bump gpuocelot

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

@@ -183,7 +183,7 @@ runs:
       uses: actions/cache@v4
       with:
         path: /var/cache/apt/archives/
-        key: ${{ runner.os }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.APT_CACHE_VERSION }}
+        key: ${{ runner.os }}-apt-${{ steps.apt-pkgs.outputs.hash }}
 
     - name: Run apt Update + Install
       if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.cuda == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true')
@@ -234,7 +234,7 @@ runs:
         cache-name: cache-gpuocelot-build
       with:
         path: ${{ github.workspace }}/gpuocelot/ocelot
-        key: ${{ runner.os }}-gpuocelot-b16039dc940dc6bc4ea0a98380495769ff35ed99-rebuild-${{ env.BUILD_CACHE_VERSION }}
+        key: ${{ runner.os }}-gpuocelot-b16039dc940dc6bc4ea0a98380495769ff35ed99-rebuild-1
     - name: Clone/compile gpuocelot
       if: inputs.ocelot == 'true' && steps.cache-build.outputs.cache-hit != 'true'
       shell: bash

.github/workflows/benchmark.yml

@@ -63,7 +63,7 @@ jobs:
     - name: Run model inference benchmark
       run: METAL=1 python3.11 test/external/external_model_benchmark.py
     - name: Test speed vs torch
-      run: BIG=2 MPS=1 python3.11 test/speed/external_test_speed_v_torch.py | tee torch_speed.txt
+      run: BIG=2 MPS=1 python3.11 test/external/external_test_speed_v_torch.py | tee torch_speed.txt
     - name: Test tensor cores
       run: METAL=1 python3.11 test/test_linearizer.py TestLinearizer.test_tensor_cores TestLinearizer.test_tensor_cores_padded TestLinearizer.test_tensor_cores_padded_uops
     - name: Test AMX tensor cores
@@ -187,7 +187,7 @@ jobs:
     - name: Run model inference benchmark
       run: NV=1 CAPTURE_PROCESS_REPLAY=0 NOCLANG=1 python3 test/external/external_model_benchmark.py
     - name: Test speed vs torch
-      run: NV=1 CAPTURE_PROCESS_REPLAY=0 HALF=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py | tee torch_speed.txt
+      run: NV=1 CAPTURE_PROCESS_REPLAY=0 HALF=1 BIG=2 TORCHCUDA=1 python3 test/external/external_test_speed_v_torch.py | tee torch_speed.txt
     - name: Test speed vs theoretical
       run: NV=1 IGNORE_BEAM_CACHE=1 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
     - name: Test benchmark allreduce
@@ -389,7 +389,7 @@ jobs:
     #- name: Test speed vs torch
     #  run: |
     #    python3 -c "import torch; print(torch.__version__)"
-    #    LD_PRELOAD="/opt/rocm/lib/libhsa-runtime64.so" HSA=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py | tee torch_speed.txt
+    #    LD_PRELOAD="/opt/rocm/lib/libhsa-runtime64.so" HSA=1 BIG=2 TORCHCUDA=1 python3 test/external/external_test_speed_v_torch.py | tee torch_speed.txt
     - name: Test speed vs theoretical
       run: AMD=1 IGNORE_BEAM_CACHE=1 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
     - name: Test tensor cores

.github/workflows/test.yml

@@ -1,10 +1,8 @@
 name: Unit Tests
 env:
   # increment this when downloads substantially change to avoid the internet
-  DOWNLOAD_CACHE_VERSION: '12'
-  PYTHON_CACHE_VERSION: '3'
-  APT_CACHE_VERSION: '1'
-  BUILD_CACHE_VERSION: '1'
+  DOWNLOAD_CACHE_VERSION: '11'
+  PYTHON_CACHE_VERSION: '2'
   CAPTURE_PROCESS_REPLAY: 1
   GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
 
@@ -32,9 +30,9 @@ jobs:
     - name: External Benchmark Schedule
       run: PYTHONPATH="." python3 test/external/external_benchmark_schedule.py
     - name: Speed Test
-      run: LLVM=1 python3 test/speed/external_test_speed_v_torch.py
+      run: LLVM=1 python3 test/external/external_test_speed_v_torch.py
     - name: Speed Test (BEAM=2)
-      run: BEAM=2 LLVM=1 python3 test/speed/external_test_speed_v_torch.py
+      run: BEAM=2 LLVM=1 python3 test/external/external_test_speed_v_torch.py
 
   docs:
     name: Docs
@@ -729,7 +727,7 @@ jobs:
         backend: [ptx, nv]
 
     name: Linux (${{ matrix.backend }})
-    runs-on: ubuntu-22.04
+    runs-on: ubuntu-24.04
     timeout-minutes: 20
 
     steps:

test/external/external_benchmark_kernel_launch.py

@@ -2,7 +2,7 @@ import time
 from tinygrad import Tensor, TinyJit, Device, Context
 from tinygrad.helpers import Profiling, Timing, GlobalCounters
 
-# python3 test/speed/external_test_speed_v_torch.py TestSpeed.test_add_a
+# python3 test/external/external_test_speed_v_torch.py TestSpeed.test_add_a
 
 @TinyJit
 def plus(a:Tensor, b:Tensor): return a+b

test/external/process_replay/process_replay.py

@@ -99,6 +99,7 @@ def diff(offset:int, fxns:dict[str, Callable[..., tuple|None]]) -> None:
     except Exception as e:
       changed += 1
       warnings.warn(f"{name=} {loc=} {e=}", ProcessReplayWarning)
+  conn.commit()
   cur.close()
 
 # *** generic runner to map rows of a table to a function in parallel
@@ -110,11 +111,12 @@ def _pmap(fxns:dict[str, Callable]) -> None:
   except sqlite3.OperationalError:
     raise RuntimeError(f"{TABLE_NAME} isn't accessible in master, did DB_VERSION change?")
   finally:
+    conn.commit()
     cur.close()
 
   with multiprocessing.get_context("spawn").Pool(multiprocessing.cpu_count()) as pool:
-    bar = tqdm(total=row_count)
-    for _ in pool.imap_unordered(functools.partial(diff, fxns=fxns), range(0, row_count, PAGE_SIZE)): bar.update(PAGE_SIZE)
+    inputs = list(range(0, row_count, PAGE_SIZE))
+    list(tqdm(pool.imap_unordered(functools.partial(diff, fxns=fxns), inputs), total=len(inputs)))
     pool.close()
     pool.join()
     pool.terminate()

test/test_compile_failures.py

@@ -1,10 +1,7 @@
-import unittest, io
-from contextlib import redirect_stdout
+import unittest
 from tinygrad import Tensor, dtypes, Device
-from tinygrad.helpers import OSX
 from tinygrad.engine.realize import lower_schedule
 from tinygrad.device import is_dtype_supported
-from tinygrad.engine.realize import get_program
 
 class TestCompileFailures(unittest.TestCase):
   def compile(self, out:Tensor):
@@ -17,17 +14,5 @@ class TestCompileFailures(unittest.TestCase):
   def test_add_max_uchar(self):
     self.compile((Tensor.empty(1024, dtype='uint8') + Tensor.empty(1024, dtype='uint8')).max())
 
-class TestDisassembly(unittest.TestCase):
-  # TODO: fails on llvm. llvm.LLVMGetHostCPUName() returns "generic"
-  @unittest.skipUnless(Device.DEFAULT in ("CPU",) and OSX, "m series cpus support fp16 arithmetic")
-  def test_float16_alu(self):
-    c = Tensor([1], dtype=dtypes.float16) + Tensor([1], dtype=dtypes.float16)
-    s = c.schedule()[-1]
-    p = get_program(s.ast, Device[Device.DEFAULT].renderer)
-    lib = Device[Device.DEFAULT].compiler.compile(p.src)
-    out = io.StringIO()
-    with redirect_stdout(out): Device[Device.DEFAULT].compiler.disassemble(lib)
-    assert "fcvt" not in out.getvalue()
-
 if __name__ == '__main__':
   unittest.main()

test/test_disassembly.py

@@ -0,0 +1,21 @@
+import unittest, io
+from tinygrad import Tensor, dtypes
+from contextlib import redirect_stdout
+from tinygrad.device import Device
+from tinygrad.helpers import OSX
+from tinygrad.engine.realize import get_program
+
+class TestDisassembly(unittest.TestCase):
+  # TODO: fails on llvm. llvm.LLVMGetHostCPUName() returns "generic"
+  @unittest.skipUnless(Device.DEFAULT in ("CPU",) and OSX, "m series cpus support fp16 arithmetic")
+  def test_float16_alu(self):
+    c = Tensor([1], dtype=dtypes.float16) + Tensor([1], dtype=dtypes.float16)
+    s = c.schedule()[-1]
+    p = get_program(s.ast, Device[Device.DEFAULT].renderer)
+    lib = Device[Device.DEFAULT].compiler.compile(p.src)
+    out = io.StringIO()
+    with redirect_stdout(out): Device[Device.DEFAULT].compiler.disassemble(lib)
+    assert "fcvt" not in out.getvalue()
+
+if __name__ == "__main__":
+  unittest.main()

test/test_softmax_fusion.py

@@ -41,7 +41,7 @@ class TestFuse(unittest.TestCase):
 
   def test_fuse_norm(self):
     a = Tensor.rand(50,50).realize()
-    self._test_fuse(lambda a: a / a.mean(axis=1), a, atol=1e-6)
+    self._test_fuse(lambda a: a / a.mean(axis=1), a)
 
   def test_fuse_argmax(self):
     a = Tensor.rand(50,50).realize()

test/test_uops_stats.py

@@ -20,7 +20,6 @@ def get_stats(x:Tensor):
   ei = lower_schedule_item(si)
   return ei.prg.estimates.ops, ei.prg.estimates.mem
 
-@unittest.skipIf(Device.DEFAULT == "WEBGPU", "webgpu does extra load/store for packed types")
 class TestMemoryCount(unittest.TestCase):
   def test_add(self):
     a = Tensor.empty(1024, 1024, dtype=dtypes.uint8)

tinygrad/renderer/__init__.py

@@ -1,7 +1,7 @@
 from __future__ import annotations
 from typing import Callable, cast, TYPE_CHECKING
-import functools
-from dataclasses import dataclass, field
+import functools, itertools
+from dataclasses import dataclass, field, replace
 from tinygrad.helpers import to_function_name, dedup, prod
 from tinygrad.uop.ops import Ops, UOp, sym_infer, sint, Variable, ssimplify, GroupOp, PatternMatcher
 from tinygrad.dtype import AddrSpace, PtrDType
@@ -23,7 +23,6 @@ class Estimates:
   def from_uops(uops:list[UOp], ignore_indexing=False) -> Estimates:
     flops: sint = 0
     lds: sint = 0
-    mem: dict[tuple[UOp, Ops], sint] = {}
     mults: sint = 1
     mult_stack: list[sint] = []
     dont_count: set[UOp] = set()
@@ -35,11 +34,6 @@ class Estimates:
         elif u.op is Ops.IF:
           dont_count = dont_count.union(u.src[0].toposort())
     for u in uops:
-      if u.op in {Ops.LOAD, Ops.STORE}:
-        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
       if u.op is Ops.RANGE:
         mult_stack.append(mults)
         mults *= cast(sint, u.src[0].ssimplify())
@@ -53,7 +47,7 @@ class Estimates:
         lds += u.src[1].dtype.itemsize * mults
       elif u.op in GroupOp.ALU and u not in dont_count: flops += (mults * (2 if u.op is Ops.MULACC else 1)) * u.dtype.count
       elif u.op is Ops.WMMA and u not in dont_count: flops += 2 * prod(u.arg[1]) // u.arg[5] * mults
-    return Estimates(flops, lds, sum(mem.values()))
+    return Estimates(flops, lds, lds) # TODO: properly track memory, lds is always a high estimate
 
 @dataclass
 class ProgramSpec:
@@ -90,9 +84,17 @@ class ProgramSpec:
       self.ins = sorted(dedup(self.ins))
       self._ran_post_init = True
 
+  @functools.cached_property
+  def mem_estimate(self) -> sint:
+    # group non-local bufs by the op type (LOAD or STORE) and the buffer arg. take the max access of that buffer in bytes
+    # TODO: these max and min don't work on symbolic, and results are very wrong.
+    return sum(max(x.src[0].dtype.nbytes() for x in group)
+      for _, group in itertools.groupby([x for x in self.ast.toposort() if x.op in {Ops.LOAD, Ops.STORE} and x.src[0].base.op is Ops.DEFINE_GLOBAL],
+                        key=lambda x: (x.op, x.src[0].base.arg)))
+
   @functools.cached_property
   def estimates(self) -> Estimates:
-    return Estimates() if self.uops is None else Estimates.from_uops(self.uops, ignore_indexing=True)
+    return replace(Estimates() if self.uops is None else Estimates.from_uops(self.uops, ignore_indexing=True), mem=self.mem_estimate)
 
   @functools.cached_property
   def function_name(self) -> str: return to_function_name(self.name)

tinygrad/schedule/kernelize.py

@@ -91,8 +91,8 @@ kernelize_sym = symbolic_simple+PatternMatcher([
   (UPat(Ops.CAST, name="cast", src=(UPat(Ops.VIEW, name="vm"),)), lambda cast,vm: vm.base.cast(cast.dtype).view(vm.st)
     if cast.dtype.itemsize <= vm.dtype.itemsize and resolve(prod(vm.shape) > vm.st.real_size()) else None),
   # put UnaryOps before EXPANDs, if it can fuse with the input
-  #(UPat(GroupOp.Unary, src=(UPat(Ops.VIEW, src=(UPat(GroupOp.All-ALWAYS_CONTIGUOUS, name="inp"),), name="v"),), name="alu"),
-  # lambda inp,v,alu: inp.alu(alu.op).view(v.st) if resolve(prod(alu.shape) > v.st.real_size()) else None),
+  (UPat(GroupOp.Unary, src=(UPat(Ops.VIEW, src=(UPat(GroupOp.All-ALWAYS_CONTIGUOUS, name="inp"),), name="v"),), name="alu"),
+   lambda inp,v,alu: inp.alu(alu.op).view(v.st) if resolve(prod(alu.shape) > v.st.real_size()) else None),
 ])
 
 # support for using a contiguous permuted view instead of the parent view if one exists