reraise is fine

it was using float16 inputs, now it's uint16 then convert to bf16

.github/actions/setup-tinygrad/action.yml

@@ -225,13 +225,22 @@ runs:
     - name: Install gpuocelot dependencies (MacOS)
       if: inputs.ocelot == 'true' && runner.os == 'macOS'
       shell: bash
-      run: brew install --quiet cmake ninja llvm@15 zlib glew flex bison boost zstd ncurses
+      run: |
+        pkgs=(cmake ninja llvm@15 zlib glew flex bison [email protected] zstd ncurses)
+        for f in "${pkgs[@]}"; do
+          brew ls --versions "$f" >/dev/null 2>&1 || brew install --quiet "$f"
+        done
+
+        # Fix boost 1.85 for gpuocelot
+        ln -s /opt/homebrew/opt/[email protected] /opt/homebrew/opt/boost || true
+        ln -s /opt/homebrew/opt/boost/lib/libboost_atomic-mt.dylib /opt/homebrew/opt/boost/lib/libboost_atomic.dylib || true
+        ln -s /opt/homebrew/opt/boost/lib/libboost_thread-mt.dylib /opt/homebrew/opt/boost/lib/libboost_thread.dylib || true
     - name: Cache gpuocelot
       if: inputs.ocelot == 'true'
       id: cache-build
       uses: actions/cache@v4
       env:
-        cache-name: cache-gpuocelot-build
+        cache-name: cache-gpuocelot-build-1
       with:
         path: ${{ github.workspace }}/gpuocelot/ocelot
         key: ${{ runner.os }}-gpuocelot-b16039dc940dc6bc4ea0a98380495769ff35ed99-rebuild-${{ env.BUILD_CACHE_VERSION }}
@@ -244,7 +253,8 @@ runs:
         git checkout b16039dc940dc6bc4ea0a98380495769ff35ed99
         mkdir build
         cd build
-        cmake .. -Wno-dev -G Ninja -DOCELOT_BUILD_TOOLS=OFF -DCMAKE_BUILD_ALWAYS=0 -DBUILD_TESTS_CUDA=OFF -DCMAKE_POLICY_VERSION_MINIMUM=3.5
+        cmake .. -Wno-dev -G Ninja -DOCELOT_BUILD_TOOLS=OFF -DCMAKE_BUILD_ALWAYS=0 -DBUILD_TESTS_CUDA=OFF \
+          -DBoost_INCLUDE_DIR=$(brew --prefix boost)/include -DBoost_LIBRARY_DIR=$(brew --prefix boost)/lib -DCMAKE_POLICY_VERSION_MINIMUM=3.5
         ninja
     - name: Install gpuocelot
       if: inputs.ocelot == 'true'

.github/workflows/benchmark.yml

@@ -688,6 +688,10 @@ jobs:
       run: DEBUG=2 AMD=1 python -m pytest -rA test/test_tiny.py
     - name: Test DISK copy time
       run: AMD=1 TESTFILE=/raid/downloads/llama3-8b-sfr/model-00001-of-00004.safetensors python3 test/external/external_benchmark_disk_raw.py
+    - name: Test CPU copy time
+      run: |
+        AMD=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyDefaulttoCPUJit
+        AMD=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyCPUtoDefaultJit
     - name: Run full CIFAR training w 1 GPU
       run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF LATEWINO=1 STEPS=1000 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee am_train_cifar_one_gpu.txt
     # TODO: enable
@@ -745,6 +749,10 @@ jobs:
       run: NV=1 ALLOW_TF32=1 python3 test/test_linearizer.py TestLinearizer.test_tensor_cores TestLinearizer.test_tensor_cores_padded TestLinearizer.test_tensor_cores_padded_uops
     - name: Test DISK copy time
       run: NV=1 TESTFILE=/raid/downloads/llama3-8b-sfr/model-00001-of-00004.safetensors python3 test/external/external_benchmark_disk_raw.py
+    - name: Test CPU copy time
+      run: |
+        NV=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyDefaulttoCPUJit
+        NV=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyCPUtoDefaultJit
     - name: Test LLAMA-3
       run: BENCHMARK_LOG=llama3_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --benchmark --temperature 0 | tee nv_llama3_beam.txt
     - name: Run full CIFAR training w 1 GPU

.github/workflows/test.yml

@@ -343,6 +343,8 @@ jobs:
       run: |
         python -m mypy --strict-equality --lineprecision-report .
         cat lineprecision.txt
+    - name: Run TYPED=1
+      run: TYPED=1 python -c "import tinygrad"
 
   unittest:
     name: Unit Tests
@@ -613,8 +615,8 @@ jobs:
           -k "not test_symbolic_arange_sym_step and not test_threefry_doesnt_use_long" \
           test/test_tiny.py test/test_rangeify.py test/test_ops.py test/test_tensor_variable.py \
           test/test_outerworld_range.py test/test_sample.py test/test_randomness.py test/test_tensor_data.py
-    - name: Test CPU=1 RANGEIFY=1 PARTIAL_CONTIG=1
-      run: PARTIAL_CONTIG=1 CPU=1 RANGEIFY=1 python3 -m pytest -n auto test/test_tiny.py test/test_rangeify.py test/test_ops.py --durations 20
+    - name: Test CPU=1 RANGEIFY=2
+      run: CPU=1 RANGEIFY=2 python3 -m pytest -n auto test/test_tiny.py test/test_rangeify.py test/test_ops.py --durations 20
     - name: Test LLVM=1 RANGEIFY=1 (slow tests)
       run: LLVM=1 RANGEIFY=1 python3 -m pytest -n auto test/models/test_mnist.py --durations 20
 

.pylintrc

@@ -54,11 +54,12 @@ confidence=
 # --enable=similarities". If you want to run only the classes checker, but have
 # no Warning level messages displayed, use"--disable=all --enable=classes
 # --disable=W"
-disable=C,R,W0613,W0511,W0212,W0201,W0106,W0603,W0621,W0703,W1201,W1203,E1136,W1514,E1101,W0221,W0105,E0401,abstract-method
+disable=C,R,W0613,W0511,W0212,W0201,W0106,W0603,W0621,W0703,W1201,W1203,E1136,W1514,E1101,W0221,W0105,E0401,abstract-method,W0707
 # E1101 for function binding
 # W0221 for Function class
 # W0105 for comment strings
 # E0401 for missing imports
+# W0707 for not reraising
 
 # Enable the message, report, category or checker with the given id(s). You can
 # either give multiple identifier separated by comma (,) or put this option

docs/tensor/elementwise.md

@@ -78,6 +78,7 @@ Elementwise ops operate on a per element basis. They don't change the shape of t
 ::: tinygrad.Tensor.minimum
 ::: tinygrad.Tensor.where
 ::: tinygrad.Tensor.copysign
+::: tinygrad.Tensor.logaddexp
 
 ## Casting Ops
 

docs/tinybox.md

@@ -6,7 +6,7 @@ If you don't have a tinybox and you want one, see [tinygrad.org](https://tinygra
 
 ## Welcome
 
-Welcome to your tinybox! The tinybox is the universal system purpose-built for all AI infrastructure and workloads, from training to inference. The red box includes six 7900XTX GPUs, and the green box includes six 4090 GPUs. Whether you bought a red one or a green one, we want you to love it.
+Welcome to your tinybox! The tinybox is the universal system purpose-built for all AI infrastructure and workloads, from training to inference. The red box includes six 7900XTX GPUs, the green box includes six 4090 GPUs, and the green v2 box includes four 5090 GPUs. Whether you bought a red one or a green one, we want you to love it.
 
 We don't have a stupid cloud service, you don't have to create a tiny account to set it up, and we aren't tracking how you use the box. We're just happy you bought one. This petaflop is your petaflop.
 

examples/beautiful_cifar.py

@@ -2,7 +2,6 @@ import time
 start_tm = time.perf_counter()
 import math
 from typing import Tuple, cast
-import numpy as np
 from tinygrad import Tensor, nn, GlobalCounters, TinyJit, dtypes, Device
 from tinygrad.helpers import partition, trange, getenv, Context
 from extra.lr_scheduler import OneCycleLR
@@ -150,13 +149,12 @@ if __name__ == "__main__":
       acc.append((out.argmax(-1) == Y).sum() / eval_batchsize)
     return Tensor.stack(*loss).mean() / (batchsize*loss_batchsize_scaler), Tensor.stack(*acc).mean()
 
-  np.random.seed(1337)
+  Tensor.manual_seed(1337)
+  num_train_samples = X_train.shape[0]
+
   for epoch in range(math.ceil(hyp['misc']['train_epochs'])):
-    # TODO: move to tinygrad
     gst = time.perf_counter()
-    idxs = np.arange(X_train.shape[0])
-    np.random.shuffle(idxs)
-    tidxs = Tensor(idxs, dtype='int')[:num_steps_per_epoch*batchsize].reshape(num_steps_per_epoch, batchsize)  # NOTE: long doesn't fold
+    tidxs = Tensor.randperm(num_train_samples, dtype='int')[:num_steps_per_epoch*batchsize].reshape(num_steps_per_epoch, batchsize)
     train_loss:float = 0
     for epoch_step in (t:=trange(num_steps_per_epoch)):
       st = time.perf_counter()

examples/hlb_cifar10.py

@@ -118,7 +118,7 @@ class SpeedyResNet:
 # hyper-parameters were exactly the same as the original repo
 bias_scaler = 58
 hyp = {
-  'seed' : 200,
+  'seed' : 201,
   'opt': {
     'bias_lr':            1.76 * bias_scaler/512,
     'non_bias_lr':        1.76 / 512,

examples/mlperf/dataloader.py

@@ -758,6 +758,27 @@ def batch_load_llama3(bs:int, samples:int, seqlen:int, base_dir:Path, seed:int=0
       batch.append(tokens)
     yield Tensor.stack(batch, dim=0)
 
+def batch_load_llama3_small(bs:int, samples:int, seqlen:int, base_dir:Path, seed:int=0, val:bool=True):
+  if val:
+    dataset = BlendedGPTDataset([
+      base_dir / "c4-validation-91205-samples.en_text_document",
+    ], [
+      1.0
+    ], samples, seqlen, seed, False)
+  else:
+    dataset = BlendedGPTDataset([
+      base_dir / "c4-train.en_6_text_document",
+    ], [
+      1.0
+    ], samples, seqlen, seed, True)
+
+  for b in range(math.ceil(samples / bs)):
+    batch = []
+    for i in range(bs):
+      tokens = dataset.get(b * bs + i)
+      batch.append(tokens)
+    yield Tensor.stack(batch, dim=0)
+
 if __name__ == "__main__":
   def load_unet3d(val):
     assert not val, "validation set is not supported due to different sizes on inputs"

examples/mlperf/model_eval.py

@@ -243,31 +243,49 @@ def eval_mrcnn():
 
 def eval_llama3():
   from extra.models.llama import Transformer
-  from examples.llama3 import MODEL_PARAMS
+  from examples.llama3 import MODEL_PARAMS, load, convert_from_huggingface
   from tinygrad.helpers import tqdm
 
-  bs = 4
-  sequence_length = 512
+  BASEDIR = Path(getenv("BASEDIR", "/raid/datasets/c4/"))
+  BS = getenv("BS", 4)
+  SMALL = getenv("SMALL", 0)
+  SEQLEN = getenv("SEQLEN", 8192)
+  MODEL_PATH = Path(getenv("MODEL_PATH", "/raid/weights/llama31_8b/"))
 
-  model = Transformer(**(MODEL_PARAMS[getenv("LLAMA3_SIZE", "8B")]["args"]|{"vocab_size": 32000}), max_context=sequence_length, jit=False, disable_kv_cache=True)
+  params = MODEL_PARAMS[getenv("LLAMA3_SIZE", "8B")]["args"]
+  params = params | {"vocab_size": 32000} if not SMALL else params
+  if (llama_layers:=getenv("LLAMA_LAYERS")) != 0: params['n_layers'] = llama_layers
+  model = Transformer(**params, max_context=SEQLEN, jit=False, disable_kv_cache=True)
+
+  # load weights
+  weights = load(str(MODEL_PATH / "model.safetensors.index.json"))
+  if "model.embed_tokens.weight" in weights:
+    print("converting from huggingface format")
+    weights = convert_from_huggingface(weights, params["n_layers"], params["n_heads"], params["n_kv_heads"])
+
+  load_state_dict(model, weights, strict=False, consume=True)
 
   @TinyJit
   def eval_step(model, tokens):
     logits:Tensor = model(tokens[:, :-1], start_pos=0, temperature=math.nan)
     loss = logits.sparse_categorical_crossentropy(tokens[:, 1:])
-    return loss.flatten()
+    return loss.flatten().float()
 
-  from examples.mlperf.dataloader import batch_load_llama3
-  iter = batch_load_llama3(bs, 5760, sequence_length, Path(getenv("BASEDIR", "/raid/datasets/c4/")), True)
+  if SMALL:
+    from examples.mlperf.dataloader import batch_load_llama3_small
+    iter = batch_load_llama3_small(BS, 5760, SEQLEN, BASEDIR, val=True)
+  else:
+    from examples.mlperf.dataloader import batch_load_llama3
+    iter = batch_load_llama3(BS, 5760, SEQLEN, BASEDIR, val=True)
 
   losses = []
-  for tokens in tqdm(iter, total=5760//bs):
+  for tokens in tqdm(iter, total=5760//BS):
     GlobalCounters.reset()
     losses += eval_step(model, tokens).tolist()
     tqdm.write(f"loss: {np.mean(losses)}")
 
-  log_perplexity = Tensor(losses).mean()
-  print(f"Log Perplexity: {log_perplexity.item()}")
+  log_perplexity = np.mean(losses)
+  print(f"Log Perplexity: {log_perplexity}")
 
 if __name__ == "__main__":
   # inference only

examples/mlperf/model_train.py

@@ -1290,12 +1290,14 @@ def train_llama3():
   from examples.mlperf.lr_schedulers import CosineAnnealingLRWithWarmup
 
   config = {}
+  BASEDIR            = config["BASEDIR"]                = Path(getenv("BASEDIR", "/raid/datasets/c4/"))
   BS                 = config["BS"]                     = getenv("BS", 16)
   grad_acc           = config["GRADIENT_ACC_STEPS"]     = getenv("GRADIENT_ACC_STEPS", 1)
   GBS                = config["GLOBAL_BATCH_SIZE"]      = BS * grad_acc
   SEED               = config["SEED"]                   = getenv("SEED", 5760)
   SEQLEN             = config["SEQLEN"]                 = getenv("SEQLEN", 8192)
   TRAIN_ON_VAL       = config["TRAIN_ON_VAL"]           = getenv("TRAIN_ON_VAL", 0)
+  SMALL              = config["SMALL"]                  = getenv("SMALL", 0)
   SAMPLES            = config["SAMPLES"]                = getenv("SAMPLES", 5_760 if TRAIN_ON_VAL else 1_200_000 * 1152)
   EVAL_FREQ          = config["EVAL_FREQ"]              = getenv("EVAL_FREQ", 46080)
   EVAL_BS            = config["EVAL_BS"]                = getenv("EVAL_BS", 16)
@@ -1311,13 +1313,14 @@ def train_llama3():
 
   opt_gradient_clip_norm = 1.0
   opt_learning_rate_warmup_steps = getenv("WARMUP_STEPS", math.ceil(8000 * 1152 / GBS))
-  opt_learning_rate_decay_steps = getenv("DECAY_STEPS", math.ceil(1_200_000 * 1152 / GBS) - opt_learning_rate_warmup_steps)
+  opt_learning_rate_decay_steps = getenv("MAX_STEPS", math.ceil(1_200_000 * 1152 / GBS)) - opt_learning_rate_warmup_steps
   opt_base_learning_rate = getenv("LR", 8e-5 * GBS / 1152)  # NOTE: cannot change for benchmark
-  opt_end_learning_rate = 8e-7
+  opt_end_learning_rate = getenv("END_LR", 8e-7)
 
   # TODO: confirm weights are in bf16
   # vocab_size from the mixtral tokenizer
-  params = MODEL_PARAMS[getenv("LLAMA3_SIZE", "8B")]["args"]|{"vocab_size": 32000}
+  params = MODEL_PARAMS[getenv("LLAMA3_SIZE", "8B")]["args"]
+  params = params | {"vocab_size": 32000} if not SMALL else params
   if (llama_layers:=getenv("LLAMA_LAYERS")) != 0: params['n_layers'] = llama_layers
   model = Transformer(**params, max_context=SEQLEN, jit=False, disable_kv_cache=True)
 
@@ -1403,21 +1406,29 @@ def train_llama3():
   # ** data iters **
   def fake_data(bs, samples):
     for _ in range(samples // bs):
-      yield Tensor.randint(bs, SEQLEN + 1, low=0, high=32000, dtype=dtypes.int32, device=Device.DEFAULT)
+      yield Tensor.randint(bs, SEQLEN + 1, low=0, high=params["vocab_size"], dtype=dtypes.int32, device=Device.DEFAULT)
 
   def get_train_iter():
     if getenv("FAKEDATA", 0):
       return fake_data(GBS, SAMPLES)
     else:
-      from examples.mlperf.dataloader import batch_load_llama3
-      return batch_load_llama3(GBS, SAMPLES, SEQLEN, Path(getenv("BASEDIR", "/raid/datasets/c4/")), seed=SEED, val=bool(TRAIN_ON_VAL))
+      if SMALL:
+        from examples.mlperf.dataloader import batch_load_llama3_small
+        return batch_load_llama3_small(GBS, SAMPLES, SEQLEN, BASEDIR, seed=SEED, val=bool(TRAIN_ON_VAL))
+      else:
+        from examples.mlperf.dataloader import batch_load_llama3
+        return batch_load_llama3(GBS, SAMPLES, SEQLEN, BASEDIR, seed=SEED, val=bool(TRAIN_ON_VAL))
 
   def get_eval_iter():
     if getenv("FAKEDATA", 0):
       return fake_data(EVAL_BS, 5760)
     else:
-      from examples.mlperf.dataloader import batch_load_llama3
-      return batch_load_llama3(EVAL_BS, 5760, SEQLEN, Path(getenv("BASEDIR", "/raid/datasets/c4/")), seed=SEED, val=True)
+      if SMALL:
+        from examples.mlperf.dataloader import batch_load_llama3_small
+        return batch_load_llama3_small(EVAL_BS, 5760, SEQLEN, BASEDIR, val=True)
+      else:
+        from examples.mlperf.dataloader import batch_load_llama3
+        return batch_load_llama3(EVAL_BS, 5760, SEQLEN, BASEDIR, val=True)
 
   iter = get_train_iter()
   i, sequences_seen = 0, 0
@@ -1426,7 +1437,7 @@ def train_llama3():
     GlobalCounters.reset()
     loss, lr = train_step(model, tokens, grad_acc)
     loss = loss.float().item()
-    # above as tqdm.write f-string
+
     tqdm.write(f"{loss:.4f} loss, {lr.item():.12f} LR, {GlobalCounters.mem_used / 1e9:.2f} GB used, {time.perf_counter()-t:.2f} s")
     if (fname:=getenv("LOSS_FILE", "")):
       with open(fname, "a") as f:

extra/amdpci/proclogs.py

@@ -37,7 +37,7 @@ def main():
   dev = PCIIface(None, 0)
   for x, y in dev.dev_impl.__dict__.items():
     if isinstance(y, AMRegister):
-      for inst, addr in y.addr.keys(): reg_names[addr] = f"{x}, xcc={inst}"
+      for inst, addr in y.addr.items(): reg_names[addr] = f"{x}, xcc={inst}"
 
   with open(sys.argv[1], 'r') as f:
     log_content = log_content_them = f.read()

extra/gemm/amd_uop_matmul.py

@@ -5,7 +5,7 @@ from tinygrad.dtype import AddrSpace
 from tinygrad.helpers import getenv, colored, prod, unwrap
 from tinygrad.shape.shapetracker import ShapeTracker, View
 from tinygrad.shape.view import strides_for_shape
-from tinygrad.codegen.opt.kernel import axis_colors
+from tinygrad.codegen.opt.kernel import axis_colors, Opt, OptOps
 from tinygrad.codegen.opt.swizzler import merge_views, view_left
 
 def to_colored(full_shape, axis_types): return '_'.join([colored(str(s), axis_colors[at]) for s,at in zip(full_shape, axis_types)])
@@ -44,13 +44,28 @@ pm = PatternMatcher([
   (UPat(Ops.VIEW, src=(UPat(Ops.REDUCE_AXIS, src=(UPat.var("src"),), name="r"),), name="view"), swizzle_reduceop),
 ])
 
+def rangeify_kernel3():
+  a = Tensor.empty(N,N)
+  b = Tensor.empty(N,N)
+  c = a@b
+  #c = c.reshape((32,2,16,4,32,2,16,4)).contiguous()
+  with Context(RANGEIFY=1):
+    sink = c.schedule()[-1].ast
+  #print(sink)
+
+  opts  = [Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.LOCAL, 0, 16), Opt(OptOps.UPCAST, 0, 2)]
+  opts += [Opt(OptOps.UPCAST, 1, 4), Opt(OptOps.LOCAL, 1, 16), Opt(OptOps.UPCAST, 1, 2)]
+  opts += [Opt(OptOps.UNROLL, 0, 8)]
+
+  return sink.replace(arg=KernelInfo(opts_to_apply=tuple(opts)))
+
 def top_spec_kernel3():
   a = Tensor.empty(N,N)
   b = Tensor.empty(N,N)
   c = a@b
   sink = c.schedule()[-1].ast
   L = 16
-  sink = sink.reshape((N//L, L, N//L, L)) #.lift({0:UOp.range(dtypes.int, N//BM, 0), 2:UOp.range(dtypes.int, N//BN, 1)})
+  sink = sink.reshape((N//L, L, N//L, L)) #.lift({0:UOp.range(N//BM, 0), 2:UOp.range(N//BN, 1)})
   sink = graph_rewrite(sink, view_left+pm)
   axis_types = (AxisType.GLOBAL, AxisType.LOCAL, AxisType.GLOBAL, AxisType.LOCAL, AxisType.REDUCE)
   return sink.replace(arg=KernelInfo(name="top_"+to_colored(sink.full_shape, axis_types), axis_types=axis_types))
@@ -171,7 +186,7 @@ def hand_spec_kernel3(kernel4=getenv("K4", 0), kernel5=getenv("K5", 0)):
 
   c_regs = UOp(Ops.DEFINE_REG, dtypes.float.ptr(TM * nbIterWaveM * TN * nbIterWaveN), arg=2)
 
-  i = UOp.range(dtypes.int, c_regs.dtype.size, 16)
+  i = UOp.range(c_regs.dtype.size, 16)
   init_store = c_regs[i].store(UOp.const(dtypes.float, 0.0), i)
 
   if kernel4:
@@ -182,53 +197,53 @@ def hand_spec_kernel3(kernel4=getenv("K4", 0), kernel5=getenv("K5", 0)):
     kId = 0
 
     # load from globals into locals
-    i = UOp.range(dtypes.int, nbReadsB, 0)
+    i = UOp.range(nbReadsB, 0)
     index_x = BN * blockIdx_x + rBIdx
     index_y = rBIdy + i * strideReadB + kId
     Bs_store = Bs[(index_y % BK) * BN + index_x % BN].store(b[N * index_y + index_x].load(), i)
 
-    i = UOp.range(dtypes.int, nbReadsA, 1)
+    i = UOp.range(nbReadsA, 1)
     index_x = rAIdx + kId
     index_y = BM * blockIdx_y + rAIdy + i * strideReadA
     As_store = As[(index_x % BK) * BM_As_stride + index_y % BM].store(a[N * index_y + index_x].load(), i)
 
     # iterate over the middle chunk
-    kId_range = UOp.range(dtypes.int, N//BK-1, 2)
+    kId_range = UOp.range(N//BK-1, 2)
     kId = kId_range*BK
 
     barrier = UOp.barrier(As_store, Bs_store)
 
     # load from globals into registers (next round)
-    i = UOp.range(dtypes.int, nbReadsB, 3)
+    i = UOp.range(nbReadsB, 3)
     index_x = BN * blockIdx_x + rBIdx
     index_y = rBIdy + i * strideReadB + kId + BK
     regB_store = regB[i].store(b[N * index_y + index_x].load(), i)
 
-    i = UOp.range(dtypes.int, nbReadsA, 4)
+    i = UOp.range(nbReadsA, 4)
     index_x = rAIdx + kId + BK
     index_y = BM * blockIdx_y + rAIdy + i * strideReadA
     regA_store = regA[i].store(a[N * index_y + index_x].load(), i)
 
     def inner_loop(first_range, inp_dep=()):
       # inner unroll
-      k = UOp.range(dtypes.int, BK, first_range+0)
+      k = UOp.range(BK, first_range+0)
 
       # load from locals into registers
-      iterWave = UOp.range(dtypes.int, nbIterWaveN, first_range+1)
-      i = UOp.range(dtypes.int, TN, first_range+2)
+      iterWave = UOp.range(nbIterWaveN, first_range+1)
+      i = UOp.range(TN, first_range+2)
       index = waveIdx * WN + iterWave * SUBWN + TN * idxInWave + i
       B_row_store = B_row[iterWave*TN + i].store(Bs[k*BN + index].load(*inp_dep), iterWave, i)
 
-      iterWave = UOp.range(dtypes.int, nbIterWaveM, first_range+3)
-      i = UOp.range(dtypes.int, TM, first_range+4)
+      iterWave = UOp.range(nbIterWaveM, first_range+3)
+      i = UOp.range(TM, first_range+4)
       index = waveIdy * WM + iterWave * SUBWM + TM * idyInWave + i
       A_col_store = A_col[iterWave*TM + i].store(As[k*BM_As_stride + index].load(*inp_dep), iterWave, i)
 
       # do the GEMM math
-      iterWaveM = UOp.range(dtypes.int, nbIterWaveM, first_range+5)
-      yt = UOp.range(dtypes.int, TM, first_range+6)
-      iterWaveN = UOp.range(dtypes.int, nbIterWaveN, first_range+7)
-      xt = UOp.range(dtypes.int, TN, first_range+8)
+      iterWaveM = UOp.range(nbIterWaveM, first_range+5)
+      yt = UOp.range(TM, first_range+6)
+      iterWaveN = UOp.range(nbIterWaveN, first_range+7)
+      xt = UOp.range(TN, first_range+8)
       x = iterWaveN * TN + xt
       y = iterWaveM * TM + yt
       c_regs_idx = c_regs[y * TN * nbIterWaveN + x]
@@ -241,12 +256,12 @@ def hand_spec_kernel3(kernel4=getenv("K4", 0), kernel5=getenv("K5", 0)):
     sink = inner_loop(5, (barrier, regB_store, regA_store)).barrier()
 
     # load from registers into locals
-    i = UOp.range(dtypes.int, nbReadsB, 14)
+    i = UOp.range(nbReadsB, 14)
     index_x = BN * blockIdx_x + rBIdx
     index_y = rBIdy + i * strideReadB + kId + BK
     Bs_store = Bs[(index_y % BK) * BN + index_x % BN].store(regB[i].load(sink), i, kId_range)
 
-    i = UOp.range(dtypes.int, nbReadsA, 15)
+    i = UOp.range(nbReadsA, 15)
     index_x = rAIdx + kId + BK
     index_y = BM * blockIdx_y + rAIdy + i * strideReadA
     As_store = As[(index_x % BK) * BM_As_stride + index_y % BM].store(regA[i].load(sink), i, kId_range)
@@ -254,40 +269,40 @@ def hand_spec_kernel3(kernel4=getenv("K4", 0), kernel5=getenv("K5", 0)):
     # final iteration without the copy
     sink = inner_loop(16, (UOp.barrier(Bs_store, As_store),))
   else:
-    kId_range = UOp.range(dtypes.int, N//BK, 0)
+    kId_range = UOp.range(N//BK, 0)
     kId = kId_range*BK
 
     # load from globals into locals
-    i = UOp.range(dtypes.int, nbReadsB, 1)
+    i = UOp.range(nbReadsB, 1)
     index_x = BN * blockIdx_x + rBIdx
     index_y = rBIdy + i * strideReadB + kId
     Bs_store = Bs[(index_y % BK) * BN + index_x % BN].store(b[N * index_y + index_x].load(), i)
 
-    i = UOp.range(dtypes.int, nbReadsA, 2)
+    i = UOp.range(nbReadsA, 2)
     index_x = rAIdx + kId
     index_y = BM * blockIdx_y + rAIdy + i * strideReadA
     As_store = As[(index_x % BK) * BM_As_stride + index_y % BM].store(a[N * index_y + index_x].load(), i)
 
     barrier = UOp.barrier(As_store, Bs_store)
 
-    k = UOp.range(dtypes.int, BK, 3)
+    k = UOp.range(BK, 3)
 
     # load from locals into registers
-    iterWave = UOp.range(dtypes.int, nbIterWaveN, 4)
-    i = UOp.range(dtypes.int, TN, 5)
+    iterWave = UOp.range(nbIterWaveN, 4)
+    i = UOp.range(TN, 5)
     index = waveIdx * WN + iterWave * SUBWN + TN * idxInWave + i
     B_row_store = B_row[iterWave*TN + i].store(Bs[k*BN + index].load(barrier), iterWave, i)
 
-    iterWave = UOp.range(dtypes.int, nbIterWaveM, 6)
-    i = UOp.range(dtypes.int, TM, 7)
+    iterWave = UOp.range(nbIterWaveM, 6)
+    i = UOp.range(TM, 7)
     index = waveIdy * WM + iterWave * SUBWM + TM * idyInWave + i
     A_col_store = A_col[iterWave*TM + i].store(As[k*BM_As_stride + index].load(barrier), iterWave, i)
 
     # do the GEMM math
-    iterWaveM = UOp.range(dtypes.int, nbIterWaveM, 8)
-    yt = UOp.range(dtypes.int, TM, 9)
-    iterWaveN = UOp.range(dtypes.int, nbIterWaveN, 10)
-    xt = UOp.range(dtypes.int, TN, 12)
+    iterWaveM = UOp.range(nbIterWaveM, 8)
+    yt = UOp.range(TM, 9)
+    iterWaveN = UOp.range(nbIterWaveN, 10)
+    xt = UOp.range(TN, 12)
     x = iterWaveN * TN + xt
     y = iterWaveM * TM + yt
     c_regs_idx = c_regs[y * TN * nbIterWaveN + x]
@@ -295,10 +310,10 @@ def hand_spec_kernel3(kernel4=getenv("K4", 0), kernel5=getenv("K5", 0)):
                             iterWaveM, iterWaveN, yt, xt, k, kId_range)
 
   # store c_regs into c
-  iterWaveM = UOp.range(dtypes.int, nbIterWaveM, 1000)
-  yt = UOp.range(dtypes.int, TM, 1001)
-  iterWaveN = UOp.range(dtypes.int, nbIterWaveN, 1002)
-  xt = UOp.range(dtypes.int, TN, 1003)
+  iterWaveM = UOp.range(nbIterWaveM, 1000)
+  yt = UOp.range(TM, 1001)
+  iterWaveN = UOp.range(nbIterWaveN, 1002)
+  xt = UOp.range(TN, 1003)
   xOut = blockIdx_x * BN + waveIdx * WN + iterWaveN * SUBWN + TN * idxInWave
   yOut = blockIdx_y * BM + waveIdy * WM + iterWaveM * SUBWM + TM * idyInWave
   indexC = N * (yOut + yt) + xOut + xt
@@ -309,10 +324,15 @@ def hand_spec_kernel3(kernel4=getenv("K4", 0), kernel5=getenv("K5", 0)):
 
 if __name__ == "__main__":
   HL = getenv("HL")
-  if HL == 2: hprg = top_spec_kernel3()
+  if HL == 3: hprg = rangeify_kernel3()
+  elif HL == 2: hprg = top_spec_kernel3()
   elif HL == 1: hprg = hl_spec_kernel3()
   else: hprg = hand_spec_kernel3()
-  prg = get_program(hprg, Device.default.renderer)
+  if HL == 3:
+    with Context(RANGEIFY=1, BLOCK_REORDER=0):
+      prg = get_program(hprg, Device.default.renderer)
+  else:
+    prg = get_program(hprg, Device.default.renderer)
   print(prg.src)
   if getenv("SRC"): exit(0)
   hrunner = CompiledRunner(prg)

extra/gemm/max_matmul.py

@@ -56,7 +56,7 @@ def randoms():
 def ast_to_cuda_prog(compiler, ast, opts):
   k = Kernel(ast)
   k.apply_opts(opts)
-  p = get_program(k.get_optimized_ast(), k.opts)
+  p = get_program(k.ast, k.opts, k.applied_opts)
   return CUDAProgram(device, p.function_name, compiler.compile(p.src))
 
 if __name__ == "__main__":

extra/gemm/tinygrad_nv_matmul.py

@@ -29,7 +29,7 @@ if __name__ == "__main__":
       Opt(op=OptOps.LOCAL, axis=0, amt=2),
     ]
   k.apply_opts(opts)
-  prg = get_program(k.get_optimized_ast(), k.opts)
+  prg = get_program(k.ast, k.opts, k.applied_opts)
   new_src = prg.src
   # can mod source here
   prg = replace(prg, src=new_src)

extra/optimization/test_beam_search.py

@@ -16,9 +16,9 @@ class TestBeamSearch(unittest.TestCase):
     BEAM.value = self.old_beam
 
   def test_variable_ast_beam(self):
-    with Context(IGNORE_OOB=1):
-      a = rand(3, 3).reshape((Variable("a", 1, 10).bind(3), 3))
-      a = (a+1).realize()
+    vi = Variable("a", 1, 10).bind(3)
+    a = rand(10, 3)[:vi]
+    a = (a+1).realize()
 
   def test_big_prime_number(self):
     a = rand(367, 367)
@@ -42,18 +42,16 @@ class TestBeamSearch(unittest.TestCase):
 
   def test_variable_big_prime_number(self):
     v = Variable("v", 1, 400).bind(367)
-    a = rand(367, 367)
-    b = rand(367, 367)
-    with Context(IGNORE_OOB=1):
-      c = (a.reshape(367, v) @ b.reshape(v, 367)).realize()
-      np.testing.assert_allclose(c.numpy(), a.numpy() @ b.numpy(), atol=1e-4, rtol=1e-4)
+    a = rand(367, 400)
+    b = rand(400, 367)
+    c = (a[:, :v] @ b[:v, :]).realize()
+    np.testing.assert_allclose(c.numpy(), a[:, :367].numpy() @ b[:367, :].numpy(), atol=1e-4, rtol=1e-4)
 
   def test_variable_shrink_prime_number(self):
     v = Variable("v", 1, 400).bind(367)
     a = rand(400, 367)
-    with Context(IGNORE_OOB=1):
-      b = (a.shrink(((0,v), None))+1).reshape(367,367).realize()
-      np.testing.assert_allclose(b.numpy(), a.numpy()[:367]+1, atol=1e-4, rtol=1e-4)
+    b = (a.shrink(((0,v), None))+1).reshape(367,367).realize()
+    np.testing.assert_allclose(b.numpy(), a.numpy()[:367]+1, atol=1e-4, rtol=1e-4)
 
   def test_no_mutate_rawbuffers(self):
     a = rand(3, 3).realize()

extra/remu/src/thread.rs

@@ -673,6 +673,7 @@ impl<'a> Thread<'a> {
                                         39 => f32::log2(s0),
                                         42 => 1.0 / s0,
                                         43 => 1.0 / s0,
+                                        46 => 1.0 / f32::sqrt(s0),
                                         51 => f32::sqrt(s0),
                                         _ => todo_instr!(instruction)?,
                                     }
@@ -1246,7 +1247,7 @@ impl<'a> Thread<'a> {
                             }
 
                             let ret = match op {
-                                257 | 259 | 299 | 260 | 261 | 264 | 272 | 392 | 426 | 531 | 537 | 540 | 551 | 567 | 796 => {
+                                257 | 259 | 299 | 260 | 261 | 264 | 272 | 392 | 426 | 430 | 531 | 537 | 540 | 551 | 567 | 796 => {
                                     let s0 = f32::from_bits(s0).negate(0, neg).absolute(0, abs);
                                     let s1 = f32::from_bits(s1).negate(1, neg).absolute(1, abs);
                                     let s2 = f32::from_bits(s2).negate(2, neg).absolute(2, abs);
@@ -1258,6 +1259,7 @@ impl<'a> Thread<'a> {
                                         272 => f32::max(s0, s1),
                                         299 => f32::mul_add(s0, s1, f32::from_bits(self.vec_reg[vdst])),
                                         426 => s0.recip(),
+                                        430 => 1.0 / f32::sqrt(s0),
                                         531 => f32::mul_add(s0, s1, s2),
                                         537 => f32::min(f32::min(s0, s1), s2),
                                         540 => f32::max(f32::max(s0, s1), s2),
@@ -2625,6 +2627,14 @@ mod test_vop1 {
         assert_eq!(thread.vec_reg[3], 1071644672);
     }
 
+    #[test]
+    fn test_v_rsq_f32() {
+        let mut thread = _helper_test_thread();
+        thread.vec_reg[0] = f32::to_bits(4.0);
+        r(&vec![0x7E005D00, END_PRG], &mut thread);
+        assert_eq!(f32::from_bits(thread.vec_reg[0]), 0.5);
+    }
+
     #[test]
     fn test_v_frexp_exp_i32_f64() {
         [(3573412790272.0, 42), (69.0, 7), (2.0, 2), (f64::NEG_INFINITY, 0)]

extra/replay_pkl.py

@@ -58,7 +58,7 @@ if __name__ == "__main__":
             GlobalCounters.kernel_count -= 1
 
         if not getenv("NOOPT"): k.apply_opts(hand_coded_optimizations(k))
-        p2 = get_program(k.get_optimized_ast(), k.opts)
+        p2 = get_program(k.ast, k.opts, k.applied_opts)
         new_ei = replace(ei, prg=CompiledRunner(p2))
         new_ei.run()
         new_jit.append(new_ei)

extra/torch_backend/backend.py

@@ -381,6 +381,7 @@ decomps = [
   aten.elu,  # elu has a scale + input_scale param
   aten.elu_backward,
   aten.softplus,
+  aten.logaddexp,
   aten.threshold,
   aten.nll_loss_forward,
   aten.nll_loss_backward,

setup.py

@@ -29,6 +29,7 @@ setup(name='tinygrad',
         'tinygrad.apps',
         'tinygrad.codegen',
         'tinygrad.codegen.opt',
+        'tinygrad.codegen.late',
         'tinygrad.engine',
         'tinygrad.frontend',
         'tinygrad.nn',
@@ -63,6 +64,7 @@ setup(name='tinygrad',
             "pre-commit",
             "ruff",
             "numpy",
+            "typeguard",
         ],
         #'mlperf': ["mlperf-logging @ git+https://github.com/mlperf/[email protected]"],
         'testing_minimal': testing_minimal,

test/external/external_benchmark_sdxl_softmax.py

@@ -24,5 +24,5 @@ if __name__ == "__main__":
   #k.apply_opt(Opt(OptOps.GROUP, 1, 32))
   #k.apply_opt(Opt(OptOps.GROUP, 0, 32))
   from tinygrad.engine.realize import CompiledRunner, ExecItem
-  run = CompiledRunner(prg:=get_program(k.get_optimized_ast(), k.opts))
+  run = CompiledRunner(prg:=get_program(k.ast, k.opts, k.applied_opts))
   ExecItem(run, si.bufs).run()

test/external/external_debug_metal_sd_conv.py

@@ -35,7 +35,7 @@ k = Kernel(ast)
 k.apply_opts(opts)
 bufs = bufs_from_lin(k)
 
-prg = CompiledRunner(get_program(k.get_optimized_ast(), k.opts))
+prg = CompiledRunner(get_program(k.ast, k.opts, k.applied_opts))
 
 for i in range(10):
   speed = prg(bufs, var_vals={}, wait=True)

test/external/external_test_onnx_backend.py

@@ -134,7 +134,6 @@ backend_test.exclude('test_simple_rnn_*')
 
 # no control flow
 # control flow uses AttributeProto.GRAPH
-backend_test.exclude('test_if_*')
 backend_test.exclude('test_loop*')
 backend_test.exclude('test_range_float_type_positive_delta_expanded_cpu') # requires loop
 backend_test.exclude('test_affine_grid_2d_align_corners_expanded_cpu')
@@ -183,6 +182,8 @@ backend_test.exclude('test_resize_downsample_scales_cubic_antialias_cpu') # anti
 backend_test.exclude('test_resize_downsample_sizes_cubic_antialias_cpu') # antialias not implemented
 backend_test.exclude('test_ai_onnx_ml_label_encoder_tensor_value_only_mapping_cpu') # bad data type string
 backend_test.exclude('test_ai_onnx_ml_label_encoder_tensor_mapping_cpu') # bad data type string
+backend_test.exclude('test_if_opt_cpu') # ValueError: 13 is not a valid AttributeType
+backend_test.exclude('test_if_seq_cpu') # NotImplementedError: op='SequenceConstruct' is not supported
 
 backend_test.exclude('test_scatternd_min_cpu') # min not yet supported
 backend_test.exclude('test_scatternd_max_cpu') # max not yet supported

test/external/external_test_onnx_ops.py

@@ -100,6 +100,25 @@ class TestMainOnnxOps(TestOnnxOps):
     self._test_resize_scales([0.01, 0.25, 0.5, 0.51, 0.6, 1.0, 1.5, 2.0, 3.5, 20.0], mode="cubic", exclude_outside=1)
     self._test_resize_scales([0.01, 0.25, 0.5, 0.51, 0.6, 1.0, 1.5, 2.0, 3.5, 20.0], mode="cubic", exclude_outside=0)
 
+  def _test_if(self, then_value, else_value):
+    then_out = onnx.helper.make_tensor_value_info("res", onnx.TensorProto.FLOAT, then_value.shape)
+    else_out = onnx.helper.make_tensor_value_info("res", onnx.TensorProto.FLOAT, else_value.shape)
+
+    then_const_node = onnx.helper.make_node("Constant", inputs=[], outputs=["res"], value=onnx.numpy_helper.from_array(then_value))
+    else_const_node = onnx.helper.make_node("Constant", inputs=[], outputs=["res"], value=onnx.numpy_helper.from_array(else_value))
+
+    then_body = onnx.helper.make_graph([then_const_node], "then_body", [], [then_out])
+    else_body = onnx.helper.make_graph([else_const_node], "else_body", [], [else_out])
+
+    self.helper_test_single_op("If", {"cond": np.array(False).astype(bool)}, {"then_branch": then_body, "else_branch": else_body}, ["res"])
+    self.helper_test_single_op("If", {"cond": np.array(True).astype(bool)}, {"then_branch": then_body, "else_branch": else_body}, ["res"])
+
+  def test_if_different_shapes_broadcastable(self):
+    self._test_if(np.array([[1], [2]]).astype(np.float32), np.array([[6, 5, 4, 3, 2, 1]]).astype(np.float32))
+
+  def test_if_different_shapes_not_broadcastable(self):
+    self._test_if(np.array([[1, 2, 3], [4, 5, 6]]).astype(np.float32), np.array([[6, 5, 4, 3, 2, 1]]).astype(np.float32))
+
   def test_resize_downsample_scales_linear_align_corners(self):
     # https://github.com/onnx/onnx/blob/main/docs/Operators.md#examples-131
     X = np.array([[[[1, 2, 3, 4], [5, 6, 7, 8]]]], dtype=np.float32)

test/external/fuzz_fast_idiv.py

@@ -1,8 +1,8 @@
 import random
 import z3
 from tinygrad import dtypes
-from tinygrad.uop.spec import z3_renderer, z3_cdiv
-from tinygrad.uop.ops import UOp, graph_rewrite
+from tinygrad.uop.spec import uops_to_z3, z3_cdiv
+from tinygrad.uop.ops import UOp
 from tinygrad.uop.decompositions import fast_idiv
 random.seed(42)
 
@@ -19,8 +19,7 @@ if __name__ == "__main__":
     if expr is None: continue
 
     solver = z3.Solver()
-    z3_sink = graph_rewrite(expr.sink(u), z3_renderer, ctx=(solver, {}))
-    z3_expr, x = z3_sink.src[0].arg, z3_sink.src[1].arg
+    z3_expr, x =uops_to_z3(solver, expr, u)
 
     if solver.check(z3_expr != z3_cdiv(x, d)) == z3.sat:
       assert False, f"Failed: {expr.render()} != x//{d} at x={solver.model()}\nx={u}\nd={d}\n{z3_expr=}\n{x/d=}"

test/external/fuzz_symbolic.py

@@ -1,8 +1,8 @@
 import random, operator
 import z3
 from tinygrad import Variable, dtypes
-from tinygrad.uop.ops import UOp, graph_rewrite
-from tinygrad.uop.spec import z3_renderer
+from tinygrad.uop.ops import UOp
+from tinygrad.uop.spec import uops_to_z3
 from tinygrad.helpers import DEBUG, Context
 
 seed = random.randint(0, 100)
@@ -57,8 +57,7 @@ if __name__ == "__main__":
 
     solver = z3.Solver()
     solver.set(timeout=5000)  # some expressions take very long verify, but its very unlikely they actually return sat
-    z3_sink = graph_rewrite(expr.sink(simplified_expr, u1, u2, u3), z3_renderer, ctx=(solver, {}))
-    z3_expr, z3_simplified_expr = z3_sink.src[0].arg, z3_sink.src[1].arg
+    z3_expr, z3_simplified_expr, v1, v2, v3 = uops_to_z3(solver, expr, simplified_expr, u1, u2, u3)
     check = solver.check(z3_simplified_expr != z3_expr)
     if check == z3.unknown and DEBUG>=1:
       skipped += 1
@@ -69,7 +68,6 @@ if __name__ == "__main__":
             f"expr = {expr.render(simplify=False)}\n")
     elif check == z3.sat:
       m = solver.model()
-      v1, v2, v3 = z3_sink.src[2].arg, z3_sink.src[3].arg, z3_sink.src[4].arg
       n1, n2, n3 = m[v1], m[v2], m[v3]
       u1_val, u2_val, u3_val = u1.const_like(n1.as_long()), u2.const_like(n2.as_long()), u3.const_like(n3.as_long())
       with Context(CORRECT_DIVMOD_FOLDING=1):

test/mockgpu/amd/amdgpu.py

@@ -1,5 +1,6 @@
 import ctypes, time
 from test.mockgpu.gpu import VirtGPU
+from test.mockgpu.helpers import _try_dlopen_remu
 from tinygrad.helpers import getbits, to_mv, init_c_struct_t
 import tinygrad.runtime.autogen.amd_gpu as amd_gpu, tinygrad.runtime.autogen.am.pm4_nv as pm4
 
@@ -24,19 +25,6 @@ WAIT_REG_MEM_FUNCTION_EQ  = 3 # ==
 WAIT_REG_MEM_FUNCTION_NEQ = 4 # !=
 WAIT_REG_MEM_FUNCTION_GEQ = 5 # >=
 
-REMU_PATHS = ["extra/remu/target/release/libremu.so", "libremu.so", "/usr/local/lib/libremu.so",
-              "extra/remu/target/release/libremu.dylib", "libremu.dylib", "/usr/local/lib/libremu.dylib", "/opt/homebrew/lib/libremu.dylib"]
-def _try_dlopen_remu():
-  for path in REMU_PATHS:
-    try:
-      remu = ctypes.CDLL(path)
-      remu.run_asm.restype = ctypes.c_int32
-      remu.run_asm.argtypes = [ctypes.c_void_p, ctypes.c_uint32, ctypes.c_uint32, ctypes.c_uint32, ctypes.c_uint32,
-        ctypes.c_uint32, ctypes.c_uint32, ctypes.c_uint32, ctypes.c_void_p]
-    except OSError: pass
-    else: return remu
-  print("Could not find libremu.so")
-  return None
 remu = _try_dlopen_remu()
 
 def create_sdma_packets():

test/mockgpu/cuda/cuda.py

@@ -2,16 +2,14 @@ from __future__ import annotations
 from typing import Any
 import ctypes, time
 from tinygrad.runtime.autogen import cuda as orig_cuda
+from test.mockgpu.helpers import _try_dlopen_gpuocelot
 from tinygrad.helpers import mv_address
 
 for attr in dir(orig_cuda):
   if not attr.startswith('__'):
     globals()[attr] = getattr(orig_cuda, attr)
 
-try:
-  gpuocelot_lib = ctypes.CDLL(ctypes.util.find_library("gpuocelot"))
-  gpuocelot_lib.ptx_run.argtypes = [ctypes.c_char_p, ctypes.c_int, ctypes.POINTER(ctypes.c_void_p), ctypes.c_int, ctypes.c_int, ctypes.c_int, ctypes.c_int, ctypes.c_int, ctypes.c_int, ctypes.c_int]  # noqa: E501
-except Exception: pass
+gpuocelot_lib = _try_dlopen_gpuocelot()
 
 # Global state
 class CUDAState:
@@ -130,7 +128,10 @@ def cuModuleUnload(hmod) -> int:
 def cuLaunchKernel(f, gx: int, gy: int, gz: int, lx: int, ly: int, lz: int, sharedMemBytes: int,
                    hStream: Any, kernelParams: Any, extra: Any) -> int:
   cargs = [ctypes.cast(getattr(extra, field[0]), ctypes.c_void_p) for field in extra._fields_]
-  gpuocelot_lib.ptx_run(ctypes.cast(f.value, ctypes.c_char_p), len(cargs), (ctypes.c_void_p*len(cargs))(*cargs), lx, ly, lz, gx, gy, gz, 0)
+  try: gpuocelot_lib.ptx_run(ctypes.cast(f.value, ctypes.c_char_p), len(cargs), (ctypes.c_void_p*len(cargs))(*cargs), lx, ly, lz, gx, gy, gz, 0)
+  except Exception as e:
+    print("Error in cuLaunchKernel:", e)
+    return orig_cuda.CUDA_ERROR_LAUNCH_FAILED
   return orig_cuda.CUDA_SUCCESS
 
 def cuDeviceComputeCapability(major, minor, dev: int) -> int:

test/mockgpu/helpers.py

@@ -0,0 +1,29 @@
+import ctypes, ctypes.util
+
+def _try_dlopen_gpuocelot():
+  GPUOCELOT_PATHS = [ctypes.util.find_library("gpuocelot")] if ctypes.util.find_library("gpuocelot") is not None else []
+  GPUOCELOT_PATHS += ["libgpuocelot.so", "/usr/local/lib/libgpuocelot.so",
+                      "libgpuocelot.dylib", "/usr/local/lib/libgpuocelot.dylib", "/opt/homebrew/lib/libgpuocelot.dylib"]
+  for path in GPUOCELOT_PATHS:
+    try:
+      gpuocelot_lib = ctypes.CDLL(path)
+      gpuocelot_lib.ptx_run.argtypes = [ctypes.c_char_p, ctypes.c_int, ctypes.POINTER(ctypes.c_void_p), ctypes.c_int, ctypes.c_int,
+        ctypes.c_int, ctypes.c_int, ctypes.c_int, ctypes.c_int, ctypes.c_int]
+    except OSError: pass
+    else: return gpuocelot_lib
+  print("Could not find libgpuocelot.so")
+  return None
+
+def _try_dlopen_remu():
+  REMU_PATHS = ["extra/remu/target/release/libremu.so", "libremu.so", "/usr/local/lib/libremu.so",
+               "extra/remu/target/release/libremu.dylib", "libremu.dylib", "/usr/local/lib/libremu.dylib", "/opt/homebrew/lib/libremu.dylib"]
+  for path in REMU_PATHS:
+    try:
+      remu = ctypes.CDLL(path)
+      remu.run_asm.restype = ctypes.c_int32
+      remu.run_asm.argtypes = [ctypes.c_void_p, ctypes.c_uint32, ctypes.c_uint32, ctypes.c_uint32, ctypes.c_uint32,
+        ctypes.c_uint32, ctypes.c_uint32, ctypes.c_uint32, ctypes.c_void_p]
+    except OSError: pass
+    else: return remu
+  print("Could not find libremu.so")
+  return None

test/mockgpu/nv/nvgpu.py

@@ -2,6 +2,7 @@ import ctypes, ctypes.util, time
 import tinygrad.runtime.autogen.nv_gpu as nv_gpu
 from enum import Enum, auto
 from test.mockgpu.gpu import VirtGPU
+from test.mockgpu.helpers import _try_dlopen_gpuocelot
 from tinygrad.helpers import to_mv, init_c_struct_t
 
 def make_qmd_struct_type():
@@ -16,10 +17,7 @@ def make_qmd_struct_type():
 qmd_struct_t = make_qmd_struct_type()
 assert ctypes.sizeof(qmd_struct_t) == 0x40 * 4
 
-try:
-  gpuocelot_lib = ctypes.CDLL(ctypes.util.find_library("gpuocelot"))
-  gpuocelot_lib.ptx_run.argtypes = [ctypes.c_char_p, ctypes.c_int, ctypes.POINTER(ctypes.c_void_p), ctypes.c_int, ctypes.c_int, ctypes.c_int, ctypes.c_int, ctypes.c_int, ctypes.c_int, ctypes.c_int]  # noqa: E501
-except Exception: pass
+gpuocelot_lib = _try_dlopen_gpuocelot()
 
 class SchedResult(Enum): CONT = auto(); YIELD = auto() # noqa: E702
 
@@ -99,7 +97,10 @@ class GPFIFO:
     cargs = [ctypes.cast(args[i], ctypes.c_void_p) for i in range(args_cnt)] + [ctypes.cast(vals[i], ctypes.c_void_p) for i in range(vals_cnt)]
     gx, gy, gz = qmd.cta_raster_width, qmd.cta_raster_height, qmd.cta_raster_depth
     lx, ly, lz = qmd.cta_thread_dimension0, qmd.cta_thread_dimension1, qmd.cta_thread_dimension2
-    gpuocelot_lib.ptx_run(ctypes.cast(prg_addr, ctypes.c_char_p), args_cnt+vals_cnt, (ctypes.c_void_p*len(cargs))(*cargs), lx, ly, lz, gx, gy, gz, 0)
+    try:
+      gpuocelot_lib.ptx_run(ctypes.cast(prg_addr, ctypes.c_char_p), args_cnt+vals_cnt,
+        (ctypes.c_void_p*len(cargs))(*cargs), lx, ly, lz, gx, gy, gz, 0)
+    except Exception as e: print("failed to execute:", e)
     if qmd.release0_enable:
       rel0 = to_mv(qmd.release0_address_lower + (qmd.release0_address_upper << 32), 0x10).cast('Q')
       rel0[0] = qmd.release0_payload_lower + (qmd.release0_payload_upper << 32)

test/speed/external_test_copy_speed.py

@@ -1,9 +1,9 @@
 import unittest, numpy as np
 from tinygrad import Tensor, Device, TinyJit
-from tinygrad.helpers import Timing, CI, OSX
+from tinygrad.helpers import Timing, CI, OSX, getenv
 import multiprocessing.shared_memory as shared_memory
 
-N = 256
+N = getenv("NSZ", 256)
 class TestCopySpeed(unittest.TestCase):
   @classmethod
   def setUpClass(cls): Device[Device.DEFAULT].synchronize()
@@ -54,22 +54,24 @@ class TestCopySpeed(unittest.TestCase):
     @TinyJit
     def _do_copy(t): return t.to('CPU').realize()
 
-    t = Tensor.randn(N, N, 4).contiguous().realize()
+    t = Tensor.randn(N, N).contiguous().realize()
+    Device[Device.DEFAULT].synchronize()
     for _ in range(5):
-      with Timing("sync:  ", on_exit=lambda ns: f" @ {t.nbytes()/ns:.2f} GB/s"):
+      with Timing(f"copy {Device.DEFAULT} -> CPU {t.nbytes()/(1024**2)}M:  ", on_exit=lambda ns: f" @ {t.nbytes()/ns:.2f} GB/s"):
         x = _do_copy(t)
         Device[Device.DEFAULT].synchronize()
       np.testing.assert_equal(t.numpy(), x.numpy())
 
-  def testCopytoCPUtoDefaultJit(self):
+  def testCopyCPUtoDefaultJit(self):
     if Device.DEFAULT == "CPU": return unittest.skip("CPU to CPU copy is a no-op")
 
     @TinyJit
-    def _do_copy(x): return t.to(Device.DEFAULT).realize()
+    def _do_copy(x): return x.to(Device.DEFAULT).realize()
 
     for _ in range(5):
-      t = Tensor.randn(N, N, 4, device="CPU").contiguous().realize()
-      with Timing("sync:  ", on_exit=lambda ns: f" @ {t.nbytes()/ns:.2f} GB/s"):
+      t = Tensor.randn(N, N, device="CPU").contiguous().realize()
+      Device["CPU"].synchronize()
+      with Timing(f"copy CPU -> {Device.DEFAULT} {t.nbytes()/(1024**2)}M:  ", on_exit=lambda ns: f" @ {t.nbytes()/ns:.2f} GB/s"):
         x = _do_copy(t)
         Device[Device.DEFAULT].synchronize()
       np.testing.assert_equal(t.numpy(), x.numpy())

test/test_const_folding.py

@@ -1,11 +1,10 @@
 import unittest, itertools, math
-from typing import Any
 from tinygrad import Tensor, Device, dtypes
-from tinygrad.dtype import DType
+from tinygrad.dtype import DType, ConstType
 from tinygrad.uop.ops import Ops, UOp
 from tinygrad.codegen import full_rewrite_to_sink
-import numpy as np
 from tinygrad.device import is_dtype_supported
+import numpy as np
 from test.helpers import not_support_multi_device
 
 def _check_ast_count(desired_count:int, t:Tensor):
@@ -25,7 +24,7 @@ class TestUnaryOpsConstFolding(unittest.TestCase):
     _check_ast_count(0, Tensor.ones(4).cast(dtypes.int16))
     _check_ast_count(0, Tensor.full(4, fill_value=-1).cast(dtypes.uint16))
 
-  @unittest.expectedFailure  # no two level fold at lazybuffer
+  @unittest.expectedFailure  # no two level fold
   def test_neg_folding(self):
     _check_ast_count(0, Tensor([1, 2, 3]).mul(-1).neg())
     _check_ast_count(0, Tensor([1, 2, 3]).neg().mul(-1))
@@ -104,7 +103,7 @@ class TestBinaryOpsConstFolding(unittest.TestCase):
 
 class TestBitcastConstFolding(unittest.TestCase):
   def test_scalar_bitcast(self):
-    def t(cases: dict[DType, Any]):
+    def t(cases: dict[DType, ConstType]):
       for (from_dt, from_v), (to_dt, to_v) in itertools.product(cases.items(), cases.items()):
         if not math.isnan(from_v):
           r = full_rewrite_to_sink(UOp.const(from_dt, from_v).bitcast(to_dt).sink()).src[0]
@@ -165,7 +164,6 @@ class TestMovedConstFolding(unittest.TestCase):
     _check_ast_count(1, Tensor([1.0, 2, 3, 4]) * Tensor.ones(2).pad(((1, 1),)))
 
   def test_cast_padded(self):
-    # NOTE: this is folded due to CAST_BEFORE_VIEW
     if is_dtype_supported(dtypes.int16):
       _check_ast_count(0, Tensor.ones(4).pad(((1, 1),)).cast(dtypes.int16))
       np.testing.assert_equal(Tensor.ones(4).pad(((1, 1),)).cast(dtypes.int16).numpy(), [0, 1, 1, 1, 1, 0])

test/test_define_reg.py

@@ -1,32 +0,0 @@
-import unittest
-from tinygrad import dtypes, Device, Tensor, Context
-from tinygrad.dtype import AddrSpace
-from tinygrad.helpers import getenv
-from tinygrad.uop.ops import UOp, Ops, KernelInfo, AxisType
-from tinygrad.shape.shapetracker import ShapeTracker
-from tinygrad.engine.realize import get_program, ExecItem, CompiledRunner
-
-class TestDefineReg(unittest.TestCase):
-  def test_simple(self, at=AxisType.UPCAST):
-    N = 16
-    bout = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(N*N), arg=0).view(ShapeTracker.from_shape((N,N)))
-    a = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(N*N), arg=1).view(ShapeTracker.from_shape((N,N)))
-    a_col = UOp(Ops.DEFINE_REG, dtypes.float.ptr(N, AddrSpace.REG), arg=0).view(ShapeTracker.from_shape((N,N), (0,1)))
-
-    out = a_col.load(a_col.store(a.load()))
-    sink = bout.store(out).sink(arg=KernelInfo(name="regcopy", axis_types=(AxisType.LOOP, at)))
-    prg = get_program(sink, Device.default.renderer)
-
-    with Context(DEBUG=0):
-      a = Tensor.randn(N, N).realize()
-      b = Tensor.empty(N, N).realize()
-    hrunner = CompiledRunner(prg)
-    ExecItem(hrunner, [b.uop.buffer, a.uop.buffer]).run(wait=True)
-    with Context(DEBUG=0):
-      self.assertEqual((b-a).mean().item(), 0.0)
-
-  @unittest.skipIf(getenv("PTX"), "ptx needs regs to be unrolled")
-  def test_simple_loop(self): self.test_simple(AxisType.LOOP)
-
-if __name__ == '__main__':
-  unittest.main()

test/test_dtype.py

@@ -4,9 +4,8 @@ import torch
 from typing import Any, List
 from tinygrad.device import is_dtype_supported
 from tinygrad.helpers import getenv, DEBUG, CI
-from tinygrad.dtype import DType, DTYPES_DICT, least_upper_dtype, fp8_to_float, float_to_fp8
+from tinygrad.dtype import DType, DTYPES_DICT, least_upper_dtype, fp8_to_float, float_to_fp8, _to_np_dtype, _to_torch_dtype
 from tinygrad import Device, Tensor, dtypes
-from tinygrad.tensor import _to_np_dtype
 from hypothesis import assume, given, settings, strategies as strat
 from test.helpers import rand_for_dtype
 from test.unit.test_dtype_spec import _assert_eq, core_dtypes, dtype_ints, dtype_floats, FP8E4M3_MAX, FP8E5M2_MAX
@@ -24,6 +23,10 @@ def get_available_cast_dtypes(dtype: DType) -> List[DType]:
   # dont cast internal dtypes
   return [v for k, v in DTYPES_DICT.items() if v != dtype and is_dtype_supported(v) and not k.startswith("_")]
 
+def _to_torch_storage_type(dtype:DType):
+  if dtype == dtypes.bfloat16: return torch.float32
+  return _to_torch_dtype(dtype)
+
 def _test_to_np(a:Tensor, np_dtype, target):
   if DEBUG >= 2: print(a)
   na = a.numpy()
@@ -46,10 +49,10 @@ def _test_cast(a:Tensor, target_dtype:DType):
 
   _test_op(lambda: a.cast(target_dtype), target_dtype, list(a.numpy().astype(_to_np_dtype(target_dtype))))
 def _test_bitcast(a:Tensor, target_dtype:DType, target=None):
-  if target_dtype == dtypes.bfloat16: raise unittest.SkipTest("no test for bf16 bitcast yet")
   if getenv("PTX") and a.dtype == dtypes.int8 and target_dtype.itemsize != a.dtype.itemsize:
     raise unittest.SkipTest("shape changing bitcast of int8 broken on PTX")
-  _test_op(lambda: a.bitcast(target_dtype), target_dtype, target or a.numpy().view(_to_np_dtype(target_dtype)).tolist())
+  expected = torch.tensor(a.tolist(), dtype=_to_torch_storage_type(a.dtype)).view(_to_torch_dtype(target_dtype))
+  _test_op(lambda: a.bitcast(target_dtype), target_dtype, target or expected.tolist())
 
 class TestDType(unittest.TestCase):
   DTYPE: Any = None
@@ -126,7 +129,7 @@ class TestDType(unittest.TestCase):
 
   def test_finfo(self):
     if self.DTYPE not in [dtypes.float16, dtypes.bfloat16, dtypes.float32, dtypes.float64]: return
-    info = np.finfo(_to_np_dtype(self.DTYPE))
+    info = ml_dtypes.finfo(ml_dtypes.bfloat16 if self.DTYPE is dtypes.bfloat16 else _to_np_dtype(self.DTYPE))
     assert info.bits == self.DTYPE.itemsize*8
     assert info.nexp == dtypes.finfo(self.DTYPE)[0]
     assert info.nmant == dtypes.finfo(self.DTYPE)[1]
@@ -299,10 +302,10 @@ class TestBitCast(unittest.TestCase):
   @given(strat.sampled_from(dtype_ints + dtype_floats), strat.sampled_from(dtype_ints + dtype_floats))
   def test_shape_change_bitcast(self, dt1, dt2):
     # NOTE: this has to be assume to prevent hypothesis from skipping all samples
-    assume(dt2 != dtypes.bfloat16 and dt1 != dtypes.bfloat16) # no test for bf16 bitcast yet
     assume(not (getenv("PTX") and dt1 == dtypes.int8)) # TODO: bitcasting int8 fails in PTX
     data = rand_for_dtype(dt1, 32).reshape(2, 2, 8)
-    _test_op(lambda: Tensor(data, dtype=dt1).bitcast(dt2), dt2, data.view(_to_np_dtype(dt2)).tolist())
+    expected = torch.tensor(data.tolist(), dtype=_to_torch_storage_type(dt1)).view(_to_torch_dtype(dt2))
+    _test_op(lambda: Tensor(data, dtype=dt1).bitcast(dt2), dt2, expected.tolist())
 
   def test_shape_change_bitcast_exceptions(self):
     with self.assertRaises(RuntimeError):
@@ -342,6 +345,9 @@ class TestUint64DType(TestDType):
 
 class TestBoolDType(TestDType): DTYPE = dtypes.bool
 
+@unittest.skipUnless(is_dtype_supported(dtypes.bfloat16), f"no bfloat16 on {Device.DEFAULT}")
+class TestBFloat16Type(TestDType): DTYPE = dtypes.bfloat16
+
 class TestPtrDType(unittest.TestCase):
   def test_vec_double(self):
     dt1 = dtypes.float.vec(4).ptr().vec(4)
@@ -418,9 +424,14 @@ class TestDtypeUsage(unittest.TestCase):
 class TestOpsBFloat16(unittest.TestCase):
   def test_cast(self):
     # TODO: helper_test_op breaks in unrelated part
-    # TODO: wrong output with GPU=1 / PYTHON=1 on mac
+    # TODO: wrong output with GPU=1 on mac
     data = [60000.0, 70000.0, 80000.0]
     np.testing.assert_allclose(Tensor(data).cast("bfloat16").numpy(), torch.tensor(data).type(torch.bfloat16).float().numpy())
 
+  def test_no_approximation(self):
+    data = [326.0, 339.0, 10603200512.0]
+    expected = torch.tensor(data, dtype=torch.bfloat16).sqrt().float().numpy()
+    np.testing.assert_allclose(Tensor(data, dtype=dtypes.bfloat16).sqrt().numpy(), expected)
+
 if __name__ == '__main__':
   unittest.main()

test/test_dtype_alu.py

@@ -1,9 +1,10 @@
 import unittest, operator, math
 from tinygrad import Tensor, dtypes, Device
 from tinygrad.dtype import DType
-from tinygrad.helpers import CI, getenv
+from tinygrad.helpers import CI, getenv, AMD_LLVM
 from tinygrad.tensor import _to_np_dtype
 from tinygrad.device import is_dtype_supported
+from tinygrad.runtime.ops_python import from_storage_scalar
 import numpy as np
 import pytest
 from hypothesis import given, strategies as strat, settings, HealthCheck
@@ -20,13 +21,13 @@ dtypes_bool = (dtypes.bool,)
 binary_operations = [operator.add, operator.sub, operator.mul, operator.lt, operator.eq]
 
 # TODO: LLVM comparing with nan is incorrect
-if Device.DEFAULT == "LLVM" or getenv("AMD_LLVM", 0):
+if (Device.DEFAULT == "LLVM") or (Device.DEFAULT == "AMD" and AMD_LLVM):
   binary_operations.remove(operator.lt)
 
 integer_binary_operations = binary_operations + [(Tensor.bitwise_xor, np.bitwise_xor), (Tensor.bitwise_and, np.bitwise_and),
                                                  (Tensor.bitwise_or, np.bitwise_or), operator.mod]
 unary_operations = [(Tensor.exp, np.exp), (Tensor.log, np.log), (Tensor.sin, np.sin),
-                    (Tensor.sqrt, np.sqrt), (Tensor.reciprocal, np.reciprocal)]
+                    (Tensor.sqrt, np.sqrt), (Tensor.reciprocal, np.reciprocal), (Tensor.cos, np.cos)]
 
 # TODO: enable this (this is a dtype issue)
 #binary_operations.append(operator.truediv)
@@ -35,13 +36,14 @@ unary_operations = [(Tensor.exp, np.exp), (Tensor.log, np.log), (Tensor.sin, np.
 #binary_operations += [(Tensor.maximum, np.maximum)]
 
 # TODO: CI CUDA segfaults on sin, WEBGPU sin is not precise enough for large numbers
-if (getenv("MOCKGPU") and Device.DEFAULT in {"NV", "CUDA"}) or Device.DEFAULT == "WEBGPU": unary_operations.remove((Tensor.sin, np.sin))
+if (getenv("MOCKGPU") and Device.DEFAULT in {"NV", "CUDA"}) or Device.DEFAULT == "WEBGPU":
+  unary_operations.remove((Tensor.sin, np.sin))
+  unary_operations.remove((Tensor.cos, np.cos))
 
 class ht:
   float64 = strat.floats(width=64, allow_subnormal=False)
   float32 = strat.floats(width=32, allow_subnormal=False)
   float16 = strat.floats(width=16, allow_subnormal=False)
-  bfloat16 = strat.floats(width=16, allow_subnormal=False)
   uint8 = strat.integers(0, 255)
   uint16 = strat.integers(0, 65535)
   uint32 = strat.integers(0, 2**32-1)
@@ -51,6 +53,7 @@ class ht:
   int32 = strat.integers(-2147483648, 2147483647)
   int64 = strat.integers(-9223372036854775808, 9223372036854775807)
   bool = strat.booleans()
+ht.bfloat16 = ht.uint16
 
 def universal_test(a, b, dtype, op):
   # The 'nan' cases only fail with Vulkan WebGPU backend (CI)
@@ -60,18 +63,22 @@ def universal_test(a, b, dtype, op):
   ta, tb = Tensor([a], dtype=dtype), Tensor([b], dtype=dtype)
   tensor_value = (op[0](ta, tb)).numpy()
   numpy_value = op[1](ta.numpy(), tb.numpy())
-  if dtype == dtypes.bfloat16: np.testing.assert_allclose(tensor_value, numpy_value, atol=1e-3, rtol=1e-2)
-  elif dtype in dtypes_float: np.testing.assert_allclose(tensor_value, numpy_value, atol=1e-10)
+  if dtype in dtypes.floats:
+    atol, rtol = {dtypes.bfloat16:(1e-3, 1e-2)}.get(dtype, (1e-10, 1e-7))
+    np.testing.assert_allclose(tensor_value, numpy_value, atol=atol, rtol=rtol)
   else: np.testing.assert_equal(tensor_value, numpy_value)
 
 def universal_test_unary(a, dtype, op):
   if not isinstance(op, tuple): op = (op, op)
   ta = Tensor([a], dtype=dtype)
+  # TODO: cos does not match for large input
+  if op[0] == Tensor.cos and abs(a) > 100: return
   out: Tensor = op[0](ta)
   tensor_value = out.numpy()
   numpy_value = op[1](ta.numpy())
-  if dtype in (dtypes.float16, dtypes.bfloat16): np.testing.assert_allclose(tensor_value, numpy_value, atol=1e-3, rtol=1e-2)
-  elif dtype in dtypes_float: np.testing.assert_allclose(tensor_value, numpy_value, atol=1e-6, rtol=1e-5)
+  if dtype in dtypes.floats:
+    atol, rtol = {dtypes.float16:(1e-3, 1e-2), dtypes.bfloat16:(1e-3, 2e-2)}.get(dtype, (1e-6, 1e-5))
+    np.testing.assert_allclose(tensor_value, numpy_value, atol=atol, rtol=rtol)
   else: np.testing.assert_equal(tensor_value, numpy_value)
 
 def universal_test_cast(a, in_dtype, dtype):
@@ -90,45 +97,45 @@ def universal_test_midcast(a, b, c, op1, op2, d1:DType, d2:DType):
   np.testing.assert_allclose(tensor_value, numpy_value, rtol=1e-6 if getenv("PTX") else 1e-7)
 
 class TestDTypeALU(unittest.TestCase):
-  @unittest.skipUnless(is_dtype_supported(dtypes.float64, Device.DEFAULT), f"no float64 on {Device.DEFAULT}")
+  @unittest.skipUnless(is_dtype_supported(dtypes.float64), f"no float64 on {Device.DEFAULT}")
   @given(ht.float64, ht.float64, strat.sampled_from(binary_operations))
   def test_float64(self, a, b, op): universal_test(a, b, dtypes.float64, op)
 
   @given(ht.float32, ht.float32, strat.sampled_from(binary_operations))
   def test_float32(self, a, b, op): universal_test(a, b, dtypes.float32, op)
 
-  @unittest.skipUnless(is_dtype_supported(dtypes.float16, Device.DEFAULT), f"no float16 on {Device.DEFAULT}")
+  @unittest.skipUnless(is_dtype_supported(dtypes.float16), f"no float16 on {Device.DEFAULT}")
   @given(ht.float16, ht.float16, strat.sampled_from(binary_operations))
   def test_float16(self, a, b, op): universal_test(a, b, dtypes.float16, op)
 
-  @unittest.skipUnless(is_dtype_supported(dtypes.bfloat16, Device.DEFAULT), f"no bfloat16 on {Device.DEFAULT}")
+  @unittest.skipUnless(is_dtype_supported(dtypes.bfloat16), f"no bfloat16 on {Device.DEFAULT}")
   @given(ht.bfloat16, ht.bfloat16, strat.sampled_from(binary_operations))
-  def test_bfloat16(self, a, b, op): universal_test(a, b, dtypes.bfloat16, op)
+  def test_bfloat16(self, a, b, op):
+    universal_test(from_storage_scalar(a, dtypes.bfloat16), from_storage_scalar(a, dtypes.bfloat16), dtypes.bfloat16, op)
 
   @given(ht.float32, strat.sampled_from(unary_operations))
   def test_float32_unary(self, a, op): universal_test_unary(a, dtypes.float32, op)
 
-  @unittest.skipUnless(is_dtype_supported(dtypes.float16, Device.DEFAULT), f"no float16 on {Device.DEFAULT}")
+  @unittest.skipUnless(is_dtype_supported(dtypes.float16), f"no float16 on {Device.DEFAULT}")
   @given(ht.float16, strat.sampled_from(unary_operations))
   def test_float16_unary(self, a, op): universal_test_unary(a, dtypes.float16, op)
 
-  @unittest.skipUnless(is_dtype_supported(dtypes.bfloat16, Device.DEFAULT), f"no bfloat16 on {Device.DEFAULT}")
+  @unittest.skipUnless(is_dtype_supported(dtypes.bfloat16), f"no bfloat16 on {Device.DEFAULT}")
   @given(ht.bfloat16, strat.sampled_from(unary_operations))
-  @unittest.skipIf(Device.DEFAULT in ["AMD"], "broken on AMD?")
-  def test_bfloat16_unary(self, a, op): universal_test_unary(a, dtypes.bfloat16, op)
+  def test_bfloat16_unary(self, a, op): universal_test_unary(from_storage_scalar(a, dtypes.bfloat16), dtypes.bfloat16, op)
 
   @given(ht.uint8, ht.uint8, strat.sampled_from(integer_binary_operations))
   def test_uint8(self, a, b, op): universal_test(a, b, dtypes.uint8, op)
 
-  @unittest.skipUnless(is_dtype_supported(dtypes.uint16, Device.DEFAULT), f"no uint16 on {Device.DEFAULT}")
+  @unittest.skipUnless(is_dtype_supported(dtypes.uint16), f"no uint16 on {Device.DEFAULT}")
   @given(ht.uint16, ht.uint16, strat.sampled_from(integer_binary_operations))
   def test_uint16(self, a, b, op): universal_test(a, b, dtypes.uint16, op)
 
-  @unittest.skipUnless(is_dtype_supported(dtypes.uint32, Device.DEFAULT), f"no uint32 on {Device.DEFAULT}")
+  @unittest.skipUnless(is_dtype_supported(dtypes.uint32), f"no uint32 on {Device.DEFAULT}")
   @given(ht.uint32, ht.uint32, strat.sampled_from(integer_binary_operations))
   def test_uint32(self, a, b, op): universal_test(a, b, dtypes.uint32, op)
 
-  @unittest.skipUnless(is_dtype_supported(dtypes.uint64, Device.DEFAULT), f"no uint64 on {Device.DEFAULT}")
+  @unittest.skipUnless(is_dtype_supported(dtypes.uint64), f"no uint64 on {Device.DEFAULT}")
   @given(ht.uint64, ht.uint64, strat.sampled_from(integer_binary_operations))
   def test_uint64(self, a, b, op): universal_test(a, b, dtypes.uint64, op)
 
@@ -141,7 +148,7 @@ class TestDTypeALU(unittest.TestCase):
   @given(ht.int32, ht.int32, strat.sampled_from(integer_binary_operations))
   def test_int32(self, a, b, op): universal_test(a, b, dtypes.int32, op)
 
-  @unittest.skipUnless(is_dtype_supported(dtypes.int64, Device.DEFAULT), f"no int64 on {Device.DEFAULT}")
+  @unittest.skipUnless(is_dtype_supported(dtypes.int64), f"no int64 on {Device.DEFAULT}")
   @given(ht.int64, ht.int64, strat.sampled_from(integer_binary_operations))
   def test_int64(self, a, b, op): universal_test(a, b, dtypes.int64, op)
 
@@ -171,7 +178,7 @@ class TestDTypeALU(unittest.TestCase):
   @settings(suppress_health_check=[HealthCheck.filter_too_much])
   @given(strat.data(), strat.sampled_from(dtypes_float), strat.sampled_from((dtypes.uint8, dtypes.uint16)))
   def test_float_cast_to_unsigned(self, a, float_dtype, unsigned_dtype):
-    if not is_dtype_supported(float_dtype, Device.DEFAULT): float_dtype = dtypes.float32
+    if not is_dtype_supported(float_dtype): float_dtype = dtypes.float32
     float_strat = {dtypes.float16: ht.float16, dtypes.float32: ht.float32, dtypes.float64: ht.float64}[float_dtype]
     float_strat = float_strat.filter(lambda x: 0 < x < dtypes.max(unsigned_dtype))
     universal_test_cast(a.draw(float_strat), float_dtype, unsigned_dtype)
@@ -179,7 +186,7 @@ class TestDTypeALU(unittest.TestCase):
   @settings(suppress_health_check=[HealthCheck.filter_too_much])
   @given(strat.data(), strat.sampled_from(dtypes_float), strat.sampled_from((dtypes.uint8, dtypes.uint16)))
   def test_float_cast_to_unsigned_overflow(self, a, float_dtype, unsigned_dtype):
-    if not is_dtype_supported(float_dtype, Device.DEFAULT): float_dtype = dtypes.float32
+    if not is_dtype_supported(float_dtype): float_dtype = dtypes.float32
     float_strat = {dtypes.float16: ht.float16, dtypes.float32: ht.float32, dtypes.float64: ht.float64}[float_dtype]
     overflow_strat = float_strat.filter(lambda x: x > dtypes.max(unsigned_dtype) and x <= dtypes.max(dtypes.int32))
     universal_test_cast(a.draw(overflow_strat), float_dtype, unsigned_dtype)
@@ -187,7 +194,7 @@ class TestDTypeALU(unittest.TestCase):
   @settings(suppress_health_check=[HealthCheck.filter_too_much])
   @given(strat.data(), strat.sampled_from(dtypes_float), strat.sampled_from((dtypes.uint8, dtypes.uint16)))
   def test_float_cast_to_unsigned_underflow(self, a, float_dtype, unsigned_dtype):
-    if not is_dtype_supported(float_dtype, Device.DEFAULT): float_dtype = dtypes.float32
+    if not is_dtype_supported(float_dtype): float_dtype = dtypes.float32
     float_strat = {dtypes.float16: ht.float16, dtypes.float32: ht.float32, dtypes.float64: ht.float64}[float_dtype]
     underflow_strat = float_strat.filter(lambda x: x < 0 and x >= dtypes.min(dtypes.int32))
     universal_test_cast(a.draw(underflow_strat), float_dtype, unsigned_dtype)

test/test_linearizer.py

@@ -12,7 +12,7 @@ from tinygrad.tensor import Tensor, _to_np_dtype
 from tinygrad.engine.realize import run_schedule, lower_schedule, CompiledRunner, get_program
 from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
 from tinygrad.helpers import prod, Context, getenv, CI, flatten, dedup, AMX, AMD_LLVM
-from tinygrad.dtype import DType, dtypes, AddrSpace
+from tinygrad.dtype import DType, dtypes, PtrDType, AddrSpace
 from tinygrad.codegen import apply_rewrites, rewrites_for_views
 
 def push_views(ast): return apply_rewrites(ast, rewrites_for_views)
@@ -33,11 +33,10 @@ def helper_tc_allclose(N:int, M:int, K:int, dtype_in:DType, dtype_out:DType, axi
   r = a.matmul(b, dtype=dtype_out)
   if dtype_in == dtypes.bfloat16: r = r.float()
   realized_ast, bufs = helper_realized_ast(r)
-  k = Kernel(realized_ast)
-  k.apply_tensor_cores(use_tensor_cores, axis=axis, tc_select=tc_select, tc_opt=tc_opt)
-  prg = CompiledRunner(replace(get_program(k.get_optimized_ast(), k.opts), device=Device.DEFAULT))
+  opts = [Opt(op=OptOps.TC, axis=axis, arg=(tc_select, tc_opt, use_tensor_cores))]
+  prg = CompiledRunner(replace(get_program(realized_ast, opts=opts), device=Device.DEFAULT))
   if use_tensor_cores == 1: assert len([uop for uop in prg.p.uops if uop.op is Ops.WMMA]) > 0, "wmma not triggered"
-  assert len([x for x in k.applied_opts if x.op is OptOps.TC]) == 1, "tensor core opt not included"
+  assert len([x for x in prg.p.uops[-1].arg.applied_opts if x.op is OptOps.TC]) == 1, "tensor core opt not included"
   prg.exec(bufs)
   if dtype_in == dtypes.half: tc_atol, tc_rtol = 1e-2, 1e-3
   elif dtype_in == dtypes.bfloat16: tc_atol, tc_rtol = 1e-2, 1e-2
@@ -117,6 +116,7 @@ class TestLinearizer(unittest.TestCase):
         if skip and i in skip: continue
         assert ranges[i-1] != u, f"multireduce nested the ranges! {ranges[i-1], {u}}"
 
+  @unittest.skip("broken. should not depends on push_views and implementation details of getitem")
   @unittest.skipIf(CI and Device.DEFAULT in {"PTX", "AMD", "NV"}, "very slow")
   def test_indexing_multireduce(self):
     dataset = Tensor.rand(16384, 256).realize()
@@ -133,7 +133,7 @@ class TestLinearizer(unittest.TestCase):
     uops = get_program(lin.get_optimized_ast(), lin.opts).uops
     ranges = [i for i,u in enumerate(uops) if u.op is Ops.RANGE]
     assert len(ranges) == 1 # NOTE: it collapses now
-    # RANGE -> LOAD -> RANGE -> ASSIGN
+    # RANGE -> LOAD -> RANGE -> STORE
     #assert any(x.op is Ops.LOAD for x in uops[ranges[0]:ranges[1]])
 
   def test_three_nested_range(self):
@@ -143,7 +143,7 @@ class TestLinearizer(unittest.TestCase):
     uops = get_program(lin.get_optimized_ast(), lin.opts).uops
     ranges = [i for i,u in enumerate(uops) if u.op is Ops.RANGE]
     assert len(ranges) == 1 # NOTE: it collapses now
-    # RANGE -> RANGE -> LOAD -> RANGE -> ASSIGN
+    # RANGE -> RANGE -> LOAD -> RANGE -> STORE
     # NOTE: nothing should toposort between the first two ranges
     #assert ranges[0]+1 == ranges[1]
     #assert any(x.op is Ops.LOAD for x in uops[ranges[1]:ranges[2]])
@@ -154,7 +154,7 @@ class TestLinearizer(unittest.TestCase):
     lin = helper_linearizer_opt(out, wanna_output=[24])[0]
     uops = get_program(lin.get_optimized_ast(), lin.opts).uops
     ranges = [i for i,u in enumerate(uops) if u.op is Ops.RANGE]
-    # RANGE -> ALU -> RANGE -> ALU + LOAD -> ASSIGN
+    # RANGE -> ALU -> RANGE -> ALU + LOAD -> STORE
     assert any(x.op in GroupOp.ALU for x in uops[ranges[0]:ranges[1]])
     assert not any(x.op is Ops.LOAD for x in uops[ranges[0]:ranges[1]])
     assert any(x.op in {*GroupOp.ALU, Ops.LOAD} for x in uops[ranges[1]:])
@@ -166,7 +166,7 @@ class TestLinearizer(unittest.TestCase):
     lin = helper_linearizer_opt(out, wanna_output=[(a.numpy()+b.numpy()[0]).sum()+b.numpy()])[0]
     uops = get_program(lin.get_optimized_ast(), lin.opts).uops
     ranges = [i for i,u in enumerate(uops) if u.op is Ops.RANGE]
-    # LOAD -> RANGE -> LOAD -> ASSIGN
+    # LOAD -> RANGE -> LOAD -> STORE
     assert len([x for x in uops[:ranges[0]] if x.op is Ops.LOAD]) == 1
 
   def test_range_outer_op_before_phi_nested_range(self):
@@ -178,11 +178,11 @@ class TestLinearizer(unittest.TestCase):
     ranges = [i for i,u in enumerate(uops) if u.op is Ops.RANGE]
     assert len(ranges) == 1 # NOTE: it collapses now
     #if getenv("PTX"):
-    # LOAD -> RANGE -> CAST -> ALU -> ALU -> LOAD -> ALU -> RANGE -> ALU -> ASSIGN
+    # LOAD -> RANGE -> CAST -> ALU -> ALU -> LOAD -> ALU -> RANGE -> ALU -> STORE
     #  assert uops[ranges[0]-2].op is Ops.LOAD
     #  assert ranges[1] == ranges[0]+6
     #  assert [x.op for x in uops[ranges[1]-2:ranges[1]]] == [Ops.LOAD, Ops.ALU]
-    # LOAD -> RANGE -> LOAD -> ALU -> RANGE -> ASSIGN
+    # LOAD -> RANGE -> LOAD -> ALU -> RANGE -> STORE
     #else:
     #  assert uops[ranges[0]-2].op is Ops.LOAD
     #  assert ranges[1] == ranges[0]+3
@@ -194,7 +194,7 @@ class TestLinearizer(unittest.TestCase):
     out = a.sum() * a.sum()
     lin = helper_linearizer_opt(out, wanna_output=[a.numpy().sum()*a.numpy().sum()])[0]
     uops = get_program(lin.get_optimized_ast(), lin.opts).uops
-    # RANGE -> LOAD -> ASSIGN -> ALU
+    # RANGE -> LOAD -> STORE -> ALU
     end = max(i for i,u in enumerate(uops) if u.op is Ops.ENDRANGE)
     # the INDEX can be first
     assert uops[end+1].op in GroupOp.ALU or uops[end+2].op in GroupOp.ALU
@@ -205,7 +205,7 @@ class TestLinearizer(unittest.TestCase):
     out = a.reshape(2, 1).expand(2, 3).sum() + a.reshape(2, 1).expand(2, 3).sum()
     lin = helper_linearizer_opt(out, wanna_output=[(np.broadcast_to(a.numpy().reshape(2, 1), (2, 3))).sum()*2])[0]
     uops = get_program(lin.get_optimized_ast(), lin.opts).uops
-    # RANGE -> LOAD -> ASSIGN -> ALU
+    # RANGE -> LOAD -> STORE -> ALU
     end = max(i for i,u in enumerate(uops) if u.op is Ops.ENDRANGE)
     # the INDEX can be first
     assert uops[end+1].op in GroupOp.ALU or uops[end+2].op in GroupOp.ALU
@@ -327,11 +327,7 @@ class TestLinearizer(unittest.TestCase):
       n, m, k = tc.dims[0], tc.dims[1], 2 if AMX else tc.dims[2]
       a, b = Tensor.rand(m, k, dtype=tc.dtype_in), Tensor.rand(k, n, dtype=tc.dtype_in)
       r = a.matmul(b, dtype=tc.dtype_out)
-      sched = r.schedule()
-      realized_ast = push_views(sched[-1].ast)
-      kernel = Kernel(realized_ast)
-      kernel.apply_tensor_cores(1, axis=0, tc_select=-1, tc_opt=2)
-      prg = get_program(kernel.get_optimized_ast(), kernel.opts)
+      prg = get_program(r.schedule()[-1].ast, opts=[Opt(op=OptOps.TC, axis=0, arg=(-1, 2, 1))])
       if Device.DEFAULT == "LLVM":
         assert "0x201000" in prg.src
       elif Device.DEFAULT == "AMD" and AMD_LLVM:
@@ -352,7 +348,7 @@ class TestLinearizer(unittest.TestCase):
   # Internal bug: zero-stride dimensions combined with a mask may produce wrong index/valid for pad == 1 on AMD
   @unittest.skipUnless((Device.DEFAULT == "AMD") or (Device.DEFAULT == "PYTHON" and getenv("EMULATE_AMD")), "test for AMD's tc")
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
-  @unittest.expectedFailure
+  @unittest.skip("warp elements not duplicated properly across lanes")
   def test_tensor_cores_padded_amd(self):
     for tc in Device[Device.DEFAULT].renderer.tensor_cores:
       if not is_dtype_supported(tc.dtype_in) or not is_dtype_supported(tc.dtype_out): continue
@@ -422,9 +418,9 @@ class TestLinearizer(unittest.TestCase):
     x, y = Tensor.rand(128, 128, dtype=tc.dtype_in), Tensor.rand(128, 128, dtype=tc.dtype_in)
     r = x.matmul(y, dtype=tc.dtype_out)
     k = helper_linearizer_opt(r, [[Opt(OptOps.UNROLL, 0, 4)]], apply_tc=True, atol=3e-2, rtol=1e-3)[-1]
-    for u in get_program(k.get_optimized_ast(), k.opts).uops:
+    for u in get_program(k.ast, k.opts, k.applied_opts).uops:
       if u.op is Ops.WMMA:
-        assert u.src[-1].src[0].op != Ops.ASSIGN
+        assert u.src[-1].src[0].op != Ops.STORE
 
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
   @unittest.skipIf(Device.DEFAULT in {"CPU", "LLVM"}, "CPU does not support using a different type for accumulation")
@@ -433,37 +429,43 @@ class TestLinearizer(unittest.TestCase):
     x, y = Tensor.rand(128, 128, dtype=tc.dtype_in), Tensor.rand(128, 128, dtype=tc.dtype_in)
     r = x.matmul(y, dtype=tc.dtype_out)
     k = helper_linearizer_opt(r, [[Opt(OptOps.UNROLL, 0, 4)]], apply_tc=True, atol=3e-2, rtol=1e-3)[-1]
-    for u in get_program(k.get_optimized_ast(), k.opts).uops:
+    for u in get_program(k.ast, k.opts, k.applied_opts).uops:
       if u.op is Ops.WMMA:
         #assert u.src[-1].dtype == dtypes.float.vec(prod(tc.thread_local_sizes[2]))
-        assert u.src[-1].src[0].op != Ops.ASSIGN
+        assert u.src[-1].src[0].op != Ops.STORE
 
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
   @unittest.skipIf(Device.DEFAULT in {"CPU", "LLVM"}, "CPU does not support using a different type for accumulation")
   def test_tensor_cores_unroll_casted_phi_with_children(self):
-    # all ASSIGN children are outside the loop
+    # all STORE children are outside the loop
     tc = [tc for tc in Device[Device.DEFAULT].renderer.tensor_cores if tc.dtype_in != tc.dtype_out][0]
     x, y = Tensor.rand(128, 128, dtype=tc.dtype_in), Tensor.rand(128, 128, dtype=tc.dtype_in)
     r = x.matmul(y, dtype=tc.dtype_out).relu()
     k = helper_linearizer_opt(r, [[Opt(OptOps.UNROLL, 0, 4)]], apply_tc=True, atol=3e-2, rtol=1e-3)[-1]
-    for u in get_program(k.get_optimized_ast(), k.opts).uops:
+    for u in get_program(k.ast, k.opts, k.applied_opts).uops:
       if u.op is Ops.WMMA:
         #assert u.src[-1].dtype == dtypes.float.vec(prod(tc.thread_local_sizes[2]))
-        assert u.src[-1].src[0].op != Ops.ASSIGN
+        assert u.src[-1].src[0].op != Ops.STORE
 
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.supports_float4, "test requires float4")
   def test_simple_unroll_no_between_phi_dependencies(self):
     x, y = Tensor.rand(128, 128), Tensor.rand(128, 128)
     r = (x@y).relu()
     k = helper_linearizer_opt(r, [[Opt(OptOps.UNROLL, 0, 4), Opt(OptOps.UPCAST, 0, 4)]])[-1]
-    # the uops graph is RANGE -> DEFINE_ACC -> 4x ALU -> 4x ASSIGN -> ENDRANGE
-    uops = get_program(k.get_optimized_ast(), k.opts).uops
+    # the uops graph is DEFINE_REG -> 4x STORE 0.0 -> RANGE -> 4x ALU -> 4x STORE -> ENDRANGE
+    uops = get_program(k.ast, k.opts, k.applied_opts).uops
+    begin_range = [i for i, x in enumerate(uops) if x.op is Ops.RANGE][-1]
+    end_range = [i for i, x in enumerate(uops) if x.op is Ops.ENDRANGE][0]
+    for i,u in enumerate(uops): print(i, u.op, [uops.index(s) for s in u.src], u.arg, u.dtype)
     for u in uops:
-      if u.op is Ops.ASSIGN:
-        assert u.src[1].op in GroupOp.ALU
-      # children of ASSIGN are placed after ENDRANGE
-      if any(x.op is Ops.ASSIGN for x in u.src):
-        end_range = [i for i, x in enumerate(uops) if x.op is Ops.ENDRANGE][0]
+      if u.op is Ops.STORE and isinstance(dt:=u.src[0].dtype, PtrDType) and dt.addrspace is AddrSpace.REG:
+        if uops.index(u) < begin_range:
+          assert u.src[1].op is Ops.CONST
+        else:
+          assert u.src[1].op in GroupOp.ALU
+          assert begin_range < uops.index(u) < end_range
+      # children of STORE are placed after ENDRANGE
+      if any(x.op is Ops.STORE and x.src[1].op in GroupOp.ALU for x in u.src):
         assert end_range < uops.index(u)
 
   def test_grouped_dims(self):
@@ -542,7 +544,7 @@ class TestLinearizer(unittest.TestCase):
     # shrink so that the dims do not collapse
     t = Tensor.ones(5, 6, 7).contiguous().realize().shrink(((0, 4), (0, 5), (0, 6)))
     k = helper_linearizer_opt(t+1)[0]
-    uops = get_program(k.get_optimized_ast(), k.opts).uops
+    uops = get_program(k.ast, k.opts, k.applied_opts).uops
     idxs = dedup([uop for uop in uops if uop.op is Ops.SPECIAL])
     idxs = sorted(idxs, key=lambda uop: uop.arg[0])
     assert idxs[0].arg == ('gidx0', 6), idxs[0].arg
@@ -582,12 +584,13 @@ class TestLinearizer(unittest.TestCase):
   def test_phi_simplification(self):
     def helper(t, max_ops=0):
       k = helper_linearizer_opt(t)[-1]
-      uops = get_program(k.get_optimized_ast(), k.opts).uops
+      uops = get_program(k.ast, k.opts, k.applied_opts).uops
       # ignore kernel optimized IF statements for now
       if if_op:=next((u for u in uops if u.op is Ops.IF), None):
         uops = uops[:uops.index(if_op)]
       assert len(set([u.op for u in uops if u.op in {Ops.RANGE, Ops.SPECIAL}])) == 1, "has either specials or ranges, not both"
-      assert len([u for u in uops if u.op is Ops.ASSIGN]) == 0, "ASSIGN should have been simplified"
+      reg_stores = [u for u in uops if u.op is Ops.STORE and isinstance(dt:=u.src[0].dtype, PtrDType) and dt.addrspace == AddrSpace.REG]
+      assert len(reg_stores) == 0, "STORE to reg should have been simplified"
       # TODO: once uops track min/max this will be fixed
       #assert len([u for u in uops if u.op is Ops.MAX]) <= max_ops, "no unnecessary MAX ops"
 
@@ -613,7 +616,7 @@ class TestLinearizer(unittest.TestCase):
     x, y = Tensor.randn(64,64), Tensor.randn(64,64)
     out = x.matmul(y)
     k = helper_linearizer_opt(out)[-1]
-    uops = get_program(k.get_optimized_ast(), k.opts).uops
+    uops = get_program(k.ast, k.opts, k.applied_opts).uops
     # check that the float4 cast collapses
     store_vals = [u.src[1] for u in uops if u.op is Ops.STORE and u.src[0].dtype.addrspace != AddrSpace.REG]
     for val in store_vals:
@@ -638,7 +641,7 @@ class TestLinearizer(unittest.TestCase):
     x = Tensor.randn((4,3,6,6)).realize()
     out = x.flip((0,1)).contiguous()
     k = helper_linearizer_opt(out)[-1]
-    store_val = [u.src[1] for u in get_program(k.get_optimized_ast(), k.opts).uops if u.op is Ops.STORE][0]
+    store_val = [u.src[1] for u in get_program(k.ast, k.opts, k.applied_opts).uops if u.op is Ops.STORE][0]
     assert store_val.dtype == dtypes.float.vec(4) and store_val.op is not Ops.VECTORIZE
 
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
@@ -651,7 +654,7 @@ class TestLinearizer(unittest.TestCase):
             Opt(OptOps.UNROLL, 0, 4), Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.UPCAST, 1, 2)] # upcast accs in both reduces
     k = helper_linearizer_opt(out, opts=[opt])[-1]
     def get_recursive(uop): return set.union(set(uop.src), [uop], *[get_recursive(v) for v in uop.src])
-    uops = get_program(k.get_optimized_ast(), k.opts).uops
+    uops = get_program(k.ast, k.opts, k.applied_opts).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]
@@ -671,7 +674,7 @@ class TestLinearizer(unittest.TestCase):
     x, y = Tensor.rand(1,128), Tensor.rand(128, 128)
     r = (x@y).relu()
     k = helper_linearizer_opt(r)[-1]
-    uops = get_program(k.get_optimized_ast(), k.opts).uops
+    uops = get_program(k.ast, k.opts, k.applied_opts).uops
     stores = [u for u in uops if u.op is Ops.STORE and u.src[0].dtype.addrspace != AddrSpace.REG]
 
     # the float4 value stores directly in lds and we skip upcast
@@ -697,7 +700,7 @@ class TestLinearizer(unittest.TestCase):
         Opt(op=OptOps.LOCAL, axis=1, arg=2), Opt(op=OptOps.UPCAST, axis=3, arg=2)
     ]
     k = helper_linearizer_ast(ast, [Tensor.randn(240*40).realize()], opts=[opt])[-1]
-    out = [u for u in get_program(k.get_optimized_ast(), k.opts).uops if u.op is Ops.STORE][0]
+    out = [u for u in get_program(k.ast, k.opts, k.applied_opts).uops if u.op is Ops.STORE][0]
     assert out.src[1].op is Ops.VECTORIZE and out.src[1].dtype == dtypes.float.vec(4)
 
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
@@ -715,7 +718,7 @@ class TestLinearizer(unittest.TestCase):
             Opt(op=OptOps.UPCAST, axis=1, arg=0), Opt(op=OptOps.UPCAST, axis=1, arg=4), Opt(op=OptOps.LOCAL, axis=0, arg=8),
             Opt(op=OptOps.UPCAST, axis=1, arg=0), Opt(op=OptOps.UPCAST, axis=0, arg=2)]
     k = helper_linearizer_ast(ast, [Tensor.randn(8*32).realize()], opts=[opt])[-1]
-    out = [u for u in get_program(k.get_optimized_ast(), k.opts).uops if u.op is Ops.STORE][0]
+    out = [u for u in get_program(k.ast, k.opts, k.applied_opts).uops if u.op is Ops.STORE][0]
     assert out.src[1].op is Ops.VECTORIZE and out.src[1].dtype.count != 1
 
 @unittest.skipUnless(Device[Device.DEFAULT].renderer.supports_float4, "need backends that support float4")
@@ -1046,7 +1049,7 @@ def _helper_linearizer_opt_ast(realized_ast:UOp, real_bufs:list[Buffer], opts=[]
   outbufs = [real_bufs[x.src[0].base.arg] for x in realized_ast.src]
   device = real_bufs[0].device
 
-  def get_prg(k:Kernel): return CompiledRunner(replace(get_program(k.get_optimized_ast(), k.opts), device=device))
+  def get_prg(k:Kernel): return CompiledRunner(replace(get_program(k.ast, k.opts, k.applied_opts), device=device))
 
   def check_opt(opts, create_k, expected_color_size):
     k = create_k()

test/test_linearizer_dumb.py

@@ -9,7 +9,6 @@ from tinygrad.uop.ops import UOp, Ops
 from tinygrad.helpers import getenv
 from tinygrad.shape.shapetracker import ShapeTracker, View
 from tinygrad.codegen.opt.search import Opt, OptOps
-from tinygrad.codegen.opt.kernel import Kernel
 from tinygrad.engine.realize import get_program
 
 class TestLinearizerDumb(unittest.TestCase):
@@ -36,9 +35,7 @@ class TestLinearizerDumb(unittest.TestCase):
           UOp(Ops.CONST, dtypes.half, arg=0.0, src=(
              x16,)),)),)),))
     opts = [Opt(op=OptOps.TC, axis=2, arg=(-1, 2, 1)), Opt(op=OptOps.UPCAST, axis=2, arg=0), Opt(op=OptOps.UNROLL, axis=1, arg=0)]
-    k = Kernel(ast, opts=Device["METAL"].renderer)
-    k.apply_opts(opts)
-    prg = get_program(k.get_optimized_ast(), k.opts)
+    prg = get_program(ast, Device["METAL"].renderer, opts)
     print(prg.src)
     Device[Device.DEFAULT].compiler.compile_cached(prg.src)
     gate_count = len([x for x in prg.src.splitlines() if "if" in x])
@@ -75,9 +72,7 @@ class TestLinearizerDumb(unittest.TestCase):
             UOp(Ops.CONST, dtypes.int, arg=1000, src=(
                x14,)),)),)),)),))
     opts = [Opt(op=OptOps.UNROLL, axis=0, arg=4), Opt(op=OptOps.LOCAL, axis=0, arg=8)]
-    k = Kernel(ast, opts=Device[Device.DEFAULT].renderer)
-    k.apply_opts(opts)
-    prg = get_program(k.get_optimized_ast(), k.opts)
+    prg = get_program(ast, Device[Device.DEFAULT].renderer, opts)
     print(prg.src)
     assert prg.uops is not None and not any(uop.op is Ops.MAX for uop in prg.uops), "leftover MAX"
 
@@ -93,9 +88,7 @@ class TestLinearizerDumb(unittest.TestCase):
             UOp(Ops.VIEW, dtypes.float.ptr(25), arg=ShapeTracker(views=(View(shape=(26, 49), strides=(0, -1), offset=48, mask=((0, 26), (24, 49)), contiguous=False), View(shape=(25, 25), strides=(1, 50), offset=0, mask=None, contiguous=False))), src=(
               UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(25), arg=1, src=()),)),)),)),)),))
     opts = [Opt(op=OptOps.GROUP, axis=0, arg=0), Opt(op=OptOps.PADTO, axis=0, arg=32), Opt(op=OptOps.LOCAL, axis=0, arg=4), Opt(op=OptOps.UPCAST, axis=0, arg=0)]
-    k = Kernel(ast, opts=Device[Device.DEFAULT].renderer)
-    k.apply_opts(opts)
-    prg = get_program(k.get_optimized_ast(), k.opts)
+    prg = get_program(ast, Device[Device.DEFAULT].renderer, opts)
     print(prg.src)
     if_uops = [u for u in prg.uops if u.op is Ops.IF]
     self.assertIn(len(if_uops), {1,2,3})
@@ -135,8 +128,7 @@ class TestLinearizerDumb(unittest.TestCase):
                 UOp(Ops.LOAD, dtypes.half, arg=None, src=(
                   UOp(Ops.VIEW, dtypes.half.ptr(131072000), arg=ShapeTracker(views=(View(shape=(4096, 32000, 1), strides=(1, 4096, 0), offset=0, mask=None, contiguous=False),)), src=(
                     UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(131072000), arg=2, src=()),)),)),)),)),)),)),)),))
-    k = Kernel(ast, opts=Device[Device.DEFAULT].renderer)
-    prg = get_program(k.get_optimized_ast(), k.opts)
+    prg = get_program(ast, Device[Device.DEFAULT].renderer)
     print(prg.src)
 
   @unittest.expectedFailure
@@ -163,11 +155,9 @@ class TestLinearizerDumb(unittest.TestCase):
               UOp(Ops.VIEW, dtypes.float.ptr(18), arg=ShapeTracker(views=(View(shape=(3, 6), strides=(6, 1), offset=0, mask=None, contiguous=True),)), src=(
                 UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(18), arg=2, src=()),)),)),)),)),)),))
     opts = [Opt(op=OptOps.UNROLL, axis=0, arg=0)]
-    k = Kernel(ast, opts=Device[Device.DEFAULT].renderer)
-    k.apply_opts(opts)
-    prg = get_program(k.get_optimized_ast(), k.opts)
+    prg = get_program(ast, Device[Device.DEFAULT].renderer, opts)
     print(prg.src)
-    load_idxs = [x.src[1] for x in k.uops if x.op is Ops.LOAD and x.src[0].arg == 2]
+    load_idxs = [x.src[1] for x in prg.uops if x.op is Ops.LOAD and x.src[0].arg == 2]
     assert load_idxs[0] < load_idxs[1], f"first loaded idx {load_idxs[0].arg} then {load_idxs[1].arg}!"
 
   @unittest.expectedFailure
@@ -187,11 +177,9 @@ class TestLinearizerDumb(unittest.TestCase):
               UOp(Ops.VIEW, dtypes.float.ptr(1040), arg=ShapeTracker(views=(View(shape=(4, 5, 13, 1, 1, 1, 4, 1, 4, 3, 3), strides=(260, 13, 1, 0, 0, 0, 65, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
                 UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(1040), arg=2, src=()),)),)),)),)),)),))
     opts = [Opt(op=OptOps.UPCAST, axis=3, arg=0), Opt(op=OptOps.UPCAST, axis=2, arg=0)]
-    k = Kernel(ast, opts=Device[Device.DEFAULT].renderer)
-    k.apply_opts(opts)
-    prg = get_program(k.get_optimized_ast(), k.opts)
+    prg = get_program(ast, Device[Device.DEFAULT].renderer, opts)
     print(prg.src)
-    store_idxs = [x.src[1] for x in k.uops if x.op is Ops.STORE]
+    store_idxs = [x.src[1] for x in prg.uops if x.op is Ops.STORE]
     for i in range(len(store_idxs) - 1):
       first_bounds = store_idxs[i].vmin+store_idxs[i].vmax
       next_bounds = store_idxs[i+1].vmin+store_idxs[i+1].vmax

test/test_memory_planner.py

@@ -120,5 +120,19 @@ class TestMemoryPlanner(unittest.TestCase):
     ]
     check_assign(bs)
 
+  def test_very_small_buffers(self):
+    bs = [
+      [b(0, pin=True), b(1, size=32)],
+      [b(3, size=4), b(4, size=6)],
+    ]
+    check_assign(bs)
+
+  def test_very_big_buffers(self):
+    bs = [
+      [b(0, pin=True), b(1, size=34359738368000)],
+      [b(3, size=1 << 128), b(4, size=1 << 64)],
+    ]
+    check_assign(bs)
+
 if __name__ == "__main__":
   unittest.main()

test/test_multitensor.py

@@ -1128,6 +1128,7 @@ class TestMultiRamUsage(unittest.TestCase):
     self.assertUsed(self.N*self.N*4) # sharding should not increase total ram usage
   def test_zeros_shard_self(self): self.test_zeros_shard((d0, d1))
 
+  @unittest.skip("flaky")
   def test_zeros_contiguous_shard(self):
     _ = Tensor.zeros(self.N, self.N).contiguous().shard(devices_2, axis=0).contiguous().realize()
     self.assertUsed(self.N*self.N*4) # sharding should not increase total ram usage

test/test_ops.py

@@ -928,6 +928,12 @@ class TestOps(unittest.TestCase):
       for j in [-1., 0., 1.]:
         helper_test_op(None, torch.copysign, Tensor.copysign, vals=[[i], [j]])
 
+  def test_logaddexp(self):
+    helper_test_op([(45,65), (45,65)], torch.logaddexp, Tensor.logaddexp)
+    helper_test_op(None, torch.logaddexp, Tensor.logaddexp, vals=[[-1.], [-1.0, 2, 3]])
+    helper_test_op(None, torch.logaddexp, Tensor.logaddexp, vals=[[-100.0, -200, -300], [-1.0, 2, 3]])
+    helper_test_op(None, torch.logaddexp, Tensor.logaddexp, vals=[[1.0, 2000, 30000], [-1.0, 2, 3]])
+
   def test_softsign(self):
     helper_test_op([(45,65)], torch.nn.functional.softsign, Tensor.softsign)
     helper_test_op([()], torch.nn.functional.softsign, Tensor.softsign)
@@ -965,8 +971,6 @@ class TestOps(unittest.TestCase):
     helper_test_op([(45,65)], torch.nn.functional.softplus, Tensor.softplus, grad_atol=1e-6)
     helper_test_op([(45,65)], lambda t: torch.nn.functional.softplus(t, beta=3), lambda t: Tensor.softplus(t, beta=3), grad_atol=1e-6)
     helper_test_op([(45,65)], lambda t: torch.nn.functional.softplus(t, beta=1/3), lambda t: Tensor.softplus(t, beta=1/3), grad_atol=1e-6)
-    helper_test_op([(45,65)], lambda t: torch.nn.functional.softplus(t, beta=3, threshold=0.5),
-                              lambda t: Tensor.softplus(t, beta=3, threshold=0.5), grad_atol=1e-6)
     helper_test_op([(45,65)], torch.nn.functional.softplus, Tensor.softplus, grad_atol=1e-6, low=300, high=400)
     helper_test_op([(45,65)], torch.nn.functional.softplus, Tensor.softplus, grad_atol=1e-6, low=-400, high=-300)
     helper_test_op([()], torch.nn.functional.softplus, Tensor.softplus, grad_atol=1e-6)
@@ -2461,6 +2465,20 @@ class TestOps(unittest.TestCase):
       lambda x: Tensor.max_unpool2d(*Tensor.max_pool2d(x, kernel_size=(2,2), return_indices=True),
                                     kernel_size=(2,2), output_size=(99,99,7,6)), forward_only=True)
 
+  def test_max_unpool2d_inf(self):
+    data = [[[[math.inf, -math.inf, math.nan], [1.0, 2.0, 3.0]]]]
+    ksz = (2,2)
+    helper_test_op((),
+      lambda: torch.nn.functional.max_unpool2d(
+        *torch.nn.functional.max_pool2d(torch.tensor(data), kernel_size=ksz, return_indices=True),
+        kernel_size=ksz
+      ),
+      lambda: Tensor.max_unpool2d(
+        *Tensor.max_pool2d(Tensor(data), kernel_size=ksz, return_indices=True),
+        kernel_size=ksz
+      ),
+      forward_only=True)
+
   def test_avg_pool2d(self):
     shape = (32,2,111,28)
     for ksz in [(2,2), (3,3), (3,2), (5,5), (5,1)]:
@@ -2694,6 +2712,10 @@ class TestOps(unittest.TestCase):
     i, j, k, o, p = [Tensor(tor.detach().cpu().numpy().astype(np.int32), requires_grad=False) for tor in [a,b,c,d,e]]
     return a,b,c,d,e,i,j,k,o,p
 
+  def test_fancy_indexing_inf(self):
+    data = [math.inf, -math.inf, math.nan]
+    helper_test_op((), lambda: torch.tensor(data)[torch.tensor([0, 1, 2])], lambda: Tensor(data)[Tensor([0, 1, 2])])
+
   def test_slice_fancy_indexing_no_dim_collapse(self):
     a,b,c,d,e,i,j,k,o,p = self._get_index_randoms()
     # no dim collapse from int or dim injection from None

test/test_opts.py

@@ -0,0 +1,22 @@
+import unittest
+from tinygrad import Tensor, Device
+from tinygrad.helpers import RANGEIFY
+from tinygrad.codegen.opt.kernel import Opt, OptOps
+from tinygrad.engine.realize import get_program
+
+@unittest.skipIf(RANGEIFY>0, "arg is partial contig in rangeify")
+class TestOpts(unittest.TestCase):
+  def test_opt_upcast(self):
+    opts = (Opt(OptOps.UPCAST, 0, 4),)
+    a = Tensor.empty(16)
+    b = Tensor.empty(16)
+    out = (a+b).contiguous(arg=opts)
+    s = out.schedule()
+    self.assertEqual(s[-1].ast.arg.opts_to_apply, opts)
+    if Device.DEFAULT in {"CPU", "GPU", "METAL"}:
+      prg = get_program(s[-1].ast)
+      self.assertIn('float4', prg.src)
+
+if __name__ == '__main__':
+  unittest.main()
+

test/test_outerworld_range.py

@@ -1,5 +1,5 @@
 import unittest
-from tinygrad import Tensor, nn, Variable, UOp, dtypes
+from tinygrad import Tensor, nn, Variable, UOp
 
 # outerworld range should support three things
 # 1. full optimizer steps (test_model_bound_range)
@@ -136,7 +136,7 @@ class TestOuterworldRange(unittest.TestCase):
   def test_model_bound_range(self):
     m, opt = get_model_and_opt()
     # TODO: should ranges be unique so you don't have to pass in the -1?
-    rng = UOp.range(dtypes.int, self.STEPS, -1)
+    rng = UOp.range(self.STEPS, -1)
     vib = Variable('i', 0, self.STEPS-1).bind(rng)
     loss = (m(self.X[vib]) - self.Y[vib]).square().mean()
     loss.backward()

test/test_pickle.py

@@ -20,7 +20,7 @@ class TestPickle(unittest.TestCase):
     self.assertEqual(pm2.rewrite(sink).key, tt.key)
 
   def test_pickle_main_pattern_matcher(self):
-    from tinygrad.codegen.devectorizer import sym
+    from tinygrad.codegen.late.devectorizer import sym
     ssym = pickle.dumps(sym)
     dsym = pickle.loads(ssym)
     self.assertEqual(dsym.patterns[0][0].location, sym.patterns[0][0].location)

test/test_rangeify.py

@@ -1,6 +1,7 @@
 import unittest
 from tinygrad import Tensor
-from tinygrad.helpers import RANGEIFY
+from tinygrad.helpers import RANGEIFY, Context, GlobalCounters
+from tinygrad.uop.ops import UOp
 
 N = 256
 
@@ -11,6 +12,26 @@ class TestRangeify(unittest.TestCase):
     ba = A.expand(N, N)
     ((ba+1).sum(axis=1) + (ba+2).sum(axis=0)).realize()
 
+  def test_partial_contig(self):
+    A = Tensor.empty(64, 64, 64)
+    ret = A.sum(axis=2).contiguous(arg=(1,)).sum(axis=1)
+    ret.realize()
+
+  def test_double_gemm_real(self):
+    def go():
+      with Context(DEBUG=0):
+        Tensor.manual_seed(1337)
+        A,B,C = [Tensor.randn(N, N) for _ in range(3)]
+        Tensor.realize(A, B, C)
+      GlobalCounters.reset()
+      return (A@B@C).realize()
+    rng = go()
+    with Context(RANGEIFY=0, DEBUG=2):
+      ref = go()
+      mse = ((rng-ref)**2).sum().item()
+    print(f"mse: {mse}")
+    self.assertLessEqual(mse, 1e-2)
+
   def test_double_gemm(self):
     A = Tensor.empty(N, N)
     B = Tensor.empty(N, N)
@@ -96,14 +117,101 @@ class TestRangeify(unittest.TestCase):
     out.realize()
 
   def test_flash_attention(self):
-    BS = 4
-    HEADS = 2
-    MATDIM = 16
-    EMB = 8
-    q = Tensor.empty(BS, HEADS, MATDIM, EMB)
-    k = Tensor.empty(BS, HEADS, MATDIM, EMB)
-    v = Tensor.empty(BS, HEADS, MATDIM, EMB)
-    q.scaled_dot_product_attention(k, v).realize()
+    BS, HEADS, SEQLEN, EMB = 4, 2, 16, 8
+
+    # bigger
+    #BS, HEADS, SEQLEN, EMB = 4, 32, 1024, 64
+
+    # llama 8B
+    #BS, HEADS, SEQLEN, EMB = 4, 32, 2048, 128
+
+    def fa():
+      Tensor.manual_seed(1337)
+      with Context(DEBUG=0): q,k,v = [Tensor.rand(BS, HEADS, SEQLEN, EMB).contiguous().realize() for _ in range(3)]
+      return q.scaled_dot_product_attention(k, v).realize()
+
+    with Context(DEBUG=4):
+      GlobalCounters.reset()
+      ret = fa()
+    with Context(RANGEIFY=0):
+      with Context(DEBUG=2):
+        GlobalCounters.reset()
+        cmp = fa()
+      with Context(DEBUG=0):
+        mse = ((cmp-ret)**2).sum().item()
+    print(f"mse: {mse}")
+    self.assertLessEqual(mse, 1e-6)
+
+# contiguous + reduce can support ranges?
+
+@unittest.skipIf(RANGEIFY<1, "tests only for RANGEIFY")
+class TestOuterworld(unittest.TestCase):
+  def test_passthrough_range(self):
+    t = Tensor.rand(10, 10).realize()
+
+    # passthrough ranges
+    a = UOp.range(10, -1)
+    sel = t[a]
+    cpy = sel.contiguous(a).realize()
+
+    self.assertTrue((t==cpy).all().item())
+
+  def test_flip_range(self):
+    t = Tensor.rand(10, 10).realize()
+
+    # passthrough ranges
+    a = UOp.range(10, -1)
+    sel = t[9-a]
+    cpy = sel.contiguous(a).realize()
+
+    self.assertTrue((t.flip(0)==cpy).all().item())
+
+  def test_vmap(self):
+    def f(x): return x.sum(axis=0)*2
+
+    x = Tensor.ones(3, 10, 2).contiguous()
+
+    # vmap across axis 0
+    a = UOp.range(3, -1)
+    out = f(x[a])
+    out = out.contiguous(a)
+
+    # 3x2 grid of 20
+    out.realize()
+    print(out.numpy())
+
+  def test_triple_gemm(self):
+    x = Tensor.rand(1, 16).realize()
+    W = Tensor.rand(3, 16, 16).realize()
+
+    manual = (x @ W[0] @ W[1] @ W[2]).contiguous().realize()
+
+    a = UOp.range(3, -1)
+    x = x.assign(x @ W[a])
+    out = x.contiguous(a)[-1].contiguous().realize()
+
+    self.assertTrue((manual==out).all().item())
+
+  def test_setitem_pyrange(self):
+    with Context(DEBUG=0):
+      t = Tensor.rand(10).realize()
+      o = Tensor.empty(10)
+    GlobalCounters.reset()
+    for i in range(10):
+      o[i] = t[i]
+    o.realize()
+    self.assertTrue((t==o).all().item())
+
+  @unittest.skip("TODO: fix this")
+  def test_setitem(self):
+    with Context(DEBUG=0):
+      t = Tensor.rand(10).realize()
+      o = Tensor.empty(10)
+    GlobalCounters.reset()
+    i = UOp.range(10, -1)
+    o[i] = t[i]
+    o.contiguous(i).realize()
+    self.assertTrue((t==o).all().item())
 
 if __name__ == '__main__':
   unittest.main()

test/test_renderer_failures.py

@@ -25,7 +25,8 @@ def _test_uop_result(inputs:List[Tensor], stores:List[UOp], local_size=None):
       initial_value=np.zeros(sz, dtype=_to_np_dtype(dtype)).data) for u in uops if u.op is Ops.STORE]
   inbufs = [cast(UOp,x.uop).base.buffer for x in inputs]
   src = Device[Device.DEFAULT].renderer.render(uops)
-  ei = CompiledRunner(ProgramSpec("test", src, Device.DEFAULT, uops[-1], uops=uops, local_size=local_size))
+  ei = CompiledRunner(ProgramSpec(uops[-1].arg.name if uops[-1].arg is not None else "test",
+                                  src, Device.DEFAULT, uops[-1], uops=uops, local_size=local_size))
   ei.exec(outbufs+inbufs)
   return [np.frombuffer(x.as_buffer(), _to_np_dtype(x.dtype)) for x in outbufs]
 

test/test_schedule.py

@@ -1050,6 +1050,14 @@ class TestSchedule(unittest.TestCase):
       compare = torch.nn.functional.scaled_dot_product_attention(torch.tensor(q.numpy()),torch.tensor(k.numpy()),torch.tensor(v.numpy()))
       np.testing.assert_allclose(out.numpy(), compare.numpy(), atol=1e-6, rtol=1e-3)
 
+    with Context(FUSE_ATTENTION=1):
+      out = Tensor.scaled_dot_product_attention(q,k,v)
+      run_schedule(check_schedule(out, 1))
+      if getenv("CHECK", 1):
+        import torch
+        compare = torch.nn.functional.scaled_dot_product_attention(torch.tensor(q.numpy()),torch.tensor(k.numpy()),torch.tensor(v.numpy()))
+        np.testing.assert_allclose(out.numpy(), compare.numpy(), atol=1e-6, rtol=1e-3)
+
   def test_ugly_reduceop_pairing(self):
     Tensor.manual_seed(0)
     a = Tensor.randn(4, 32).realize()

test/test_symbolic_jit.py

@@ -2,50 +2,41 @@ import unittest
 
 from test.helpers import assert_jit_cache_len
 from tinygrad import Variable, Tensor, TinyJit
-from tinygrad.helpers import Context
 import numpy as np
 
 class TestSymbolicJit(unittest.TestCase):
-  def setUp(self):
-    # A lot of these test are out of bounds, so we ignore the bounds check
-    self.context = Context(IGNORE_OOB=1)
-    self.context.__enter__()
-
-  def tearDown(self):
-    self.context.__exit__(None, None, None)
-
   def test_plus1(self):
     def f(a): return (a+1).realize()
     jf = TinyJit(f)
+    a = Tensor.rand(3, 10)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(3, i)
-      symbolic = jf(a.reshape(3, vi)).reshape(3, i).numpy()
-      expected = f(a).numpy()
+      symbolic = jf(a[:, :vi]).reshape(3, i).numpy()
+      expected = f(a[:, :i]).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
     assert_jit_cache_len(jf, 1)
 
   def test_add(self):
     def f(a, b): return (a+b).realize()
     jf = TinyJit(f)
+    a = Tensor.rand(3, 10)
+    b = Tensor.rand(3, 10)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(3, i)
-      b = Tensor.rand(3, i)
-      symbolic = jf(a.reshape(3, vi), b.reshape(3, vi)).reshape(3, i).numpy()
-      expected = f(a, b).numpy()
+      symbolic = jf(a[:, :vi], b[:, :vi]).reshape(3, i).numpy()
+      expected = f(a[:, :i], b[:, :i]).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
     assert_jit_cache_len(jf, 1)
 
   def test_matmul(self):
     def f(a, b): return (a@b).realize()
     jf = TinyJit(f)
+    a = Tensor.rand(3, 10)
+    b = Tensor.rand(10, 5)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(3, i)
-      b = Tensor.rand(i, 5)
-      symbolic = jf(a.reshape(3, vi), b.reshape(vi, 5)).numpy()
-      expected = f(a, b).numpy()
+      symbolic = jf(a[:, :vi], b[:vi, :]).numpy()
+      expected = f(a[:, :i], b[:i, :]).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
     assert_jit_cache_len(jf, 1)
 
@@ -55,119 +46,119 @@ class TestSymbolicJit(unittest.TestCase):
       s = (s+s).realize() # this one does not have symbols in input
       return s
     jf = TinyJit(f)
+    a = Tensor.rand(3, 10)
+    b = Tensor.rand(10, 5)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(3, i)
-      b = Tensor.rand(i, 5)
-      symbolic = jf(a.reshape(3, vi), b.reshape(vi, 5)).numpy()
-      expected = f(a, b).numpy()
+      symbolic = jf(a[:, :vi], b[:vi, :]).numpy()
+      expected = f(a[:, :i], b[:i, :]).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
     assert_jit_cache_len(jf, 2)
 
   def test_attention(self):
     def f(q, k, v): return Tensor.scaled_dot_product_attention(q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2)).realize()
     jf = TinyJit(f)
+    q = Tensor.rand(2, 1, 4, 8)
+    k = Tensor.rand(2, 10, 4, 8)
+    v = Tensor.rand(2, 10, 4, 8)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      q = Tensor.rand(2, 1, 4, 8)
-      k = Tensor.rand(2, i, 4, 8)
-      v = Tensor.rand(2, i, 4, 8)
-      symbolic = jf(q, k.reshape(2, vi, 4, 8), v.reshape(2, vi, 4, 8)).reshape(2, 4, 1, 8).numpy()
-      expected = f(q, k, v).numpy()
+      symbolic = jf(q, k[:, :vi], v[:, :vi]).reshape(2, 4, 1, 8).numpy()
+      expected = f(q, k[:, :i], v[:, :i]).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
     assert_jit_cache_len(jf, 5)
 
   def test_cat_dim0(self):
     def f(a, b): return a.cat(b, dim=0).realize()
     jf = TinyJit(f)
+    a = Tensor.rand(10, 3)
+    b = Tensor.rand(2, 3)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(i, 3)
-      b = Tensor.rand(2, 3)
-      symbolic = jf(a.reshape(vi, 3), b).reshape(i+2, 3).numpy()
-      expected = f(a, b).numpy()
+      symbolic = jf(a[:vi], b).reshape(i+2, 3).numpy()
+      expected = f(a[:i], b).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
     assert_jit_cache_len(jf, 1)
 
   def test_cat_dim1(self):
     def f(a, b): return a.cat(b, dim=1).realize()
     jf = TinyJit(f)
+    a = Tensor.rand(3, 10)
+    b = Tensor.rand(3, 2)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(3, i)
-      b = Tensor.rand(3, 2)
-      symbolic = jf(a.reshape(3, vi), b).reshape(3, i+2).numpy()
-      expected = f(a, b).numpy()
+      symbolic = jf(a[:, :vi], b).reshape(3, i+2).numpy()
+      expected = f(a[:, :i], b).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
     assert_jit_cache_len(jf, 1)
 
   def test_cat_dim0_two_vars(self):
     def f(a, b): return a.cat(b, dim=0).realize()
     jf = TinyJit(f)
+    a = Tensor.rand(10, 3)
+    b = Tensor.rand(10, 3)
     for i in range(1, 5):
       for j in range(1, 5):
         vi = Variable("i", 1, 10).bind(i)
         vj = Variable("j", 1, 10).bind(j)
-        a = Tensor.rand(i, 3)
-        b = Tensor.rand(j, 3)
-        symbolic = jf(a.reshape(vi, 3), b.reshape(vj, 3)).reshape(i+j, 3).numpy()
-        expected = f(a, b).numpy()
+        symbolic = jf(a[:vi], b[:vj]).reshape(i+j, 3).numpy()
+        expected = f(a[:i], b[:j]).numpy()
         np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
     assert_jit_cache_len(jf, 1)
 
   def test_cat_dim1_two_vars(self):
     def f(a, b): return a.cat(b, dim=1).realize()
     jf = TinyJit(f)
+    a = Tensor.rand(3, 10)
+    b = Tensor.rand(3, 10)
     for i in range(1, 5):
       for j in range(1, 5):
         vi = Variable("i", 1, 10).bind(i)
         vj = Variable("j", 1, 10).bind(j)
-        a = Tensor.rand(3, i)
-        b = Tensor.rand(3, j)
-        symbolic = jf(a.reshape(3, vi), b.reshape(3, vj)).reshape(3, i+j).numpy()
-        expected = f(a, b).numpy()
+        symbolic = jf(a[:, :vi], b[:, :vj]).reshape(3, i+j).numpy()
+        expected = f(a[:, :i], b[:, :j]).numpy()
         np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
     assert_jit_cache_len(jf, 1)
 
   def test_two_vars_plus1_ij(self):
     def f(a, b): return (a@b+1).realize()
     jf = TinyJit(f)
+    a = Tensor.rand(10, 3)
+    b = Tensor.rand(3, 10)
     for i in range(1, 5):
       for j in range(1, 5):
         vi = Variable("i", 1, 10).bind(i)
         vj = Variable("j", 1, 10).bind(j)
-        a = Tensor.rand(i, 3)
-        b = Tensor.rand(3, j)
-        symbolic = jf(a.reshape(vi, 3), b.reshape(3, vj)).reshape(i, j).numpy()
-        expected = f(a, b).numpy()
+        symbolic = jf(a[:vi, :], b[:, :vj]).reshape(i, j).numpy()
+        expected = f(a[:i, :], b[:, :j]).numpy()
         np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
     assert_jit_cache_len(jf, 1)
 
   def test_two_vars_plus1_ji(self):
     def f(a, b): return (a@b+1).realize()
     jf = TinyJit(f)
+    a = Tensor.rand(10, 3)
+    b = Tensor.rand(3, 10)
     for i in range(1, 5):
       for j in range(1, 5):
         vi = Variable("i", 1, 10).bind(i)
         vj = Variable("j", 1, 10).bind(j)
-        a = Tensor.rand(j, 3)
-        b = Tensor.rand(3, i)
-        symbolic = jf(a.reshape(vj, 3), b.reshape(3, vi)).reshape(j, i).numpy()
-        expected = f(a, b).numpy()
+        symbolic = jf(a[:vj, :], b[:, :vi]).reshape(j, i).numpy()
+        expected = f(a[:j, :], b[:, :i]).numpy()
         np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
     assert_jit_cache_len(jf, 1)
 
   def test_jit_symbolic_shape_mismatch(self):
     @TinyJit
     def add(a, b): return (a+b).realize()
+    a = Tensor.rand(3, 10)
+    b = Tensor.rand(3, 10)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(3, i).reshape(3, vi)
-      b = Tensor.rand(3, i).reshape(3, vi)
-      add(a, b)
+      add(a[:, :vi], b[:, :vi])
     vi2 = Variable("i", 1, 10).bind(7)
-    a = Tensor.rand(3, 7).reshape(3, vi2)
-    bad = Tensor.rand(4, 7).reshape(4, vi2)
+    a = Tensor.rand(3, 7)[:, :vi2]
+    bad = Tensor.rand(4, 7)[:, :vi2]
     with self.assertRaises(AssertionError):
       add(a, bad)
 
@@ -175,9 +166,9 @@ class TestSymbolicJit(unittest.TestCase):
     # shrink is a movement, so we pair it with a simple function to test the JIT interaction
     def f(a): return (a+1).realize()
     jf = TinyJit(f)
+    a = Tensor.rand(7, 11)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(7, 11)
       symbolic = a.shrink(((3,5),(vi,vi+2)))
       symbolic = jf(symbolic).numpy()
       expected = f(a.shrink(((3,5),(i,i+2)))).numpy()
@@ -188,9 +179,9 @@ class TestSymbolicJit(unittest.TestCase):
     # slice is a movement, so we pair it with a simple function to test the JIT interaction
     def f(a): return (a+1).realize()
     jf = TinyJit(f)
+    a = Tensor.rand(7, 11)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(7, 11)
       symbolic = a[3:5, vi:vi+2]
       symbolic = jf(symbolic).numpy()
       expected = f(a[3:5, i:i+2]).numpy()
@@ -212,11 +203,11 @@ class TestSymbolicJit(unittest.TestCase):
   def test_ones_sum(self):
     def f(a): return a.sum().realize()
     jf = TinyJit(f)
+    t = Tensor.ones(10)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      t = Tensor.ones(i)
-      symbolic = jf(t.reshape(vi)).item()
-      expected = f(t).item()
+      symbolic = jf(t[:vi]).item()
+      expected = f(t[:i]).item()
       np.testing.assert_equal(symbolic, expected)
 
   def test_mean(self):
@@ -226,22 +217,22 @@ class TestSymbolicJit(unittest.TestCase):
     jf = TinyJit(f)
     jf0 = TinyJit(f0)
     jf1 = TinyJit(f1)
+    a = Tensor.rand(10, 3)
+    b = Tensor.rand(10, 3)
+    c = Tensor.rand(10, 3)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      # aixs = None
-      a = Tensor.rand(i, 3)
-      symbolic = jf(a.reshape(vi, 3)).numpy()
-      expected = a.mean().numpy()
+      # axis = None
+      symbolic = jf(a[:vi]).numpy()
+      expected = a[:i].mean().numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
-      # aixs = 0
-      a = Tensor.rand(i, 3)
-      symbolic = jf0(a.reshape(vi, 3)).numpy()
-      expected = a.mean(0).numpy()
+      # axis = 0
+      symbolic = jf0(b[:vi]).numpy()
+      expected = b[:i].mean(0).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
-      # aixs = 1
-      a = Tensor.rand(i, 3)
-      symbolic = jf1(a.reshape(vi, 3)).reshape(i).numpy()
-      expected = a.mean(1).numpy()
+      # axis = 1
+      symbolic = jf1(c[:vi]).reshape(i).numpy()
+      expected = c[:i].mean(1).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_mean_2d(self):
@@ -251,24 +242,24 @@ class TestSymbolicJit(unittest.TestCase):
     jf = TinyJit(f)
     jf0 = TinyJit(f0)
     jf1 = TinyJit(f1)
+    a = Tensor.rand(10, 10)
+    b = Tensor.rand(10, 10)
+    c = Tensor.rand(10, 10)
     for i in range(1, 5):
       for j in range(1, 5):
         vi = Variable("i", 1, 10).bind(i)
         vj = Variable("j", 1, 10).bind(j)
-        # aixs = None
-        a = Tensor.rand(i, j)
-        symbolic = jf(a.reshape(vi, vj)).numpy()
-        expected = a.mean().numpy()
+        # axis = None
+        symbolic = jf(a[:vi, :vj]).numpy()
+        expected = a[:i, :j].mean().numpy()
         np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
-        # aixs = 0
-        a = Tensor.rand(i, j)
-        symbolic = jf0(a.reshape(vi, vj)).reshape(j).numpy()
-        expected = a.mean(0).numpy()
+        # axis = 0
+        symbolic = jf0(b[:vi, :vj]).reshape(j).numpy()
+        expected = b[:i, :j].mean(0).numpy()
         np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
-        # aixs = 1
-        a = Tensor.rand(i, j)
-        symbolic = jf1(a.reshape(vi, vj)).reshape(i).numpy()
-        expected = a.mean(1).numpy()
+        # axis = 1
+        symbolic = jf1(c[:vi, :vj]).reshape(i).numpy()
+        expected = c[:i, :j].mean(1).numpy()
         np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_var(self):
@@ -278,22 +269,22 @@ class TestSymbolicJit(unittest.TestCase):
     jf = TinyJit(f)
     jf0 = TinyJit(f0)
     jf1 = TinyJit(f1)
+    a = Tensor.rand(10, 3)
+    b = Tensor.rand(10, 3)
+    c = Tensor.rand(10, 3)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      # aixs = None
-      a = Tensor.rand(i, 3)
-      symbolic = jf(a.reshape(vi, 3)).numpy()
-      expected = a.var().numpy()
+      # axis = None
+      symbolic = jf(a[:vi]).numpy()
+      expected = a[:i].var().numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
-      # aixs = 0
-      a = Tensor.rand(i, 3)
-      symbolic = jf0(a.reshape(vi, 3)).numpy()
-      expected = a.var(0).numpy()
+      # axis = 0
+      symbolic = jf0(b[:vi]).numpy()
+      expected = b[:i].var(0).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
-      # aixs = 1
-      a = Tensor.rand(i, 3)
-      symbolic = jf1(a.reshape(vi, 3)).reshape(i).numpy()
-      expected = a.var(1).numpy()
+      # axis = 1
+      symbolic = jf1(c[:vi]).reshape(i).numpy()
+      expected = c[:i].var(1).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_var_2d(self):
@@ -303,24 +294,24 @@ class TestSymbolicJit(unittest.TestCase):
     jf = TinyJit(f)
     jf0 = TinyJit(f0)
     jf1 = TinyJit(f1)
+    a = Tensor.rand(10, 10)
+    b = Tensor.rand(10, 10)
+    c = Tensor.rand(10, 10)
     for i in range(1, 5):
       for j in range(1, 5):
         vi = Variable("i", 1, 10).bind(i)
         vj = Variable("j", 1, 10).bind(j)
-        # aixs = None
-        a = Tensor.rand(i, j)
-        symbolic = jf(a.reshape(vi, vj)).numpy()
-        expected = a.var().numpy()
+        # axis = None
+        symbolic = jf(a[:vi, :vj]).numpy()
+        expected = a[:i, :j].var().numpy()
         np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
-        # aixs = 0
-        a = Tensor.rand(i, j)
-        symbolic = jf0(a.reshape(vi, vj)).reshape(j).numpy()
-        expected = a.var(0).numpy()
+        # axis = 0
+        symbolic = jf0(b[:vi, :vj]).reshape(j).numpy()
+        expected = b[:i, :j].var(0).numpy()
         np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
-        # aixs = 1
-        a = Tensor.rand(i, j)
-        symbolic = jf1(a.reshape(vi, vj)).reshape(i).numpy()
-        expected = a.var(1).numpy()
+        # axis = 1
+        symbolic = jf1(c[:vi, :vj]).reshape(i).numpy()
+        expected = c[:i, :j].var(1).numpy()
         np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
 if __name__ == '__main__':

test/test_symbolic_ops.py

@@ -1,62 +1,53 @@
 import unittest
-from tinygrad import Tensor, Variable
+from tinygrad import Tensor, Variable, GlobalCounters
 from tinygrad.shape.shapetracker import View
-from tinygrad.helpers import Context, GlobalCounters
 from tinygrad.uop.ops import sym_infer
 from tinygrad.dtype import dtypes
-from tinygrad.device import Device
+from tinygrad.device import is_dtype_supported
 from examples.gpt2 import Attention
 import numpy as np
 
 class TestSymbolicOps(unittest.TestCase):
-  def setUp(self):
-    # A lot of these test are out of bounds, so we ignore the bounds check
-    self.context = Context(IGNORE_OOB=1)
-    self.context.__enter__()
-
-  def tearDown(self):
-    self.context.__exit__(None, None, None)
-
   def test_plus1(self):
     def f(a): return (a+1).realize()
+    a = Tensor.rand(3, 10)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(3, i)
-      symbolic = f(a.reshape(3, vi)).reshape(3, i).numpy()
-      expected = f(a).numpy()
+      symbolic = f(a[:, :vi]).reshape(3, i).numpy()
+      expected = f(a[:, :i]).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_add(self):
     def f(a, b): return (a+b).realize()
+    a = Tensor.rand(3, 10)
+    b = Tensor.rand(3, 10)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(3, i)
-      b = Tensor.rand(3, i)
-      symbolic = f(a.reshape(3, vi), b.reshape(3, vi)).reshape(3, i).numpy()
-      expected = f(a, b).numpy()
+      symbolic = f(a[:, :vi], b[:, :vi]).reshape(3, i).numpy()
+      expected = f(a[:, :i], b[:, :i]).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_matmul(self):
     def f(a, b): return (a@b).realize()
+    a = Tensor.rand(3, 10)
+    b = Tensor.rand(10, 5)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(3, i)
-      b = Tensor.rand(i, 5)
-      symbolic = f(a.reshape(3, vi), b.reshape(vi, 5)).numpy()
-      expected = f(a, b).numpy()
+      symbolic = f(a[:, :vi], b[:vi, :]).numpy()
+      expected = f(a[:, :i], b[:i, :]).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_attention(self, dropout_p=0.0, imin=1, imax=5, use_symbolic=True):
     def f(q, k, v): return Tensor.scaled_dot_product_attention(q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2), dropout_p=dropout_p).realize()
+    q = Tensor.rand(2, 1, 4, 8)
+    k = Tensor.rand(2, 10, 4, 8)
+    v = Tensor.rand(2, 10, 4, 8)
     for i in range(imin, imax):
       vi = Variable("i", 1, 10).bind(i) if use_symbolic else i
-      q = Tensor.rand(2, 1, 4, 8)
-      k = Tensor.rand(2, i, 4, 8)
-      v = Tensor.rand(2, i, 4, 8)
       Tensor.realize(q, k, v)
       GlobalCounters.reset()
-      symbolic = f(q, k.reshape(2, vi, 4, 8), v.reshape(2, vi, 4, 8)).reshape(2, 4, 1, 8).numpy()
-      expected = f(q, k, v).numpy()
+      symbolic = f(q, k[:, :vi, :, :], v[:, :vi, :, :]).reshape(2, 4, 1, 8).numpy()
+      expected = f(q, k[:, :i, :, :], v[:, :i, :, :]).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_attention_cmp_symbolic(self):
@@ -90,73 +81,89 @@ class TestSymbolicOps(unittest.TestCase):
 
   def test_cat_dim0(self):
     def f(a, b): return a.cat(b, dim=0).realize()
+    a = Tensor.rand(10, 3)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(i, 3)
       b = Tensor.rand(2, 3)
-      symbolic = f(a.reshape(vi, 3), b).reshape(i+2, 3).numpy()
-      expected = f(a, b).numpy()
+      symbolic = f(a[:vi, :], b).reshape(i+2, 3).numpy()
+      expected = f(a[:i, :], b).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_cat_dim1(self):
     def f(a, b): return a.cat(b, dim=1).realize()
+    a = Tensor.rand(3, 10)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(3, i)
       b = Tensor.rand(3, 2)
-      symbolic = f(a.reshape(3, vi), b).reshape(3, i+2).numpy()
-      expected = f(a, b).numpy()
+      symbolic = f(a[:, :vi], b).reshape(3, i+2).numpy()
+      expected = f(a[:, :i], b).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_cat_dim0_two_vars(self):
     def f(a, b): return a.cat(b, dim=0).realize()
+    a = Tensor.rand(10, 3)
+    b = Tensor.rand(10, 3)
     for i in range(1, 5):
       for j in range(1, 5):
         vi = Variable("i", 1, 10).bind(i)
         vj = Variable("j", 1, 10).bind(j)
-        a = Tensor.rand(i, 3)
-        b = Tensor.rand(j, 3)
-        symbolic = f(a.reshape(vi, 3), b.reshape(vj, 3)).reshape(i+j, 3).numpy()
-        expected = f(a, b).numpy()
+        symbolic = f(a[:vi, :], b[:vj, :]).reshape(i+j, 3).numpy()
+        expected = f(a[:i, :], b[:j, :]).numpy()
         np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_cat_dim1_two_vars(self):
     def f(a, b): return a.cat(b, dim=1).realize()
+    a = Tensor.rand(3, 10)
+    b = Tensor.rand(3, 10)
     for i in range(1, 5):
       for j in range(1, 5):
         vi = Variable("i", 1, 10).bind(i)
         vj = Variable("j", 1, 10).bind(j)
-        a = Tensor.rand(3, i)
-        b = Tensor.rand(3, j)
-        symbolic = f(a.reshape(3, vi), b.reshape(3, vj)).reshape(3, i+j).numpy()
-        expected = f(a, b).numpy()
+        symbolic = f(a[:, :vi], b[:, :vj]).reshape(3, i+j).numpy()
+        expected = f(a[:, :i], b[:, :j]).numpy()
         np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_two_vars_plus1_ij(self):
     def f(a, b): return (a@b+1).realize()
+    a = Tensor.rand(10, 3)
+    b = Tensor.rand(3, 10)
     for i in range(1, 5):
       for j in range(1, 5):
         vi = Variable("i", 1, 10).bind(i)
         vj = Variable("j", 1, 10).bind(j)
-        a = Tensor.rand(i, 3)
-        b = Tensor.rand(3, j)
-        symbolic = f(a.reshape(vi, 3), b.reshape(3, vj)).reshape(i, j).numpy()
-        expected = f(a, b).numpy()
+        symbolic = f(a[:vi, :], b[:, :vj]).reshape(i, j).numpy()
+        expected = f(a[:i, :], b[:, :j]).numpy()
         np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_two_vars_plus1_ji(self):
     # reverse the order of variables
     def f(a, b): return (a@b+1).realize()
+    a = Tensor.rand(10, 3)
+    b = Tensor.rand(3, 10)
     for i in range(1, 5):
       for j in range(1, 5):
         vi = Variable("i", 1, 10).bind(i)
         vj = Variable("j", 1, 10).bind(j)
-        a = Tensor.rand(j, 3)
-        b = Tensor.rand(3, i)
-        symbolic = f(a.reshape(vj, 3), b.reshape(3, vi)).reshape(j, i).numpy()
-        expected = f(a, b).numpy()
+        symbolic = f(a[:vj, :], b[:, :vi]).reshape(j, i).numpy()
+        expected = f(a[:j, :], b[:, :i]).numpy()
         np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
+  def test_reshape_from_symbolic(self):
+    a = Tensor.rand(30)
+    for i in range(3, 5):
+      vi = Variable("i", 3, 10).bind(i)
+      symbolic = a[:vi*3].reshape((3, 3)).numpy()
+      # To match symbolic reshape (potential implicit shrink), we need a shrink
+      expected = a[:i*3].shrink(((0, 9),)).reshape((3, 3)).numpy()
+      np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
+
+  def test_invalid_symbolic_reshape(self):
+    a = Tensor.rand(30)
+    for i in range(1, 5):
+      vi = Variable("i", 1, 10).bind(i)
+      # Cannot reshape into symbolic from non-symbolic
+      with self.assertRaises(AssertionError): a.reshape((3, vi))
+
   def test_shrink(self):
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
@@ -176,11 +183,10 @@ class TestSymbolicOps(unittest.TestCase):
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_slice_no_start(self):
+    a = Tensor.rand(7, 11)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(7, 11)
-      symbolic = a[3:5, :vi:1].reshape(2,i)
-      symbolic = symbolic.numpy()
+      symbolic = a[3:5, :vi:1].reshape(2, i).numpy()
       expected = a[3:5, :i:1].numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
@@ -201,31 +207,31 @@ class TestSymbolicOps(unittest.TestCase):
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_ones_sum(self):
+    t = Tensor.ones(10)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      t = Tensor.ones(i)
-      symbolic = t.reshape(vi).sum().item()
-      expected = t.sum().item()
+      symbolic = t[:vi].sum().item()
+      expected = t[:i].sum().item()
       np.testing.assert_equal(symbolic, expected)
 
   def test_mean(self):
+    a = Tensor.rand(10, 3)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
       for axis in [None, 0, 1]:
-        a = Tensor.rand(i, 3)
-        expected = a.mean(axis).numpy()
-        symbolic = a.reshape(vi, 3).mean(axis).reshape(expected.shape).numpy()
+        expected = a[:i].mean(axis).numpy()
+        symbolic = a[:vi].mean(axis).reshape(expected.shape).numpy()
         np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_mean_2d(self):
+    a = Tensor.rand(10, 10)
     for i in range(1, 5):
       for j in range(1, 5):
         vi = Variable("i", 1, 10).bind(i)
         vj = Variable("j", 1, 10).bind(j)
         for axis in [None, 0, 1]:
-          a = Tensor.rand(i, j)
-          expected = a.mean(axis).numpy()
-          symbolic = a.reshape(vi, vj).mean(axis).reshape(expected.shape).numpy()
+          expected = a[:i, :j].mean(axis).numpy()
+          symbolic = a[:vi, :vj].mean(axis).reshape(expected.shape).numpy()
           np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_var(self):
@@ -233,43 +239,43 @@ class TestSymbolicOps(unittest.TestCase):
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
       for axis in [None, 0, 1]:
-        expected = a[:i, :].var(axis).numpy()
-        symbolic = a[:vi, :].var(axis).reshape(expected.shape).numpy()
+        expected = a[:i].var(axis).numpy()
+        symbolic = a[:vi].var(axis).reshape(expected.shape).numpy()
         np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_var_2d(self):
+    a = Tensor.rand(10, 10)
     for i in range(1, 5):
       for j in range(1, 5):
         vi = Variable("i", 1, 10).bind(i)
         vj = Variable("j", 1, 10).bind(j)
         for axis in [None, 0, 1]:
-          a = Tensor.rand(i, j)
-          expected = a.var(axis).numpy()
-          symbolic = a.reshape(vi, vj).var(axis).reshape(expected.shape).numpy()
+          expected = a[:i, :j].var(axis).numpy()
+          symbolic = a[:vi, :vj].var(axis).reshape(expected.shape).numpy()
           np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_bitcast_down(self):
+    a = Tensor.rand(10, 3)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(i, 3)
-      expected = a.bitcast(dtypes.uint8).numpy()
-      symbolic = a.reshape(vi, 3).bitcast(dtypes.uint8).reshape(expected.shape).numpy()
+      expected = a[:i].bitcast(dtypes.uint8).numpy()
+      symbolic = a[:vi].bitcast(dtypes.uint8).reshape(expected.shape).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=0)
 
-  @unittest.skipIf(Device.DEFAULT == "WEBGPU", "no uint64")
+  @unittest.skipUnless(is_dtype_supported(dtypes.uint64), "no uint64")
   def test_bitcast_up(self):
+    a = Tensor.rand(10, 4)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(i, 4)
-      expected = a.bitcast(dtypes.uint64).numpy()
-      symbolic = a.reshape(vi, 4).bitcast(dtypes.uint64).reshape(expected.shape).numpy()
+      expected = a[:i].bitcast(dtypes.uint64).numpy()
+      symbolic = a[:vi].bitcast(dtypes.uint64).reshape(expected.shape).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=0)
 
   @unittest.expectedFailure
   def test_conv2d_ceildiv_edge_case(self):
     v = Variable('v', 11, 50_000)
     val = 39601
-    x = Tensor.randn(1, 22, 39601).reshape(1, 22, v.bind(val))
+    x = Tensor.randn(1, 22, 50_000)[:, :, :v.bind(val)]
     weight = Tensor.randn(256, 22, 12)
 
     result = x.conv2d(weight=weight, groups=1, stride=6, dilation=1, padding=(3, 3))

test/test_tensor.py

@@ -415,6 +415,21 @@ class TestTinygrad(unittest.TestCase):
         data = _generate_data(depth)
         np.testing.assert_allclose(Tensor(data).numpy(), np.array(data))
 
+  def test_tensor_list_implicit_cast(self):
+    data = [True, False]
+    np.testing.assert_equal(Tensor(data, dtype=dtypes.int).numpy(), torch.tensor(data, dtype=torch.int).numpy())
+    np.testing.assert_equal(Tensor(data, dtype=dtypes.uint8).numpy(), torch.tensor(data, dtype=torch.uint8).numpy())
+    np.testing.assert_equal(Tensor(data, dtype=dtypes.float).numpy(), torch.tensor(data, dtype=torch.float).numpy())
+    data = [-1, 0, 1, 2, 3]
+    np.testing.assert_equal(Tensor(data, dtype=dtypes.int).numpy(), torch.tensor(data, dtype=torch.int).numpy())
+    np.testing.assert_equal(Tensor(data, dtype=dtypes.uint8).numpy(), torch.tensor(data, dtype=torch.uint8).numpy())
+    np.testing.assert_equal(Tensor(data, dtype=dtypes.float).numpy(), torch.tensor(data, dtype=torch.float).numpy())
+    data = [-3.5, -2.5, -1.5, 0, 1.5, 2.5, 3.5]
+    np.testing.assert_equal(Tensor(data, dtype=dtypes.int).numpy(), torch.tensor(data, dtype=torch.int).numpy())
+    # NOTE: torch and jax raise OverflowError: Python integer -3 out of bounds for uint8
+    # np.testing.assert_equal(Tensor(data, dtype=dtypes.uint8).numpy(), torch.tensor(data, dtype=torch.uint8).numpy())
+    np.testing.assert_equal(Tensor(data, dtype=dtypes.float).numpy(), torch.tensor(data, dtype=torch.float).numpy())
+
   def test_tensor_list_special_values(self):
     if is_dtype_supported(dtypes.float16):
       data = [math.nan, -math.inf, 65504, 65519, 65519.999, 65520, 65520.1]

test/test_tensor_variable.py

@@ -1,7 +1,6 @@
 import unittest
 import numpy as np
 from tinygrad import Tensor, Variable
-from tinygrad.helpers import Context
 
 class TestTensorVariable(unittest.TestCase):
   def test_add_tvar(self):
@@ -23,43 +22,38 @@ class TestTensorVariable(unittest.TestCase):
     assert (Tensor(3) * (vv * 4)).item() == 24
 
   def test_symbolic_mean(self):
-    with Context(IGNORE_OOB=1):
-      vv = Variable("a", 1, 10).bind(2)
-      t = Tensor.ones(2, 2).contiguous().reshape(2, vv)
-      ret = t.mean().item()
-      assert ret == 1
+    vv = Variable("a", 1, 10).bind(2)
+    t = Tensor.ones(2, 10).contiguous()[:, :vv]
+    ret = t.mean().item()
+    assert ret == 1
 
   def test_symbolic_mean_2d(self):
-    with Context(IGNORE_OOB=1):
-      vv = Variable("a", 1, 10).bind(2)
-      vv2 = Variable("b", 1, 10).bind(2)
-      t = Tensor.ones(2, 2).contiguous().reshape(vv2, vv)
-      ret = t.mean().item()
-      assert ret == 1
+    vv = Variable("a", 1, 10).bind(2)
+    vv2 = Variable("b", 1, 10).bind(2)
+    t = Tensor.ones(10, 10).contiguous()[:vv2, :vv]
+    ret = t.mean().item()
+    assert ret == 1
 
   def test_symbolic_mean_2d_axis_1(self):
-    with Context(IGNORE_OOB=1):
-      vv = Variable("a", 1, 10).bind(2)
-      vv2 = Variable("b", 1, 10).bind(2)
-      t = Tensor.ones(2, 2).contiguous().reshape(vv2, vv)
-      ret = t.mean(axis=1).reshape(2, 1).numpy()
-      assert np.all(ret == 1)
+    vv = Variable("a", 1, 10).bind(2)
+    vv2 = Variable("b", 1, 10).bind(2)
+    t = Tensor.ones(10, 10).contiguous()[:vv2, :vv]
+    ret = t.mean(axis=1).reshape(2, 1).numpy()
+    assert np.all(ret == 1)
 
   def test_symbolic_mean_2d_add(self):
-    with Context(IGNORE_OOB=1):
-      add_term = Variable("c", 0, 10).bind(1)
-      vv = Variable("a", 1, 10).bind(1)
-      vv2 = Variable("b", 1, 10).bind(1)
-      t = Tensor.ones(2, 2).contiguous().reshape(vv2+add_term, vv+add_term)
-      ret = t.mean().item()
-      assert ret == 1
+    add_term = Variable("c", 0, 10).bind(1)
+    vv = Variable("a", 1, 10).bind(1)
+    vv2 = Variable("b", 1, 10).bind(1)
+    t = Tensor.ones(20, 20).contiguous()[:vv2+add_term, :vv+add_term]
+    ret = t.mean().item()
+    assert ret == 1
 
   def test_symbolic_var(self):
-    with Context(IGNORE_OOB=1):
-      vv = Variable("a", 1, 10).bind(2)
-      t = Tensor.ones(2, 2).contiguous().reshape(2, vv)
-      ret = t.var().item()
-      assert ret == 0
+    vv = Variable("a", 1, 10).bind(2)
+    t = Tensor.ones(2, 10).contiguous()[:, :vv]
+    ret = t.var().item()
+    assert ret == 0
 
   def test_symbolic_pad(self):
     vv = Variable("a", 1, 10).bind(2)
@@ -92,5 +86,15 @@ class TestTensorVariable(unittest.TestCase):
     ret = Tensor.arange(begin.bind(4), end.bind(7))
     self.assertListEqual(ret.reshape(3).tolist(), [4,5,6])
 
+  def test_variable_empty(self):
+    v = Variable("i", 1, 10)
+    # TODO: Tensor creation from unbound variable should assert
+    # with self.assertRaises(AssertionError): t = Tensor.empty(3, v)
+    vb = v.bind(3)
+    t = Tensor.empty(3, vb)
+    assert t.uop.base.buffer.size == 30
+    assert t.uop.st.shape == (3, vb)
+
+
 if __name__ == '__main__':
   unittest.main()

test/test_tiny.py

@@ -1,7 +1,7 @@
 # basic self-contained tests of the external functionality of tinygrad
 import unittest, random
 from tinygrad import Tensor, Context, Variable, TinyJit, dtypes, Device, nn
-from tinygrad.helpers import IMAGE, CI
+from tinygrad.helpers import IMAGE, CI, getenv
 
 class TestTiny(unittest.TestCase):
 
@@ -27,10 +27,21 @@ class TestTiny(unittest.TestCase):
     out = Tensor.ones(256).contiguous().sum()
     self.assertEqual(out.item(), 256)
 
-  def test_gemm(self, N=64, out_dtype=dtypes.float):
+  def test_gemm(self, N=getenv("GEMM_N", 64), out_dtype=dtypes.float):
     a = Tensor.ones(N,N).contiguous()
     b = Tensor.eye(N).contiguous()
-    self.assertListEqual((out:=a@b).flatten().tolist(), [1.0]*(N*N))
+    lst = (out:=a@b).tolist()
+    for y in range(N):
+      for x in range(N):
+        self.assertEqual(lst[y][x], 1.0, msg=f"mismatch at ({y},{x})")
+    if IMAGE < 2: self.assertEqual(out.dtype, out_dtype)
+
+  def test_gemv(self, N=getenv("GEMV_N", 64), out_dtype=dtypes.float):
+    a = Tensor.ones(1,N).contiguous()
+    b = Tensor.eye(N).contiguous()
+    lst = (out:=a@b).tolist()
+    for x in range(N):
+      self.assertEqual(lst[0][x], 1.0, msg=f"mismatch at {x}")
     if IMAGE < 2: self.assertEqual(out.dtype, out_dtype)
 
   # *** randomness ***

test/test_uop_graph.py

@@ -6,7 +6,7 @@ from tinygrad.helpers import DEBUG, Context
 from tinygrad.uop.ops import Ops, UOp, UPat, PatternMatcher, track_rewrites, graph_rewrite, GroupOp
 from tinygrad.uop.symbolic import sym
 from tinygrad.codegen import full_rewrite, full_rewrite_to_sink
-from tinygrad.codegen.expander import expander
+from tinygrad.codegen.late.expander import expander
 
 simple_pm = PatternMatcher([
   (UPat.cvar('x', dtypes.int), lambda x: UOp.const(dtypes.float, 1.0) + UOp.const(dtypes.float, 2.0)),
@@ -441,18 +441,16 @@ class TestUOpGraph(unittest.TestCase):
       ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(Variable("i", 0, 20)),))
       with self.assertRaises(RuntimeError): to_uops_list([ld0])
 
-  @unittest.skip("outdated")
   def test_in_out_of_bounds_access_gated_store(self):
     with Context(IGNORE_OOB=0):
-      glbl0 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(16), (), 0)
+      glbl0 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(16), src=(), arg=0)
       v = Variable("v", 0, 20)
-      st0 = UOp(Ops.STORE, dtypes.void, (glbl0.index(v), UOp.const(dtypes.int, 0), v<16))
+      st0 = UOp(Ops.STORE, dtypes.void, src=(glbl0.index(v), UOp.const(dtypes.int, 0), UOp(Ops.IF, src=(v<16,))))
       to_uops_list([st0])
 
       st1 = UOp(Ops.STORE, dtypes.void, (glbl0.index(v), v, v<20))
       with self.assertRaises(RuntimeError): to_uops_list([st1])
 
-  @unittest.skip("outdated")
   def test_in_bounds_access_gated_local(self):
     with Context(IGNORE_OOB=0):
       # Define buffers
@@ -465,7 +463,7 @@ class TestUOpGraph(unittest.TestCase):
 
       gate = (gidx<400) & (lidx<8)
 
-      local_store = UOp(Ops.STORE, dtypes.void, (sbuf.index(lidx), UOp.const(dtypes.uint, 1), lidx<8))
+      local_store = UOp(Ops.STORE, dtypes.void, (sbuf.index(lidx), UOp.const(dtypes.uint, 1), UOp(Ops.IF, src=(lidx<8,))))
 
       barrier = UOp(Ops.BARRIER, dtypes.void, (local_store,))
       if_barrier = UOp(Ops.IF, dtypes.void, (gate, barrier))
@@ -477,6 +475,34 @@ class TestUOpGraph(unittest.TestCase):
       global_store = UOp(Ops.STORE, dtypes.void, (gbuf.index(gidx), local_load))
       to_uops_list([global_store])
 
+  def test_load_with_float_in_index(self):
+    with Context(IGNORE_OOB=0):
+      ridx = UOp.range(20, 0)
+      glbl0 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(16), (), 0)
+      i = (ridx.cast(dtypes.float)*0.68).trunc().cast(dtypes.int)
+      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(i, ((0<=i)&(i<16))),))
+      to_uops_list([ld0])
+      glblfloat = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(20), (), 0)
+      ldfloat = UOp(Ops.LOAD, dtypes.float, (glblfloat.index(ridx),))
+      i = (ldfloat+3.14).cast(dtypes.int)
+      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(i, ((0<=i)&(i<16))),))
+
+  def test_load_cast_to_bool(self):
+    with Context(IGNORE_OOB=0):
+      glbl0 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(1), (), 0)
+      ridx = UOp.range(20, 0)
+      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(ridx, ridx.cast(dtypes.bool).logical_not()),))
+      to_uops_list([ld0])
+
+  @unittest.skip("Bool load is not supported yet")
+  def test_load_mask(self):
+    with Context(IGNORE_OOB=0):
+      glbl0 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(16), (), 0)
+      mask = UOp(Ops.DEFINE_GLOBAL, dtypes.bool.ptr(16), (), 0)
+      ridx = UOp.range(20, 0)
+      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(UOp.const(ridx, ridx<16&mask),)))
+      to_uops_list([ld0])
+
   def test_out_of_bounds_off_by_one_access(self):
     with Context(IGNORE_OOB=0):
       glbl0 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(16), (), 0)
@@ -565,10 +591,9 @@ class TestUOpGraph(unittest.TestCase):
 
   def test_switched_range_order(self):
     glbl = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(), (), 0)
-    c2 = UOp.const(dtypes.int, 2)
     cf = UOp.const(dtypes.float, 0.0)
-    r1 = UOp(Ops.RANGE, dtypes.int, (c2,), 0)
-    r2 = UOp(Ops.RANGE, dtypes.int, (c2,), 1)
+    r1 = UOp.range(2, 0)
+    r2 = UOp.range(2, 1)
     alu = UOp(Ops.MUL, dtypes.int, (r2, r1))
     store = UOp(Ops.STORE, dtypes.void, (glbl.index(alu), cf))
     uops = to_uops_list([store])

test/test_uops.py

@@ -22,7 +22,7 @@ def _uops_to_prg(uops_list):
   uops = full_rewrite(ast:=UOp.sink(*uops_list), opts=Device[Device.DEFAULT].renderer)
   src = Device[Device.DEFAULT].renderer.render(uops)
   has_local = Device[Device.DEFAULT].renderer.has_local
-  return CompiledRunner(ProgramSpec("test", src, Device.DEFAULT, ast, uops=uops,
+  return CompiledRunner(ProgramSpec(uops[-1].arg.name if uops[-1].arg is not None else "test", src, Device.DEFAULT, ast, uops=uops,
                                 global_size=[1,1,1] if has_local else None, local_size=[1,1,1] if has_local else None))
 
 def uop(uops:list[UOp], uop:Ops, dtype:Optional[DType], src:tuple[UOp, ...], arg:Any=None) -> UOp:
@@ -177,6 +177,31 @@ class TestBoolUOps(TestUOps):
   def test_cmplt_bool(self): self._test_bop_bool_fxn(Ops.CMPLT, lambda a,b: a < b)
   def test_where_bool(self): self._test_top_bool_fxn(Ops.WHERE, lambda a,b,c: b if a else c)
 
+class TestSafeCast(TestUOps):
+  def test_cast_folds(self):
+    a = UOp.variable("a", 1, 10, dtype=dtypes.int32)
+    self.assertEqual(a.cast(dtypes.int64).cast(dtypes.int32).simplify(), a)
+    self.assertEqual(a.cast(dtypes.double).cast(dtypes.int32).simplify(), a)
+    a = UOp.variable("a", 1, 10, dtype=dtypes.uint8)
+    self.assertEqual(a.cast(dtypes.int64).cast(dtypes.uint8).simplify(), a)
+    self.assertEqual(a.cast(dtypes.uint32).cast(dtypes.uint8).simplify(), a)
+
+  def test_remove_intermediate_cast(self):
+    a = UOp.variable("a", 0., 100., dtype=dtypes.half)
+    self.assertEqual(a.cast(dtypes.double).cast(dtypes.float).simplify(), a.cast(dtypes.float))
+    a = UOp.variable("a", 1, 10, dtype=dtypes.int32)
+    # TODO: double preserves certain int dtypes
+    self.assertEqual(a.cast(dtypes.double).cast(dtypes.float).simplify(), a.cast(dtypes.float))
+    self.assertEqual(a.cast(dtypes.int64).cast(dtypes.int16).simplify(), a.cast(dtypes.int16))
+    a = UOp.variable("a", 1, 10, dtype=dtypes.uint8)
+    self.assertEqual(a.cast(dtypes.int64).cast(dtypes.int32).simplify(), a.cast(dtypes.int32))
+
+  def test_safe_cast_using_bounds(self):
+    a = UOp.variable("a", 1, 10, dtype=dtypes.uint64)
+    self.assertEqual(a.cast(dtypes.int16).cast(dtypes.int).simplify(), a.cast(dtypes.int))
+    a = UOp.variable("a", -10, 10, dtype=dtypes.int32)
+    self.assertEqual(a.cast(dtypes.int8).cast(dtypes.int64).simplify(), a.cast(dtypes.int64))
+
 class TestExecALU(TestUOps):
   def test_sqrt(self):
     self.assertEqual(exec_alu(Ops.SQRT, dtypes.float, (0.0,)), 0.0)
@@ -402,6 +427,14 @@ class TestAssembly(unittest.TestCase):
     self.assertIn(Ops.SHR, ops)
     self.assertNotIn(Ops.IDIV, ops)
 
+  def test_fast_idiv_remove_powers_of_two(self):
+    ridx = UOp.range(2**20, 0)
+    uops = to_uops_list([ridx//(7*64)], opts=Device[Device.DEFAULT].renderer)
+    ops = [x.op for x in uops]
+    # this requires shifting out the powers of two before doing fast_idiv
+    # (((ridx0>>6)*18725)>>17) instead of (int)((((long)(ridx0)*1198373)>>29))
+    self.assertNotIn(Ops.CAST, ops)
+
   def test_mulacc_unrolled(self):
     # test that     acc = acc + a0*b0 + a1*b1 + a2*b2 + a3*b3
     # is not        acc = acc + (a0*b0 + a1*b1 + a2*b2 + a3*b3)
@@ -439,7 +472,7 @@ class TestUOpMethod(unittest.TestCase):
   def test_uop_variables(self):
     a = UOp.variable("a", 1, 10)
     uop_var = Tensor(a.bind(1))
-    st_var = Tensor.empty((2, 1)).reshape((2, a.bind(1)))
+    st_var = Tensor.empty((2, 10))[:, :a.bind(1)]
     _, var_vals = (uop_var+st_var).schedule_with_vars()
     self.assertEqual(len(var_vals), 1)
     self.assertEqual(list(var_vals)[0], a)

test/unit/test_attention.py

@@ -1,5 +1,6 @@
 import unittest
-from tinygrad import Tensor, dtypes
+from tinygrad import Tensor, dtypes, TinyJit, UOp
+from tinygrad.apps.llm import apply_rope
 
 # TODO: test_scheduler, but just in uint
 class TestAttention(unittest.TestCase):
@@ -16,5 +17,29 @@ class TestAttention(unittest.TestCase):
     for si in softmax_inputs:
       assert all(b.dtype == dtypes.half for b in si.bufs), f"non half {si.bufs=}"
 
+  def test_apply_rope(self):
+    x = Tensor.randn(1, 2, 4, 8, dtype=dtypes.float32)
+    result = apply_rope(x, 0)
+    self.assertEqual(result.shape, x.shape)
+    self.assertEqual(result.dtype, x.dtype)
+    self.assertGreater((result - apply_rope(x, 5)).abs().max().item(), 1e-6)
+    with self.assertRaises(AssertionError): apply_rope(Tensor.randn(1, 1, 4, 7, dtype=dtypes.float32), 0)
+
+  def test_apply_rope_jit_prune(self):
+    def rope_fn(x_in, pos): return apply_rope(x_in, pos)
+    rope_noprune = TinyJit(rope_fn)
+    rope_prune = TinyJit(rope_fn, prune=True)
+
+    v_pos = UOp.variable("start_pos", 0, 100)
+    for _ in range(3):
+      rope_noprune(Tensor.randn(1, 2, 4, 8, dtype=dtypes.float32), v_pos.bind(1))
+      rope_prune(Tensor.randn(1, 2, 4, 8, dtype=dtypes.float32), v_pos.bind(1))
+    noprune_size = len(rope_noprune.captured.jit_cache)
+    prune_size = len(rope_prune.captured.jit_cache)
+
+    self.assertGreater(noprune_size, prune_size)
+    self.assertGreaterEqual(noprune_size, 3)
+    self.assertEqual(prune_size, 1)
+
 if __name__ == '__main__':
   unittest.main()

test/unit/test_block_reorder.py

@@ -1,7 +1,7 @@
 import unittest, random
 from tinygrad.dtype import dtypes
 from tinygrad.uop.ops import print_uops, UOp, Ops
-from tinygrad.codegen.linearize import block_reorder
+from tinygrad.codegen.late.linearize import block_reorder
 from tinygrad.renderer.cstyle import OpenCLRenderer
 
 def is_toposorted(lst:list[UOp]):

test/unit/test_dtype.py

@@ -56,6 +56,7 @@ class TestCastConvenienceMethod(unittest.TestCase):
 class TestDtypeTolist(unittest.TestCase):
   def test_bfloat16(self):
     self.assertEqual(Tensor([-60000, 1.5, 3.1, 60000], device="PYTHON", dtype=dtypes.bfloat16).tolist(), [-59904.0, 1.5, 3.09375, 59904.0])
+  def test_fp8(self):
     # 448
     self.assertEqual(Tensor([-30000, 1.5, 3.1, 30000], device="PYTHON", dtype=dtypes.fp8e4m3).tolist(), [-448.0, 1.5, 3.0, 448.0])
     # 57344

test/unit/test_dtype_spec.py

@@ -1,6 +1,6 @@
-import unittest, math, operator, subprocess
+import unittest, math, operator, subprocess, struct
 from tinygrad.tensor import Tensor, dtypes, Device
-from tinygrad.dtype import DType, DTYPES_DICT, truncate, truncate_fp16, truncate_bf16, _to_np_dtype, least_upper_dtype, least_upper_float
+from tinygrad.dtype import DType, DTYPES_DICT, truncate, truncate_fp16, float_to_bf16, _to_np_dtype, least_upper_dtype, least_upper_float
 from tinygrad.device import is_dtype_supported
 from tinygrad.helpers import getenv, CI, DEBUG
 from hypothesis import given, settings, strategies as strat
@@ -26,6 +26,9 @@ def _assert_eq(tensor:Tensor, target_dtype:DType, target, tol_target_dtype:float
   except AssertionError as e:
     raise AssertionError(f"\ntensor {tensor.numpy()} dtype {tensor.dtype} does not match target {target} with dtype {target_dtype}") from e
 
+def u32_to_f32(u): return struct.unpack('f', struct.pack('I', u))[0]
+def f32_to_u32(f): return struct.unpack('I', struct.pack('f', f))[0]
+
 class TestHelpers(unittest.TestCase):
   signed_ints = (dtypes.int8, dtypes.int16, dtypes.int32, dtypes.int64)
   uints = (dtypes.uint8, dtypes.uint16, dtypes.uint32, dtypes.uint64)
@@ -97,23 +100,89 @@ class TestHelpers(unittest.TestCase):
         np.testing.assert_equal(dt.min, False)
         np.testing.assert_equal(dt.max, True)
 
+  def test_dtype_range_vec(self):
+    for dt in core_dtypes:
+      self.assertEqual(dt.min, dt.vec(4).min)
+      self.assertEqual(dt.max, dt.vec(4).max)
+
   def test_truncate_fp16(self):
     self.assertEqual(truncate_fp16(1), 1)
     self.assertEqual(truncate_fp16(65504), 65504)
     self.assertEqual(truncate_fp16(65519.999), 65504)
     self.assertEqual(truncate_fp16(65520), math.inf)
+    self.assertEqual(truncate_fp16(1e-8), 0.0)
+    self.assertEqual(truncate_fp16(-65504), -65504)
+    self.assertEqual(truncate_fp16(-65519.999), -65504)
+    self.assertEqual(truncate_fp16(-65520), -math.inf)
+    self.assertTrue(math.isnan(truncate_fp16(math.nan)))
 
-  def test_truncate_bf16(self):
-    self.assertEqual(truncate_bf16(1), 1)
-    self.assertAlmostEqual(truncate_bf16(1.1), 1.09375, places=7)
-    for a in [1234, 23456, -777.777]:
-      self.assertEqual(truncate_bf16(a), torch.tensor([a], dtype=torch.bfloat16).item())
-    # TODO: torch bfloat 1.1 gives 1.1015625 instead of 1.09375
+  def test_float_to_bf16(self):
+    # TODO: fuzz this better
     max_bf16 = torch.finfo(torch.bfloat16).max
-    self.assertEqual(truncate_bf16(max_bf16), max_bf16)
-    self.assertEqual(truncate_bf16(min_bf16:=-max_bf16), min_bf16)
-    self.assertEqual(truncate_bf16(max_bf16 * 1.00001), math.inf)
-    self.assertEqual(truncate_bf16(min_bf16 * 1.00001), -math.inf)
+    for a in [1, 1.1, 1234, 23456, -777.777, max_bf16, max_bf16 * 1.00001, -max_bf16, -max_bf16 * 1.00001, math.inf, -math.inf]:
+      self.assertEqual(float_to_bf16(a), torch.tensor([a], dtype=torch.bfloat16).item())
+    self.assertTrue(math.isnan(float_to_bf16(math.nan)))
+
+  def test_float_to_bf16_nan(self):
+    # In f32, NaN = exp 0xFF and mantissa ≠ 0. Quiet-vs-signaling is bit 22 of the mantissa: 1 = qNaN, 0 = sNaN.
+    # qNaN(+/-), sNaN(+/-) overflow(+/-)
+    patterns = [0x7FC00001, 0xFFC00001, 0x7F800001, 0xFF800001, 0x7FFFFFFF, 0xFFFFFFFF]
+    for u in patterns:
+      x = u32_to_f32(u)
+      y = float_to_bf16(x)
+      t = torch.tensor([x], dtype=torch.bfloat16).item()
+      self.assertTrue(math.isnan(y))
+      self.assertTrue(math.isnan(t))
+
+  def test_float_to_bf16_round(self):
+    # round_to_nearest_even
+    uppers = [0x3f800000, 0x41230000, 0xC1460000] # 1.0, 10.1875, -12.375
+    for upper in uppers:
+      base = upper & 0xFFFF0000
+      base_f32 = u32_to_f32(base)
+      base_f32_round_up = u32_to_f32(base + 0x00010000)
+
+      # low < 0x8000(0.5ULP) -> round down
+      x = u32_to_f32(base | 0x00007000)
+      self.assertEqual(float_to_bf16(x), base_f32)
+      self.assertEqual(torch.tensor([x], dtype=torch.bfloat16).item(), base_f32)
+
+      # low > 0x8000(0.5ULP) -> round up
+      x = u32_to_f32(base | 0x0000C000)
+      self.assertEqual(float_to_bf16(x), base_f32_round_up)
+      self.assertEqual(torch.tensor([x], dtype=torch.bfloat16).item(), base_f32_round_up)
+
+      # low == 0x8000(0.5ULP) and LSB even -> round down
+      if ((upper >> 16) & 1) == 0:
+        x = u32_to_f32(base | 0x00008000)
+        self.assertEqual(float_to_bf16(x), base_f32)
+        self.assertEqual(torch.tensor([x], dtype=torch.bfloat16).item(), base_f32)
+      # low == 0x8000(0.5ULP) and LSB odd -> round up
+      else:
+        x = u32_to_f32(base | 0x00008000)
+        self.assertEqual(float_to_bf16(x), base_f32_round_up)
+        self.assertEqual(torch.tensor([x], dtype=torch.bfloat16).item(), base_f32_round_up)
+
+  def test_float_to_bf16_boundary(self):
+    # bf16 max finite: exp=0xFE, faction=0x7F => 0x7F7F0000(f32)
+    # bf16 inf(+/-):   exp=0xFF
+    base = 0x7F7F0000
+    inf_u32 = 0x7F800000
+
+    # low < 0.5ULP
+    x = u32_to_f32(base | 0x00007FFF)
+    self.assertEqual(f32_to_u32(float_to_bf16(x)), base)
+    self.assertEqual(f32_to_u32(torch.tensor([x], dtype=torch.bfloat16).item()), base)
+
+    # low > 0.5ULP -> overflows to +inf
+    x = u32_to_f32(base | 0x0000C000)
+    self.assertEqual(f32_to_u32(float_to_bf16(x)), inf_u32)
+    self.assertEqual(f32_to_u32(torch.tensor([x], dtype=torch.bfloat16).item()), inf_u32)
+
+    # low == 0.5ULP and LSB odd -> overflows to +inf
+    x = u32_to_f32(base | 0x00008000)
+    self.assertEqual(f32_to_u32(float_to_bf16(x)), inf_u32)
+    self.assertEqual(f32_to_u32(torch.tensor([x], dtype=torch.bfloat16).item()), inf_u32)
 
   @given(strat.floats(width=32, allow_subnormal=True, allow_nan=True, allow_infinity=True))
   def test_truncate_fp8e4m3(self, x):
@@ -549,4 +618,4 @@ class TestAutoCastType(unittest.TestCase):
     np.testing.assert_allclose(out.numpy(), tt.log_softmax(0).numpy(), rtol=1e-3)
     out = t.log_softmax(0, dtype=dtypes.float)
     self.assertEqual(out.dtype, dtypes.float)
-    np.testing.assert_allclose(out.numpy(), tt.log_softmax(0, dtype=torch.float).numpy(), rtol=1e-3)
+    np.testing.assert_allclose(out.numpy(), tt.log_softmax(0, dtype=torch.float).numpy(), rtol=1e-3)

test/unit/test_gguf.py

@@ -53,11 +53,37 @@ class TestGGUF(unittest.TestCase):
   def test_load_tinyllama_q4_0(self): self._test_gguf_load("https://huggingface.co/ggml-org/models/resolve/main/tinyllamas/stories15M-q4_0.gguf?download=true")
   def test_load_gpt2_q4_1(self): self._test_gguf_load("https://huggingface.co/PrunaAI/gpt2-GGUF-smashed/resolve/main/gpt2.Q4_1.gguf?download=true")
   def test_load_sample_q6_k(self): self._test_gguf_load("https://huggingface.co/Isotr0py/test-gguf-sample/resolve/main/Quant_Q6_K_1024.gguf?download=true")
+  def test_load_sample_mxfp4(self): self._test_gguf_load("https://huggingface.co/ngxson/boring-testing-tiny/resolve/main/stories260K-mxfp4.gguf?download=true")
 
   def test_dequantization_q4_0(self): self._test_dequantization(ggml.GGML_TYPE_Q4_0)
   def test_dequantization_q4_1(self): self._test_dequantization(ggml.GGML_TYPE_Q4_1)
   def test_dequantization_q8_0(self): self._test_dequantization(ggml.GGML_TYPE_Q8_0)
   def test_dequantization_q6_k(self): self._test_dequantization(ggml.GGML_TYPE_Q6_K)
+  def test_dequantization_mxfp4(self):
+    MXFP4 = 39
+
+    def encode(nibbles, E):
+      packed = [(low & 0xF) | ((high & 0xF) << 4) for low, high in zip(nibbles[:16], nibbles[16:])]
+      return np.array([E] + packed, dtype=np.uint8)
+
+    def decode(code, E):
+      sign = -1.0 if code * 0b1000 else 1.0
+      exp = (code >> 1) & 0b11
+      mant = code & 0b1
+      val = (1.0 + 0.5 * mant) * np.exp2(exp - 1) if exp else 0.5 * mant
+      scale = np.exp2(E - 128) if E >= 2 else np.exp2(-127 if E == 1 else -128)
+      return sign * val * scale
+
+    blocks, expected = [], []
+    rng = np.random.default_rng(42)
+    for _ in range(4):
+      E = rng.integers(0, 256)
+      codes = rng.integers(0, 16, size=32, dtype=np.uint8)
+      blocks.append(encode(codes, E))
+      expected.extend(decode(c, E) for c in codes)
+    tensor = Tensor(np.concatenate(blocks))
+    out = ggml_data_to_tensor(tensor, len(expected), MXFP4)
+    self.assertListEqual(out.numpy().tolist(), np.array(expected, dtype=np.float32).tolist())
 
   def test_expected_failure_unknown_type(self):
     with self.assertRaises(ValueError):

test/unit/test_graph_rewrite.py

@@ -65,21 +65,21 @@ class TestFoldingAndReduction(unittest.TestCase):
   def test_full_graph_rewrite_reduction_with_unused_range(self):
     const1 = UOp.const(dtypes.int32, 15)
     const2 = UOp.const(dtypes.int32, 25)
-    rng = UOp.range(dtypes.int32, 10, idx=0)
+    rng = UOp.range(10, idx=0)
     optimized_sink = apply_rewrite((const1 + const2).reduce(Ops.ADD, rng))
     expected_sum = 10 * (15 + 25)
     self.assertEqual(optimized_sink.arg, expected_sum)
 
   @unittest.skip("currently failing")
   def test_full_graph_rewrite_range_reduction(self):
-    simple_range = UOp.range(dtypes.int32, 5, idx=0)
+    simple_range = UOp.range(5, idx=0)
     optimized_sink = apply_rewrite(simple_range.reduce(Ops.ADD, simple_range))
     expected_sum = sum(range(5))
     self.assertEqual(optimized_sink.arg, expected_sum)
 
   @unittest.skip("currently failing")
   def test_full_graph_rewrite_simple_reduction_folding(self):
-    simple_range = UOp.range(dtypes.int32, 4, idx=0)
+    simple_range = UOp.range(4, idx=0)
     add_uop = simple_range + UOp.const(dtypes.int32, 1)
     optimized_sink = apply_rewrite(add_uop.reduce(Ops.ADD, simple_range))
     expected_sum = sum(i + 1 for i in range(4))
@@ -87,8 +87,8 @@ class TestFoldingAndReduction(unittest.TestCase):
 
   @unittest.skip("currently failing")
   def test_full_graph_rewrite_nested_loop_collapse(self):
-    outer_range = UOp.range(dtypes.int32, 8, 0)
-    inner_range = UOp.range(dtypes.int32, 4, 1)
+    outer_range = UOp.range(8, 0)
+    inner_range = UOp.range(4, 1)
     expr = (outer_range * 10) + inner_range
     optimized_reduce_uop = apply_rewrite(expr.reduce(Ops.ADD, outer_range, inner_range))
     self.assertEqual(optimized_reduce_uop.op, Ops.CONST)

test/unit/test_shapetracker.py

@@ -6,7 +6,7 @@ from tinygrad.helpers import prod
 from tinygrad.shape.shapetracker import ShapeTracker, View
 from tinygrad import Variable
 from tinygrad.uop.ops import UOp, Ops, graph_rewrite
-from tinygrad.codegen.devectorizer import sym
+from tinygrad.codegen.late.devectorizer import sym
 from itertools import product
 
 def shapetracker_getitem(st:ShapeTracker, val:int):
@@ -839,25 +839,22 @@ class TestRender(unittest.TestCase):
     self.assertEqual(idx.render(), "((ridx0*3)+ridx1)")
     self.assertEqual(valid.render(), "(ridx0<2)")
 
-class TestVariableReshape(unittest.TestCase):
-  def test_reshape(self):
-    st = ShapeTracker.from_shape((3,))
-    st = st.reshape((Variable("i", 1, 10),))
+class TestVariableShrink(unittest.TestCase):
+  def test_shrink(self):
+    st = ShapeTracker.from_shape((10,))
+    st = st.shrink(((0, Variable("i", 1, 10)),))
     assert len(st.views) == 1
 
-  def test_reshape_stride_0(self):
-    st = ShapeTracker.from_shape((3,), (0,))
-    st = st.reshape((Variable("i", 1, 10).bind(3),))
-    assert len(st.views) == 1, f"multiview {st}"
-
-  def test_reshape_bound(self):
-    st = ShapeTracker.from_shape((3,))
-    st = st.reshape((Variable("i", 1, 10).bind(3),))
+  def test_shrink_bound(self):
+    st = ShapeTracker.from_shape((10,))
+    st = st.shrink(((0, Variable("i", 1, 10).bind(3)),))
     assert len(st.views) == 1
 
-  def test_add(self):
-    st1 = ShapeTracker.from_shape((3,))
-    st2 = ShapeTracker.from_shape((Variable("i", 1, 10),))
+class TestVariableMerge(unittest.TestCase):
+  def test_add_reshape(self):
+    vi = Variable("i", 1, 10)
+    st1 = ShapeTracker.from_shape((vi,))
+    st2 = ShapeTracker.from_shape((1, vi,))
     st = st1+st2
     assert len(st.views) == 1
 
@@ -867,15 +864,17 @@ class TestVariableReshape(unittest.TestCase):
     st = st1+st2
     assert len(st.views) == 1, f"multiview {st}"
 
-  def test_add_bound(self):
-    st1 = ShapeTracker.from_shape((3,))
-    st2 = ShapeTracker.from_shape((Variable("i", 1, 10).bind(3),))
+  def test_add_reshape_bound(self):
+    vi = Variable("i", 1, 10).bind(3)
+    st1 = ShapeTracker.from_shape((vi,))
+    st2 = ShapeTracker.from_shape((1, vi,))
     st = st1+st2
     assert len(st.views) == 1
 
   def test_simplify(self):
-    st1 = ShapeTracker.from_shape((3,))
-    st2 = ShapeTracker.from_shape((Variable("i", 1, 10).bind(3),))
+    vi = Variable("i", 1, 10).bind(3)
+    st1 = ShapeTracker.from_shape((vi,))
+    st2 = ShapeTracker.from_shape((1, vi,))
     st = ShapeTracker((st1.views[0], st2.views[0]))
     st = st.simplify()
     assert len(st.views) == 1

test/unit/test_shapetracker_math.py

@@ -87,20 +87,6 @@ class TestShapeTrackerAdd(unittest.TestCase):
     assert not (st_equal(st1, st2))
 
 class TestShapeTrackerAddVariable(unittest.TestCase):
-  def test_self_add(self):
-    j = Variable("j", 0, 20).bind(10)
-    a = ShapeTracker.from_shape((10,10))
-    x = a.reshape((10, j))
-    out = x + x
-    assert out == x
-
-  def test_self_add_reshape(self):
-    j = Variable("j", 0, 20).bind(10)
-    a = ShapeTracker.from_shape((10,10))
-    x = a.reshape((10, j))
-    out = x.reshape((5, 2, j)) + x
-    assert out == x
-
   def test_merge_symbolic_views(self):
     var_i = Variable('i', 1, 10)
     var_j = Variable('i', 1, 10)

test/unit/test_simplify_valid_idx.py

@@ -19,7 +19,7 @@ def get_load_image_uop(image_shape:tuple[int, ...], valid:UOp, idx:tuple[UOp, UO
 
 def Special(expr, nmax): return UOp(Ops.SPECIAL, dtypes.int, (), (expr, nmax))
 def Variable(expr, nmin, nmax): return UOp.variable(expr, nmin, nmax)
-def Range(n, nmax): return UOp(Ops.RANGE, dtypes.int, arg=n, src=(UOp.const(dtypes.int, nmax),))
+def Range(n, nmax): return UOp.range(nmax, n)
 
 class TestHelpers(unittest.TestCase):
   def test_is_increasing(self):

test/unit/test_symbolic_shapetracker.py

@@ -48,11 +48,11 @@ class TestSymbolic(unittest.TestCase):
     i = Variable("i", 1, 5).bind(3)
     j = Variable("j", 1, 5).bind(3)
     k = Variable("k", 1, 5).bind(3)
-    t = Tensor.rand(3, 4).reshape(i, 4).cat(Tensor.rand(3, 4).reshape(j, 4), dim=0).cat(Tensor.rand(3, 4).reshape(k, 4), dim=0)
+    t = Tensor.rand(5, 4)[:i].cat(Tensor.rand(5, 4)[:j], dim=0).cat(Tensor.rand(5, 4)[:k], dim=0)
     st = t.uop.st
     self.assert_tuple_equal(st.shape, (i+j+k, 4))
     assert st.real_strides() == (4, 1)
-    t = Tensor.rand(3, 3).reshape(i, 3).cat(Tensor.rand(3, 3).reshape(i, 3), dim=0).cat(Tensor.rand(3, 3), dim=0)
+    t = Tensor.rand(5, 3)[:i].cat(Tensor.rand(5, 3)[:i], dim=0).cat(Tensor.rand(3, 3), dim=0)
     st = t.uop.st
     self.assert_tuple_equal(st.shape, (2*i+3, 3))
     assert st.real_strides() == (3, 1)
@@ -61,7 +61,7 @@ class TestSymbolic(unittest.TestCase):
     i = Variable("i", 1, 5).bind(4)
     j = Variable("j", 1, 5).bind(4)
     k = Variable("k", 1, 5).bind(4)
-    t = Tensor.rand(3, 4).reshape(3, i).cat(Tensor.rand(3, 4).reshape(3, j), dim=1).cat(Tensor.rand(3, 4).reshape(3, k), dim=1)
+    t = Tensor.rand(3, 5)[:, :i].cat(Tensor.rand(3, 5)[:, :j], dim=1).cat(Tensor.rand(3, 5)[:, :k], dim=1)
     st = t.uop.st
     self.assert_tuple_equal(st.shape, (3, i+j+k))
     self.assert_tuple_equal(st.real_strides(), (i+j+k, 1))
@@ -109,60 +109,44 @@ class TestShapeTrackerUnbind(unittest.TestCase):
     assert unbound_view == View.create(shape=(v, 4))
     assert var_val == {v: 3}
 
-  def test_reshape_unbind(self):
-    v = Variable("v", 1, 100)
-    bv = Variable("v", 1, 100).bind(3)
-    t = Tensor.rand(3, 4).reshape(bv, 4)
-    unbound_st, var_val = t.uop.st.unbind()
-    assert unbound_st == ShapeTracker((View.create(shape=(v, 4)),))
-    assert var_val == {v: 3}
-
   def test_shrink_unbind(self):
     v = Variable("v", 1, 100)
     bv = Variable("v", 1, 100).bind(2)
+    t = Tensor.rand(3, 4).shrink(((0,bv),(0,4)))
+    unbound_st, var_val = t.uop.st.unbind()
+    assert unbound_st == ShapeTracker((View.create(shape=(v, 4)),))
+    assert var_val == {v: 2}
     t = Tensor.rand(3, 4).shrink(((bv, bv+1), (0, 4)))
     unbound_st, var_val = t.uop.st.unbind()
     assert unbound_st == ShapeTracker((View.create(shape=(1, 4), offset=4*v),))
     assert var_val == {v: 2}
 
-class TestSymbolicReshapeFromContiguous(unittest.TestCase):
-  def test_reshape_into_symbols_simple(self):
+class TestSymbolicReshape(unittest.TestCase):
+  def test_reshape(self):
+    a = Tensor.rand(5, 4)
+    b = Tensor.rand(5, 6)
     for i in range(1, 6):
       vi = Variable("i", 1, 5).bind(i)
-      t = Tensor.rand(i, 4).reshape(vi, 4)
-      assert t.shape == (vi, 4)
-      t = Tensor.rand(i, 6).reshape(vi, 2, 3)
-      assert t.shape == (vi, 2, 3)
-
-  def test_reshape_symbols_reshape_ints(self):
-    for i in range(1, 6):
-      vi = Variable("i", 1, 5).bind(i)
-      t = Tensor.rand(i, 4).reshape(vi, 4)
-      assert t.shape == (vi, 4)
-      t = t.reshape(i, 4)
-      assert t.shape == (i, 4)
-
-  @unittest.skip("works now")
-  def test_reshape_into_symbols_bad_shape(self):
-    vi = Variable("i", 1, 10).bind(4)
-    # TODO: this never actually worked, it relied on lazy
-    #with self.assertRaises(ValueError):
-    #  Tensor.rand(4, 6).reshape(vi, 6).reshape(1, 77) # reshape to a different size new shape through symbolic shape
-    with self.assertRaises(AssertionError):
-      Tensor.rand(3, 4).reshape(3, (vi+1)) # reshape into non-Variable Node
+      ret = a[:vi]
+      ret = ret.reshape((vi, 4))
+      assert ret.shape == (vi, 4)
+      ret = b[:vi]
+      ret = ret.reshape((vi, 2, 3))
+      assert ret.shape == (vi, 2, 3)
 
   def test_two_symbol_reshape(self):
+    t = Tensor.rand(5, 5)
     for i in range(1, 6):
       for j in range(1, 6):
         vi = Variable("i", 1, 5).bind(i)
         vj = Variable("j", 1, 5).bind(j)
-        t = Tensor.rand(i, j).reshape(vi, vj)
-        assert t.shape == (vi, vj)
-        # NOTE: this is currently not allowed
-        # t = t.reshape(1, vi*vj)
-        # assert t.shape == (1, vi*vj)
-        t = t.reshape(vj, vi)
-        assert t.shape == (vj, vi)
+        ret = t[:vi, :vj]
+        ret = ret.reshape(vj, vi)
+        assert ret.shape == (vj, vi)
+        ret = ret.reshape(vi, vj)
+        assert ret.shape == (vi, vj)
+        ret = ret.reshape(1, vi*vj)
+        assert ret.shape == (1, vi*vj)
 
   def test_symbolic_mask(self):
     # taken from gpt2 single kvcache
@@ -175,41 +159,6 @@ class TestSymbolicReshapeFromContiguous(unittest.TestCase):
     new_shape = (2, (Variable('start_pos', 1, 128)+1), 16, 64)
     assert view.reshape(new_shape) is None
 
-class TestSymbolicReshapeFromNonContiguous(unittest.TestCase):
-  def test_reshape_from_const(self):
-    vi = Variable("i", 1, 5).bind(4)
-    t = Tensor.ones(3, 4).reshape(3, vi)
-    assert t.shape == (3, vi)
-    assert not t.uop.st.contiguous
-    assert len(t.uop.st.views) == 1
-
-  def test_reshape_not_allowed(self):
-    vi = Variable("i", 1, 5).bind(4)
-    with self.assertRaises(ValueError):
-      # different shape length  # TODO: cases where contractions matched might be fine
-      Tensor.ones(3, 4, 1).reshape(3, vi)
-    with self.assertRaises(ValueError):
-      # size matched, but dimensions do not match
-      Tensor.ones(4, 3).reshape(3, vi)
-
-  def test_reshape_from_padded(self):
-    vi = Variable("i", 1, 5).bind(4)
-    t = Tensor.ones(3, 4).contiguous().expand(2, 3, 4).pad(((1, 1), None, None)).shrink((None, None, (1, 3)))
-    st = t.uop.st
-    assert len(st.views) == 1
-    view = st.views[0]
-    assert view.shape == (4, 3, 2)
-    t = t.reshape(vi, 3, 2)
-    st2 = t.uop.st
-    assert len(st2.views) == 1
-    view2 = st2.views[0]
-    # check only shape changed. strides, offset, mask, contiguous remained the same
-    assert view2.shape == (vi, 3, 2)
-    assert view.strides == view2.strides == (0, 4, 1)
-    assert view.offset == view2.offset == 1
-    assert view.mask == view2.mask == ((1, 3), (0, 3), (0, 2))
-    assert not view.contiguous and not view2.contiguous
-
 class TestSymbolicExpand(unittest.TestCase):
   def test_expand_into_symbols(self):
     vi = Variable("i", 1, 5).bind(3)
@@ -220,11 +169,12 @@ class TestSymbolicExpand(unittest.TestCase):
     assert a.shape == (3, vi, vj)
 
   def test_plus_expands_constant(self):
+    a = Tensor.rand(3, 5)
     for i in range(1, 6):
       vi = Variable("i", 1, 5).bind(i)
-      a = Tensor.rand(3, i).reshape(3, vi)
-      a = a + 1
-      self.assertTupleEqual(a.shape, (3, vi))
+      ret = a[:, :vi]
+      ret = ret + 1
+      self.assertTupleEqual(ret.shape, (3, vi))
 
   def test_pad_then_expand_into_symbols(self):
     vi = Variable("i", 1, 10).bind(3)
@@ -234,6 +184,11 @@ class TestSymbolicExpand(unittest.TestCase):
     self.assertEqual(a.reshape(vi*25).shape, (vi*25,))
 
 class TestSymbolicShrink(unittest.TestCase):
+  def test_shrink_symbols_simple(self):
+    vi = Variable("i", 1, 5)
+    t = Tensor.rand(5, 5).shrink(((0, 5),(0,vi)))
+    assert t.shape == (5, vi)
+
   def test_shrink_symbols(self):
     vi = Variable("i", 1, 5)
     t = Tensor.rand(3, 5).shrink(((0, 2), (vi, vi+1)))
@@ -242,10 +197,10 @@ class TestSymbolicShrink(unittest.TestCase):
 class TestSymbolicPad(unittest.TestCase):
   def test_pad(self):
     v = Variable("v", 1, 100).bind(5)
-    t = Tensor.ones(5).reshape(v).pad(((4, 0),)).reshape(9)
-    assert t.shape == (9,)
-    st = t.uop.st
-    print(st)
+    t = Tensor.ones(100)[:v].pad(((4, 0),))
+    t = t.reshape(9)
+    assert t.tolist() == [0,0,0,0,1,1,1,1,1]
+
 
 if __name__ == '__main__':
   unittest.main()

test/unit/test_tensor_uop_representation.py

@@ -97,7 +97,7 @@ class TestTensorUopRepresentation(unittest.TestCase):
     is_pattern(a, UPat(Ops.RESHAPE, src=(UPat(Ops.BUFFER),)))
     vi = UOp.variable("i", 1, 3).bind(1)
     a = Tensor.empty(3, vi)
-    is_pattern(a, UPat(Ops.RESHAPE, src=(UPat(Ops.BUFFER),)))
+    is_pattern(a, UPat(Ops.RESHAPE, src=(UPat(Ops.SHRINK, src=(UPat(Ops.BUFFER),))),))
     self.assertEqual(a.uop.base.buffer.size, 9)
 
 if __name__ == '__main__':

test/unit/test_uop_spec.py

@@ -81,5 +81,16 @@ class TestUOpSpec(unittest.TestCase):
     with self.assertRaisesRegex(RuntimeError, "UOp verification failed"):
       type_verify([a], tensor_uop_spec)
 
+class TestUOpSink(unittest.TestCase):
+  def test_0(self):
+    s = UOp.sink()
+    self.assertEqual(len(s.src), 0)
+
+  def test_1(self):
+    a = UOp.const(dtypes.int, 0)
+    s1 = UOp.sink(a)
+    s2 = a.sink()
+    self.assertIs(s1, s2)
+
 if __name__ == '__main__':
   unittest.main()

test/unit/test_uop_symbolic.py

@@ -4,11 +4,11 @@ import z3
 
 from tinygrad.dtype import dtypes, ConstType
 from tinygrad.codegen import full_rewrite
-from tinygrad.codegen.devectorizer import sym
+from tinygrad.codegen.late.devectorizer import sym
 from tinygrad.helpers import Context
 from tinygrad.uop.ops import UOp, Ops, graph_rewrite, sym_infer
 from tinygrad import Variable
-from tinygrad.uop.spec import z3_renderer
+from tinygrad.uop.spec import uops_to_z3
 
 def render(self) -> tuple[str, ConstType, ConstType]:
   # NOTE: we need STORE so the ALU op has children
@@ -32,9 +32,8 @@ class TestSymbolic(unittest.TestCase):
   def helper_test_variable(self, v, n, m, s, test_z3:bool=True):
     if test_z3:
       solver = z3.Solver()
-      z3_sink = graph_rewrite(v.sink(v.simplify()), z3_renderer, ctx=(solver, {}))
-      expr, epxr_simplified = z3_sink.src[0].arg, z3_sink.src[1].arg
-      self.assertEqual(solver.check(expr != epxr_simplified), z3.unsat, "simplified expression not equal to original")
+      expr, expr_simplified = uops_to_z3(solver, v, v.simplify())
+      self.assertEqual(solver.check(expr != expr_simplified), z3.unsat, "simplified expression not equal to original")
     rendered, nmin, nmax = render(v)
     if isinstance(s, tuple): self.assertIn(rendered, s)
     else: self.assertEqual(rendered, s)
@@ -128,6 +127,8 @@ class TestSymbolic(unittest.TestCase):
     b = Variable("b", 0, 8)
     self.helper_test_variable(a+a, 0, 16, "(a*2)")
     self.helper_test_variable((a+b)+b, 0, 24, "(a+(b*2))")
+    self.helper_test_variable((a*3+b)+a, 0, 40, "(b+(a*4))")
+    self.helper_test_variable((a+b)+a*3, 0, 40, "(b+(a*4))")
 
   def test_sub_self(self):
     a = Variable("a", 0, 8)
@@ -162,10 +163,6 @@ class TestSymbolic(unittest.TestCase):
   def test_div_remove(self):
     self.helper_test_variable(Variable("a", 0, 7) // 20, 0, 0, "0")
 
-  def test_div_min_max(self):
-    self.helper_test_variable(Variable("a", 1, 7) // 2, 0, 3, "(a//2)")
-    self.helper_test_variable(Variable("a", 0, 6) // 2, 0, 3, "(a//2)")
-
   def test_div_neg_min_max(self):
     self.helper_test_variable(Variable("a", 1, 7) // -2, -3, 0, "((a//2)*-1)")
     self.helper_test_variable(Variable("a", 0, 6) // -2, -3, 0, "((a//2)*-1)")
@@ -211,6 +208,28 @@ class TestSymbolic(unittest.TestCase):
     self.assertEqual((Variable("x", -10, 0)%Variable("y", -10, -1))._min_max, (-9, 0))
     self.assertEqual((Variable("x", -10, 0)%Variable("y", 1, 10))._min_max, (-9, 0))
 
+  def test_range_div_its_symbolic_bound(self):
+    a = Variable("a", 1, 10)
+    ridx0 = UOp.range(a+2, 0)
+    self.helper_test_variable(ridx0//(a+2), 0, 0, "0")
+
+  def test_range_mod_its_symbolic_bound(self):
+    a = Variable("a", 1, 10)
+    ridx = UOp.range(a+2, 0)
+    self.helper_test_variable(ridx%(a+2), 0, 11, "ridx0")
+
+  def test_div_min_max(self):
+    self.helper_test_variable(Variable("a", 2, 7) // 2, 1, 3, "(a//2)")
+    self.helper_test_variable(Variable("a", 0, 6) // 2, 0, 3, "(a//2)")
+
+    self.helper_test_variable(Variable("x", 0, 10)//Variable("y", 1, 10), 0, 10, "(x//y)")
+    self.helper_test_variable(Variable("x", -10, 0)//Variable("y", 1, 10), -10, 0, "(((x*-1)//y)*-1)")
+    self.helper_test_variable(Variable("x", 0, 10)//Variable("y", -10, -1), -10, 0, "((x//(y*-1))*-1)")
+    self.helper_test_variable(Variable("x", -10, 0)//Variable("y", -10, -1), 0, 10, "((x*-1)//(y*-1))")
+
+    self.helper_test_variable(Variable("x", -10, 10)//Variable("y", 1, 10), -10, 10, "(x//y)")
+    self.helper_test_variable(Variable("x", -10, 10)//Variable("y", -10, -1), -10, 10, "((x//(y*-1))*-1)")
+
   def test_mod_factor(self):
     self.helper_test_variable(usum([Variable("a", 0, 7)*100, Variable("b", 0, 3)*50]) % 100, 0, 50, "((b%2)*50)")
 
@@ -440,7 +459,8 @@ class TestSymbolic(unittest.TestCase):
     self.helper_test_variable((-Variable("a", 10, 10))%7, -3, -3, "-3")
 
   def test_div_numerator_negative(self):
-    self.helper_test_variable((Variable("idx", 0, 9)*-10)//11, -8, 0, "(((idx*10)//11)*-1)")
+    with Context(CORRECT_DIVMOD_FOLDING=1):
+      self.helper_test_variable((Variable("idx", 0, 9)*-10)//11, -8, 0, "(((idx*10)//11)*-1)")
 
   def test_nest_div_negative_factor(self):
     ridx0=UOp.variable("ridx0", 0, 9)
@@ -629,15 +649,16 @@ class TestSymbolic(unittest.TestCase):
     cond = Variable("x", 0, 3) < 2
     a = Variable("a", 0, 3)
     b = Variable("b", 0, 3)
+    c = Variable("c", 0, 3)
     aa = cond.where(a, a.ufix(0))
     bb = cond.where(b, b.ufix(1))
     self.helper_test_variable(aa, 0, 3, "(a if (x<2) else 0)")
     self.helper_test_variable(bb, 0, 3, "(b if (x<2) else 1)")
     self.helper_test_variable(aa+bb, 0, 6, "((a+b) if (x<2) else 1)")
     self.helper_test_variable(aa.maximum(bb), 0, 3, "(max(a, b) if (x<2) else 1)")
+    self.helper_test_variable((c+aa)+bb, 0, 9, "(c+((a+b) if (x<2) else 1))")
 
     # not combining because it increased total ALU
-    c = Variable("c", 0, 3)
     cc = cond.where(c, c+1)
     self.helper_test_variable(bb+cc, 0, 7, "((b if (x<2) else 1)+(c if (x<2) else (c+1)))")
 

test/unit/test_uop_vmin_vmax.py

@@ -60,7 +60,7 @@ class TestVminVmaxProperties(unittest.TestCase):
   def test_vmin_vmax_variable_inside_special(self):
     uop = UOp(Ops.SPECIAL, dtypes.int, arg=('gidx0', UOp(Ops.DEFINE_VAR, dtypes.int, arg=('i', 1, 10))))
     self.assertEqual(uop.vmin, 0)
-    self.assertEqual(uop.vmax, 10)
+    self.assertEqual(uop.vmax, 9)
 
   def test_vmin_vmax_multiplication_0_inf(self):
     # vmin and vmax for multiplication with a variable

test/unit/test_viz.py

@@ -1,11 +1,11 @@
-import unittest, decimal, json
+import unittest, decimal, json, struct
 from dataclasses import dataclass
 
 from tinygrad.uop.ops import UOp, UPat, Ops, PatternMatcher, TrackedPatternMatcher
 from tinygrad.uop.ops import graph_rewrite, track_rewrites, TRACK_MATCH_STATS
 from tinygrad.uop.symbolic import sym
 from tinygrad.dtype import dtypes
-from tinygrad.helpers import PROFILE, colored, ansistrip, flatten, TracingKey, ProfileRangeEvent, Context
+from tinygrad.helpers import PROFILE, colored, ansistrip, flatten, TracingKey, ProfileRangeEvent, ProfileEvent, Context
 from tinygrad.device import Buffer
 
 @track_rewrites(name=True)
@@ -16,10 +16,9 @@ def exec_rewrite(sink:UOp, pm_lst:list[PatternMatcher], names:None|list[str]=Non
 
 # real VIZ=1 pickles these tracked values
 from tinygrad.uop.ops import tracked_keys, tracked_ctxs, uop_fields, active_rewrites, _name_cnt
-from tinygrad.viz import serve
-serve.contexts = (tracked_keys, tracked_ctxs, uop_fields)
+traces = [(tracked_keys, tracked_ctxs, uop_fields)]
 from tinygrad.viz.serve import get_metadata, uop_to_json, get_details
-def get_viz_list(): return get_metadata(tracked_keys, tracked_ctxs)
+def get_viz_list(): return get_metadata(traces)
 
 class BaseTestViz(unittest.TestCase):
   def setUp(self):
@@ -142,6 +141,8 @@ class TestViz(BaseTestViz):
     z = UOp.const(dtypes.int, 0)
     alu = a*z
     exec_rewrite(alu, [sym])
+    lst = get_viz_list()
+    self.assertEqual(len(lst), 1)
     graphs = [x["graph"] for x in get_details(tracked_ctxs[0][0])]
     # embed const in the parent node when possible
     self.assertEqual(list(graphs[0]), [id(a), id(alu)])
@@ -252,31 +253,74 @@ class TestVizIntegration(BaseTestViz):
 from tinygrad.device import ProfileDeviceEvent, ProfileGraphEvent, ProfileGraphEntry
 from tinygrad.viz.serve import get_profile
 
+class TinyUnpacker:
+  def __init__(self, buf): self.buf, self.offset = buf, 0
+  def __call__(self, fmt:str) -> tuple:
+    ret = struct.unpack_from(fmt, self.buf, self.offset)
+    self.offset += struct.calcsize(fmt)
+    return ret
+
+# 0 means None, otherwise it's an enum value
+def option(i:int) -> int|None: return None if i == 0 else i-1
+
+def load_profile(lst:list[ProfileEvent]) -> dict:
+  ret = get_profile(lst)
+  u = TinyUnpacker(ret)
+  total_dur, global_peak, index_len, layout_len = u("<IQII")
+  strings, dtypes = json.loads(ret[u.offset:u.offset+index_len]).values()
+  u.offset += index_len
+  layout:dict[str, dict] = {}
+  for _ in range(layout_len):
+    klen = u("<B")[0]
+    k = ret[u.offset:u.offset+klen].decode()
+    u.offset += klen
+    layout[k] = v = {"events":[]}
+    event_type, event_count = u("<BI")
+    if event_type == 0:
+      for _ in range(event_count):
+        name, ref, st, dur, _ = u("<IIIfI")
+        v["events"].append({"name":strings[name], "ref":option(ref), "st":st, "dur":dur})
+    else:
+      v["peak"] = u("<Q")[0]
+      for _ in range(event_count):
+        alloc, ts, key = u("<BII")
+        if alloc: v["events"].append({"event":"alloc", "ts":ts, "key":key, "arg": {"dtype":strings[u("<I")[0]], "sz":u("<Q")[0]}})
+        else: v["events"].append({"event":"free", "ts":ts, "key":key})
+  return {"dur":total_dur, "peak":global_peak, "layout":layout}
+
 class TestVizProfiler(unittest.TestCase):
   def test_perfetto_node(self):
     prof = [ProfileRangeEvent(device='NV', name='E_2', st=decimal.Decimal(1000), en=decimal.Decimal(1010), is_copy=False),
             ProfileDeviceEvent(device='NV', comp_tdiff=decimal.Decimal(-1000), copy_tdiff=decimal.Decimal(-100))]
 
-    j = json.loads(get_profile(prof))
+    j = load_profile(prof)
 
-    dev_events = j['layout']['NV']['shapes']
+    dev_events = j['layout']['NV']['events']
     self.assertEqual(len(dev_events), 1)
     event = dev_events[0]
     self.assertEqual(event['name'], 'E_2')
     self.assertEqual(event['st'], 0)
     self.assertEqual(event['dur'], 10)
+    assert event['ref'] is None
 
   def test_perfetto_copy_node(self):
     prof = [ProfileRangeEvent(device='NV', name='COPYxx', st=decimal.Decimal(1000), en=decimal.Decimal(1010), is_copy=True),
-            ProfileDeviceEvent(device='NV', comp_tdiff=decimal.Decimal(-1000), copy_tdiff=decimal.Decimal(-100))]
+            ProfileRangeEvent(device='NV:2', name='COPYxx', st=decimal.Decimal(1000), en=decimal.Decimal(1010), is_copy=True),
+            ProfileDeviceEvent(device='NV', comp_tdiff=decimal.Decimal(-1000), copy_tdiff=decimal.Decimal(-100)),
+            ProfileDeviceEvent(device='NV:2', comp_tdiff=decimal.Decimal(-800), copy_tdiff=decimal.Decimal(-80))]
 
-    j = json.loads(get_profile(prof))
+    j = load_profile(prof)
 
-    event = j['layout']['NV']['shapes'][0]
+    event = j['layout']['NV']['events'][0]
     self.assertEqual(event['name'], 'COPYxx')
-    self.assertEqual(event['st'], 900) # diff clock
+    self.assertEqual(event['st'], 0)   # first event
     self.assertEqual(event['dur'], 10)
 
+    event2 = j['layout']['NV:2']['events'][0]
+    self.assertEqual(event2['st'], 20) # second event, diff clock
+
+    self.assertEqual(j["dur"], (event2["st"]+event2["dur"])-event["st"])
+
   def test_perfetto_graph(self):
     prof = [ProfileDeviceEvent(device='NV', comp_tdiff=decimal.Decimal(-1000), copy_tdiff=decimal.Decimal(-100)),
             ProfileDeviceEvent(device='NV:1', comp_tdiff=decimal.Decimal(-500), copy_tdiff=decimal.Decimal(-50)),
@@ -285,28 +329,44 @@ class TestVizProfiler(unittest.TestCase):
                               deps=[[], [0]],
                               sigs=[decimal.Decimal(1000), decimal.Decimal(1002), decimal.Decimal(1004), decimal.Decimal(1008)])]
 
-    j = json.loads(get_profile(prof))
+    j = load_profile(prof)
 
     tracks = list(j['layout'])
     self.assertEqual(tracks[0], 'NV Graph')
-    self.assertEqual(tracks[2], 'NV')
-    self.assertEqual(tracks[4], 'NV:1')
+    self.assertEqual(tracks[1], 'NV')
+    self.assertEqual(tracks[2], 'NV:1')
 
-    nv_events = j['layout']['NV']['shapes']
+    nv_events = j['layout']['NV']['events']
     self.assertEqual(nv_events[0]['name'], 'E_25_4n2')
     self.assertEqual(nv_events[0]['st'], 0)
     self.assertEqual(nv_events[0]['dur'], 2)
     #self.assertEqual(j['devEvents'][6]['pid'], j['devEvents'][0]['pid'])
 
-    nv1_events = j['layout']['NV:1']['shapes']
+    nv1_events = j['layout']['NV:1']['events']
     self.assertEqual(nv1_events[0]['name'], 'NV -> NV:1')
     self.assertEqual(nv1_events[0]['st'], 954)
     #self.assertEqual(j['devEvents'][7]['pid'], j['devEvents'][3]['pid'])
 
-    graph_events = j['layout']['NV Graph']['shapes']
+    graph_events = j['layout']['NV Graph']['events']
     self.assertEqual(graph_events[0]['st'], nv_events[0]['st'])
     self.assertEqual(graph_events[0]['st']+graph_events[0]['dur'], nv1_events[0]['st']+nv1_events[0]['dur'])
 
+  def test_bytes_per_kernel(self):
+    step = 10
+    n_events = 1_000
+    prof = [ProfileRangeEvent("CPU", name="k_test", st=decimal.Decimal(ts:=i*step), en=decimal.Decimal(ts)+step) for i in range(n_events)]
+    sz = len(get_profile(prof))
+    self.assertLessEqual(sz/n_events, 26)
+
+  # can pack up to 1hr 11 min of trace events
+  def test_trace_duration(self):
+    dur_mins = 72
+    n_events = 1_000
+    step = decimal.Decimal(dur_mins*60*1e6//n_events)
+    prof = [ProfileRangeEvent("CPU", name="k_test", st=decimal.Decimal(ts:=i*step), en=decimal.Decimal(ts)+step) for i in range(n_events)]
+    with self.assertRaises(struct.error):
+      get_profile(prof)
+
 def _alloc(b:int):
   a = Tensor.empty(b, device="NULL", dtype=dtypes.char)
   a.uop.buffer.allocate()
@@ -316,38 +376,29 @@ class TestVizMemoryLayout(BaseTestViz):
   def test_double_alloc(self):
     a = _alloc(1)
     _b = _alloc(1)
-    profile_ret = json.loads(get_profile(Buffer.profile_events))
+    profile_ret = load_profile(Buffer.profile_events)
     ret = profile_ret["layout"][f"{a.device} Memory"]
     self.assertEqual(ret["peak"], 2)
-    self.assertEqual(ret["shapes"][0]["x"], [0, 2])
-    self.assertEqual(ret["shapes"][1]["x"], [1, 2])
+    self.assertEqual(len(ret["events"]), 2)
 
   def test_del_once(self):
     a = _alloc(1)
     del a
     b = _alloc(1)
-    profile_ret = json.loads(get_profile(Buffer.profile_events))
+    profile_ret = load_profile(Buffer.profile_events)
     ret = profile_ret["layout"][f"{b.device} Memory"]
     self.assertEqual(ret["peak"], 1)
-    self.assertEqual(ret["shapes"][0]["x"], [0, 2])
-    self.assertEqual(ret["shapes"][1]["x"], [2, 3])
-    self.assertEqual(ret["shapes"][0]["y"], [0, 0])
-    self.assertEqual(ret["shapes"][1]["y"], [0, 0])
+    self.assertEqual(len(ret["events"]), 3)
 
   def test_alloc_free(self):
     a = _alloc(1)
     _b = _alloc(1)
     del a
     c = _alloc(1)
-    profile_ret = json.loads(get_profile(Buffer.profile_events))
+    profile_ret = load_profile(Buffer.profile_events)
     ret = profile_ret["layout"][f"{c.device} Memory"]
     self.assertEqual(ret["peak"], 2)
-    self.assertEqual(ret["shapes"][0]["x"], [0, 3])
-    self.assertEqual(ret["shapes"][1]["x"], [1, 3, 3, 4])
-    self.assertEqual(ret["shapes"][0]["y"], [0, 0])
-    self.assertEqual(ret["shapes"][1]["y"], [1, 1, 0, 0])
-    self.assertEqual(ret["shapes"][2]["x"], [3, 4])
-    self.assertEqual(ret["shapes"][2]["y"], [1, 1])
+    self.assertEqual(len(ret["events"]), 4)
 
 if __name__ == "__main__":
   unittest.main()

tinygrad/apps/llm.py

@@ -53,17 +53,15 @@ class SimpleTokenizer:
     try: return [ self._normal_tokens[p] for p in parts ]
     except KeyError: raise RuntimeError("token not found")
 
-def apply_rope(x:Tensor, start_pos:int|UOp, base:int=10000):
+def apply_rope(x:Tensor, start_pos:int|UOp, base:float = 10000.0) -> Tensor:
   B, H, T, Hd = x.shape
-  # NOTE: this is usually in a RoPE cache, but tinygrad JIT should prune it outside the kernel
-  # TODO: make it do that
-  freq = base ** (-Tensor.arange(0, 1, 2/Hd, dtype='float32'))
-  angles = Tensor.arange(start_pos, start_pos+T, dtype='float32')[None, None, :, None] * freq
-  cos, sin = angles.cos(), angles.sin()
-  x = x.reshape(B, H, T, Hd // 2, 2)    # split into pairs
-  y1 = x[..., 0] * cos - x[..., 1] * sin
-  y2 = x[..., 0] * sin + x[..., 1] * cos
-  return Tensor.stack(y1, y2, dim=-1).reshape(B, H, T, Hd)
+  assert (Hd & 1) == 0, "RoPE requires an even head dimension"
+  half = Hd // 2
+  angles = (Tensor.arange(T, dtype="float32") + start_pos)[:, None] * (base ** (-(Tensor.arange(half, dtype="float32") / half)))[None, :]
+  cos, sin = angles.cos().reshape(1, 1, T, half).cast(x.dtype), angles.sin().reshape(1, 1, T, half).cast(x.dtype)
+  x_pairs = x.reshape(B, H, T, half, 2)
+  return Tensor.stack(x_pairs[..., 0] * cos - x_pairs[..., 1] * sin,
+                      x_pairs[..., 0] * sin + x_pairs[..., 1] * cos, dim=-1).reshape(B, H, T, Hd)
 
 class TransformerBlock:
   def __init__(self, dim:int, hidden_dim:int, n_heads:int, n_kv_heads:int, norm_eps:float, max_context:int=0):

tinygrad/codegen/__init__.py

@@ -1,7 +1,7 @@
 from typing import Any, Callable
 import functools
 from dataclasses import dataclass
-from tinygrad.helpers import QUANTIZE, DEVECTORIZE, TRANSCENDENTAL
+from tinygrad.helpers import QUANTIZE, DEVECTORIZE, TRANSCENDENTAL, RANGEIFY, POSTOPT
 from tinygrad.uop.ops import PatternMatcher, graph_rewrite, UOp
 from tinygrad.uop.spec import type_verify
 from tinygrad.renderer import Renderer
@@ -12,12 +12,14 @@ from tinygrad.codegen.quantize import pm_quant
 from tinygrad.codegen.gpudims import pm_add_gpudims
 from tinygrad.uop.symbolic import sym, symbolic_simple, gep_pushing
 from tinygrad.uop.decompositions import get_late_rewrite_patterns
-from tinygrad.codegen.expander import migrate_indexing, expander
-from tinygrad.codegen.devectorizer import load_store_folding, load_store_indexing, devectorize, pm_reduce, \
+from tinygrad.codegen.late.expander import migrate_indexing, expander, pm_pre_expander
+from tinygrad.codegen.late.devectorizer import load_store_folding, load_store_indexing, devectorize, pm_reduce, \
   ReduceContext, correct_load_store, pm_render
-from tinygrad.codegen.linearize import block_create, pm_blockend_merge, block_merge, pm_finalize, BlockContext
-from tinygrad.codegen.opt import pm_optimize
+from tinygrad.codegen.late.linearize import block_create, pm_blockend_merge, block_merge, pm_finalize, BlockContext
+from tinygrad.codegen.opt import pm_get_optimization, pm_do_optimize
 from tinygrad.codegen.opt.swizzler import view_left, view_right, fix_kernel_ops
+from tinygrad.codegen.opt.postrange import pm_postrange_opt
+from tinygrad.schedule.rangeify import pm_add_buffers_local, rangeify_codegen
 
 @dataclass
 class RewriteStep:
@@ -44,10 +46,10 @@ rewrites_for_linearizer = [
 
 def get_rewrites_for_renderer(opts:Renderer, linearizer:bool=True) -> list[RewriteStep]:
   # cache with the values of the context vars
-  return _get_rewrites_for_renderer(opts, linearizer, QUANTIZE.value, DEVECTORIZE.value, TRANSCENDENTAL.value)
+  return _get_rewrites_for_renderer(opts, linearizer, QUANTIZE.value, DEVECTORIZE.value, TRANSCENDENTAL.value, RANGEIFY.value, POSTOPT.value)
 
 @functools.cache
-def _get_rewrites_for_renderer(opts:Renderer, linearizer:bool, _QUANTIZE, _DEVECTORIZE, _TRANSCENDENTAL) -> list[RewriteStep]:
+def _get_rewrites_for_renderer(opts:Renderer, linearizer:bool, _QUANTIZE, _DEVECTORIZE, _TRANSCENDENTAL, _RANGEIFY, _POSTOPT) -> list[RewriteStep]:
   # ** lowerer (rewrite_shapetracker_with_index) **
   ret: list[RewriteStep] = []
 
@@ -55,22 +57,28 @@ def _get_rewrites_for_renderer(opts:Renderer, linearizer:bool, _QUANTIZE, _DEVEC
   ret.extend(rewrites_for_views)
 
   # this is kernel.py
-  ret.append(RewriteStep(pm_optimize, ctx=lambda _: opts, name="optimize ast"))
+  if not _RANGEIFY: ret.append(RewriteStep(pm_get_optimization, ctx=lambda _: opts, name="get optimization"))
+  if not _POSTOPT and not _RANGEIFY: ret.append(RewriteStep(pm_do_optimize, ctx=lambda _: opts, name="optimize ast"))
 
   if _QUANTIZE and opts.device in {"CPU", "DSP"}: ret.append(RewriteStep(pm_quant, name="quantize"))
   ret.append(RewriteStep(pm_lowerer, get_index, name="lowerer", bottom_up=True))
 
+  if _POSTOPT or _RANGEIFY: ret.append(RewriteStep(pm_postrange_opt, ctx=lambda _: opts, name="post optimize ast"))
+
   # ** expander (expand_rewrite) **
   ret.append(RewriteStep(sym+migrate_indexing, name="initial symbolic"))
 
   # expand
-  ret.append(RewriteStep(sym+expander, name="expander"))
+  ret.append(RewriteStep(sym+pm_pre_expander+expander, name="expander"))
+
+  # add locals
+  ret.append(RewriteStep(pm_add_buffers_local+rangeify_codegen, name="add local buffers"))
 
   # ** devectorizer (full_graph_rewrite) **
   # remove reduce
   ret.append(RewriteStep(pm_reduce+gep_pushing, lambda _: ReduceContext(), name="remove_reduce"))
 
-  # add gpu dims (late)
+  # add gpu dims (late). this works after devectorize, but it's faster here
   ret.append(RewriteStep(pm_add_gpudims, lambda _: opts, name="add gpudims"))
 
   # devectorize (TODO: does this need opts?)

tinygrad/codegen/gpudims.py

@@ -1,6 +1,6 @@
 import math
 from tinygrad.uop.ops import UOp, Ops, sint, PatternMatcher, UPat, KernelInfo, ssimplify, AxisType
-from tinygrad.helpers import all_int
+from tinygrad.helpers import all_int, dedup
 from tinygrad.dtype import dtypes
 from tinygrad.shape.view import get_contraction
 from tinygrad.renderer import Renderer
@@ -52,20 +52,24 @@ def get_grouped_dims(prefix, dims:tuple[sint, ...], max_sizes:tuple[int, ...]|No
 
 def add_gpudims(ctx:Renderer, s:UOp):
   if s.arg is None: return None
-  ki: KernelInfo = s.arg
-  global_dims = [i for i,x in enumerate(ki.axis_types) if x is AxisType.GLOBAL]
-  local_dims = [i for i,x in enumerate(ki.axis_types) if x in (AxisType.LOCAL, AxisType.GROUP_REDUCE)]
-  if not global_dims and not local_dims: return None
   s_topo = list(s.toposort())
   if any(x.op is Ops.SPECIAL for x in s_topo): return None
 
+  # get ranges
+  all_ranges = {x.arg[0:-1]:x for x in s_topo if x.op is Ops.RANGE}
+
+  # extract global/local dims
+  global_dims = sorted(dedup([x.arg[0:-1] for x in all_ranges.values() if x.arg[-1] is AxisType.GLOBAL]))
+  local_dims = sorted(dedup([x.arg[0:-1] for x in all_ranges.values() if x.arg[-1] in (AxisType.LOCAL, AxisType.GROUP_REDUCE)]))
+  if not global_dims and not local_dims: return None
+
   # get global and local shape
-  all_ranges = {x.arg%1000:x for x in s_topo if x.op is Ops.RANGE}
   ranges = [all_ranges[r] for r in global_dims+local_dims if r in all_ranges]
-  global_shape = tuple([ssimplify(r.src[0]) for r in ranges if r.arg%1000 in global_dims])
-  local_shape = tuple([ssimplify(r.src[0]) for r in ranges if r.arg%1000 in local_dims])
+  global_shape = tuple([ssimplify(r.src[0]) for r in ranges if r.arg[0:-1] in global_dims])
+  local_shape = tuple([ssimplify(r.src[0]) for r in ranges if r.arg[0:-1] in local_dims])
 
   # get the idxs
+  ki: KernelInfo = s.arg
   if ki.dont_use_locals:
     assert not local_dims, "can't use locals if there's no local dims"
     idxs = get_grouped_dims("idx", global_shape, ctx.global_max, reverse=True)
@@ -78,12 +82,13 @@ def add_gpudims(ctx:Renderer, s:UOp):
   for r in s_topo:
     if r.op is not Ops.RANGE: continue
     try:
-      ii = (global_dims+local_dims).index(r.arg%1000)
-      if r.arg < 2000 and ki.axis_types[r.arg%1000] == AxisType.GROUP_REDUCE: continue
+      ii = (global_dims+local_dims).index(r.arg[0:-1])
+      if r.arg[1] == AxisType.REDUCE: continue
       subs[r] = idxs[ii]
     except ValueError: continue
   return s.substitute(subs)
 
 pm_add_gpudims = PatternMatcher([
+  # add gpudims must be last
   (UPat(Ops.SINK, name="s"), add_gpudims),
 ])

tinygrad/codegen/late/devectorizer.py

@@ -2,7 +2,7 @@ from typing import Any, cast
 import functools, operator, itertools
 from collections import defaultdict
 from dataclasses import dataclass
-from tinygrad.dtype import dtypes, ImageDType, PtrDType, DType, AddrSpace
+from tinygrad.dtype import dtypes, ImageDType, DType, AddrSpace
 from tinygrad.uop.ops import UOp, Ops, UPat, PatternMatcher, graph_rewrite, GroupOp, identity_element
 from tinygrad.uop.symbolic import split_uop, uop_given_valid, parse_valid, simplify_valid, sym, symbolic_flat
 from tinygrad.helpers import getenv, flatten, AMX, prod, partition
@@ -80,9 +80,6 @@ def expand_index(buf:UOp, vec:UOp, mask:UOp|None=None):
     if len(midx.src[i].src) == 3: root_src = (midx.src[i].src[2], root_src)
     offsets_rootsrc[root_src].setdefault(arg, []).append(i)
 
-  # the buf.dtype is always a pointer
-  ptrdtype = cast(PtrDType, buf.dtype)
-
   # then rewrite everything we can into groups
   ret = []
   idxs: list[int|None] = [None]*vec.dtype.count
@@ -92,7 +89,7 @@ def expand_index(buf:UOp, vec:UOp, mask:UOp|None=None):
     for grp in grouped_offsets:
       # get the index offset for this element. using [0] is okay, because they are the same
       lidx = midx.src[offsets[grp[0]][0]]
-      if len(grp) > 1: lidx = lidx.cast(ptrdtype.base.vec(len(grp)).ptr(size=ptrdtype.size, addrspace=ptrdtype.addrspace))
+      if len(grp) > 1: lidx = lidx.cast(buf.ptrdtype.base.vec(len(grp)).ptr(size=buf.ptrdtype.size, addrspace=buf.ptrdtype.addrspace))
       # set the idxs of the output
       for i,g in enumerate(grp):
         for oo in offsets[g]: idxs[oo] = global_offset+i
@@ -101,7 +98,7 @@ def expand_index(buf:UOp, vec:UOp, mask:UOp|None=None):
       global_offset += len(grp)
   assert None not in idxs, f"some idxs are missing {idxs}"
   # this base thing is for image, we want the CAT to be a normal pointer
-  post_cat = UOp(Ops.PTRCAT, ptrdtype.base.ptr(size=ptrdtype.size, addrspace=ptrdtype.addrspace).vec(vec.dtype.count), tuple(ret))
+  post_cat = UOp(Ops.PTRCAT, buf.ptrdtype.base.ptr(size=buf.ptrdtype.size, addrspace=buf.ptrdtype.addrspace).vec(vec.dtype.count), tuple(ret))
   return post_cat.gep(tuple(cast(list[int], idxs)))
 
 def cat_after_store(cat:UOp, data:UOp, sto:UOp):
@@ -154,7 +151,7 @@ def split_load_store(ctx:Renderer|None, ls:UOp, idx:UOp):
     must_divide = False
   elif buf.dtype.base != dtypes.float and buf.dtype.base != dtypes.half and not isinstance(buf.dtype, ImageDType):
     pass
-  elif cast(PtrDType, buf.dtype).addrspace == AddrSpace.REG:
+  elif buf.ptrdtype.addrspace == AddrSpace.REG:
     pass
   elif isinstance(buf.dtype, ImageDType):
     lengths = [4]
@@ -169,13 +166,12 @@ def split_load_store(ctx:Renderer|None, ls:UOp, idx:UOp):
   # split based on the fold lengths
   global_offset = 0
   ret = []
-  ptrdtype = cast(PtrDType, buf.dtype)
   while global_offset < sz:
     # with 1 at the end of the lengths list, this will always hit
     for fold_length in lengths:
       if global_offset+fold_length > sz: continue
       lidx = buf.index(idx.src[1] + global_offset, idx.src[2] if len(idx.src) > 2 else None)
-      if fold_length > 1: lidx = lidx.cast(ptrdtype.base.vec(fold_length).ptr(size=ptrdtype.size, addrspace=ptrdtype.addrspace))
+      if fold_length > 1: lidx = lidx.cast(buf.ptrdtype.base.vec(fold_length).ptr(size=buf.ptrdtype.size, addrspace=buf.ptrdtype.addrspace))
       if ls.op is Ops.STORE: ret.append(ls.replace(src=(lidx,ls.src[1].gep(tuple(range(global_offset, global_offset+fold_length))))+ls.src[2:]))
       else: ret.append(ls.replace(src=(lidx,)+ls.src[1:], dtype=ls.dtype.scalar().vec(fold_length)))
       global_offset += fold_length
@@ -232,17 +228,20 @@ def no_vectorized_alu(alu:UOp):
   alus = tuple(UOp(alu.op, alu.dtype.scalar(), tuple(s.gep(i) for s in alu.src), alu.arg) for i in range(alu.dtype.vcount))
   return UOp(Ops.VECTORIZE, alu.dtype, alus)
 
-def no_vectorized_acc(acc:UOp, c:UOp):
-  if acc.dtype.count == 1: return None
-  assert c.arg == 0, "this only supports index 0"
-  new_acc = acc.replace(dtype=acc.dtype.base.scalar().ptr(acc.dtype.count, cast(PtrDType, acc.dtype).addrspace))
-  return UOp(Ops.PTRCAT, acc.dtype, tuple([new_acc.index(UOp.const(dtypes.int, i)) for i in range(acc.dtype.count)]))
+def no_vectorized_buf(buf:UOp):
+  return buf.replace(dtype=buf.ptrdtype.base.scalar().ptr(buf.ptrdtype.size*buf.ptrdtype.count, buf.ptrdtype.addrspace)).cast(buf.dtype)
+
+def no_vectorized_index(buf:UOp, cast:UOp, idx:UOp):
+  cnt = cast.dtype.count
+  assert idx.dtype.count == 1, f"idx dtype must be 1 {idx.dtype}"
+  return buf.broadcast(cnt).index(idx.broadcast(cnt)*cnt+UOp.const(dtypes.int.vec(cnt), tuple(range(cnt))))
 
 devectorize = PatternMatcher([
   # no ALU on vectorized dtypes
   (UPat((*GroupOp.ALU, Ops.CAST, Ops.BITCAST), name="alu"), no_vectorized_alu),
   (UPat(Ops.WMMA, name="wmma"), no_vectorized_wmma),
-  (UPat(Ops.DEFINE_REG, name="acc").index(UPat.cvar("c")), no_vectorized_acc),
+  (UPat((Ops.DEFINE_LOCAL, Ops.DEFINE_REG), name="buf"), no_vectorized_buf),
+  (UPat((Ops.DEFINE_LOCAL, Ops.DEFINE_REG), name="buf").cast(name="cast").index(UPat.var("idx")), no_vectorized_index),
 ])
 
 pm_render = PatternMatcher([

tinygrad/codegen/late/expander.py

@@ -1,8 +1,9 @@
 # this converts a lowerer program into a vectorized program
 
 import functools, itertools, operator
-from tinygrad.helpers import AMX, dedup, flatten, all_same, prod
-from tinygrad.uop.ops import UOp, Ops, UPat, PatternMatcher, GroupOp
+from tinygrad.dtype import dtypes, PtrDType, AddrSpace
+from tinygrad.helpers import AMX, dedup, flatten, all_same, prod, partition
+from tinygrad.uop.ops import UOp, Ops, UPat, PatternMatcher, GroupOp, AxisType
 
 def _expand_arg_to_idx(args:tuple[tuple[int, int], ...], rpk:dict[int, int]) -> int:
   idx, mul = 0, 1
@@ -46,11 +47,13 @@ def do_expand(root:UOp):
         new_srcs.append(src.src[0].gep(tuple(lst)))
     else:
       # non-UNROLL input
-      if root.op is Ops.IF:
+      if root.op is Ops.IF or src.op is Ops.IF:
         # for the first arg of IF, just pass them through ignoring UNROLLS
         new_srcs.append(src)
-      elif root.op in {Ops.REDUCE, Ops.STORE} and src.op is Ops.RANGE:
-        # for any range args of REDUCE, pass them through
+      elif (root.op is Ops.STORE and i >= 2) or (root.op in {Ops.REDUCE, Ops.BUFFERIZE} and i >= 1) or (root.op is Ops.WMMA and i >= 3):
+        # for any range args of STORE/REDUCE, pass them through
+        new_srcs.append(src)
+      elif root.op is Ops.INDEX and i >= 1 and not isinstance(root.dtype, PtrDType):
         new_srcs.append(src)
       elif src.dtype.count > 1:
         # put any input dtype > 1 grouped together
@@ -72,7 +75,7 @@ def do_contract(con:UOp):
   # CONTRACT without UNROLL repeats the element VECTORIZED
   if ex.op is not Ops.UNROLL: return UOp(Ops.VECTORIZE, con.dtype, con.src*con.dtype.count)
   # CONTRACT may remove several axes from UNROLL
-  assert con.dtype.count == prod([x[1] for x in con.arg]), "dtype is wrong"
+  assert con.dtype == dtypes.void or con.dtype.count == prod([x[1] for x in con.arg]), "dtype is wrong"
   idxs = []
   for rpk in _choices_from_args(new_ex_args:=tuple(x for x in ex.arg if x not in con.arg)):
     idxs += [_expand_arg_to_idx(ex.arg, {**rpk, **lrpk}) for lrpk in _choices_from_args(con.arg)]
@@ -83,7 +86,7 @@ expander = PatternMatcher([
   (UPat(Ops.UNROLL, name="outer", src=(UPat(Ops.UNROLL, name="inner"),)),
    lambda outer, inner: UOp(Ops.UNROLL, outer.dtype, (inner.src[0],), inner.arg+outer.arg)),
   # do expansion
-  (UPat((*GroupOp.ALU, Ops.CAST, Ops.BITCAST, Ops.GEP, Ops.WMMA, Ops.LOAD, Ops.STORE, Ops.INDEX,
+  (UPat((*GroupOp.ALU, Ops.CAST, Ops.BITCAST, Ops.GEP, Ops.WMMA, Ops.LOAD, Ops.STORE, Ops.INDEX, Ops.BUFFERIZE,
          Ops.VECTORIZE, Ops.IF, Ops.REDUCE), name="root", custom_early_reject=set([Ops.UNROLL])), do_expand),
   (UPat(Ops.CONTRACT, name="con"), do_contract),
   # BARRIERs aren't actually expanded
@@ -111,3 +114,49 @@ migrate_indexing = PatternMatcher([
   # create gate MUST BE BEFORE expander
   (UPat(Ops.STORE, name="root"), create_gate),
 ])
+
+# ****
+
+def fix_reduce_unroll(x:UOp):
+  reduce_range, reduce_expand = partition(x.src[1:], lambda y: y.op is Ops.RANGE)
+  if len(reduce_expand) == 0: return None
+  reduce_expand = [x for x in reduce_expand if x.op is not Ops.CONST]
+  assert all(x.op is Ops.UNROLL for x in reduce_expand), f"not all UNROLLS in {reduce_expand}"
+  ret = x.src[0]
+  if len(contract_axis:=flatten(x.arg for x in reduce_expand)):
+    ret = UOp(Ops.CONTRACT, x.dtype.vec(prod(x[1] for x in contract_axis)), (ret,), tuple(contract_axis), tag=1)
+  # REDUCE supports both "horizontal" reduction and range reduction. the horizontal elements are taken in the nearest group
+  return x.replace(src=(ret,)+tuple(reduce_range))
+
+def fix_store_unroll(x:UOp):
+  store_expand, store_range = partition(x.src[2:], lambda y: y.op is Ops.UNROLL)
+  if len(store_expand) == 0: return None
+  return UOp(Ops.CONTRACT, dtypes.void, (x.replace(src=x.src[:2]+tuple(store_range)),), tuple(flatten(x.arg for x in store_expand)), tag=1)
+
+def fix_group_for_reduce(x:UOp):
+  reduce_gfr, reduce_r = partition(x.src[1:], lambda u: u.op is Ops.RANGE and u.arg[1] == AxisType.GROUP_REDUCE)
+  if len(reduce_gfr) == 0: return None
+
+  # NOTE: if there's other locals here, we need them in the buffer too
+  upstream_locals = [u for u in x.toposort() if u.op is Ops.RANGE and u.arg[1] == AxisType.LOCAL]
+
+  # do only the non grouped reduces early
+  ret = x.replace(src=(x.src[0],)+tuple(reduce_r))
+  reduce_loop = [x.replace(arg=(x.arg[0]+100, AxisType.REDUCE)) for x in reduce_gfr]
+  buf = ret.bufferize(*upstream_locals, *reduce_gfr, arg=(AddrSpace.LOCAL, reduce_gfr[0].arg[0])).index(*upstream_locals, *reduce_loop)
+
+  # gate with an if on the store + do the final reduce
+  buf = UOp(Ops.IF, dtype=buf.dtype, src=(functools.reduce(operator.and_, [x.eq(0) for x in reduce_gfr]), buf))
+  return buf.reduce(*reduce_loop, arg=x.arg)
+
+pm_pre_expander = PatternMatcher([
+  # rewrite UPCAST/UNROLL range to something to be expanded
+  (UPat(Ops.RANGE, name="r"),
+   lambda r: UOp(Ops.UNROLL, dtypes.int, (UOp.const(dtypes.int.vec(s:=r.vmax+1), tuple(range(s))),), ((r.arg[0],s),)) \
+    if r.arg[1] in {AxisType.UNROLL, AxisType.UPCAST} else None),
+  # fix REDUCEs with UNROLLs
+  (UPat(Ops.REDUCE, name="x"), fix_reduce_unroll),
+  (UPat(Ops.STORE, name="x"), fix_store_unroll),
+  # fix group for reduce
+  (UPat(Ops.REDUCE, name="x"), fix_group_for_reduce),
+])

tinygrad/codegen/late/linearize.py

@@ -3,7 +3,7 @@ import heapq
 from collections import defaultdict
 from dataclasses import dataclass, replace
 from tinygrad.uop.ops import UOp, Ops, PatternMatcher, UPat, GroupOp
-from tinygrad.helpers import dedup, all_same, flatten, getenv
+from tinygrad.helpers import dedup, all_same, flatten, BLOCK_REORDER
 
 # NOTE: any toposort should be valid here, unlike last time this isn't required, it's just for speed
 def block_reorder(lst:list[UOp]) -> list[UOp]:
@@ -150,7 +150,7 @@ def make_block_bottom_up(ctx:BlockContext, x:UOp):
     srcs.append(add_blockends(base_block, new_ctx, current_ctx))
 
   lst = lst[::-1]
-  if getenv("BLOCK_REORDER", 1): lst = block_reorder(lst)
+  if BLOCK_REORDER: lst = block_reorder(lst)
   bb = BasicBlock(tuple(lst), ctx=current_ctx, cnt=child_count, child_ctx=child_ctx)
   return UOp(Ops.BLOCK, src=tuple(srcs), arg=bb)
 

tinygrad/codegen/lowerer.py

@@ -1,10 +1,6 @@
 # the job of the lowerer is to do indexing
-import functools, operator
-from typing import cast
 from dataclasses import dataclass
-from tinygrad.dtype import dtypes, AddrSpace, PtrDType
-from tinygrad.uop.ops import KernelInfo, UOp, Ops, PatternMatcher, UPat, sint_to_uop, AxisType, graph_rewrite
-from tinygrad.helpers import prod, partition, flatten
+from tinygrad.uop.ops import KernelInfo, UOp, Ops, PatternMatcher, UPat, sint_to_uop, AxisType, graph_rewrite, resolve
 
 # ***** indexing *****
 
@@ -15,20 +11,12 @@ class IndexContext:
   start: int = 0
 
 def shape_to_idx(s, axis_types, start=0):
-  # indexes
-  idxs = []
-  for i, (s, at) in enumerate(zip(s, axis_types)):
-    if at in (AxisType.UPCAST, AxisType.UNROLL):
-      assert isinstance(s, int), "needs to be int to upcast/unroll"
-      idxs.append(UOp(Ops.UNROLL, dtypes.int, (UOp.const(dtypes.int.vec(s), tuple(range(s))),), ((i,s),), tag=1))
-    else:
-      # all others are RANGES
-      idxs.append(UOp(Ops.RANGE, dtypes.int, (sint_to_uop(s),), start+i))
-  return idxs
+  return [UOp.range(sint_to_uop(s), start+i, at) for i, (s, at) in enumerate(zip(s, axis_types))]
 
 def get_index(ast:UOp) -> IndexContext:
   axis_types = ast.arg.axis_types if isinstance(ast.arg, KernelInfo) else ()
-  if len(ast.full_shape) != len(axis_types): axis_types = (AxisType.LOOP,)*len(ast.full_shape)
+  if len(ast.full_shape) != len(axis_types) and ast.st is not None:
+    axis_types = tuple([AxisType.REDUCE if resolve(s != fs) else AxisType.LOOP for s,fs in zip(ast.shape, ast.full_shape)])
   return IndexContext(axis_types, [], 0)
 
 # ***** lowering (given index) *****
@@ -42,16 +30,8 @@ def lower_reduce_axis(ctx: IndexContext, x: UOp):
   new_idxs = shape_to_idx(x.src[0].shape, ctx.axis_types, ctx.start)
   full_new_idx = list(ctx.idxs)
   for a in x.axis_arg: full_new_idx[a] = new_idxs[a]
-
   ret = subblock(ctx, full_new_idx, x.src[0])
-
-  # NOTE: always using ridxs is fine here
-  reduce_range, reduce_expand = partition([full_new_idx[i] for i in x.axis_arg], lambda y: y.op is Ops.RANGE)
-  assert all(x.op is Ops.UNROLL for x in reduce_expand), f"not all UNROLLS in {reduce_expand} for {x.axis_arg}"
-  if len(contract_axis:=flatten(x.arg for x in reduce_expand)):
-    ret = UOp(Ops.CONTRACT, x.dtype.vec(prod(x[1] for x in contract_axis)), (ret,), tuple(contract_axis), tag=1)
-  # REDUCE supports both "horizontal" reduction and range reduction. the horizontal elements are taken in the nearest group
-  return UOp(Ops.REDUCE, x.dtype, (ret,)+tuple(reduce_range), x.arg[0])
+  return UOp(Ops.REDUCE, x.dtype, (ret,)+tuple([full_new_idx[i] for i in x.axis_arg]), x.arg[0])
 
 def lower_store(ctx: IndexContext, x: UOp, buf: UOp):
   # TODO: reenable after REDUCE_AXIS is fixed
@@ -67,15 +47,7 @@ def lower_store(ctx: IndexContext, x: UOp, buf: UOp):
 
   stored = subblock(ctx, real_new_idxs, x.src[1])
   used_ranges = [x for x in used_idxs if x.op is Ops.RANGE]
-  ret = buf.index(idx, valid).store(stored, *used_ranges)
-
-  # insert BARRIER if we are ending a LOCAL, IF if we are ending a GROUP_REDUCE
-  if cast(PtrDType, buf.dtype).addrspace == AddrSpace.LOCAL and \
-      any(ctx.axis_types[x.arg%1000] in {AxisType.GROUP_REDUCE, AxisType.LOCAL} for x in used_ranges):
-    ret = ret.barrier()
-    range_gates = [x.eq(0) for x in used_ranges if ctx.axis_types[x.arg%1000] == AxisType.GROUP_REDUCE]
-    if len(range_gates): ret = UOp(Ops.IF, src=(functools.reduce(operator.and_, range_gates), ret))
-  return ret
+  return buf.index(idx, valid).store(stored, *used_ranges)
 
 def fixup_wmma(ctx:IndexContext, x:UOp):
   if x.tag is not None: return None
@@ -86,8 +58,8 @@ def fixup_wmma(ctx:IndexContext, x:UOp):
   srcs = subblock(ctx, full_new_idx, UOp.sink(*x.src)).src
 
   # NOTE: this assumes these are expanded. which now shouldn't change anything
-  new_x_arg_m2 = tuple([tuple([(full_new_idx[a].arg[0][0], sz) for a,sz in v]) for v in x.arg[-2]])
-  new_x_arg_m1 = tuple([full_new_idx[a].arg[0][0] for a in x.arg[-1]])
+  new_x_arg_m2 = tuple([tuple([(full_new_idx[a].arg[0], sz) for a,sz in v]) for v in x.arg[-2]])
+  new_x_arg_m1 = tuple([full_new_idx[a].arg[0] for a in x.arg[-1]])
   return x.replace(src=srcs, arg=x.arg[:-2]+(new_x_arg_m2, new_x_arg_m1), tag=1)
 
 pm_lowerer = PatternMatcher([
@@ -110,5 +82,5 @@ pm_lowerer = PatternMatcher([
 
   # axis fixups for WMMA
   (UPat((Ops.CONTRACT, Ops.UNROLL), name="x"),
-   lambda ctx,x: x.replace(tag=1, arg=tuple([(ctx.idxs[a].arg[0][0], sz) for a,sz in x.arg])) if x.tag is None else None),
+   lambda ctx,x: x.replace(tag=1, arg=tuple([(ctx.idxs[a].arg[0], sz) for a,sz in x.arg])) if x.tag is None else None),
 ])

tinygrad/codegen/opt/__init__.py

@@ -2,12 +2,12 @@
 
 from tinygrad.codegen.opt.kernel import Kernel
 from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
-from tinygrad.uop.ops import UOp, PatternMatcher, UPat, Ops
+from tinygrad.uop.ops import UOp, PatternMatcher, UPat, Ops, KernelInfo
 from tinygrad.helpers import NOOPT, BEAM, USE_TC, getenv
 from tinygrad.renderer import Renderer
 from tinygrad.uop.spec import type_verify
 
-def get_optimized_ast(ast:UOp, renderer:Renderer) -> UOp:
+def get_optimized_ast(ast:UOp, renderer:Renderer) -> UOp|None:
   """
   Optimize an AST based on heuristics or BEAM search.
 
@@ -19,20 +19,33 @@ def get_optimized_ast(ast:UOp, renderer:Renderer) -> UOp:
     The Ops.SINK rooted AST transformed to apply the opts and with a KernelInfo in the arg.
   """
 
+  # no shape, no opt
+  if ast.src[0].st is None: return None
+  new_arg = ast.arg
+  if new_arg is None:
+    k = Kernel(ast, opts=renderer)
+    if not NOOPT:
+      if not k.apply_tensor_cores(USE_TC.value): k.apply_opts(hand_coded_optimizations(k))
+      if BEAM >= 1:
+        from tinygrad.codegen.opt.search import beam_search, bufs_from_lin
+        kb = Kernel(ast, opts=renderer)
+        rawbufs = bufs_from_lin(kb, allocate=False)
+        k = beam_search(kb, rawbufs, BEAM.value, bool(getenv("BEAM_ESTIMATE", 1)))
+    new_arg = KernelInfo(opts_to_apply=tuple(k.applied_opts))
+  elif len(new_arg.applied_opts): return None
+  return Kernel(ast.replace(arg=None), opts=renderer).get_optimized_ast().replace(arg=new_arg)
+
+pm_get_optimization = PatternMatcher([
+  (UPat(Ops.SINK, name="ast"), lambda ctx,ast: get_optimized_ast(ast, ctx)),
+])
+
+def apply_opt(ast:UOp, renderer:Renderer):
   k = Kernel(ast, opts=renderer)
-  if ast.arg is not None and ast.arg.opts_to_apply is not None: k.apply_opts(ast.arg.opts_to_apply)
-  elif not NOOPT:
-    if not k.apply_tensor_cores(USE_TC.value): k.apply_opts(hand_coded_optimizations(k))
-    if BEAM >= 1:
-      from tinygrad.codegen.opt.search import beam_search, bufs_from_lin
-      kb = Kernel(ast, opts=renderer)
-      rawbufs = bufs_from_lin(kb, allocate=False)
-      k = beam_search(kb, rawbufs, BEAM.value, bool(getenv("BEAM_ESTIMATE", 1)))
+  k.apply_opts(ast.arg.opts_to_apply)
   ret = k.get_optimized_ast()
   if __debug__: type_verify(list(ret.toposort()))
   return ret
 
-pm_optimize = PatternMatcher([
-  (UPat(Ops.SINK, name="ast"), lambda ctx,ast:
-   get_optimized_ast(ast, ctx) if (ast.arg is None or ast.arg.opts_to_apply is not None) and ast.src[0].st is not None else None),
+pm_do_optimize = PatternMatcher([
+  (UPat(Ops.SINK, name="ast"), lambda ctx,ast: apply_opt(ast, ctx) if ast.arg is not None and ast.arg.opts_to_apply is not None else None),
 ])

tinygrad/codegen/opt/heuristic.py

@@ -28,7 +28,7 @@ def hand_coded_optimizations(k:Kernel) -> list[Opt]:
           return k.applied_opts
 
   # are we grouping? (requires local shape support)
-  if resolve(prod(k.sts[0].shape[i] for i in k.upcastable_dims) <= 2048, False):
+  if resolve(prod(k.output_shape[i] for i in k.upcastable_dims) <= 2048, False):
     for sz in [16]:
       try:
         k.apply_opt(Opt(OptOps.GROUPTOP, 0, sz))
@@ -62,7 +62,7 @@ def hand_coded_optimizations(k:Kernel) -> list[Opt]:
   # potentially do more upcasts of non reduce axes based on a heuristic
   is_dsp = k.opts is not None and k.opts.device == "DSP"
   upcasted_axis: set[int] = set()
-  while resolve(prod(k.sts[0].shape[i] for i in k.upcastable_dims) >= 1024):
+  while resolve(prod(k.output_shape[i] for i in k.upcastable_dims) >= 1024):
     xb_choices = []
     # consider all upcastable axes with 3 or 4 upcast (128 on the DSP)
     for axis, upcast_amount in itertools.product(k.upcastable_dims, ([128] if not len(upcasted_axis) else []) if is_dsp else [3,4]):

tinygrad/codegen/opt/kernel.py

@@ -10,7 +10,7 @@ from tinygrad.uop.spec import type_verify, ast_spec
 from tinygrad.device import Device
 from tinygrad.codegen.opt.tc import TensorCore
 from tinygrad.renderer import Renderer
-from tinygrad.dtype import ImageDType, AddrSpace
+from tinygrad.dtype import ImageDType
 from tinygrad.helpers import all_same, colored, ansilen, dedup, prod, round_up, to_function_name, unwrap, argfix, DEBUG, TC_SELECT, TC_OPT, AMX
 from tinygrad.shape.shapetracker import ShapeTracker
 from tinygrad.shape.view import strides_for_shape, get_contraction
@@ -60,7 +60,7 @@ class Kernel:
 
     self.vars: list[Variable] = self.ast.variables()
     # NOTE: this requires a specific order with the [::-1], this is likely a bug
-    self.bufs: list[UOp] = [x for x in self.ast.toposort() if x.op in GroupOp.Buffer][::-1]
+    self.bufs: list[UOp] = [x for x in self.ast.toposort() if x.op in GroupOp.Buffer and x.st is not None][::-1]
 
     # create new shapetrackers inside this kernel, we will permute them
     self.sts: list[ShapeTracker] = [x.st_arg for x in self.bufs]
@@ -122,7 +122,7 @@ class Kernel:
   @property
   def output_shape(self) -> tuple[sint, ...]: return self.sts[0].shape
   @property
-  def shape_len(self) -> int: return len(self.sts[0].shape)
+  def shape_len(self) -> int: return len(self.full_shape)
 
   def axes_of(self, *axis_type:AxisType) -> list[int]: return [i for i,t in enumerate(self.axis_types) if t in argfix(axis_type)]
   @property
@@ -174,7 +174,7 @@ class Kernel:
   # amount : the amount to take
   # top : if you want to pull that amount from the top
   # insert_at : place to insert the new stuff
-  def shift_to(self, axis:int, amount:int, new_type:AxisType, top:bool=False, insert_at:int|None=None):
+  def shift_to(self, axis:int, amount:int, new_type:AxisType, top:bool=False, insert_at:int|None=None) -> int:
     if insert_at is None: insert_at = self.shape_len
     self.axis_types.insert(insert_at, new_type)
     move_axis = axis if top else axis+1
@@ -183,6 +183,7 @@ class Kernel:
     new_axes = [i for i in range(insert_at) if i != move_axis]+[move_axis]+[i for i in range(insert_at, self.shape_len+1) if i != move_axis]
     self.reshape(new_shape_fxn)
     self.permute(new_axes)
+    return insert_at
 
   # ******************** complex simplifiers ********************
 
@@ -244,11 +245,11 @@ class Kernel:
       if axis is None: return -1
       if op is OptOps.UNROLL: return self.unrollable_dims[axis]
       if op in {OptOps.GROUP, OptOps.GROUPTOP}: return self.axes_of(AxisType.REDUCE)[axis]
-      check(axis < self.shape_len, "invalid axis")
+      check(axis < self.shape_len, f"invalid axis on {axis=} {op=} {self.shape_len=}")
       return axis
     except IndexError as e: raise KernelOptError from e
 
-  def apply_opt(self, opt:Opt, append_opt:bool=True):
+  def apply_opt(self, opt:Opt, append_opt:bool=True) -> int|None:
     if self.finalized: raise RuntimeError("can't optimize Kernel after it's finalized")
     if self.dont_use_locals: check(opt.op not in {OptOps.LOCAL, OptOps.GROUP, OptOps.GROUPTOP}, "not using locals")
 
@@ -262,7 +263,7 @@ class Kernel:
       check(0 < (use_tensor_cores:=cast(tuple, opt.arg)[2]) <= 2, "use_tensor_cores value is not valid")
       check(self._apply_tc_opt(use_tensor_cores, cast(int, opt.axis), tc_select, tc_opt), "no tensor core available")
       self.applied_opts.append(opt)
-      return
+      return None
 
     axis = self.real_axis(opt.op, opt.axis)
 
@@ -285,28 +286,30 @@ class Kernel:
       smem_sz = amt*acc_sz*upcast_sz*local_sz
       check(smem_sz <= self.opts.shared_max, f"exceeds maximum shared memory size: needs {smem_sz}, max {self.opts.shared_max}")
 
+    new_axis = None
     if opt.op is OptOps.LOCAL:    # cyan
       # NOTE: LLVM/CPU can use locals too, but they are treated the same as globals (still helpful for L1 cache)
       # it's disabled for now since it makes BEAM slow for little gain
       check(self.opts.has_local, "target does not support local")
       check(self.axis_types[axis] is AxisType.GLOBAL, "local is for globals")
-      self.shift_to(axis, amt, AxisType.LOCAL, insert_at=max(self.axes_of(AxisType.GLOBAL, AxisType.LOCAL))+1)
+      new_axis = self.shift_to(axis, amt, AxisType.LOCAL, insert_at=max(self.axes_of(AxisType.GLOBAL, AxisType.LOCAL))+1)
     elif opt.op in {OptOps.GROUP, OptOps.GROUPTOP}:   # green
       check(self.opts.has_local and self.opts.has_shared, "target does not support local or shared mem")
       check(self.axis_types[axis] is AxisType.REDUCE, "must be reduce axis to group")
       check(not self.tensor_core, "can't group with tensor cores")
       check(len(reduce_axes:=[i for r in self.reduceops for i in r.axis_arg]) == len(set(reduce_axes)), "can't group with parallel reduces")
-      self.shift_to(axis, amt, AxisType.GROUP_REDUCE, top=(opt.op is OptOps.GROUPTOP), insert_at=min(self.axes_of(AxisType.REDUCE)))
+      new_axis = self.shift_to(axis, amt, AxisType.GROUP_REDUCE, top=(opt.op is OptOps.GROUPTOP), insert_at=min(self.axes_of(AxisType.REDUCE)))
     elif opt.op is OptOps.UNROLL:                     # purple
       check(self.axis_types[axis] not in (AxisType.UPCAST, AxisType.UNROLL), "can't upcasted already upcasted")
       check(amt <= 32, "don't unroll more than 32")
-      self.shift_to(axis, amt, AxisType.UNROLL, insert_at=None)
+      new_axis = self.shift_to(axis, amt, AxisType.UNROLL, insert_at=None)
     elif opt.op is OptOps.UPCAST:                     # yellow
       check(axis in self.upcastable_dims, f"{axis=} not in {self.upcastable_dims=}")
       # NOTE: assume the first get_local_axes() LOCAL are for TC
       check(not (self.tensor_core and axis in self.axes_of(AxisType.LOCAL)[:len(self.tensor_core.get_local_axes())]), "can't upcast TC locals")
       check((self.opts is not None and self.opts.device == "DSP") or amt <= 16, "don't upcast more than 16")
-      self.shift_to(axis, amt, AxisType.UPCAST, insert_at=max(self.axes_of(AxisType.GLOBAL, AxisType.LOCAL, AxisType.LOOP, AxisType.UPCAST))+1)
+      new_axis = self.shift_to(axis, amt, AxisType.UPCAST,
+                               insert_at=max(self.axes_of(AxisType.GLOBAL, AxisType.LOCAL, AxisType.LOOP, AxisType.UPCAST))+1)
     elif opt.op is OptOps.NOLOCALS:
       check(self.opts.has_local and not self.dont_use_locals, "NOLOCALS is meaningless if target does not support local or already not using locals")
       check(AxisType.LOCAL not in self.axis_types and self.group_for_reduces == 0, "can't have no locals with locals")
@@ -336,6 +339,7 @@ class Kernel:
     if append_opt: self.applied_opts.append(opt)
     if self.simplify_ones() and self.tensor_core_opts:
       self.tensor_core_opts.fix_axes(axis) # fix up axes in TC opts if required after simplify_ones()
+    return new_axis
 
   def apply_opts(self, opts:Sequence[Opt]) -> Kernel:
     for opt in opts: self.apply_opt(opt)
@@ -460,8 +464,7 @@ class Kernel:
       if op.op is Ops.REDUCE_AXIS:
         reduce_idx = len(self.bufs) + self.reduceops.index(op) * 2
         changed = tuple(i for i in range(self.shape_len) if resolve(self.sts[reduce_idx].shape[i] != self.sts[reduce_idx + 1].shape[i]))
-        axes = tuple(i for i in self.axes_of(AxisType.REDUCE, AxisType.UNROLL) if i in changed)
-        grouped_axes = tuple(i for i in self.axes_of(AxisType.GROUP_REDUCE) if i in changed)
+        axes = tuple(i for i in self.axes_of(AxisType.REDUCE, AxisType.GROUP_REDUCE, AxisType.UNROLL) if i in changed)
         if (tc := self.tensor_core) and self.use_tensor_cores == 1:
           # get reduce/upcast axes for the tensor cores
           tc_reduce_axes = self.shape_str_to_axis([f"r{i}" for i in range(len(tc.get_reduce_axes()))])
@@ -486,23 +489,6 @@ class Kernel:
           return ret.replace(src=(tc_uop,), arg=(Ops.ADD, new_axes)) if (new_axes := tuple(i for i in axes if i not in tc_reduce_axes)) else tc_uop
 
         ret = ret.replace(arg = (op.arg[0], axes))
-        if self.group_for_reduces and grouped_axes:
-          local_axes = tuple([i for i,t in enumerate(self.axis_types) if t in (AxisType.LOCAL, AxisType.UPCAST) or i in grouped_axes])
-          slocal, supcast, sgroup = sorted(self.axes_of(AxisType.LOCAL)), sorted(self.axes_of(AxisType.UPCAST)), sorted(grouped_axes)
-          # NOTE: start with UPCAST at the end so it has stride 1 and can merge
-          base_shape = tuple([self.full_shape[i] for i in slocal] + [self.full_shape[i] for i in sgroup] + [self.full_shape[i] for i in supcast])
-          permute_axes = tuple([local_axes.index(i) for i in slocal+sgroup+supcast])
-          local_shape = tuple([s if i in local_axes else 1 for i,s in enumerate(self.full_shape)])
-          local_src_shape = tuple([self.full_shape[i] if i in self.axes_of(AxisType.GLOBAL) else s for i,s in enumerate(local_shape)])
-          st = ShapeTracker.from_shape(base_shape).permute(permute_axes).reshape(local_shape).expand(local_src_shape)
-          local_size = st.real_size()
-          local_buffer = UOp(Ops.DEFINE_LOCAL, op.dtype.ptr(local_size, addrspace=AddrSpace.LOCAL), (), f"temp{self.reduceops.index(op)}")
-          local_load = local_buffer.view(st).load(local_buffer.view(st).store(ret))
-          grouped_reduce = UOp(Ops.REDUCE_AXIS, op.dtype, (local_load,), arg=(op.arg[0], grouped_axes))
-          if op is self.reduceops[-1]: return grouped_reduce
-          st = ShapeTracker.from_shape(tuple([1 if i in grouped_axes else s for i,s in enumerate(local_shape)]))
-          return local_buffer.view(st).load(local_buffer.view(st).store(grouped_reduce))
-
       return ret
     self.finalized = True
     fixed_ast = fixup_ast(self.ast)

tinygrad/codegen/opt/postrange.py

@@ -0,0 +1,18 @@
+from dataclasses import replace
+from tinygrad.uop.ops import PatternMatcher, UPat, Ops, UOp, KernelInfo
+from tinygrad.helpers import colored
+from tinygrad.codegen.opt.kernel import axis_colors
+
+def rename_sink(s:UOp):
+  if s.arg is not None and s.arg.name != "test": return None
+
+  # get all ranges (sorted)
+  rngs = sorted([u for u in s.parents if u.op is Ops.RANGE], key=lambda x: x.arg[0:-1])
+
+  # add name to kernel
+  name = "k" + colored('_', 'BLACK').join(['']+[colored(x.src[0].render(), axis_colors[x.arg[-1]]) for x in rngs])
+  return s.replace(arg=KernelInfo(name=name) if s.arg is None else replace(s.arg, name=name))
+
+pm_postrange_opt = PatternMatcher([
+  (UPat(Ops.SINK, name="s"), rename_sink),
+])

tinygrad/codegen/opt/search.py

@@ -1,5 +1,5 @@
-from typing import cast, Callable
-import itertools, functools, random, math, time, multiprocessing, traceback, signal, atexit
+from typing import cast
+import functools, math, time, multiprocessing, traceback, signal, atexit
 from collections import defaultdict
 from dataclasses import replace
 from tinygrad.uop.ops import UOp, Ops, Variable, sym_infer, AxisType
@@ -201,15 +201,3 @@ def beam_search(lin:Kernel, rawbufs:list[Buffer], amt:int, allow_test_size=True,
   if CACHELEVEL >= 1: diskcache_put("beam_search", key, beam[0][0].applied_opts)
   if BEAM_DEBUG: print(f"BEAM_SEARCH: final tm={time_to_str(beam[0][1], w=0)}, applied_opts={beam[0][0].applied_opts}")
   return beam[0][0]
-
-def optimize_local_size(_prg:Callable, global_size:list[int], rawbufs:list[Buffer]) -> list[int]:
-  test_rawbuffers = [Buffer(rawbufs[0].device, rawbufs[0].size, rawbufs[0].dtype).allocate(), *rawbufs[1:]] if rawbufs[0] in rawbufs[1:] else rawbufs
-  MAX_WORKGROUP = 1024
-  local_dims = [[x for x in set([sz, 1, 2, 4, 8, 16, 32, 64, 128, 256, MAX_WORKGROUP]) if x<=sz] for sz in global_size]
-  local_sizes = [list(x) for x in itertools.product(*local_dims) if prod(x) <= MAX_WORKGROUP] * 2  # try each valid size twice
-  def try_exec(local_size):
-    try: return _prg(*[x._buf for x in test_rawbuffers], global_size=[g//l if g%l == 0 else g/l for g,l in zip(global_size, local_size)], local_size=local_size, wait=True)  # noqa: E501
-    except Exception: return float('inf')
-  ret = min([(try_exec(local_size), local_size) for local_size in random.sample(local_sizes, len(local_sizes))])
-  assert not math.isinf(ret[0]), "all optimize_local_size exec failed"
-  return ret[1]

tinygrad/codegen/opt/swizzler.py

@@ -128,7 +128,8 @@ fix_kernel_ops = view_left_through_load+PatternMatcher([
   (UPat(Ops.VIEW, src=(UPat.cvar(),), name="self"),
    lambda self: UOp.where(UOp(Ops.VALID, dtypes.bool, (UOp(Ops.VIEW, arg=self.st),)), self.const_like(self.base.arg), 0)),
   # no ImageDType after index
-  (UPat(GroupOp.All-{Ops.DEFINE_GLOBAL, Ops.VIEW}, name="x"), lambda x: x.replace(dtype=x.dtype.base) if isinstance(x.dtype, ImageDType) else None),
+  (UPat(GroupOp.All-{Ops.DEFINE_GLOBAL, Ops.VIEW, Ops.INDEX}, name="x"),
+   lambda x: x.replace(dtype=x.dtype.base) if isinstance(x.dtype, ImageDType) else None),
   # if this kernel also assigns to the loaded buffer, ensure we can index it correctly
   (UPat(Ops.LOAD, src=(UPat.var("glbl").view(name="view"),)), check_load_st),
 ])

tinygrad/codegen/opt/tc.py

@@ -22,6 +22,15 @@ class TensorCore: # D = A * B + C, A is (M x K), B is (K x N), C and D are (M x
   def permutes_for_shape_str(self, shape_str:list[str]) -> tuple[tuple[int, ...], tuple[int, ...]]:
     ret = [[shape_str.index(remap[ss]) if ss in remap else i for i,ss in enumerate(shape_str)] for remap in self._remaps()]
     return tuple(ret[0]), tuple(ret[1])
+  @functools.cache  # pylint: disable=method-cache-max-size-none
+  def base_shape_str(self) -> list[str]:
+    ret = []
+    cnt = {'u': 0, 'l': 0}
+    for opt in self.opts:
+      ret.append(f"{opt[0]}{cnt[opt[0]]}")
+      cnt[opt[0]] += 1
+    # assumes you do the UNROLL after the opts
+    return ret + [f"r{i}" for i in range(len(self.get_reduce_axes()))]
   def get_reduce_axes(self): return [(i, 2) for i in range(int(math.log2(self.dims[2])))]
   def get_upcast_axes(self): return [opt for opt in self.opts if opt[0] == "u"]
   def get_local_axes(self): return [opt for opt in self.opts if opt[0] == "l"]

tinygrad/device.py

@@ -2,7 +2,7 @@ from __future__ import annotations
 from dataclasses import dataclass, replace
 from collections import defaultdict
 from typing import Any, Generic, TypeVar, Iterator
-import importlib, inspect, functools, pathlib, os, platform, contextlib, sys, re, atexit, pickle, decimal, time
+import importlib, inspect, functools, pathlib, os, platform, contextlib, sys, re, atexit, pickle, decimal
 from tinygrad.helpers import CI, OSX, LRU, getenv, diskcache_get, diskcache_put, DEBUG, GlobalCounters, flat_mv, PROFILE, temp, colored, \
                              Context, DISABLE_COMPILER_CACHE, ALLOW_DEVICE_USAGE, MAX_BUFFER_SIZE, cpu_events, ProfileEvent, ProfilePointEvent, dedup
 from tinygrad.dtype import DType, ImageDType, PtrDType, dtypes, _to_np_dtype
@@ -138,15 +138,14 @@ class Buffer:
       if not self.device.startswith("DISK"): GlobalCounters.mem_used += self.nbytes
       if PROFILE:
         self._prof_num = num = len(Buffer.profile_events)
-        ts = decimal.Decimal(time.perf_counter_ns())/1000
-        Buffer.profile_events.append(ProfilePointEvent(self.device, "alloc", ts, num, {"dtype":str(self.dtype),"sz":self.size,"nbytes":self.nbytes}))
+        Buffer.profile_events.append(ProfilePointEvent(self.device, "alloc", num, {"dtype":self.dtype, "sz":self.size}))
     return self
   def deallocate(self):
     assert hasattr(self, '_buf'), "buffer must be allocated to deallocate"
     if DEBUG is not None and DEBUG >= 7: print(f"buffer: deallocate {self.nbytes} bytes on {self.device}")
     if self._base is None and (self.options is None or self.options.external_ptr is None):
       if GlobalCounters is not None and not self.device.startswith("DISK"): GlobalCounters.mem_used -= self.nbytes
-      if PROFILE: Buffer.profile_events.append(ProfilePointEvent(self.device, "free", decimal.Decimal(time.perf_counter_ns())/1000, self._prof_num))
+      if PROFILE: Buffer.profile_events.append(ProfilePointEvent(self.device, "free", self._prof_num))
       self.allocator.free(self._buf, self.nbytes, self.options)
     elif self._base is not None: self._base.allocated_views -= 1
     del self._buf
@@ -304,7 +303,7 @@ def is_dtype_supported(dtype:DType, device:str|None=None) -> bool:
     if device == "METAL": return not CI
     if device in {"CUDA", "NV"}: return not CI and not getenv("PTX")
     if device in {"CPU", "LLVM"}: return not CI and platform.machine() in {"arm", "arm64", "aarch64", "x86_64", "amd64"}
-    return device == "AMD"
+    return device in {"AMD", "PYTHON"}
   if dtype in dtypes.fp8s:
     # not supported yet - in progress
     return False

tinygrad/dtype.py

@@ -108,17 +108,16 @@ class dtypes:
     if isinstance(val, tuple):
       assert len(val) == dtype.count, f"mismatch {val} {dtype}"
       return tuple(dtypes.as_const(x, dtype) for x in val)
-    # TODO: should truncate here
     return int(val) if dtypes.is_int(dtype) else float(val) if dtypes.is_float(dtype) else bool(val)
   @staticmethod
   @functools.cache
   def min(dtype:DType):
-    if dtypes.is_int(dtype): return 0 if dtypes.is_unsigned(dtype) else -2**(dtype.itemsize*8-1)
+    if dtypes.is_int(dtype): return 0 if dtypes.is_unsigned(dtype) else -2**(dtype.scalar().itemsize*8-1)
     return -float("inf") if dtypes.is_float(dtype) else False
   @staticmethod
   @functools.cache
   def max(dtype:DType):
-    if dtypes.is_int(dtype): return 2**(dtype.itemsize*8)-1+dtypes.min(dtype)
+    if dtypes.is_int(dtype): return 2**(dtype.scalar().itemsize*8)-1+dtypes.min(dtype)
     return float("inf") if dtypes.is_float(dtype) else True
   @staticmethod
   def finfo(dtype:DType) -> tuple[int, int]:
@@ -199,8 +198,9 @@ def can_safe_cast(dt0:DType, dt1:DType) -> bool:
   # https://numpy.org/doc/stable/reference/generated/numpy.can_cast.html
   if dt0 == dt1 or dt0 == dtypes.bool: return True
   match dt1:
-    case dtypes.double: return dt0 in (dtypes.float, dtypes.half, dtypes.bfloat16)
-    case dtypes.float: return dt0 in (dtypes.half, dtypes.bfloat16)
+    case dtypes.double: return dt0 in (dtypes.float, dtypes.half, dtypes.bfloat16,
+      dtypes.uint32, dtypes.uint16, dtypes.uint8, dtypes.int32, dtypes.int16, dtypes.int8)
+    case dtypes.float: return dt0 in (dtypes.half, dtypes.bfloat16, dtypes.uint16, dtypes.uint8, dtypes.int16, dtypes.int8)
     case dtypes.uint64: return dt0 in (dtypes.uint32, dtypes.uint16, dtypes.uint8)
     case dtypes.uint32: return dt0 in (dtypes.uint16, dtypes.uint8)
     case dtypes.int64: return dt0 in (dtypes.uint32, dtypes.uint16, dtypes.uint8, dtypes.int32, dtypes.int16, dtypes.int8)
@@ -215,15 +215,14 @@ def sum_acc_dtype(dt:DType):
   return least_upper_dtype(dt, to_dtype(getenv("SUM_DTYPE", "float32")))
 
 def truncate_fp16(x):
-  try: return struct.unpack("@e", struct.pack("@e", float(x)))[0]
+  try: return struct.unpack('e', struct.pack('e', float(x)))[0]
   except OverflowError: return math.copysign(math.inf, x)
 
-def truncate_bf16(x):
-  max_bf16 = struct.unpack('f', struct.pack('I', 0x7f7f0000))[0]
-  if abs(x) > max_bf16: return math.copysign(math.inf, x)
-  f32_int = struct.unpack('I', struct.pack('f', x))[0]
-  bf = struct.unpack('f', struct.pack('I', f32_int & 0xFFFF0000))[0]
-  return bf
+def float_to_bf16(x):
+  if not math.isfinite(x): return x
+  u = struct.unpack('I', struct.pack('f', x))[0]
+  u = (u + 0x7FFF + ((u >> 16) & 1)) & 0xFFFF0000
+  return struct.unpack('f', struct.pack('I', u))[0]
 
 # fp8-float conversions based on https://gitlab.com/nvidia/headers/cuda-individual/cudart/-/blob/main/cuda_fp8.hpp
 def float_to_fp8(x: float, dtype: DType) -> int:
@@ -288,7 +287,7 @@ def fp8_to_float(x: int, dtype: DType) -> float:
   return float(float32_val)
 
 truncate: dict[DType, Callable] = {dtypes.bool: bool,
-  dtypes.float16: truncate_fp16, dtypes.bfloat16: truncate_bf16,
+  dtypes.float16: truncate_fp16, dtypes.bfloat16: lambda x: float_to_bf16(float(x)),
   **{fp8: (lambda x, dtype=fp8: fp8_to_float(float_to_fp8(x, dtype), dtype)) for fp8 in dtypes.fp8s},
   dtypes.float32: lambda x: ctypes.c_float(x).value, dtypes.float64: lambda x: ctypes.c_double(x).value,
   dtypes.uint8: lambda x: ctypes.c_uint8(x).value, dtypes.uint16: lambda x: ctypes.c_uint16(x).value,
@@ -300,6 +299,7 @@ truncate: dict[DType, Callable] = {dtypes.bool: bool,
 
 def _to_np_dtype(dtype:DType) -> type|None:
   import numpy as np
+  if dtype == dtypes.bfloat16: return np.float32
   return np.dtype(dtype.fmt).type if dtype.fmt is not None else None
 def _from_np_dtype(npdtype:'np.dtype') -> DType: # type: ignore [name-defined] # noqa: F821
   import numpy as np
@@ -308,9 +308,11 @@ def _from_np_dtype(npdtype:'np.dtype') -> DType: # type: ignore [name-defined] #
 @functools.cache
 def _to_torch_dtype(dtype:DType) -> 'torch.dtype'|None:  # type: ignore [name-defined] # noqa: F821
   import numpy as np, torch
+  if dtype == dtypes.uint64: return torch.uint64
+  if dtype == dtypes.bfloat16: return torch.bfloat16
   # NOTE: torch doesn't expose this mapping with a stable API
   try: return torch.from_numpy(np.array([], dtype=_to_np_dtype(dtype))).dtype
   except TypeError: return None
 @functools.cache
 def _from_torch_dtype(torchdtype:'torch.dtype') -> DType: # type: ignore [name-defined] # noqa: F821
-  return {v:k for k in dtypes.all if (v:=_to_torch_dtype(k)) is not None}[torchdtype]
+  return {v:k for k in dtypes.all if (v:=_to_torch_dtype(k)) is not None}[torchdtype]

tinygrad/engine/memory.py

@@ -23,12 +23,13 @@ def _internal_memory_planner(buffers:list[list[Buffer]], noopt_buffers=None, ign
   # Sort buffer operations in timeline order. Two events: buffer is allocated or buffer is freed.
   buffer_requests = sorted([((first_appearance[buf], True), buf) for buf in first_appearance.keys()] + \
                            [((last_appearance[buf] + 1, False), buf) for buf in first_appearance.keys()], key=lambda x: x[0])
+  total_memory = sum(round_up(buf.nbytes, min_block_size:=0x1000) for buf in first_appearance.keys()) * 2 # *2 for fragmentation (which is about 15%)
 
   # Try to suballocate from a shared buffer managed by global_planner using TLSFAllocator.
   # Also track buffer replacements for buffers that do not support suballocation.
   buffer_replace:dict[Buffer, tuple[Buffer|None, int|None]] = {}
   reuse_buffers:dict[tuple, list[Buffer]] = defaultdict(list)
-  global_planner:dict[str, tuple[int, TLSFAllocator]] = defaultdict(lambda: (0, TLSFAllocator(1 << 44, block_size=0x1000, lv2_cnt=32)))
+  global_planner:dict[str, tuple[int, TLSFAllocator]] = defaultdict(lambda: (0, TLSFAllocator(total_memory, block_size=min_block_size, lv2_cnt=32)))
   for (_, is_open_ev), buf in buffer_requests:
     # Check if suballocation is possible for the given buffer and device.
     if hasattr(Device[buf.device].allocator, "_offset") and not isinstance(buf.dtype, ImageDType):

tinygrad/engine/realize.py

@@ -1,8 +1,8 @@
-from typing import cast, Generator
-import time, pprint, decimal
+from typing import cast, Generator, Callable
+import time, pprint, random, itertools, math
 from dataclasses import dataclass, replace, field
 from tinygrad.helpers import all_same, colored, DEBUG, GlobalCounters, ansilen, BEAM, NOOPT, all_int, CAPTURING, Metadata, TRACEMETA, TracingKey
-from tinygrad.helpers import DEVECTORIZE, time_to_str, VALIDATE_WITH_CPU, getenv, cpu_profile, PROFILE, ProfilePointEvent, cpu_events
+from tinygrad.helpers import DEVECTORIZE, time_to_str, VALIDATE_WITH_CPU, getenv, cpu_profile, PROFILE, ProfilePointEvent, cpu_events, prod
 from tinygrad.uop.ops import Ops, PatternMatcher, UOp, UPat, Variable, sym_infer, graph_rewrite, print_uops, track_rewrites, KernelInfo
 from tinygrad.device import Device, Buffer
 from tinygrad.renderer import Renderer, ProgramSpec, Estimates
@@ -34,8 +34,9 @@ def get_program(ast:UOp, renderer:Renderer|None=None, opts:list[Opt]|None=None)
     ast = ast.replace(arg=KernelInfo(opts_to_apply=tuple(opts)))
   try:
     uops = full_rewrite(ast, renderer)
-  except RuntimeError:
+  except RuntimeError as e:
     print("***** LINEARIZE FAILURE *****")
+    print(e)
     print(f"ast = {ast}")
     raise
   assert uops[-1].op is Ops.SINK, "last uop must be sink"
@@ -59,13 +60,27 @@ class Runner:
   def __call__(self, rawbufs:list[Buffer], var_vals:dict[Variable, int], wait=False) -> float|None:
     raise NotImplementedError("override this")
 
+def optimize_local_size(_prg:Callable, global_size:list[int], rawbufs:list[Buffer]) -> list[int]:
+  test_rawbuffers = [Buffer(rawbufs[0].device, rawbufs[0].size, rawbufs[0].dtype).allocate(), *rawbufs[1:]] if rawbufs[0] in rawbufs[1:] else rawbufs
+  MAX_WORKGROUP = 1024
+  local_dims = [[x for x in set([sz, 1, 2, 4, 8, 16, 32, 64, 128, 256, MAX_WORKGROUP]) if x<=sz] for sz in global_size]
+  local_sizes = [list(x) for x in itertools.product(*local_dims) if prod(x) <= MAX_WORKGROUP] * 2  # try each valid size twice
+  def try_exec(local_size):
+    try:
+      return _prg(*[x._buf for x in test_rawbuffers],global_size=[g//l if g%l == 0 else g/l for g,l in zip(global_size, local_size)],
+                  local_size=local_size, wait=True)
+    except Exception: return float('inf')
+  ret = min([(try_exec(local_size), local_size) for local_size in random.sample(local_sizes, len(local_sizes))])
+  assert not math.isinf(ret[0]), "all optimize_local_size exec failed"
+  return ret[1]
+
 class CompiledRunner(Runner):
   def __init__(self, p:ProgramSpec, precompiled:bytes|None=None, prg=None):
     if DEBUG >= 4: print(p.src)
     self.p:ProgramSpec = p
     if precompiled is not None: self.lib = precompiled
     else:
-      with cpu_profile(TracingKey(f"compile {p.name}", (p.function_name,), cat="compiler"), "TINY"):
+      with cpu_profile(TracingKey(f"compile {p.name}", (p.function_name,)), "TINY"):
         self.lib = Device[p.device].compiler.compile_cached(p.src)
     if DEBUG >= 7: Device[p.device].compiler.disassemble(self.lib)
     self._prg = Device[p.device].runtime(p.function_name, self.lib) if prg is None else prg
@@ -76,8 +91,6 @@ class CompiledRunner(Runner):
   def __call__(self, rawbufs:list[Buffer], var_vals:dict[Variable, int], wait=False) -> float|None:
     global_size, local_size = self.p.launch_dims(var_vals)
     if global_size is not None and local_size is None and all_int(self.p.global_size): # type: ignore[arg-type]
-      # TODO: this is copied from get_program
-      from tinygrad.codegen.opt.search import optimize_local_size
       local_size = optimize_local_size(self._prg, global_size, rawbufs)
       global_size = [g//l if g%l == 0 else g/l for g,l in zip(global_size, local_size)]
       self.p = replace(self.p, global_size=global_size, local_size=local_size)
@@ -149,8 +162,7 @@ class ExecItem:
   def run(self, _var_vals:dict[Variable, int]|None=None, wait=False, jit=False, do_update_stats=True) -> float|None:
     var_vals = self.fixedvars if _var_vals is None else (_var_vals|self.fixedvars)
     bufs = [cast(Buffer, x) for x in self.bufs] if jit else [cast(Buffer, x).ensure_allocated() for x in self.bufs]
-    if PROFILE: cpu_events.append(ProfilePointEvent(self.prg.device, "exec", decimal.Decimal(time.perf_counter_ns())/1000, self.prg.display_name,
-                                                    {"metadata":self.metadata, "var_vals":var_vals}))
+    if PROFILE: cpu_events.append(ProfilePointEvent(self.prg.device, "exec", self.prg.display_name, {"metadata":self.metadata, "var_vals":var_vals}))
     et = self.prg(bufs, var_vals, wait=wait or DEBUG >= 2)
     if do_update_stats:
       GlobalCounters.kernel_count += 1
@@ -160,10 +172,15 @@ class ExecItem:
       if DEBUG >= 2:
         lds_est = sym_infer(self.prg.estimates.lds, var_vals)
         mem_est = min(mem_est, lds_est)   # there can't be more memory accessed than loads/stores. remove this when symbolic is fixed
+        header_color = 'magenta' if jit else ('green' if self.prg.first_run else None)
         ptm = colored(time_to_str(et, w=9), "yellow" if et > 0.01 else None) if et is not None else ""
-        print(f"{colored(f'*** {self.prg.device[:7]:7s} {GlobalCounters.kernel_count:4d}', 'magenta' if jit else ('green' if self.prg.first_run else None))} {self.prg.display_name+' '*(44-ansilen(self.prg.display_name))} arg {len(bufs):2d} mem {GlobalCounters.mem_used/1e9:5.2f} GB " +  # noqa: E501
-              (str() if et is None else f"tm {ptm}/{GlobalCounters.time_sum_s*1e3:9.2f}ms ({op_est/((et or 1e-20)*1e9):9.2f} GFLOPS {mem_est/((et or 1e-20)*1e9):6.1f}|{lds_est/((et or 1e-20)*1e9):<7.1f} GB/s)" +  # noqa: E501
-               f" {[repr(m) if TRACEMETA >= 2 else str(m) for m in self.metadata] if self.metadata else ''}"))
+        flops, membw, ldsbw = op_est/(et or 1e-20), mem_est/(et or 1e-20), lds_est/(et or 1e-20)
+        flops_str = f"{flops*1e-9:9.2f} GFLOPS" if flops < 1e14 else colored(f"{flops*1e-12:9.2f} TFLOPS", 'green')
+        mem_str = f"{membw*1e-9:6.1f}|{ldsbw*1e-9:<7.1f} GB/s" if membw < 1e13 else colored(f"{membw*1e-12:6.1f}|{ldsbw*1e-12:<7.1f} TB/s", 'green')
+        print(f"{colored(f'*** {self.prg.device[:7]:7s} {GlobalCounters.kernel_count:4d}', header_color)}"+
+          f" {self.prg.display_name+' '*(44-ansilen(self.prg.display_name))} arg {len(bufs):2d} mem {GlobalCounters.mem_used/1e9:5.2f} GB"+
+          ("" if et is None else f" tm {ptm}/{GlobalCounters.time_sum_s*1e3:9.2f}ms ({flops_str} {mem_str})")+
+          f" {[repr(m) if TRACEMETA >= 2 else str(m) for m in self.metadata] if self.metadata else ''}")
       self.prg.first_run = False
     return et
 

tinygrad/frontend/onnx.py

@@ -21,9 +21,9 @@ class AttributeType(enum.IntEnum):
   ONNX attribute type identifiers.
   Reference: https://github.com/onnx/onnx/blob/rel-1.18.0/onnx/onnx.proto3#L128-L145
   """
-  FLOAT = 1; INT = 2; STRING = 3; TENSOR = 4; FLOATS = 6; INTS = 7; STRINGS = 8 # noqa: E702
+  FLOAT = 1; INT = 2; STRING = 3; TENSOR = 4; GRAPH = 5; FLOATS = 6; INTS = 7; STRINGS = 8 # noqa: E702
 
-  def to_field_name(self) -> str: return {1: "f", 2: "i", 3: "s", 4: "t", 6: "floats", 7: "ints", 8: "strings"}[self.value]
+  def to_field_name(self) -> str: return {1: "f", 2: "i", 3: "s", 4: "t", 5: "g", 6: "floats", 7: "ints", 8: "strings"}[self.value]
 
 class OnnxDataType(enum.IntEnum):
   """
@@ -266,6 +266,7 @@ class OnnxPBParser:
         case 3: obj["i"] = self.reader.read_int64()
         case 4: obj["s"] = self.reader.read_bytes().data().tobytes().decode("utf8")
         case 5: obj["t"] = self._parse_TensorProto()['parsed_tensor']
+        case 6: obj["g"] = OnnxRunner._from_subgraph(self._parse_GraphProto())
         case 7: obj["floats"].append(self.reader.read_float())
         case 8: obj["ints"].append(self.reader.read_int64())
         case 9: obj["strings"].append(self.reader.read_bytes().data().tobytes().decode("utf8"))
@@ -401,8 +402,11 @@ class OnnxRunner:
   """
   def __init__(self, model_path: Tensor | str | pathlib.Path):
     model = OnnxPBParser(model_path, load_external_data=True).parse()
-    graph = model["graph"]
+    self._init_from_graph(model["graph"])
+
+  def _init_from_graph(self, graph: dict, is_subgraph: bool = False):
     self.is_training = any(n['parsed_node'].opset_id.domain in {Domain.AI_ONNX_TRAINING, Domain.AI_ONNX_PREVIEW_TRAINING} for n in graph["node"])
+    self.graph_name = graph["name"] if is_subgraph else ""
     self.graph_values = {"": None, **{i["name"]: i["parsed_tensor"] for i in graph["initializer"]}}
     self.graph_inputs = {i["name"]: i["parsed_type"] for i in graph["input"] if i["name"] not in self.graph_values}
     self.graph_outputs = tuple(o["name"] for o in graph["output"])
@@ -414,6 +418,12 @@ class OnnxRunner:
     self.variable_dims: dict[str, int] = {}
     self.onnx_ops = onnx_ops
 
+  @classmethod
+  def _from_subgraph(cls, graph: dict) -> "OnnxRunner":
+    subgraph = cls.__new__(cls)
+    subgraph._init_from_graph(graph, is_subgraph=True)
+    return subgraph
+
   def _parse_input(self, name: str, value: Any, spec: OnnxValue):
     if spec.is_optional and value is None: return None
     if spec.is_sequence:
@@ -445,9 +455,10 @@ class OnnxRunner:
     return {name:Tensor.empty(*spec.shape, device=device, dtype=dtype or spec.dtype) for name, spec in self.graph_inputs.items()}
 
   def to(self, device:str|None):
-    self.graph_values = {k:v.to(device) if isinstance(v, Tensor) else v for k,v in self.graph_values.items()}
+    self.graph_values = {k: (v.to(device) if isinstance(v, Tensor) else v) for k,v in self.graph_values.items()}
     self.graph_nodes = tuple(OnnxNode(n.op, n.opset_id, tuple(n.inputs), tuple(n.outputs),
-                                      {k:v.to(device) if isinstance(v, Tensor) else v for k,v in n.opts.items()}) for n in self.graph_nodes)
+                                      {k: (v.to(device) if isinstance(v, (Tensor, OnnxRunner)) else v) for k,v in n.opts.items()})
+                                      for n in self.graph_nodes)
     return self
 
   def __call__(self, inputs:dict[str, Any], debug=debug):
@@ -461,9 +472,9 @@ class OnnxRunner:
 
       # provide additional opts
       if node.op == "Split" and 'num_outputs' not in opts: opts['num_outputs'] = len(node.outputs)
-      if node.op == "Gradient": opts['intermediate_tensors'] = self.graph_values
+      if node.op in {"Gradient", "If"}: opts['intermediate_tensors'] = self.graph_values
 
-      if debug >= 1: print(f"{num}: op '{node.op}' opt {opts}")
+      if debug >= 1: print((f"[{self.graph_name}] " if self.graph_name else "") + f"{num}: op '{node.op}' opt {opts}")
       if debug >= 2 and node.inputs: print("\tinputs:\n" + "\n".join(f"\t\t{x} - {i!r}" for x,i in zip(node.inputs, inps)))
       ret = self._select_op(node.op, node.opset_id)(*inps, **opts)
       ret = ret if isinstance(ret, tuple) else (ret,)
@@ -543,6 +554,23 @@ def get_onnx_ops() -> dict[str, types.FunctionType|dict[OpSetId, types.FunctionT
     return __decorator
 
   # ***** Property/Graph Ops *****
+  def If(condition:Tensor, else_branch:OnnxRunner, then_branch:OnnxRunner, intermediate_tensors:dict[str, Tensor]):
+    def run_branch(branch:OnnxRunner):
+      branch.graph_values.update(intermediate_tensors)
+      out = branch({k:intermediate_tensors[k] for k in branch.graph_inputs.keys()})
+      # dereference intermediate tensors so Buffer can be deallocated
+      for k in intermediate_tensors: del branch.graph_values[k]
+      return out
+    # both branch must be ran before the condition can be evaluated
+    else_out, then_out = run_branch(else_branch), run_branch(then_branch)
+    assert len(else_out) == len(then_out), f"else_out and then_out must have the same number of outputs: {len(else_out)} != {len(then_out)}"
+    # can use where op when output shape is the same
+    if all(t.shape == e.shape for t,e in zip(then_out.values(), else_out.values())):
+      return tuple(condition.where(t,e) for t,e in zip(then_out.values(), else_out.values()))
+    # otherwise, use condition to select the output in python
+    cond = _resolve_const(_cached_to_python_const(condition))
+    return tuple(t if cond else e for t,e in zip(then_out.values(), else_out.values()))
+
   def Identity(x:Tensor): return x
   def Constant(sparse_value:Tensor|None=None, value:Tensor|None=None, value_float:float|None=None, value_floats:list[float]|None=None,
               value_int:int|None=None, value_ints:list[int]|None=None, value_string:str|None=None, value_strings:list[str]|None=None):

tinygrad/gradient.py

@@ -22,11 +22,10 @@ pm_gradient = PatternMatcher([
   (UPat(Ops.SQRT, name="ret"), lambda ctx, ret: (ctx / (ret*2),)),
   (UPat((Ops.CMPLT, Ops.CMPNE)), lambda: (None, None)),
   (UPat(Ops.ADD), lambda ctx: (ctx, ctx)),
-  (UPat(Ops.POW, name="ret"), lambda ctx, ret:
-   (ctx*(ret.src[0].eq(0) & ret.src[1].eq(0)).where(ret.src[1], ret.src[1]*ret.src[0].pow(ret.src[1]-1)),
-    ctx*ret.src[0].eq(0).where((ret.src[1]<0).where(ret.const_like(-math.inf), ret.const_like(0)), ret*ret.src[0].log2()*math.log(2.0)))),
-  (UPat(Ops.MAX, name="ret"), lambda ctx, ret: ((ret.src[0]>ret.src[1]).where(ctx, (ret.src[0]!=ret.src[1]).where(ctx.const_like(0), ctx * 0.5)),
-                                                (ret.src[0]<ret.src[1]).where(ctx, (ret.src[0]!=ret.src[1]).where(ctx.const_like(0), ctx * 0.5)))),
+  (UPat(Ops.POW, name="ret", src=(UPat.var("b"), UPat.var("e"))), lambda ctx, ret, b, e:
+    (ctx * (b.eq(0)&e.eq(0)).where(e, e*b.pow(e-1)), ctx * b.eq(0).where((e<0).where(ret.const_like(-math.inf), 0), ret*b.log2()*math.log(2.0)))),
+  (UPat(Ops.MAX, name="ret", src=(UPat.var("x"), UPat.var("y"))), lambda ctx, ret, x, y:
+    ((x>y).where(ctx, (x.eq(y)).where(ctx * 0.5, 0)), (x<y).where(ctx, (x.eq(y)).where(ctx * 0.5, 0)))),
   (UPat(Ops.MUL, name="ret"), lambda ctx, ret: (ret.src[1]*ctx, ret.src[0]*ctx)),
   (UPat(Ops.WHERE, name="ret"), lambda ctx, ret: (None, ret.src[0].where(ctx, ctx.const_like(0)), ret.src[0].where(ctx.const_like(0), ctx))),
   (UPat(Ops.REDUCE_AXIS, name="ret"), reduce_gradient),

tinygrad/helpers.py

@@ -135,12 +135,12 @@ FUSE_ARANGE, FUSE_CONV_BW = ContextVar("FUSE_ARANGE", 1), ContextVar("FUSE_CONV_
 SPLIT_REDUCEOP, NO_MEMORY_PLANNER, RING = ContextVar("SPLIT_REDUCEOP", 1), ContextVar("NO_MEMORY_PLANNER", 0), ContextVar("RING", 1)
 PICKLE_BUFFERS, PROFILE, LRU = ContextVar("PICKLE_BUFFERS", 1), ContextVar("PROFILE", getenv("VIZ")), ContextVar("LRU", 1)
 CACHELEVEL, IGNORE_BEAM_CACHE, DEVECTORIZE = ContextVar("CACHELEVEL", 2), ContextVar("IGNORE_BEAM_CACHE", 0), ContextVar("DEVECTORIZE", 1)
-DISABLE_COMPILER_CACHE = ContextVar("DISABLE_COMPILER_CACHE", 0)
+DISABLE_COMPILER_CACHE, BLOCK_REORDER = ContextVar("DISABLE_COMPILER_CACHE", 0), ContextVar("BLOCK_REORDER", 1)
 DONT_REALIZE_EXPAND, DONT_GROUP_REDUCES = ContextVar("DONT_REALIZE_EXPAND", 0), ContextVar("DONT_GROUP_REDUCES", 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, AMD_LLVM = ContextVar("ALLOW_DEVICE_USAGE", 1), ContextVar("MAX_BUFFER_SIZE", 0), ContextVar("AMD_LLVM", 1)
-RANGEIFY, PARTIAL_CONTIG = ContextVar("RANGEIFY", 0), ContextVar("PARTIAL_CONTIG", 0)
+RANGEIFY, POSTOPT, FUSE_ATTENTION = ContextVar("RANGEIFY", 0), ContextVar("POSTOPT", 0), ContextVar("FUSE_ATTENTION", 0)
 
 @dataclass(frozen=True)
 class Metadata:
@@ -192,12 +192,12 @@ class Profiling(contextlib.ContextDecorator):
               colored(_format_fcn(fcn).ljust(50), "yellow"),
               colored(f"<- {(scallers[0][1][2]/tottime)*100:3.0f}% {_format_fcn(scallers[0][0])}", "BLACK") if scallers else '')
 
+def perf_counter_us() -> decimal.Decimal: return decimal.Decimal(time.perf_counter_ns())/1000
 
 @dataclass(frozen=True)
 class TracingKey:
   display_name:str                       # display name of this trace event
-  keys:tuple[str, ...]=()                # optional keys to search for related traces
-  cat:str|None=None                      # optional category to color this by
+  keys:tuple[Any, ...]=()                # optional keys to search for related traces
   ret:Any=None
 
 class ProfileEvent: pass
@@ -206,15 +206,16 @@ class ProfileEvent: pass
 class ProfileRangeEvent(ProfileEvent): device:str; name:str|TracingKey; st:decimal.Decimal; en:decimal.Decimal|None=None; is_copy:bool=False # noqa: E702
 
 @dataclass(frozen=True)
-class ProfilePointEvent(ProfileEvent): device:str; name:str; ts:decimal.Decimal; key:Any; arg:dict=field(default_factory=dict) # noqa: E702
+class ProfilePointEvent(ProfileEvent): device:str; name:str; key:Any; arg:dict=field(default_factory=dict); \
+    ts:decimal.Decimal=field(default_factory=perf_counter_us) # noqa: E702
 
 cpu_events:list[ProfileEvent] = []
 @contextlib.contextmanager
 def cpu_profile(name:str|TracingKey, device="CPU", is_copy=False, display=True) -> Generator[ProfileRangeEvent, None, None]:
-  res = ProfileRangeEvent(device, name, decimal.Decimal(time.perf_counter_ns()) / 1000, is_copy=is_copy)
+  res = ProfileRangeEvent(device, name, perf_counter_us(), is_copy=is_copy)
   try: yield res
   finally:
-    res.en = decimal.Decimal(time.perf_counter_ns()) / 1000
+    res.en = perf_counter_us()
     if PROFILE and display: cpu_events.append(res)
 
 # *** universal database cache ***

tinygrad/nn/state.py

@@ -274,9 +274,9 @@ def ggml_data_to_tensor(t: Tensor, n: int, ggml_type: int) -> Tensor:
   Converts ggml tensor data to a tinygrad tensor.
 
   Supported native types: float32 (id: 0), float16 (id: 1), int8 (id: 16), int16 (id: 17), int32 (id: 18)
-  Supported quantized types: Q4_0 (id: 2), Q4_1 (id: 3), Q8_0 (id: 8), Q6_K (id: 14)
+  Supported quantized types: Q4_0 (id: 2), Q4_1 (id: 3), Q8_0 (id: 8), Q6_K (id: 14), MXFP4 (id: 39)
   """
-  # https://github.com/ggerganov/ggml/blob/6dccc647264f5429df2624f36138f601e7ce23e5/include/ggml.h#L356
+  # https://github.com/ggerganov/ggml/blob/323951f1bdcdfbd5b5ff3a9a7c3770e63b1a560e/include/ggml.h#L356
 
   # native types
   if (dtype := { 0: dtypes.float32, 1: dtypes.float16, 16: dtypes.int8, 17: dtypes.int16, 18: dtypes.int32 }.get(ggml_type)) is not None:
@@ -288,7 +288,7 @@ def ggml_data_to_tensor(t: Tensor, n: int, ggml_type: int) -> Tensor:
     return t.unsqueeze(-1).expand((*t.shape,8//b)).idiv(shift_tensor).bitwise_and(bitmask).transpose(-1, -2).flatten(-2)
 
   # map to (number of elements, number of bytes)
-  if (nelements_nbytes := { 2: (32, 18), 3: (32, 20), 14: (256, 210), 8: (32, 34) }.get(ggml_type)) is not None:
+  if (nelements_nbytes := { 2: (32, 18), 3: (32, 20), 14: (256, 210), 8: (32, 34), 39: (32, 17) }.get(ggml_type)) is not None:
     blocks = t[:(n//nelements_nbytes[0])*nelements_nbytes[1]].reshape((-1, nelements_nbytes[1]))
     if ggml_type == 2: return (q_to_uint8(blocks[:,2:], 4).bitcast(dtypes.int8) - 8) * blocks[:,:2].bitcast(dtypes.float16).cast(dtypes.float32)
     if ggml_type == 3:
@@ -300,6 +300,17 @@ def ggml_data_to_tensor(t: Tensor, n: int, ggml_type: int) -> Tensor:
       scales = blocks[:,192:208].bitcast(dtypes.int8).unsqueeze(-1).expand((-1, 16, 16)).reshape((-1, 256))
       d = blocks[:,-2:].bitcast(dtypes.float16).cast(dtypes.float32).expand((-1, 256))
       return d * (xl.bitwise_or(xh).bitcast(dtypes.int8) - 32).flatten(-2) * scales
+    if ggml_type == 39:
+      e_int = blocks[:, 0].cast(dtypes.int32)
+      d = ((e_int >= 2).cast(dtypes.float32) * (e_int.cast(dtypes.float32) - 128).exp2() +
+           (e_int == 1).cast(dtypes.float32) * 2.0**(-127) +
+           (e_int == 0).cast(dtypes.float32) * 2.0**(-128)).unsqueeze(-1)
+      codes = q_to_uint8(blocks[:, 1:17], 4)
+      sign = 1.0 - codes.rshift(3).cast(dtypes.float32) * 2.0
+      exp, mant = codes.rshift(1).bitwise_and(0x3).cast(dtypes.float32), codes.bitwise_and(0x1).cast(dtypes.float32)
+      fp4_val = sign * ((exp != 0).cast(dtypes.float32) * (1.0 + 0.5 * mant) * (exp - 1.0).exp2() +
+                        (exp == 0).cast(dtypes.float32) * 0.5 * mant)
+      return (fp4_val * d).flatten(-2)[:n]
   raise ValueError(f"GGML type '{ggml_type}' is not supported!")
 
 @accept_filename

tinygrad/renderer/__init__.py

@@ -39,7 +39,7 @@ class Estimates:
         buf = u
         while len(buf.src): buf = buf.src[0]
         if buf.op is Ops.DEFINE_GLOBAL: # assume all DEFINE_GLOBAL memory is accessed
-          mem[(buf, u.op)] = cast(PtrDType, buf.dtype).size * buf.dtype.itemsize
+          mem[(buf, u.op)] = buf.ptrdtype.size * buf.dtype.itemsize
       if u.op is Ops.RANGE:
         mult_stack.append(mults)
         mults *= cast(sint, u.src[0].ssimplify())

tinygrad/renderer/cstyle.py

@@ -2,11 +2,11 @@ from typing import Literal, Callable, cast
 import os, math, sys
 from collections import defaultdict, Counter
 from tinygrad.codegen.opt import tc
-from tinygrad.uop.ops import GroupOp, Ops, UOp, PatternMatcher, UPat
+from tinygrad.uop.ops import GroupOp, Ops, UOp, PatternMatcher, UPat, sint_to_uop
 from tinygrad.helpers import strip_parens, getenv, prod, dedup, AMX
 from tinygrad.dtype import ImageDType, dtypes, DType, PtrDType, AddrSpace, truncate
 from tinygrad.renderer import Renderer
-from tinygrad.codegen.devectorizer import no_vectorized_alu
+from tinygrad.codegen.late.devectorizer import no_vectorized_alu
 
 base_rewrite = PatternMatcher([
   (UPat(Ops.DEFINE_REG, name="x"), lambda ctx,x: f"{ctx.render_dtype(x.dtype.base)} {ctx[x]}[{x.dtype.size}];"),
@@ -26,7 +26,7 @@ base_rewrite = PatternMatcher([
   (UPat(Ops.DEFINE_LOCAL, name="x"), lambda ctx,x: f"{ctx.smem_align}{ctx.smem_prefix}{ctx.render_dtype(x.dtype.base)} {ctx[x]}[{x.dtype.size}];"),
   (UPat(Ops.BARRIER), lambda ctx: ctx.barrier),
   (UPat(Ops.PRECAST, name="x"), lambda ctx,x: ctx[x.src[0]]),
-  (UPat(Ops.SPECIAL, name="x"), lambda ctx,x: f"{ctx.code_for_workitem[x.arg[0][0]](x.arg[0][-1])}; /* {x.arg[1]} */"),
+  (UPat(Ops.SPECIAL, name="x"), lambda ctx,x: f"{ctx.code_for_workitem[x.arg[0][0]](x.arg[0][-1])}; /* {sint_to_uop(x.arg[1]).render()} */"),
   # const
   (UPat(Ops.CONST, arg=math.inf, name="x"), lambda ctx, x: f"({ctx.render_cast(x.dtype, ctx.infinity)})"),
   (UPat(Ops.CONST, arg=-math.inf, name="x"), lambda ctx, x: f"({ctx.render_cast(x.dtype, f'-{ctx.infinity}')})"),
@@ -145,7 +145,7 @@ class CStyleLanguage(Renderer):
         if u.arg is not None: name = u.arg.function_name
         continue
       if u.op in (Ops.DEFINE_GLOBAL, Ops.DEFINE_VAR):
-        r[u] = (f"data{u.arg}_{sz}" if (sz:=cast(PtrDType, u.dtype).size) > 0 else f"data{u.arg}") if u.op is Ops.DEFINE_GLOBAL else u.arg[0]
+        r[u] = (f"data{u.arg}_{sz}" if (sz:=u.ptrdtype.size) > 0 else f"data{u.arg}") if u.op is Ops.DEFINE_GLOBAL else u.arg[0]
         bufs[u] = (r[u], (u.dtype, False))
         continue
 
@@ -157,7 +157,7 @@ class CStyleLanguage(Renderer):
       # naming
       prefix = None
       if u.op is Ops.SPECIAL: r[u] = u.arg[0]
-      elif u.op is Ops.RANGE: r[u] = f"ridx{u.arg}"
+      elif u.op is Ops.RANGE: r[u] = "ridx"+'_'.join([str(x) if x >= 0 else "m"+str(-x) for x in u.arg[0:-1]])
       else:
         prefix = {Ops.WMMA: "wmma", Ops.DEFINE_LOCAL: "temp", Ops.CONST: "const",
                   Ops.CAST: "cast", Ops.BITCAST: "cast", Ops.GEP: "gep", Ops.VECTORIZE: "cast", Ops.PRECAST: "precast",
@@ -169,7 +169,7 @@ class CStyleLanguage(Renderer):
 
       if u.op in {Ops.ENDIF, Ops.ENDRANGE}: depth -= 1
       if (u.op is not Ops.CAST or u.dtype.vcount == 1) and (u.op in {Ops.CONST, Ops.GEP, Ops.INDEX, Ops.CUSTOMI} or \
-        (u.op is Ops.LOAD and cast(PtrDType, u.src[0].dtype).addrspace == AddrSpace.REG) or \
+        (u.op is Ops.LOAD and u.src[0].ptrdtype.addrspace == AddrSpace.REG) or \
         (u.op is Ops.CAST and isinstance(u.dtype, PtrDType)) or \
         (u.op in {Ops.VECTORIZE, *(GroupOp.ALU-{Ops.WHERE}), Ops.CAST, Ops.BITCAST} and child_count[u] == 1 and not getenv("EXPAND_SSA"))):
         r[u] = l

tinygrad/renderer/llvmir.py

@@ -101,13 +101,13 @@ base_rewrite = PatternMatcher([
 
   # range
   (UPat(Ops.RANGE, name="x"), lambda ctx,x:
-   f"  br label %loop_entry_{x.arg}\nloop_entry_{x.arg}:\n"
-   f"  br label %loop_body_{x.arg}\nloop_body_{x.arg}:\n"
-   f"  {ctx[x]} = phi {ldt(x.dtype)} [ 0, %loop_entry_{x.arg} ], [ {ctx[x]}phi, %loop_latch_{x.arg} ]"),
+   f"  br label %loop_entry_{x.arg[0]}\nloop_entry_{x.arg[0]}:\n"
+   f"  br label %loop_body_{x.arg[0]}\nloop_body_{x.arg[0]}:\n"
+   f"  {ctx[x]} = phi {ldt(x.dtype)} [ 0, %loop_entry_{x.arg[0]} ], [ {ctx[x]}phi, %loop_latch_{x.arg[0]} ]"),
   (UPat(Ops.ENDRANGE, name="x"), lambda ctx,x:
-   f"  br label %loop_latch_{x.src[0].arg}\nloop_latch_{x.src[0].arg}:\n"
+   f"  br label %loop_latch_{x.src[0].arg[0]}\nloop_latch_{x.src[0].arg[0]}:\n"
    f"  {ctx[x.src[0]]}phi = add i32 {ctx[x.src[0]]}, 1\n  {ctx[x]} = icmp ult i32 {ctx[x.src[0]]}phi, {ctx[x.src[0].src[0]]}\n"
-   f"  br i1 {ctx[x]}, label %loop_body_{x.src[0].arg}, label %loop_exit_{x.src[0].arg}\nloop_exit_{x.src[0].arg}:"),
+   f"  br i1 {ctx[x]}, label %loop_body_{x.src[0].arg[0]}, label %loop_exit_{x.src[0].arg[0]}\nloop_exit_{x.src[0].arg[0]}:"),
 
   # if
   (UPat(Ops.IF, name="x"), lambda ctx,x: f"  br i1 {ctx[x.src[0]]}, label %ifbody_{ctx[x][1:]}, label %ifskip_{ctx[x][1:]}\nifbody_{ctx[x][1:]}:"),
@@ -196,14 +196,20 @@ class LLVMRenderer(Renderer):
 barrier = 'fence syncscope("workgroup") release\ntail call void @llvm.amdgcn.s.barrier()\nfence syncscope("workgroup") acquire\n'
 code_for_workitem = {"g": lambda x: f"tail call i32 @llvm.amdgcn.workgroup.id.{chr(120+int(x))}()",
                      "l": lambda x: f"tail call i32 @llvm.amdgcn.workitem.id.{chr(120+int(x))}()"}
+# https://rocm.docs.amd.com/projects/llvm-project/en/latest/LLVM/llvm/html/AMDGPUUsage.html#llvm-ir-intrinsics
+# llvm.log2/llvm.exp2 don't support double
+llvm_intrinsics = {Ops.SQRT: "sqrt"}
 class AMDLLVMRenderer(LLVMRenderer):
   device = "AMD"
   has_local = True
   shared_max = AMDRenderer.shared_max
   global_max = AMDRenderer.global_max
   abi = "amdgpu_kernel"
+  code_for_op = {**LLVMRenderer.code_for_op, **{op: lambda: None for op in llvm_intrinsics}}
   string_rewrite = PatternMatcher([
     (UPat(Ops.SPECIAL, name="x"), lambda ctx, x: f"  {ctx[x]} = " + f"{ code_for_workitem[x.arg[0][0]](x.arg[0][-1])}; "),
+    (UPat(tuple(llvm_intrinsics), name="x"),
+    lambda ctx, x: f"  {ctx[x]} = call {ldt(x.dtype)} @llvm.{llvm_intrinsics[x.op]}.{ldt(x.dtype.scalar())}({ldt(x.src[0].dtype)} {ctx[x.src[0]]})"),
     (UPat(Ops.BARRIER), lambda ctx: barrier),
   ]) + base_rewrite
   extra_matcher = LLVMRenderer.extra_matcher + PatternMatcher([

tinygrad/renderer/ptx.py

@@ -119,7 +119,7 @@ string_rewrite = PatternMatcher([
     ctx.code_for_op[Ops.CMPLT](ctx.r[x], ctx.r[x.src[0]], ctx.r[src0.src[0]], dtypes.int, ctx.types[dtypes.int]),
     f"@{ctx.r[x]} bra LOOP_{ctx.r[src0][1:]};"]),
   (UPat(Ops.DEFINE_LOCAL, name="x"),
-   lambda ctx, x: [f".shared .align 16 .b8 {x.arg}[{x.dtype.size*x.dtype.itemsize}];", f"mov.u64 {ctx.r[x]}, {x.arg}[0];"]),
+   lambda ctx, x: [f".shared .align 16 .b8 local{x.arg}[{x.dtype.size*x.dtype.itemsize}];", f"mov.u64 {ctx.r[x]}, local{x.arg}[0];"]),
   (UPat(Ops.IF, name="x"), lambda ctx, x: f"@!{ctx.r[x.src[0]]} bra IF_{ctx.r[x.src[0]][1:]}_{ctx.uops.index(x)};"),
   (UPat(Ops.ENDIF, name="x"), lambda ctx, x: f"IF_{ctx.r[x.src[0].src[0]][1:]}_{ctx.uops.index(x.src[0])}:"),
   (UPat(Ops.WMMA, name="x"), lambda ctx, x: list(render_wmma(ctx, x))),
@@ -190,7 +190,7 @@ class PTXRenderer(Renderer):
         r[u] = r[u.src[0]]
         continue
       if u.op is Ops.DEFINE_REG:
-        r[u] = [ssa("reg", u, self.types[u.dtype.base.scalar()]) for _ in range(cast(PtrDType, u.dtype).size)]
+        r[u] = [ssa("reg", u, self.types[u.dtype.base.scalar()]) for _ in range(u.ptrdtype.size)]
         continue
       if u.op in {Ops.INDEX, Ops.LOAD, Ops.STORE} and isinstance(u.src[0].dtype, PtrDType) and u.src[0].dtype.addrspace == AddrSpace.REG:
         if u.op is Ops.INDEX:
@@ -215,7 +215,7 @@ class PTXRenderer(Renderer):
                        [ssa("wmma_acc", dtype="b32") for _ in range(0, len(r[u.src[2]]), 4 // u.dtype.scalar().itemsize)]]
         r[u] = [ssa("wmma", dtype=self.types[u.dtype.scalar()]) for _ in range(u.dtype.count)]
       prefix, dtype = {Ops.CAST: ("cast", None), Ops.BITCAST: ("cast", None), Ops.ENDRANGE: ("pred", "pred"), Ops.RANGE: ("ridx", None),
-        Ops.DEFINE_VAR: ("dat", None), Ops.CONST: ("const", None), Ops.DEFINE_LOCAL:("local",self.types[dtypes.ulong]),
+        Ops.DEFINE_VAR: ("dat", None), Ops.CONST: ("const", None), Ops.DEFINE_LOCAL: ("local",self.types[dtypes.ulong]),
         Ops.DEFINE_GLOBAL: ("dat", self.types[dtypes.ulong]), **{op: ("alu", None) for op in GroupOp.ALU}}.get(u.op, (None, None))
       if prefix: r[u] = ssa(prefix, u, dtype)
 

tinygrad/runtime/ops_amd.py

@@ -4,10 +4,10 @@ import os, ctypes, ctypes.util, struct, hashlib, functools, importlib, mmap, err
 assert sys.platform != 'win32'
 from dataclasses import dataclass
 from tinygrad.runtime.support.hcq import HCQCompiled, HCQAllocator, HCQBuffer, HWQueue, CLikeArgsState, HCQSignal, HCQProgram, FileIOInterface
-from tinygrad.runtime.support.hcq import MMIOInterface
+from tinygrad.runtime.support.hcq import MMIOInterface, BumpAllocator
 from tinygrad.uop.ops import sint
 from tinygrad.device import Compiled, DMAFdRef, BufferSpec
-from tinygrad.helpers import getenv, to_mv, round_up, data64_le, all_same, flatten, DEBUG, AMD_LLVM, PROFILE, ProfileEvent, suppress_finalizing
+from tinygrad.helpers import getenv, to_mv, round_up, data64_le, DEBUG, AMD_LLVM, PROFILE, ProfileEvent, suppress_finalizing, lo32, hi32
 from tinygrad.renderer.cstyle import AMDRenderer
 from tinygrad.renderer.llvmir import AMDLLVMRenderer
 from tinygrad.runtime.autogen import kfd, hsa, pci, sqtt
@@ -24,6 +24,8 @@ EVENT_INDEX_PARTIAL_FLUSH = 4 # based on a comment in nvd.h
 WAIT_REG_MEM_FUNCTION_EQ  = 3 # ==
 WAIT_REG_MEM_FUNCTION_NEQ = 4 # !=
 WAIT_REG_MEM_FUNCTION_GEQ = 5 # >=
+AQL_HDR = (1 << hsa.HSA_PACKET_HEADER_BARRIER) | (hsa.HSA_FENCE_SCOPE_SYSTEM << hsa.HSA_PACKET_HEADER_SCACQUIRE_FENCE_SCOPE) \
+        | (hsa.HSA_FENCE_SCOPE_SYSTEM << hsa.HSA_PACKET_HEADER_SCRELEASE_FENCE_SCOPE)
 
 class AMDSignal(HCQSignal):
   def __init__(self, *args, **kwargs): super().__init__(*args, **{**kwargs, 'timestamp_divider': 100})
@@ -106,17 +108,6 @@ class AMDComputeQueue(HWQueue):
 
     self.pkt3(self.pm4.PACKET3_RELEASE_MEM, event_dw | cache_flags_dw, memsel_dw, *data64_le(address), *data64_le(value), ctxid)
 
-  def xcc_barrier(self):
-    if self.dev.xcc_sync is None: return self
-    assert self.dev.xccs == 8, 'only 8 XCCs supported'
-    a, b = self.dev.xcc_sync
-    mem_eq = self.pm4.WAIT_REG_MEM_FUNCTION(WAIT_REG_MEM_FUNCTION_EQ) | self.pm4.WAIT_REG_MEM_MEM_SPACE(1)
-    self.pkt3(self.pm4.PACKET3_ATOMIC_MEM, self.soc.TC_OP_ATOMIC_ADD_RTN_32, *data64_le(a.value_addr), *data64_le(1), *data64_le(0), 0x10) # a += 1
-    self.pkt3(self.pm4.PACKET3_WAIT_REG_MEM, mem_eq, *data64_le(a.value_addr), 0, 0b111, 0x80) # a == 0 (mod 8) via bitmask
-    self.pkt3(self.pm4.PACKET3_ATOMIC_MEM, self.soc.TC_OP_ATOMIC_ADD_RTN_32, *data64_le(b.value_addr), *data64_le(1), *data64_le(0), 0x10) # b += 1
-    self.pkt3(self.pm4.PACKET3_WAIT_REG_MEM, mem_eq, *data64_le(b.value_addr), 0, 0b111, 0x80) # b == 0 (mod 8) via bitmask
-    return self
-
   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],
@@ -124,13 +115,6 @@ class AMDComputeQueue(HWQueue):
     self.acquire_mem()
     return self
 
-  def xcc_config(self):
-    self.wreg(self.gc.regCOMPUTE_TG_CHUNK_SIZE, 1)
-    for xcc_id in range(self.dev.xccs):
-      with self.pred_exec(xcc_mask=1 << xcc_id):
-        self.wreg(self.dev.regCOMPUTE_CURRENT_LOGIC_XCC_ID, xcc_id)
-    return self
-
   def spi_config(self, tracing:bool):
     self.wreg(self.gc.regSPI_CONFIG_CNTL, ps_pkr_priority_cntl=3, exp_priority_order=3, gpr_write_priority=0x2c688,
               enable_sqg_bop_events=int(tracing), enable_sqg_top_events=int(tracing))
@@ -275,16 +259,10 @@ class AMDComputeQueue(HWQueue):
 
     if prg.dev.sqtt_enabled: self.pkt3(self.pm4.PACKET3_EVENT_WRITE, self.pm4.EVENT_TYPE(self.soc.THREAD_TRACE_MARKER) | self.pm4.EVENT_INDEX(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))
-
-    if self.dev.xccs > 1:
-      self.release_mem(cache_flush=True)
-      self.acquire_mem(gli=0)
-      self.xcc_barrier()
     return self
 
   def wait(self, signal:AMDSignal, value:sint=0):
     self.wait_reg_mem(mem=signal.value_addr, value=value, mask=0xffffffff)
-    if self.dev.xccs > 1: self.xcc_barrier()
     return self
 
   def timestamp(self, signal:AMDSignal):
@@ -329,6 +307,41 @@ class AMDComputeQueue(HWQueue):
     dev.compute_queue.put_value += len(cmds)
     dev.compute_queue.signal_doorbell(dev)
 
+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)
+    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))
+    self.bind_sints_to_mem(*local_size, mem=(pkt_view:=MMIOInterface(addr=ctypes.addressof(pkt), nbytes=ctypes.sizeof(pkt))), fmt='H', offset=4)
+    self.bind_sints_to_mem(*[l * g for l,g in zip(local_size, global_size)], mem=pkt_view, fmt='I', offset=12)
+
+  def bind(self, dev:AMDDevice): pass # not supported
+  def _submit(self, dev:AMDDevice):
+    pm4_batch:list[int] = []
+    aql_bytes = bytes()
+
+    def flush_pm4_batch():
+      nonlocal pm4_batch
+      if not pm4_batch: return bytes()
+      dev.pm4_ibs.cpu_view().view(off:=dev.pm4_ib_alloc.alloc(len(pm4_batch) * 4), fmt='I')[:len(pm4_batch)] = array.array('I', pm4_batch)
+      pkt = [AQL_HDR | (hsa.HSA_PACKET_TYPE_VENDOR_SPECIFIC << hsa.HSA_PACKET_HEADER_TYPE) | (1 << 16),
+        self.pm4.PACKET3(self.pm4.PACKET3_INDIRECT_BUFFER, 2), *data64_le(dev.pm4_ibs.va_addr+off), len(pm4_batch)|self.pm4.INDIRECT_BUFFER_VALID, 10]
+      pm4_batch.clear()
+      return bytes(array.array('I', pkt + [0] * 10))
+
+    for cmd in self._q:
+      if isinstance(cmd, hsa.hsa_kernel_dispatch_packet_t): aql_bytes += flush_pm4_batch() + bytes(cmd)
+      else: pm4_batch.append(cmd)
+    aql_bytes += flush_pm4_batch()
+
+    assert len(aql_bytes) < dev.compute_queue.ring.nbytes, "submit is too large for the queue"
+    cp_bytes = min(len(aql_bytes), (dev.compute_queue.ring.nbytes - (dev.compute_queue.put_value * 64) % dev.compute_queue.ring.nbytes))
+    dev.compute_queue.ring.view(offset=(dev.compute_queue.put_value * 64) % dev.compute_queue.ring.nbytes, fmt='B')[:cp_bytes] = aql_bytes[:cp_bytes]
+    if (tail_bytes:=(len(aql_bytes) - cp_bytes)) > 0: dev.compute_queue.ring.view(offset=0, fmt='B')[:tail_bytes] = aql_bytes[cp_bytes:]
+    dev.compute_queue.put_value += len(aql_bytes) // 64
+    dev.compute_queue.signal_doorbell(dev, doorbell_value=dev.compute_queue.put_value-1)
+
 class AMDCopyQueue(HWQueue):
   def __init__(self, dev, max_copy_size=0x40000000):
     self.dev, self.sdma, self.internal_cmd_sizes, self.max_copy_size = dev, dev.sdma, [], max_copy_size
@@ -426,14 +439,19 @@ class AMDProgram(HCQProgram):
     # TODO; this API needs the type signature of the function and global_size/local_size
     self.dev, self.name, self.lib = dev, name, lib
 
-    image, sections, _ = elf_loader(self.lib)
+    image, sections, relocs = elf_loader(self.lib)
+
+    rodata_entry = next((sh.header.sh_addr for sh in sections if sh.name == ".rodata"), -1)
+    assert rodata_entry >= 0, ".rodata section not found"
+
+    for apply_image_offset, rel_sym_offset, typ, addent in relocs:
+      if typ == 5: image[apply_image_offset:apply_image_offset+8] = struct.pack('<q', rel_sym_offset - apply_image_offset + addent) # R_AMDGPU_REL64
+      else: raise RuntimeError(f"unknown AMD reloc {typ}")
+
     self.lib_gpu = self.dev.allocator.alloc(round_up(image.nbytes, 0x1000), buf_spec:=BufferSpec(cpu_access=True, nolru=True))
     self.dev.allocator._copyin(self.lib_gpu, image)
     self.dev.synchronize()
 
-    rodata_entry = next((sh.header.sh_addr for sh in sections if sh.name == ".rodata"), -1)
-    text_entry = next((sh.header.sh_addr for sh in sections if sh.name == ".text"), -1)
-    assert rodata_entry >= 0 and text_entry >= 0, ".text or .rodata section not found"
     self.group_segment_size = image[rodata_entry:rodata_entry+4].cast("I")[0]
     self.private_segment_size = image[rodata_entry+4:rodata_entry+8].cast("I")[0]
     self.kernargs_segment_size = image[rodata_entry+8:rodata_entry+12].cast("I")[0]
@@ -451,8 +469,8 @@ class AMDProgram(HCQProgram):
     self.rsrc1: int = code.compute_pgm_rsrc1 | ((1 << 20) if (11,0,0) <= self.dev.target < (12,0,0) else 0)
     self.rsrc2: int = code.compute_pgm_rsrc2 | (lds_size << 15)
     self.rsrc3: int = image[rodata_entry+44:rodata_entry+48].cast("I")[0] # NOTE: kernel descriptor, not in amd_kernel_code_t struct
+    self.aql_prog_addr: int = self.lib_gpu.va_addr + rodata_entry
     self.prog_addr: int = self.lib_gpu.va_addr + rodata_entry + code.kernel_code_entry_byte_offset
-    if code.kernel_code_entry_byte_offset == 0: self.prog_addr = self.lib_gpu.va_addr + text_entry
     # Some programs use hsa_kernel_dispatch_packet_t to read workgroup sizes during execution.
     # The packet is represented as a pointer and set up in SGPRs. Space for the packet is allocated as part of the kernel arguments.
     self.enable_dispatch_ptr: int = code.kernel_code_properties & hsa.AMD_KERNEL_CODE_PROPERTIES_ENABLE_SGPR_DISPATCH_PTR
@@ -495,13 +513,7 @@ class AMDQueueDesc:
   @property
   def read_ptr(self): return min(p[0] for p in self.read_ptrs)
 
-  @classmethod
-  def multi(cls, *queues: AMDQueueDesc):
-    assert all_same([(q.ring.addr, q.put_value) for q in queues]), f"All queues must have the same ring and put_value: {queues}"
-    return cls(ring=queues[0].ring, put_value=queues[0].put_value, doorbells=flatten(q.doorbells for q in queues),
-               read_ptrs=flatten(q.read_ptrs for q in queues), write_ptrs=flatten(q.write_ptrs for q in queues))
-
-  def signal_doorbell(self, dev):
+  def signal_doorbell(self, dev, doorbell_value:int|None=None):
     for write_ptr in self.write_ptrs: write_ptr[0] = self.put_value
 
     # Ensure all prior writes are visible to the GPU.
@@ -509,7 +521,7 @@ class AMDQueueDesc:
 
     # Flush hdp if queue is in dev mem.
     if dev.is_am() and not dev.is_usb(): dev.iface.dev_impl.gmc.flush_hdp()
-    for doorbell in self.doorbells: doorbell[0] = self.put_value
+    for doorbell in self.doorbells: doorbell[0] = self.put_value if doorbell_value is None else doorbell_value
 
 class KFDIface:
   kfd:FileIOInterface|None = None
@@ -612,12 +624,12 @@ class KFDIface:
     stm = kfd.AMDKFD_IOC_MAP_MEMORY_TO_GPU(self.kfd, handle=mem.meta.handle, device_ids_array_ptr=ctypes.addressof(c_gpus), n_devices=1)
     assert stm.n_success == 1
 
-  def create_queue(self, queue_type, ring, gart, eop_buffer=None, cwsr_buffer=None, ctl_stack_size=0, ctx_save_restore_size=0, xcc_id=0):
+  def create_queue(self, queue_type, ring, gart, rptr, wptr, eop_buffer=None, cwsr_buffer=None, ctl_stack_size=0, ctx_save_restore_size=0, xcc_id=0):
     queue = kfd.AMDKFD_IOC_CREATE_QUEUE(KFDIface.kfd, ring_base_address=ring.va_addr, ring_size=ring.size, gpu_id=self.gpu_id,
       queue_type=queue_type, queue_percentage=kfd.KFD_MAX_QUEUE_PERCENTAGE|(xcc_id<<8), queue_priority=kfd.KFD_MAX_QUEUE_PRIORITY,
       eop_buffer_address=eop_buffer.va_addr if eop_buffer else 0, eop_buffer_size=eop_buffer.size if eop_buffer else 0, ctl_stack_size=ctl_stack_size,
       ctx_save_restore_address=cwsr_buffer.va_addr if cwsr_buffer else 0, ctx_save_restore_size=ctx_save_restore_size,
-      write_pointer_address=gart.va_addr, read_pointer_address=gart.va_addr + 8 * (xcc_id + 1))
+      write_pointer_address=gart.va_addr+wptr, read_pointer_address=gart.va_addr+rptr+8*xcc_id)
 
     if not hasattr(self, 'doorbells'):
       self.doorbells_base = queue.doorbell_offset & (~0x1fff) # doorbell is two pages
@@ -662,18 +674,19 @@ class PCIIface(PCIIfaceBase):
       'max_slots_scratch_cu': self.dev_impl.gc_info.gc_max_scratch_slots_per_cu, 'max_waves_per_simd': self.dev_impl.gc_info.gc_max_waves_per_simd,
       'simd_arrays_per_engine': self.dev_impl.gc_info.gc_num_sa_per_se, 'lds_size_in_kb': self.dev_impl.gc_info.gc_lds_size}
 
-  def create_queue(self, queue_type, ring, gart, eop_buffer=None, cwsr_buffer=None, ctl_stack_size=0, ctx_save_restore_size=0, xcc_id=0):
+  def create_queue(self, queue_type, ring, gart, rptr, wptr, eop_buffer=None, cwsr_buffer=None, ctl_stack_size=0, ctx_save_restore_size=0, xcc_id=0):
     assert cwsr_buffer is None, "no cwsr buffer for am"
 
     if queue_type == kfd.KFD_IOC_QUEUE_TYPE_SDMA:
-      self.dev_impl.sdma.setup_ring(ring_addr=ring.va_addr, ring_size=ring.size, rptr_addr=gart.va_addr, wptr_addr=gart.va_addr+0x10,
+      self.dev_impl.sdma.setup_ring(ring_addr=ring.va_addr, ring_size=ring.size, rptr_addr=gart.va_addr+rptr, wptr_addr=gart.va_addr+wptr,
                                     doorbell=(doorbell_index:=am.AMDGPU_NAVI10_DOORBELL_sDMA_ENGINE0), pipe=0, queue=0)
     else:
-      self.dev_impl.gfx.setup_ring(ring_addr=ring.va_addr, ring_size=ring.size, rptr_addr=gart.va_addr, wptr_addr=gart.va_addr+0x10,
-        eop_addr=eop_buffer.va_addr, eop_size=eop_buffer.size, doorbell=(doorbell_index:=am.AMDGPU_NAVI10_DOORBELL_MEC_RING0), pipe=0, queue=0)
+      self.dev_impl.gfx.setup_ring(ring_addr=ring.va_addr, ring_size=ring.size, rptr_addr=gart.va_addr+rptr, wptr_addr=gart.va_addr+wptr,
+        eop_addr=eop_buffer.va_addr, eop_size=eop_buffer.size, doorbell=(doorbell_index:=am.AMDGPU_NAVI10_DOORBELL_MEC_RING0), pipe=0, queue=0,
+        aql=(queue_type==kfd.KFD_IOC_QUEUE_TYPE_COMPUTE_AQL))
 
     return AMDQueueDesc(ring=ring.cpu_view().view(fmt='I'), doorbells=[self.dev_impl.doorbell64.view(doorbell_index * 8, 8, fmt='Q')],
-      read_ptrs=[gart.cpu_view().view(size=8, fmt='Q')], write_ptrs=[gart.cpu_view().view(offset=0x10, size=8, fmt='Q')])
+      read_ptrs=[gart.cpu_view().view(offset=rptr, size=8, fmt='Q')], write_ptrs=[gart.cpu_view().view(offset=wptr, size=8, fmt='Q')])
 
   def sleep(self, timeout):
     if self.pci_dev.irq_poller is not None and (events_cnt:=len(self.pci_dev.irq_poller.poll(timeout))):
@@ -715,9 +728,9 @@ class USBIface(PCIIface):
     return HCQBuffer(am_mapping.va_addr, size, meta=PCIAllocationMeta(am_mapping, has_cpu_mapping=False),
       view=USBMMIOInterface(self.usb, self.bars[0][0] + am_mapping.paddrs[0][0], size, fmt='B') if cpu_access else None, owner=self.dev)
 
-  def create_queue(self, queue_type, ring, gart, eop_buffer=None, cwsr_buffer=None, ctl_stack_size=0, ctx_save_restore_size=0, xcc_id=0):
+  def create_queue(self, queue_type, ring, gart, rptr, wptr, eop_buffer=None, cwsr_buffer=None, ctl_stack_size=0, ctx_save_restore_size=0, xcc_id=0):
     if queue_type == kfd.KFD_IOC_QUEUE_TYPE_COMPUTE: self.usb._pci_cacheable += [(ring.cpu_view().addr, ring.size)]
-    return super().create_queue(queue_type, ring, gart, eop_buffer, cwsr_buffer, ctl_stack_size, ctx_save_restore_size, xcc_id)
+    return super().create_queue(queue_type, ring, gart, rptr, wptr, eop_buffer, cwsr_buffer, ctl_stack_size, ctx_save_restore_size, xcc_id)
 
   def sleep(self, timeout): pass
 
@@ -756,14 +769,17 @@ class AMDDevice(HCQCompiled):
     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])
 
-    # Define the regCOMPUTE_CURRENT_LOGIC_XCC_ID register, which is missing from the asic_regs files.
-    if self.target[:2] in {(9,4),(9,5)}: self.regCOMPUTE_CURRENT_LOGIC_XCC_ID = AMDReg("regCOMPUTE_CURRENT_LOGIC_XCC_ID", 0xe25, 0, {}, self.gc.bases)
-
     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.compute_queue = self.create_queue(kfd.KFD_IOC_QUEUE_TYPE_COMPUTE, 0x2000 if self.is_usb() else (16 << 20), eop_buffer_size=0x1000,
+    self.is_aql = getenv("AMD_AQL", int(self.xccs > 1))
+    if self.is_aql:
+      self.pm4_ibs = self.iface.alloc(0x2000 if self.is_usb() else (16 << 20), uncached=True, cpu_access=True)
+      self.pm4_ib_alloc = BumpAllocator(self.pm4_ibs.size, wrap=True)
+
+    self.compute_queue = self.create_queue(kfd.KFD_IOC_QUEUE_TYPE_COMPUTE_AQL if self.is_aql else kfd.KFD_IOC_QUEUE_TYPE_COMPUTE,
+      0x2000 if self.is_usb() else (16 << 20), eop_buffer_size=0x1000,
       ctx_save_restore_size=0 if self.is_am() else wg_data_size + ctl_stack_size, ctl_stack_size=ctl_stack_size, debug_memory_size=debug_memory_size)
 
     max_copy_size = 0x40000000 if self.iface.ip_versions[am.SDMA0_HWIP][0] >= 5 else 0x400000
@@ -771,20 +787,14 @@ class AMDDevice(HCQCompiled):
 
     super().__init__(device, AMDAllocator(self), AMDLLVMRenderer(self.arch) if AMD_LLVM else AMDRenderer(self.arch),
                      AMDLLVMCompiler(self.arch) if AMD_LLVM else HIPCompiler(self.arch), functools.partial(AMDProgram, self),
-                     AMDSignal, functools.partial(AMDComputeQueue, self), functools.partial(AMDCopyQueue, self, max_copy_size=max_copy_size),
+                     AMDSignal, functools.partial(AMDComputeAQLQueue if self.is_aql else AMDComputeQueue, self),
+                     functools.partial(AMDCopyQueue, self, max_copy_size=max_copy_size),
                      kernargs_size=(8 << 10) if self.is_usb() else (16 << 20), sigalloc_size=0x100 if self.is_usb() else 0x1000)
 
     # Scratch setup
     self.max_private_segment_size = 0
     self._ensure_has_local_memory(128) # set default scratch size to 128 bytes per thread
 
-    # XCC setup
-    self.xcc_sync: tuple[AMDSignal, AMDSignal]|None = None
-    if self.xccs > 1:
-      self.xcc_sync_area = self.allocator.alloc(0x1000, BufferSpec(nolru=True, cpu_access=True))
-      self.xcc_sync = (AMDSignal(base_buf=self.xcc_sync_area), AMDSignal(base_buf=self.xcc_sync_area.offset(256)))
-      AMDComputeQueue(self).xcc_config().submit(self)
-
     # 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))
     if self.sqtt_enabled:
@@ -798,19 +808,26 @@ class AMDDevice(HCQCompiled):
       self.sqtt_buffers = [self.allocator.alloc(SQTT_BUFFER_SIZE*1024*1024, BufferSpec(cpu_access=True, nolru=True)) for _ in range(SQTT_NUM)]
       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.cmd_id = 0
-      AMDComputeQueue(self).sqtt_start(self.sqtt_buffers, self.sqtt_itrace_se_mask).submit(self)
+      cast(AMDComputeQueue, self.hw_compute_queue_t()).sqtt_start(self.sqtt_buffers, self.sqtt_itrace_se_mask).submit(self)
 
   def create_queue(self, queue_type, ring_size, ctx_save_restore_size=0, eop_buffer_size=0, ctl_stack_size=0, debug_memory_size=0):
     ring = self.iface.alloc(ring_size, uncached=True, cpu_access=True)
     gart = self.iface.alloc(0x100, uncached=True, cpu_access=True)
 
+    if queue_type == kfd.KFD_IOC_QUEUE_TYPE_COMPUTE_AQL:
+      aql_desc = hsa.amd_queue_t(queue_properties=hsa.AMD_QUEUE_PROPERTIES_IS_PTR64 | hsa.AMD_QUEUE_PROPERTIES_ENABLE_PROFILING,
+        read_dispatch_id_field_base_byte_offset=getattr(hsa.amd_queue_t, 'read_dispatch_id').offset,
+        max_cu_id=self.max_cu_id, max_wave_id=self.max_wave_id)
+      gart.cpu_view().view(fmt='B')[:ctypes.sizeof(aql_desc)] = bytes(aql_desc)
+      self.aql_desc = hsa.amd_queue_t.from_address(gart.va_addr)
+
     cwsr_buffer_size = round_up((ctx_save_restore_size + debug_memory_size) * self.iface.props.get('num_xcc', 1), mmap.PAGESIZE)
     cwsr_buffer = self.iface.alloc(cwsr_buffer_size) if ctx_save_restore_size else None
     eop_buffer = self.iface.alloc(eop_buffer_size) if eop_buffer_size else None
 
-    return AMDQueueDesc.multi(*(self.iface.create_queue(queue_type, ring, gart, eop_buffer=eop_buffer, cwsr_buffer=cwsr_buffer, xcc_id=xcc_id,
-                                                            ctx_save_restore_size=ctx_save_restore_size, ctl_stack_size=ctl_stack_size)
-                                for xcc_id in range(self.xccs if queue_type == kfd.KFD_IOC_QUEUE_TYPE_COMPUTE else 1)))
+    return (self.iface.create_queue(queue_type, ring, gart, rptr=getattr(hsa.amd_queue_t, 'read_dispatch_id').offset,
+            wptr=getattr(hsa.amd_queue_t, 'write_dispatch_id').offset, eop_buffer=eop_buffer, cwsr_buffer=cwsr_buffer,
+            ctx_save_restore_size=ctx_save_restore_size, ctl_stack_size=ctl_stack_size))
 
   def _ensure_has_local_memory(self, required):
     if self.max_private_segment_size >= required: return
@@ -828,8 +845,16 @@ class AMDDevice(HCQCompiled):
       self.tmpring_size = waves << 12 | wavesize
       self.max_private_segment_size = required
 
+      if hasattr(self, 'aql_desc'):
+        self.aql_desc.scratch_backing_memory_location = self.scratch.va_addr
+        self.aql_desc.scratch_backing_memory_byte_size = self.scratch.size
+        self.aql_desc.scratch_wave64_lane_byte_size = self.max_private_segment_size * (self.aql_desc.max_wave_id + 1) // 64
+        self.aql_desc.scratch_resource_descriptor[:] = [lo32(self.scratch.va_addr), hi32(self.scratch.va_addr) | (1 << 30), lo32(self.scratch.size),
+          0x20814fac] # FORMAT=BUF_FORMAT_32_UINT,OOB_SELECT=2,ADD_TID_ENABLE=1,TYPE=SQ_RSRC_BUF,SQ_SELs
+        self.aql_desc.compute_tmpring_size = self.tmpring_size
+
   def invalidate_caches(self):
-    AMDComputeQueue(self).memory_barrier().signal(self.timeline_signal, self.next_timeline()).submit(self)
+    self.hw_compute_queue_t().memory_barrier().signal(self.timeline_signal, self.next_timeline()).submit(self)
     self.synchronize()
 
   def on_device_hang(self): self.iface.on_device_hang()
@@ -838,7 +863,8 @@ class AMDDevice(HCQCompiled):
     if self.sqtt_enabled:
       wptrs_buf = self.allocator.alloc(round_up(len(self.sqtt_buffers), 0x1000), BufferSpec(cpu_access=True, nolru=True))
       wptrs = to_mv(wptrs_buf.va_addr, wptrs_buf.size)
-      AMDComputeQueue(self).sqtt_stop(len(self.sqtt_buffers), wptrs_buf).signal(self.timeline_signal, self.next_timeline()).submit(self)
+      cast(AMDComputeQueue, self.hw_compute_queue_t()).sqtt_stop(len(self.sqtt_buffers), wptrs_buf) \
+                                                      .signal(self.timeline_signal, self.next_timeline()).submit(self)
       self.synchronize()
       if DEBUG>=2: print('Saving SQTT in profile...')
       for i,buf0 in enumerate(self.sqtt_buffers):

tinygrad/runtime/ops_null.py

@@ -1,28 +1,33 @@
+import functools
 from tinygrad.device import Compiled, Compiler, Allocator
 from tinygrad.engine.jit import MultiGraphRunner
 from tinygrad.renderer.cstyle import CStyleLanguage
 from tinygrad.uop.ops import Ops
+from tinygrad.helpers import cpu_profile
 
 class NullRenderer(CStyleLanguage):
   device = "NULL"
   has_local = False
   float4 = "float4"
+  barrier = "// BARRIER"
   code_for_op = {**CStyleLanguage.code_for_op, Ops.THREEFRY: lambda a,b,dtype: f"threefry({a},{b})", Ops.MAX: lambda a,b,dtype: f"max({a},{b})"}
 
 class NullProgram:
-  def __init__(self, name:str, lib:bytes): pass
+  def __init__(self, device:str, name:str, lib:bytes): self.device, self.name = device, name
   def __call__(self, *bufs, global_size:tuple[int,int,int]=(1,1,1), local_size:tuple[int,int,int]=(1,1,1), vals:tuple[int, ...]=(), wait=False):
-    return 1e-4
+    with cpu_profile(self.name, self.device): return 1e-4
 
 class NullAllocator(Allocator['NullDevice']):
   def _alloc(self, size, options): pass
   def _copyin(self, dest, src:memoryview): pass
   def _copyout(self, dest:memoryview, src): pass
-  def _transfer(self, dest, src, sz:int, src_dev, dest_dev): pass
+  def _transfer(self, dest, src, sz:int, src_dev, dest_dev):
+    with cpu_profile(f"{src_dev.device} -> {dest_dev.device}", self.dev.device): pass
   def _offset(self, buf, offset:int, size:int): pass
 
 class NullGraph(MultiGraphRunner):
   def __call__(self, input_rawbuffers, var_vals, wait=False) -> float|None: return 1e-3
 
 class NullDevice(Compiled):
-  def __init__(self, device:str): super().__init__(device, NullAllocator(self), NullRenderer(), Compiler(), NullProgram, NullGraph)
+  def __init__(self, device:str): super().__init__(device, NullAllocator(self), NullRenderer(), Compiler(), functools.partial(NullProgram, device),
+                                                   NullGraph)

tinygrad/runtime/ops_python.py

@@ -4,25 +4,35 @@
 # this is the (living) definition of uops
 from typing import Any, TYPE_CHECKING
 import pickle, base64, itertools, time, struct, sys
-from tinygrad.dtype import DType, dtypes, ImageDType, PtrDType, truncate
+from tinygrad.dtype import DType, dtypes, ImageDType, PtrDType, truncate, float_to_bf16
 from tinygrad.helpers import all_same, getenv, flatten, get_single_element
 from tinygrad.device import Compiled, Compiler, Allocator
 from tinygrad.codegen.opt import tc
-from tinygrad.uop.ops import exec_alu, Ops, UOp, GroupOp
+from tinygrad.uop.ops import exec_alu, python_alu, Ops, UOp, GroupOp
 from tinygrad.renderer import Renderer
 
-def _load(m, i):
+def storage_fmt_for_dtype(dtype: DType): return 'H' if dtype == dtypes.bfloat16 else dtype.fmt
+
+def to_storage_scalar(x, dtype: DType):
+  if dtype == dtypes.bfloat16: return (struct.unpack('I', struct.pack('f', float_to_bf16(x)))[0] >> 16) & 0xFFFF
+  return x
+
+def from_storage_scalar(x, dtype: DType):
+  if dtype == dtypes.bfloat16: return struct.unpack('f', struct.pack('I', (x & 0xFFFF) << 16))[0]
+  return x
+
+def _load(m, i, dtype: DType):
   if i is None: return 0.0
   if i < 0 or i >= len(m): raise IndexError(f"load out of bounds, size is {len(m)} and access is {i}")
-  return m[i]
+  return from_storage_scalar(m[i], dtype)
 
-def load(inp, j=0):
-  if len(inp) == 2: return [_load(m, x+j if x is not None else None) if gate else default for (m,x,gate),default in zip(*inp)]
-  return [_load(m, x+j if x is not None else None) for m,x,_ in inp[0]]
+def load(inp, j, dtype: DType):
+  if len(inp) == 2: return [_load(m, x+j if x is not None else None, dtype) if gate else default for (m,x,gate),default in zip(*inp)]
+  return [_load(m, x+j if x is not None else None, dtype) for m,x,_ in inp[0]]
 
-def _store(m, i, v):
+def _store(m, i, v, dtype: DType):
   if i < 0 or i >= len(m): raise IndexError(f"store out of bounds, size is {len(m)}, access is {i}, value is {v}")
-  m[i] = v
+  m[i] = to_storage_scalar(v, dtype)
 
 class PythonProgram:
   def __init__(self, name:str, lib:bytes):
@@ -57,19 +67,20 @@ class PythonProgram:
         if uop is Ops.STORE:
           for j,val in enumerate(inp[1] if dtp[1].count > 1 else [inp[1]]):
             for (m,o,g),v in zip(inp[0], val):
-              if g: _store(m, o+j, v)
+              if g: _store(m, o+j, v, dtp[1].scalar())
           i += 1
           continue
         if uop in {Ops.DEFINE_GLOBAL, Ops.DEFINE_LOCAL, Ops.DEFINE_REG}:
           assert isinstance(dtype, PtrDType), dtype
-          if dtype.fmt is None: raise RuntimeError(f"{dtype=} is not supported")
-          if TYPE_CHECKING or sys.version_info < (3, 12): assert dtype.fmt != "e"
+          storage_fmt = storage_fmt_for_dtype(dtype.base.scalar())
+          if storage_fmt is None: raise RuntimeError(f"{dtype=} is not supported")
+          if TYPE_CHECKING or sys.version_info < (3, 12): assert storage_fmt != "e"
           if uop is Ops.DEFINE_REG:
             # REGs are per thread
-            ul[i] = [memoryview(bytearray(dtype.size*dtype.itemsize)).cast(dtype.fmt) for _ in range(warp_size)]
+            ul[i] = [memoryview(bytearray(dtype.size*dtype.itemsize)).cast(storage_fmt) for _ in range(warp_size)]
           else:
             buf = memoryview(bytearray(dtype.size*dtype.itemsize)) if uop is not Ops.DEFINE_GLOBAL else pbufs.pop(0)
-            ul[i] = [buf.cast(dtype.fmt)] * warp_size
+            ul[i] = [buf.cast(storage_fmt)] * warp_size
         elif uop is Ops.DEFINE_VAR:
           ul[i] = [pvals.pop(0)] * warp_size
         elif uop is Ops.SPECIAL:
@@ -98,16 +109,17 @@ class PythonProgram:
               continue
         elif uop is Ops.VECTORIZE: ul[i] = inp
         elif uop is Ops.BITCAST:
-          assert dtp[0].fmt and dtype.fmt
-          pack_format, unpack_format = str(warp_size) + dtp[0].fmt, str(warp_size) + dtype.fmt
-          ul[i] = list(struct.unpack(unpack_format, struct.pack(pack_format, *inp[0])))
+          packed = struct.pack(str(warp_size) + storage_fmt_for_dtype(dtp[0].scalar()), *[to_storage_scalar(x, dtp[0].scalar()) for x in inp[0]])
+          ul[i] = list(struct.unpack(str(warp_size) +  storage_fmt_for_dtype(dtype.scalar()), packed))
+          ul[i] = [from_storage_scalar(x, dtype.scalar()) for x in ul[i]]
         elif uop is Ops.CAST:
           ul[i] = [truncate.get(dtype, lambda dt: dt)(dtypes.as_const(x, dtype)) for x in inp[0]]
         elif uop is Ops.LOAD:
           if dtype.count > 1:
-            ul[i] = [load([inp[i][j] if i != 0 and dtp[i].count > 1 else inp[i] for i in range(len(inp))], j) for j in range(dtype.count)]
+            ul[i] = [load([inp[i][j] if i != 0 and dtp[i].count > 1 else inp[i] for i in range(len(inp))], j, dtype.scalar()) \
+              for j in range(dtype.count)]
           else:
-            ul[i] = load(inp)
+            ul[i] = load(inp, 0, dtype)
         elif uop is Ops.GEP: ul[i] = inp[0][get_single_element(arg)]
         elif uop is Ops.WMMA:
           # here are the models for the WMMA instruction on the different hardware
@@ -188,7 +200,7 @@ class PythonProgram:
           else: raise NotImplementedError(f"unimplemented tensor core {arg}")
         elif uop in GroupOp.ALU:
           assert all_same([len(x) for x in inp]), f"{[len(x) for x in inp]} doesn't match on {uop}"
-          assert all_same([dtype] + dtp) or uop in {Ops.CMPNE, Ops.CMPLT, Ops.WHERE}, f"dtype mismatch on {uop}"
+          assert all_same([dtype] + dtp) or uop in {*GroupOp.Comparison, Ops.WHERE}, f"dtype mismatch on {uop}"
           ul[i] = [exec_alu(uop, dtype, p) for p in zip(*inp)]
         assert i in ul, (uop, dtype, idp, arg)
         i += 1
@@ -196,6 +208,7 @@ class PythonProgram:
 
 class PythonRenderer(Renderer):
   device = "PYTHON"
+  code_for_op = python_alu
   def __init__(self):
     if getenv("EMULATE_METAL"): self.device, self.tensor_cores = "METAL", tc.metal
     if getenv("EMULATE_AMD"): self.device, self.tensor_cores = "AMD", tc.amd_rdna3

tinygrad/runtime/support/am/amdev.py

@@ -104,7 +104,7 @@ class AMPageTableEntry:
   def entry(self, entry_id:int) -> int: return self.entries[entry_id]
   def valid(self, entry_id:int) -> bool: return (self.entries[entry_id] & am.AMDGPU_PTE_VALID) != 0
   def address(self, entry_id:int) -> int: return self.entries[entry_id] & 0x0000FFFFFFFFF000
-  def is_huge_page(self, entry_id:int) -> bool: return self.lv == am.AMDGPU_VM_PTB or self.adev.gmc.is_pte_huge_page(self.entries[entry_id])
+  def is_page(self, entry_id:int) -> bool: return self.lv == am.AMDGPU_VM_PTB or self.adev.gmc.is_pte_huge_page(self.entries[entry_id])
   def supports_huge_page(self, paddr:int): return self.lv >= am.AMDGPU_VM_PDB2
 
 class AMMemoryManager(MemoryManager):

tinygrad/runtime/support/am/ip.py

@@ -224,7 +224,8 @@ class AM_GFX(AM_IP):
     self._grbm_select()
     self.adev.regGCVM_CONTEXT0_CNTL.write(0)
 
-  def setup_ring(self, ring_addr:int, ring_size:int, rptr_addr:int, wptr_addr:int, eop_addr:int, eop_size:int, doorbell:int, pipe:int, queue:int):
+  def setup_ring(self, ring_addr:int, ring_size:int, rptr_addr:int, wptr_addr:int, eop_addr:int, eop_size:int, doorbell:int, pipe:int, queue:int,
+                 aql:bool):
     mqd = self.adev.mm.valloc(0x1000, uncached=True, contiguous=True)
 
     struct_t = getattr(am, f"struct_v{self.adev.ip_ver[am.GC_HWIP][0]}_compute_mqd")
@@ -235,9 +236,10 @@ class AM_GFX(AM_IP):
       cp_hqd_pq_rptr_report_addr_lo=lo32(rptr_addr), cp_hqd_pq_rptr_report_addr_hi=hi32(rptr_addr),
       cp_hqd_pq_wptr_poll_addr_lo=lo32(wptr_addr), cp_hqd_pq_wptr_poll_addr_hi=hi32(wptr_addr),
       cp_hqd_pq_doorbell_control=self.adev.regCP_HQD_PQ_DOORBELL_CONTROL.encode(doorbell_offset=doorbell*2, doorbell_en=1),
-      cp_hqd_pq_control=self.adev.regCP_HQD_PQ_CONTROL.encode(rptr_block_size=5, unord_dispatch=0, queue_size=(ring_size//4).bit_length()-2),
+      cp_hqd_pq_control=self.adev.regCP_HQD_PQ_CONTROL.encode(rptr_block_size=5, unord_dispatch=0, queue_size=(ring_size//4).bit_length()-2,
+        **({'queue_full_en':1, 'slot_based_wptr':2, 'no_update_rptr':1} if aql else {})),
       cp_hqd_ib_control=self.adev.regCP_HQD_IB_CONTROL.encode(min_ib_avail_size=0x3), cp_hqd_hq_status0=0x20004000,
-      cp_mqd_control=self.adev.regCP_MQD_CONTROL.encode(priv_state=1), cp_hqd_vmid=0,
+      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))
 

tinygrad/runtime/support/hcq.py

@@ -383,15 +383,20 @@ class HCQCompiled(Compiled, Generic[SignalType]):
     self.kernargs_buf:HCQBuffer = self.allocator.alloc(kernargs_size, BufferSpec(cpu_access=True))
     self.kernargs_offset_allocator:BumpAllocator = BumpAllocator(self.kernargs_buf.size, wrap=True)
 
+    self.error_state:Exception|None = None # Exception if error is unrecoverable and sync will always fail
+
     if self._is_cpu(): HCQCompiled.cpu_devices.append(self)
 
   def synchronize(self):
+    if self.error_state is not None: raise self.error_state
+
     # If we have any work on CPU devices, need to synchronize them. This is just an optimization to release GIL allowing to finish faster.
     if not self._is_cpu():
       for dev in HCQCompiled.cpu_devices: dev.synchronize()
 
     try: self.timeline_signal.wait(self.timeline_value - 1)
     except RuntimeError as e:
+      self.error_state = e
       if hasattr(self, 'on_device_hang'): self.on_device_hang()
       else: raise e