites, graph_rewrite, identity_element, sint, AxisType
+from tinygrad.uop.symbolic import symbolic_simple
+from tinygrad.helpers import argsort, prod, all_same, pluralize, getenv, flatten, dedup, unwrap, all_int, DEBUG, SPLIT_REDUCEOP, Metadata
 from tinygrad.codegen.simplify import pm_flatten_range, pm_reduce_unparented
 from tinygrad.codegen.opt import Opt
-from tinygrad.schedule.indexing import run_rangeify, BufferizeOpts, IndexingContext
+from tinygrad.schedule.indexing import run_rangeify, BufferizeOpts, ALWAYS_CONTIGUOUS, IndexingContext, apply_movement_op
 
 # creation can recurse a lot
 import sys
@@ -19,10 +16,6 @@ sys.setrecursionlimit(10000)
 # *****************
 # 0. do some cleanup rewrites, mostly copied from the old stuff
 
-ALWAYS_CONTIGUOUS: set[Ops] = {Ops.CONTIGUOUS, Ops.ASSIGN, Ops.COPY, Ops.BUFFER, Ops.BUFFER_VIEW,
-                     Ops.CONST, Ops.BIND, Ops.DEVICE, Ops.MSELECT, Ops.MSTACK, Ops.DEFINE_GLOBAL,
-                     Ops.DEFINE_LOCAL, Ops.DEFINE_REG, Ops.LOAD, Ops.KERNEL}
-
 def find_permutes(a:UOp, b:UOp, assign:UOp):
   if not (permutes:=[s for s in b.toposort(gate=lambda s:s.op not in ALWAYS_CONTIGUOUS)
                      if s.op in GroupOp.Movement and s.op not in {Ops.RESHAPE, Ops.EXPAND, Ops.PAD, Ops.SHRINK}]): return
@@ -104,285 +97,13 @@ earliest_rewrites = PatternMatcher([
   (UPat(Ops.CONTIGUOUS, name="root", src=(UPat(Ops.BUFFER),)), lambda root: root.src[0].forced_reshape(root.shape).rtag(root.tag)),
 ])
 
-# *****************
-# 1. add realize where we have to
-
-def realize(ctx:dict[UOp, None], tr:UOp) -> None: ctx[tr] = None
-
-def realize_parents(ctx:dict[UOp, None], rb:UOp) -> None:
-  for s in rb.src:
-    if s.base.op not in ALWAYS_CONTIGUOUS: ctx[s] = None
-
-def realize_assign(ctx:dict[UOp, None], a:UOp) -> None:
-  if a.src[1].op not in ALWAYS_CONTIGUOUS: ctx[a.src[1]] = None
-  # if it's a kernel, we don't realize it
-  if a.src[1].op is not Ops.KERNEL: ctx[a] = None
-
-do_realize = PatternMatcher([
-  # always realize SINK parents
-  (UPat(Ops.SINK, name="s"), lambda ctx,s: ctx.update((x.base, None) for x in s.src if x.base.op not in ALWAYS_CONTIGUOUS)),
-  # always realize ASSIGN/COPY/BUFFER_VIEW/CONTIGUOUS
-  (UPat({Ops.COPY, Ops.BUFFER_VIEW, Ops.CONTIGUOUS}, name="tr"), realize),
-  # realize parents of COPY, MSELECT, MSTACK
-  (UPat((Ops.COPY, Ops.MSELECT, Ops.MSTACK), name="rb"), realize_parents),
-  # realize input to assign (might be optimized out)
-  (UPat(Ops.ASSIGN, name="a"), realize_assign),
-])
-
-class WrappedContig:
-  def __init__(self, x): self.x = x
-  def __repr__(self): return f"C({self.x})"
-add_contiguous = PatternMatcher([(UPat(GroupOp.All, name="x"), lambda ctx,x: x.replace(tag=WrappedContig(x.tag)).realize() if x in ctx else None),])
-remove_contig_tags = PatternMatcher([(UPat(GroupOp.All, name="x"), lambda x: x.replace(tag=x.tag.x) if isinstance(x.tag, WrappedContig) else None)])
-
-# *****************
-# 2. mark all children
-
-@dataclass
-class ChildrenContext: children: dict[UOp, list[UOp]]|None = None
-def extract_children(ctx:ChildrenContext, x:UOp):
-  if ctx.children is not None: return
-  children_map = x.get_consumer_map()
-  ctx.children = {}
-  for k,v in children_map.items():
-    # NOTE: we treat mstack children like sink here
-    non_sink_children = [u for u in v if u.op not in {Ops.SINK, Ops.MSTACK}]
-    if len(non_sink_children) <= 1: continue
-    # NOTE: this gate shouldn't be here
-    if k.op_in_backward_slice_with_self(Ops.REDUCE_AXIS) and k.op_in_backward_slice_with_self(Ops.BUFFER, Ops.CONTIGUOUS):
-      ctx.children[k] = non_sink_children
-
-def mark_children(ctx:ChildrenContext, x:UOp):
-  assert ctx.children is not None
-  new_srcs = [(UOp(Ops.CHILD, s.dtype, src=(UOp(Ops.CHILDREN, s.dtype, (s,), arg=len(ctx.children[s])),),
-                   arg=(ctx.children[s].index(x), len(ctx.children[s]))) if s in ctx.children else s) for s in x.src]
-  return x.replace(src=tuple(new_srcs))
-
-pm_children = PatternMatcher([
-  (UPat(Ops.SINK, name="x"), extract_children),
-  (UPat(GroupOp.All-{Ops.CHILD, Ops.CHILDREN, Ops.SINK}, name="x"), mark_children),
-])
-
 # *****************
 # 3a. rangeify (movement)
 
-@dataclass
-class RangeifyContext:
-  # block on parent until all children have been seen
-  seen_children: dict[UOp, dict[int, UOp]] = field(default_factory=dict)
-  seen_child: dict[UOp, Any] = field(default_factory=dict)
-  progress: int = 0
-
-  # create ranges
-  range_idx: Iterator[int] = field(default_factory=itertools.count)
-  def new_range(self, s:sint, axistype:AxisType=AxisType.LOOP):
-    return UOp.range(s, next(self.range_idx), axistype) if resolve(s!=1) else UOp.const(dtypes.index, 0)
-
-def map_reshape(idx:UOp, r:UOp):
-  acc = 1
-  to_sum = []
-  for s,src in list(zip(idx.shape, idx.src[1:]))[::-1]:
-    to_sum.append(acc*src)
-    acc *= s
-  mish = sum(to_sum, start=UOp.const(dtypes.index, 0))
-  ret:list[UOp] = []
-  for s in r.src[0].shape[::-1]:
-    ret.append(mish % s) # NOTE: simplify will turn this to CONST
-    mish //= s
-  tret = UOp.sink(*ret[::-1]).simplify().src
-  return r.src[0].index(*tret, dtype=idx.dtype, arg=idx.arg)
-
-def map_pad(idx:UOp, r:UOp):
-  ret = list(idx.src[1:])
-  bigwhere = UOp.const(dtypes.bool, True)
-  for i,(sh,(s,e)) in enumerate(zip(r.shape, r.arg)):
-    if s == 0 and e == 0: continue
-    where = UOp.const(dtypes.bool, True)
-    if resolve(e > 0): where = where & (ret[i] < (sh-e))
-    if resolve(s > 0): where = where & (ret[i] >= s)
-    bigwhere = bigwhere & where
-    with Context(TRACK_MATCH_STATS=0):
-      ret[i] = graph_rewrite(where.where(ret[i]-s, UOp.invalid()), sym)
-  # PAD is with 0
-  return bigwhere.simplify().where(r.src[0].index(*ret, dtype=idx.dtype, arg=idx.arg), UOp.const(r.dtype, 0))
-
-def map_expand(r:UOp, idx:UOp):
-  new_rngs = []
-  ending_ranges = []
-  non_ending_ranges = []
-  for a,x,y in zip(idx.src[1:], r.src[0].shape, r.shape):
-    axis_to_range = [u for u in a.toposort() if u.op is Ops.RANGE]
-    if resolve(x==y, False):
-      non_ending_ranges.extend(axis_to_range)
-      new_rngs.append(a)
-    else:
-      ending_ranges.extend(axis_to_range)
-      new_rngs.append(a.const_like(0))
-  # if RANGEIFY >= 2, we are aggressive about not ending ranges
-  if RANGEIFY >= 2: ending_ranges = [x.arg for x in ending_ranges if x not in non_ending_ranges]
-  # if RANGEIFY=1, if it's ending at all we end it
-  else: ending_ranges = [x.arg for x in ending_ranges]
-  if idx.arg is not None: ending_ranges.append(idx.arg)
-  return r.src[0].index(*new_rngs, arg=min(ending_ranges) if ending_ranges else None)
-
+# movement op on INDEX as a PatternMatcher
 pm_mops = PatternMatcher([
-  # this is like the definitions of these
-  (UPat(Ops.SHRINK, name="r").f(Ops.INDEX, allow_any_len=True, name="idx"),
-   lambda r,idx: r.src[0].index(*[a+ss if resolve(ss != 0) else a for a,(ss,_) in zip(idx.src[1:], r.arg)], dtype=idx.dtype, arg=idx.arg)),
-  (UPat(Ops.PERMUTE, name="r").f(Ops.INDEX, allow_any_len=True, name="idx"),
-   lambda r,idx: r.src[0].index(*[idx.src[1+p] for p in argsort(idx.src[0].arg)], dtype=idx.dtype, arg=idx.arg)),
-  (UPat(Ops.FLIP, name="r").f(Ops.INDEX, allow_any_len=True, name="idx"),
-   lambda r,idx: r.src[0].index(*[((s-1)-a) if f else a for a,s,f in zip(idx.src[1:], r.shape, r.arg)], dtype=idx.dtype, arg=idx.arg)),
-  # expand needs to end ranges
-  (UPat(Ops.EXPAND, name="r").f(Ops.INDEX, allow_any_len=True, name="idx"), map_expand),
-  # reshape does a lot of symbolic stuff
-  (UPat(Ops.RESHAPE, name="r").f(Ops.INDEX, allow_any_len=True, name="idx"), map_reshape),
-  # pad adds min and max
-  (UPat(Ops.PAD, name="r").f(Ops.INDEX, allow_any_len=True, name="idx"), map_pad),
-])
-
-# *****************
-# 3b. rangeify (ops)
-
-# bufferization can happen in three ways
-#   1. there's an explicit REALIZE in the graph
-#   2. the ranges from the children don't match and we have to create a buffer (only on children)
-#   3. might_end_axis triggers because we should be closing a loop to save compute
-
-def map_partial_realize(ctx:RangeifyContext, x:UOp, idx:UOp):
-  if x.arg is None: return None  # map_contiguous can handle this
-  # NOTE: all partial contiguous can safely be replaced by full contiguous. we should be able to match old functionality like this
-  if not (RANGEIFY > 1): return idx.replace(src=(x.replace(arg=None),)+idx.src[1:])
-  ranges = []
-  new_ranges = []
-  passthrough_idx = []
-  for i,s in enumerate(x.shape):
-    if i not in x.arg:
-      ranges.append(idx.src[1+i])
-      continue
-    passthrough_idx.append(idx.src[1+i])
-    ranges.append(ctx.new_range(s))
-    new_ranges.append(ranges[-1])
-  # TODO: this should be able to be global or local
-  ret = x.src[0].index(*ranges).bufferize(*[x for x in new_ranges if x.op is not Ops.CONST],
-                                          arg=BufferizeOpts(device=None, addrspace=AddrSpace.LOCAL))
-  return ret.index(*passthrough_idx)
-
-def map_realize(ctx:RangeifyContext, x:UOp):
-  if x.arg is not None: return None
-  ranges = [ctx.new_range(s) for s in x.shape]
-  return x.src[0].index(*ranges).bufferize(*x.src[1:], *ranges, arg=BufferizeOpts(device=x.device), tag=x.src[0].tag)
-
-def map_reduce(ctx:RangeifyContext, idx:UOp, red:UOp):
-  rngs = list(idx.src[1:])
-  new_ranges = []
-  for i,s in enumerate(red.src[0].shape):
-    if i in red.arg[1]:
-      rngs[i] = ctx.new_range(s, axistype=AxisType.REDUCE)
-      new_ranges.append(rngs[i])
-  return UOp(Ops.REDUCE, red.dtype, src=(red.src[0].index(*rngs),)+tuple(new_ranges), arg=red.arg[0], tag=red.tag)
-
-def index_child(ctx:RangeifyContext, c:UOp, x:UOp, idx:UOp):
-  if c not in ctx.seen_children: ctx.seen_children[c] = {}
-  ctx.seen_children[c][x.arg[0]] = idx
-  # wait here until we have seen all the children
-  if len(ctx.seen_children[c]) != x.arg[1]:
-    ctx.progress += 1
-    if ctx.progress > 10000: raise RuntimeError("children not making progress")
-    raise RewriteNotReady
-  ctx.progress = 0
-
-  if c not in ctx.seen_child:
-    all_rngs = list(zip(*[ch.src[1:] for ch in ctx.seen_children[c].values()]))
-    out_rngs = []
-    end_ranges = []
-    idx_ranges = []
-    # NOTE: locals aren't working, so we only fully bufferize here (unless RANGEIFY > 1)
-    rngs_valids = []
-    for valid_rngs in all_rngs:
-      rngs, valids = zip(*[(r.get_idx(), r.get_valid()) for r in valid_rngs])
-      # if a range has a 1 src, it's the same as UOp.const(dtypes.index, 0)
-      same_rngs = [x if x.op is not Ops.RANGE or resolve(x.src[0] != 1) else UOp.const(dtypes.index, 0) for x in rngs]
-      rngs_valids.append((rngs, valids, all_same(same_rngs)))
-    all_all_same = all(same_rngs for _,_,same_rngs in rngs_valids)
-    for i,(rngs,valids,same_rngs) in enumerate(rngs_valids):
-      # we compare the ranges without their valids
-      if same_rngs and (all_all_same or RANGEIFY > 1):
-        # the new valid is the OR of all the children valids
-        minimum_valid = functools.reduce(operator.or_, valids, UOp.const(dtypes.bool, False))
-        out_rngs.append(minimum_valid.where(rngs[0], UOp.invalid()).simplify())
-      else:
-        out_rngs.append(ctx.new_range(c.shape[i]))
-        end_ranges.append(out_rngs[-1])
-        idx_ranges.append(i)
-    ctx.seen_child[c] = (out_rngs, idx_ranges, end_ranges)
-  else:
-    out_rngs, idx_ranges, end_ranges = ctx.seen_child[c]
-    for i,nr in zip(idx_ranges, end_ranges): out_rngs[i] = nr
-  # index based on the shared ranges
-  ret = c.index(*out_rngs)
-  # if all ranges aren't the same between children, we have to bufferize
-  if len(idx_ranges) > 0:
-    if len(idx_ranges) == len(out_rngs):
-      # this is a global bufferize
-      ret = ret.bufferize(*end_ranges, arg=BufferizeOpts(device=x.device))
-    else:
-      assert RANGEIFY > 1, "this isn't supported with RANGEIFY=1"
-      ret = ret.bufferize(*end_ranges, arg=BufferizeOpts(device=None, addrspace=AddrSpace.LOCAL))
-    ret = ret.index(*[idx.src[1+i] for i in idx_ranges])
-  return ret
-
-def children_gate(ctx:RangeifyContext, idx:UOp, c:UOp):
-  if len(ctx.seen_children[c]) != c.arg: raise RuntimeError("all children should have been seen by now")
-  return idx.replace(src=(idx.src[0].src[0],)+idx.src[1:])
-
-def might_end_axis(idx:UOp):
-  if idx.arg is None: return None
-  # TODO: write a proper cost function here
-  if not idx.op_in_backward_slice_with_self(Ops.BUFFER, Ops.REALIZE, Ops.BUFFERIZE): return None
-  if not idx.op_in_backward_slice_with_self(Ops.REDUCE_AXIS): return None
-  to_end_axis = []
-  for i,a in enumerate(idx.src[1:]):
-    # in RANGEIFY=1, always realize
-    if not (RANGEIFY > 1) or any(x.arg > idx.arg for x in a.toposort() if x.op is Ops.RANGE):
-      to_end_axis.append(i)
-  if to_end_axis: return idx.replace(src=(idx.src[0].realize(arg=tuple(to_end_axis)),)+idx.src[1:], arg=None)
-  return idx.replace(arg=None)
-
-def unprocessed_index(x:UOp): raise RuntimeError(f"unprocessed index on {x.src[0].op}")
-
-pm_rangeify = pm_mops+PatternMatcher([
-  # sink contigs to kick it off
-  (UPat(Ops.REALIZE, src=(UPat(),), name="x", allow_any_len=True), map_realize),
-  # if there's an INDEX it can support partial contig
-  (UPat(Ops.INDEX, src=(UPat(Ops.REALIZE, src=(UPat(),), name="x"),), allow_any_len=True, name="idx"), map_partial_realize),
-
-  # if there are new ended children, tag the SINK
-  (UPat(Ops.INDEX, src=(UPat(Ops.CHILD, src=(UPat(name="c"), ), name="x"),), allow_any_len=True, name="idx"), index_child),
-  (UPat(Ops.INDEX, src=(UPat(Ops.CHILDREN, name="c"),), allow_any_len=True, name="idx"), children_gate),
-
-  # if we come across this, remove it. it was a CHILD unused in an INDEX
-  (UPat(Ops.CHILD, src=(UPat(Ops.CHILDREN, src=(UPat.var("x"),)),)), lambda x: x),
-
-  # CONST (or DEFINE_VAR) can't have axes. remove INDEX when we get here
-  (UPat(Ops.INDEX, src=(UPat((Ops.CONST, Ops.DEFINE_VAR), name="c"),)), lambda c: c.replace(src=())),
-
-  # handle arg on any op with weight. old endrange stuff
-  (UPat(Ops.INDEX, src=(UPat(GroupOp.Elementwise.union({Ops.REDUCE_AXIS})),), allow_any_len=True, name="idx"), might_end_axis),
-
-  # handle assign
-  (UPat(Ops.INDEX, src=(UPat(Ops.ASSIGN, name="assign"),), allow_any_len=True, name="x"),
-    lambda x,assign: assign.replace(src=tuple([s.index(*x.src[1:]) for s in assign.src])+(assign.src[0],)) \
-        if assign.src[1].op is not Ops.KERNEL else None),
-
-  # move MAP through elementwise ALU / reduce. these are the items with cost
-  (UPat(Ops.INDEX, src=(UPat(GroupOp.Elementwise.union(
-    {Ops.STORE, Ops.COPY, Ops.BUFFER_VIEW, Ops.DEVICE, Ops.BIND, Ops.CONTIGUOUS, Ops.NOOP})),), allow_any_len=True, name="x"),
-   lambda x: x.src[0].replace(src=tuple([s.index(*x.src[1:]) for s in x.src[0].src]))),
-  (UPat(Ops.INDEX, src=(UPat(Ops.REDUCE_AXIS, name="red"),), allow_any_len=True, name="idx"), map_reduce),
-
-  # assert if there's any index we didn't process
-  (UPat(GroupOp.All-{Ops.REALIZE, Ops.BUFFERIZE, Ops.MSELECT, Ops.MSTACK}).f(Ops.INDEX, name="x"), unprocessed_index),
+  (UPat(GroupOp.Movement, name="r").f(Ops.INDEX, allow_any_len=True, name="idx"),
+   lambda r,idx: r.src[0].index(*apply_movement_op(r, idx.src[1:]), dtype=idx.dtype, arg=idx.arg)),
 ])
 
 # *****************
@@ -500,7 +221,7 @@ to_bufferview = PatternMatcher([
 ])
 
 DEVICE_MAX_BUFS = {"METAL": 31, "WEBGPU": 8} # TODO: get from device?
-def limit_bufs(ctx:RangeifyContext, root:UOp):
+def limit_bufs(ctx:IndexingContext, root:UOp):
   if (device:=root._device) is None: return None # no device, index related calculations
   device = device if isinstance(device, str) else device[0].split(":")[0]
   if not (MAX_BUFS:=getenv("MAX_KERNEL_BUFFERS", DEVICE_MAX_BUFS.get(device, 0))): return None
@@ -760,19 +481,9 @@ def get_rangeify_map(sink:UOp) -> dict[UOp, UOp]:
   tsink = graph_rewrite(sink, add_tags, ctx=uop_list, bottom_up=True, name="number the uops")
 
   tsink = graph_rewrite(tsink, earliest_rewrites+replace_contiguous, ctx={}, name="earliest rewrites")
-  realize_map: dict[UOp, None] = {}
-  graph_rewrite(tsink, do_realize, ctx=realize_map, name="Input Graph")
 
-  FAST = getenv("FAST", 1)
-  if FAST:
-    rctx: RangeifyContext|IndexingContext
-    tsink, rctx = run_rangeify(tsink, realize_map, FAST > 1)
-  else:
-    # NOTE: we don't use contiguous here, contiguous is a user op
-    tsink = graph_rewrite(tsink, add_contiguous, ctx=realize_map, bottom_up=True, name="add realize")
-    tsink = graph_rewrite(tsink, remove_contig_tags, name="remove contiguous tags")
-    tsink = graph_rewrite(tsink, pm_children, ctx=ChildrenContext(), bottom_up=True, name="get children")
-    tsink = graph_rewrite(tsink, pm_rangeify, ctx=(rctx:=RangeifyContext()), bottom_up=True, name="rangeify")
+  # convert movement ops to ranges
+  tsink, rctx = run_rangeify(tsink, getenv("DEBUG_RANGEIFY", 0))
 
   # NOTE: sym (vs symbolic_simple) breaks things here because ranges with len 1 aren't handled right
   tsink = graph_rewrite(tsink, symbolic_simple+pm_reduce_unparented, name="symbolic")  # this supports const folding

tinygrad/shape/shapetracker.py

@@ -12,7 +12,6 @@ from tinygrad.uop.ops import UOp, Ops, graph_rewrite, Variable, sint, sint_to_uo
 def views_to_valid_uop(views: tuple[View, ...], _idxs:tuple[UOp, ...]|None=None) -> UOp:
   idx = views[-1].to_valid_uop(_idxs)
   for view in reversed(views[0:-1]):
-    view = view.minify()
     idx = view.to_valid_uop([sint_to_uop(i) for i in unravel(view.shape, idx)])
   with Context(TRACK_MATCH_STATS=0):
     return graph_rewrite(idx, sym, name="indexing sym @ 1")
@@ -72,7 +71,6 @@ class ShapeTracker:
 
   def real_strides(self, ignore_valid=False) -> tuple[sint|None, ...]:
     with Context(TRACK_MATCH_STATS=0): return views_to_real_strides(self.views, ignore_valid)
-  def unit_stride_axes(self, ignore_valid=False) -> list[int]: return [i for i,st in enumerate(self.real_strides(ignore_valid)) if st == 1]
 
   def simplify(self) -> ShapeTracker:
     if len(self.views) >= 2 and (new_view := self.views[-2] + self.views[-1]) is not None:

tinygrad/shape/view.py

@@ -4,7 +4,7 @@ from dataclasses import dataclass
 from typing import cast, Sequence
 from tinygrad.dtype import dtypes
 from tinygrad.uop.ops import resolve, UOp, Variable, sint, smax, smin, sint_to_uop, Ops, ssimplify
-from tinygrad.helpers import prod, all_int, argsort, flatten, ceildiv
+from tinygrad.helpers import prod, all_int, flatten, ceildiv
 
 # returns the axes to create new_shape if new_shape can be created by combining axis from old_shape
 def get_contraction(old_shape:tuple[sint, ...], new_shape:tuple[sint, ...]) -> list[list[int]]|None:
@@ -13,24 +13,6 @@ def get_contraction(old_shape:tuple[sint, ...], new_shape:tuple[sint, ...]) -> l
   except ValueError: return None
   return [list(range(st,ed)) for st,ed in zip([0]+split[:-1], split[:-1]+[len(old_shape)])]
 
-def get_contraction_with_reduce(old_shape:tuple[sint, ...], new_shape:tuple[sint, ...], reduce_axis:tuple[int, ...]) -> list[list[int]]|None:
-  if (contraction:=get_contraction(old_shape, new_shape)) is None: return None
-  # contraction returns the 1s as right justified as possible
-  # normally this contraction is good, but sometimes the reduce dim is empty. borrow from the next one, leaving one
-  # this ensures there's always ones available in the reduce dimension. this is also a valid contraction
-  for i in range(len(contraction)):
-    if i in reduce_axis and len(contraction[i]) == 0:
-      take_from = i+1
-      while take_from < len(contraction) and len(contraction[take_from]) == 0:
-        assert new_shape[take_from] == 1
-        take_from += 1
-      if take_from == len(contraction) or new_shape[take_from] != 1: return None # nothing to take
-      for j in range(take_from, i, -1):
-        assert len(contraction[j]) > 0
-        contraction[j-1] = contraction[j][:-1]
-        contraction[j] = contraction[j][-1:]
-  return contraction
-
 @functools.cache
 def canonicalize_strides(shape:tuple[sint, ...], strides:tuple[sint, ...]) -> tuple[sint, ...]:
   return tuple(0 if s == 1 else st for s, st in zip(shape, strides))
@@ -244,18 +226,6 @@ class View:
 
     return View.create(vm1.shape, tuple(strides), ssimplify(sum(o * s for o, s in zip(origin, vm2.strides)) + vm2.offset))
 
-  @functools.cache  # pylint: disable=method-cache-max-size-none
-  def invert(self, out_shape:tuple[sint, ...]) -> View|None:
-    ret = View.create(self.shape)
-    if self.mask: ret = ret.shrink(self.mask)
-    ret = ret.flip(tuple(x < 0 for x in self.strides)).permute(argsort(tuple(-x if x > 0 else x for x in self.strides)))
-    return ret if prod(ret.shape) == prod(out_shape) else None   # don't support shrink, expand, or stride != (-1, 1)
-
-  @functools.cache  # pylint: disable=method-cache-max-size-none
-  def minify(self):
-    min_shape = tuple(x[0] for x in merge_dims(self.shape, self.strides, self.mask))
-    return nv if (nv := self.reshape(min_shape)) else self
-
   def __unsafe_resize(self, arg: tuple[tuple[sint, sint], ...], mask=None) -> View:
     offset = sum([s * x[0] for s, x in zip(self.strides,arg)])
     if self.mask:

tinygrad/uop/__init__.py

@@ -12,9 +12,6 @@ class Ops(FastEnum):
   NOOP = auto(); SINK = auto(); UNIQUE = auto(); DEVICE = auto(); KERNEL = auto(); PRECAST = auto(); REWRITE_ERROR = auto()  # noqa: E702
   SENTINEL = auto()
 
-  # track children
-  CHILD = auto(); CHILDREN = auto() # noqa: E702
-
   # buffer ops
   COPY = auto(); BUFFER = auto(); BUFFER_VIEW = auto(); MSELECT = auto(); MSTACK = auto() # noqa: E702
 
@@ -24,7 +21,6 @@ class Ops(FastEnum):
 
   # ops that adjust the behavior of the scheduler
   CONTIGUOUS = auto(); CONTIGUOUS_BACKWARD = auto(); DETACH = auto(); FUSE = auto() # noqa: E702
-  REALIZE = auto()
 
   # blocks in linearizer (only used there)
   BLOCK = auto(); BLOCKSTART = auto(); BLOCKEND = auto(); BLOCKFINAL = auto() # noqa: E702
@@ -112,6 +108,4 @@ class GroupOp:
   # do not preserve f(0) = 0
   UnsafePad = {Ops.RECIP, Ops.LOG2, Ops.EXP2, Ops.IDIV, Ops.POW}
 
-  Meta = {Ops.COPY, Ops.BUFFER_VIEW}
-
   All = set(Ops)

tinygrad/uop/ops.py

@@ -23,9 +23,6 @@ range_start = {Ops.BUFFERIZE: 1, Ops.REDUCE: 1, Ops.STORE: 2, Ops.WMMA: 3}
 # https://en.wikipedia.org/wiki/Identity_element
 def identity_element(op:Ops, dt:DType) -> ConstType: return dtypes.as_const({Ops.ADD:0, Ops.MUL:1, Ops.MAX:dtypes.min(dt)}[op], dt)
 
-def can_pad(root:UOp, edges:dict[UOp, None]) -> bool:
-  return all(u.op not in GroupOp.UnsafePad for u in root.toposort(gate=lambda x:x not in edges))
-
 # With True as the default, this matches the old symbolic behavior
 def resolve(x:UOp|bool, default:bool=True):
   if isinstance(x, bool): return x
@@ -223,7 +220,10 @@ class UOp(MathTrait, metaclass=UOpMetaClass):
       case Ops.BITCAST:
         shape = src_sts[0].shape
         if self.dtype.itemsize != (input_sz:=self.src[0].dtype.itemsize): shape = shape[:-1]+((shape[-1]*input_sz) // self.dtype.itemsize,)
-      case Ops.REDUCE_AXIS | Ops.WMMA: shape = src_sts[0].reduce(self.axis_arg)
+      case Ops.REDUCE_AXIS | Ops.WMMA:
+        axis_arg = self.arg[1] if self.op is Ops.REDUCE_AXIS else self.arg[7]
+        assert isinstance(axis_arg, tuple) and all(isinstance(x, int) for x in axis_arg), f"invalid type for axis: {axis_arg}"
+        shape = src_sts[0].reduce(axis_arg)
       case _: shape = src_sts[0].shape
     return ShapeTracker.from_shape(shape)
 
@@ -286,16 +286,6 @@ class UOp(MathTrait, metaclass=UOpMetaClass):
 
   # *** uop syntactic sugar ***
 
-  @property
-  def st_arg(self) -> ShapeTracker:
-    assert self.op in GroupOp.Buffer, f"st_arg called on {self.op}"
-    return unwrap(self.st)
-  @property
-  def axis_arg(self) -> tuple[int, ...]:
-    assert self.op in {Ops.REDUCE_AXIS, Ops.WMMA}, f"axis_arg called on {self.op}"
-    ret = self.arg[1] if self.op is Ops.REDUCE_AXIS else self.arg[7]
-    assert isinstance(ret, tuple) and all(isinstance(x, int) for x in ret), f"axis_arg trying to return {ret}"
-    return ret
   def sink(*srcs:UOp|None, **kwargs):  # pylint: disable=no-self-argument
     return UOp(Ops.SINK, dtypes.void, tuple([x for x in srcs if x is not None]), **kwargs)
   def detach(self): return UOp(Ops.DETACH, self.dtype, (self,))
@@ -367,7 +357,6 @@ class UOp(MathTrait, metaclass=UOpMetaClass):
     return self.src[0] if self.op is Ops.WHERE and self.src[2].arg is Invalid else UOp.const(dtypes.bool, self.arg is not Invalid)
   def reduce(self, *src:UOp, **kwargs): return UOp(Ops.REDUCE, kwargs.pop('dtype', self.dtype), src=(self,)+src, **kwargs)
   def contiguous(self, *args, **kwargs): return UOp(Ops.CONTIGUOUS, dtype=self.dtype, src=(self,)+args, **kwargs)
-  def realize(self, *args, **kwargs): return UOp(Ops.REALIZE, dtype=self.dtype, src=(self,)+args, **kwargs)
   def contiguous_backward(self): return self.alu(Ops.CONTIGUOUS_BACKWARD)
   def bufferize(self, *args, **kwargs): return UOp(Ops.BUFFERIZE, dtype=self.dtype, src=(self,)+args, **kwargs)
   def fuse(self): return self.alu(Ops.FUSE)
@@ -502,7 +491,7 @@ class UOp(MathTrait, metaclass=UOpMetaClass):
   def buffer(self) -> Buffer|MultiBuffer:
     from tinygrad.device import Buffer, MultiBuffer
     if self is not self.base:
-      assert unwrap(self.st).contiguous, "VIEW only works here if it's contiguous"
+      assert self.op is Ops.RESHAPE, f"can only be RESHAPE {self}"
       return self.src[0].buffer
     if self.op is Ops.MSELECT:
       ret = self.src[0].buffer
@@ -880,11 +869,11 @@ match_stats:dict[UPat, list[int|float]] = dict()
 
 @dataclass(frozen=True)
 class TrackedGraphRewrite:
-  loc:tuple[str, int]                    # location that called graph_rewrite
-  sink:int                               # the sink input to graph_rewrite
-  matches:list[tuple[int, int, tuple]]   # before/after UOp, UPat location
-  name:str|None                          # optional name of the rewrite
-  depth:int                              # depth if it's a subrewrite
+  loc:tuple[str, int]                           # location that called graph_rewrite
+  sink:int                                      # the sink input to graph_rewrite
+  matches:list[tuple[int, int, tuple, float]]   # before/after UOp, UPat location and time
+  name:str|None                                 # optional name of the rewrite
+  depth:int                                     # depth if it's a subrewrite
   bottom_up:bool
 
 tracked_keys:list[TracingKey] = []
@@ -954,16 +943,16 @@ class TrackedPatternMatcher(PatternMatcher):
         continue
       match_stats[p][1] += 1
       try: ret = match(uop, ctx)
-      except Exception as e:
-        if TRACK_MATCH_STATS >= 2 and active_rewrites and not isinstance(e, RewriteNotReady):
-          active_rewrites[-1].matches.append((track_uop(uop), track_uop(UOp(Ops.REWRITE_ERROR, src=uop.src, arg=str(sys.exc_info()[1]))), p.location))
+      except Exception:
+        if TRACK_MATCH_STATS >= 2 and active_rewrites:
+          active_rewrites[-1].matches.append((track_uop(uop), track_uop(UOp(Ops.REWRITE_ERROR,src=uop.src,arg=str(sys.exc_info()[1]))),p.location,0))
         raise
       if ret is not None and ret is not uop:
         match_stats[p][0] += 1
         match_stats[p][3] += (et:=time.perf_counter()-st)
         if TRACK_MATCH_STATS >= 3: print(f"{et*1e6:7.2f} us -- ", printable(p.location))
         if TRACK_MATCH_STATS >= 2 and isinstance(ret, UOp) and active_rewrites:
-          active_rewrites[-1].matches.append((track_uop(uop), track_uop(ret), p.location))
+          active_rewrites[-1].matches.append((track_uop(uop), track_uop(ret), p.location, et))
         return ret
       match_stats[p][2] += time.perf_counter()-st
     return None
@@ -993,12 +982,11 @@ if TRACK_MATCH_STATS or PROFILE:
     if not int(os.getenv("VIZ", "0")) and not int(os.getenv("PROFILE", "0")) and not int(os.getenv("SQTT", "0")):
       args = ['--kernels', getenv("VIZ_DATA", "")] if getenv("VIZ_DATA", "") else []
       args += ['--profile', getenv("PROFILE_DATA", "")] if getenv("PROFILE_DATA", "") else []
-      os.execv(sys.executable, [sys.executable] + [os.path.join(os.path.dirname(__file__), "../", "viz", "serve.py")] + args)
+      os.execv(sys.executable, [sys.executable] + [pathlib.Path(__file__).resolve().parent.parent / "viz" / "serve.py"] + args)
 
 # *** simple graph rewrite engine ***
 
 with Context(SPEC=0): SENTINEL = UOp(Ops.SENTINEL)
-class RewriteNotReady(Exception): pass
 class BottomUpGate(Exception): pass
 class RewriteContext:
   def __init__(self, pm, bpm, ctx=None):
@@ -1021,7 +1009,6 @@ class RewriteContext:
 
   def unified_rewrite(self, root:UOp) -> UOp:
     stack: collections.deque[tuple[UOp, int, UOp]] = collections.deque([(root, 0, root)])
-    on_stack = {root}  # all UOps either on the stack or in self.replace, i.e. dont have to be placed again
     REWRITE_STACK_LIMIT = getenv("REWRITE_STACK_LIMIT", 250000)
     while stack:
       if len(stack) > REWRITE_STACK_LIMIT: raise RuntimeError("infinite loop in graph_rewrite (stack too big)")
@@ -1038,48 +1025,28 @@ class RewriteContext:
               if test_n in seen: raise RuntimeError("infinite loop in fixed_point_rewrite")
               seen.add(test_n)
               new_n, test_n = test_n, self.cached_bpm_rewrite(test_n)
-          except RewriteNotReady:
-            # try the full thing again later
-            stack.appendleft((n, 0, n))
-            continue
           except BottomUpGate:
             # if the bpm matching raised a gate, we are done with this node and dont continue down the srcs
             self.replace[n] = new_n
             continue
         stack.append((n, 1, new_n))
-        for x in reversed(new_n.src):
-          if x in on_stack: continue
-          stack.append((x, 0, x))
-          on_stack.add(x)
+        for x in reversed(new_n.src): stack.append((x, 0, x))
       elif stage == 1:
-        tmp = []
-        for x in new_n.src:
-          if (rx:=self.replace.get(x, SENTINEL)) is SENTINEL:
-            # if some new sources aren't ready, we try this again later
-            stack.appendleft((n, 1, new_n))
-            break
-          tmp.append(rx)
+        # in stage 1, once all srcs are rewritten, rebuild (if changed) or run top-down rewrite
+        if (new_src:=tuple([self.replace[x] for x in new_n.src])) == new_n.src:
+          # if top down, do the rewrite. if no rewrite or bottom up, we are done rewriting this node so we add it to the dict
+          if self.pm is None or (new_src_n:=self.cached_pm_rewrite(new_n)) is None:
+            self.replace[n] = new_n
+            continue
         else:
-          # in stage 1, once all srcs are rewritten, rebuild (if changed) or run top-down rewrite
-          if (new_src:=tuple(tmp)) == new_n.src:
-            # if top down, do the rewrite. if no rewrite or bottom up, we are done rewriting this node so we add it to the dict
-            if self.pm is None or (new_src_n:=self.cached_pm_rewrite(new_n)) is None:
-              self.replace[n] = new_n
-              continue
-          else:
-            # if srcs changed from rewrites, construct a new UOp with the new srcs
-            new_src_n = UOp(new_n.op, new_n.dtype, new_src, new_n.arg, new_n.tag)
-          # trigger a rewrite of new_src_n, then after that rewrite is done, link it back to n
-          stack.append((n, 2, new_src_n))
-          stack.append((new_src_n, 0, new_src_n))
+          # if srcs changed from rewrites, construct a new UOp with the new srcs
+          new_src_n = UOp(new_n.op, new_n.dtype, new_src, new_n.arg, new_n.tag)
+        # trigger a rewrite of new_src_n, then after that rewrite is done, link it back to n
+        stack.append((n, 2, new_src_n))
+        stack.append((new_src_n, 0, new_src_n))
       else:
         # in stage 2, we link the result of new_n to the result of n
-        if (replaced_new_n:=self.replace.get(new_n, SENTINEL)) is SENTINEL:
-          # not ready, try the link later
-          stack.appendleft((n, 2, new_n))
-        else:
-          # otherwise we are done
-          self.replace[n] = replaced_new_n
+        self.replace[n] = self.replace[new_n]
     return self.replace[root]
 
 @track_matches

tinygrad/uop/spec.py

@@ -239,11 +239,6 @@ full_spec = PatternMatcher([
   # where on index in rhs position is fine
   (UPat(Ops.WHERE, src=(UPat(dtype=dtypes.bool), UPat(), UPat(dtype=dtypes.index))), lambda: True),
 
-  # all children is fine
-  (UPat(Ops.CHILDREN), lambda: True),
-  # child must have CHILDREN parent
-  (UPat(Ops.CHILD, src=(UPat(Ops.CHILDREN),)), lambda: True),
-
   # all rewrite error are okay
   (UPat(Ops.REWRITE_ERROR), lambda: True),
 
@@ -251,8 +246,6 @@ full_spec = PatternMatcher([
   (UPat(Ops.BUFFER_VIEW, src=(UPat((Ops.INDEX, Ops.LOAD)),)), lambda: True),
   # bufferize (must be on ranges)
   (UPat(Ops.BUFFERIZE, src=(UPat(),), allow_any_len=True, name="x"), lambda x: all(y.op in {Ops.RANGE, Ops.CONST} for y in x.src[1:])),
-  # realize with one src is fine
-  (UPat(Ops.REALIZE, src=(UPat(),)), lambda: True),
   # intermediate index
   (UPat(Ops.INDEX, src=(UPat(),), allow_any_len=True, name="x"), lambda x: all(y.dtype == dtypes.index for y in x.src[1:]) or None),
   (UPat(Ops.REDUCE, src=(UPat(),), allow_any_len=True, name="x"), lambda x: all(y.dtype == dtypes.index for y in x.src[1:])),

tinygrad/viz/js/index.js

@@ -225,7 +225,7 @@ async function renderProfiler() {
         else if (ref != null) {
           const start = ref.step>0 ? ref.step+1 : 0;
           const stepIdx = ctxs[ref.ctx+1].steps.findIndex((s, i) => i >= start && s.name == e.name);
-          ref = stepIdx === -1 ? null : {ctx:ref.ctx, step:stepIdx};
+          ref = {ctx:ref.ctx, step:stepIdx};
         }
         const htmlLabel = label.map(({color, st}) => `<span style="color:${color}">${st}</span>`).join('');
         const arg = { tooltipText:htmlLabel+"\n"+formatTime(e.dur)+(e.info != null ? "\n"+e.info : ""), ...ref };

tinygrad/viz/serve.py

@@ -19,9 +19,8 @@ uops_colors = {Ops.LOAD: "#ffc0c0", Ops.STORE: "#87CEEB", Ops.CONST: "#e0e0e0",
                Ops.INDEX: "#e8ffa0", Ops.WMMA: "#efefc0", Ops.MULTI: "#f6ccff", Ops.KERNEL: "#3e7f55",
                **{x:"#D8F9E4" for x in GroupOp.Movement}, **{x:"#ffffc0" for x in GroupOp.ALU}, Ops.THREEFRY:"#ffff80", Ops.BUFFER_VIEW: "#E5EAFF",
                Ops.BLOCK: "#C4A484", Ops.BLOCKEND: "#C4A4A4", Ops.BUFFER: "#B0BDFF", Ops.COPY: "#a040a0", Ops.FUSE: "#FFa500",
-               Ops.ALLREDUCE: "#ff40a0", Ops.MSELECT: "#d040a0", Ops.MSTACK: "#d040a0", Ops.CONTIGUOUS: "#FFC14D", Ops.REALIZE: "#C1C14D",
-               Ops.CHILDREN: "#80ffc0", Ops.CHILD: "#80fff0", Ops.BUFFERIZE: "#FF991C", Ops.REWRITE_ERROR: "#ff2e2e",
-               Ops.SUBSTITUTE: "#ffff00"}
+               Ops.ALLREDUCE: "#ff40a0", Ops.MSELECT: "#d040a0", Ops.MSTACK: "#d040a0", Ops.CONTIGUOUS: "#FFC14D",
+               Ops.BUFFERIZE: "#FF991C", Ops.REWRITE_ERROR: "#ff2e2e", Ops.SUBSTITUTE: "#ffff00"}
 
 # VIZ API
 
@@ -98,12 +97,13 @@ def _reconstruct(a:int, i:int):
 def get_details(ctx:TrackedGraphRewrite, i:int=0) -> Generator[GraphRewriteDetails, None, None]:
   yield {"graph":uop_to_json(next_sink:=_reconstruct(ctx.sink, i)), "uop":str(next_sink), "changed_nodes":None, "diff":None, "upat":None}
   replaces: dict[UOp, UOp] = {}
-  for u0_num,u1_num,upat_loc in tqdm(ctx.matches):
+  for u0_num,u1_num,upat_loc,dur in tqdm(ctx.matches):
     replaces[u0:=_reconstruct(u0_num, i)] = u1 = _reconstruct(u1_num, i)
     try: new_sink = next_sink.substitute(replaces)
     except RuntimeError as e: new_sink = UOp(Ops.NOOP, arg=str(e))
+    match_repr = f"# {dur*1e6:.2f} us\n"+printable(upat_loc)
     yield {"graph":(sink_json:=uop_to_json(new_sink)), "uop":str(new_sink), "changed_nodes":[id(x) for x in u1.toposort() if id(x) in sink_json],
-           "diff":list(difflib.unified_diff(str(u0).splitlines(), str(u1).splitlines())), "upat":(upat_loc, printable(upat_loc))}
+           "diff":list(difflib.unified_diff(str(u0).splitlines(),str(u1).splitlines())), "upat":(upat_loc, match_repr)}
     if not ctx.bottom_up: next_sink = new_sink
 
 # encoder helpers