start disabling rules there's no tests for

* no ast for mem estimate

* skip for webgpu

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

@@ -121,7 +121,7 @@ runs:
         echo 'Acquire::GzipIndexes "true";' | sudo tee /etc/apt/apt.conf.d/gzip
         echo 'Acquire::http::Pipeline-Depth "5";' | sudo tee -a /etc/apt/apt.conf.d/99parallel
         echo 'Binary::apt::APT::Keep-Downloaded-Packages "true";' | sudo tee -a /etc/apt/apt.conf.d/99keep-debs
-    
+
     - name: Add OpenCL Repo
       if: inputs.opencl == 'true' && runner.os == 'Linux'
       shell: bash
@@ -174,7 +174,7 @@ runs:
         if [[ "${{ inputs.llvm }}" == "true" ]]; then
           pkgs+=" libllvm20 clang-20 lld-20"
         fi
-        
+
         echo "pkgs=$pkgs" >> "$GITHUB_OUTPUT"
         echo "hash=$(echo -n "$pkgs" | sha256sum | cut -d' ' -f1)" >> "$GITHUB_OUTPUT"
 
@@ -183,21 +183,21 @@ runs:
       uses: actions/cache@v4
       with:
         path: /var/cache/apt/archives/
-        key: ${{ runner.os }}-apt-${{ steps.apt-pkgs.outputs.hash }}
+        key: ${{ runner.os }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.APT_CACHE_VERSION }}
 
     - name: Run apt Update + Install
       if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.cuda == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true')
       shell: bash
       run: |
         sudo apt -qq update || true
-        
+
         # ******** do install ********
         if [[ -n "${{ steps.apt-pkgs.outputs.pkgs }}" ]]; then
           sudo apt-get -y --allow-unauthenticated --no-install-recommends install ${{ steps.apt-pkgs.outputs.pkgs }}
         fi
-        
+
         sudo chown -R $USER:$USER /var/cache/apt/archives/
-    
+
     # **** AMD ****
     - name: Setup AMD (Linux)
       if: inputs.amd == 'true' && runner.os == 'Linux'
@@ -234,7 +234,7 @@ runs:
         cache-name: cache-gpuocelot-build
       with:
         path: ${{ github.workspace }}/gpuocelot/ocelot
-        key: ${{ runner.os }}-gpuocelot-b16039dc940dc6bc4ea0a98380495769ff35ed99-rebuild-0
+        key: ${{ runner.os }}-gpuocelot-b16039dc940dc6bc4ea0a98380495769ff35ed99-rebuild-${{ env.BUILD_CACHE_VERSION }}
     - 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/test_speed_v_torch.py | tee torch_speed.txt
+      run: BIG=2 MPS=1 python3.11 test/speed/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/test_speed_v_torch.py | tee torch_speed.txt
+      run: NV=1 CAPTURE_PROCESS_REPLAY=0 HALF=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py | tee torch_speed.txt
     - name: Test speed vs theoretical
       run: NV=1 IGNORE_BEAM_CACHE=1 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
     - 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/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/speed/external_test_speed_v_torch.py | tee torch_speed.txt
     - name: Test speed vs theoretical
       run: AMD=1 IGNORE_BEAM_CACHE=1 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
     - name: Test tensor cores

.github/workflows/test.yml

@@ -1,8 +1,10 @@
 name: Unit Tests
 env:
   # increment this when downloads substantially change to avoid the internet
-  DOWNLOAD_CACHE_VERSION: '10'
-  PYTHON_CACHE_VERSION: '2'
+  DOWNLOAD_CACHE_VERSION: '12'
+  PYTHON_CACHE_VERSION: '3'
+  APT_CACHE_VERSION: '1'
+  BUILD_CACHE_VERSION: '1'
   CAPTURE_PROCESS_REPLAY: 1
   GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
 
@@ -30,9 +32,9 @@ jobs:
     - name: External Benchmark Schedule
       run: PYTHONPATH="." python3 test/external/external_benchmark_schedule.py
     - name: Speed Test
-      run: LLVM=1 python3 test/test_speed_v_torch.py
+      run: LLVM=1 python3 test/speed/external_test_speed_v_torch.py
     - name: Speed Test (BEAM=2)
-      run: BEAM=2 LLVM=1 python3 test/test_speed_v_torch.py
+      run: BEAM=2 LLVM=1 python3 test/speed/external_test_speed_v_torch.py
 
   docs:
     name: Docs
@@ -46,6 +48,11 @@ jobs:
       with:
         deps: docs
         pydeps: "capstone"
+    - name: Build wheel and show size
+      run: |
+        pip install build
+        python -m build --wheel --outdir dist
+        ls -lh dist/*.whl
     - name: Use as an external package
       run: |
         mkdir $HOME/test_external_dir
@@ -453,7 +460,7 @@ jobs:
   testopenpilot:
     name: 'openpilot Compile Tests'
     runs-on: ubuntu-22.04
-    timeout-minutes: 10
+    timeout-minutes: 15
     env:
       IGNORE_OOB: 0
     steps:
@@ -584,6 +591,29 @@ jobs:
       - name: Run process replay tests
         uses: ./.github/actions/process-replay
 
+  testdevectorize:
+    name: Linux (devectorize)
+    runs-on: ubuntu-24.04
+    timeout-minutes: 15
+    env:
+      IGNORE_OOB: 0
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v4
+    - name: Setup Environment
+      uses: ./.github/actions/setup-tinygrad
+      with:
+        key: devectorize-minimal
+        deps: testing_minimal
+        pydeps: "pillow"
+        llvm: "true"
+    - name: Test LLVM=1 DEVECTORIZE=0
+      run: LLVM=1 DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/test_ops.py -k "not test_avg_pool3d_failure"
+    - name: Test LLVM=1 DEVECTORIZE=0 for model
+      run: PYTHONPATH="." LLVM=1 DEVECTORIZE=0 python3 test/models/test_efficientnet.py
+    - name: Test CPU=1 DEVECTORIZE=0
+      run: CPU=1 DEVECTORIZE=0 FUSE_ARANGE=0 python3 -m pytest -n auto test/test_tiny.py test/test_ops.py -k "not test_avg_pool3d_failure"
+
   testdsp:
     name: Linux (DSP)
     runs-on: ubuntu-24.04
@@ -619,12 +649,6 @@ jobs:
       run: CC=clang-20 PYTHONPATH="." DEBUG=2 DSP=1 python test/test_transcendental.py TestTranscendentalVectorized
     - name: Test quantize onnx
       run: PYTHONPATH="." DEBUG=2 DSP=1 python3 test/test_quantize_onnx.py
-    - name: Test LLVM=1 DEVECTORIZE=0
-      run: LLVM=1 DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/test_ops.py -k "not test_avg_pool3d_failure"
-    - name: Test LLVM=1 DEVECTORIZE=0 for model
-      run: PYTHONPATH="." LLVM=1 DEVECTORIZE=0 python3 test/models/test_efficientnet.py
-    - name: Test CPU=1 DEVECTORIZE=0
-      run: CPU=1 DEVECTORIZE=0 FUSE_ARANGE=0 python3 -m pytest -n auto test/test_tiny.py test/test_ops.py -k "not test_avg_pool3d_failure"
 
   testwebgpu:
     name: Linux (WebGPU)
@@ -684,9 +708,9 @@ jobs:
           DEBUG=5 FORWARD_ONLY=1 python3 test/test_ops.py TestOps.test_add
       - name: Run LLVM test
         if: matrix.backend=='amdllvm'
-        run: python test/test_amd_llvm.py
+        run: python test/device/test_amd_llvm.py
       - name: Run pytest (amd)
-        run: python -m pytest -n=auto test/test_ops.py test/test_dtype.py test/test_dtype_alu.py test/test_linearizer.py test/test_randomness.py test/test_jit.py test/test_graph.py test/test_multitensor.py test/test_hcq.py --durations=20
+        run: python -m pytest -n=auto test/test_ops.py test/test_dtype.py test/test_dtype_alu.py test/test_linearizer.py test/test_randomness.py test/test_jit.py test/test_graph.py test/test_multitensor.py test/device/test_hcq.py --durations=20
       - name: Run pytest (amd)
         run: python -m pytest test/external/external_test_am.py --durations=20
       - name: Run TRANSCENDENTAL math
@@ -811,14 +835,14 @@ jobs:
         AMD: 1
         FORWARD_ONLY: 1
       run: |
-        python3 -m pytest -n=auto test/test_hcq.py test/test_tiny.py --durations=20
+        python3 -m pytest -n=auto test/device/test_hcq.py test/test_tiny.py --durations=20
     - name: Run pytest (amd with llvm backend)
       env:
         MOCKGPU: 1
         AMD: 1
         FORWARD_ONLY: 1
       run: |
-        python -m pytest -n=auto test/test_hcq.py test/test_tiny.py test/test_amd_llvm.py --durations=20
+        python -m pytest -n=auto test/device/test_hcq.py test/test_tiny.py test/device/test_amd_llvm.py --durations=20
     - name: Run pytest (ptx)
       env:
         MOCKGPU: 1
@@ -826,7 +850,7 @@ jobs:
         NV: 1
         FORWARD_ONLY: 1
       run: |
-        python3 -m pytest -n=auto test/test_hcq.py test/test_tiny.py --durations=20
+        python3 -m pytest -n=auto test/device/test_hcq.py test/test_tiny.py --durations=20
     - name: Run process replay tests
       uses: ./.github/actions/process-replay
 
@@ -937,18 +961,18 @@ jobs:
         env:
           HOST: 127.0.0.1:6667*6,127.0.0.1:6668*6
         run: |
-          python3 -m pytest test/test_tiny.py test/test_jit.py test/test_subbuffer.py test/test_graph.py test/test_multitensor.py test/test_remote.py test/test_tensor_variable.py
+          python3 -m pytest test/test_tiny.py test/test_jit.py test/test_subbuffer.py test/test_graph.py test/test_multitensor.py test/test_remote.py test/test_tensor_variable.py --durations 20
       - name: Run REMOTE=1 Test (GPU)
         env:
           HOST: 127.0.0.1:7667*6
         run: |
-          python3 -m pytest test/test_tiny.py test/test_image_dtype.py test/test_jit.py
+          python3 -m pytest test/test_tiny.py test/test_image_dtype.py test/test_jit.py --durations 20
           IMAGE=2 python3 -m pytest test/test_tiny.py test/test_image_dtype.py
       - name: Run REMOTE=1 Test (CPU)
         env:
           HOST: 127.0.0.1:8667*6
         run: |
-          python3 -m pytest test/test_tiny.py test/test_jit.py test/test_multitensor.py
+          python3 -m pytest test/test_tiny.py test/test_jit.py test/test_multitensor.py --durations 20
       - name: Show remote server logs
         if: always()
         run: |

autogen_stubs.sh

@@ -198,11 +198,7 @@ generate_amd() {
   clang2py -k cdefstum \
     extra/hip_gpu_driver/sdma_registers.h \
     extra/hip_gpu_driver/nvd.h \
-    extra/hip_gpu_driver/kfd_pm4_headers_ai.h \
-    extra/hip_gpu_driver/soc21_enum.h \
-    extra/hip_gpu_driver/sdma_v6_0_0_pkt_open.h \
     extra/hip_gpu_driver/gc_11_0_0_offset.h \
-    extra/hip_gpu_driver/gc_10_3_0_offset.h \
     extra/hip_gpu_driver/sienna_cichlid_ip_offset.h \
     --clang-args="-I/opt/rocm/include -x c++" \
     -o $BASE/amd_gpu.py
@@ -376,26 +372,6 @@ generate_am() {
     -o $BASE/am/pm4_nv.py
   fixup $BASE/am/pm4_nv.py
 
-  clang2py -k cdefstum \
-    $AMKERN_INC/vega10_enum.h \
-    -o $BASE/am/vega10.py
-  fixup $BASE/am/vega10.py
-
-  clang2py -k cdefstum \
-    $AMKERN_INC/navi10_enum.h \
-    -o $BASE/am/navi10.py
-  fixup $BASE/am/navi10.py
-
-  clang2py -k cdefstum \
-    $AMKERN_INC/soc21_enum.h \
-    -o $BASE/am/soc21.py
-  fixup $BASE/am/soc21.py
-
-  clang2py -k cdefstum \
-    $AMKERN_INC/soc24_enum.h \
-    -o $BASE/am/soc24.py
-  fixup $BASE/am/soc24.py
-
   clang2py -k cdefstum \
     extra/hip_gpu_driver/sdma_registers.h \
     $AMKERN_AMD/amdgpu/vega10_sdma_pkt_open.h \

examples/beautiful_mnist.py

@@ -29,8 +29,7 @@ if __name__ == "__main__":
     opt.zero_grad()
     samples = Tensor.randint(getenv("BS", 512), high=X_train.shape[0])
     loss = model(X_train[samples]).sparse_categorical_crossentropy(Y_train[samples]).backward()
-    opt.step()
-    return loss
+    return loss.realize(*opt.schedule_step())
 
   @TinyJit
   def get_test_acc() -> Tensor: return (model(X_test).argmax(axis=1) == Y_test).mean()*100

pytest.ini

@@ -0,0 +1,2 @@
+[pytest]
+norecursedirs = extra

setup.py

@@ -18,7 +18,7 @@ testing_minimal = [
 ]
 
 setup(name='tinygrad',
-      version='0.10.3',
+      version='0.11.0',
       description='You like pytorch? You like micrograd? You love tinygrad! <3',
       author='George Hotz',
       license='MIT',

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/test_speed_v_torch.py TestSpeed.test_add_a
+# python3 test/speed/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

@@ -99,7 +99,6 @@ 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
@@ -111,12 +110,11 @@ 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:
-    inputs = list(range(0, row_count, PAGE_SIZE))
-    list(tqdm(pool.imap_unordered(functools.partial(diff, fxns=fxns), inputs), total=len(inputs)))
+    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)
     pool.close()
     pool.join()
     pool.terminate()

test/mockgpu/amd/amdgpu.py

@@ -1,7 +1,7 @@
 import ctypes, time
 from test.mockgpu.gpu import VirtGPU
 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, tinygrad.runtime.autogen.am.soc21 as soc21
+import tinygrad.runtime.autogen.amd_gpu as amd_gpu, tinygrad.runtime.autogen.am.pm4_nv as pm4
 
 SDMA_MAX_COPY_SIZE = 0x400000
 
@@ -14,6 +14,9 @@ regSQ_THREAD_TRACE_BUF0_SIZE = 0x39e9 + amd_gpu.GC_BASE__INST0_SEG1
 regSQ_THREAD_TRACE_WPTR = 0x39ef + amd_gpu.GC_BASE__INST0_SEG1
 regSQ_THREAD_TRACE_STATUS = 0x39f4 + amd_gpu.GC_BASE__INST0_SEG1
 
+class SQTT_EVENTS:
+  THREAD_TRACE_FINISH = 0x00000037
+
 CACHE_FLUSH_AND_INV_TS_EVENT = 0x14
 
 WAIT_REG_MEM_FUNCTION_ALWAYS = 0
@@ -208,7 +211,7 @@ class PM4Executor(AMDQueue):
     assert n == 0
     event_dw = self._next_dword()
     match (event_dw & 0xFF): # event type
-      case soc21.THREAD_TRACE_FINISH:
+      case SQTT_EVENTS.THREAD_TRACE_FINISH:
         old_idx = self.gpu.regs.grbm_index
         for se in range(self.gpu.regs.n_se):
           self.gpu.regs.grbm_index = 0b011 << 29 | se << 16 # select se, broadcast sa and instance

test/models/test_onnx.py

@@ -32,6 +32,7 @@ OPENPILOT_MODEL = "https://github.com/commaai/openpilot/raw/v0.9.4/selfdrive/mod
 np.random.seed(1337)
 
 class TestOnnxModel(unittest.TestCase):
+  @unittest.skip("this isn't a test, it can't fail")
   def test_benchmark_openpilot_model(self):
     onnx_model = fetch(OPENPILOT_MODEL)
     run_onnx = OnnxRunner(onnx_model)

test/models/test_whisper.py

@@ -16,7 +16,7 @@ TRANSCRIPTION_2 = "a slightly longer audio file so that we can test batch transc
 TEST_FILE_3_URL = 'https://homepage.ntu.edu.tw/~karchung/miniconversations/mc45.mp3'
 TRANSCRIPTION_3 = "Just lie back and relax. Is the level of pressure about right? Yes, it's fine, and I'd like conditioner please. Sure. I'm going to start the second lathering now. Would you like some Q-tips? How'd you like it cut? I'd like my bangs and the back trimmed, and I'd like the rest thinned out a bit and layered. Where would you like the part? On the left, right about here. Here, have a look. What do you think? It's fine. Here's a thousand anti-dollars. It's 30-ant extra for the rants. Here's your change and receipt. Thank you, and please come again. So how do you like it? It could have been worse, but you'll notice that I didn't ask her for her card. Hmm, yeah. Maybe you can try that place over there next time."   # noqa: E501
 
-@unittest.skipIf(CI and Device.DEFAULT in ["CPU"], "slow")
+@unittest.skipIf(Device.DEFAULT in ["CPU", "LLVM"], "slow")
 @unittest.skipUnless(is_dtype_supported(dtypes.float16), "need float16 support")
 class TestWhisper(unittest.TestCase):
   @classmethod

test/speed/external_test_copy_speed.py

@@ -3,7 +3,7 @@ from tinygrad import Tensor, Device, TinyJit
 from tinygrad.helpers import Timing, CI, OSX
 import multiprocessing.shared_memory as shared_memory
 
-N = 256 if CI else 4096
+N = 256
 class TestCopySpeed(unittest.TestCase):
   @classmethod
   def setUpClass(cls): Device[Device.DEFAULT].synchronize()

test/test_compile_failures.py

@@ -1,7 +1,10 @@
-import unittest
+import unittest, io
+from contextlib import redirect_stdout
 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):
@@ -14,5 +17,17 @@ 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

@@ -1,21 +0,0 @@
-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_multitensor.py

@@ -373,11 +373,11 @@ class TestMultiTensor(unittest.TestCase):
     np.testing.assert_allclose(y.numpy(), y_shard.numpy(), atol=1e-6, rtol=1e-6)
 
   # NOTE: this is failing on LLVM CI, no idea why. Works locally.
-  @unittest.skipIf(CI and REAL_DEV in ("CUDA", "NV", "LLVM", "CPU"), "slow, and flaky on LLVM/CPU")
+  @unittest.skipIf(CI and REAL_DEV in ("CUDA", "NV", "LLVM", "CPU", "AMD"), "slow, and flaky on LLVM/CPU")
   def test_data_parallel_resnet(self):
     from extra.models.resnet import ResNet18
 
-    fake_image = Tensor.rand((2, 3, 224//8, 224//8))
+    fake_image = Tensor.rand((2, 3, 224//16, 224//16))
     fake_image_sharded = fake_image.shard(devices_2, axis=0)
     m = ResNet18()
     m.load_from_pretrained()
@@ -409,10 +409,10 @@ class TestMultiTensor(unittest.TestCase):
     # sometimes there is zeros in these grads... why?
     np.testing.assert_allclose(grad, shard_grad, atol=1e-5, rtol=1e-5)
 
-  @unittest.skipIf(CI and REAL_DEV in ("CUDA", "NV", "LLVM", "CPU"), "slow, and flaky on LLVM/CPU")
+  @unittest.skipIf(CI and REAL_DEV in ("CUDA", "NV", "LLVM", "CPU", "AMD"), "slow, and flaky on LLVM/CPU")
   def test_data_parallel_resnet_train_step(self):
     from extra.models.resnet import ResNet18
-    fake_image = Tensor.rand((2, 3, 224//8, 224//8))
+    fake_image = Tensor.rand((2, 3, 224//16, 224//16))
     labels = Tensor.randint(2, low=0, high=1000)
     m = ResNet18()
     self._test_model_train_step(m, fake_image, labels)

test/test_net_speed.py

@@ -1,87 +0,0 @@
-#!/usr/bin/env python
-import time
-import unittest
-import torch
-from tinygrad import Tensor, Device
-from tinygrad.helpers import Profiling, CI
-
-@unittest.skipIf(CI and Device.DEFAULT in {"CUDA", "NV"}, "slow")
-class TestConvSpeed(unittest.TestCase):
-
-  def test_mnist(self):
-    # https://keras.io/examples/vision/mnist_convnet/
-    conv = 3
-    inter_chan, out_chan = 32, 64
-
-    # ****** torch baseline *******
-
-    torch.backends.mkldnn.enabled = False
-
-    conv = 3
-    inter_chan, out_chan = 32, 64
-    c1 = torch.randn(inter_chan,1,conv,conv, requires_grad=True)
-    c2 = torch.randn(out_chan,inter_chan,conv,conv, requires_grad=True)
-    l1 = torch.randn(out_chan*5*5, 10, requires_grad=True)
-
-    c2d = torch.nn.functional.conv2d
-    mp = torch.nn.MaxPool2d((2,2))
-    lsm = torch.nn.LogSoftmax(dim=1)
-
-    cnt = 5
-    fpt, bpt = 0.0, 0.0
-    for i in range(cnt):
-      et0 = time.time()
-      x = torch.randn(128, 1, 28, 28, requires_grad=True)
-      x = mp(c2d(x,c1).relu())
-      x = mp(c2d(x,c2).relu())
-      x = x.reshape(x.shape[0], -1)
-      out = lsm(x.matmul(l1))
-      out = out.mean()
-      et1 = time.time()
-      out.backward()
-      et2 = time.time()
-      fpt += (et1-et0)
-      bpt += (et2-et1)
-
-    fpt_baseline = (fpt*1000/cnt)
-    bpt_baseline = (bpt*1000/cnt)
-    print("torch forward pass:  %.3f ms" % fpt_baseline)
-    print("torch backward pass: %.3f ms" % bpt_baseline)
-
-    # ****** tinygrad compare *******
-
-    c1 = Tensor(c1.detach().numpy(), requires_grad=True)
-    c2 = Tensor(c2.detach().numpy(), requires_grad=True)
-    l1 = Tensor(l1.detach().numpy(), requires_grad=True)
-
-    cnt = 5
-    fpt, bpt = 0.0, 0.0
-    for i in range(1+cnt):
-      et0 = time.time()
-      x = Tensor.randn(128, 1, 28, 28)
-      x = x.conv2d(c1).relu().avg_pool2d()
-      x = x.conv2d(c2).relu().max_pool2d()
-      x = x.reshape(shape=(x.shape[0], -1))
-      out = x.dot(l1).log_softmax()
-      out = out.mean()
-      out.backward()  # NOTE: we have to now compute this here, but it doesn't realize
-      out.realize()
-      et1 = time.time()
-      [x.grad.realize() for x in [c1, c2, l1]]
-      et2 = time.time()
-      if i == 0:
-        pr = Profiling(sort='time', frac=0.2)
-        pr.__enter__()
-      else:
-        fpt += (et1-et0)
-        bpt += (et2-et1)
-
-    pr.__exit__()
-    fpt = (fpt*1000/cnt)
-    bpt = (bpt*1000/cnt)
-    print("forward pass:  %.3f ms, %.2fx off baseline %.3f ms" % (fpt, fpt/fpt_baseline, fpt_baseline))
-    print("backward pass: %.3f ms, %.2fx off baseline %.3f ms" % (bpt, bpt/bpt_baseline, bpt_baseline))
-
-
-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)
+    self._test_fuse(lambda a: a / a.mean(axis=1), a, atol=1e-6)
 
   def test_fuse_argmax(self):
     a = Tensor.rand(50,50).realize()
@@ -117,6 +117,14 @@ class TestFuse(unittest.TestCase):
     c = (a.sum(axis=1) + b.sum(axis=1)).fuse()
     self.assertListEqual(c.tolist(), [30]*16)
 
+  @unittest.skipUnless(Device.DEFAULT == "METAL", "METAL TC")
+  def test_fuse_and_tc_opt(self):
+    A = Tensor.randn(8, 8).realize()
+    B = Tensor.randn(8, 8).realize()
+    C = Tensor.ones(1, 8, 8).pad(((1,1), None, None),).sum(0)
+    out = (C + (A @ B)).fuse()
+    out.realize()
+
 class TestSoftmaxFusion(unittest.TestCase):
   @classmethod
   def setUpClass(cls):

test/test_tiny.py

@@ -106,6 +106,24 @@ class TestTiny(unittest.TestCase):
     probs = Tensor.rand(1, 1, 28, 28).sequential(layers).tolist()
     self.assertEqual(len(probs[0]), 10)
 
+  # TODO: this is failing because of how swizzling rewrites the ShapeTracker of the final STORE
+  @unittest.skipIf(IMAGE>0 or (CI and Device.DEFAULT == "DSP"), "failing because of make things that can't be images not images")
+  def test_mnist_backward(self):
+    # NOTE: we don't have the whole model here for speed
+    layers = [
+      nn.Conv2d(1, 32, 5), Tensor.relu,
+      nn.Conv2d(32, 32, 5), Tensor.relu]
+
+    # replace random weights with ones
+    # TODO: there's a bug here where it's tying two of the biases together. we need UNIQUE const
+    #Tensor.realize(*[p.replace(Tensor.ones_like(p).contiguous()) for p in nn.state.get_parameters(layers)])
+    for p in nn.state.get_parameters(layers): p.replace(Tensor.empty(p.shape))
+
+    # realize gradients
+    for x in nn.state.get_parameters(layers): x.requires_grad_()
+    Tensor.empty(4, 1, 28, 28).sequential(layers).sum().backward()
+    Tensor.realize(*[x.grad for x in nn.state.get_parameters(layers) if x.grad is not None])
+
   # *** image ***
 
   @unittest.skipIf(Device.DEFAULT != "GPU", "image only supported on GPU")

test/test_uops_stats.py

@@ -20,6 +20,7 @@ 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)

test/unit/test_dtype.py

@@ -53,5 +53,13 @@ class TestCastConvenienceMethod(unittest.TestCase):
       self.assertEqual(t.float().dtype, dtypes.float)
       self.assertEqual(t.double().dtype, dtypes.double)
 
+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])
+    # 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
+    self.assertEqual(Tensor([-30000, 1.5, 3.1, 30000], device="PYTHON", dtype=dtypes.fp8e5m2).tolist(), [-28672.0, 1.5, 3.0, 28672.0])
+
 if __name__ == "__main__":
   unittest.main()

test/unit/test_hashing.py

@@ -2,7 +2,6 @@ from typing_extensions import Callable
 import hashlib, random, unittest
 from tinygrad import Tensor, Device, getenv, dtypes
 from tinygrad.device import is_dtype_supported
-from tinygrad.helpers import CI
 
 @unittest.skipUnless(is_dtype_supported(dtypes.uint8) and is_dtype_supported(dtypes.uint64), "Device must support uint8 and uint64")
 @unittest.skipIf(getenv("MOCKGPU") and Device.DEFAULT == "NV", "crashes in NV CI")
@@ -12,7 +11,7 @@ class TestHashing(unittest.TestCase):
     chunk_hashes = [hashlib.shake_128(chunk).digest(16) for chunk in chunks]
     return hashlib.shake_128(b''.join(chunk_hashes)).digest(16)
 
-  @unittest.skipIf(CI, "very slow")
+  @unittest.skip("very slow")
   def test_abc(self):
     expected = self._python_hash_1mb(b"abc" + b"\x00" * (2**20 - 3))
     out = Tensor(b"abc").hash()
@@ -65,7 +64,7 @@ class TestKeccak(unittest.TestCase):
     data = b"\x00" * 4
     self.assertEqual(bytes(Tensor(data).keccak("shake_128").tolist()), hashlib.shake_128(data).digest(16))
 
-    data = b"\x00" * (1000 if CI else 4096)
+    data = b"\x00" * 1000
     self.assertEqual(bytes(Tensor(data).keccak("shake_128").tolist()), hashlib.shake_128(data).digest(16))
 
 if __name__ == "__main__":

test/unit/test_uop_symbolic.py

@@ -622,6 +622,8 @@ class TestSymbolic(unittest.TestCase):
     self.helper_test_variable(cond, 0, 1, "(a<2)")
     self.helper_test_variable(cond.where(u1, u0), 0, 1, "(a<2)")
     self.helper_test_variable(cond.where(u1, u0).where(u1, u0), 0, 1, "(a<2)")
+    self.helper_test_variable(cond.where(u0, u1), 0, 1, "((a<2)!=True)")
+    self.helper_test_variable(cond.where(u0, u1).where(u0, u1), 0, 1, "(a<2)")
 
   def test_where_combine(self):
     cond = Variable("x", 0, 3) < 2

test/unit/test_viz.py

@@ -250,7 +250,7 @@ class TestVizProfiler(unittest.TestCase):
 
     j = json.loads(get_profile(prof))
 
-    dev_events = j['layout']['NV']['timeline']['shapes']
+    dev_events = j['layout']['NV']['shapes']
     self.assertEqual(len(dev_events), 1)
     event = dev_events[0]
     self.assertEqual(event['name'], 'E_2')
@@ -263,7 +263,7 @@ class TestVizProfiler(unittest.TestCase):
 
     j = json.loads(get_profile(prof))
 
-    event = j['layout']['NV']['timeline']['shapes'][0]
+    event = j['layout']['NV']['shapes'][0]
     self.assertEqual(event['name'], 'COPYxx')
     self.assertEqual(event['st'], 900) # diff clock
     self.assertEqual(event['dur'], 10)
@@ -278,23 +278,23 @@ class TestVizProfiler(unittest.TestCase):
 
     j = json.loads(get_profile(prof))
 
-    devices = list(j['layout'])
-    self.assertEqual(devices[0], 'NV Graph')
-    self.assertEqual(devices[1], 'NV')
-    self.assertEqual(devices[2], 'NV:1')
+    tracks = list(j['layout'])
+    self.assertEqual(tracks[0], 'NV Graph')
+    self.assertEqual(tracks[2], 'NV')
+    self.assertEqual(tracks[4], 'NV:1')
 
-    nv_events = j['layout']['NV']['timeline']['shapes']
+    nv_events = j['layout']['NV']['shapes']
     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']['timeline']['shapes']
+    nv1_events = j['layout']['NV:1']['shapes']
     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']['timeline']['shapes']
+    graph_events = j['layout']['NV Graph']['shapes']
     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'])
 
@@ -308,7 +308,7 @@ class TestVizMemoryLayout(BaseTestViz):
     a = _alloc(1)
     _b = _alloc(1)
     profile_ret = json.loads(get_profile(Buffer.profile_events))
-    ret = profile_ret["layout"][a.device]["mem"]
+    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])
@@ -318,7 +318,7 @@ class TestVizMemoryLayout(BaseTestViz):
     del a
     b = _alloc(1)
     profile_ret = json.loads(get_profile(Buffer.profile_events))
-    ret = profile_ret["layout"][b.device]["mem"]
+    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])
@@ -331,7 +331,7 @@ class TestVizMemoryLayout(BaseTestViz):
     del a
     c = _alloc(1)
     profile_ret = json.loads(get_profile(Buffer.profile_events))
-    ret = profile_ret["layout"][c.device]["mem"]
+    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])

test_driven_development.sh

@@ -1,8 +0,0 @@
-#!/bin/bash
-python3 test/external/process_replay/reset.py
-CAPTURE_PROCESS_REPLAY=1 pytest -n auto test/test_tiny.py test/test_uop_graph.py test/test_ops.py test/test_linearizer.py
-while true; do
-  if python3 test/test_tiny.py; then
-    PYTHONPATH="." python3 test/external/process_replay/process_replay.py
-  fi
-done

tinygrad/codegen/__init__.py

@@ -87,7 +87,7 @@ def _get_rewrites_for_renderer(opts:Renderer, linearizer:bool, _QUANTIZE, _DEVEC
 
   # decompositions
   pm_decomp = symbolic_simple+get_late_rewrite_patterns(supported_ops, _TRANSCENDENTAL>=2)
-  ret.append(RewriteStep(pm_decomp, name="decompositions"))
+  ret.append(RewriteStep(pm_decomp, lambda _: opts.device, name="decompositions"))
 
   # final rules for the renderer (without sym)
   pm_final_rewrite = pm_decomp+pm_render+extra_matcher

tinygrad/codegen/devectorizer.py

@@ -364,8 +364,7 @@ def reduce_collapse(red:UOp):
         replaces[s] = UOp(Ops.DEFINE_VAR, dtype=s.dtype, arg=(f'in{len(replaces)}', s.vmin, s.vmax))
   collapse_fxn = red.substitute(replaces)
   sink = graph_rewrite(collapse_fxn, pm_reduce_collapse, name="reduce_collapse")
-  # TODO: why is REDUCE needed here and just RANGE isn't enough?
-  if any(x.op in {Ops.REDUCE, Ops.RANGE} for x in sink.toposort()): return None
+  if any(x.op is Ops.RANGE for x in sink.toposort()): return None
   return sink.substitute({v:k for k,v in replaces.items()})
 
 def reduce_unparented(red:UOp):

tinygrad/codegen/opt/kernel.py

@@ -378,9 +378,9 @@ class Kernel:
       tensor_cores = self.opts.tensor_cores if tc_select == -1 else [self.opts.tensor_cores[tc_select]]
       for tc in tensor_cores:
         tensor_core_opts = [self._create_tc_opts(reduceop, tc, axis, opt_level) for reduceop in self.reduceops]
+        if tensor_core_opts[0] is None: continue
         # can only fuse reduces with the same tc options
         assert all_same(tensor_core_opts)
-        if tensor_core_opts[0] is None: continue
         self.tensor_core_opts = tc_opts = tensor_core_opts[0]
 
         # attempt to pad the tensor axes that require it

tinygrad/frontend/onnx.py

@@ -5,7 +5,7 @@ from io import BufferedReader
 from tinygrad.nn.state import TensorIO
 from tinygrad.tensor import Tensor, _broadcast_shape, ReductionStr
 from tinygrad.helpers import getenv, DEBUG, all_same, prod, flatten, make_tuple, argsort, is_numpy_ndarray, get_single_element, polyN
-from tinygrad.dtype import DType, ConstType, dtypes, _from_np_dtype
+from tinygrad.dtype import DType, ConstType, dtypes, _from_np_dtype, truncate
 from tinygrad.device import is_dtype_supported, Device
 
 # ***** protobuf definitions ******
@@ -105,9 +105,7 @@ class PBBufferedReader(BufferedReader):
   def read_bytes(self) -> Tensor: return self.read_delimited(use_tensor=True)
   def read_float(self) -> float: return struct.unpack("<f", self.read(4))[0]
   def read_packed_floats(self) -> Tensor: return self.read_delimited(use_tensor=True)
-  def read_int64(self) -> int:
-    val = self.decode_varint()
-    return val - 2**64 if val & (1 << 63) else val
+  def read_int64(self) -> int: return truncate[dtypes.int64](self.decode_varint())
   def read_packed_int64s(self) -> list[int]:
     total_bytes_len = self.decode_varint()
     old_pos = self.tell()
@@ -1074,7 +1072,7 @@ def get_onnx_ops() -> dict[str, types.FunctionType|dict[OpSetId, types.FunctionT
     if X.ndim == 4: X = X.permute(0, 2, 1, 3)
     elif X.ndim == 3:
       assert num_heads is not None, "num_heads must be provided for 3D input"
-      X = X.reshape(*X.shape[:-1], num_heads, X.shape[-1] // num_heads)
+      X = X.unflatten(-1, (num_heads, X.shape[-1] // num_heads))
 
     head_size = cast(int, X.shape[-1])
     rot_dim = rotary_embedding_dim or head_size
@@ -1085,16 +1083,10 @@ def get_onnx_ops() -> dict[str, types.FunctionType|dict[OpSetId, types.FunctionT
     cos = cos[..., :rot_dim//2].unsqueeze(2)
     sin = sin[..., :rot_dim//2].unsqueeze(2)
 
-    if interleaved:
-      x1, x2 = x_rotate[..., ::2], x_rotate[..., 1::2]
-      real = x1 * cos - x2 * sin
-      imag = x1 * sin + x2 * cos
-      x_rotated = Tensor.stack(real, imag, dim=-1).flatten(start_dim=-2)
-    else:
-      x1, x2 = x_rotate.chunk(2, dim=-1)
-      real = x1 * cos - x2 * sin
-      imag = x1 * sin + x2 * cos
-      x_rotated = real.cat(imag, dim=-1)
+    x1, x2 = (x_rotate[..., ::2], x_rotate[..., 1::2]) if interleaved else x_rotate.chunk(2, dim=-1)
+    real = x1 * cos - x2 * sin
+    imag = x1 * sin + x2 * cos
+    x_rotated = real.stack(imag, dim=-1).flatten(start_dim=-2) if interleaved else real.cat(imag, dim=-1)
 
     output = x_rotated.cat(x_pass, dim=-1)
     return output.flatten(start_dim=2) if len(original_input_shape) == 3 else output.permute(0, 2, 1, 3)

tinygrad/helpers.py

@@ -137,7 +137,7 @@ PICKLE_BUFFERS, PROFILE, LRU = ContextVar("PICKLE_BUFFERS", 1), ContextVar("PROF
 CACHELEVEL, IGNORE_BEAM_CACHE, DEVECTORIZE = ContextVar("CACHELEVEL", 2), ContextVar("IGNORE_BEAM_CACHE", 0), ContextVar("DEVECTORIZE", 1)
 DISABLE_COMPILER_CACHE = ContextVar("DISABLE_COMPILER_CACHE", 0)
 DONT_REALIZE_EXPAND, DONT_GROUP_REDUCES = ContextVar("DONT_REALIZE_EXPAND", 0), ContextVar("DONT_GROUP_REDUCES", 0)
-QUANTIZE, VALIDATE_WITH_CPU = ContextVar("QUANTIZE", 0), ContextVar("VALIDATE_WITH_CPU", 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)
 
@@ -293,7 +293,7 @@ def fetch(url:str, name:pathlib.Path|str|None=None, subdir:str|None=None, gunzip
   else: fp = _ensure_downloads_dir() / (subdir or "") / ((name or hashlib.md5(url.encode('utf-8')).hexdigest()) + (".gunzip" if gunzip else ""))
   if not fp.is_file() or not allow_caching:
     (_dir := fp.parent).mkdir(parents=True, exist_ok=True)
-    with urllib.request.urlopen(urllib.request.Request(url, headers={"User-Agent": "tinygrad 0.10.3"}), timeout=10) as r:
+    with urllib.request.urlopen(urllib.request.Request(url, headers={"User-Agent": "tinygrad 0.11.0"}), timeout=10) as r:
       assert r.status == 200, r.status
       length = int(r.headers.get('content-length', 0)) if not gunzip else None
       readfile = gzip.GzipFile(fileobj=r) if gunzip else r

tinygrad/renderer/__init__.py

@@ -1,7 +1,7 @@
 from __future__ import annotations
 from typing import Callable, cast, TYPE_CHECKING
-import functools, itertools
-from dataclasses import dataclass, field, replace
+import functools
+from dataclasses import dataclass, field
 from tinygrad.helpers import to_function_name, dedup, prod
 from tinygrad.uop.ops import Ops, UOp, sym_infer, sint, Variable, ssimplify, GroupOp, PatternMatcher
 from tinygrad.dtype import AddrSpace, PtrDType
@@ -23,6 +23,7 @@ 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()
@@ -34,6 +35,11 @@ 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())
@@ -47,7 +53,7 @@ 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, lds) # TODO: properly track memory, lds is always a high estimate
+    return Estimates(flops, lds, sum(mem.values()))
 
 @dataclass
 class ProgramSpec:
@@ -84,17 +90,9 @@ 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 replace(Estimates() if self.uops is None else Estimates.from_uops(self.uops, ignore_indexing=True), mem=self.mem_estimate)
+    return Estimates() if self.uops is None else Estimates.from_uops(self.uops, ignore_indexing=True)
 
   @functools.cached_property
   def function_name(self) -> str: return to_function_name(self.name)

tinygrad/renderer/cstyle.py

@@ -63,8 +63,6 @@ extra_pm = PatternMatcher([
   # insert a PRECAST before BITCAST to force it to be rendered. not needed on all backends?
   (UPat(Ops.BITCAST, name="x"), lambda x: UOp(Ops.BITCAST, x.dtype, (UOp(Ops.PRECAST, x.src[0].dtype, x.src),))
    if x.src[0].op not in {Ops.PRECAST, Ops.LOAD, Ops.CUSTOM} else None),
-  # rewrite MAX to CMPLT + WHERE (max function is annoying on many cstyle backends)
-  (UPat(Ops.MAX, name="m"), lambda m: (m.src[0] < m.src[1]).where(m.src[1], m.src[0])),
   # devectorize any bools
   (UPat((*GroupOp.ALU, Ops.CAST, Ops.BITCAST, Ops.INDEX), dtype=dtypes.bool, name="alu"), no_vectorized_alu),
   # CAST (from bool) can't be vectorized

tinygrad/renderer/llvmir.py

@@ -206,12 +206,13 @@ class AMDLLVMRenderer(LLVMRenderer):
   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(Ops.BARRIER), lambda ctx: barrier),
-    (UPat(Ops.CAST, name="x", dtype=dtypes.half.vec(16), src=UPat.var("y", dtypes.half.vec(8))), lambda ctx, x, y: f"  {ctx[x]} = shufflevector "\
-      f"<8 x half> {ctx[y]}, <8 x half> zeroinitializer, <16 x i32> <{', '.join([f'i32 {i}, i32 {j}' for i, j in zip(range(0, 8), range(8, 16))])}>"),
-    (UPat(Ops.CAST, name="x", dtype=dtypes.half.vec(8), src=UPat.var("y", dtypes.half.vec(16))), lambda ctx, x, y:
-      f"  {ctx[x]}= shufflevector <16 x half> {ctx[y]}, <16 x half> undef, <8 x i32> <{', '.join([f'i32 {x}' for x in range(0, 16, 2)])}>"),
   ]) + base_rewrite
-  extra_matcher = LLVMRenderer.extra_matcher
+  extra_matcher = LLVMRenderer.extra_matcher + PatternMatcher([
+    (UPat(Ops.CAST, name="x", dtype=dtypes.half.vec(16), src=UPat.var("y", dtypes.half.vec(8))),
+      lambda x, y: UOp(Ops.VECTORIZE, dtypes.half.vec(16), tuple(y.gep(i // 2) if i % 2 == 0 else UOp.const(dtypes.half, 0.0) for i in range(16)))),
+    (UPat(Ops.CAST, name="x", dtype=dtypes.half.vec(8), src=UPat.var("y", dtypes.half.vec(16))),
+      lambda x, y: UOp(Ops.VECTORIZE, dtypes.half.vec(8), tuple(y.gep(i * 2) for i in range(8)))),
+  ])
   def _render_footer(self, uops: list[UOp]) -> str:
     # TODO: this is copied from cstyle
     requiredMaxThreadsPerBlock = prod(u.arg[1] for u in uops if u.op is Ops.SPECIAL and u.arg[0][0] == "l")

tinygrad/runtime/ops_amd.py

@@ -15,7 +15,7 @@ from tinygrad.runtime.autogen.am import am
 from tinygrad.runtime.support.compiler_amd import HIPCompiler, AMDLLVMCompiler
 from tinygrad.runtime.support.elf import elf_loader
 from tinygrad.runtime.support.am.amdev import AMDev, AMMemoryManager
-from tinygrad.runtime.support.amd import AMDReg, AMDIP, import_module, setup_pci_bars
+from tinygrad.runtime.support.amd import AMDReg, AMDIP, import_module, import_soc, setup_pci_bars
 from tinygrad.runtime.support.system import System, PCIIfaceBase, PCIAllocationMeta, MAP_FIXED, MAP_NORESERVE
 from tinygrad.runtime.support.usb import ASM24Controller, USBMMIOInterface
 if getenv("IOCTL"): import extra.hip_gpu_driver.hip_ioctl  # noqa: F401 # pylint: disable=unused-import
@@ -751,7 +751,7 @@ class AMDDevice(HCQCompiled):
     debug_memory_size = round_up((self.max_cu_id + 1 if self.target >= (10,1,0) else 1) * (self.max_wave_id + 1) * 32, 64)
     if self.target[0] == 10: ctl_stack_size = min(ctl_stack_size, 0x7000)
 
-    self.soc = importlib.import_module(f"tinygrad.runtime.autogen.am.{({9: 'vega10', 10: 'navi10', 11: 'soc21', 12: 'soc24'}[self.target[0]])}")
+    self.soc = import_soc(self.target)
     self.pm4 = importlib.import_module(f"tinygrad.runtime.autogen.am.pm4_{'nv' if self.target[0] >= 10 else 'soc15'}")
     self.sdma = import_module('sdma', min(self.iface.ip_versions[am.SDMA0_HWIP], (6, 0, 0)))
     self.gc = AMDIP('gc', self.iface.ip_versions[am.GC_HWIP], self.iface.ip_offsets[am.GC_HWIP])

tinygrad/runtime/ops_null.py

@@ -1,12 +1,13 @@
-from tinygrad.device import Compiled, Compiler, Renderer, Allocator
-from tinygrad.uop.ops import Ops
+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
 
-class NullRenderer(Renderer):
+class NullRenderer(CStyleLanguage):
   device = "NULL"
-  code_for_op = {k:lambda:None for k in [Ops.EXP2, Ops.LOG2, Ops.SIN, Ops.SQRT]}
   has_local = False
-  def render(self, uops:list) -> str: return ""
+  float4 = "float4"
+  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

tinygrad/runtime/support/am/amdev.py

@@ -29,7 +29,7 @@ class AMFirmware:
     # Load SOS firmware
     self.sos_fw = {}
 
-    blob, sos_hdr = self.load_fw(f"psp_{fmt_ver(am.MP0_HWIP)}_sos.bin", am.struct_psp_firmware_header_v2_0)
+    blob, sos_hdr = self.load_fw(f"psp_{fmt_ver(am.MP0_HWIP)}_sos.bin", versioned_header='struct_psp_firmware_header')
     fw_bin = sos_hdr.psp_fw_bin
 
     for fw_i in range(sos_hdr.psp_fw_bin_count):
@@ -45,11 +45,11 @@ class AMFirmware:
     self.smu_psp_desc = self.desc(blob, hdr.header.ucode_array_offset_bytes, hdr.header.ucode_size_bytes, am.GFX_FW_TYPE_SMU)
 
     # SDMA firmware
-    blob, hdr, hdr_v3 = self.load_fw(f"sdma_{fmt_ver(am.SDMA0_HWIP)}.bin", am.struct_sdma_firmware_header_v2_0, am.struct_sdma_firmware_header_v3_0)
+    blob, hdr = self.load_fw(f"sdma_{fmt_ver(am.SDMA0_HWIP)}.bin", versioned_header='struct_sdma_firmware_header')
     if hdr.header.header_version_major < 3:
       self.descs += [self.desc(blob, hdr.ctl_ucode_offset, hdr.ctl_ucode_size_bytes, am.GFX_FW_TYPE_SDMA_UCODE_TH1)]
       self.descs += [self.desc(blob, hdr.header.ucode_array_offset_bytes, hdr.ctx_ucode_size_bytes, am.GFX_FW_TYPE_SDMA_UCODE_TH0)]
-    else: self.descs += [self.desc(blob, hdr_v3.header.ucode_array_offset_bytes, hdr_v3.ucode_size_bytes, am.GFX_FW_TYPE_SDMA_UCODE_TH0)]
+    else: self.descs += [self.desc(blob, hdr.header.ucode_array_offset_bytes, hdr.ucode_size_bytes, am.GFX_FW_TYPE_SDMA_UCODE_TH0)]
 
     # PFP, ME, MEC firmware
     for (fw_name, fw_cnt) in ([('PFP', 1), ('ME', 1)] if self.adev.ip_ver[am.GC_HWIP] >= (12,0,0) else []) + [('MEC', 1)]:
@@ -83,10 +83,13 @@ class AMFirmware:
 
     self.descs += [self.desc(blob, hdr0.header.ucode_array_offset_bytes, hdr0.header.ucode_size_bytes, am.GFX_FW_TYPE_RLC_G)]
 
-  def load_fw(self, fname:str, *headers):
+  def load_fw(self, fname:str, *headers, versioned_header:str|None=None):
     fpath = fetch(f"https://gitlab.com/kernel-firmware/linux-firmware/-/raw/45f59212aebd226c7630aff4b58598967c0c8c91/amdgpu/{fname}", subdir="fw")
     blob = memoryview(bytearray(fpath.read_bytes()))
     if AM_DEBUG >= 1: print(f"am {self.adev.devfmt}: loading firmware {fname}: {hashlib.sha256(blob).hexdigest()}")
+    if versioned_header:
+      chdr = am.struct_common_firmware_header.from_address(mv_address(blob))
+      headers += (getattr(am, versioned_header + f"_v{chdr.header_version_major}_{chdr.header_version_minor}"),)
     return tuple([blob] + [hdr.from_address(mv_address(blob)) for hdr in headers])
 
   def desc(self, blob:memoryview, offset:int, size:int, *types:int) -> tuple[list[int], memoryview]: return (list(types), blob[offset:offset+size])
@@ -223,7 +226,7 @@ class AMDev(PCIDevImplBase):
 
     self.bhdr = am.struct_binary_header.from_buffer(bytearray(self.vram.view(tmr_offset, tmr_size)[:]))
     ihdr = am.struct_ip_discovery_header.from_address(ctypes.addressof(self.bhdr) + self.bhdr.table_list[am.IP_DISCOVERY].offset)
-    assert ihdr.signature == am.DISCOVERY_TABLE_SIGNATURE and not ihdr.base_addr_64_bit, f"0x{ihdr.signature:X} != 0x{am.DISCOVERY_TABLE_SIGNATURE:X}"
+    assert self.bhdr.binary_signature == am.BINARY_SIGNATURE and ihdr.signature == am.DISCOVERY_TABLE_SIGNATURE, "discovery signatures mismatch"
 
     # Mapping of HW IP to Discovery HW IP
     hw_id_map = {am.__dict__[x]: int(y) for x,y in am.hw_id_map}
@@ -256,4 +259,3 @@ class AMDev(PCIDevImplBase):
     for prefix, hwip in mods:
       self.__dict__.update(import_asic_regs(prefix, self.ip_ver[hwip], cls=functools.partial(AMRegister, adev=self, bases=self.regs_offset[hwip])))
     self.__dict__.update(import_asic_regs('mp', (11, 0), cls=functools.partial(AMRegister, adev=self, bases=self.regs_offset[am.MP1_HWIP])))
-

tinygrad/runtime/support/am/ip.py

@@ -1,7 +1,8 @@
-import ctypes, time, contextlib, importlib, functools
+import ctypes, time, contextlib, functools
 from typing import Literal
-from tinygrad.runtime.autogen.am import am
 from tinygrad.helpers import to_mv, data64, lo32, hi32, DEBUG, wait_cond
+from tinygrad.runtime.autogen.am import am
+from tinygrad.runtime.support.amd import import_soc
 
 class AM_IP:
   def __init__(self, adev): self.adev = adev
@@ -11,9 +12,7 @@ class AM_IP:
   def set_clockgating_state(self): pass # Set clockgating state for this IP
 
 class AM_SOC(AM_IP):
-  def init_sw(self):
-    self.soc_ver = 24 if self.adev.ip_ver[am.GC_HWIP] >= (12,0,0) else 21
-    self.module = importlib.import_module(f"tinygrad.runtime.autogen.am.soc{self.soc_ver}")
+  def init_sw(self): self.module = import_soc(self.adev.ip_ver[am.GC_HWIP])
 
   def init_hw(self):
     self.adev.regRCC_DEV0_EPF2_STRAP2.update(strap_no_soft_reset_dev0_f2=0x0)

tinygrad/runtime/support/amd.py

@@ -43,12 +43,22 @@ def fixup_ip_version(ip:str, version:tuple[int, ...]) -> list[tuple[int, ...]]:
 
   return [version, version[:2], version[:2]+(0,), version[:1]+(0, 0)]
 
+def header_download(file, name=None, subdir="defines") -> str:
+  url = "https://gitlab.com/linux-kernel/linux-next/-/raw/cf6d949a409e09539477d32dbe7c954e4852e744/drivers/gpu/drm/amd"
+  return fetch(f"{url}/{file}", name=name, subdir=subdir).read_text()
+
+def import_header(path:str):
+  t = re.sub(r'//.*|/\*.*?\*/','', header_download(path, subdir="defines"), flags=re.S)
+  return {k:int(v,0) for k,v in re.findall(r'\b([A-Za-z_]\w*)\s*=\s*(0x[0-9A-Fa-f]+|\d+)', t)}
+
 def import_module(name:str, version:tuple[int, ...], version_prefix:str=""):
   for ver in fixup_ip_version(name, version):
     try: return importlib.import_module(f"tinygrad.runtime.autogen.am.{name}_{version_prefix}{'_'.join(map(str, ver))}")
     except ImportError: pass
   raise ImportError(f"Failed to load autogen module for {name.upper()} {'.'.join(map(str, version))}")
 
+def import_soc(ip): return type("SOC", (object,), import_header(f"include/{({9: 'vega10', 10: 'navi10', 11: 'soc21', 12: 'soc24'}[ip[0]])}_enum.h"))
+
 def import_asic_regs(prefix:str, version:tuple[int, ...], cls=AMDReg) -> dict[str, AMDReg]:
   def _split_name(name): return name[:(pos:=next((i for i,c in enumerate(name) if c.isupper()), len(name)))], name[pos:]
   def _extract_regs(txt):
@@ -56,8 +66,7 @@ def import_asic_regs(prefix:str, version:tuple[int, ...], cls=AMDReg) -> dict[st
   def _download_file(ver, suff) -> str:
     dir_prefix = {"osssys": "oss"}.get(prefix, prefix)
     fetch_name, file_name = f"{prefix}_{'_'.join(map(str, ver))}_{suff}.h", f"{prefix}_{'_'.join(map(str, version))}_{suff}.h"
-    url = "https://gitlab.com/linux-kernel/linux-next/-/raw/cf6d949a409e09539477d32dbe7c954e4852e744/drivers/gpu/drm/amd/include/asic_reg"
-    return fetch(f"{url}/{dir_prefix}/{fetch_name}", name=file_name, subdir="asic_regs").read_text()
+    return header_download(f"include/asic_reg/{dir_prefix}/{fetch_name}", name=file_name, subdir="asic_regs")
 
   for ver in fixup_ip_version(prefix, version):
     try: offs, sh_masks = _extract_regs(_download_file(ver, "offset")), _extract_regs(_download_file(ver, "sh_mask"))

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

tinygrad/schedule/multi.py

@@ -13,6 +13,10 @@ def handle_allreduce_multirank(buf:UOp, red:UOp) -> UOp|None:
   for i,dev in enumerate(buf.device):
     groups.setdefault(Device[dev].group_id, []).append(buf.mselect(i))
 
+  # Put reduce leader of each group first
+  reduce_leaders = set(getenv("REDUCE_LEADERS", "").split(","))
+  groups = {gid: sorted(bufs, key=lambda x: (x.device not in reduce_leaders, x.device)) for gid,bufs in groups.items()}
+
   # Skip if only one group or if every group has only one buffer
   if len(groups) <= 1 or not any(len(g) > 1 for g in groups.values()): return None
 

tinygrad/tensor.py

@@ -252,7 +252,7 @@ class Tensor(MathTrait):
     # create the schedule
     schedule, var_vals = create_schedule_with_vars(sink)
     schedule = memory_planner(schedule)
-    if DEBUG >= 1 and len(schedule) >= 10: print(f"scheduled {len(schedule)} kernels in {(time.perf_counter()-st)*1000:.2f} ms")
+    if DEBUG >= 1 and len(schedule) > 1: print(f"scheduled {len(schedule)} kernels in {(time.perf_counter()-st)*1000:.2f} ms")
     return schedule, var_vals
 
   def schedule(self, *lst:Tensor) -> list[ScheduleItem]:
@@ -345,6 +345,7 @@ class Tensor(MathTrait):
     print(t.tolist())
     ```
     """
+    if self.dtype in (dtypes.bfloat16, *dtypes.fp8s): return self.cast(dtypes.float32).tolist()
     return self.data().tolist()
 
   def numpy(self) -> 'np.ndarray':  # type: ignore [name-defined] # noqa: F821
@@ -1127,11 +1128,12 @@ class Tensor(MathTrait):
     if (isinstance(indices, list) and all_int(indices)) or not isinstance(indices, (tuple, list)): indices = [indices]
     x, indices = self, list(indices)
 
-    # filter ellipsis and fill with slice(None) or fill rest of indices with slice(None)
+    # fill ellipsis or rest of indices with slice(None)
     if len(ellipsis_idx := [dim for dim, i in enumerate(indices) if i is Ellipsis]) > 1: raise IndexError("indices can only have a single ellipsis")
-    fill_idx = ellipsis_idx[0] if ellipsis_idx else len(indices)
+    # NOTE: None adds a dim later
     num_indices = len(indices) - len(ellipsis_idx) - sum(1 for i in indices if i is None)
     if num_indices > self.ndim: raise IndexError(f"too many {num_indices=} for {self.ndim=}")
+    fill_idx = ellipsis_idx[0] if ellipsis_idx else len(indices)
     indices[fill_idx:fill_idx+1] = [slice(None)] * (self.ndim - num_indices)
 
     indices_parsed, dim = [], 0
@@ -1147,6 +1149,7 @@ class Tensor(MathTrait):
           index = Tensor([i+size if i<0 else i for i in fully_flatten(index)], self.device, requires_grad=False).reshape(ti.shape)
         case int() | UOp(): # sint
           if index >= size or index < -size: raise IndexError(f"{index=} is out of bounds with {size=}")
+          # TODO: is this right for (negative) symbolic?
           boundary = [index, index+1] if index >= 0 else [index+size, index+size+1]
         case slice():
           if index.step == 0: raise ValueError(f"{index=} cannot have 0 as step")
@@ -1160,9 +1163,9 @@ class Tensor(MathTrait):
             elif stride < 0: boundary = [boundary[1] + 1, boundary[0] + 1]
             # update size for slice
             size = ceildiv((boundary[1] - boundary[0]), abs(stride))
-          elif (step == 1) and isinstance(step, int) and all(isinstance(s,(int,UOp)) for s in (start, stop)) and resolve((stop-start) > 0, False):
+          elif resolve(step == 1, False) and all(isinstance(s,sint) for s in (start, stop)) and resolve((stop-start) > 0, False):
             # simple symbolic slice
-            size = cast(UOp|int, cast(UOp, (stop - start)).ssimplify())
+            size = cast(sint, cast(UOp, (stop - start)).ssimplify())
           else: raise TypeError(f"slice {index=} is not supported")
         case None: pass # do nothing
         case _: raise IndexError(f"{type(index).__name__} indexing is not supported")
@@ -1174,9 +1177,9 @@ class Tensor(MathTrait):
       # flip negative strides
       shrinks, strides = zip(*((i['boundary'], i['stride']) for i in mops))
       x = x.shrink(shrinks).flip(tuple(i for i,st in enumerate(strides) if st < 0))
-      # handle stride != 1 or -1
-      if any(abs(st) != 1 for st in strides):
-        strides = tuple(abs(s) for s in strides)
+      strides = tuple(map(abs, strides))
+      # apply stride
+      if any(st != 1 for st in strides):
         # pad shape to multiple of stride
         if not all_int(x.shape): raise RuntimeError("symbolic shape not supported")
         x = x.pad(tuple((0, round_up(s, st) - s) for s, st in zip(x.shape, strides)))
@@ -2680,7 +2683,7 @@ class Tensor(MathTrait):
     print(t.triu(diagonal=-1).numpy())
     ```
     """
-    return Tensor._tri(self.shape[-2], self.shape[-1], diagonal=diagonal, device=self.device, dtype=dtypes.bool).where(self, 0).cast(self.dtype)
+    return Tensor._tri(self.shape[-2], self.shape[-1], diagonal=diagonal, device=self.device, dtype=dtypes.bool).where(self, self.zeros_like())
 
   def tril(self, diagonal:int=0) -> Tensor:
     """
@@ -2703,7 +2706,7 @@ class Tensor(MathTrait):
     print(t.tril(diagonal=-1).numpy())
     ```
     """
-    return Tensor._tri(self.shape[-2], self.shape[-1], diagonal=diagonal+1, device=self.device, dtype=dtypes.bool).where(0, self).cast(self.dtype)
+    return Tensor._tri(self.shape[-2], self.shape[-1], diagonal=diagonal+1, device=self.device, dtype=dtypes.bool).where(self.zeros_like(), self)
 
   def interpolate(self, size:tuple[int, ...], mode:str="linear", align_corners:bool=False) -> Tensor:
     """

tinygrad/uop/__init__.py

@@ -12,11 +12,14 @@ class Ops(FastEnum):
   NOOP = auto(); SINK = auto(); UNIQUE = auto(); DEVICE = auto(); KERNEL = auto(); PRECAST = auto(); REWRITE_ERROR = auto()  # noqa: E702
 
   # track children
-  CHILD = auto()
+  CHILD = auto(); CHILDREN = auto() # noqa: E702
 
   # buffer ops
   COPY = auto(); BUFFER = auto(); BUFFER_VIEW = auto(); MSELECT = auto(); MSTACK = auto() # noqa: E702
 
+  # create buffer
+  BUFFERIZE = auto()
+
   # ops that adjust the behavior of the scheduler
   CONTIGUOUS = auto(); CONTIGUOUS_BACKWARD = auto(); DETACH = auto(); FUSE = auto() # noqa: E702
 

tinygrad/uop/decompositions.py

@@ -2,7 +2,7 @@ from typing import Callable
 import math, functools
 from tinygrad.dtype import dtypes, DType, promo_lattice
 from tinygrad.device import is_dtype_supported
-from tinygrad.helpers import polyN, getenv
+from tinygrad.helpers import polyN, DISABLE_FAST_IDIV
 from tinygrad.uop.ops import UOp, UPat, Ops, PatternMatcher
 
 TRANSCENDENTAL_SUPPORTED_DTYPES = (dtypes.float16, dtypes.float32, dtypes.float64)
@@ -317,8 +317,10 @@ powers_of_two = {2**i:i for i in range(64)}
 def get_late_rewrite_patterns(ops:tuple[Ops, ...], force_transcendental=False):
   pat: list[tuple[UPat, Callable]] = [(UPat(op, dtype=TRANSCENDENTAL_SUPPORTED_DTYPES, src=(UPat.var("d"),)), f) for op,f in \
            ((Ops.EXP2, xexp2), (Ops.LOG2, xlog2), (Ops.SIN, xsin)) if op not in ops or force_transcendental]
-  # no real hardware supports THREEFRY
-  pat.append((UPat(Ops.THREEFRY, dtype=dtypes.uint64, src=(UPat.var("x"), UPat.var("key"))), threefry2x32))
+  # no real hardware supports THREEFRY, but NullRenderer does
+  if Ops.THREEFRY not in ops: pat.append((UPat(Ops.THREEFRY, dtype=dtypes.uint64, src=(UPat.var("x"), UPat.var("key"))), threefry2x32))
+  # MAX can be rewritten as CMPLT + WHERE (max function is annoying on many cstyle backends)
+  if Ops.MAX not in ops and Ops.CMPLT in ops: pat.append((UPat(Ops.MAX, name="m"), lambda m: (m.src[0] < m.src[1]).where(m.src[1], m.src[0])))
   # rewrite SQRT to xpow 0.5
   if Ops.SQRT not in ops: pat.append((UPat(Ops.SQRT, src=UPat.var("d")), lambda d: xpow(d, d.const_like(0.5))))
   # rewrite MOD to AND (which should always be supported, but not for generic in tests): x % (2**y) -> x & (2**y-1)
@@ -330,7 +332,7 @@ def get_late_rewrite_patterns(ops:tuple[Ops, ...], force_transcendental=False):
     pat += [(UPat.var("x", dtypes.uints)//UPat.cvar("c"), lambda x,c: x >> v if (v:=powers_of_two.get(c.arg, 0)) else None)]
     pat += [(UPat.var("x", dtypes.ints)//UPat.cvar("c"), lambda x,c: (x+(l.const_like(l.vmin) if (l:=(x<0)).vmin==l.vmax else l).where(
       c-1, 0)) >> v if (v:=powers_of_two.get(c.arg, 0)) else None)]  # (x+(x<0).where(c-1, 0)) >> v
-    if not getenv("DISABLE_FAST_IDIV"):
+    if not DISABLE_FAST_IDIV:
       pat += [(UPat.var("x", dtypes.ints)//UPat.cvar("d"), lambda ctx, x, d: fast_idiv(ctx, x, d.arg))]
       pat += [(UPat.var("x", dtypes.ints)%UPat.var("d"), lambda x, d: x-d*(x//d))]
   if Ops.NEG in ops:

tinygrad/uop/ops.py

@@ -140,6 +140,8 @@ class UOp(MathTrait, metaclass=UOpMetaClass):
     from tinygrad.shape.shapetracker import ShapeTracker
     # VIEW and MovementOps define a new ShapeTracker from the arg
     if self.op is Ops.VIEW: return self.arg
+    # allow reshape from nothing
+    if self.op is Ops.RESHAPE and self.src[0].st is None: return ShapeTracker.from_shape(self.arg)
     if self.op in GroupOp.Movement: return unwrap(self.src[0].st).mop(self.op, self.arg)
     # CONST with a DEVICE has a shape of ()
     if self.op is Ops.CONST and len(self.src) and self.src[0].op is Ops.DEVICE: return ShapeTracker.from_shape(())
@@ -186,10 +188,15 @@ class UOp(MathTrait, metaclass=UOpMetaClass):
   @functools.cached_property
   def ranges(self) -> dict[UOp, None]:
     if self.op is Ops.RANGE: return {self:None}
-    if self.op in {Ops.CONTIGUOUS, Ops.REDUCE, Ops.STORE}:
+    if self.op in {Ops.BUFFERIZE, Ops.REDUCE}:
       ret = self.src[0].ranges.copy()
       for s in self.src[1:]:
         if s in ret: del ret[s]
+    elif self.op in {Ops.STORE}:
+      ret = self.src[0].ranges.copy()
+      ret.update(self.src[1].ranges)
+      for s in self.src[2:]:
+        if s in ret: del ret[s]
     else:
       ret = {}
       for s in self.src: ret.update(s.ranges)
@@ -293,6 +300,7 @@ class UOp(MathTrait, metaclass=UOpMetaClass):
   def reduce(self, *src:UOp, **kwargs): return UOp(Ops.REDUCE, kwargs.pop('dtype', self.dtype), src=(self,)+src, **kwargs)
   def contiguous(self, *args, **kwargs): return UOp(Ops.CONTIGUOUS, dtype=self.dtype, src=(self,)+args, **kwargs)
   def contiguous_backward(self): return self.alu(Ops.CONTIGUOUS_BACKWARD)
+  def bufferize(self, *args, **kwargs): return UOp(Ops.BUFFERIZE, dtype=self.dtype, src=(self,)+args, **kwargs)
   def fuse(self): return self.alu(Ops.FUSE)
   def allreduce(self, op, device:str|tuple[str, ...]|UOp):
     assert isinstance(self.device, tuple), f"allreduce must be on tuple {self.device} isn't"
@@ -364,6 +372,7 @@ class UOp(MathTrait, metaclass=UOpMetaClass):
     if self.st == ret.st: return self  # ignore NOOPs, also check ret.st
     return ret
 
+  def forced_reshape(self, arg:tuple[sint, ...]): return UOp(Ops.RESHAPE, self.dtype, src=(self,), arg=arg)
   def reshape(self, arg:tuple[sint, ...]): return self._mop(Ops.RESHAPE, arg)
   def pad(self, arg:tuple[tuple[sint, sint], ...]): return self._mop(Ops.PAD, arg)
   def expand(self, arg:tuple[sint, ...]): return self._mop(Ops.EXPAND, arg)
@@ -658,6 +667,9 @@ class UPat(MathTrait):
   @staticmethod
   def const(dtype:DType|tuple[DType, ...]|None, b:ConstType): return UPat(Ops.CONST, dtype=dtype, arg=b)
 
+  # lil helper
+  def f(self, op, **kwargs): return UPat(op, src=(self,), **kwargs)
+
   # copied from UOp
   def sink(self, *srcs:UPat|None, **kwargs): return UPat(Ops.SINK, dtypes.void, (self,)+tuple([x for x in srcs if x is not None]), **kwargs)
   def index(self, idx:UPat, valid:UPat|None=None): return UPat(Ops.INDEX, self.dtype, (self,idx,valid) if valid is not None else (self,idx))
@@ -671,6 +683,7 @@ class UPat(MathTrait):
   def reduce(self, *src:UPat, **kwargs): return UPat(Ops.REDUCE, self.dtype, src=(self,)+src, **kwargs)
   def fuse(self): return self.alu(Ops.FUSE)
   def or_broadcasted(self, **kwargs): return UPat.any(self, UPat(Ops.VECTORIZE, self.dtype, src=self, **kwargs))
+  def contiguous(self, *args, **kwargs): return UPat(Ops.CONTIGUOUS, dtype=self.dtype, src=(self,)+args, **kwargs)
 
   def const_like(self, b:ConstLike): return UPat.const(self.dtype, cast(ConstType, b))
   def alu(self, op:Ops, *src:UPat):

tinygrad/uop/symbolic.py

@@ -41,6 +41,7 @@ symbolic_simple = PatternMatcher([
   (UPat(GroupOp.Idempotent, src=(UPat.var("x"), UPat.var("x"))), lambda x: x),
   (UPat.var("x", dtype=dtypes.bool).logical_not().logical_not(), lambda x: x),
   (UPat.var("x", dtype=dtypes.bool).where(UPat.const(dtypes.bool, True), UPat.const(dtypes.bool, False)), lambda x: x),
+  (UPat.var("x", dtype=dtypes.bool).where(UPat.const(dtypes.bool, False), UPat.const(dtypes.bool, True)), lambda x: x.logical_not()),
   (UPat.var("x", dtype=dtypes.ints+(dtypes.bool,)).trunc(), lambda x: x),
   # ** zero folding **
   (UPat.var("x") < UPat.var("x"), lambda x: x.const_like(False).cast(dtypes.bool.vec(x.dtype.count))), # x < x -> False
@@ -242,18 +243,18 @@ def gep_through_wmma(gep:UOp, wmma:UOp):
 
 gep_pushing = PatternMatcher([
   # GEP/VECTORIZE, GEP/GEP, GEP/CONST, GEP/VCONST
-  (UPat(Ops.GEP, src=(UPat(Ops.GEP, name='g2'),), name='g1'),
+  (UPat(Ops.GEP, name='g2').f(Ops.GEP, name='g1'),
    lambda g1, g2: g2.src[0].gep(tuple(g2.arg[g1.arg[i]] for i in range(len(g1.arg))))),
-  (UPat(Ops.GEP, src=(UPat(Ops.VECTORIZE, name="vec"),), name="gep"),
+  (UPat(Ops.VECTORIZE, name='vec').f(Ops.GEP, name='gep'),
    lambda gep, vec: UOp(Ops.VECTORIZE, gep.dtype, tuple(vec.src[i] for i in gep.arg)) if len(gep.arg) > 1 else vec.src[gep.arg[0]]),
-  (UPat(Ops.GEP, src=(UPat.cvar("c", vec=False),), name="gep"), lambda gep, c: gep.const_like(c.arg)),
-  (UPat(Ops.GEP, src=(UPat(Ops.VCONST, name="c"),), name="gep"), lambda gep, c: gep.const_like(tuple(c.arg[x] for x in gep.arg))),
+  (UPat.cvar("c", vec=False).f(Ops.GEP, name="gep"), lambda gep, c: gep.const_like(c.arg)),
+  (UPat(Ops.VCONST, name="c").f(Ops.GEP, name="gep"), lambda gep, c: gep.const_like(tuple(c.arg[x] for x in gep.arg))),
   # GEP on void is skipped
   (UPat(Ops.GEP, src=(UPat(dtype=dtypes.void, name="x"),)), lambda x: x),
   # GEP in order is removed
   (UPat(Ops.GEP, name="g"), lambda g: g.src[0] if not isinstance(g.dtype, PtrDType) and g.arg == tuple(range(g.src[0].dtype.count)) else None),
   # push all GEPs through ALUs (fix arange stuff)
-  (UPat(Ops.GEP, src=(UPat((*GroupOp.ALU, Ops.CAST, Ops.BITCAST), name='alu'),), name='gep'),
+  (UPat((*GroupOp.ALU, Ops.CAST, Ops.BITCAST), name='alu').f(Ops.GEP, name='gep'),
    lambda gep,alu: UOp(alu.op, alu.dtype.scalar().vec(gep.dtype.count), tuple(x.gep(gep.arg) for x in alu.src), alu.arg) \
      if not isinstance(gep.dtype, PtrDType) else None),
   # CAT can't be rendered. it's a VECTORIZE on vectors, we expand to a single VECTORIZEs with GEPs (TODO: move this later)
@@ -262,7 +263,7 @@ gep_pushing = PatternMatcher([
   # VECTORIZE on same GEP
   (UPat(Ops.VECTORIZE, name="v", src=UPat(Ops.GEP, src=(UPat.var("x"),))), lambda v,x: x.gep(tuple(get_single_element(i.arg) for i in v.src))),
   # push some GEPs through WMMAs
-  (UPat(Ops.GEP, src=(UPat(Ops.WMMA, name="wmma"),), name="gep"), gep_through_wmma),
+  (UPat(Ops.WMMA, name="wmma").f(Ops.GEP, name="gep"), gep_through_wmma),
 ])
 
 commutative = PatternMatcher([

tinygrad/viz/index.html

@@ -228,7 +228,7 @@
   }
   #device-list > div {
     min-height: 32px;
-    max-width: 100px;
+    max-width: 132px;
     overflow-x: auto;
     overflow-y: hidden;
     white-space: nowrap;

tinygrad/viz/js/index.js

@@ -122,17 +122,18 @@ const colorScheme = {TINY:["#1b5745", "#354f52", "#354f52", "#1d2e62", "#63b0cd"
   CATEGORICAL:["#ff8080", "#F4A261", "#C8F9D4", "#8D99AE", "#F4A261", "#ffffa2", "#ffffc0", "#87CEEB"],}
 const cycleColors = (lst, i) => lst[i%lst.length];
 
-const createPolygons = (source, area) => {
-  const shapes = [];
-  const yscale = d3.scaleLinear().domain([0, source.peak]).range([area, 0]);
-  for (const [i,e] of source.shapes.entries()) {
-    const x = e.x.map((i,_) => (source.timestamps[i] ?? data.et)-data.st);
-    const y0 = e.y.map(yscale);
-    const y1 = e.y.map(y => yscale(y+e.arg.nbytes));
-    const arg = { tooltipText:`${e.arg.dtype} len:${formatUnit(e.arg.sz)}\n${formatUnit(e.arg.nbytes, "B")}` };
-    shapes.push({ x, y0, y1, arg, fillColor:cycleColors(colorScheme.BUFFER, i) });
+const rescaleTrack = (source, tid, k) => {
+  for (const e of source.shapes) {
+    for (let i=0; i<e.y0.length; i++) {
+      e.y0[i] = e.y0[i]*k;
+      e.y1[i] = e.y1[i]*k;
+    }
   }
-  return shapes;
+  const change = (source.height*k)-source.height;
+  const div = document.getElementById(tid);
+  div.style.height = rect(div).height+change+"px";
+  source.height = source.height*k;
+  return change;
 }
 
 const drawLine = (ctx, x, y) => {
@@ -150,77 +151,68 @@ async function renderProfiler() {
   // layout once!
   if (data != null) return;
   const profiler = d3.select(".profiler").html("");
-  const deviceList = profiler.append("div").attr("id", "device-list").node();
+  const { layout, st, et } = await (await fetch("/get_profile")).json();
+  // place devices on the y axis and set vertical positions
+  const [tickSize, padding] = [10, 8];
+  const deviceList = profiler.append("div").attr("id", "device-list").style("padding-top", tickSize+padding+"px");
   const canvas = profiler.append("canvas").attr("id", "timeline").node();
   // NOTE: scrolling via mouse can only zoom the graph
   canvas.addEventListener("wheel", e => (e.stopPropagation(), e.preventDefault()), { passive:false });
-  const profileRet = await (await fetch("/get_profile")).json()
-  const { layout, st, et } = profileRet;
-  // place devices on the y axis and set vertical positions
-  const [tickSize, padding] = [10, 8];
-  deviceList.style.paddingTop = `${tickSize+padding}px`;
   const ctx = canvas.getContext("2d");
   const canvasTop = rect(canvas).top;
   // color by key (name/category/device)
   const colorMap = new Map();
   data = {tracks:new Map(), axes:{}, st, et};
-  const areaScale = d3.scaleLinear().domain([0, Object.entries(layout).reduce((peak, [_,d]) => Math.max(peak, d.mem.peak), 0)]).range([4,maxArea=100]);
-  for (const [k, { timeline, mem }] of Object.entries(layout)) {
-    if (timeline.shapes.length === 0 && mem.shapes.length == 0) continue;
-    const div = deviceList.appendChild(document.createElement("div"));
-    div.innerText = k;
-    div.style.padding = `${padding}px`;
-    div.onclick = () => { // TODO: make this feature more visible
-      const prevScroll = profiler.node().scrollTop;
-      let newOffset = null;
-      for (const [track, v] of data.tracks) {
-        if (track === `${k} memory`)  {
-          // expand the y axis or reset to default size
-          const pick = [areaScale(mem.peak), maxArea*4];
-          const expand = k !== focusedDevice;
-          const [newArea, prevArea] = expand ? pick.reverse() : pick;
-          focusedDevice = expand ? k : null;
-          data.axes.y = expand ? { domain:[0, mem.peak], range:[v.offsetY+newArea, v.offsetY], fmt:"B" } : null;
-          // either way update all offsets
-          v.shapes = createPolygons(mem, newArea);
-          newOffset = newArea-prevArea;
-          v.div.style.height = rect(v.div).height+newOffset+"px";
-        } else if (newOffset != null) v.offsetY += newOffset;
-      }
-      d3.select(canvas).call(canvasZoom.transform, zoomLevel);
-      if (prevScroll) profiler.node().scrollTop = prevScroll;
-    }
-    const { y:baseY, height:baseHeight } = rect(div);
-    const levelHeight = baseHeight-padding;
+  const heightScale = d3.scaleLinear().domain([0, Object.entries(layout).reduce((peak, [_,d]) => Math.max(peak, d.peak||0), 0)]).range([4,maxheight=100]);
+  for (const [k, v] of Object.entries(layout)) {
+    if (v.shapes.length === 0) continue;
+    const div = deviceList.append("div").attr("id", k).text(k).style("padding", padding+"px");
+    const { y:baseY, height:baseHeight } = rect(div.node());
     const offsetY = baseY-canvasTop+padding/2;
-    const shapes = [];
-    data.tracks.set(k, { shapes, offsetY });
-    let colorKey, ref;
-    for (const e of timeline.shapes) {
-      if (e.depth === 0) colorKey = e.cat ?? e.name;
-      if (!colorMap.has(colorKey)) colorMap.set(colorKey, cycleColors(colorScheme[k] ?? colorScheme.DEFAULT, colorMap.size));
-      const fillColor = d3.color(colorMap.get(colorKey)).brighter(e.depth).toString();
-      const label = parseColors(e.name).map(({ color, st }) => ({ color, st, width:ctx.measureText(st).width }));
-      if (e.ref != null) ref = {ctx:e.ref, step:0};
-      else if (ref != null) {
-        const start = ref.step>0 ? ref.step+1 : 0;
-        const stepIdx = ctxs[ref.ctx+1].steps.findIndex((s, i) => i >= start && s.name == e.name);
-        ref = stepIdx === -1 ? null : {ctx:ref.ctx, step:stepIdx};
+    if (v.shapes[0].dur != null) {
+      const levelHeight = baseHeight-padding;
+      const shapes = [];
+      data.tracks.set(k, { shapes, offsetY });
+      let colorKey, ref;
+      for (const e of v.shapes) {
+        if (e.depth === 0) colorKey = e.cat ?? e.name;
+        if (!colorMap.has(colorKey)) colorMap.set(colorKey, cycleColors(colorScheme[k] ?? colorScheme.DEFAULT, colorMap.size));
+        const fillColor = d3.color(colorMap.get(colorKey)).brighter(e.depth).toString();
+        const label = parseColors(e.name).map(({ color, st }) => ({ color, st, width:ctx.measureText(st).width }));
+        if (e.ref != null) ref = {ctx:e.ref, step:0};
+        else if (ref != null) {
+          const start = ref.step>0 ? ref.step+1 : 0;
+          const stepIdx = ctxs[ref.ctx+1].steps.findIndex((s, i) => i >= start && s.name == e.name);
+          ref = stepIdx === -1 ? null : {ctx:ref.ctx, step:stepIdx};
+        }
+        const arg = { tooltipText:formatTime(e.dur)+(e.info != null ? "\n"+e.info : ""), ...ref };
+        // offset y by depth
+        shapes.push({x:e.st-st, y:levelHeight*e.depth, width:e.dur, height:levelHeight, arg, label, fillColor });
       }
-      const arg = { tooltipText:formatTime(e.dur)+(e.info != null ? "\n"+e.info : ""), ...ref };
-      // offset y by depth
-      shapes.push({x:e.st-st, y:levelHeight*e.depth, width:e.dur, height:levelHeight, arg, label, fillColor });
+      div.style("height", levelHeight*v.maxDepth+padding+"px").style("pointerEvents", "none");
+    } else {
+      const height = heightScale(v.peak);
+      const yscale = d3.scaleLinear().domain([0, v.peak]).range([height, 0]);
+      const shapes = [];
+      for (const [i,e] of v.shapes.entries()) {
+        const x = e.x.map(tsIdx => v.timestamps[tsIdx]-st);
+        const arg = {tooltipText:`${e.arg.dtype} len:${formatUnit(e.arg.sz)}\n${formatUnit(e.arg.nbytes, "B")}`};
+        shapes.push({ x, y0:e.y.map(yscale), y1:e.y.map(y => yscale(y+e.arg.nbytes)), arg, fillColor:cycleColors(colorScheme.BUFFER, i) });
+      }
+      data.tracks.set(k, { shapes, offsetY, height, peak:v.peak, scaleFactor:maxheight*4/height });
+      div.style("height", height+padding+"px").style("cursor", "pointer").on("click", (e) => {
+        const newFocus = e.currentTarget.id === focusedDevice ? null : e.currentTarget.id;
+        let offset = 0;
+        for (const [tid, track] of data.tracks) {
+          track.offsetY += offset;
+          if (tid === newFocus) offset += rescaleTrack(track, tid, track.scaleFactor);
+          else if (tid === focusedDevice) offset += rescaleTrack(track, tid, 1/track.scaleFactor);
+        }
+        data.axes.y = newFocus != null ? { domain:[0, (t=data.tracks.get(newFocus)).peak], range:[t.offsetY+t.height, t.offsetY], fmt:"B" } : null;
+        focusedDevice = newFocus;
+        return resize();
+      });
     }
-    // position shapes on the canvas and scale to fit fixed area
-    let area = mem.shapes.length === 0 ? 0 : areaScale(mem.peak);
-    if (area === 0) div.style.pointerEvents = "none";
-    else {
-      const startY = offsetY+(levelHeight*timeline.maxDepth)+padding/2;
-      data.tracks.set(`${k} memory`, { shapes:createPolygons(mem, area), offsetY:startY, div });
-      div.style.cursor = "pointer";
-    }
-    // lastly, adjust device rect by number of levels
-    div.style.height = `${Math.max(levelHeight*timeline.maxDepth, baseHeight)+area+padding}px`;
   }
   updateProgress({ "show":false });
   // draw events on a timeline

tinygrad/viz/serve.py

@@ -19,7 +19,7 @@ uops_colors = {Ops.LOAD: "#ffc0c0", Ops.STORE: "#87CEEB", Ops.CONST: "#e0e0e0",
                **{x:"#D8F9E4" for x in GroupOp.Movement}, **{x:"#ffffc0" for x in GroupOp.ALU}, Ops.THREEFRY:"#ffff80", Ops.BUFFER_VIEW: "#E5EAFF",
                Ops.BLOCK: "#C4A484", Ops.BLOCKEND: "#C4A4A4", Ops.BUFFER: "#B0BDFF", Ops.COPY: "#a040a0", Ops.FUSE: "#FFa500",
                Ops.ALLREDUCE: "#ff40a0", Ops.MSELECT: "#d040a0", Ops.MSTACK: "#d040a0", Ops.CONTIGUOUS: "#FFC14D",
-               Ops.CHILD: "#80fff0", Ops.REWRITE_ERROR: "#ff2e2e"}
+               Ops.CHILDREN: "#80ffc0", Ops.CHILD: "#80fff0", Ops.BUFFERIZE: "#FF991C", Ops.REWRITE_ERROR: "#ff2e2e"}
 
 # VIZ API
 
@@ -73,8 +73,8 @@ def uop_to_json(x:UOp) -> dict[int, dict]:
     if u.dtype != dtypes.void: label += f"\n{u.dtype}"
     for idx,x in enumerate(u.src):
       if x in excluded:
-        if x.op is Ops.CONST and dtypes.is_float(u.dtype): label += f"\nCONST{idx} {x.arg:g}"
-        else: label += f"\n{x.op.name}{idx} {x.arg}"
+        arg = f"{x.arg:g}" if x.op is Ops.CONST and dtypes.is_float(u.dtype) else f"{x.arg}"
+        label += f"\n{x.op.name}{idx} {arg}" + (f" {x.src[0].op}" if len(x.src) else "")
     try:
       if u.op not in {Ops.VIEW, Ops.BUFFER, Ops.KERNEL, Ops.ASSIGN, Ops.COPY, Ops.SINK, *GroupOp.Buffer} and u.st is not None:
         label += f"\n{shape_to_str(u.shape)}"
@@ -144,7 +144,7 @@ def timeline_layout(events:list[tuple[int, int, float, DevEvent]]) -> dict:
     shapes.append({"name":name, "ref":ref, "st":st, "dur":dur, "depth":depth, "cat":cat, "info":info})
   return {"shapes":shapes, "maxDepth":len(levels)}
 
-def mem_layout(events:list[tuple[int, int, float, DevEvent]]) -> dict:
+def mem_layout(events:list[tuple[int, int, float, DevEvent]], max_ts:int) -> dict:
   step, peak, mem = 0, 0, 0
   shps:dict[int, dict] = {}
   temp:dict[int, dict] = {}
@@ -170,9 +170,10 @@ def mem_layout(events:list[tuple[int, int, float, DevEvent]]) -> dict:
   for v in temp.values():
     v["x"].append(step)
     v["y"].append(v["y"][-1])
+  timestamps.append(max_ts)
   return {"shapes":list(shps.values()), "peak":peak, "timestamps":timestamps}
 
-def get_profile(profile:list[ProfileEvent]):
+def get_profile(profile:list[ProfileEvent]) -> bytes|None:
   # start by getting the time diffs
   for ev in profile:
     if isinstance(ev,ProfileDeviceEvent): device_ts_diffs[ev.device] = (ev.comp_tdiff, ev.copy_tdiff if ev.copy_tdiff is not None else ev.comp_tdiff)
@@ -184,10 +185,14 @@ def get_profile(profile:list[ProfileEvent]):
     dev_events.setdefault(e.device,[]).append((st:=int(ts), et:=int(en), float(en-ts), e))
     if min_ts is None or st < min_ts: min_ts = st
     if max_ts is None or et > max_ts: max_ts = et
+  if min_ts is None: return None
   # return layout of per device events
-  for events in dev_events.values(): events.sort(key=lambda v:v[0])
-  dev_layout = {k:{"timeline":timeline_layout(v), "mem":mem_layout(v)} for k,v in dev_events.items()}
-  return json.dumps({"layout":dev_layout, "st":min_ts, "et":max_ts}).encode("utf-8")
+  layout:dict[str, dict] = {}
+  for k,v in dev_events.items():
+    v.sort(key=lambda e:e[0])
+    layout[k] = timeline_layout(v)
+    layout[f"{k} Memory"] = mem_layout(v, unwrap(max_ts))
+  return json.dumps({"layout":layout, "st":min_ts, "et":max_ts}).encode("utf-8")
 
 def get_runtime_stats(key) -> list[dict]:
   ret:list[dict] = []