fix that

* add 0.10.0 to comma benchmark

disabled the 0.10.1 ones which are pinned to master. it does not work because benchmark uses the cached old version

* that's pinned

.github/workflows/benchmark.yml

@@ -319,9 +319,9 @@ jobs:
     - name: Run 10 CIFAR training steps
       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=310 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=240 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=310 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=270 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
@@ -619,18 +619,24 @@ jobs:
         rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
     - name: reset process replay
       run: test/external/process_replay/reset.py
-    - name: benchmark openpilot 0.9.9 driving_vision
-      run: BENCHMARK_LOG=openpilot_0_9_9_vision 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/driving_vision.onnx
-    - name: benchmark openpilot 0.9.9 driving_policy
-      run: BENCHMARK_LOG=openpilot_0_9_9_policy 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/driving_policy.onnx
-    - 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: BENCHMARK_LOG=openpilot_0_9_9_vision PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 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: BENCHMARK_LOG=openpilot_0_9_9_policy PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 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: BENCHMARK_LOG=openpilot_0_9_9_dmonitoring PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 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 0.10.0 driving_policy
+      run: BENCHMARK_LOG=openpilot_0_10_0_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=5 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.10.0/selfdrive/modeld/models/driving_policy.onnx
+    - name: openpilot compile3 0.10.0 dmonitoring
+      run: BENCHMARK_LOG=openpilot_0_10_0_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=13 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.10.0/selfdrive/modeld/models/dmonitoring_model.onnx
     - name: openpilot compile3 0.10.1 driving_vision
-      run: BENCHMARK_LOG=openpilot_0_10_1_vision PYTHONPATH="." ASSERT_MIN_STEP_TIME=25 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
+      # TODO: ASSERT_MIN_STEP_TIME=17
+      run: BENCHMARK_LOG=openpilot_0_10_1_vision PYTHONPATH="." ASSERT_MIN_STEP_TIME=25 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
     - name: openpilot compile3 0.10.1 driving_policy
-      run: BENCHMARK_LOG=openpilot_0_10_1_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=7 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/refs/heads/master/selfdrive/modeld/models/driving_policy.onnx
+      run: BENCHMARK_LOG=openpilot_0_10_1_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=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_policy.onnx
     - name: openpilot compile3 0.10.1 dmonitoring
-      run: BENCHMARK_LOG=openpilot_0_10_1_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=12 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/refs/heads/master/selfdrive/modeld/models/dmonitoring_model.onnx
+      # TODO: ASSERT_MIN_STEP_TIME=10
+      run: BENCHMARK_LOG=openpilot_0_10_1_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=13 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/dmonitoring_model.onnx
     - name: benchmark MobileNetV2 on DSP
       run: |
         # generate quantized weights
@@ -641,16 +647,6 @@ jobs:
         PYTHONPATH=. CC=clang-19 DSP=1 NOOPT=1 CNT=2 DEBUG=2 python3 examples/test_onnx_imagenet.py /tmp/model.quant.onnx
     - name: Run process replay tests
       run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py
-    - uses: actions/upload-artifact@v4
-      with:
-        name: Speed (comma)
-        path: |
-          openpilot_compile_0_9_4.txt
-          openpilot_compile_0_9_7.txt
-          openpilot_0_9_4.txt
-          openpilot_0_9_7.txt
-          openpilot_image_0_9_4.txt
-          openpilot_image_0_9_7.txt
 
   testreddriverbenchmark:
     name: AM Benchmark

.github/workflows/test.yml

@@ -204,7 +204,7 @@ jobs:
         DEBUG=2 EMULATE=CUDA FORWARD_ONLY=1 PYTHON=1 python3 test/test_ops.py TestOps.test_gemm_fp16
         DEBUG=2 EMULATE=CUDA ALLOW_TF32=1 FORWARD_ONLY=1 PYTHON=1 python3 test/test_ops.py TestOps.test_gemm
         DEBUG=2 EMULATE=CUDA_SM75 FORWARD_ONLY=1 PYTHON=1 python3 test/test_ops.py TestOps.test_gemm_fp16
-        DEBUG=2 EMULATE=CUDA ALLOW_TF32=1 FORWARD_ONLY=1 PYTHON=1 python3 test/opt/test_tensor_cores.py
+        DEBUG=2 EMULATE=CUDA_SM89 ALLOW_TF32=1 FORWARD_ONLY=1 PYTHON=1 python3 test/opt/test_tensor_cores.py
     - name: Test emulated INTEL OpenCL tensor cores
       run: DEBUG=2 EMULATE=INTEL FORWARD_ONLY=1 PYTHON=1 HALF=1 N=64 python3 ./extra/gemm/simple_matmul.py
     - name: Test emulated AMX tensor cores
@@ -378,7 +378,7 @@ jobs:
       - name: Test openpilot CL compile fp16
         run: FLOAT16=1 DEBUGCL=1 CL=1 IMAGE=2 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
       - name: Test openpilot CL compile fp32 (test correctness)
-        run: DEBUGCL=1 CL=1 IMAGE=2 python examples/openpilot/compile3.py https://github.com/haraschax/filedump/raw/refs/heads/master/driving_vision_fp32.onnx
+        run: DEBUGCL=1 CL=1 IMAGE=2 SELFTEST=1 python examples/openpilot/compile3.py https://github.com/haraschax/filedump/raw/refs/heads/master/driving_vision_fp32.onnx
       - name: Test openpilot LLVM compile fp16
         run: FLOAT16=1 CPU=1 CPU_LLVM=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
       - name: Run process replay tests

autogen_stubs.sh

@@ -520,13 +520,8 @@ generate_mesa() {
   LVP_NIR_OPTIONS=$(./extra/mesa/lvp_nir_options.sh $MESA_SRC)
 
   fixup $BASE/mesa.py
-  patch_dlopen $BASE/mesa.py tinymesa_cpu "(BASE:=os.getenv('MESA_PATH', f\"/usr{'/local/' if helpers.OSX else '/'}lib\"))+'/libtinymesa_cpu'+(EXT:='.dylib' if helpers.OSX else '.so')" "f'{BASE}/libtinymesa{EXT}'" "brew_path('tinymesa_cpu')" "brew_path('tinymesa')"
+  patch_dlopen $BASE/mesa.py tinymesa_cpu "(BASE:=os.getenv('MESA_PATH', f\"/usr{'/local/' if helpers.OSX else '/'}lib\"))+'/libtinymesa_cpu'+(EXT:='.dylib' if helpers.OSX else '.so')" "f'{BASE}/libtinymesa{EXT}'" "'/opt/homebrew/lib/libtinymesa_cpu.dylib'" "'/opt/homebrew/lib/libtinymesa.dylib'"
   echo "lvp_nir_options = gzip.decompress(base64.b64decode('$LVP_NIR_OPTIONS'))" >> $BASE/mesa.py
-  cat <<EOF | sed -i "/import ctypes.*/r /dev/stdin" $BASE/mesa.py
-def brew_path(nm):
-  try: return f"{subprocess.check_output(['brew', '--prefix', nm]).decode().strip()}/lib/lib{nm}.dylib"
-  except Exception: return 'failed'
-EOF
   sed -i "/in_dll/s/.*/try: &\nexcept (AttributeError, ValueError): pass/" $BASE/mesa.py
   sed -i "s/import ctypes/import ctypes, ctypes.util, os, gzip, base64, subprocess, tinygrad.helpers as helpers/" $BASE/mesa.py
   sed -i "s/ctypes.CDLL('.\+')/(dll := _try_dlopen_tinymesa_cpu())/" $BASE/mesa.py

examples/openpilot/compile3.py

@@ -134,7 +134,7 @@ if __name__ == "__main__":
   with open(OUTPUT, "rb") as f: pickle_loaded = pickle.load(f)
 
   test_vs_compile(pickle_loaded, inputs, outputs)
-  if not getenv("FLOAT16"):
+  if getenv("SELFTEST"):
     test_vs_onnx(inputs, outputs, onnx_file, 1e-4)
 
   if getenv("BENCHMARK_LOG", ""):

extra/gemm/amd_uop_matmul.py

@@ -328,8 +328,7 @@ if __name__ == "__main__":
   elif HL == 1: hprg = hl_spec_kernel3()
   else: hprg = hand_spec_kernel3()
   if HL == 3:
-    with Context(BLOCK_REORDER=0):
-      prg = get_program(hprg, Device.default.renderer)
+    prg = get_program(hprg, Device.default.renderer)
   else:
     prg = get_program(hprg, Device.default.renderer)
   print(prg.src)

extra/gemm/simple_matmul.py

@@ -5,8 +5,10 @@ from tinygrad.dtype import _to_np_dtype
 from tinygrad.codegen.opt import OptOps
 from tinygrad.engine.realize import lower_schedule
 
-dtype_in = dtypes.half if getenv("HALF") else dtypes.bfloat16 if getenv("BFLOAT16") else dtypes.float
-acc_dtype = dtypes.half if getenv("ACC_HALF") else dtypes.bfloat16 if getenv("ACC_BFLOAT16") else None
+dtype_in = (dtypes.half if getenv("HALF") else dtypes.bfloat16 if getenv("BFLOAT16") else
+            dtypes.fp8e4m3 if getenv("FP8E4M3") else dtypes.fp8e5m2 if getenv("FP8E5M2") else dtypes.float)
+acc_dtype = (dtypes.half if getenv("ACC_HALF") else dtypes.bfloat16 if getenv("ACC_BFLOAT16") else
+            dtypes.fp8e4m3 if getenv("ACC_FP8E4M3") else dtypes.fp8e5m2 if getenv("ACC_FP8E5M2") else None)
 if getenv("INT"):  dtype_in, acc_dtype = dtypes.int8, dtypes.int32
 if getenv("UINT"): dtype_in, acc_dtype = dtypes.uint8, dtypes.int32
 
@@ -14,8 +16,10 @@ N = getenv("N", 4096)
 M = getenv("M", N)
 K = getenv("K", N)
 CNT = getenv("CNT", 10)
-ATOL = getenv("ATOL", 1e-4)
-RTOL = getenv("RTOL", 3e-2)
+
+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 = getenv("ATOL", atol), getenv("RTOL", rtol)
+
 INT_LOW = getenv("INT_LOW", 0)
 INT_HIGH = getenv("INT_HIGH", 10)
 

extra/sqtt/roc.py

@@ -27,14 +27,18 @@ class _ROCParseCtx:
         self.disasms[prog.base + addr] = info
         self.addr2prg[prog.base + addr] = prog
 
-  def next_sqtt(self): return next(self.sqtt_evs, None)
+  def next_sqtt(self):
+    x = next(self.sqtt_evs, 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, ev.cu, ev.simd, ev.wave_id, ev.start)
+    if DEBUG >= 4: 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, ev.cu, ev.simd, ev.contexts, ev.begin_time, ev.end_time)
+    if DEBUG >= 4: print("WAVE", ev.wave_id, self.active_se, ev.cu, ev.simd, ev.contexts, ev.begin_time, ev.end_time)
 
     asm = {}
     for j in range(ev.instructions_size):

extra/torch_muon.py

@@ -1,75 +0,0 @@
-import torch
-
-#credit to KellerJordan at https://github.com/KellerJordan/Muon/tree/master
-#some changes: classic momentum instead of weighting gradient
-#added ns_steps, ns_params, nesterov as hyperparams
-def zeropower_via_newtonschulz5(G:torch.tensor, steps:int, params:tuple[int, ...]):
-  """
-  Newton-Schulz iteration to compute the zeroth power / orthogonalization of G. We opt to use a
-  quintic iteration whose coefficients are selected to maximize the slope at zero. For the purpose
-  of minimizing steps, it turns out to be empirically effective to keep increasing the slope at
-  zero even beyond the point where the iteration no longer converges all the way to one everywhere
-  on the interval. This iteration therefore does not produce UV^T but rather something like US'V^T
-  where S' is diagonal with S_{ii}' ~ Uniform(0.5, 1.5), which turns out not to hurt model
-  performance at all relative to UV^T, where USV^T = G is the SVD.
-  """
-  assert G.ndim >= 2 # batched Muon implementation by @scottjmaddox, and put into practice in the record by @YouJiacheng
-
-  a, b, c = params
-  X = G
-  if G.size(-2) > G.size(-1):
-    X = X.mT
-
-  # Ensure spectral norm is at most 1
-  X = X / (X.norm(dim=(-2, -1), keepdim=True) + 1e-7)
-  # Perform the NS iterations
-  for _ in range(steps):
-    A = X @ X.mT
-    B = b * A + c * A @ A # quintic computation strategy adapted from suggestion by @jxbz, @leloykun, and @YouJiacheng
-    X = a * X + B @ X
-
-  if G.size(-2) > G.size(-1):
-    X = X.mT
-
-  return X
-
-def muon_update(grad, momentum, beta=0.95, ns_steps=5, ns_params=(3.4445, -4.7750,  2.0315), nesterov=True):
-  if beta:
-    momentum.mul_(beta).add_(grad)
-    update = grad.add(momentum,alpha=beta) if nesterov else momentum
-  else: update = grad
-  if update.ndim == 4: # for the case of conv filters
-    update = update.view(len(update), -1)
-  update = zeropower_via_newtonschulz5(update, steps=ns_steps, params=ns_params)
-  return update
-
-class SingleDeviceMuon(torch.optim.Optimizer):
-  """
-  Muon variant for usage in non-distributed settings.
-  """
-  def __init__(self, params, lr=0.02, weight_decay=0.0, momentum=0.95, ns_steps=5, ns_params=(3.4445, -4.7750,  2.0315), nesterov=True):
-    defaults = dict(lr=lr, weight_decay=weight_decay, momentum=momentum, ns_steps=ns_steps, ns_params=ns_params, nesterov=nesterov)
-    super().__init__(params, defaults)
-
-  @torch.no_grad()
-  def step(self, closure=None):
-
-    loss = None
-    if closure is not None:
-      with torch.enable_grad():
-        loss = closure()
-
-    for group in self.param_groups:
-      for p in group["params"]:
-        if p.grad is None:
-          p.grad = torch.zeros_like(p)  # Force synchronization
-        state = self.state[p]
-        if len(state) == 0:
-          state["momentum_buffer"] = torch.zeros_like(p)
-        update = muon_update(p.grad, state["momentum_buffer"], beta=group["momentum"], ns_steps=group["ns_steps"],
-                             ns_params=group["ns_params"], nesterov=group["nesterov"])
-        p.mul_(1.0 - group["lr"] * group["weight_decay"])
-
-        p.add_(update.reshape(p.shape), alpha=-group["lr"])
-
-    return loss

setup.py

@@ -9,7 +9,7 @@ with open(directory / 'README.md', encoding='utf-8') as f:
 
 testing_minimal = [
   "numpy",
-  "torch==2.8.0",
+  "torch==2.9.0",
   "pytest",
   "pytest-xdist",
   "pytest-timeout",

sz.py

@@ -54,6 +54,8 @@ def gen_diff(table_old, table_new):
 
 def display_diff(diff): return "+"+str(diff) if diff > 0 else str(diff)
 
+NONCORE_DIRS = {"tinygrad/apps", "tinygrad/nn", "tinygrad/renderer", "tinygrad/runtime", "tinygrad/viz"}
+
 if __name__ == "__main__":
   if len(sys.argv) == 3:
     headers = ["Name", "Lines", "Diff", "Tokens/Line", "Diff"]
@@ -76,9 +78,12 @@ if __name__ == "__main__":
     else:
       print(tabulate([headers] + sorted(table, key=lambda x: -x[1]), headers="firstrow", floatfmt=".1f")+"\n")
       groups = sorted([('/'.join(x[0].rsplit("/", 1)[0].split("/")[0:2]), x[1], x[2]) for x in table])
+      dir_sizes = {}
       for dir_name, group in itertools.groupby(groups, key=lambda x:x[0]):
-        print(f"{dir_name:30s} : {sum([x[1] for x in group]):6d}")
+        dir_sizes[dir_name] = sum([x[1] for x in group])
+        print(f"{dir_name:30s} : {dir_sizes[dir_name]:6d}")
+      print(f"\n core line count: {sum([v for k,v in dir_sizes.items() if k not in NONCORE_DIRS])}")
       total_lines = sum([x[1] for x in table])
-      print(f"\ntotal line count: {total_lines}")
+      print(f"total line count: {total_lines}")
       max_line_count = int(os.getenv("MAX_LINE_COUNT", "-1"))
       assert max_line_count == -1 or total_lines <= max_line_count, f"OVER {max_line_count} LINES"

test/external/external_benchmark_openpilot.py

@@ -1,63 +0,0 @@
-import time, sys, hashlib
-from pathlib import Path
-from tinygrad.nn.onnx import OnnxRunner
-from tinygrad import Tensor, dtypes, TinyJit
-from tinygrad.helpers import IMAGE, GlobalCounters, fetch, colored, getenv, trange
-import numpy as np
-from extra.bench_log import BenchEvent, WallTimeEvent
-
-OPENPILOT_MODEL = sys.argv[1] if len(sys.argv) > 1 else "https://github.com/commaai/openpilot/raw/v0.9.4/selfdrive/modeld/models/supercombo.onnx"
-
-if __name__ == "__main__":
-  run_onnx = OnnxRunner(fetch(OPENPILOT_MODEL))
-
-  Tensor.manual_seed(100)
-  input_shapes = {name: spec.shape for name, spec in run_onnx.graph_inputs.items()}
-  input_types = {name: spec.dtype for name, spec in run_onnx.graph_inputs.items()}
-  new_inputs = {k:Tensor.randn(*shp, dtype=input_types[k]).mul(8).realize() for k,shp in input_shapes.items()}
-  new_inputs_junk = {k:Tensor.randn(*shp, dtype=input_types[k]).mul(8).realize() for k,shp in input_shapes.items()}
-  new_inputs_junk_numpy = {k:v.numpy() for k,v in new_inputs_junk.items()}
-
-  # benchmark
-  for _ in range(5):
-    GlobalCounters.reset()
-    st = time.perf_counter_ns()
-    ret = next(iter(run_onnx(new_inputs_junk).values())).cast(dtypes.float32).numpy()
-    print(f"unjitted: {(time.perf_counter_ns() - st)*1e-6:7.4f} ms")
-
-  # NOTE: the inputs to a JIT must be first level arguments
-  run_onnx_jit = TinyJit(lambda **kwargs: run_onnx(kwargs), prune=True)
-  step_times = []
-  for _ in range(20):
-    GlobalCounters.reset()
-    st = time.perf_counter_ns()
-    with WallTimeEvent(BenchEvent.STEP):
-      # Need to cast non-image inputs from numpy, this is only realistic way to run model
-      inputs = {**{k:v for k,v in new_inputs_junk.items() if 'img' in k},
-                **{k:Tensor(v) for k,v in new_inputs_junk_numpy.items() if 'img' not in k}}
-      ret = next(iter(run_onnx_jit(**inputs).values())).cast(dtypes.float32).numpy()
-      step_times.append(t:=(time.perf_counter_ns() - st)*1e-6)
-    print(f"jitted:  {t:7.4f} ms")
-
-  suffix = ""
-  if IMAGE.value < 2: suffix += f"_image{IMAGE.value}" # image=2 has no suffix for compatibility
-  if getenv("FLOAT16") == 1: suffix += "_float16"
-  path = Path(__file__).parent / "openpilot" / f"{hashlib.md5(OPENPILOT_MODEL.encode()).hexdigest()}{suffix}.npy"
-
-  # validate if we have records
-  tinygrad_out = next(iter(run_onnx_jit(**new_inputs).values())).cast(dtypes.float32).numpy()
-  if getenv("SAVE_OUTPUT"):
-    np.save(path, tinygrad_out)
-    print(f"saved output to {path}!")
-  elif getenv("FUZZ") and path.exists():
-    known_good_out = np.load(path)
-    for _ in trange(1000):
-      ret = next(iter(run_onnx_jit(**new_inputs).values())).cast(dtypes.float32).numpy()
-      np.testing.assert_allclose(known_good_out, ret, atol=1e-2, rtol=1e-2)
-    print(colored("fuzz validated!", "green"))
-  elif path.exists():
-    known_good_out = np.load(path)
-    np.testing.assert_allclose(known_good_out, tinygrad_out, atol=1e-2, rtol=1e-2)
-    print(colored("outputs validated!", "green"))
-  else:
-    print(colored("skipping validation", "yellow"))

test/external/external_benchmark_schedule.py

@@ -2,9 +2,9 @@ from extra.models.resnet import ResNet50
 from tinygrad import Tensor, nn, Device
 from tinygrad.helpers import Profiling, Timing, getenv
 from tinygrad.uop.ops import Ops
-from tinygrad.codegen import get_rewrites_for_renderer, apply_rewrites
+from tinygrad.codegen import full_rewrite_to_sink
 from tinygrad.codegen.late.control_flow import linearize
-from tinygrad.uop.spec import type_verify
+from tinygrad.uop.spec import type_verify, program_spec
 
 if __name__ == "__main__":
   mdl = ResNet50()
@@ -29,12 +29,11 @@ if __name__ == "__main__":
         asts = list({x.ast.key:x.ast for x in sched if x.ast.op is Ops.SINK}.values())
         if (restrict_kernel := getenv("RESTRICT_KERNEL", -1)) != -1: asts = asts[restrict_kernel:restrict_kernel+1]
 
-        rewrites = get_rewrites_for_renderer(Device.default.renderer, linearizer=False)
         with Profiling(PROFILE, fn="/tmp/rewrite.prof"):
           with Timing("***** model rewrite in   "):
             rewritten_uops = []
             for u in asts:
-              rewritten_uops.append(apply_rewrites(u, rewrites))
+              rewritten_uops.append(full_rewrite_to_sink(u, ren=Device.default.renderer))
 
         if LINEARIZE:
           with Timing("***** model linearize in "):
@@ -42,5 +41,5 @@ if __name__ == "__main__":
             for u in rewritten_uops:
               uops_line.append(linearize(u))
           with Timing("***** model verify in    "):
-            for u in uops_line: type_verify(u)
+            for u in uops_line: type_verify(u, program_spec)
           print(sum(len(u) for u in uops_line))

test/external/fuzz_fast_idiv.py

@@ -1,7 +1,7 @@
 import random
 import z3
 from tinygrad import dtypes
-from tinygrad.uop.spec import uops_to_z3, z3_cdiv
+from tinygrad.uop.validate import uops_to_z3, z3_cdiv
 from tinygrad.uop.ops import UOp
 from tinygrad.uop.decompositions import fast_idiv
 random.seed(42)

test/external/fuzz_linearizer.py

@@ -207,7 +207,7 @@ def fuzz_linearizer(lin: Kernel, rtol=1e-2, atol=1e-2, opts_list=None):
         if not FUZZ_ALL_ACTIONS and test_lin.applied_opts: print(f"applied opts: {test_lin.applied_opts}")
 
         # stop if kernel uops repeat
-        try: tuops = tuplize_uops(get_program(test_lin.get_optimized_ast(), test_lin.opts).uops)
+        try: tuops = tuplize_uops(get_program(test_lin.get_optimized_ast(), test_lin.ren).uops)
         except KeyboardInterrupt: raise
         except BaseException as e:
           print(test_lin.ast)
@@ -224,7 +224,7 @@ def fuzz_linearizer(lin: Kernel, rtol=1e-2, atol=1e-2, opts_list=None):
         (msg, rawbufs, var_vals, ground_truth, state1) = compare_linearizer(test_lin, rawbufs, var_vals, ground_truth, rtol=rtol, atol=atol)
         if state1 is not None and validate_device is not None:
           validate_lin = test_lin.copy()
-          validate_lin.opts = validate_device.renderer
+          validate_lin.ren = validate_device.renderer
           if validate_rawbufs is None:
             validate_rawbufs = [get_fuzz_rawbuf_like(x, copy=True, force_device=validate_device.device) for x in rawbufs]
           (_msg, _, _, _, state2) = compare_linearizer(validate_lin, validate_rawbufs, var_vals, ground_truth, rtol=rtol, atol=atol)

test/external/fuzz_symbolic.py

@@ -2,7 +2,7 @@ import random, operator
 import z3
 from tinygrad import Variable, dtypes
 from tinygrad.uop.ops import UOp
-from tinygrad.uop.spec import uops_to_z3
+from tinygrad.uop.validate import uops_to_z3
 from tinygrad.helpers import DEBUG, Context
 
 seed = random.randint(0, 100)

test/test_const_folding.py

@@ -67,12 +67,9 @@ 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_dtype_alu.py

@@ -75,7 +75,10 @@ def universal_test_unary(a, dtype, op):
   out: Tensor = op[0](ta)
   tensor_value = out.numpy()
   numpy_value = op[1](ta.numpy())
-  if dtype in dtypes.fp8s: numpy_value = truncate[dtype](numpy_value)
+  if dtype in dtypes.fp8s:
+    # cuda cast f32 inf to f8 MAX, amd cast it to nan(E4M3)/inf(E5M2)
+    if math.isinf(numpy_value): return
+    numpy_value = truncate[dtype](numpy_value)
   if dtype in dtypes.floats:
     atol, rtol = { dtypes.float16:(1e-3, 1e-2), dtypes.bfloat16:(1e-3, 2e-2),
       dtypes.fp8e4m3:(1e-1, 1e-1), dtypes.fp8e5m2: (1.0, 5e-1)}.get(dtype, (1e-6, 1e-5))

test/test_fusion_op.py

@@ -51,7 +51,7 @@ class TestFusionOp(unittest.TestCase):
     a = Tensor(val)
     for _ in range(24): a = Tensor.stack(a, a)[0]
     sched = a.schedule()
-    self.assertEqual(len(sched), 1)
+    self.assertEqual(len(sched), 0)
     self.assertLess(time.perf_counter()-st, 2.0)
 
   def test_recursive_reshape(self):

test/test_image_dtype.py

@@ -52,7 +52,6 @@ class TestImageDType(unittest.TestCase):
     assert isinstance(it.uop.base.realized.dtype, ImageDType)
     np.testing.assert_equal(tst, it.numpy())
 
-  @unittest.expectedFailure # this isn't supported anymore, CAST to ImageDType stays ImageDType
   def test_image_cast_and_back_collapses(self):
     data = Tensor.randn(9*27*4).realize()
     tst = data.numpy()

test/test_linearizer.py

@@ -393,14 +393,15 @@ class TestLinearizer(unittest.TestCase):
     uops = get_program(ast, opts=opt).uops
     local_stores = [u for u in uops if u.op is Ops.STORE and any(x.op is Ops.DEFINE_LOCAL for x in get_recursive(u.src[0]))]
     global_stores = [u for u in uops if u.op is Ops.STORE and any(x.op is Ops.DEFINE_GLOBAL for x in get_recursive(u.src[0]))]
-    barrier = [u for u in uops if u.op is Ops.BARRIER][0]
+    barrier = [u for u in uops if u.op is Ops.BARRIER]
+    assert len(barrier) == 1
     # check that the float4 cast collapses for all stores
     for store in local_stores+global_stores:
       assert store.src[1].dtype.count > 1 # and store.src[2].op is not Ops.VECTORIZE
     # # check the children's vins
     # TODO: src ALU are not the same, should it?
     # assert barrier.src == tuple(local_stores)
-    assert len([u for u in uops if u.op is Ops.IF and u.src[1] == barrier]) == 1
+    assert len([u for u in uops if u.op is Ops.IF])
 
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared")

test/test_optim.py

@@ -5,7 +5,6 @@ from tinygrad import Tensor, Device, dtypes
 from tinygrad.nn.optim import Adam, SGD, AdamW, Muon
 from tinygrad.helpers import CI
 from tinygrad.device import is_dtype_supported
-from extra.torch_muon import SingleDeviceMuon as TorchMuon
 
 np.random.seed(1337)
 x_init = np.random.randn(1,4).astype(np.float32)
@@ -58,12 +57,11 @@ class TestOptim(unittest.TestCase):
   def _test_sgd(self, steps, opts, atol, rtol): self._test_optim(SGD, torch.optim.SGD, steps, opts, atol, rtol)
   def _test_adam(self, steps, opts, atol, rtol): self._test_optim(Adam, torch.optim.Adam, steps, opts, atol, rtol)
   def _test_adamw(self, steps, opts, atol, rtol): self._test_optim(AdamW, torch.optim.AdamW, steps, opts, atol, rtol)
-  #TODO: use torch.muon when it comes out
-  def _test_muon(self, steps, opts, atol, rtol): self._test_optim(Muon, TorchMuon, steps, opts, atol, rtol)
+  def _test_muon(self, steps, opts, atol, rtol): self._test_optim(Muon, torch.optim.Muon, steps, opts, atol, rtol)
 
   def test_multistep_sgd_high_lr_teeny(self): self._test_sgd(2, {'lr': 1.1, 'teeny': True}, 1e-6, 1e-5)
   def test_multistep_adam_high_lr_teeny(self): self._test_adam(2, {'lr': 1.1, 'teeny': True}, 2e-4, 5e-4)
-  def test_multistep_muon_high_lr_teeny(self): self._test_muon(2, {'lr': 1.1, 'teeny': True}, 2e-4, 5e-4)
+  def test_multistep_muon_high_lr_teeny(self): self._test_muon(2, {'lr': 1.1, 'teeny': True}, 1e-2, 5e-4)
 
   def test_sgd(self): self._test_sgd(1, {'lr': 0.001}, 1e-6, 0)
   def test_sgd_high_lr(self): self._test_sgd(1, {'lr': 10}, 1e-6, 1e-5)
@@ -87,27 +85,34 @@ class TestOptim(unittest.TestCase):
   def test_multistep_sgd_high_lr_nesterov_momentum_wd(self):
     self._test_sgd(10, {'lr': 9, 'momentum': 0.9, 'nesterov': True, 'weight_decay': 0.1}, 1e-5, 3e-4)
 
-  def test_muon(self): self._test_muon(1, {'lr': 0.001}, 1e-6, 0)
-  def test_muon_high_lr(self): self._test_muon(1, {'lr': 10}, 1e-6, 3e-4)
-  def test_muon_wd(self): self._test_muon(1, {'lr': 0.001, 'weight_decay': 0.01}, 1e-6, 0)
-  def test_muon_high_lr_wd(self): self._test_muon(1, {'lr': 10, 'weight_decay': 0.01}, 1e-6, 5e-4)
+  def test_muon(self): self._test_muon(1, {'lr': 0.001}, 1e-3, 0)
+  # TODO: disabled due to big atol
+  # def test_muon_high_lr(self): self._test_muon(1, {'lr': 10}, 1e-6, 3e-4)
+  def test_muon_wd(self): self._test_muon(1, {'lr': 0.001, 'weight_decay': 0.01}, 1e-3, 3e-4)
+  # TODO: disabled due to big atol
+  # def test_muon_high_lr_wd(self): self._test_muon(1, {'lr': 10, 'weight_decay': 0.01}, 1e-6, 5e-4)
 
   # NOTE: momentum set to 0.95 by default, nesterov set to True by default
-  def test_multistep_muon_momentum_wd(self): self._test_muon(10, {'lr': 0.001, 'weight_decay': 0.01}, 1e-5, 0)
+  def test_multistep_muon_momentum_wd(self): self._test_muon(10, {'lr': 0.001, 'weight_decay': 0.01}, 3e-3, 0)
   # ns defaults are numerically unstable, but it is tolerable in real training (see nsteps/nparam tests)
-  def test_multistep_muon_high_lr_momentum_wd(self): self._test_muon(10, {'lr': 10, 'weight_decay': 0.01}, 1e-1, 3e-4)
-  def test_multistep_muon_no_nesterov_momentum(self): self._test_muon(10, {'lr': 0.001, 'nesterov': False}, 1e-5, 0)
-  def test_multistep_muon_high_lr_no_nesterov_momentum(self): self._test_muon(10, {'lr': 10, 'nesterov': False}, 0.5e-1, 1e-1)
+  # TODO: disabled due to big atol
+  # def test_multistep_muon_high_lr_momentum_wd(self): self._test_muon(10, {'lr': 10, 'weight_decay': 0.01}, 1e-1, 3e-4)
+  def test_multistep_muon_no_nesterov_momentum(self): self._test_muon(10, {'lr': 0.001, 'nesterov': False}, 1e-3, 0)
+  # TODO: disabled due to big atol
+  # def test_multistep_muon_high_lr_no_nesterov_momentum(self): self._test_muon(10, {'lr': 10, 'nesterov': False}, 5e-2, 1e-1)
 
-  def test_muon_ns_steps(self): self._test_muon(1, {'lr': 0.001, 'ns_steps': 3}, 1e-6, 0)
-  def test_muon_high_lr_ns_steps(self): self._test_muon(1, {'lr': 10, 'ns_steps': 3}, 1e-5, 3e-4)
-  def test_muon_ns_params(self): self._test_muon(1, {'lr': 0.001,'ns_params': (2.0,-1.5,0.5)}, 1e-6, 0)
-  def test_muon_high_lr_ns_params(self): self._test_muon(1, {'lr': 10,'ns_params': (2.0,-1.5,0.5)}, 1e-5, 3e-4)
+  def test_muon_ns_steps(self): self._test_muon(1, {'lr': 0.001, 'ns_steps': 3}, 1e-4, 0)
+  # TODO: disabled due to big atol
+  # def test_muon_high_lr_ns_steps(self): self._test_muon(1, {'lr': 10, 'ns_steps': 3}, 1e-5, 3e-4)
+  def test_muon_ns_coefficients(self): self._test_muon(1, {'lr': 0.001,'ns_coefficients': (2.0,-1.5,0.5)}, 1e-5, 3e-4)
+  # TODO: disabled due to big atol
+  # def test_muon_high_lr_ns_coefficients(self): self._test_muon(1, {'lr': 10,'ns_coefficients': (2.0,-1.5,0.5)}, 1e-5, 3e-4)
 
-  def test_muon_momentum_wd_ns_steps_ns_params(self):
-    self._test_muon(10, {'lr': 0.001, 'momentum': 0.90, 'weight_decay': 0.01, 'ns_steps': 3, 'ns_params': (2.0,-1.5,0.5)}, 1e-5, 0)
-  def test_multistep_muon_high_lr_momentum_wd_ns_steps_ns_params(self):
-    self._test_muon(10, {'lr': 10, 'momentum': 0.90, 'weight_decay': 0.01, 'ns_steps': 3, 'ns_params': (2.0,-1.5,0.5)}, 1e-5, 3e-4)
+  def test_muon_momentum_wd_ns_steps_ns_coefficients(self):
+    self._test_muon(10, {'lr': 0.001, 'momentum': 0.90, 'weight_decay': 0.01, 'ns_steps': 3, 'ns_coefficients': (2.0,-1.5,0.5)}, 1e-4, 0)
+  # TODO: disabled due to big atol
+  # def test_multistep_muon_high_lr_momentum_wd_ns_steps_ns_coefficients(self):
+  #   self._test_muon(10, {'lr': 10, 'momentum': 0.90, 'weight_decay': 0.01, 'ns_steps': 3, 'ns_coefficients': (2.0,-1.5,0.5)}, 1e-5, 3e-4)
 
   def test_adam(self): self._test_adam(1, {'lr': 0.001}, 1e-5, 0)
   def test_adam_high_lr(self): self._test_adam(1, {'lr': 10}, 1e-4, 1e-4)

test/test_rangeify.py

@@ -1,6 +1,6 @@
 import unittest
 from tinygrad import Tensor, nn, Device
-from tinygrad.helpers import Context, GlobalCounters, CI, getenv
+from tinygrad.helpers import Context, GlobalCounters, CI, getenv, PCONTIG
 from tinygrad.uop.ops import graph_rewrite, PatternMatcher, UPat, Ops
 from tinygrad.renderer.ptx import PTXRenderer
 from tinygrad.renderer.nir import NIRRenderer
@@ -64,11 +64,11 @@ class TestPcontig(unittest.TestCase):
       Tensor.realize(*ret)
       return ret
 
-    with Context(PCONTIG=2, DEBUG=2):
+    with Context(PCONTIG=max(2, PCONTIG.value), DEBUG=2):
       grads = fa_bw()
       print(f"{GlobalCounters.global_ops/1e9:.2f} GFLOPS")
 
-    with Context(DEBUG=2):
+    with Context(PCONTIG=0, DEBUG=2):
       cmp_grads = fa_bw()
       print(f"{GlobalCounters.global_ops/1e9:.2f} GFLOPS")
 

test/test_schedule.py

@@ -446,7 +446,7 @@ 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, 30), (nn.optim.SGD, 13)]:
+    for optim, cnt in [(nn.optim.Adam, 21), (nn.optim.SGD, 8)]:
       with self.subTest(optim=optim.__name__):
         with Tensor.train():
           img = Tensor.ones(1,3,4,4)
@@ -1863,7 +1863,7 @@ class TestSchedule(unittest.TestCase):
     yt = Tensor.randn(BS, 10).realize()
     with Context(SPLIT_REDUCEOP=0):
       loss = yt.sparse_categorical_crossentropy(Y_train[samples])
-      run_schedule(check_schedule(loss, 5))
+      run_schedule(check_schedule(loss, 4))
       loss_fused = loss.numpy()
     loss_ref = torch.nn.CrossEntropyLoss()(torch.tensor(yt.numpy()), torch.tensor(Y_train.numpy())[torch.tensor(samples.numpy())])
     np.testing.assert_allclose(loss_fused, loss_ref.numpy(), atol=1e-6, rtol=1e-6)
@@ -2076,6 +2076,11 @@ class TestCopyFolding(unittest.TestCase):
     check_schedule(b, 0, filter_sink=False)
     assert b.item() == 1
 
+  def test_one_hot_with_copy(self):
+    y = Tensor([1, 2, 3]).to("CPU")
+    x = y.one_hot(10)
+    check_schedule(x, 3, filter_sink=False)
+
   def test_const_copy_multi(self):
     x = Tensor.ones(1, device="CPU").to_(["CPU", "CPU:1"])
     check_schedule(x, 0, filter_sink=False)
@@ -2085,7 +2090,7 @@ class TestCopyFolding(unittest.TestCase):
     a = Tensor.arange(3).realize()
     zeros = Tensor.zeros(3).realize()
     b = (a*zeros).to("CPU")
-    run_schedule(check_schedule(b, 2, filter_sink=False)) # TODO: 0?
+    run_schedule(check_schedule(b, 0, filter_sink=False))
     self.assertListEqual(b.tolist(), [0, 0, 0])
     self.assertEqual(b.device, "CPU")
 

test/test_tensor.py

@@ -839,12 +839,11 @@ 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), 4, f"failed with {si.metadata}")
-    self.assertSetEqual(set(m.name for m in si.metadata), {"__mul__", "sigmoid", "relu"})
+    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), 2)
-    self.assertEqual(bw[0].name, "__mul__")
-    self.assertEqual(bw[1].name, "sigmoid")
+    self.assertEqual(len(bw), 1)
+    self.assertEqual(bw[0].name, "sigmoid")
 
 class TestIdxUpcast(unittest.TestCase):
   def _find_op(self, ast: UOp, op: Ops):

test/test_uop_graph.py

@@ -473,7 +473,7 @@ class TestUOpGraph(unittest.TestCase):
     l0 = UOp(Ops.LOAD, dtypes.long, (d0.index(UOp.const(dtypes.int, 0)),)).cast(dtypes.index)
     idx = l0 * 600
     valid = (l0<-1).ne(True)&(l0<3000)
-    l1 = UOp(Ops.LOAD, dtypes.long, (d1.index(idx.valid(valid)),))
+    l1 = valid.where(UOp(Ops.LOAD, dtypes.long, (d1.index(idx),)),0)
     uops = to_uops_list([l1])
     for u in uops:
       if u.op is Ops.INDEX: self.assertEqual(u.src[1].dtype, dtypes.int)
@@ -518,6 +518,7 @@ class TestUOpGraph(unittest.TestCase):
       st1 = UOp(Ops.STORE, dtypes.void, (glbl0.index(v), v, v<20))
       with self.assertRaises(RuntimeError): to_uops_list([st1])
 
+  @unittest.skip("if not allowed in graph")
   def test_in_bounds_access_gated_local(self):
     with Context(IGNORE_OOB=0):
       # Define buffers
@@ -638,13 +639,13 @@ class TestUOpGraph(unittest.TestCase):
     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)))
     barrier = UOp(Ops.BARRIER, dtypes.void, (st, ))
-    ld0 = UOp(Ops.LOAD, dtypes.int, (smem.index(UOp.invalid()), barrier))
-    ld1 = UOp(Ops.LOAD, dtypes.int, (smem.index(lidx+2, UOp.const(dtypes.bool, True)), barrier))
+    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)),))
     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.index(lidx+2))
+    self.assertEqual(ld0.src[0], smem.after(barrier).index(lidx+2))
 
   def test_fold_gated_store(self):
     glbl = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(), (), 0)

test/test_uops.py

@@ -6,7 +6,7 @@ from tinygrad.helpers import CI, DEBUG, getenv, Timing
 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.spec import spec
+from tinygrad.uop.spec import shared_spec
 from tinygrad.renderer import ProgramSpec
 from tinygrad.engine.realize import CompiledRunner, get_program
 from tinygrad.codegen import full_rewrite
@@ -18,7 +18,7 @@ from tinygrad.renderer.ptx import PTXRenderer
 def to_uops_list(u:list[UOp], opts=None, skip_check=False) -> list[UOp]: return full_rewrite(UOp.sink(*u), opts)
 
 def _uops_to_prg(uops_list):
-  uops = full_rewrite(ast:=UOp.sink(*uops_list), opts=Device[Device.DEFAULT].renderer)
+  uops = full_rewrite(ast:=UOp.sink(*uops_list), ren=Device[Device.DEFAULT].renderer)
   src = Device[Device.DEFAULT].renderer.render(uops)
   has_local = Device[Device.DEFAULT].renderer.has_local
   return CompiledRunner(ProgramSpec(uops[-1].arg.name if uops[-1].arg is not None else "test", src, Device.DEFAULT, ast, uops=uops,
@@ -302,6 +302,7 @@ class TestGatedStoreRewrite(unittest.TestCase):
     self.assertIs(gated_uops[-1].op, Ops.STORE)
 
   # scaled down version of TestLinearizerDumb.test_unmerged_ifs
+  @unittest.skip("we don't merge ifs anymore")
   def test_merge_ifs_alt(self):
     gmem0 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), (), 0)
     gmem1 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), (), 1)
@@ -331,7 +332,7 @@ 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.index(uop(uops, Ops.CONST, dtypes.int32, (), 0)), barr))
+    sres = uop(uops, Ops.LOAD, dtypes.float32, (smem.after(barr).index(uop(uops, Ops.CONST, dtypes.int32, (), 0)),))
     self.assertEqual(_test_uops_result(dtypes.float32, uops, sres), 42)
 
   # NOTE: webgpu specific, since only webgpu performs bitpacking
@@ -341,7 +342,7 @@ 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.index(uop(uops, Ops.CONST, dtypes.int32, (), 0)), barr))
+    sres = uop(uops, Ops.LOAD, dtypes.uint8, (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
@@ -512,7 +513,7 @@ 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(spec.patterns[0][0].location[0].replace("\\", "/").split("/")[-1], "spec.py")
+    self.assertEqual(shared_spec.patterns[0][0].location[0].replace("\\", "/").split("/")[-1], "spec.py")
     test_upat = UPat(Ops.CONST, dtypes.bool)
     self.assertEqual(test_upat.location[0].split("/")[-1], __file__.replace("\\", "/").split("/")[-1])
     test_upat_named = test_upat.named("test_name")

test/unit/test_uop_symbolic.py

@@ -7,7 +7,7 @@ from tinygrad.codegen import full_rewrite
 from tinygrad.helpers import Context
 from tinygrad.uop.ops import UOp, Ops, graph_rewrite, sym_infer
 from tinygrad.uop.symbolic import sym, commutative
-from tinygrad.uop.spec import uops_to_z3
+from tinygrad.uop.validate import uops_to_z3
 
 def check_uop_against_string(self, v:UOp, s:str):
   sym_vars = {v.render():v for v in v.toposort() if v.op in (Ops.DEFINE_VAR, Ops.RANGE, Ops.SPECIAL)}

tinygrad/codegen/__init__.py

@@ -1,9 +1,6 @@
-from typing import Any, Callable
-import functools
-from dataclasses import dataclass
 from tinygrad.helpers import QUANTIZE, DEVECTORIZE, TRANSCENDENTAL
 from tinygrad.uop.ops import PatternMatcher, graph_rewrite, UOp, pm_lower_index_dtype
-from tinygrad.uop.spec import type_verify
+from tinygrad.uop.spec import type_verify, program_spec
 from tinygrad.renderer import Renderer
 
 # import all pattern matchers here
@@ -14,109 +11,91 @@ 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.devectorizer import load_store_folding, load_store_indexing, devectorize, pm_reduce, \
   ReduceContext, correct_load_store, pm_render
-from tinygrad.codegen.opt.postrange import pm_postrange_opt
+from tinygrad.codegen.opt.postrange import apply_opts
 from tinygrad.codegen.simplify import pm_simplify_ranges, pm_reduce_simplify, pm_flatten_range, pm_split_ranges
 from tinygrad.schedule.rangeify import pm_add_buffers, rangeify_codegen
-from tinygrad.codegen.late.control_flow import CFGContext, pm_merge_ends, pm_add_control_flow, linearize
+from tinygrad.codegen.late.control_flow import CFGContext, pm_add_ends, pm_add_control_flow, linearize
 
-@dataclass
-class RewriteStep:
-  pm: PatternMatcher
-  ctx: Callable[[UOp], Any]|None = None
-  name: str|None = None
-  bottom_up: bool = False
-  def __call__(self, sink:UOp):
-    return graph_rewrite(sink, self.pm, ctx=self.ctx(sink) if self.ctx is not None else None, name=self.name, bottom_up=self.bottom_up)
-
-def apply_rewrites(sink:UOp, rewrites:list[RewriteStep]): return functools.reduce(lambda x,f: f(x), rewrites, sink)
-
-def get_rewrites_for_renderer(opts:Renderer, optimize:bool=True, linearizer:bool=True) -> list[RewriteStep]:
-  # cache with the values of the context vars
-  return _get_rewrites_for_renderer(opts, optimize, linearizer, QUANTIZE.value, DEVECTORIZE.value, TRANSCENDENTAL.value)
-
-@functools.cache
-def _get_rewrites_for_renderer(opts:Renderer, optimize:bool, linearizer:bool, _QUANTIZE, _DEVECTORIZE, _TRANSCENDENTAL) -> list[RewriteStep]:
-  # ** lowerer **
-  ret: list[RewriteStep] = []
+def full_rewrite_to_sink(sink:UOp, ren:Renderer|None=None, optimize:bool=True) -> UOp:
+  if ren is None: ren = Renderer()
 
+  # first we optimize
   if optimize:
-
-    # lowerer first
-    if _QUANTIZE and opts.device in {"CPU", "DSP"}: ret.append(RewriteStep(pm_quant, name="quantize"))
+    if QUANTIZE and ren.device in {"CPU", "DSP"}: sink = graph_rewrite(sink, pm_quant, name="quantize")
 
     # split ranges
-    ret.append(RewriteStep(pm_split_ranges+pm_flatten_range, ctx=lambda _: {}, name="split ranges"))
+    sink = graph_rewrite(sink, pm_split_ranges+pm_flatten_range, ctx={}, name="split ranges")
 
     # symbolic (NOTE: this is a requirement for pm_simplify_ranges to be correct)
-    ret.append(RewriteStep(sym+pm_flatten_range, name="initial symbolic"))
+    sink = graph_rewrite(sink, sym+pm_flatten_range, name="initial symbolic")
 
     # optimize (schedule) the AST
-    ret.append(RewriteStep(pm_simplify_ranges, name="simplify ranges"))
-    ret.append(RewriteStep(pm_reduce_simplify, name="simplify reduces"))
-    ret.append(RewriteStep(pm_postrange_opt, ctx=lambda _: opts, name="post optimize ast"))
+    sink = graph_rewrite(sink, pm_simplify_ranges, name="simplify ranges")
+    sink = graph_rewrite(sink, pm_reduce_simplify, name="simplify reduces")
+
+    # do postrange optimization, BEAM or hand_coded_optimizations
+    sink = apply_opts(sink, ren)
 
   # ** expander (expand_rewrite) **
-  ret.append(RewriteStep(sym+migrate_indexing+pm_move_where_on_load, name="postopt symbolic"))
+  sink = graph_rewrite(sink, sym+migrate_indexing+pm_move_where_on_load, name="postopt symbolic")
 
   # expand
-  ret.append(RewriteStep(sym+pm_pre_expander+pm_group_for_reduce+expander, name="expander"))
+  sink = graph_rewrite(sink, sym+pm_pre_expander+pm_group_for_reduce+expander, name="expander")
 
   # add locals
-  ret.append(RewriteStep(pm_add_buffers+rangeify_codegen, name="add local buffers"))
+  sink = graph_rewrite(sink, pm_add_buffers+rangeify_codegen, name="add local buffers")
 
   # ** devectorizer (full_graph_rewrite) **
   # remove reduce
-  ret.append(RewriteStep(pm_reduce+gep_pushing, lambda _: ReduceContext(), name="remove_reduce"))
+  sink = graph_rewrite(sink, pm_reduce+gep_pushing, ctx=ReduceContext(), name="remove_reduce")
 
   # add gpu dims (late). this works after devectorize, but it's faster here
-  ret.append(RewriteStep(pm_add_gpudims, lambda _: opts, name="add gpudims"))
+  sink = graph_rewrite(sink, pm_add_gpudims, ctx=ren, name="add gpudims")
+
+  # add ends (after reduces are removed, as long as we have reduces we can have stores)
+  sink = graph_rewrite(sink, pm_add_ends, name="add ends of ranges")
 
   # 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
+  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
   else: pm_devectorize = sym+load_store_folding+correct_load_store+load_store_indexing
-  ret.append(RewriteStep(pm_devectorize, lambda _: opts, name="devectorize"))
-
-  supported_ops = tuple(opts.code_for_op.keys())
-  extra_matcher = opts.extra_matcher if opts.extra_matcher is not None else PatternMatcher([])
+  sink = graph_rewrite(sink, pm_devectorize, ctx=ren, name="devectorize")
 
   # lower the index dtype to a concrete int
-  ret.append(RewriteStep(pm_lower_index_dtype+load_store_indexing, lambda _: opts.device, name="lower all index dtypes"))
-  ret.append(RewriteStep(symbolic, name="post index symbolic"))
+  sink = graph_rewrite(sink, pm_lower_index_dtype+load_store_indexing, ctx=ren.device, name="lower all index dtypes")
+  sink = graph_rewrite(sink, symbolic, name="post index symbolic")
 
   # optional pre matcher
-  if opts.pre_matcher is not None: ret.append(RewriteStep(opts.pre_matcher, name="pre_matcher"))
+  if ren.pre_matcher is not None: sink = graph_rewrite(sink, ren.pre_matcher, name="pre_matcher")
 
   # decompositions
-  pm_decomp = symbolic_simple+get_late_rewrite_patterns(supported_ops, _TRANSCENDENTAL>=2)
-  ret.append(RewriteStep(pm_decomp, lambda _: opts.device, name="decompositions"))
+  supported_ops = tuple(ren.code_for_op.keys())
+  pm_decomp = symbolic_simple+get_late_rewrite_patterns(supported_ops, TRANSCENDENTAL>=2)
+  sink = graph_rewrite(sink, pm_decomp, ctx=ren.device, name="decompositions")
 
   # final rules for the renderer (without sym)
+  extra_matcher = ren.extra_matcher if ren.extra_matcher is not None else PatternMatcher([])
   pm_final_rewrite = pm_decomp+pm_render+extra_matcher
-  ret.append(RewriteStep(pm_final_rewrite, lambda _: opts.device, name="final rewrite"))
+  sink = graph_rewrite(sink, pm_final_rewrite, ctx=ren.device, name="final rewrite")
 
   # this was the linearizer
-  ret.append(RewriteStep(pm_merge_ends, name="merge ends"))
-  ret.append(RewriteStep(pm_add_control_flow, CFGContext, name="add control flow starts", bottom_up=True))
+  sink = graph_rewrite(sink, pm_add_control_flow, ctx=CFGContext(sink), name="add control flow", bottom_up=True)
 
-  # return the list
-  return ret
+  # return the rewritten sink
+  return sink
 
-def full_rewrite_to_sink(sink:UOp, opts:Renderer|None=None, optimize:bool=True) -> UOp:
-  return apply_rewrites(sink, get_rewrites_for_renderer(opts if opts is not None else Renderer(), optimize))
-
-def full_rewrite(sink:UOp, opts:Renderer|None=None) -> list[UOp]:
+def full_rewrite(sink:UOp, ren:Renderer|None=None) -> list[UOp]:
   """
   Function to transform the Kernel UOp graph into a linearized program.
 
   Args:
     sink: The Ops.SINK rooting the Kernel graph.
-    opts: The Renderer (can change how things are processed, fix this).
+    ren: The Renderer (can change how things are processed, fix this).
 
   Returns:
     Linear program in UOps.
   """
 
-  lst = linearize(full_rewrite_to_sink(sink, opts, optimize=sink.tag is None))
-  if __debug__: type_verify(lst)
+  lst = linearize(full_rewrite_to_sink(sink, ren, optimize=sink.tag is None))
+  if __debug__: type_verify(lst, program_spec)
   return lst

tinygrad/codegen/gpudims.py

@@ -1,7 +1,7 @@
 import math, functools, operator
 from tinygrad.uop.ops import UOp, Ops, sint, PatternMatcher, UPat, KernelInfo, ssimplify, AxisType, sint_to_uop
 from tinygrad.helpers import all_int, dedup, get_contraction
-from tinygrad.dtype import dtypes
+from tinygrad.dtype import dtypes, AddrSpace, Invalid
 from tinygrad.renderer import Renderer
 
 def _group_dims(dims:tuple[sint, ...], max_sizes:tuple[int, ...]):
@@ -79,6 +79,14 @@ def add_gpudims(ctx:Renderer, s:UOp):
   # apply to multiple ranges
   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:
+      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):
+        assert len(idx.src) == 2, "index has 2 sources"
+        mask: UOp = functools.reduce(operator.and_, [x.eq(0) for x in missing_locals])
+        subs[idx] = idx.replace(src=(idx.src[0], mask.broadcast(idx.src[1].dtype.count).where(idx.src[1], Invalid)))
     if r.op is not Ops.RANGE: continue
     try:
       ii = (global_dims+local_dims).index(r.arg[0:-1])
@@ -87,15 +95,7 @@ def add_gpudims(ctx:Renderer, s:UOp):
     except ValueError: continue
   return s.substitute(subs)
 
-def add_barrier_and_if(buf:UOp, e:UOp):
-  # TODO: this is not generic
-  local_ranges = [x for x in e.ended_ranges if x.op is Ops.RANGE and x.arg[-1] == AxisType.GROUP_REDUCE]
-  if len(local_ranges) == 0: return None
-  return buf.after(UOp(Ops.IF, dtype=dtypes.void, src=(functools.reduce(operator.and_, [x.eq(0) for x in local_ranges]), e.barrier())))
-
 pm_add_gpudims = PatternMatcher([
   # add gpudims must be last
   (UPat(Ops.SINK, name="s"), add_gpudims),
-  # add barrier and if
-  (UPat(Ops.AFTER, src=(UPat(Ops.DEFINE_LOCAL, name="buf"), UPat(Ops.END, name="e"))), add_barrier_and_if),
 ])

tinygrad/codegen/late/control_flow.py

@@ -1,6 +1,29 @@
-import heapq
+import heapq, functools
+from typing import cast
 from collections import defaultdict
+from tinygrad.dtype import dtypes
 from tinygrad.uop.ops import PatternMatcher, UOp, Ops, UPat
+from tinygrad.helpers import panic
+
+# only needed if device doesn't support gated stores
+pm_linearize_cleanups = PatternMatcher([
+  # if statements are not allowed in the graph
+  (UPat((Ops.IF, Ops.ENDIF)), lambda: panic(RuntimeError("if not allowed in graph"))),
+  # gated INDEX becomes IF-STORE-ENDIF. this is the only use of IF-ENDIF
+  (UPat(Ops.STORE, name="u", src=(UPat(Ops.INDEX, src=(UPat(), UPat(), UPat(name="gate", dtype=dtypes.bool))).or_casted(), UPat()),
+        allow_any_len=True), lambda u, gate: (u, [mif:=UOp(Ops.IF, src=(gate, u.src[0])), u, UOp(Ops.ENDIF, src=(mif,))]))
+])
+
+# requires lst be toposorted. like graph rewrite, but for lines
+def line_rewrite(lst:list[UOp], pm:PatternMatcher) -> list[UOp]:
+  newlst = []
+  replaced: dict[UOp, UOp] = {}
+  for u in lst:
+    nu = u.replace(src=tuple([replaced[x] for x in u.src]))
+    ret: tuple[UOp, list[UOp]] = cast(tuple[UOp, list[UOp]]|None, pm.rewrite(nu)) or (nu, [nu])
+    replaced[u] = ret[0]
+    newlst.extend(ret[1])
+  return newlst
 
 def linearize(u:UOp) -> list[UOp]:
   lst = list(u.toposort())
@@ -40,7 +63,7 @@ def linearize(u:UOp) -> list[UOp]:
       if in_degree[v] == 0: heapq.heappush(heap, (nkey[v],v))
 
   assert len(newlst) == len(lst), f"len mismatch {len(newlst)} != {len(lst)}"
-  return newlst
+  return line_rewrite(newlst, pm_linearize_cleanups)
 
 class CFGContext:
   def __init__(self, sink:UOp):
@@ -49,20 +72,23 @@ class CFGContext:
     # dependent, meaning endrange y is a dependency of endrange x and range x is not a dependency of endrange y
     # independent, endrange y is not a dependency of endrange x
     # everything is nested inside the sink
-    deps: dict[UOp, set[UOp]] = {}
+    deps: dict[UOp, dict[UOp, None]] = {}
     nesting: dict[UOp, UOp] = {}
     for u in sink.toposort():
-      deps[u] = set().union(*(deps[s] for s in u.src))
-      if u.op in (Ops.END, Ops.ENDIF, Ops.SINK):
-        nesting |= {x:u for x in deps[u] if x.op in (Ops.END, Ops.ENDIF) and (u.op is Ops.SINK or u.src[0] in deps[x]) and x not in nesting}
-      if u.op in (Ops.RANGE, Ops.END, Ops.IF, Ops.ENDIF): deps[u] |= {u}
+      # get the deps from the src
+      deps[u] = {}
+      for s in u.src: deps[u] |= deps[s]
+
+      if u.op in (Ops.END, Ops.SINK):
+        nesting |= {x:u for x in deps[u] if x.op is Ops.END and (u.op is Ops.SINK or u.src[0] in deps[x]) and x not in nesting}
+      if u.op in (Ops.RANGE, Ops.END): deps[u][u] = None
 
     self.edges: dict[UOp, UOp] = {}
     siblings: dict[UOp, list[UOp]] = {}
     for k,vv in nesting.items(): siblings.setdefault(vv, []).append(k)
     for k,v in siblings.items():
       # range/if that have dependencies on other siblings need to run after them
-      order = sorted(v, key=lambda x: len(deps[x].intersection(v)))
+      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[0]] + order, order)
       for x,y in zipped:
         # TODO: is this check correct?
@@ -70,33 +96,11 @@ class CFGContext:
           self.edges[y.src[0]] = x
 
 pm_add_control_flow = PatternMatcher([
-  (UPat((Ops.RANGE, Ops.IF), name="x"), lambda ctx,x: x.replace(src=x.src+(y,)) if (y:=ctx.edges.get(x)) is not None else None),
+  (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),
 ])
 
-def do_merge_ends(s:UOp):
-  # NOTE: this can fail
-  stacked: dict[UOp, list[UOp]] = {}
-  dangling_ifs = []
-  for x in s.toposort():
-    if x.op in {Ops.END, Ops.ENDIF}:
-      assert x.op is not Ops.END or x.arg == 1, "ends must be single ends for linearizer"
-      stacked.setdefault(x.src[0], []).append(x)
-    if x.op is Ops.IF: dangling_ifs.append(x)
-  dangling_ifs = [x for x in dangling_ifs if x not in stacked]
-  replaces = {}
-  for k,v in stacked.items():
-    if len(v) == 1: continue
-    rep = UOp(v[0].op, src=tuple([k] + [y for x in v for y in x.src[1:]]), arg=x[0].arg)
-    for x in v: replaces[x] = rep
-  if not len(replaces) and not len(dangling_ifs): return None
-  ret = s.substitute(replaces)
-  if len(dangling_ifs):
-    assert len(dangling_ifs) == 1, "we only support 1 dangling if"
-    ret = ret.replace(src=(UOp(Ops.ENDIF, src=(dangling_ifs[0], *ret.src)),))
-  return ret
-
-pm_merge_ends = PatternMatcher([
-  # for renderering and linearizing, all ends must end one loop
-  (UPat(Ops.END, name="e"), lambda e: e.replace(src=e.src[e.arg-1:], arg=1).end(ends=e.src[:e.arg-1]) if e.arg > 1 else None),
-  (UPat(Ops.SINK, name="s"), do_merge_ends),
-])
+pm_add_ends = PatternMatcher([
+  # put the end on the store
+  (UPat(Ops.STORE, name="s"), lambda s:
+    functools.reduce(lambda x,y: y.end(x), [x for x in s.src[2:] if x.op is Ops.RANGE][::-1], s.replace(src=s.src[:2]))),
+])

tinygrad/codegen/late/devectorizer.py

@@ -268,10 +268,6 @@ pm_render = PatternMatcher([
     UPat.var("a")), lambda c,idx,l,a: l.replace(src=(l.src[0], a.cast(l.dtype))+l.src[2:]).cast(a.dtype)),
   (UPat.var("c").where(UPat.var("a"), UPat(Ops.LOAD, src=(UPat().index(UPat.var("idx"), UPat.var("c").logical_not()).or_casted(),),
     allow_any_len=True, name="l").or_casted()), lambda c,idx,l,a: l.replace(src=(l.src[0], a.cast(l.dtype))+l.src[2:]).cast(a.dtype)),
-  # gate any stores that aren't gated with if/endif pairs
-  (UPat(Ops.STORE, src=(UPat(src=(UPat(), UPat(), UPat(dtype=dtypes.bool)), name="idx").or_casted(), UPat()), name="store", allow_any_len=True),
-    lambda store,idx: UOp(Ops.ENDIF, src=(uif:=UOp(Ops.IF, src=(idx.src[2],)), UOp(Ops.STORE, src=store.src[:2]+(uif,)+store.src[2:]))) if \
-      len(store.src) <= 2 or store.src[2].op != Ops.IF else None),
 ])
 
 # *** Ops.REDUCE -> Ops.DEFINE_ACC ***
@@ -295,7 +291,7 @@ def reduce_to_acc(ctx:ReduceContext, red:UOp):
   # if we have a range
   if len(reduce_range) != 0:
     topo = inp.toposort()
-    ended_ranges = flatten([x.src[:x.arg] for x in topo if x.op is Ops.END])
+    ended_ranges = flatten([x.ended_ranges for x in topo if x.op is Ops.STORE])
     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,))
@@ -305,7 +301,7 @@ def reduce_to_acc(ctx:ReduceContext, red:UOp):
     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(ends=reduce_range[::-1])).index(UOp.const(dtypes.int, 0)).load()
+  return acc.after(acc.index(UOp.const(dtypes.int, 0)).store(ret, *reduce_range)).index(UOp.const(dtypes.int, 0)).load()
 
 pm_reduce = PatternMatcher([
   # REDUCE -> DEFINE_ACC+ASSIGN

tinygrad/codegen/opt/heuristic.py

@@ -62,8 +62,8 @@ def hand_coded_optimizations(k:Scheduler) -> Scheduler:
 
   # should use matvec - TODO: adjust/tune based on the wide vs tall/large vs small mat
   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.opts.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.opts.has_shared and \
+  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()
     if k.ranges_of(AxisType.REDUCE):
@@ -103,7 +103,7 @@ def hand_coded_optimizations(k:Scheduler) -> Scheduler:
   for axis in to_upcast[::-1]: k.apply_opt(Opt(OptOps.UPCAST, axis, 0))
 
   # potentially do more upcasts of non reduce axes based on a heuristic
-  is_dsp = k.opts is not None and k.opts.device == "DSP"
+  is_dsp = k.ren is not None and k.ren.device == "DSP"
   upcasted_axis: set[int] = set()
   while resolve(prod(k.output_shape[i] for i in k.upcastable_dims) >= 1024):
     xb_choices = []
@@ -155,7 +155,7 @@ def hand_coded_optimizations(k:Scheduler) -> Scheduler:
 
   # **** local groups ****
 
-  if k.opts.has_local:
+  if k.ren.has_local:
     if NOLOCALS:
       k.apply_opt(Opt(OptOps.NOLOCALS))
     else:
@@ -176,10 +176,10 @@ def hand_coded_optimizations(k:Scheduler) -> Scheduler:
 
   # **** threading ****
 
-  if k.opts.has_threads and k.opts.global_max is not None:
+  if k.ren.has_threads and k.ren.global_max is not None:
     for threads in [32,16,12,8,6,5,4,3,2]:
       # Skip if too many threads. Heuristic: use about 128K ops per thread
-      if threads > k.opts.global_max[0] or resolve(prod(k.full_shape) // (128 << 10) < threads): continue
+      if threads > k.ren.global_max[0] or resolve(prod(k.full_shape) // (128 << 10) < threads): continue
       for axis in k.axes_of(AxisType.LOOP):
         if k.full_shape[axis] % threads == 0:
           k.apply_opt(Opt(OptOps.THREAD, axis, threads))

tinygrad/codegen/opt/postrange.py

@@ -4,8 +4,8 @@ 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.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
+from tinygrad.dtype import dtypes, ImageDType, AddrSpace
+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.simplify import pm_flatten_range
 from tinygrad.renderer import Renderer
@@ -17,8 +17,8 @@ axis_to_pos = {AxisType.LOOP: -1, AxisType.THREAD: 0, AxisType.GLOBAL: 0, AxisTy
                AxisType.GROUP_REDUCE: 2, AxisType.REDUCE: 4, AxisType.UNROLL: 5}
 
 class Scheduler:
-  def __init__(self, ast:UOp, opts:Renderer):
-    self.ast, self.opts = ast, opts
+  def __init__(self, ast:UOp, ren:Renderer):
+    self.ast, self.ren = ast, ren
     self.dont_use_locals = self.ast.arg.dont_use_locals if self.ast.arg is not None else False
     self.applied_opts = list(self.ast.arg.applied_opts) if self.ast.arg is not None else []
 
@@ -46,7 +46,7 @@ class Scheduler:
   def shape_str_to_axis(self, nms:list[str]) -> tuple[int, ...]: return tuple([self.shape_str().index(x) for x in nms])
 
   def copy(self):
-    ret = Scheduler(self.ast, self.opts)
+    ret = Scheduler(self.ast, self.ren)
     ret.dont_use_locals = self.dont_use_locals
     ret.applied_opts = self.applied_opts[:]
     return ret
@@ -64,11 +64,22 @@ class Scheduler:
     return self.ast.replace(arg=KernelInfo(name=name, applied_opts=tuple(self.applied_opts), dont_use_locals=self.dont_use_locals), tag=1)
 
   def _globalizable_rngs(self) -> list[UOp]:
-    # all ranges that end before any STOREs
-    return [x for x in self.ast.toposort(lambda x: x.op is not Ops.STORE) if x.op is Ops.RANGE and x not in self.ast.ranges]
+    store_rngs = self.ast.src[0].src[2:]
+    # filter any not in local stores
+    local_store_rngs = [x.ranges for x in self.ast.toposort() if (x.op is Ops.STORE and x.src[0].ptrdtype.addrspace == AddrSpace.LOCAL) \
+                        or (x.op is Ops.BUFFERIZE and x.arg == AddrSpace.LOCAL)]
+    for ls in local_store_rngs: store_rngs = tuple([x for x in store_rngs if x in ls])
+
+    # filter any not in reduces
+    # TODO: enable this
+    """
+    reduce_rngs = [x.ranges for x in self.ast.toposort() if x.op is Ops.REDUCE]
+    for ls in reduce_rngs: store_rngs = tuple([x for x in store_rngs if x in ls])
+    """
+    return [x for x in UOp.sink(*store_rngs).toposort() if x.op is Ops.RANGE and x.arg[-1] == AxisType.LOOP] if store_rngs else []
 
   def convert_loop_to_global(self):
-    if not self.opts.has_local: return None
+    if not self.ren.has_local: return None
 
     globalizible_rngs = self._globalizable_rngs()
     rng = [x.replace(arg=x.arg[0:-1]+(AxisType.GLOBAL,)) if x in globalizible_rngs else x for x in self.rngs]
@@ -76,11 +87,11 @@ class Scheduler:
     self.ast = self.ast.substitute(dict(zip(self.rngs, rng)))
 
   def colors(self) -> list[str]:
-    globalizible_rngs = self._globalizable_rngs()
+    output_rngs = flatten([s.src[2:] for s in self.ast.src])
     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 globalizible_rngs and x == AxisType.LOOP: ret.append("BLACK")
+      elif r not in output_rngs and x == AxisType.LOOP: ret.append("BLACK")
       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())])
@@ -122,7 +133,7 @@ class Scheduler:
       return
 
     if opt.op in {OptOps.LOCAL, OptOps.GROUP, OptOps.GROUPTOP}:
-      check(self.opts.has_local, "locals needed for opt")
+      check(self.ren.has_local, "locals needed for opt")
 
     rng = self.rngs[real_axis] if (real_axis:=self.real_axis(opt.op, opt.axis)) >= 0 else UOp(Ops.NOOP)
 
@@ -140,7 +151,7 @@ class Scheduler:
                                         (self.group_for_reduces and opt.op not in {OptOps.NOLOCALS, OptOps.PADTO})):
         upcast_local_sz = prod([self.full_shape[a] for a in self.axes_of(AxisType.UPCAST, AxisType.WARP, AxisType.LOCAL, AxisType.GROUP_REDUCE)])
         smem_sz = amt*upcast_local_sz*self.reduceop.dtype.itemsize
-        check(smem_sz <= self.opts.shared_max, f"exceeds maximum shared memory size: needs {smem_sz}, max {self.opts.shared_max}")
+        check(smem_sz <= self.ren.shared_max, f"exceeds maximum shared memory size: needs {smem_sz}, max {self.ren.shared_max}")
       if self.reduceop is not None and (opt.op in {OptOps.GROUP, OptOps.GROUPTOP}):
         # We currently dont support a group within another rudece, TODO: fix if-contexts
         reduce = [u for u in self.ast.backward_slice if u.op is Ops.REDUCE and rng in merge_dicts([r.ranges for r in u.src[1:]])][0]
@@ -151,14 +162,14 @@ class Scheduler:
         check(amt <= 32, "don't unroll more than 32")
         check(rng.arg[-1] in {AxisType.GROUP_REDUCE, AxisType.REDUCE}, "unroll is for GROUP_REDUCE/REDUCE")
       if opt.op is OptOps.UPCAST:
-        check((self.opts is not None and self.opts.device == "DSP") or amt <= 16, "don't upcast more than 16")
+        check((self.ren is not None and self.ren.device == "DSP") or amt <= 16, "don't upcast more than 16")
         check(rng.arg[-1] in {AxisType.GLOBAL, AxisType.LOCAL, AxisType.LOOP}, f"upcast is for GLOBAL/LOCAL/LOOP, not {rng.arg[-1]}")
       if opt.op is OptOps.LOCAL:
         check(not self.dont_use_locals, "can't use locals")
         check(rng.arg[-1] in {AxisType.GLOBAL, AxisType.LOOP}, "local is for globals")
       if opt.op is OptOps.THREAD:
-        check(self.opts is not None and self.opts.has_threads, "target does not support threads")
-        check(self.opts is not None and self.opts.global_max is not None and amt <= self.opts.global_max[0], "too many threads")
+        check(self.ren is not None and self.ren.has_threads, "target does not support threads")
+        check(self.ren is not None and self.ren.global_max is not None and amt <= self.ren.global_max[0], "too many threads")
         check(all(x is not AxisType.THREAD for x in self.axis_types), "already threaded")
         check(rng in self._globalizable_rngs(), "can't apply range to this dim")
       if opt.op in {OptOps.GROUP, OptOps.GROUPTOP}:
@@ -170,7 +181,7 @@ class Scheduler:
       check(len(self.applied_opts) == 0, "tensor core opts must be first") # TODO: remove the need for this by having warps
       check(opt.axis is not None, "tensor core opts must have an axis")
       check(opt.arg is not None and isinstance(opt.arg, tuple) and len(opt.arg) == 3, "tensor core opts must have valid arg")
-      check(-1 <= (tc_select:=cast(tuple, opt.arg)[0]) < len(self.opts.tensor_cores), "tensor core opts must have valid tc_select")
+      check(-1 <= (tc_select:=cast(tuple, opt.arg)[0]) < len(self.ren.tensor_cores), "tensor core opts must have valid tc_select")
       check(0 <= (tc_opt:=cast(tuple, opt.arg)[1]) <= 2, "tensor core opts must have valid tc_opt")
       check(0 < (use_tensor_cores:=cast(tuple, opt.arg)[2]) <= 2, "use_tensor_cores value is not valid")
       try: ret = self._apply_tc_opt(use_tensor_cores, cast(int, opt.axis), tc_select, tc_opt)
@@ -217,7 +228,7 @@ class Scheduler:
       if mul.op is not Ops.MUL: return None
       in0, in1 = mul.src
       try:
-        tensor_cores = self.opts.tensor_cores if tc_select == -1 else [self.opts.tensor_cores[tc_select]]
+        tensor_cores = self.ren.tensor_cores if tc_select == -1 else [self.ren.tensor_cores[tc_select]]
       except IndexError:
         raise KernelOptError(f"invalid tensor core choice {tc_select}")
       for tc in tensor_cores:
@@ -288,7 +299,7 @@ class Scheduler:
             # TODO: remove tc_upcast_axes from the arg
             # do the reduce_axes always disappear? i think they don't
             # they need to be moved into the WMMA srcs
-            wmma_arg = (str(tc), tc.dims, tc.dtype_in, tc.dtype_out, self.opts.device, tc.threads, tc_upcast_axes, ()) #, tc_reduce_axes)
+            wmma_arg = (str(tc), tc.dims, tc.dtype_in, tc.dtype_out, self.ren.device, tc.threads, tc_upcast_axes, ()) #, tc_reduce_axes)
             wmma = UOp(Ops.WMMA, dtype=tc.dtype_out.vec(tc.elements_per_thread[2]), src=(
               UOp(Ops.CONTRACT, dtype=srcs[0].dtype.vec(tc.elements_per_thread[0]), src=(srcs[0],), arg=tc_upcast_axes[0], tag=1),
               UOp(Ops.CONTRACT, dtype=srcs[1].dtype.vec(tc.elements_per_thread[1]), src=(srcs[1],), arg=tc_upcast_axes[1], tag=1),
@@ -322,15 +333,15 @@ def bufs_from_ast(ast:UOp, dname:str) -> list[Buffer]:
   glbls = sorted([x for x in ast.backward_slice if x.op is Ops.DEFINE_GLOBAL], key=lambda x: x.arg)
   return [Buffer(dname, x.ptrdtype.size, x.dtype.base if not isinstance(x.dtype, ImageDType) else x.dtype) for x in glbls]
 
-def apply_opts(ctx:Renderer, ast:UOp):
-  if ast.tag is not None: return None
-  k = Scheduler(ast, ctx)
+def apply_opts(ast:UOp, ren:Renderer) -> UOp:
+  if ast.tag is not None: return ast
+  k = Scheduler(ast, ren)
   k.convert_loop_to_global()
   if ast.arg is not None and ast.arg.opts_to_apply is not None:
     for opt in ast.arg.opts_to_apply: k.apply_opt(opt)
   elif BEAM >= 1:
     from tinygrad.codegen.opt.search import beam_search
-    rawbufs = bufs_from_ast(ast, ctx.device)
+    rawbufs = bufs_from_ast(ast, ren.device)
     k = beam_search(k, rawbufs, BEAM.value, bool(getenv("BEAM_ESTIMATE", 1)))
   elif not NOOPT and (ast.arg is None or ast.arg.applied_opts == ()):
     from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
@@ -338,7 +349,3 @@ def apply_opts(ctx:Renderer, ast:UOp):
     if not any(u.op is Ops.AFTER and u.src[0].op is Ops.DEFINE_LOCAL 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)
-
-pm_postrange_opt = PatternMatcher([
-  (UPat(Ops.SINK, name="ast"), apply_opts),
-])

tinygrad/codegen/opt/search.py

@@ -66,7 +66,7 @@ def _try_compile_linearized_w_idx(x:tuple[int,Scheduler], compiler:Compiler) ->
     signal.alarm(getenv("BEAM_TIMEOUT_SEC", 10))
   ret = None
   try:
-    p = get_program(x[1].copy().get_optimized_ast(name_override="test"), x[1].opts)
+    p = get_program(x[1].copy().get_optimized_ast(name_override="test"), x[1].ren)
     assert p.uops is not None, "uop list wasn't generated?"
     if len(p.uops) >= (uops_max:=getenv("BEAM_UOPS_MAX", 3000)) > 0:
       if getenv("BEAM_LOG_SURPASS_MAX"): print(f"too many uops. {len(p.uops)=}, {uops_max=}")
@@ -119,7 +119,7 @@ def get_kernel_actions(lin:Scheduler, include_0=True, candidates:list[Opt]|None=
 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):
   global beam_pool
-  key = {"ast": lin.ast.key, "amt": amt, "allow_test_size": allow_test_size, "device": lin.opts.device, "suffix": lin.opts.suffix}
+  key = {"ast": lin.ast.key, "amt": amt, "allow_test_size": allow_test_size, "device": lin.ren.device, "suffix": lin.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)
@@ -128,7 +128,7 @@ def beam_search(lin:Scheduler, rawbufs:list[Buffer], amt:int, allow_test_size=Tr
   beam: list[tuple[Scheduler, float]] = [(lin, float("inf"))]
   seen_libs = set()
 
-  default_parallel = multiprocessing.cpu_count() if lin.opts.device in {"CUDA", "AMD", "NV", "METAL", "HIP"} else 0
+  default_parallel = multiprocessing.cpu_count() if lin.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
@@ -144,7 +144,7 @@ def beam_search(lin:Scheduler, rawbufs:list[Buffer], amt:int, allow_test_size=Tr
     rawbufs = _ensure_buffer_alloc(rawbufs)
     var_vals: dict[str, int] = {k.expr:int(k.vmax+k.vmin)//2 for k in lin.ast.variables()}
     exiting, st = False, time.perf_counter()
-    dev = Device[lin.opts.device]
+    dev = Device[lin.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]] = []

tinygrad/codegen/opt/tc.py

@@ -80,6 +80,10 @@ cuda_81616 = [TensorCore(dims=(8,16,16), threads=32, elements_per_thread=(8,4,4)
   swizzle=((('r1', 'r2', 'l2', 'l3', 'l4'), ('u1', 'r3'), ('l0', 'l1', 'u0', 'r0')),
            (('r1', 'r2', 'u0', 'l0', 'l1'), ('r0', 'r3'), ('l2', 'l3', 'l4', 'u1'))))
   for di,do in [(dtypes.half,dtypes.float), (dtypes.bfloat16,dtypes.float), (dtypes.half,dtypes.half)]]
+cuda_81632_f8 = [TensorCore(dims=(8,16,32), threads=32, elements_per_thread=(16,8,4), dtype_in=di, dtype_out=do, opts=cuda_tc_opts,
+  swizzle=((('r2', 'r3', 'l2', 'l3', 'l4'), ('u1', 'r4'), ('l0', 'l1', 'u0', 'r0', 'r1')),
+           (('r2', 'r3', 'u0', 'l0', 'l1'), ('r1', 'r4'), ('l2', 'l3', 'l4', 'u1', 'r0'))))
+  for di,do in [(dtypes.fp8e4m3,dtypes.float),(dtypes.fp8e5m2,dtypes.float)]]
 cuda_8168_f16 = [TensorCore(dims=(8,16,8), threads=32, elements_per_thread=(4,2,4), dtype_in=di, dtype_out=do, opts=cuda_tc_opts,
   swizzle=((('r1', 'r2', 'l2', 'l3', 'l4'), ('r0', 'u1'), ('l0', 'l1', 'u0')),
            (('r1', 'r2', 'u0', 'l0', 'l1'), ('u1', 'r0'), ('l2', 'l3', 'l4'))))
@@ -87,9 +91,10 @@ cuda_8168_f16 = [TensorCore(dims=(8,16,8), threads=32, elements_per_thread=(4,2,
 cuda_8168_tf32 = [TensorCore(dims=(8,16,8), threads=32, elements_per_thread=(4,2,4), dtype_in=dtypes.float, dtype_out=dtypes.float, opts=cuda_tc_opts,
   swizzle=((('r0', 'r1', 'l2', 'l3', 'l4'), ('u1', 'r2'), ('l0', 'l1', 'u0')),
            (('r0', 'r1', 'u0', 'l0', 'l1'), ('u1', 'r2'), ('l2', 'l3', 'l4'))))]
+cuda_sm75: list[TensorCore] = cuda_8168_f16
 cuda_sm80: list[TensorCore] = cuda_81616 + cuda_8168_f16
 if getenv("ALLOW_TF32", 0): cuda_sm80 += cuda_8168_tf32
-cuda_sm75: list[TensorCore] = cuda_8168_f16
+cuda_sm89: list[TensorCore] = cuda_sm80 + cuda_81632_f8
 
 # ***** AMD *****
 
@@ -112,6 +117,14 @@ amd_cdna = [TensorCore(dims=(16,16,16), threads=64, elements_per_thread=(4,4,4),
            (('l0', 'l1', 'l2', 'l3', 'r2', 'r3'), ('r0', 'r1'), ('l4', 'l5', 'u0', 'u1'))))
   for di,do in [(dtypes.half,dtypes.float),(dtypes.bfloat16,dtypes.float)]]
 
+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)]]
+
+amd_cdna4 = amd_cdna_161632 + amd_cdna
+
 # ***** Apple Metal *****
 
 metal = [TensorCore(dims=(8,8,8), threads=32, elements_per_thread=(2,2,2), dtype_in=di, dtype_out=do,

tinygrad/codegen/simplify.py

@@ -13,8 +13,6 @@ def flatten_range(r:UOp):
 pm_flatten_range = PatternMatcher([
   # real ranges only
   (UPat((Ops.REDUCE, Ops.STORE), name="r"), flatten_range),
-  # END is only on RANGES. TODO: this is copied from symbolic
-  (UPat(Ops.END, name="e"), lambda e: UOp.end(*e.src[e.arg:], ends=sorted(UOp.sink(*e.src[:e.arg]).ranges, key=lambda x: x.arg))),
 ])
 
 def count_divmod(x:UOp): return len([u for u in x.toposort() if u.op in {Ops.IDIV, Ops.MOD}])
@@ -41,7 +39,7 @@ def simplify_merge_adjacent(u:UOp) -> UOp|None:
   return u
 
 pm_simplify_ranges = PatternMatcher([
-  (UPat((Ops.END, Ops.REDUCE), name="u"), simplify_merge_adjacent),
+  (UPat((Ops.STORE, Ops.REDUCE), name="u"), simplify_merge_adjacent),
 ])
 
 def mark_range_mod(ctx, r:UOp, c:UOp):

tinygrad/device.py

@@ -331,7 +331,9 @@ def is_dtype_supported(dtype:DType, device:str|None=None) -> bool:
     if device in {"CUDA", "NV"}: return not CI and not getenv(f"{device}_PTX") and not getenv("NV_NAK")
     if device in {"CPU"}: return not CI and platform.machine() in {"arm", "arm64", "aarch64", "x86_64", "amd64"} and not getenv("CPU_LVP")
     return device in {"AMD", "PYTHON", "NULL"}
-  if dtype in dtypes.fp8s: return device in {"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")
+    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]
   # for CI GPU and OSX, cl_khr_fp16 isn't supported

tinygrad/helpers.py

@@ -85,6 +85,7 @@ 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
 
 @functools.cache
 def canonicalize_strides(shape:tuple[T, ...], strides:tuple[T, ...]) -> tuple[T, ...]:
@@ -157,7 +158,7 @@ TRANSCENDENTAL, NOLOCALS = ContextVar("TRANSCENDENTAL", 1), ContextVar("NOLOCALS
 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)
-DISABLE_COMPILER_CACHE, BLOCK_REORDER = ContextVar("DISABLE_COMPILER_CACHE", 0), ContextVar("BLOCK_REORDER", 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)
 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)
@@ -170,6 +171,7 @@ SPEC = ContextVar("SPEC", 0)
 # TODO: disable by default due to speed
 IGNORE_OOB = ContextVar("IGNORE_OOB", 1)
 PCONTIG = ContextVar("PCONTIG", 0)  # partial contiguous in rangeify
+DEBUG_RANGEIFY = ContextVar("DEBUG_RANGEIFY", 0)
 
 @dataclass(frozen=True)
 class Metadata:

tinygrad/nn/optim.py

@@ -80,7 +80,7 @@ def SGD(params: list[Tensor], lr=0.001, momentum=0.0, weight_decay=0.0, nesterov
   return LARS(params, lr, momentum, weight_decay, 0, None, nesterov, classic=classic, pre_wd=True, tcoef=0.0, fused=fused)
 
 # Muon applies the newton schulz algorithm on gradient. also can include momentum, nesterov, and weight decay
-def Muon(params: list[Tensor], lr=0.02, momentum=0.95, weight_decay=0.0, ns_steps=5, ns_params=(3.4445, -4.775, 2.0315),
+def Muon(params: list[Tensor], lr=0.001, momentum=0.95, weight_decay=0.1, ns_steps=5, ns_coefficients=(3.4445, -4.775, 2.0315),
          nesterov=True, fused=FUSE_OPTIM):
   """
   SGD with newton-schulz iteration and post momentum weight decay.
@@ -89,7 +89,7 @@ def Muon(params: list[Tensor], lr=0.02, momentum=0.95, weight_decay=0.0, ns_step
   - Paper: https://arxiv.org/pdf/2502.16982
   """
   assert not fused, "FUSE_OPTIM not allowed for Muon optimizer"
-  return LARS(params, lr, momentum, weight_decay, ns_steps, ns_params, nesterov, classic=False, pre_wd=False, tcoef=0.0, fused=fused)
+  return LARS(params, lr, momentum, weight_decay, ns_steps, ns_coefficients, nesterov, classic=False, pre_wd=False, tcoef=0.0, fused=fused)
 
 class LARS(Optimizer):
   """
@@ -97,10 +97,10 @@ class LARS(Optimizer):
 
   - Paper: https://arxiv.org/abs/1708.03888v3
   """
-  def __init__(self, params:list[Tensor], lr=0.001, momentum=0.9, weight_decay=1e-4, ns_steps=0, ns_params=None,
+  def __init__(self, params:list[Tensor], lr=0.001, momentum=0.9, weight_decay=1e-4, ns_steps=0, ns_coefficients=None,
                nesterov=False, classic=True, pre_wd=True, tcoef=0.001, fused=FUSE_OPTIM):
     super().__init__(params, lr, fused)
-    self.momentum, self.wd, self.ns_steps, self.ns_params  = momentum, weight_decay, ns_steps, ns_params
+    self.momentum, self.wd, self.ns_steps, self.ns_coefficients  = momentum, weight_decay, ns_steps, ns_coefficients
     self.nesterov, self.classic, self.pre_wd, self.tcoef = nesterov, classic, pre_wd, tcoef
     self.b = self._new_optim_param() if self.momentum else []
 
@@ -118,7 +118,7 @@ class LARS(Optimizer):
       if self.momentum:
         self.b[i].assign(self.momentum * self.b[i] + g)  # NOTE: self.b[i] is zero on the first run, no if required
         g = (g + self.momentum * self.b[i]) if self.nesterov else self.b[i]
-      if self.ns_params: g = g.reshape(g.shape[0], -1).newton_schulz(self.ns_steps, self.ns_params).reshape(g.shape)
+      if self.ns_coefficients: g = g.reshape(g.shape[0], -1).newton_schulz(self.ns_steps, self.ns_coefficients).reshape(g.shape)
       # muon does post momentum weight decay
       if not self.pre_wd and self.wd > 0: t = t.detach() * (1.0 - self.wd * self.lr)
       # popular momentum does pre learning rate update

tinygrad/renderer/cstyle.py

@@ -37,7 +37,7 @@ base_rewrite = PatternMatcher([
   (UPat(Ops.CONST, dtype=dtypes.uint32, name="x"), lambda ctx,x: f"{truncate[x.dtype](x.arg)}u"),
   (UPat(Ops.CONST, dtype=dtypes.bool, name="x"), lambda ctx,x: "1" if x.arg else "0"),
   # consts are rendered to larger type and casted
-  (UPat(Ops.CONST, (dtypes.bfloat16, dtypes.half), name="x"), lambda ctx,x: f"({ctx.render_cast(x.dtype, f'{x.arg}f')})"),
+  (UPat(Ops.CONST, (*dtypes.fp8s, dtypes.bfloat16, dtypes.half), name="x"), lambda ctx,x: f"({ctx.render_cast(x.dtype, f'{x.arg}f')})"),
   (UPat(Ops.CONST, (dtypes.uint8, dtypes.uint16), name="x"), lambda ctx,x: f"({ctx.render_cast(x.dtype, f'{x.arg}u')})"),
   (UPat(Ops.CONST, (dtypes.int8, dtypes.int16), name="x"), lambda ctx,x: f"({ctx.render_cast(x.dtype, str(x.arg))})"),
   # default const render
@@ -269,7 +269,8 @@ class OpenCLRenderer(CStyleLanguage):
       lambda ctx,buf,idx,var,gate: f"({ctx[gate]}?read_imagef({ctx[buf]}, smp, {ctx[idx]}):{ctx[var]})"),
     (UPat(Ops.LOAD, dtype=dtypes.float.vec(4), src=(UPat.var('buf').index(UPat.var('idx', dtypes.int.vec(2))),)),
       lambda ctx,buf,idx: f"read_imagef({ctx[buf]}, smp, {ctx[idx]})"),
-    (UPat(Ops.STORE, src=(UPat.var('buf').index(UPat.var('idx', dtypes.int.vec(2))), UPat.var("var", dtypes.float.vec(4))), allow_any_len=True),
+    (UPat(Ops.STORE, src=(UPat.var('buf').index(UPat.var('idx', dtypes.int.vec(2)), allow_any_len=True),
+                          UPat.var("var", dtypes.float.vec(4))), allow_any_len=True),
       lambda ctx,buf,idx,var: f"write_imagef({ctx[buf]}, {ctx[idx]}, {ctx[var]});"),
   ]) + base_rewrite
 
@@ -345,7 +346,8 @@ class CUDARenderer(CStyleLanguage):
   shared_max = 49152
 
   def __init__(self, arch:str):
-    self.tensor_cores, self.arch = tc.cuda_sm80 if int(arch[3:]) >= 80 else tc.cuda_sm75 if int(arch[3:]) >= 75 else [], arch
+    self.arch = arch
+    self.tensor_cores = tc.cuda_sm89 if int(arch[3:]) >= 89 else tc.cuda_sm80 if int(arch[3:]) >= 80 else tc.cuda_sm75 if int(arch[3:]) >= 75 else []
   def __reduce__(self): return self.__class__, (self.arch,)
 
   # language options
@@ -364,8 +366,14 @@ class CUDARenderer(CStyleLanguage):
     Ops.EXP2: lambda x,dtype: f"hexp2({x})" if dtype in (dtypes.half, dtypes.bfloat16) else f"exp2({x})",
     Ops.SQRT: lambda x,dtype: f"hsqrt({x})" if dtype in (dtypes.half, dtypes.bfloat16) else f"sqrt({x})",
     Ops.RECIP: lambda x,dtype: f"hrcp({x})" if dtype in (dtypes.half, dtypes.bfloat16) else f"(1/{x})" }
-  type_map = {dtypes.bfloat16: "nv_bfloat16"}
-
+  type_map = {dtypes.bfloat16: "nv_bfloat16", dtypes.fp8e4m3: "__nv_fp8_e4m3", dtypes.fp8e5m2: "__nv_fp8_e5m2"}
+  extra_matcher = 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()),
     elems, header = ', '.join(_nms[:dt.count]), ', '.join([f"{scal} {x}" for x in _nms[:dt.count]])
@@ -376,11 +384,12 @@ class CUDARenderer(CStyleLanguage):
     prefix = ["#define INFINITY (__int_as_float(0x7f800000))","#define NAN (__int_as_float(0x7fffffff))"]
 
     used_dtypes = uops_to_dtypes(uops)
+    if any(dt.scalar() in dtypes.fp8s for dt in used_dtypes): prefix.append("#include <cuda_fp8.h>")
     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}]
-
-    dt_map_in = { dtypes.float: "tf32", dtypes.half: "f16", dtypes.bfloat16: "bf16" }
+    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)]
+    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):
       upcast_sizes = [prod(size for _, size in upcast) for upcast in upcast_axes]
@@ -414,7 +423,7 @@ class AMDRenderer(CStyleLanguage):
 
   @staticmethod
   def get_tensor_cores(arch):
-    return {"gfx942": tc.amd_cdna, "gfx950": tc.amd_cdna, "gfx1200": tc.amd_rdna4, "gfx1201": tc.amd_rdna4}.get(arch.split(":")[0], tc.amd_rdna3)
+    return {"gfx942": tc.amd_cdna, "gfx950": tc.amd_cdna4, "gfx1200": tc.amd_rdna4, "gfx1201": tc.amd_rdna4}.get(arch.split(":")[0], tc.amd_rdna3)
   def __init__(self, arch:str): # gfx942 => MI300, gfx1100 => RX 7900, gfx1201 => RX 9700
     self.arch = arch
     self.tensor_cores = self.get_tensor_cores(arch)

tinygrad/renderer/llvmir.py

@@ -49,9 +49,11 @@ def render_wmma_amx(ctx, wmma: UOp) -> str:
 def render_wmma_amd(ctx, wmma: UOp, cdna=False) -> str:
   dt_map = {dtypes.half: "f16", dtypes.float: "f32", dtypes.ushort: "bf16.1k" if cdna else "bf16", dtypes.bfloat16: "bf16.1k" if cdna else "bf16"}
   # https://github.com/llvm/llvm-project/blob/main/clang/test/CodeGenOpenCL/builtins-amdgcn-mfma.cl
+  N,M,K = wmma.arg[1]
   if cdna:
+    if K == 32: dt_map.update({dtypes.half: ".f16", dtypes.bfloat16: ".bf16"})
     return f"  {ctx[wmma]} = call {ldt(wmma.dtype)} @llvm.amdgcn.mfma.{dt_map[wmma.src[-1].dtype.scalar()]}" + \
-           f".16x16x16{dt_map[wmma.src[0].dtype.scalar()]}(" + ", ".join([f"{ldt(w.dtype)} {ctx[w]}" for w in wmma.src]) + ", i32 0, i32 0, i32 0)"
+           f".{N}x{M}x{K}{dt_map[wmma.src[0].dtype.scalar()]}(" + ", ".join([f"{ldt(w.dtype)} {ctx[w]}" for w in wmma.src]) + ", i32 0, i32 0, i32 0)"
   # https://github.com/llvm/llvm-project/blob/main/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.wmma_32.ll
   # example: %wmma0 = call <8 x float> @llvm.amdgcn.wmma.f32.16x16x16.f16(<16 x half> %v99,<16 x half> %v100,<8 x float> %v101)
   return f"  {ctx[wmma]} = call {ldt(wmma.dtype)} @llvm.amdgcn.wmma.{dt_map[wmma.src[-1].dtype.scalar()]}.16x16x16." + \

tinygrad/renderer/ptx.py

@@ -49,21 +49,23 @@ ptx_matcher = PatternMatcher([
    lambda x: UOp(x.op, dtypes.uint8, x.src[0:1] + ((x.src[1].cast(dtypes.uint8),) if len(x.src) >= 2 else ()) + x.src[2:]).cast(dtypes.bool)),
   (UPat(Ops.STORE, src=(UPat(dtype=dtypes.int64), UPat(dtype=dtypes.bool)), name="x", allow_any_len=True),
    lambda x: UOp(x.op, dtypes.void, x.src[0:1] + (x.src[1].cast(dtypes.uint8),) + x.src[2:])),
+  # indexing on PTX is in uint64, we do the math while it's still in the graph
+  (UPat(Ops.INDEX, src=(UPat.var("buf"), UPat.var("idx")), name="op", allow_any_len=True), lambda buf,idx,op:
+    UOp(Ops.INDEX, dtype=dtypes.int64, src=(buf, buf.cast(dtypes.int64)+idx.cast(dtypes.int64)*buf.dtype.itemsize)+op.src[2:]) \
+      if op.dtype != dtypes.int64 and buf.dtype.addrspace != AddrSpace.REG else None),
   # load/store use pointer arithmetic, and the cast does nothing
-  (UPat(Ops.INDEX, src=(UPat.var("buf"), UPat.var("idx"))),
-   lambda buf,idx: (buf.cast(dtypes.int64) + idx.cast(dtypes.int64)*buf.dtype.itemsize) if buf.dtype.addrspace != AddrSpace.REG else None),
   (UPat(Ops.CAST, name="x"), lambda x: x.src[0] if isinstance(x.dtype, PtrDType) or x.src[0].dtype == dtypes.void else None),
-  # move mask from INDEX to the load/store to enable pointer arithmetic
-  (UPat(Ops.LOAD, src=(UPat(Ops.INDEX, src=(UPat.var("buf"), UPat.var("idx"), UPat.var("gate"))), UPat.var("alt")), allow_any_len=True, name="l"),
-   lambda buf,idx,gate,alt,l: UOp(Ops.LOAD, alt.dtype, (buf.index(idx), alt, gate, *l.src[2:]))),
-  (UPat(Ops.STORE, src=(UPat(Ops.INDEX, src=(UPat.var("buf"), UPat.var("idx"), UPat())), UPat.var("val"), UPat.var("gate")), allow_any_len=True),
-   lambda buf,idx,val,gate: UOp.store(buf.index(idx), val, gate)),
   # ptx shr and shl instructions require y to be uint
   (UPat.var("x") << UPat.var("y"), lambda x,y: UOp(Ops.SHL, x.dtype, (x,y.cast(dtypes.uint))) if y.dtype != dtypes.uint else None),
   (UPat.var("x") >> UPat.var("y"), lambda x,y: UOp(Ops.SHR, x.dtype, (x,y.cast(dtypes.uint))) if y.dtype != dtypes.uint else None),
 ])
 
-def mem_type(x: UOp): return 'shared' if any(_x.op is Ops.DEFINE_LOCAL for _x in x.src[0].toposort()) else 'global'
+def mem_type(x:UOp) -> str:
+  match x.op:
+    case Ops.AFTER: return mem_type(x.src[0])
+    case Ops.DEFINE_LOCAL: return 'shared'
+    case Ops.DEFINE_GLOBAL: return 'global'
+    case _: raise RuntimeError(f"{x.op} needs to be memory")
 
 def render_wmma(ctx: "PTXRenderer", wmma: UOp):
   assert ctx.wmma_r, "registry values for wmma must be populated"
@@ -88,9 +90,6 @@ def modifier(a: DType, b: DType): return '.rzi' if dtypes.is_int(a) and dtypes.i
 string_rewrite = PatternMatcher([
   (UPat.cvar("x", dtypes.bool), lambda ctx, x: f"setp.ne.s16 {ctx.r[x]}, {render_val(x.arg, x.dtype)}, 0;"),
   (UPat.cvar("x"), lambda ctx, x: f"mov.b{ctx.types[x.dtype][1:]} {ctx.r[x]}, {render_val(x.arg, x.dtype)};"),
-  (UPat(Ops.STORE, name="x", src=(UPat.var('bidx'), UPat.var("var")), allow_any_len=True), lambda ctx, x, bidx, var: f"st.{mem_type(bidx)}" + \
-    f"{f'.v{cnt}' if ((cnt:=var.dtype.count)>1) else ''}.{ctx.mem_types[var.dtype.scalar()]} " + \
-    f"[{ctx.r[bidx]}+0], {('{' + ', '.join(ctx.r[var]) + '}') if var.dtype.count > 1 else ctx.r[var]};"),
   (UPat(Ops.SPECIAL, name="x"), lambda ctx,x: f"mov.u32 %{x.arg}, %{'ctaid' if x.arg[0] == 'g' else 'tid'}.{chr(120+int(x.arg[-1]))};"),
   (UPat(Ops.DEFINE_GLOBAL, name="x"), lambda ctx, x: f"ld.param.{ctx.types[dtypes.ulong]} {ctx.r[x]}, [data{x.arg}+0];"),
   (UPat((Ops.CMPLT, Ops.CMPNE, Ops.CMPEQ), name="x", allow_any_len=True, src=(UPat.var("src0"),)),
@@ -103,16 +102,22 @@ string_rewrite = PatternMatcher([
    lambda ctx, x, a: f"setp.ne.b{ctx.types[a.dtype][1:]} {ctx.r[x]}, {ctx.r[a]}, {render_val(0, a.dtype)};"),
   (UPat(Ops.CAST, name="x", src=(UPat.var("a"),)),
    lambda ctx, x, a: f"cvt{modifier(x.dtype, a.dtype)}.{ctx.cast_types[x.dtype]}.{ctx.cast_types[a.dtype]} {ctx.r[x]}, {ctx.r[a]};"),
-  (UPat(Ops.LOAD, name="x", src=(UPat.var('loc'), UPat(name='alt'), UPat(name="gate", op=GroupOp.ALU)), allow_any_len=True),
-    lambda ctx, x, loc, alt, gate: flatten([
+  # store / gated load / load
+  (UPat(Ops.STORE, src=(UPat(Ops.INDEX, src=(UPat.var("buf"), UPat.var("loc")), allow_any_len=True), UPat.var("var"))),
+   lambda ctx, loc, var, buf: f"st.{mem_type(buf)}" + \
+    f"{f'.v{cnt}' if ((cnt:=var.dtype.count)>1) else ''}.{ctx.mem_types[var.dtype.scalar()]} " + \
+    f"[{ctx.r[loc]}+0], {('{' + ', '.join(ctx.r[var]) + '}') if var.dtype.count > 1 else ctx.r[var]};"),
+  (UPat(Ops.LOAD, name="x", src=(UPat(Ops.INDEX, src=(UPat.var("buf"), UPat.var("loc"), UPat.var("gate"))), UPat.var("alt")), allow_any_len=True),
+    lambda ctx, x, loc, alt, gate, buf: flatten([
     [f"mov.{ctx.mem_types[x.dtype.scalar()]} {v}, {render_val(0, x.dtype.scalar())};" for v in ctx.r[x]],
-    [f"@{ctx.r[gate]} ld.{mem_type(x)}.v{x.dtype.count}.{ctx.mem_types[x.dtype.scalar()]} {{{', '.join(ctx.r[x])}}}, [{ctx.r[loc]}+0];"]
+    [f"@{ctx.r[gate]} ld.{mem_type(buf)}.v{x.dtype.count}.{ctx.mem_types[x.dtype.scalar()]} {{{', '.join(ctx.r[x])}}}, [{ctx.r[loc]}+0];"]
   ]) if alt.dtype.count > 1 else [
-    f"@{ctx.r[gate]} ld.{mem_type(x)}.{ctx.mem_types[x.dtype.scalar()]} {ctx.r[x]}, [{ctx.r[loc]}+0];",
+    f"@{ctx.r[gate]} ld.{mem_type(buf)}.{ctx.mem_types[x.dtype.scalar()]} {ctx.r[x]}, [{ctx.r[loc]}+0];",
     f"@!{ctx.r[gate]} mov.b{ctx.types[x.dtype.scalar()][1:]} {ctx.r[x]}, {ctx.r[alt]};"]),
-  (UPat(Ops.LOAD, name="x", src=(UPat.var('loc'),), allow_any_len=True),
-   lambda ctx, x, loc: f"ld.{mem_type(x)}.v{x.dtype.count}.{ctx.mem_types[x.dtype.scalar()]} {{{', '.join(ctx.r[x])}}}, [{ctx.r[loc]}+0];" \
-     if x.dtype.count > 1 else f"ld.{mem_type(x)}.{ctx.mem_types[x.dtype]} {ctx.r[x]}, [{ctx.r[loc]}+0];"),
+  (UPat(Ops.LOAD, name="x", src=(UPat(Ops.INDEX, src=(UPat.var("buf"), UPat.var("loc"))),), allow_any_len=True),
+    lambda ctx, x, loc, buf: f"ld.{mem_type(buf)}.v{x.dtype.count}.{ctx.mem_types[x.dtype.scalar()]} {{{', '.join(ctx.r[x])}}}, [{ctx.r[loc]}+0];" \
+     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="x"), lambda ctx, x: [f"mov.u32 {ctx.r[x]}, 0;", "LOOP_" + f"{ctx.r[x][1:]}:"]),
   (UPat(Ops.END, name="x", src=(UPat.var("src0"),), allow_any_len=True), lambda ctx, x, src0: [
@@ -207,6 +212,7 @@ class PTXRenderer(Renderer):
             typ = "pred" if u.src[1].dtype == dtypes.bool else ("b"+self.types[u.src[1].dtype][1:])
             kernel.append(f"mov.{typ} {self.r[u.src[0]]}, {self.r[u.src[1]]};")
         continue
+      if u.op is Ops.INDEX: continue  # other index we can skip
       if u.op is Ops.SPECIAL: r[u] = "%" + u.arg
       elif u.op is Ops.DEFINE_VAR: bufs.append((u.arg[0], u.dtype))
       elif u.op is Ops.LOAD:

tinygrad/runtime/autogen/mesa.py

@@ -7,14 +7,11 @@
 # LONGDOUBLE_SIZE is: 16
 #
 import ctypes, ctypes.util, os, gzip, base64, subprocess, tinygrad.helpers as helpers
-def brew_path(nm):
-  try: return f"{subprocess.check_output(['brew', '--prefix', nm]).decode().strip()}/lib/lib{nm}.dylib"
-  except Exception: return 'failed'
 PATHS_TO_TRY = [
   (BASE:=os.getenv('MESA_PATH', f"/usr{'/local/' if helpers.OSX else '/'}lib"))+'/libtinymesa_cpu'+(EXT:='.dylib' if helpers.OSX else '.so'),
   f'{BASE}/libtinymesa{EXT}',
-  brew_path('tinymesa_cpu'),
-  brew_path('tinymesa'),
+  '/opt/homebrew/lib/libtinymesa_cpu.dylib',
+  '/opt/homebrew/lib/libtinymesa.dylib',
 ]
 def _try_dlopen_tinymesa_cpu():
   library = ctypes.util.find_library("tinymesa_cpu")

tinygrad/runtime/ops_amd.py

@@ -15,11 +15,12 @@ from tinygrad.runtime.autogen.am import am
 from tinygrad.runtime.support.compiler_amd import HIPCompiler, AMDLLVMCompiler
 from tinygrad.runtime.support.elf import elf_loader
 from tinygrad.runtime.support.am.amdev import AMDev, AMMemoryManager
-from tinygrad.runtime.support.amd import AMDReg, AMDIP, import_module, import_soc, setup_pci_bars
+from tinygrad.runtime.support.amd import AMDReg, AMDIP, import_module, import_soc, import_ip_offsets, setup_pci_bars
 from tinygrad.runtime.support.system import System, PCIIfaceBase, PCIAllocationMeta, MAP_FIXED, MAP_NORESERVE
 from tinygrad.runtime.support.usb import ASM24Controller, USBMMIOInterface
 if getenv("IOCTL"): import extra.hip_gpu_driver.hip_ioctl  # noqa: F401 # pylint: disable=unused-import
 
+SQTT = getenv("SQTT", 0)
 EVENT_INDEX_PARTIAL_FLUSH = 4 # based on a comment in nvd.h
 WAIT_REG_MEM_FUNCTION_EQ  = 3 # ==
 WAIT_REG_MEM_FUNCTION_NEQ = 4 # !=
@@ -66,11 +67,15 @@ class AMDComputeQueue(HWQueue):
     if self.dev.xccs > 1:
       self._q[prev_len-1] |= (len(self._q) - prev_len)
 
-  def wait_reg_mem(self, value, mask=0xffffffff, mem=None, reg_req=None, reg_done=None):
-    wrm_info_dw = self.pm4.WAIT_REG_MEM_MEM_SPACE(int(mem is not None)) | self.pm4.WAIT_REG_MEM_OPERATION(int(mem is None)) \
-                | self.pm4.WAIT_REG_MEM_FUNCTION(WAIT_REG_MEM_FUNCTION_GEQ) | self.pm4.WAIT_REG_MEM_ENGINE(0)
+  def set_grbm_broadcast(self):
+    self.wreg(self.gc.regGRBM_GFX_INDEX, **{f'{f}_broadcast_writes': 1 for f in ['se', 'sh' if self.dev.target[0] == 9 else 'sa', 'instance']})
+  def set_grbm_se(self, se): self.wreg(self.gc.regGRBM_GFX_INDEX, se_index=se, instance_broadcast_writes=1)
 
-    self.pkt3(self.pm4.PACKET3_WAIT_REG_MEM, wrm_info_dw, *(data64_le(mem) if mem is not None else (reg_req, reg_done)), value, mask, 4)
+  def wait_reg_mem(self, value, mask=0xffffffff, mem=None, reg=None, reg_done=0, op=WAIT_REG_MEM_FUNCTION_GEQ):
+    wrm_info_dw = self.pm4.WAIT_REG_MEM_MEM_SPACE(int(mem is not None)) | self.pm4.WAIT_REG_MEM_OPERATION(int(mem is None and reg_done > 0)) \
+                | self.pm4.WAIT_REG_MEM_FUNCTION(op) | self.pm4.WAIT_REG_MEM_ENGINE(0)
+
+    self.pkt3(self.pm4.PACKET3_WAIT_REG_MEM, wrm_info_dw, *(data64_le(mem) if mem is not None else (reg, reg_done)), value, mask, 4)
 
   def acquire_mem(self, addr=0x0, sz=(1 << 64)-1, gli=1, glm=1, glk=1, glv=1, gl1=1, gl2=1):
     if self.dev.target >= (10,0,0):
@@ -113,7 +118,7 @@ class AMDComputeQueue(HWQueue):
 
   def memory_barrier(self):
     pf = '' if self.nbio.version[0] == 2 else '0' if self.nbio.version[:2] != (7, 11) else '1'
-    self.wait_reg_mem(reg_req=getattr(self.nbio, f'regBIF_BX_PF{pf}_GPU_HDP_FLUSH_REQ').addr[0],
+    self.wait_reg_mem(reg=getattr(self.nbio, f'regBIF_BX_PF{pf}_GPU_HDP_FLUSH_REQ').addr[0],
                       reg_done=getattr(self.nbio, f'regBIF_BX_PF{pf}_GPU_HDP_FLUSH_DONE').addr[0], value=0xffffffff)
     self.acquire_mem()
     return self
@@ -124,6 +129,22 @@ class AMDComputeQueue(HWQueue):
 
   ### SQTT ###
 
+  def sqtt_setup_exec(self, prg, global_size):
+    self.sqtt_userdata(sqtt.struct_rgp_sqtt_marker_pipeline_bind(
+      _0=sqtt.union_rgp_sqtt_marker_pipeline_bind_0(_0=sqtt.struct_rgp_sqtt_marker_pipeline_bind_0_0(
+        identifier=sqtt.RGP_SQTT_MARKER_IDENTIFIER_BIND_PIPELINE, bind_point=(__BIND_POINT_COMPUTE:=1))),
+      _1=sqtt.union_rgp_sqtt_marker_pipeline_bind_1(api_pso_hash=data64_le(prg.libhash[0]))))
+
+    self.sqtt_userdata(sqtt.struct_rgp_sqtt_marker_event(
+      _0=sqtt.union_rgp_sqtt_marker_event_0(_0=sqtt.struct_rgp_sqtt_marker_event_0_0(has_thread_dims=1)),
+      _2=sqtt.union_rgp_sqtt_marker_event_2(cmd_id=next(prg.dev.sqtt_next_cmd_id))), *global_size)
+
+    for xcc in range(self.dev.xccs):
+      with self.pred_exec(xcc_mask=1 << xcc):
+        for i in range(8 if prg.dev.target >= (11,0,0) else 4):
+          self.wreg(getattr(self.gc, f'regCOMPUTE_STATIC_THREAD_MGMT_SE{i}'),
+            ((prg.dev.sqtt_itrace_se_mask >> ((self.dev.se_cnt // self.dev.xccs) * xcc + i)) & 0b1) if SQTT >= 2 else 0xffffffff)
+
   def sqtt_userdata(self, data, *extra_dwords):
     data_ints = [x[0] for x in struct.iter_unpack('<I', bytes(data))] + list(extra_dwords)
     for i in range(0, len(data_ints), 2):
@@ -134,87 +155,97 @@ class AMDComputeQueue(HWQueue):
     self.wreg(self.gc.regSQ_THREAD_TRACE_CTRL, draw_event_en=1, spi_stall_en=1, sq_stall_en=1, reg_at_hwm=2, hiwater=1, util_timer=1,
       mode=int(tracing), **trace_ctrl)
 
-  # Magic values from mesa/src/amd/vulkan/radv_sqtt.c:radv_emit_spi_config_cntl and src/amd/common/ac_sqtt.c:ac_sqtt_emit_start
   def sqtt_start(self, buf0s:list[HCQBuffer], se_mask:int):
     self.memory_barrier()
-    self.spi_config(tracing=True)
-    # One buffer for one SE, mesa does it with a single buffer and ac_sqtt_get_data_offset, but this is simpler and should work just as well
-    for se in range(len(buf0s)):
-      self.wreg(self.gc.regGRBM_GFX_INDEX, se_index=se, instance_broadcast_writes=1)
-      buf0_lo, buf0_hi = data64_le(buf0s[se].va_addr >> 12)
-      if self.dev.target >= (12,0,0):
-        self.wreg(self.gc.regSQ_THREAD_TRACE_BUF0_SIZE, size=buf0s[se].size >> 12)
-        self.wreg(self.gc.regSQ_THREAD_TRACE_BUF0_BASE_LO, base_lo=buf0_lo)
-        self.wreg(self.gc.regSQ_THREAD_TRACE_BUF0_BASE_HI, base_hi=buf0_hi)
-      else:
-        self.wreg(self.gc.regSQ_THREAD_TRACE_BUF0_SIZE, base_hi=buf0_hi, size=buf0s[se].size >> 12)
-        self.wreg(self.gc.regSQ_THREAD_TRACE_BUF0_BASE, base_lo=buf0_lo)
-      # NOTE: SQTT can only trace instructions on one simd per se, this selects first simd in first wgp in first sa.
-      # For RGP to display instruction trace it has to see it on first SE. Howerver ACE/MEC/whatever does the dispatching starting with second se,
-      # and on amdgpu/non-AM it also does weird things with dispatch order inside se: around 7 times out of 10 it starts from the last cu, but
-      # sometimes not, especially if the kernel has more than one wavefront which means that kernels with small global size might get unlucky and
-      # be dispatched on something else and not be seen in instruction tracing tab. You can force the wavefronts of a kernel to be dispatched on the
-      # CUs you want to by disabling other CUs via bits in regCOMPUTE_STATIC_THREAD_MGMT_SE<x> and trace even kernels that only have one wavefront.
-      cs_wtype = (1 << 6) if self.dev.target >= (12,0,0) else self.soc.SQ_TT_WTYPE_INCLUDE_CS_BIT
-      self.wreg(self.gc.regSQ_THREAD_TRACE_MASK, wtype_include=cs_wtype, simd_sel=0, wgp_sel=0, sa_sel=0)
-      reg_include = self.soc.SQ_TT_TOKEN_MASK_SQDEC_BIT | self.soc.SQ_TT_TOKEN_MASK_SHDEC_BIT | self.soc.SQ_TT_TOKEN_MASK_GFXUDEC_BIT | \
-                    self.soc.SQ_TT_TOKEN_MASK_COMP_BIT | self.soc.SQ_TT_TOKEN_MASK_CONTEXT_BIT
-      token_exclude = (1 << self.soc.SQ_TT_TOKEN_EXCLUDE_PERF_SHIFT) if self.dev.target < (12,0,0) else 0
+    if self.dev.target[0] == 9:
+      self.set_grbm_broadcast()
+      self.wreg(self.gc.regSQ_THREAD_TRACE_MASK, simd_en=0xf, cu_sel=0, sq_stall_en=1, spi_stall_en=1, reg_stall_en=1, vm_id_mask=0)
+      for se in range(len(buf0s)):
+        mask = (__SQTT_MISC:=1<<0) | (__SQTT_TIME:=1<<1) | (__SQTT_REG:=1<<2) | (__SQTT_WAVE_START:=1<<3) | (__SQTT_WAVE_END:=1<<6) \
+             | (__SQTT_USERDATA:=1<<12) | (__SQTT_REG_CS:=1<<5) | (__SQTT_REG_CS_PRIV:=1<<15)
+        if (se_mask >> se) & 0b1: mask |= (__SQTTINST:=1<<10) | (__SQTT_INST_PC:=1<<11) | (__SQTT_ISSUE:=1<<13)
 
-      # disable tracing
-      if not (se_mask >> se) & 0b1:
-        # gfx12 doesn't have enums with all fields, so it's hardcoded, but it's the same as gfx11.
-        token_exclude |= (1 << self.soc.SQ_TT_TOKEN_EXCLUDE_VMEMEXEC_SHIFT | 1 << self.soc.SQ_TT_TOKEN_EXCLUDE_ALUEXEC_SHIFT | \
-                         1 << self.soc.SQ_TT_TOKEN_EXCLUDE_VALUINST_SHIFT | 1 << self.soc.SQ_TT_TOKEN_EXCLUDE_IMMEDIATE_SHIFT | \
-                         1 << self.soc.SQ_TT_TOKEN_EXCLUDE_INST_SHIFT) if self.dev.target < (12,0,0) else 0x927
+        with self.pred_exec(xcc_mask=1<<(se // (ses_per_xcc:=(self.dev.se_cnt // self.dev.xccs)))):
+          self.set_grbm_se(se % ses_per_xcc)
+          self.wreg(self.gc.regSQ_THREAD_TRACE_TOKEN_MASK, reg_mask=0xf, token_mask=mask)
+          self.wreg(self.gc.regSQ_THREAD_TRACE_TOKEN_MASK2, inst_mask=0xffffffff)
+          self.wreg(self.gc.regSQ_THREAD_TRACE_BASE, addr=lo32(buf0s[se].va_addr >> 12))
+          self.wreg(self.gc.regSQ_THREAD_TRACE_BASE2, addr_hi=hi32(buf0s[se].va_addr >> 12))
+          self.wreg(self.gc.regSQ_THREAD_TRACE_SIZE, size=buf0s[se].size >> 12)
+          self.wreg(self.gc.regSQ_THREAD_TRACE_CTRL, reset_buffer=1)
+          self.wreg(self.gc.regSQ_THREAD_TRACE_MODE, mask_cs=1, autoflush_en=1, mode=1)
+    else:
+      self.spi_config(tracing=True)
+      # One buffer for one SE, mesa does it with a single buffer and ac_sqtt_get_data_offset, but this is simpler and should work just as well
+      for se in range(len(buf0s)):
+        self.set_grbm_se(se)
 
-      token_mask = {} if self.dev.target < (12,0,0) else {'exclude_barrier_wait': 1}
-      self.wreg(self.gc.regSQ_THREAD_TRACE_TOKEN_MASK, reg_include=reg_include, token_exclude=token_exclude, bop_events_token_include=1, **token_mask)
-      # Enable SQTT
-      self.sqtt_config(tracing=True)
-    # Restore global broadcasting
-    self.wreg(self.gc.regGRBM_GFX_INDEX, se_broadcast_writes=1, sa_broadcast_writes=1, instance_broadcast_writes=1)
-    self.wreg(self.gc.regCOMPUTE_THREAD_TRACE_ENABLE, 1)
+        buf0_lo, buf0_hi = data64_le(buf0s[se].va_addr >> 12)
+        if self.dev.target >= (12,0,0):
+          self.wreg(self.gc.regSQ_THREAD_TRACE_BUF0_SIZE, size=buf0s[se].size >> 12)
+          self.wreg(self.gc.regSQ_THREAD_TRACE_BUF0_BASE_LO, base_lo=buf0_lo)
+          self.wreg(self.gc.regSQ_THREAD_TRACE_BUF0_BASE_HI, base_hi=buf0_hi)
+        else:
+          self.wreg(self.gc.regSQ_THREAD_TRACE_BUF0_SIZE, base_hi=buf0_hi, size=buf0s[se].size >> 12)
+          self.wreg(self.gc.regSQ_THREAD_TRACE_BUF0_BASE, base_lo=buf0_lo)
+        # NOTE: SQTT can only trace instructions on one simd per se, this selects first simd in first wgp in first sa.
+        # For RGP to display instruction trace it has to see it on first SE. Howerver ACE/MEC/whatever does the dispatching starting with second se,
+        # and on amdgpu/non-AM it also does weird things with dispatch order inside se: around 7 times out of 10 it starts from the last cu, but
+        # sometimes not, especially if the kernel has more than one wavefront which means that kernels with small global size might get unlucky and
+        # be dispatched on something else and not be seen in instruction tracing tab. You can force the wavefronts of a kernel to be dispatched on the
+        # CUs you want to by disabling other CUs via bits in regCOMPUTE_STATIC_THREAD_MGMT_SE<x> and trace even kernels that only have one wavefront.
+        cs_wtype = (1 << 6) if self.dev.target >= (12,0,0) else self.soc.SQ_TT_WTYPE_INCLUDE_CS_BIT
+        self.wreg(self.gc.regSQ_THREAD_TRACE_MASK, wtype_include=cs_wtype, simd_sel=0, wgp_sel=0, sa_sel=0)
+        reg_include = self.soc.SQ_TT_TOKEN_MASK_SQDEC_BIT | self.soc.SQ_TT_TOKEN_MASK_SHDEC_BIT | self.soc.SQ_TT_TOKEN_MASK_GFXUDEC_BIT | \
+                      self.soc.SQ_TT_TOKEN_MASK_COMP_BIT | self.soc.SQ_TT_TOKEN_MASK_CONTEXT_BIT
+        token_exclude = (1 << self.soc.SQ_TT_TOKEN_EXCLUDE_PERF_SHIFT) if self.dev.target < (12,0,0) else 0
+
+        # disable instr tracing
+        if not (se_mask >> se) & 0b1:
+          # gfx12 doesn't have enums with all fields, so it's hardcoded, but it's the same as gfx11.
+          token_exclude |= (1 << self.soc.SQ_TT_TOKEN_EXCLUDE_VMEMEXEC_SHIFT | 1 << self.soc.SQ_TT_TOKEN_EXCLUDE_ALUEXEC_SHIFT | \
+                            1 << self.soc.SQ_TT_TOKEN_EXCLUDE_VALUINST_SHIFT | 1 << self.soc.SQ_TT_TOKEN_EXCLUDE_IMMEDIATE_SHIFT | \
+                            1 << self.soc.SQ_TT_TOKEN_EXCLUDE_INST_SHIFT) if self.dev.target < (12,0,0) else 0x927
+
+        self.wreg(self.gc.regSQ_THREAD_TRACE_TOKEN_MASK, reg_include=reg_include, token_exclude=token_exclude, bop_events_token_include=1,
+                  **({} if self.dev.target < (12,0,0) else {'exclude_barrier_wait': 1}))
+        self.sqtt_config(tracing=True)
+
+    self.set_grbm_broadcast()
+    if self.dev.target[0] > 9: self.wreg(self.gc.regCOMPUTE_THREAD_TRACE_ENABLE, 1)
     self.memory_barrier()
     return self
 
   # Magic values from src/amd/common/ac_sqtt.c:ac_sqtt_emit_stop and src/amd/common/ac_sqtt.c:ac_sqtt_emit_wait
-  def sqtt_stop(self, ses: int, wptrs: HCQBuffer):
+  def sqtt_stop(self, ses:int, wptrs:HCQBuffer):
     self.memory_barrier()
+    self.set_grbm_broadcast()
+
     # Start shutting everything down
-    self.wreg(self.gc.regCOMPUTE_THREAD_TRACE_ENABLE, 0)
-    self.pkt3(self.pm4.PACKET3_EVENT_WRITE, self.pm4.EVENT_TYPE(self.soc.THREAD_TRACE_FINISH) | self.pm4.EVENT_INDEX(0))
+    if self.dev.target[0] == 9: self.wreg(self.gc.regSQ_THREAD_TRACE_MODE, mask_cs=1, autoflush_en=1, mode=0)
+    else:
+      self.wreg(self.gc.regCOMPUTE_THREAD_TRACE_ENABLE, 0)
+      self.pkt3(self.pm4.PACKET3_EVENT_WRITE, self.pm4.EVENT_TYPE(self.soc.THREAD_TRACE_FINISH) | self.pm4.EVENT_INDEX(0))
+
     # For each SE wait for finish to complete and copy regSQ_THREAD_TRACE_WPTR to know where in the buffer trace data ends
     for se in range(ses):
-      self.wreg(self.gc.regGRBM_GFX_INDEX, se_index=se, instance_broadcast_writes=1)
-      # Wait for FINISH_PENDING==0
-      self.pkt3(self.pm4.PACKET3_WAIT_REG_MEM, self.pm4.WAIT_REG_MEM_FUNCTION(WAIT_REG_MEM_FUNCTION_EQ),
-                self.gc.regSQ_THREAD_TRACE_STATUS.addr[0], 0, 0, self.gc.regSQ_THREAD_TRACE_STATUS.fields_mask('finish_pending'), 4)
-      # Disable SQTT
-      self.sqtt_config(tracing=False)
-      # Wait for BUSY==0
-      self.pkt3(self.pm4.PACKET3_WAIT_REG_MEM, self.pm4.WAIT_REG_MEM_FUNCTION(WAIT_REG_MEM_FUNCTION_EQ),
-                self.gc.regSQ_THREAD_TRACE_STATUS.addr[0], 0, 0, self.gc.regSQ_THREAD_TRACE_STATUS.fields_mask('busy'), 4)
+      self.set_grbm_se(se)
+
+      status_reg = self.gc.regSQ_THREAD_TRACE_STATUS.addr[0] - (self.pm4.PACKET3_SET_UCONFIG_REG_START if self.dev.target[0] == 9 else 0)
+      if self.dev.target >= (10, 0, 0):
+        self.wait_reg_mem(reg=status_reg, mask=self.gc.regSQ_THREAD_TRACE_STATUS.fields_mask('finish_pending'), op=WAIT_REG_MEM_FUNCTION_EQ, value=0)
+        self.sqtt_config(tracing=False)
+      self.wait_reg_mem(reg=status_reg, mask=self.gc.regSQ_THREAD_TRACE_STATUS.fields_mask('busy'), op=WAIT_REG_MEM_FUNCTION_EQ, value=0)
+      self.pkt3(self.pm4.PACKET3_EVENT_WRITE, self.pm4.EVENT_TYPE(self.soc.CS_PARTIAL_FLUSH) | self.pm4.EVENT_INDEX(EVENT_INDEX_PARTIAL_FLUSH))
+
       # Copy WPTR to memory (src_sel = perf, dst_sel = tc_l2, wr_confirm = True)
       self.pkt3(self.pm4.PACKET3_COPY_DATA, 1 << 20 | 2 << 8 | 4, self.gc.regSQ_THREAD_TRACE_WPTR.addr[0], 0, *data64_le(wptrs.va_addr+(se*4)))
-    # Restore global broadcasting
-    self.wreg(self.gc.regGRBM_GFX_INDEX, se_broadcast_writes=1, sa_broadcast_writes=1, instance_broadcast_writes=1)
-    self.spi_config(tracing=False)
+
+    self.set_grbm_broadcast()
+    if self.dev.target[0] > 9: self.spi_config(tracing=False)
     self.memory_barrier()
     return self
 
-  def sqtt_prg_marker(self, prg:AMDProgram, global_size:tuple[sint, ...]):
-    BIND_POINT_COMPUTE = 1
-
-    self.sqtt_userdata(sqtt.struct_rgp_sqtt_marker_pipeline_bind(
-      _0=sqtt.union_rgp_sqtt_marker_pipeline_bind_0(_0=sqtt.struct_rgp_sqtt_marker_pipeline_bind_0_0(
-        identifier=sqtt.RGP_SQTT_MARKER_IDENTIFIER_BIND_PIPELINE, bind_point=BIND_POINT_COMPUTE)),
-      _1=sqtt.union_rgp_sqtt_marker_pipeline_bind_1(api_pso_hash=data64_le(prg.libhash[0]))))
-
-    self.sqtt_userdata(sqtt.struct_rgp_sqtt_marker_event(
-      _0=sqtt.union_rgp_sqtt_marker_event_0(_0=sqtt.struct_rgp_sqtt_marker_event_0_0(has_thread_dims=1)),
-      _2=sqtt.union_rgp_sqtt_marker_event_2(cmd_id=next(prg.dev.sqtt_next_cmd_id))), *global_size)
-
   def exec(self, prg:AMDProgram, args_state:CLikeArgsState, global_size:tuple[sint, ...], local_size:tuple[sint, ...]):
     self.bind_args_state(args_state)
 
@@ -238,7 +269,7 @@ class AMDComputeQueue(HWQueue):
 
     user_regs += [*data64_le(args_state.buf.va_addr)]
 
-    if prg.dev.sqtt_enabled: self.sqtt_prg_marker(prg, global_size)
+    if prg.dev.sqtt_enabled: self.sqtt_setup_exec(prg, global_size)
 
     self.wreg(self.gc.regCOMPUTE_PGM_LO, *data64_le(prg.prog_addr >> 8))
     self.wreg(self.gc.regCOMPUTE_PGM_RSRC1, prg.rsrc1, prg.rsrc2)
@@ -254,13 +285,8 @@ class AMDComputeQueue(HWQueue):
     if (10,0,0) <= prg.dev.target < (11,0,0): self.wreg(self.gc.mmCP_COHER_START_DELAY, 0x20)
 
     self.wreg(self.gc.regCOMPUTE_RESTART_X, 0, 0, 0)
-    self.wreg(self.gc.regCOMPUTE_STATIC_THREAD_MGMT_SE0, 0xFFFFFFFF, 0xFFFFFFFF)
-    self.wreg(self.gc.regCOMPUTE_STATIC_THREAD_MGMT_SE2, 0xFFFFFFFF, 0xFFFFFFFF)
-    if prg.dev.target >= (11,0,0): self.wreg(self.gc.regCOMPUTE_STATIC_THREAD_MGMT_SE4, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF)
-
     self.wreg(self.gc.regCOMPUTE_USER_DATA_0, *user_regs)
     self.wreg(self.gc.regCOMPUTE_RESOURCE_LIMITS, 0)
-
     self.wreg(self.gc.regCOMPUTE_START_X, 0, 0, 0, *local_size, 0, 0)
 
     gfx10p = {'cs_w32_en': int(prg.wave32)} if prg.dev.target >= (10,0,0) else {}
@@ -321,6 +347,7 @@ class AMDComputeQueue(HWQueue):
 class AMDComputeAQLQueue(AMDComputeQueue):
   def exec(self, prg:AMDProgram, args_state:CLikeArgsState, global_size:tuple[sint, ...], local_size:tuple[sint, ...]):
     self.bind_args_state(args_state)
+    if prg.dev.sqtt_enabled: self.sqtt_setup_exec(prg, global_size)
     self._q.append(pkt:=hsa.hsa_kernel_dispatch_packet_t(header=AQL_HDR | (hsa.HSA_PACKET_TYPE_KERNEL_DISPATCH << hsa.HSA_PACKET_HEADER_TYPE),
       setup=3<<hsa.HSA_KERNEL_DISPATCH_PACKET_SETUP_DIMENSIONS, private_segment_size=prg.private_segment_size,
       group_segment_size=prg.group_segment_size, kernel_object=prg.aql_prog_addr, kernarg_address=args_state.buf.va_addr))
@@ -561,8 +588,6 @@ class KFDIface:
     id2ip = {am.GC_HWID: am.GC_HWIP, am.SDMA0_HWID: am.SDMA0_HWIP, am.NBIF_HWID: am.NBIF_HWIP}
     ip_hw = [(id2ip[int(hwid)], int(hwid)) for hwid in FileIOInterface(ip_base).listdir() if hwid.isnumeric() and int(hwid) in id2ip]
     self.ip_versions = {ip:tuple(int(FileIOInterface(f'{ip_base}/{hw}/0/{part}').read()) for part in ['major','minor','revision']) for ip,hw in ip_hw}
-    self.ip_offsets = {ip:{int(i):tuple(int(x, 16) for x in FileIOInterface(f'{ip_base}/{hw}/{i}/base_addr').read().splitlines())
-      for i in FileIOInterface(f'{ip_base}/{hw}').listdir()} for ip,hw in ip_hw }
     self.drm_fd = FileIOInterface(f"/dev/dri/renderD{self.props['drm_render_minor']}", os.O_RDWR)
 
     self.kfd_ver = ((ver_st:=kfd.AMDKFD_IOC_GET_VERSION(KFDIface.kfd)).major_version, ver_st.minor_version)
@@ -678,7 +703,7 @@ class PCIIface(PCIIfaceBase):
 
   def _setup_adev(self, name, vram:MMIOInterface, doorbell:MMIOInterface, mmio:MMIOInterface, dma_regions:list[tuple[int, MMIOInterface]]|None=None):
     self.dev_impl:AMDev = AMDev(name, vram, doorbell, mmio, dma_regions)
-    self.ip_offsets, self.ip_versions = self.dev_impl.regs_offset, self.dev_impl.ip_ver
+    self.ip_versions = self.dev_impl.ip_ver
 
     gfxver = int(f"{self.dev_impl.ip_ver[am.GC_HWIP][0]:02d}{self.dev_impl.ip_ver[am.GC_HWIP][1]:02d}{self.dev_impl.ip_ver[am.GC_HWIP][2]:02d}")
     array_count = self.dev_impl.gc_info.gc_num_sa_per_se * self.dev_impl.gc_info.gc_num_se
@@ -778,14 +803,15 @@ class AMDDevice(HCQCompiled):
     debug_memory_size = round_up((self.max_cu_id + 1 if self.target >= (10,1,0) else 1) * (self.max_wave_id + 1) * 32, 64)
     if self.target[0] == 10: ctl_stack_size = min(ctl_stack_size, 0x7000)
 
+    self.ip_off = import_ip_offsets(self.target)
     self.soc = import_soc(self.target)
     self.pm4 = importlib.import_module(f"tinygrad.runtime.autogen.am.pm4_{'nv' if self.target[0] >= 10 else 'soc15'}")
     self.sdma = import_module('sdma', min(self.iface.ip_versions[am.SDMA0_HWIP], (6, 0, 0)))
-    self.gc = AMDIP('gc', self.iface.ip_versions[am.GC_HWIP], self.iface.ip_offsets[am.GC_HWIP])
+    self.gc = AMDIP('gc', self.iface.ip_versions[am.GC_HWIP],
+                    bases={i: tuple(getattr(self.ip_off, f'GC_BASE__INST{i}_SEG{s}', 0) for s in range(6)) for i in range(6)})
 
-    nbio_name = 'nbio' if self.target[0] < 12 else 'nbif'
-    nbio_pad = (0,) if self.target[0] == 9 else ()
-    self.nbio = AMDIP(nbio_name, self.iface.ip_versions[am.NBIF_HWIP], {i:nbio_pad+x for i,x in self.iface.ip_offsets[am.NBIF_HWIP].items()})
+    self.nbio = AMDIP('nbio' if self.target[0] < 12 else 'nbif', self.iface.ip_versions[am.NBIF_HWIP],
+                      bases={i: tuple(getattr(self.ip_off, f'NBIO_BASE__INST{i}_SEG{s}', 0) for s in range(9)) for i in range(6)})
 
     self.is_aql = getenv("AMD_AQL", int(self.xccs > 1))
     if self.is_aql:
@@ -812,16 +838,16 @@ class AMDDevice(HCQCompiled):
     self._ensure_has_local_memory(128) # set default scratch size to 128 bytes per thread
 
     # SQTT is disabled by default because of runtime overhead and big file sizes (~200mb to Tensor.full() two 4096x4096 tensors and matmul them)
-    self.sqtt_enabled = PROFILE and bool(getenv("SQTT", 0))
+    self.sqtt_enabled = PROFILE and SQTT > 0
     if self.sqtt_enabled:
-      if self.target[0] < 11: raise RuntimeError(f'SQ Thread Tracing is not supported on gc:{self.target}')
+      if self.target[0] not in {9, 11, 12}: raise RuntimeError(f'SQ Thread Tracing is not supported on gc:{self.target}')
       if not self.is_am() and (ppfeaturemask:=int(FileIOInterface('/sys/module/amdgpu/parameters/ppfeaturemask', os.O_RDONLY).read(), 16))&0x8000:
         raise RuntimeError("SQTT can't be enabled because of hardware bug, to workaround either use AMD_IFACE=PCI or add "
                            f"ppfeaturemask={(ppfeaturemask&~0x8000):#x} (current {ppfeaturemask=:#x} & ~PP_GFXOFF_MASK) to amdgpu module parameters\n"
                            "For more information read https://github.com/tinygrad/tinygrad/blob/master/extra/sqtt/README.md")
       SQTT_BUFFER_SIZE = getenv("SQTT_BUFFER_SIZE", 256) # in mb, per shader engine
       self.sqtt_buffers = [self.allocator.alloc(SQTT_BUFFER_SIZE*1024*1024, BufferSpec(nolru=True)) for _ in range(self.se_cnt)]
-      self.sqtt_itrace_se_mask = getenv("SQTT_ITRACE_SE_MASK", 2) # -1 enable all, 0 disable all, >0 bitmask for where to enable instruction tracing
+      self.sqtt_itrace_se_mask = getenv("SQTT_ITRACE_SE_MASK", -1 if SQTT >= 2 else (1 << 1)) # se bitmask: -1 enable all, 0 disable all
       self.sqtt_next_cmd_id = itertools.count(0)
       cast(AMDComputeQueue, self.hw_compute_queue_t()).sqtt_start(self.sqtt_buffers, self.sqtt_itrace_se_mask).submit(self)
 
@@ -881,12 +907,13 @@ class AMDDevice(HCQCompiled):
       self.synchronize()
       if DEBUG >= 2: print(f'{self.device}: Saving SQTT in profile...')
       for i,buf0 in enumerate(self.sqtt_buffers):
-        wptr = ((wptrs_buf.cpu_view().view(fmt='I')[i] & 0x1FFFFFFF) - (((buf0.va_addr//32) & 0x1FFFFFFF) if self.target < (12,0,0) else 0)) * 32
+        wptr = ((wptrs_buf.cpu_view().view(fmt='I')[i] & 0x1FFFFFFF) - (((buf0.va_addr//32) & 0x1FFFFFFF) if self.target[0] == 11 else 0)) * 32
         if DEBUG >= 2: print(f'\t{self.device}: SE {i} blob size {wptr:#x}')
         assert wptr >= 0 and wptr <= buf0.size, f"{wptr} > {buf0.size}, should never happen"
         # When sqtt buffer overflows, wptr stops at the last dword
         if wptr >= buf0.size - 32:
           print(colored(f"{self.device}: Warning: SQTT buffer is full (SE {i})! Increase SQTT buffer with SQTT_BUFFER_SIZE=X (in MB)", "yellow"))
         self.allocator._copyout(sqtt_buf:=memoryview(bytearray(wptr)), buf0)
+        if self.target[0] == 9: sqtt_buf = memoryview(bytearray(b'\x11\x80\x1f\x00\x00\x00\x00\x00') + sqtt_buf)
         Compiled.profile_events += [ProfileSQTTEvent(self.device, i, self.iface.props, bytes(sqtt_buf), bool((self.sqtt_itrace_se_mask >> i) & 0b1))]
     super()._at_profile_finalize()

tinygrad/runtime/ops_cuda.py

@@ -5,7 +5,7 @@ from tinygrad.device import Compiled, BufferSpec, LRUAllocator, CompilerPairT
 from tinygrad.renderer.cstyle import CUDARenderer
 from tinygrad.renderer.ptx import PTXRenderer
 from tinygrad.runtime.autogen import cuda
-from tinygrad.runtime.support.compiler_cuda import pretty_ptx, CUDACompiler, PTXCompiler
+from tinygrad.runtime.support.compiler_cuda import pretty_ptx, CUDACompiler, PTXCompiler, NVCCCompiler
 if getenv("IOCTL"): import extra.nv_gpu_driver.nv_ioctl  # noqa: F401  # pylint: disable=unused-import
 if MOCKGPU:=getenv("MOCKGPU"): from test.mockgpu.cuda import cuda # type: ignore # pylint: disable=reimported
 
@@ -118,7 +118,8 @@ class CUDADevice(Compiled):
 
     from tinygrad.runtime.graph.cuda import CUDAGraph
     compilers:list[CompilerPairT] = [(functools.partial(CUDARenderer, self.arch), functools.partial(CUDACompiler, self.arch)),
-                                     (functools.partial(PTXRenderer, self.arch), functools.partial(PTXCompiler, self.arch))]
+                                     (functools.partial(PTXRenderer, self.arch), functools.partial(PTXCompiler, self.arch)),
+                                     (functools.partial(CUDARenderer, self.arch), functools.partial(NVCCCompiler, self.arch))]
     super().__init__(device, CUDAAllocator(self), compilers, functools.partial(CUDAProgram, self), None if MOCKGPU else CUDAGraph)
 
   def synchronize(self):

tinygrad/runtime/ops_python.py

@@ -150,10 +150,10 @@ class PythonProgram:
             def c_map(lane, elem): return (elem + ((lane%2)*2) + ((lane//8)%2)*4, ((lane//2)%4) + (lane//16)*4)
             ul[i] = wmma_helper(32, 8, 2, 2, 2, a_b_elem, a_b_elem, c_map)
           elif device == "AMD" and threads == 64:
-            def a_elem(x, k, row, goff): return x[k%4][goff + (k//4)*16 + row]
-            def b_elem(x, col, k, goff): return a_elem(x, k, col, goff) # pylint: disable=arguments-out-of-order
+            def a_elem(x, k, row, goff): return x[k%(dims[2]//4)][goff + (k//(dims[2]//4))*16 + row]
+            def b_elem(x, col, k, goff): return a_elem(x, k, col, goff)  # pylint: disable=arguments-out-of-order
             def c_map(lane, elem): return (lane%16, (lane//16)*4 + elem)
-            ul[i] = wmma_helper(64, 16, 4, 4, 4, a_elem, b_elem, c_map)
+            ul[i] = wmma_helper(64, dims[2], len(inp[0]), len(inp[1]), len(inp[2]), a_elem, b_elem, c_map)
           elif device == "AMD" and len(inp[0]) == 8: # RDNA4
             def a_elem(x, k, row, goff): return x[k - [0, 4, 4, 8][k//4]][goff + row + [0, 16, 0, 16][k//4]]
             def b_elem(x, col, k, goff): return a_elem(x, k, col, goff)
@@ -177,6 +177,11 @@ class PythonProgram:
               def b_elem(x, col, k, goff): return x[k%2 + (k//8)*2][goff + (k//2)%4 + col*4]
               ul[i] = wmma_helper(32, 16, 8, 4, 4, a_elem, b_elem, c_map)
 
+            elif dims == (8,16,32):
+              def a_elem(x, k, row, goff): return x[k%4 + (row//8)*4 + (k//16)*8][goff + (k//4)%4 + (row%8)*4]
+              def b_elem(x, col, k, goff): return x[k%4 + (k//16)*4][goff + (k//4)%4  + col*4]
+              ul[i] = wmma_helper(32, 32, 16, 8, 4, a_elem, b_elem, c_map)
+
             elif dims == (8,16,8) and dtype_in == dtypes.half:
               def a_elem(x, k, row, goff): return x[k%2 + (row//8)*2][goff + k//2 + (row%8)*4]
               def b_elem(x, col, k, goff): return x[k%2][goff + k//2 + col*4]
@@ -216,10 +221,11 @@ class PythonRenderer(Renderer):
     match cast(str, EMULATE.value):
       case "METAL": self.device, self.tensor_cores = "METAL", tc.metal
       case "AMD": self.device, self.tensor_cores = "AMD", tc.amd_rdna3
-      case "AMD_MFMA": self.device, self.tensor_cores = "AMD", tc.amd_cdna
+      case "AMD_MFMA": self.device, self.tensor_cores = "AMD", tc.amd_cdna4
       case "AMD_RDNA4": self.device, self.tensor_cores = "AMD", tc.amd_rdna4
       case "CUDA": self.device, self.tensor_cores = "CUDA", tc.cuda_sm80
       case "CUDA_SM75": self.device, self.tensor_cores = "CUDA", tc.cuda_sm75
+      case "CUDA_SM89": self.device, self.tensor_cores = "CUDA", tc.cuda_sm89
       case "INTEL": self.device, self.suffix, self.tensor_cores = "INTEL", "INTEL", tc.intel
       case "AMX": self.device, self.tensor_cores = "CPU", tc.amx
       case "": pass

tinygrad/runtime/support/am/ip.py

@@ -242,6 +242,7 @@ class AM_GFX(AM_IP):
       cp_mqd_control=self.adev.regCP_MQD_CONTROL.encode(priv_state=1), cp_hqd_vmid=0, cp_hqd_aql_control=int(aql),
       cp_hqd_eop_base_addr_lo=lo32(eop_addr>>8), cp_hqd_eop_base_addr_hi=hi32(eop_addr>>8),
       cp_hqd_eop_control=self.adev.regCP_HQD_EOP_CONTROL.encode(eop_size=(eop_size//4).bit_length()-2))
+    for se in range(8): setattr(mqd_struct, f'compute_static_thread_mgmt_se{se}', 0xffffffff)
 
     # Copy mqd into memory
     self.adev.vram.view(mqd.paddrs[0][0], ctypes.sizeof(mqd_struct))[:] = memoryview(mqd_struct).cast('B')

tinygrad/runtime/support/amd.py

@@ -49,7 +49,9 @@ def header_download(file, name=None, subdir="defines", url=None) -> str:
 
 def import_header(path:str, url=None):
   t = re.sub(r'//.*|/\*.*?\*/','', header_download(path, subdir="defines", url=url), flags=re.S)
-  return {k:int(v,0) for k,v in re.findall(r'\b([A-Za-z_]\w*)\s*=\s*(0x[0-9A-Fa-f]+|\d+)', t)}
+  # TODO: refactor when clang2py is replaced
+  return {k:int(v,0) for k,v in re.findall(r'\b([A-Za-z_]\w*)\s*=\s*(0x[0-9A-Fa-f]+|\d+)', t) + \
+                                re.findall(r'^\s*#\s*define\s+([A-Za-z_0-9]\w*)\s+(0x[0-9A-Fa-f]+|\d+)', t, re.M)}
 
 def import_module(name:str, version:tuple[int, ...], version_prefix:str=""):
   for ver in fixup_ip_version(name, version):
@@ -62,6 +64,8 @@ def import_soc(ip):
   url = "https://raw.githubusercontent.com/ROCm/rocm-systems/cccc350dc620e61ae2554978b62ab3532dc10bd9/projects"
   return type("SOC", (object,), import_header(f"aqlprofile/linux/{({9: 'vega10', 10: 'navi10', 11: 'soc21', 12: 'soc24'}[ip[0]])}_enum.h", url=url))
 
+def import_ip_offsets(ip): return type("IPOFF", (object,), import_header(f"include/{('sienna_cichlid' if ip[0] > 9 else 'vega20')}_ip_offset.h"))
+
 def import_asic_regs(prefix:str, version:tuple[int, ...], cls=AMDReg) -> dict[str, AMDReg]:
   def _split_name(name): return name[:(pos:=next((i for i,c in enumerate(name) if c.isupper()), len(name)))], name[pos:]
   def _extract_regs(txt):

tinygrad/runtime/support/compiler_cuda.py

@@ -60,6 +60,19 @@ class NVCompiler(CUDACompiler):
   def __init__(self, arch:str): super().__init__(arch, cache_key="nv")
   def compile(self, src:str) -> bytes: return self._compile_program(src, nvrtc.nvrtcGetCUBIN, nvrtc.nvrtcGetCUBINSize)
 
+class NVCCCompiler(Compiler):
+  def __init__(self, arch:str, extra_options:list[str]=[]):
+    self.arch, self.extra_options = arch, extra_options
+    super().__init__(f"compile_nvcc_{self.arch}_{hashlib.sha256(' '.join(extra_options).encode()).hexdigest()[:8]}")
+  def compile(self, src:str) -> bytes:
+    with tempfile.NamedTemporaryFile(suffix=".cu") as srcf, tempfile.NamedTemporaryFile(suffix=".ptx") as libf:
+      srcf.write(src.encode())
+      srcf.flush()
+      subprocess.run(["nvcc", f"-arch={self.arch}", "-ptx", "-o", libf.name, srcf.name] + self.extra_options,
+                 check=True)
+      return libf.read()
+  def disassemble(self, lib:bytes): cuda_disassemble(lib, self.arch)
+
 class PTXCompiler(Compiler):
   def __init__(self, arch:str, cache_key="ptx"):
     self.arch = arch

tinygrad/schedule/indexing.py

@@ -141,6 +141,7 @@ def apply_movement_op(op:Ops, in_shape:tuple[sint,...], arg:tuple, rngs:tuple[UO
 
 @profile_matches
 def run_rangeify(tsink:UOp, debug:bool=False) -> tuple[UOp, IndexingContext]:
+  if debug: print("**************************")
   rctx = IndexingContext()
 
   # get ops to realize

tinygrad/schedule/rangeify.py

@@ -4,7 +4,8 @@ from tinygrad.dtype import dtypes, PtrDType, ImageDType, AddrSpace
 from tinygrad.uop.ops import PatternMatcher, UPat, Ops, UOp, resolve, GroupOp, _substitute, ssimplify, KernelInfo
 from tinygrad.uop.ops import track_rewrites, graph_rewrite, identity_element, sint, AxisType, BottomUpGate
 from tinygrad.uop.symbolic import symbolic_flat
-from tinygrad.helpers import argsort, prod, all_same, pluralize, getenv, flatten, dedup, all_int, DEBUG, SPLIT_REDUCEOP, Metadata
+from tinygrad.helpers import argsort, prod, all_same, pluralize, getenv, flatten, dedup, all_int, DEBUG, SPLIT_REDUCEOP, Metadata, DEBUG_RANGEIFY
+from tinygrad.helpers import PCONTIG
 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, ALWAYS_CONTIGUOUS, IndexingContext, apply_movement_op
@@ -157,16 +158,19 @@ def remove_bufferize(src:UOp, buf:UOp, idx:UOp):
     accessed_buffers: list[UOp] = []
     reduces: list[UOp] = []
     def red_gate(x:UOp):
-      if x.op is Ops.INDEX:
+      if x.op is Ops.BUFFERIZE and x.arg.addrspace == AddrSpace.GLOBAL:
         accessed_buffers.append(x)
         return False
+      if x.op is Ops.BUFFER:
+        accessed_buffers.append(x)
       if x.op is Ops.REDUCE: reduces.append(x)
       return True
     src.toposort(gate=red_gate)
     del red_gate
+    accessed_buffers = dedup(accessed_buffers)
 
     # if this is generated from multiple buffers, don't remove this buffer
-    if len(dedup([x.src[0] for x in accessed_buffers])) > 2: return None
+    if len(accessed_buffers) > 2 and not (PCONTIG > 2): return None
 
     # if any reduces access a buffer, don't remove this buffer
     buffer_in_reduce = False
@@ -176,40 +180,54 @@ def remove_bufferize(src:UOp, buf:UOp, idx:UOp):
       return not buffer_in_reduce
     UOp.sink(*[x.src[0] for x in reduces]).toposort(gate=buf_gate)
     del buf_gate
-    if buffer_in_reduce: return None
+    if buffer_in_reduce:
+      if PCONTIG > 2:
+        out_in_ratio = (prod(buf.shape)+1) / (sum([x.size for x in accessed_buffers])+1)
+        if out_in_ratio < 10: return None
+      else:
+        return None
 
   # if it makes it here, the bufferize is removed
   # this is the ranges replaced
   # NOTE: if buf src is a const, we don't replace it
   return src.substitute({k:v for k,v in zip(buf.src[1:], idx.src[1:]) if k.op is not Ops.CONST}, extra_pm=pm_gate_substitute)
 
-def pre_bufferize(b:UOp, x:UOp, copy:UOp):
-  nb = b.replace(src=(b.src[0].contiguous(),)+b.src[1:])
-  return copy.replace(src=(x.replace(src=(nb,)+x.src[1:]), copy.src[1]))
+def remove_noop_bufferize(idx,b2):
+  if idx.src[1:] != b2.src[1:] or idx.src[0].op is Ops.BUFFER_VIEW: return None
+  new_tag = (idx.src[0].tag or ()) + (b2.tag or ()) or None
+  return idx.src[0].rtag(new_tag).shrink(tuple((0, s) for s in b2.shape)) if b2.shape else idx.src[0].rtag(new_tag)
 
-pm_cleanups = pm_mops+PatternMatcher([
+pm_const_buffer_folding = pm_mops+PatternMatcher([
   (UPat(Ops.BUFFERIZE, name="b"), cleanup_dead_axes),
   (UPat(GroupOp.All-{Ops.BUFFERIZE, Ops.BUFFER}, name="x"), lambda x: x.replace(dtype=x.dtype.base) if isinstance(x.dtype, ImageDType) else None),
   (UPat((Ops.BUFFERIZE), name="x"), lambda x: x.replace(dtype=x.dtype.base) if isinstance(x.dtype, ImageDType)
     and (resolve(prod(x.dtype.shape)!=prod(x.shape)) or x.shape[-1]%4!=0) else None),
   # remove noop buffers. if we look at the next index we can remove even more of these
-  # NOTE: this is mostly the same case as below, but if there's no INDEX this gets more
-  (UPat(Ops.INDEX, name="idx").f(Ops.BUFFERIZE, allow_any_len=True, name="b2"),
-   lambda idx,b2: idx.src[0].replace(tag=nt if len(nt:=(idx.src[0].tag or ()) + (b2.tag or ())) else None) if idx.src[1:] == b2.src[1:] \
-       and idx.src[0].op is not Ops.BUFFER_VIEW else None),
-  # remove reindexing with cost function
-  (UPat.var("src").f(Ops.BUFFERIZE, allow_any_len=True, name="buf").f(Ops.INDEX, allow_any_len=True, name="idx"), remove_bufferize),
+  (UPat(Ops.INDEX, name="idx").f(Ops.BUFFERIZE, allow_any_len=True, name="b2"), remove_noop_bufferize),
   # no buffers for const
   (UPat(Ops.CONST, name='c').f(Ops.BUFFERIZE, allow_any_len=True, name="b"), lambda c,b: b.const_like(c.arg).rtag(b.tag)),
+  # indexing a const is a const
+  (UPat(Ops.INDEX, src=(UPat(Ops.CONST, name="c"),),), lambda c: c),
   # copy on CONST is CONST
   (UPat(Ops.COPY, src=(UPat.cvar("x"), UPat()), name="copy"), lambda copy,x: copy.const_like(x.arg)),
-  (UPat(Ops.COPY, src=(UPat(GroupOp.All-{Ops.CONTIGUOUS, Ops.COPY}).f(Ops.BUFFERIZE, allow_any_len=True, name="b")
-                       .f(Ops.INDEX, allow_any_len=True, name="x"), UPat()), name="copy"), pre_bufferize),
+  # hack if a noop turned to a const
+  (UPat.cvar("c").f(Ops.NOOP).f(Ops.BUFFERIZE, allow_any_len=True, name="buf"), lambda c,buf: buf.replace(src=(c,)+buf.src[1:])),
   # mstack on CONST is CONST
   (UPat(Ops.MSTACK, src=(UPat.var("s"),), allow_any_len=True).f(Ops.INDEX, allow_any_len=True),
    lambda s: UOp.const(c.dtype, c.arg) if (c:=s.base).op is Ops.CONST else None),
 ])
 
+def pre_bufferize(b:UOp, x:UOp, copy:UOp):
+  nb = b.replace(src=(b.src[0].contiguous(),)+b.src[1:])
+  return copy.replace(src=(x.replace(src=(nb,)+x.src[1:]), copy.src[1]))
+pm_remove_bufferize = PatternMatcher([
+  # hack so remove_bufferize doesnt remove the buffer before a copy
+  (UPat(Ops.COPY, src=(UPat(GroupOp.All-{Ops.CONTIGUOUS, Ops.COPY}).f(Ops.BUFFERIZE, allow_any_len=True, name="b")
+                      .f(Ops.INDEX, allow_any_len=True, name="x"), UPat()), name="copy"), pre_bufferize),
+  # remove reindexing with cost function
+  (UPat.var("src").f(Ops.BUFFERIZE, allow_any_len=True, name="buf").f(Ops.INDEX, allow_any_len=True, name="idx"), remove_bufferize),
+])
+
 def late_buffer_view(t:UOp, b:UOp):
   if isinstance(b.device, str) and (b.device.startswith("DISK") or b.device.startswith("TINYFS")):
     rngs = b.src[1:]
@@ -276,7 +294,7 @@ def bufferize_to_store(x:UOp):
     assert assign_target.op is Ops.INDEX, f"{assign_target.op} is not index"
     # in assign, this is the buffer size, not the bufferize size
     # TODO: assign_mops here
-    do_store = assign_target.replace(dtype=sdtype).store(assign_src).replace(tag=x.tag).end(ends=[x for x in rngs if x.op is Ops.RANGE])
+    do_store = assign_target.replace(dtype=sdtype).store(assign_src, *rngs).replace(tag=x.tag)
     ret = assign_target.src[0].after(do_store)
     mops = []
     walk = assign_mops
@@ -289,7 +307,7 @@ def bufferize_to_store(x:UOp):
   # NOTE: the DEFINE_LOCAL needs to be disambiguated here
   if sdtype.addrspace == AddrSpace.GLOBAL:
     buf = UOp.new_buffer(x.arg.device, size, x.dtype)
-    do_store = buf.reshape(shape).index(*rngs, dtype=sdtype).store(x.src[0]).replace(tag=x.tag).end(ends=[x for x in rngs if x.op is Ops.RANGE])
+    do_store = buf.reshape(shape).index(*rngs, dtype=sdtype).store(x.src[0], *rngs).replace(tag=x.tag)
     ret = buf.after(do_store).forced_reshape(shape)
     # TODO: is this right? what if it's offset
     if any(r.op is Ops.RANGE and r.src[0].op is not Ops.CONST for r in rngs):
@@ -301,8 +319,8 @@ def bufferize_to_store(x:UOp):
   tag = x.arg.device
   if tag is None: tag = UOp.unique().arg # TODO: hack
   buf = UOp(Ops.DEFINE_LOCAL, sdtype, arg=tag)
-  do_store = buf.reshape(shape).index(*rngs, dtype=sdtype).store(x.src[0]).end(ends=[x for x in rngs if x.op is Ops.RANGE])
-  return buf.after(do_store).reshape(shape)
+  do_store = buf.reshape(shape).index(*rngs, dtype=sdtype).store(x.src[0], *rngs)
+  return buf.after(do_store.barrier()).reshape(shape)
 
 pm_add_buffers = pm_mops+to_bufferview+PatternMatcher([
   (UPat(Ops.BUFFERIZE, name="x"), bufferize_to_store),
@@ -437,7 +455,7 @@ def split_store(ctx:list[UOp], x:UOp) -> UOp|None:
   return kernel
 
 split_kernels = PatternMatcher([
-  (UPat((Ops.STORE, Ops.END), name="x"), split_store),
+  (UPat(Ops.STORE, name="x"), split_store),
 ])
 
 def tag_uop(ctx:list[UOp], x:UOp):
@@ -477,11 +495,10 @@ def get_rangeify_map(sink:UOp) -> dict[UOp, UOp]:
   tsink = graph_rewrite(tsink, earliest_rewrites+replace_contiguous, ctx={}, name="earliest rewrites")
 
   # convert movement ops to ranges
-  tsink, rctx = run_rangeify(tsink, getenv("DEBUG_RANGEIFY", 0))
+  tsink, rctx = run_rangeify(tsink, DEBUG_RANGEIFY)
 
-  # NOTE: sym (vs symbolic_simple) breaks things here because ranges with len 1 aren't handled right
-  tsink = graph_rewrite(tsink, symbolic_flat+pm_reduce_unparented, name="symbolic")  # this supports const folding
-  tsink = graph_rewrite(tsink, pm_cleanups, bottom_up=True, name="remove costly buffers")
+  tsink = graph_rewrite(tsink, symbolic_flat+pm_reduce_unparented+pm_const_buffer_folding, name="symbolic")  # this supports const folding
+  tsink = graph_rewrite(tsink, pm_remove_bufferize, bottom_up=True, name="remove bufferize with cost function")
   tsink = graph_rewrite(tsink, pm_limit_bufs, ctx=rctx, name="limit buffers")
 
   # rebuild the sink with all the BUFFERIZEs with tags, this is what's ending up in the tensor graph

tinygrad/tensor.py

@@ -11,7 +11,7 @@ from tinygrad.helpers import suppress_finalizing
 from tinygrad.gradient import compute_gradient
 from tinygrad.uop.mathtraits import MathTrait
 from tinygrad.uop.ops import smax, smin, resolve, UOp, Ops, sint, identity_element, all_metadata, _index_to_concrete_int, sint_to_uop, srender
-from tinygrad.uop.spec import tensor_uop_spec, type_verify
+from tinygrad.uop.spec import type_verify, tensor_spec
 from tinygrad.device import Device, Buffer
 from tinygrad.engine.realize import run_schedule
 from tinygrad.engine.memory import memory_planner
@@ -229,7 +229,7 @@ class Tensor(MathTrait):
     big_sink = UOp.sink(*[x.uop for x in (self,)+lst])
 
     # verify Tensors match the spec
-    if __debug__: type_verify(list(big_sink.toposort()), tensor_uop_spec)
+    if __debug__: type_verify(list(big_sink.toposort()), tensor_spec)
 
     if any(isinstance(x._device, tuple) for x in big_sink.toposort()):
       _apply_map_to_tensors(get_multi_map(big_sink), "Apply Multi Map")
@@ -4149,10 +4149,10 @@ class Tensor(MathTrait):
     ```
     """
     assert self.ndim > 1, "NS only works for two or more dims"
+    if self.shape[-2] > self.shape[-1]: return self.transpose(-2, -1).newton_schulz(steps, params, eps).transpose(-2, -1)
     G = self / (self.square().sum(axis=(-2, -1), keepdim=True).sqrt() + eps)
-    if (swap := self.shape[-2] > self.shape[-1]): G = G.transpose(-2, -1)
     for _ in range(steps): G = sum(p * functools.reduce(lambda x, y: (y @ y.transpose(-2, -1)) @ x, [G]*i, G) for i,p in enumerate(params))
-    return G.transpose(-2, -1) if swap else G
+    return G
 
   def qr(self) -> tuple[Tensor, Tensor]:
     assert self.ndim > 1, f"expected two or more dimensions, got {self.ndim}"

tinygrad/uop/ops.py

@@ -268,20 +268,24 @@ class UOp(MathTrait, metaclass=UOpMetaClass):
   @property
   def size(self) -> int: return prod([int(x.vmax) if isinstance(x, UOp) else x for x in self.shape])
 
+  @functools.cached_property
+  def ended_ranges(self):
+    # copy of range_start
+    match self.op:
+      case Ops.REDUCE | Ops.BUFFERIZE: return self.src[1:]
+      case Ops.STORE: return self.src[2:]
+      case Ops.WMMA: return self.src[3:]
+      case Ops.END: return self.src[:1]
+      case _: return ()
+
   # determine what ranges this is in
   @recursive_property
   def _ranges(self) -> dict[UOp, None]:
     ret: dict[UOp, None] = {}
-    if self.op in range_start.keys():
-      for s in self.src[:range_start[self.op]]: ret.update(s.ranges)
-      for s in UOp.sink(*self.src[range_start[self.op]:]).ranges:
+    for s in self.src: ret.update(s.ranges)
+    if (er:=self.ended_ranges):
+      for s in UOp.sink(*er).ranges:
         if s in ret: del ret[s]
-    elif self.op is Ops.END:
-      for s in self.src[self.arg:]: ret.update(s.ranges)
-      for s in UOp.sink(*self.src[:self.arg]).ranges:
-        if s in ret: del ret[s]
-    else:
-      for s in self.src: ret.update(s.ranges)
     return ret
 
   @property
@@ -289,13 +293,6 @@ class UOp(MathTrait, metaclass=UOpMetaClass):
     if self.op is Ops.RANGE: return {self:None}
     return self._ranges
 
-  @functools.cached_property
-  def ended_ranges(self):
-    match self.op:
-      case Ops.REDUCE: return self.src[1:]
-      case Ops.END: return self.src[:self.arg]
-      case _: raise RuntimeError(f"{self.op} doesn't end ranges")
-
   # *** uop evaluation ***
 
   def simplify(self, tracked=False, full_symbolic=True):
@@ -361,9 +358,9 @@ class UOp(MathTrait, metaclass=UOpMetaClass):
     return UOp(Ops.GEP, self.dtype.scalar().vec(len(i)) if len(i) > 1 else self.dtype.scalar(), (self,), i)
   def load(self, *src:UOp, **kwargs): return UOp(Ops.LOAD, dtype=kwargs.pop("dtype", self.dtype.base), src=(self,)+src, **kwargs)
   def store(self, *src:UOp, **kwargs): return UOp(Ops.STORE, kwargs.pop("dtype", dtypes.void), (self,)+src, **kwargs)
-  def end(self, *src:UOp, ends:Sequence[UOp]):
-    if len(ends) == 0: return self
-    return UOp(Ops.END, src=(*ends, self, *src), arg=len(ends))
+  def end(self, *src:UOp):
+    assert self.op is Ops.RANGE, "end only ends ranges"
+    return UOp(Ops.END, src=(self,)+src)
   def after(self, *src:UOp): return UOp(Ops.AFTER, self.dtype, (self,)+src)
   def assign(self, x:UOp): return UOp(Ops.ASSIGN, self.dtype, (self, x))
   def barrier(self, *src:UOp): return UOp(Ops.BARRIER, src=(self,)+src)
@@ -555,8 +552,8 @@ class UOp(MathTrait, metaclass=UOpMetaClass):
     if self.op is Ops.BUFFER: return self
     if self.op is Ops.MSELECT: return self.src[0].buf_uop.mselect(self.arg)
     if self.op is Ops.MSTACK: return UOp(Ops.MSTACK, self.dtype, src=tuple(x.buf_uop for x in self.src))
-    assert self.op is Ops.AFTER, f"must be AFTER {self.op}"
-    return self.src[0].buf_uop.base
+    assert self.base.op is Ops.AFTER, f"must be AFTER {self.base.op}"
+    return self.base.src[0].buf_uop.base
 
   def as_buf(self) -> UOp:
     if self.op is Ops.MSELECT: return self.src[0].as_buf().mselect(self.arg)
@@ -845,7 +842,8 @@ class UPat(MathTrait):
 
   # copied from UOp
   def sink(self, *srcs:UPat|None, **kwargs): return UPat(Ops.SINK, dtypes.void, (self,)+tuple([x for x in srcs if x is not None]), **kwargs)
-  def index(self, idx:UPat, valid:UPat|None=None): return UPat(Ops.INDEX, self.dtype, (self,idx,valid) if valid is not None else (self,idx))
+  def index(self, idx:UPat, valid:UPat|None=None, **kwargs):
+    return UPat(Ops.INDEX, self.dtype, (self,idx,valid) if valid is not None else (self,idx), **kwargs)
   def cast(self, dtype=None, **kwargs): return UPat(Ops.CAST, dtype, (self,), **kwargs)
   def bitcast(self, dtype=None): return UPat(Ops.BITCAST, dtype, (self,))
   def gep(self, i:int|None=None, **kwargs): return UPat(Ops.GEP, None, (self,), (i,) if i is not None else None, **kwargs)
@@ -1191,6 +1189,8 @@ pm_lower_index_dtype = PatternMatcher([
     lambda s: s.replace(src=s.src[:2]+tuple(u.src[0] for u in s.src[2:]))),
   # TODO: this is only triggering if they are all casts, correct?
   (UPat((Ops.SINK, Ops.NOOP), src=UPat().cast(dtypes.index), name="n"), lambda n: n.replace(src=tuple(s.src[0] for s in n.src))),
+  # no CAST on END
+  (UPat(Ops.END, src=(UPat(Ops.CAST),), allow_any_len=True, name="e"), lambda e: e.replace(src=(e.src[0].src[0],)+e.src[1:])),
 ])
 def _index_to_concrete_int(u:UOp): return graph_rewrite(u.sink(), pm_lower_index_dtype).src[0]
 

tinygrad/uop/spec.py

@@ -1,60 +1,45 @@
-from typing import cast, Callable
-from tinygrad.uop.ops import PatternMatcher, UPat, GroupOp, Ops, UOp, print_uops, python_alu, graph_rewrite, AxisType
+from typing import cast
+from tinygrad.uop.ops import PatternMatcher, UPat, GroupOp, Ops, UOp, print_uops, AxisType
 from tinygrad.dtype import DType, ImageDType, dtypes, PtrDType, AddrSpace, Invalid
-from tinygrad.helpers import all_same, prod, DEBUG, IGNORE_OOB, Context, cpu_profile
-try:
-  import z3
-  # older versions of z3 dont have some operators like & overloaded
-  if z3.get_version() < (4, 12, 4, 0): raise ImportError
+from tinygrad.helpers import DEBUG, Context
+from tinygrad.uop.validate import validate_index
 
-  # IDIV is truncated division but z3 does euclidian division (floor if b>0 ceil otherwise); mod by power of two sometimes uses Ops.AND
-  def z3_cdiv(a, b):return z3.If((a<0), z3.If(0<b, (a+(b-1))/b, (a-(b+1))/b), a/b)
-  def z3_xor(a,b):
-    if isinstance(a, z3.BoolRef): return a^b
-    assert a==-1 or b==-1, "xor can only be used in indexing if one of the aruments is -1"
-    return -a-1 if b==-1 else -b-1
-  z3_alu: dict[Ops, Callable] = python_alu | {Ops.MOD: lambda a,b: a-z3_cdiv(a,b)*b, Ops.IDIV: z3_cdiv, Ops.SHR: lambda a,b: a/(2**b.as_long()),
-    Ops.SHL: lambda a,b: a*(2**b.as_long()), Ops.AND: lambda a,b: a%(b+1) if isinstance(b, z3.ArithRef) else a&b, Ops.WHERE: z3.If, Ops.XOR: z3_xor,
-    Ops.MAX: lambda a,b: z3.If(a<b, b, a), Ops.TRUNC: lambda a: a if a.is_int() else z3.ToReal(z3.If(a >= 0, z3.ToInt(a), -z3.ToInt(-a)))}
-  def create_bounded(name:str, vmin, vmax, solver:z3.Solver) -> z3.ArithRef:
-    s = z3.Int(name, ctx=solver.ctx)
-    solver.add(vmin <= s, s <= vmax)
-    return s
+# four specs:
+#   shared_spec  -- usable anywhere
+#   tensor_spec  -- usable in tensor graph
+#   program_spec -- usable in linearized program
+#   full_spec    -- all uops ever created
 
-  # ctx is (solver, load_number_dict)
-  # each uop gets rewritten to NOOP(arg=(solver, z3_object)), the arg has the solver first due to UOpMetaClass caching. z3 objects from different
-  # contexts can have the same hash but error on comparison
-  z3_renderer = PatternMatcher([
-    (UPat(Ops.SPECIAL, src=UPat(Ops.NOOP), name="x"), lambda x,ctx: UOp(Ops.NOOP, arg=(ctx[0],create_bounded(x.arg, 0, x.src[0].arg[1]-1, ctx[0])))),
-    (UPat(Ops.DEFINE_VAR, name="x"), lambda x,ctx: UOp(Ops.NOOP, arg=(ctx[0],create_bounded(x.arg[0], x.arg[1], x.arg[2], ctx[0])))),
-    (UPat(Ops.RANGE, name="x"), lambda x,ctx: UOp(Ops.NOOP, arg=(ctx[0],create_bounded(f"ridx{x.arg}", 0, x.src[0].arg[1]-1, ctx[0])))),
-    # loaded bools become a z3 int with min max of 0-1
-    (UPat(Ops.LOAD, dtypes.ints+(dtypes.bool,), name="x"), lambda x,ctx:
-      UOp(Ops.NOOP, arg=(ctx[0],create_bounded(f"load{ctx[1].setdefault(x, len(ctx[1]))}", x.dtype.min, x.dtype.max, ctx[0]))).cast(x.dtype)),
-    (UPat(Ops.CONST, dtype=dtypes.ints+(dtypes.bool,dtypes.index), name="x"),
-      lambda x,ctx: UOp(Ops.NOOP, arg=(ctx[0],(z3.BoolVal if dtypes.is_bool(x.dtype) else z3.IntVal)(x.arg, ctx=ctx[0].ctx)))),
-    # z3 can cast from bool to int automatically
-    (UPat(Ops.CAST, dtype=dtypes.ints+(dtypes.index,), src=UPat(Ops.NOOP), name="x"), lambda x: x.src[0]),
-    (UPat(Ops.CAST, dtype=dtypes.bool, src=UPat(Ops.NOOP), name="x"), lambda x,ctx: UOp(Ops.NOOP, arg=(ctx[0], x.src[0].arg[1]!=0))),
-    # if the source of the cast is not a noop it means that it is a float and so we create a new variable
-    (UPat(Ops.CAST, dtype=dtypes.ints+(dtypes.index,), name="x"), lambda x,ctx:
-      UOp(Ops.NOOP, arg=(ctx[0], create_bounded(f"cast{ctx[1].setdefault(x, len(ctx[1]))}", x.dtype.min, x.dtype.max, ctx[0])))),
-    (UPat(Ops.CAST, dtype=dtypes.bool, name="x"), lambda x,ctx:
-      UOp(Ops.NOOP, arg=(ctx[0], z3.Bool(f"cast{ctx[1].setdefault(x, len(ctx[1]))}",ctx=ctx[0].ctx)))),
-    (UPat(GroupOp.ALU, src=UPat(Ops.NOOP), name="x"), lambda x,ctx: UOp(Ops.NOOP, arg=(ctx[0], z3_alu[x.op](*(s.arg[1] for s in x.src))))),
-    # A comparison between floats introduces a new bool variable
-    (UPat(GroupOp.Comparison, src=UPat(dtype=dtypes.floats), name="x"), lambda x,ctx:
-      UOp(Ops.NOOP, arg=(ctx[0], z3.Bool(f"float_cmp{ctx[1].setdefault(x, len(ctx[1]))}",ctx=ctx[0].ctx)))),
-  ])
+# *** these uops work anywhere ***
 
-  def uops_to_z3(solver, *uops: UOp) -> 'list[z3.ExprRef]':
-    with Context(TRACK_MATCH_STATS=0, SPEC=0):  # cant pickle z3 objects, and these UOps don't follow spec
-      return [s.arg[1] for s in graph_rewrite(uops[0].sink(*uops[1:]), z3_renderer, ctx=(solver, {})).src]
+shared_spec = PatternMatcher([
+  (UPat(Ops.SINK, dtypes.void), lambda: True), # NOTE: for testing, we let sinks be anything
 
-  z3_imported = True
-except (ImportError, AttributeError): z3_imported = False
+  # CONST/DEFINE_VAR are everywhere
+  (UPat(Ops.CONST, src=(), name="x"), lambda x: type(x.arg) is type(dtypes.as_const(x.arg, x.dtype))),
+  (UPat(Ops.DEFINE_VAR, name="x"), lambda x: isinstance(x.arg[1], int) and isinstance(x.arg[2], int)),
 
-buffer_spec = PatternMatcher([
+  # ALUs: most ALUs have all matching dtypes, except CMPLT, CMPNE, and WHERE
+  (UPat(Ops.WHERE, name="w", src=(UPat(dtype=dtypes.bool), UPat.var("x"), UPat.var("y"))), lambda w,x,y: w.dtype == x.dtype == y.dtype),
+  (UPat((Ops.CMPLT, Ops.CMPNE, Ops.CMPEQ), dtype=dtypes.bool, src=(UPat.var("x"), UPat.var("y"))), lambda x,y: x.dtype.base == y.dtype.base),
+  # and SHL/SHR, the shift distance can be an int
+  (UPat((Ops.SHL, Ops.SHR), src=(UPat.var("x"), UPat.var("y")), name="a"), lambda a,x,y: a.dtype == x.dtype and y.dtype in (x.dtype, dtypes.uint)),
+  (UPat((Ops.IDIV, Ops.MOD), name="x"), lambda x: None if dtypes.is_int(x.dtype) else False),
+  (UPat(GroupOp.ALU, name="x"), lambda x: all(x.dtype.base == y.dtype.base for y in x.src)),
+
+  # CAST
+  (UPat((Ops.BITCAST, Ops.CAST), src=(UPat(),), name="x"), lambda x: x.arg is None),
+
+  # RANGE can be in the big graph now
+  (UPat(Ops.RANGE, src=(UPat.var("x"),), allow_any_len=True, name="rng"), lambda rng,x:
+    rng.dtype == x.dtype and isinstance(rng.arg, tuple) and len(rng.arg) >= 2 and \
+      all(isinstance(ra, int) for ra in rng.arg[0:-1]) and isinstance(rng.arg[-1], AxisType)),
+])
+
+# ***** UOp spec in the Tensor graph *****
+
+tensor_spec = PatternMatcher([
+  # buffer spec
   (UPat(Ops.UNIQUE, dtypes.void, ()), lambda: True),
   (UPat(Ops.DEVICE, dtypes.void, (), name="d"), lambda d:
    isinstance(d.arg, str) or (isinstance(d.arg, tuple) and all(isinstance(s, str) for s in d.arg))),
@@ -63,9 +48,7 @@ buffer_spec = PatternMatcher([
   (UPat(Ops.BUFFER_VIEW, src=(UPat(Ops.BUFFER),), name="buf_view"),
    lambda buf_view: isinstance(buf_view.arg, tuple) and len(buf_view.arg) == 2 and all(isinstance(arg, (int, UOp)) for arg in buf_view.arg)),
   (UPat(Ops.BUFFER_VIEW, src=(UPat(Ops.MSTACK, src=UPat(Ops.BUFFER)),)), lambda: True),
-])
 
-assign_spec = PatternMatcher([
   # KERNEL can attach to an AFTER to describe the compute required to realize a BUFFER
   (UPat(Ops.KERNEL, src=UPat((Ops.BUFFER, Ops.BUFFER_VIEW, Ops.AFTER, Ops.MSELECT, Ops.MSTACK, Ops.BIND))), lambda: True),
 
@@ -77,11 +60,7 @@ assign_spec = PatternMatcher([
 
   # MSTACK combines buffers into multi
   (UPat(Ops.MSTACK, name="x"), lambda x: all(isinstance(x.device, str) for x in x.src)),
-])
 
-# *** this is the spec of a Tensor in UOp ***
-
-tensor_uop_spec = buffer_spec+assign_spec+PatternMatcher([
   (UPat((Ops.RESHAPE, Ops.EXPAND), name="mv", src=(UPat.var("x"), UPat(dtype=dtypes.index))), lambda mv,x: True),
   (UPat((Ops.PAD, Ops.SHRINK), name="mv", src=(UPat.var("x"), UPat(dtype=dtypes.index), UPat(dtype=dtypes.index))), lambda mv,x: True),
   (UPat((Ops.PERMUTE, Ops.FLIP), name="mv", src=(UPat.var("x"),)), lambda mv,x: isinstance(mv.arg, tuple)),
@@ -109,130 +88,69 @@ tensor_uop_spec = buffer_spec+assign_spec+PatternMatcher([
   (UPat(Ops.ALLREDUCE, name="red", src=(UPat.var("x"), UPat(Ops.DEVICE))), lambda red,x: red.dtype == x.dtype and isinstance(red.arg, Ops)),
   (UPat(Ops.MULTI, name="multi"), lambda multi: all(x.dtype == multi.dtype for x in multi.src) and isinstance(multi.arg, int)),
 
+  # REDUCE_AXIS is the reduce in the tensor graph
+  (UPat(Ops.REDUCE_AXIS, name="x"), lambda x: isinstance(x.arg, tuple) and len(x.arg) >= 2 and x.arg[0] in {Ops.ADD, Ops.MUL, Ops.MAX}),
+
   # REDUCE with an outerworld range
   (UPat(Ops.REDUCE, src=(UPat(),), allow_any_len=True, name="x"), lambda x: all(y.dtype == dtypes.index for y in x.src[1:])),
 
   # AFTER if things were kernelized
-  (UPat(Ops.AFTER, src=(UPat((Ops.BUFFER, Ops.AFTER)),), allow_any_len=True), lambda: True)
-])
+  (UPat(Ops.AFTER, src=(UPat((Ops.BUFFER, Ops.AFTER)),), allow_any_len=True), lambda: True),
+])+shared_spec
 
-# ***** uop type spec *****
+# ***** UOp spec in linearized programs *****
 
-def validate_index(idx:UOp, gate:UOp|None=None):
-  if gate is None: gate = UOp.const(dtypes.bool, True)
-  # TODO: check for overflow
-  if IGNORE_OOB or isinstance(idx.dtype, ImageDType) or (sz := idx.src[0].ptrdtype.size) == -1: return True
-  # We can use UOp min/max to do a faster check, but it can give false positive since its not an exact bound and doesn't consider the mask
-  if 0<=idx.src[1].vmin and idx.src[1].vmax<sz: return True
-  mask = idx.src[2]&gate if len(idx.src)==3 else gate
-
-  # WEBGPU has a BITCAST in the index. TODO: fix
-  if any(x.op is Ops.BITCAST for x in idx.toposort()): return True
-
-  if not z3_imported: raise ImportError("z3 >= 4.12.4 is required for bounds checking, try IGNORE_OOB=0 or \"pip install 'z3-solver>=4.12.4\"")
-  solver = z3.Solver(ctx=z3.Context())
-  z3_idx, z3_mask = uops_to_z3(solver, idx.src[1], mask)
-  solver.add(z3_mask)
-  with cpu_profile("validate index with z3", "TINY"):
-    if solver.check((z3_idx<0)|(sz<=z3_idx)) == z3.sat:
-      print(f"idx={idx.src[1].render(simplify=False)}")
-      print(f"mask & gate={mask.render(simplify=False)}")
-      print(f"# OUT OF BOUNDS ACCESS: at {solver.model()} INDEX not in 0 - {sz}\nconstraints = {solver}")
-      return False
-  return True
-
-def validate_store(idx:UOp, val:UOp, gate:UOp|None=None):
-  if gate is None: gate = UOp.const(dtypes.bool, True)
-  if gate.op is Ops.IF: gate = gate.src[0]
-  # we need to find the implicit gates, inverse of delete_redundant_gates
-  for u in val.toposort():
-    if u.op is Ops.IF: gate &= u.src[0]
-  return validate_index(idx, gate)
-
-index_pat = UPat(Ops.INDEX, name="idx").or_casted()
-
-# this is the matcher for the final rendered UOps
-# matcher functions returns True or False (or None to not match)
-spec = PatternMatcher([
+program_spec = PatternMatcher([
+  # DEFINEs
   (UPat(Ops.DEFINE_GLOBAL, name="x"), lambda x: isinstance(x.dtype, (PtrDType, ImageDType)) and x.dtype.addrspace == AddrSpace.GLOBAL),
   (UPat(Ops.DEFINE_LOCAL, name="x"), lambda x: isinstance(x.dtype, PtrDType) and x.dtype.addrspace == AddrSpace.LOCAL),
   (UPat(Ops.DEFINE_REG, src=()), lambda: True),
-  (UPat(Ops.DEFINE_VAR, name="x"), lambda x: isinstance(x.arg[1], int) and isinstance(x.arg[2], int)),
-
-  (UPat(Ops.RANGE, src=(UPat.var("x"),), allow_any_len=True, name="rng"), lambda rng,x:
-    rng.dtype == x.dtype and isinstance(rng.arg, tuple) and len(rng.arg) >= 2 and \
-      all(isinstance(ra, int) for ra in rng.arg[0:-1]) and isinstance(rng.arg[-1], AxisType)),
-  (UPat(Ops.SPECIAL, src=(UPat.var("x"),), name="s"), lambda s,x: s.dtype == x.dtype == dtypes.int32 and isinstance(s.arg, str)),
-
-  (UPat(Ops.CONST, src=(), name="x"), lambda x: type(x.arg) is type(dtypes.as_const(x.arg, x.dtype))),
 
   # allow AFTER on buffers
   (UPat(Ops.AFTER, src=(UPat(GroupOp.Defines),), allow_any_len=True), lambda: True),
 
-  # **** new style load/store ****
+  # INDEX is used in new style load/store
+  (UPat(Ops.INDEX, src=(UPat(GroupOp.Defines).or_after(), UPat(), UPat(dtype=dtypes.bool))), lambda: True),
+  (UPat(Ops.INDEX, src=(UPat(GroupOp.Defines).or_after(), UPat())), lambda: True),
+
+  # LOAD (idx, alt_value) / LOAD(idx) / STORE(idx, val)
+  (UPat(Ops.LOAD,  src=(UPat(Ops.INDEX, name="idx").or_casted(), UPat())), validate_index),
+  (UPat(Ops.LOAD,  src=(UPat(Ops.INDEX, name="idx").or_casted(), )), validate_index),
+  (UPat(Ops.STORE, src=(UPat(Ops.INDEX, name="idx").or_casted(), UPat())), validate_index),
+
+  # RANGE/SPECIAL define loops, END closes them
+  (UPat(Ops.SPECIAL, src=(UPat.var("x"),), name="s"), lambda s,x: s.dtype == x.dtype == dtypes.int32 and isinstance(s.arg, str)),
+  (UPat(Ops.END, src=(UPat(Ops.RANGE), UPat()), dtype=dtypes.void), lambda: True),
 
   # make sure all index dtypes have been lowered
   (UPat(GroupOp.All, dtype=dtypes.index), lambda: False),
   (UPat(Ops.CONST, arg=Invalid), lambda: False),
   (UPat(Ops.VCONST, name="x"), lambda x: all(v is not Invalid for v in x.src)),
 
-  # INDEX is used in new style load/store
-  # INDEX takes a <buf, alu, gate?>
-  (UPat(Ops.INDEX, src=(UPat(GroupOp.Defines).or_after(), UPat())), lambda: True),
-  (UPat(Ops.INDEX, src=(UPat(GroupOp.Defines).or_after(), UPat(), UPat(dtype=dtypes.bool))), lambda: True),
-
-  # LOAD takes a <bufidx, alt?, barrier?>
-  (UPat(Ops.LOAD, src=(index_pat, UPat(Ops.IF, name="cond")), allow_any_len=True), lambda idx,cond: validate_index(idx,cond.src[0])),
-  (UPat(Ops.LOAD, src=(index_pat,), allow_any_len=True), validate_index),
-
-  # STORE takes a <bufidx, val, ranges...>
-  (UPat(Ops.STORE, src=(index_pat, UPat(name="val")), allow_any_len=True), validate_store),
-
-  # most ALUs have all matching dtypes, except CMPLT, CMPNE, and WHERE
-  (UPat(Ops.WHERE, name="w", src=(UPat(dtype=dtypes.bool), UPat.var("x"), UPat.var("y"))), lambda w,x,y: w.dtype == x.dtype == y.dtype),
-  (UPat((Ops.CMPLT, Ops.CMPNE, Ops.CMPEQ), dtype=dtypes.bool, src=(UPat.var("x"), UPat.var("y"))), lambda x,y: x.dtype.base == y.dtype.base),
-  # and SHL/SHR, the shift distance can be an int
-  (UPat((Ops.SHL, Ops.SHR), src=(UPat.var("x"), UPat.var("y")), name="a"), lambda a,x,y: a.dtype == x.dtype and y.dtype in (x.dtype, dtypes.uint)),
-  (UPat((Ops.IDIV, Ops.MOD), name="x"), lambda x: None if dtypes.is_int(x.dtype) else False),
-  (UPat(GroupOp.ALU, name="x"), lambda x: all(x.dtype.base == y.dtype.base for y in x.src)),
-
-  (UPat(Ops.END, dtype=dtypes.void), lambda: True),
-
   # WMMA has a <a, b, acc>
   (UPat(Ops.WMMA, src=(UPat(), UPat(), UPat()), name="x"), lambda x: isinstance(x.arg, tuple) and len(x.arg) == 8),
-  (UPat(Ops.CONTRACT, name="x"), lambda x: x.dtype.count == prod(y[1] for y in x.arg)),
-  (UPat(Ops.UNROLL, name="x"), lambda x: x.src[0].dtype.count == prod(y[1] for y in x.arg)),
 
-  # if has a <gate, barrier?>
-  (UPat(Ops.IF, dtype=dtypes.void, src=(UPat(),), allow_any_len=True), lambda: True),
-  (UPat(Ops.ENDIF, dtype=dtypes.void, src=(UPat(Ops.IF),), allow_any_len=True), lambda: True),
+  # if has a <gate, index_for_dedup>
+  (UPat(Ops.IF, dtype=dtypes.void, src=(UPat(dtype=dtypes.bool), UPat((Ops.CAST, Ops.INDEX)))), lambda: True),
+  (UPat(Ops.ENDIF, dtype=dtypes.void, src=(UPat(Ops.IF),)), lambda: True),
 
-  (UPat(Ops.REDUCE_AXIS, name="x"), lambda x: isinstance(x.arg, tuple) and len(x.arg) >= 2 and x.arg[0] in {Ops.ADD, Ops.MUL, Ops.MAX}),
-  (UPat(Ops.GEP, src=(UPat.var("src"),), name="gep"), lambda gep,src: gep.dtype == src.dtype.scalar()),
+  # VECTORIZE/GEP
   (UPat(Ops.VECTORIZE, name="x"), lambda x: len(x.src)>1 and len(x.src) == x.dtype.vcount and all(x.dtype == y.dtype.vec(len(x.src)) for y in x.src)),
-  (UPat((Ops.BITCAST, Ops.CAST), src=(UPat(),), name="x"), lambda x: x.arg is None),
+  (UPat(Ops.GEP, src=(UPat.var("src"),), name="gep"), lambda gep,src: gep.dtype == src.dtype.scalar()),
+
+  # BARRIER
   (UPat(Ops.BARRIER, dtypes.void, src=UPat(Ops.STORE, allow_any_len=True)), lambda: True), # NOTE: all pointers must be local
   (UPat(Ops.BARRIER, dtypes.void), lambda: True), # BARRIERs can also happen at the end of loops
 
-  # NOTE: for testing, we let sinks be anything
-  #(UPat(Ops.SINK, src=UPat(Ops.STORE)), lambda: True),
-  (UPat(Ops.SINK, dtypes.void), lambda: True),
   (UPat((Ops.NOOP, Ops.CUSTOMI, Ops.CUSTOM, Ops.PRECAST)), lambda: True),
-
-  # PTX LOAD/STORE
-  (UPat((Ops.LOAD, Ops.STORE), src=(UPat(dtype=dtypes.int64),), allow_any_len=True), lambda: True),
-])
-
-# *** this is the UOp AST spec ***
-
-ast_spec = PatternMatcher([
-  # all parent UOps must have the same shape
-  (UPat(GroupOp.All-{Ops.SINK}, name="root"), lambda root: all_same([x.shape for x in root.src if x.st is not None])),
-])
+])+shared_spec
 
 # *** this spec should match all UOps ever created ***
 
 full_spec = PatternMatcher([
+  # any END
+  (UPat(Ops.END), lambda: True),
+
   # SENTINEL should never be in the graph
   (UPat(Ops.SENTINEL), lambda: False),
 
@@ -257,6 +175,8 @@ full_spec = PatternMatcher([
   (UPat(Ops.ASSIGN, src=(UPat(), UPat(), UPat(GroupOp.Movement))), lambda: True),
 
   # expander: unroll/contract/gep/ptrcat/cat
+  #(UPat(Ops.CONTRACT, name="x"), lambda x: x.dtype.count == prod(y[1] for y in x.arg)),
+  #(UPat(Ops.UNROLL, name="x"), lambda x: x.src[0].dtype.count == prod(y[1] for y in x.arg)),
   (UPat((Ops.UNROLL, Ops.CONTRACT), src=(UPat(),)), lambda: True),
   # GEP multi is supported here
   (UPat(Ops.GEP, name="gep"), lambda gep: gep.dtype is dtypes.void or gep.dtype.vcount == len(gep.arg)),
@@ -284,12 +204,11 @@ full_spec = PatternMatcher([
   (UPat(Ops.RESHAPE, src=(UPat(Ops.STORE),)), lambda: True),
   # allow any AFTER
   (UPat(Ops.AFTER, src=(UPat(),), allow_any_len=True), lambda: True),
-])+tensor_uop_spec+spec
+])+tensor_spec+program_spec
 
 # ***** uop helpers *****
 
-def type_verify(uops:list[UOp], extra_spec:PatternMatcher|None=None):
-  check_spec = (extra_spec+spec) if extra_spec is not None else spec
+def type_verify(uops:list[UOp], check_spec:PatternMatcher):
   for i,u in enumerate(uops):
     with Context(TRACK_MATCH_STATS=0): ret = check_spec.rewrite(u)
     if cast(bool|None, ret) is not True:

tinygrad/uop/symbolic.py

@@ -382,8 +382,6 @@ symbolic = symbolic_simple+commutative+PatternMatcher([
     tuple(flatten([(y,) if y.op in {Ops.RANGE, Ops.IF, Ops.STORE, Ops.KERNEL, Ops.BARRIER, Ops.END} else y.src for y in x.src[1:]])))),
   # after with 1 src is just src[0]
   (UPat(Ops.AFTER, src=(UPat.var("s"),)), lambda s: s),
-  # END is only on RANGES
-  (UPat(Ops.END, name="e"), lambda e: UOp.end(*e.src[e.arg:], ends=sorted(UOp.sink(*e.src[:e.arg]).ranges, key=lambda x: x.arg))),
 ])+gep_pushing
 
 symbolic_flat = symbolic+PatternMatcher([
@@ -489,7 +487,7 @@ def where_on_load(l, c1, buf, x):
   # we move the condition from the where to the load _as long as_ the condtition doesn't have some range that would place it inside of a new range
   # also no data dependent loads!
   moved_clauses = [c for c in c1.split_uop(Ops.AND) if c not in duplicate_clauses and all(r in x.ranges for r in c.ranges)
-    and not c.op_in_backward_slice_with_self(Ops.LOAD)]
+    and all(u in x.backward_slice_with_self for u in c.backward_slice_with_self if u.op is Ops.LOAD)]
   if not (removed:=moved_clauses+duplicate_clauses): return None
   # aditionally we can drop the clause on the where if it already exists in the load
   remaining_clause = UOp.const(dtypes.bool, True).prod(*[c for c in c1.split_uop(Ops.AND) if c not in removed])

tinygrad/uop/validate.py

@@ -0,0 +1,79 @@
+from typing import Callable
+from tinygrad.uop.ops import PatternMatcher, UPat, GroupOp, Ops, UOp, python_alu, graph_rewrite
+from tinygrad.dtype import ImageDType, dtypes
+from tinygrad.helpers import IGNORE_OOB, Context, cpu_profile
+
+try:
+  import z3
+  # older versions of z3 dont have some operators like & overloaded
+  if z3.get_version() < (4, 12, 4, 0): raise ImportError
+
+  # IDIV is truncated division but z3 does euclidian division (floor if b>0 ceil otherwise); mod by power of two sometimes uses Ops.AND
+  def z3_cdiv(a, b):return z3.If((a<0), z3.If(0<b, (a+(b-1))/b, (a-(b+1))/b), a/b)
+  def z3_xor(a,b):
+    if isinstance(a, z3.BoolRef): return a^b
+    assert a==-1 or b==-1, "xor can only be used in indexing if one of the aruments is -1"
+    return -a-1 if b==-1 else -b-1
+  z3_alu: dict[Ops, Callable] = python_alu | {Ops.MOD: lambda a,b: a-z3_cdiv(a,b)*b, Ops.IDIV: z3_cdiv, Ops.SHR: lambda a,b: a/(2**b.as_long()),
+    Ops.SHL: lambda a,b: a*(2**b.as_long()), Ops.AND: lambda a,b: a%(b+1) if isinstance(b, z3.ArithRef) else a&b, Ops.WHERE: z3.If, Ops.XOR: z3_xor,
+    Ops.MAX: lambda a,b: z3.If(a<b, b, a), Ops.TRUNC: lambda a: a if a.is_int() else z3.ToReal(z3.If(a >= 0, z3.ToInt(a), -z3.ToInt(-a)))}
+  def create_bounded(name:str, vmin, vmax, solver:z3.Solver) -> z3.ArithRef:
+    s = z3.Int(name, ctx=solver.ctx)
+    solver.add(vmin <= s, s <= vmax)
+    return s
+
+  # ctx is (solver, load_number_dict)
+  # each uop gets rewritten to NOOP(arg=(solver, z3_object)), the arg has the solver first due to UOpMetaClass caching. z3 objects from different
+  # contexts can have the same hash but error on comparison
+  z3_renderer = PatternMatcher([
+    (UPat(Ops.SPECIAL, src=UPat(Ops.NOOP), name="x"), lambda x,ctx: UOp(Ops.NOOP, arg=(ctx[0],create_bounded(x.arg, 0, x.src[0].arg[1]-1, ctx[0])))),
+    (UPat(Ops.DEFINE_VAR, name="x"), lambda x,ctx: UOp(Ops.NOOP, arg=(ctx[0],create_bounded(x.arg[0], x.arg[1], x.arg[2], ctx[0])))),
+    (UPat(Ops.RANGE, name="x"), lambda x,ctx: UOp(Ops.NOOP, arg=(ctx[0],create_bounded(f"ridx{x.arg}", 0, x.src[0].arg[1]-1, ctx[0])))),
+    # loaded bools become a z3 int with min max of 0-1
+    (UPat(Ops.LOAD, dtypes.ints+(dtypes.bool,), name="x"), lambda x,ctx:
+      UOp(Ops.NOOP, arg=(ctx[0],create_bounded(f"load{ctx[1].setdefault(x, len(ctx[1]))}", x.dtype.min, x.dtype.max, ctx[0]))).cast(x.dtype)),
+    (UPat(Ops.CONST, dtype=dtypes.ints+(dtypes.bool,dtypes.index), name="x"),
+      lambda x,ctx: UOp(Ops.NOOP, arg=(ctx[0],(z3.BoolVal if dtypes.is_bool(x.dtype) else z3.IntVal)(x.arg, ctx=ctx[0].ctx)))),
+    # z3 can cast from bool to int automatically
+    (UPat(Ops.CAST, dtype=dtypes.ints+(dtypes.index,), src=UPat(Ops.NOOP), name="x"), lambda x: x.src[0]),
+    (UPat(Ops.CAST, dtype=dtypes.bool, src=UPat(Ops.NOOP), name="x"), lambda x,ctx: UOp(Ops.NOOP, arg=(ctx[0], x.src[0].arg[1]!=0))),
+    # if the source of the cast is not a noop it means that it is a float and so we create a new variable
+    (UPat(Ops.CAST, dtype=dtypes.ints+(dtypes.index,), name="x"), lambda x,ctx:
+      UOp(Ops.NOOP, arg=(ctx[0], create_bounded(f"cast{ctx[1].setdefault(x, len(ctx[1]))}", x.dtype.min, x.dtype.max, ctx[0])))),
+    (UPat(Ops.CAST, dtype=dtypes.bool, name="x"), lambda x,ctx:
+      UOp(Ops.NOOP, arg=(ctx[0], z3.Bool(f"cast{ctx[1].setdefault(x, len(ctx[1]))}",ctx=ctx[0].ctx)))),
+    (UPat(GroupOp.ALU, src=UPat(Ops.NOOP), name="x"), lambda x,ctx: UOp(Ops.NOOP, arg=(ctx[0], z3_alu[x.op](*(s.arg[1] for s in x.src))))),
+    # A comparison between floats introduces a new bool variable
+    (UPat(GroupOp.Comparison, src=UPat(dtype=dtypes.floats), name="x"), lambda x,ctx:
+      UOp(Ops.NOOP, arg=(ctx[0], z3.Bool(f"float_cmp{ctx[1].setdefault(x, len(ctx[1]))}",ctx=ctx[0].ctx)))),
+  ])
+
+  def uops_to_z3(solver, *uops: UOp) -> 'list[z3.ExprRef]':
+    with Context(TRACK_MATCH_STATS=0, SPEC=0):  # cant pickle z3 objects, and these UOps don't follow spec
+      return [s.arg[1] for s in graph_rewrite(uops[0].sink(*uops[1:]), z3_renderer, ctx=(solver, {})).src]
+
+  z3_imported = True
+except (ImportError, AttributeError): z3_imported = False
+
+def validate_index(idx:UOp, gate:UOp|None=None):
+  if gate is None: gate = UOp.const(dtypes.bool, True)
+  # TODO: check for overflow
+  if IGNORE_OOB or isinstance(idx.dtype, ImageDType) or (sz := idx.src[0].ptrdtype.size) == -1: return True
+  # We can use UOp min/max to do a faster check, but it can give false positive since its not an exact bound and doesn't consider the mask
+  if 0<=idx.src[1].vmin and idx.src[1].vmax<sz: return True
+  mask = idx.src[2]&gate if len(idx.src)==3 else gate
+
+  # WEBGPU has a BITCAST in the index. TODO: fix
+  if any(x.op is Ops.BITCAST for x in idx.toposort()): return True
+
+  if not z3_imported: raise ImportError("z3 >= 4.12.4 is required for bounds checking, try IGNORE_OOB=0 or \"pip install 'z3-solver>=4.12.4\"")
+  solver = z3.Solver(ctx=z3.Context())
+  z3_idx, z3_mask = uops_to_z3(solver, idx.src[1], mask)
+  solver.add(z3_mask)
+  with cpu_profile("validate index with z3", "TINY"):
+    if solver.check((z3_idx<0)|(sz<=z3_idx)) == z3.sat:
+      print(f"idx={idx.src[1].render(simplify=False)}")
+      print(f"mask & gate={mask.render(simplify=False)}")
+      print(f"# OUT OF BOUNDS ACCESS: at {solver.model()} INDEX not in 0 - {sz}\nconstraints = {solver}")
+      return False
+  return True

tinygrad/viz/index.html

@@ -41,6 +41,7 @@
     color: #4a90e2;
     text-decoration: underline;
     cursor: pointer;
+    display: block;
   }
   ul {
     padding: 0;
@@ -148,10 +149,10 @@
     position: relative;
     height: 100%;
   }
-  .metadata > * + *, .rewrite-container > * + *, .ctx-list > * + * {
+  .metadata > * + *, .info > * + *, .rewrite-container > * + *, .ctx-list > * + * {
     margin-top: 12px;
   }
-  ul > * + * {
+  ul > * + *, .args > * + * {
     margin-top: 4px;
   }
   .graph {

tinygrad/viz/js/index.js

@@ -72,7 +72,7 @@ function renderDag(graph, additions, recenter) {
     d3.select("#graph-svg").on("click", () => d3.selectAll(".highlight").classed("highlight", false));
     const nodes = d3.select("#nodes").selectAll("g").data(g.nodes().map(id => g.node(id)), d => d).join("g").attr("class", d => d.className ?? "node")
       .attr("transform", d => `translate(${d.x},${d.y})`).classed("clickable", d => d.ref != null).on("click", (e,d) => {
-        if (d.ref != null) return setCtxWithHistory(d.ref);
+        if (d.ref != null) return switchCtx(d.ref);
         const parents = g.predecessors(d.id);
         const children = g.successors(d.id);
         if (parents == null && children == null) return;
@@ -246,18 +246,14 @@ async function renderProfiler() {
           const stepIdx = ctxs[ref.ctx+1].steps.findIndex((s, i) => i >= start && s.name == e.name);
           if (stepIdx !== -1) { ref.step = stepIdx; shapeRef = ref; }
         }
-        const html = document.createElement("div");
-        html.appendChild(tabulate([["Name", colored(e.name)], ["Duration", formatTime(e.dur)], ["Start Time", formatTime(e.st)]]).node());
-        const argsDiv = document.createElement("div"); argsDiv.id = "args"; html.appendChild(document.createElement("br")); html.appendChild(argsDiv);
-        if (e.info != null) html.appendChild(document.createElement("p")).innerText = "\n"+e.info;
-        if (shapeRef != null) {
-          const a = html.appendChild(document.createElement("a"));
-          a.innerText = "\nView codegen rewrite";
-          a.onclick = () => setCtxWithHistory(shapeRef.ctx, shapeRef.step);
-        }
+        const html = d3.create("div").classed("info", true);
+        html.append(() => tabulate([["Name", colored(e.name)], ["Duration", formatTime(e.dur)], ["Start Time", formatTime(e.st)]]).node());
+        html.append("div").classed("args", true);
+        if (e.info != null) html.append("p").style("white-space", "pre-wrap").text(e.info);
+        if (shapeRef != null) html.append("a").text("View codegen rewrite").on("click", () => switchCtx(shapeRef.ctx, shapeRef.step));
         // tiny device events go straight to the rewrite rule
         const key = k.startsWith("TINY") ? null : `${k}-${j}`;
-        if (key != null) shapeMetadata.set(key, html);
+        if (key != null) shapeMetadata.set(key, html.node());
         const arg = { tooltipText:colored(e.name).outerHTML+"\n"+formatTime(e.dur)+(e.info != null ? "\n"+e.info : ""), key, ...shapeRef };
         if (e.key != null) shapeMap.set(e.key, arg);
         // offset y by depth
@@ -298,23 +294,22 @@ async function renderProfiler() {
       for (const [num, {dtype, sz, nbytes, y, x:steps, users}] of buf_shapes) {
         const x = steps.map(s => timestamps[s]);
         const dur = x.at(-1)-x[0];
-        const html = document.createElement("div");
+        const html = d3.create("div").classed("info", true);
         const rows = [["DType", dtype], ["Len", formatUnit(sz)], ["Size", formatUnit(nbytes, "B")], ["Lifetime", formatTime(dur)]];
         if (users != null) rows.push(["Users", users.length]);
-        const info = html.appendChild(tabulate(rows).node());
-        const arg = {tooltipText:info.outerHTML, key:`${k}-${num}`};
+        const info = html.append(() => tabulate(rows).node());
+        const arg = {tooltipText:info.node().outerHTML, key:`${k}-${num}`};
+        const kernels = html.append("div").classed("args", true);
         for (let u=0; u<users?.length; u++) {
-          const p = html.appendChild(document.createElement("p")); p.style.marginTop = "4px";
           const { repr, num, mode, shape } = users[u];
           const bufInfo = `${mode == 2 ? 'read+write' : mode == 1 ? 'write' : 'read'}@data${num}`
-          p.appendChild(colored(`[${u}] ${repr} ${bufInfo}`));
+          const p = kernels.append("p").append(() => colored(`[${u}] ${repr} ${bufInfo}`));
           const metadata = shape?.tooltipText?.split("\n").at(-1);
-          if (metadata != null) p.appendChild(document.createElement("span")).innerText = "\n"+metadata;
+          if (metadata != null) p.append("span").text(" "+metadata);
           if (shape != null) {
-            p.style.cursor = "pointer";
-            p.onclick = () => focusShape(shape);
-            const args = shapeMetadata.get(shape.key).querySelector("#args");
-            const bufArg = d3.create("p").text(`${bufInfo} ${rows[2][1]}`).style("cursor", "pointer").style("margin-top", "4px").on("click", () => {
+            p.style("cursor", "pointer").on("click", () => focusShape(shape))
+            const args = shapeMetadata.get(shape.key).querySelector(".args");
+            const bufArg = d3.create("p").text(`${bufInfo} ${rows[2][1]}`).style("cursor", "pointer").on("click", () => {
               const device = document.getElementById(k);
               if (!isExpanded(device)) device.click();
               focusShape(arg);
@@ -325,7 +320,7 @@ async function renderProfiler() {
             args.insertBefore(bufArg, before);
           }
         }
-        shapeMetadata.set(arg.key, html)
+        shapeMetadata.set(arg.key, html.node())
         shapes.push({ x, y0:y.map(yscale), y1:y.map(y0 => yscale(y0+nbytes)), arg, fillColor:cycleColors(colorScheme.BUFFER, shapes.length) });
       }
       // generic polygon merger
@@ -487,7 +482,7 @@ async function renderProfiler() {
   canvas.addEventListener("click", e => {
     e.preventDefault();
     const foundRect = findRectAtPosition(e.clientX, e.clientY);
-    if (foundRect?.step != null && foundRect?.key == null) { return setCtxWithHistory(foundRect.ctx, foundRect.step); }
+    if (foundRect?.step != null && foundRect?.key == null) { return switchCtx(foundRect.ctx, foundRect.step); }
     if (foundRect?.key != focusedShape) { focusShape(foundRect); }
   });
 
@@ -585,7 +580,10 @@ function setState(ns) {
   // update element styles if needed
   const { ctx, step } = select(state.currentCtx, state.currentStep);
   toggleCls(prevCtx, ctx, "expanded", state.expandSteps);
-  if (ctx?.id !== prevCtx?.id) toggleCls(prevCtx, ctx, "active");
+  if (ctx?.id !== prevCtx?.id) {
+    saveToHistory({ currentCtx:deselect(prevCtx).ctx, currentRewrite:0, currentStep:0, expandSteps:false });
+    toggleCls(prevCtx, ctx, "active");
+  }
   if (ctx?.id !== prevCtx?.id || step?.id !== prevStep?.id) {
     toggleCls(prevStep, step, "active");
     // walk the tree back until all parents expanded so that the child is visible
@@ -607,11 +605,8 @@ function saveToHistory(ns) {
   history.pushState(ns, "");
 }
 
-// set a new context and keep the old one in browser history
-function setCtxWithHistory(newCtx, step=0) {
-  saveToHistory(state);
-  setState({ expandSteps:true, currentCtx:newCtx+1, currentStep:step, currentRewrite:0 });
-}
+// switch to the start of a new graph and expand all the steps
+const switchCtx = (newCtx, step) => setState({ expandSteps:true, currentCtx:newCtx+1, currentStep:step ?? 0, currentRewrite:0 });
 
 window.addEventListener("popstate", (e) => {
   if (e.state?.shape != null) return focusShape({ key:e.state?.shape });