fix later

mixin
use Self type
2025-11-02 13:18:41 +08:00 · 2025-11-02 13:05:26 +08:00 · 2025-11-02 13:00:35 +08:00
183 changed files with 215766 additions and 145350 deletions
@@ -1,3 +0,0 @@
-[run]
-source = tinygrad
-branch = True
@@ -61,7 +61,7 @@ runs:
      uses: actions/cache@v4
      with:
        path: ${{ github.workspace }}/.venv
-        key: venv-${{ runner.os }}-python-${{ steps.setup-python.outputs.python-version }}-${{ inputs.deps }}-${{ inputs.pydeps }}-${{ hashFiles('**/pyproject.toml') }}-${{ env.CACHE_VERSION }}
+        key: venv-${{ runner.os }}-python-${{ steps.setup-python.outputs.python-version }}-${{ inputs.deps }}-${{ inputs.pydeps }}-${{ hashFiles('**/setup.py') }}-${{ env.PYTHON_CACHE_VERSION }}

    # **** Caching downloads ****

@@ -70,13 +70,13 @@ runs:
      uses: actions/cache@v4
      with:
        path: ~/.cache/tinygrad/downloads/
-        key: downloads-cache-${{ inputs.key }}-${{ env.CACHE_VERSION }}
+        key: downloads-cache-${{ inputs.key }}-${{ env.DOWNLOAD_CACHE_VERSION }}
    - name: Cache downloads (macOS)
      if: inputs.key != '' && runner.os == 'macOS'
      uses: actions/cache@v4
      with:
        path: ~/Library/Caches/tinygrad/downloads/
-        key: osx-downloads-cache-${{ inputs.key }}-${{ env.CACHE_VERSION }}
+        key: osx-downloads-cache-${{ inputs.key }}-${{ env.DOWNLOAD_CACHE_VERSION }}

    # **** Python deps ****

@@ -187,7 +187,7 @@ runs:
      uses: actions/cache@v4
      with:
        path: /var/cache/apt/archives/
-        key: ${{ runner.os }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.CACHE_VERSION }}
+        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')
@@ -247,7 +247,7 @@ runs:
        cache-name: cache-gpuocelot-build-1
      with:
        path: ${{ github.workspace }}/gpuocelot/ocelot
-        key: ${{ runner.os }}-gpuocelot-b16039dc940dc6bc4ea0a98380495769ff35ed99-rebuild-${{ env.CACHE_VERSION }}
+        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
@@ -1,7 +1,10 @@
 name: Autogen
 env:
  # increment this when downloads substantially change to avoid the internet
-  CACHE_VERSION: '13'
+  DOWNLOAD_CACHE_VERSION: '12'
+  PYTHON_CACHE_VERSION: '4'
+  APT_CACHE_VERSION: '1'
+  BUILD_CACHE_VERSION: '1'
  CAPTURE_PROCESS_REPLAY: 1
  GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
  PYTHONPATH: ${{ github.workspace }}
@@ -11,15 +14,15 @@ on:
    branches:
      - master
  pull_request:
-    paths:
+    paths: 
    - 'tinygrad/runtime/autogen/**/*'
  workflow_dispatch:
-    paths:
+    paths: 
    - 'tinygrad/runtime/autogen/**/*'

 jobs:
  autogen:
-    name: In-tree Autogen
+    name: Autogen
    runs-on: ubuntu-24.04
    timeout-minutes: 15
    steps:
@@ -31,149 +34,64 @@ jobs:
        opencl: 'true'
        amd: 'true'
        cuda: 'true'
-        llvm: 'true'
        webgpu: 'true'
-        mesa: 'true'
-        pydeps: 'clang>=20 pyyaml mako'
+        llvm: 'true'
+        pydeps: 'pyyaml mako'
    - name: Install autogen support packages
-      run: sudo apt-get install -y --no-install-recommends libclang-20-dev llvm-20-dev hip-dev libusb-1.0-0-dev
+      run: sudo apt-get install -y --no-install-recommends llvm-14-dev libclang-14-dev llvm-20-dev
    - name: Verify OpenCL autogen
      run: |
-        mv tinygrad/runtime/autogen/opencl.py /tmp/opencl.py.bak
-        python3 -c "from tinygrad.runtime.autogen import opencl"
+        cp tinygrad/runtime/autogen/opencl.py /tmp/opencl.py.bak
+        ./autogen_stubs.sh opencl
        diff /tmp/opencl.py.bak tinygrad/runtime/autogen/opencl.py
    - name: Verify CUDA autogen
      run: |
-        mv tinygrad/runtime/autogen/cuda.py /tmp/cuda.py.bak
-        mv tinygrad/runtime/autogen/nvrtc.py /tmp/nvrtc.py.bak
-        mv tinygrad/runtime/autogen/nvjitlink.py /tmp/nvjitlink.py.bak
-        mv tinygrad/runtime/autogen/nv_570.py /tmp/nv_570.py.bak
-        mv tinygrad/runtime/autogen/nv.py /tmp/nv.py.bak
-        python3 -c "from tinygrad.runtime.autogen import cuda, nvrtc, nvjitlink, nv_570, nv"
+        cp tinygrad/runtime/autogen/cuda.py /tmp/cuda.py.bak
+        cp tinygrad/runtime/autogen/nv_gpu.py /tmp/nv_gpu.py.bak
+        ./autogen_stubs.sh cuda
+        ./autogen_stubs.sh nv
        diff /tmp/cuda.py.bak tinygrad/runtime/autogen/cuda.py
-        diff /tmp/nvrtc.py.bak tinygrad/runtime/autogen/nvrtc.py
-        diff /tmp/nvjitlink.py.bak tinygrad/runtime/autogen/nvjitlink.py
-        diff /tmp/nv_570.py.bak tinygrad/runtime/autogen/nv_570.py
-        diff /tmp/nv.py.bak tinygrad/runtime/autogen/nv.py
+        diff /tmp/nv_gpu.py.bak tinygrad/runtime/autogen/nv_gpu.py
    - name: Verify AMD autogen
      run: |
-        mv tinygrad/runtime/autogen/comgr.py /tmp/comgr.py.bak
-        mv tinygrad/runtime/autogen/hsa.py /tmp/hsa.py.bak
-        mv tinygrad/runtime/autogen/hip.py /tmp/hip.py.bak
-        mv tinygrad/runtime/autogen/amd_gpu.py /tmp/amd_gpu.py.bak
-        mv tinygrad/runtime/autogen/sqtt.py /tmp/sqtt.py.bak
-        mv tinygrad/runtime/autogen/rocprof.py /tmp/rocprof.py.bak
-        mv tinygrad/runtime/autogen/am/am.py /tmp/am_am.py.bak
-        mv tinygrad/runtime/autogen/am/pm4_soc15.py /tmp/am_pm4_soc15.py.bak
-        mv tinygrad/runtime/autogen/am/pm4_nv.py /tmp/am_pm4_nv.py.bak
-        mv tinygrad/runtime/autogen/am/sdma_4_0_0.py /tmp/am_sdma_4_0_0.py.bak
-        mv tinygrad/runtime/autogen/am/sdma_5_0_0.py /tmp/am_sdma_5_0_0.py.bak
-        mv tinygrad/runtime/autogen/am/sdma_6_0_0.py /tmp/am_sdma_6_0_0.py.bak
-        mv tinygrad/runtime/autogen/am/smu_v13_0_0.py /tmp/am_smu_v13_0_0.py.bak
-        mv tinygrad/runtime/autogen/am/smu_v14_0_2.py /tmp/am_smu_v14_0_2.py.bak
-        python3 -c "from tinygrad.runtime.autogen import comgr, hsa, hip, amd_gpu, sqtt, rocprof; from tinygrad.runtime.autogen.am import am, pm4_soc15, pm4_nv, sdma_4_0_0, sdma_5_0_0, sdma_6_0_0, smu_v13_0_0, smu_v14_0_2"
-        diff /tmp/comgr.py.bak tinygrad/runtime/autogen/comgr.py
+        cp tinygrad/runtime/autogen/hsa.py /tmp/hsa.py.bak
+        cp tinygrad/runtime/autogen/kfd.py /tmp/kfd.py.bak
+        cp tinygrad/runtime/autogen/comgr.py /tmp/comgr.py.bak
+        cp tinygrad/runtime/autogen/amd_gpu.py /tmp/amd_gpu.py.bak
+        cp tinygrad/runtime/autogen/sqtt.py /tmp/sqtt.py.bak
+        ./autogen_stubs.sh hsa
+        ./autogen_stubs.sh kfd
+        ./autogen_stubs.sh comgr
+        ./autogen_stubs.sh amd
+        ./autogen_stubs.sh sqtt
        diff /tmp/hsa.py.bak tinygrad/runtime/autogen/hsa.py
-        diff /tmp/hip.py.bak tinygrad/runtime/autogen/hip.py
+        diff /tmp/kfd.py.bak tinygrad/runtime/autogen/kfd.py
+        diff /tmp/comgr.py.bak tinygrad/runtime/autogen/comgr.py
        diff /tmp/amd_gpu.py.bak tinygrad/runtime/autogen/amd_gpu.py
        diff /tmp/sqtt.py.bak tinygrad/runtime/autogen/sqtt.py
-        diff /tmp/rocprof.py.bak tinygrad/runtime/autogen/rocprof.py
-        diff /tmp/am_am.py.bak tinygrad/runtime/autogen/am/am.py
-        diff /tmp/am_pm4_soc15.py.bak tinygrad/runtime/autogen/am/pm4_soc15.py
-        diff /tmp/am_pm4_nv.py.bak tinygrad/runtime/autogen/am/pm4_nv.py
-        diff /tmp/am_sdma_4_0_0.py.bak tinygrad/runtime/autogen/am/sdma_4_0_0.py
-        diff /tmp/am_sdma_5_0_0.py.bak tinygrad/runtime/autogen/am/sdma_5_0_0.py
-        diff /tmp/am_sdma_6_0_0.py.bak tinygrad/runtime/autogen/am/sdma_6_0_0.py
-        diff /tmp/am_smu_v13_0_0.py.bak tinygrad/runtime/autogen/am/smu_v13_0_0.py
-        diff /tmp/am_smu_v14_0_2.py.bak tinygrad/runtime/autogen/am/smu_v14_0_2.py
    - name: Verify Linux autogen
      run: |
-        mv tinygrad/runtime/autogen/libc.py /tmp/libc.py.bak
-        mv tinygrad/runtime/autogen/kfd.py /tmp/kfd.py.bak
-        mv tinygrad/runtime/autogen/io_uring.py /tmp/io_uring.py.bak
-        mv tinygrad/runtime/autogen/ib.py /tmp/ib.py.bak
-        mv tinygrad/runtime/autogen/pci.py /tmp/pci.py.bak
-        mv tinygrad/runtime/autogen/vfio.py /tmp/vfio.py.bak
-        python3 -c "from tinygrad.runtime.autogen import libc, kfd, io_uring, ib, pci, vfio"
+        cp tinygrad/runtime/autogen/libc.py /tmp/libc.py.bak
+        cp tinygrad/runtime/autogen/io_uring.py /tmp/io_uring.py.bak
+        cp tinygrad/runtime/autogen/ib.py /tmp/ib.py.bak
+        ./autogen_stubs.sh libc
+        ./autogen_stubs.sh io_uring
+        ./autogen_stubs.sh ib
        diff /tmp/libc.py.bak tinygrad/runtime/autogen/libc.py
-        diff /tmp/kfd.py.bak tinygrad/runtime/autogen/kfd.py
        diff /tmp/io_uring.py.bak tinygrad/runtime/autogen/io_uring.py
        diff /tmp/ib.py.bak tinygrad/runtime/autogen/ib.py
-        diff /tmp/pci.py.bak tinygrad/runtime/autogen/pci.py
-        diff /tmp/vfio.py.bak tinygrad/runtime/autogen/vfio.py
-    - name: Verify LLVM autogen
-      run: |
-        mv tinygrad/runtime/autogen/llvm.py /tmp/llvm.py.bak
-        python3 -c "from tinygrad.runtime.autogen import llvm"
-        diff /tmp/llvm.py.bak tinygrad/runtime/autogen/llvm.py
    - name: Verify WebGPU autogen
      run: |
-        mv tinygrad/runtime/autogen/webgpu.py /tmp/webgpu.py.bak
-        python3 -c "from tinygrad.runtime.autogen import webgpu"
+        cp tinygrad/runtime/autogen/webgpu.py /tmp/webgpu.py.bak
+        ./autogen_stubs.sh webgpu
        diff /tmp/webgpu.py.bak tinygrad/runtime/autogen/webgpu.py
-    - name: Verify Qualcomm autogen
+    - name: Verify LLVM autogen
      run: |
-        mv tinygrad/runtime/autogen/kgsl.py /tmp/kgsl.py.bak
-        mv tinygrad/runtime/autogen/adreno.py /tmp/adreno.py.bak
-        mv tinygrad/runtime/autogen/qcom_dsp.py /tmp/qcom_dsp.py.bak
-        python3 -c "from tinygrad.runtime.autogen import kgsl, adreno, qcom_dsp"
-        diff /tmp/kgsl.py.bak tinygrad/runtime/autogen/kgsl.py
-        diff /tmp/adreno.py.bak tinygrad/runtime/autogen/adreno.py
-        diff /tmp/qcom_dsp.py.bak tinygrad/runtime/autogen/qcom_dsp.py
-    - name: Verify libusb autogen
-      run: |
-        mv tinygrad/runtime/autogen/libusb.py /tmp/libusb.py.bak
-        python3 -c "from tinygrad.runtime.autogen import libusb"
-        diff /tmp/libusb.py.bak tinygrad/runtime/autogen/libusb.py
+        cp tinygrad/runtime/autogen/llvm.py /tmp/llvm.py.bak
+        ./autogen_stubs.sh llvm
+        diff /tmp/llvm.py.bak tinygrad/runtime/autogen/llvm.py
    - name: Verify mesa autogen
      run: |
-        mv tinygrad/runtime/autogen/mesa.py /tmp/mesa.py.bak
-        python3 -c "from tinygrad.runtime.autogen import mesa"
+        cp tinygrad/runtime/autogen/mesa.py /tmp/mesa.py.bak
+        ./autogen_stubs.sh mesa
        diff /tmp/mesa.py.bak tinygrad/runtime/autogen/mesa.py
-    - name: Verify libclang autogen
-      run: |
-        mv tinygrad/runtime/autogen/libclang.py /tmp/libclang.py.bak
-        python3 -c "from tinygrad.runtime.autogen import libclang"
-        diff /tmp/libclang.py.bak tinygrad/runtime/autogen/libclang.py
-  autogen-mac:
-    name: In-tree Autogen (macos)
-    runs-on: macos-14
-    timeout-minutes: 15
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v4
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        llvm: 'true'
-        pydeps: 'clang>=20'
-    - name: Verify macos autogen
-      run: |
-        mv tinygrad/runtime/autogen/metal.py /tmp/metal.py.bak
-        LIBCLANG_PATH=/opt/homebrew/opt/llvm@20/lib/libclang.dylib python3 -c "from tinygrad.runtime.autogen import metal"
-        diff /tmp/metal.py.bak tinygrad/runtime/autogen/metal.py
-  autogen-comgr-3:
-    name: In-tree Autogen (comgr 3)
-    runs-on: ubuntu-24.04
-    timeout-minutes: 15
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v4
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        pydeps: 'clang>=20'
-    - name: Install autogen support packages
-      run: |
-        wget https://repo.radeon.com/rocm/rocm.gpg.key -O - | gpg --dearmor | sudo tee /etc/apt/keyrings/rocm.gpg > /dev/null
-        sudo tee /etc/apt/sources.list.d/rocm.list <<EOF
-        deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/6.4 $(lsb_release -cs) main
-        EOF
-        echo -e 'Package: *\nPin: release o=repo.radeon.com\nPin-Priority: 600' | sudo tee /etc/apt/preferences.d/rocm-pin-600
-        sudo apt -qq update || true
-        sudo apt-get install -y --no-install-recommends libclang-20-dev comgr
-    - name: Verify comgr (3) autogen
-      run: |
-        mv tinygrad/runtime/autogen/comgr_3.py /tmp/comgr_3.py.bak
-        python3 -c "from tinygrad.runtime.autogen import comgr_3"
-        diff /tmp/comgr_3.py.bak tinygrad/runtime/autogen/comgr_3.py
@@ -54,7 +54,7 @@ jobs:
    - name: Print macOS version
      run: sw_vers
    - name: Run Stable Diffusion
-      run: BENCHMARK_LOG=stable_diffusion JIT=1 ASSERT_MIN_STEP_TIME=800 python3.11 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
+      run: BENCHMARK_LOG=stable_diffusion JIT=1 ASSERT_MIN_STEP_TIME=720 python3.11 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
    - name: Run Stable Diffusion without fp16
      run: BENCHMARK_LOG=stable_diffusion_fp32 JIT=1 ASSERT_MIN_STEP_TIME=800 python3.11 examples/stable_diffusion.py --seed 0 --noshow --timing | tee sd_no_fp16.txt
    - name: Run Stable Diffusion v2
@@ -199,7 +199,7 @@ jobs:
    - 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
    - name: Test speed vs theoretical
-      run: NV=1 IGNORE_BEAM_CACHE=1 CCACHE=0 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
+      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
      run: NV=1 python test/external/external_benchmark_multitensor_allreduce.py
    - name: Test tensor cores
@@ -320,20 +320,19 @@ jobs:
    #   run: NV=1 NV_PTX=1 M_START=12 M_STOP=20 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 python3 ./extra/gemm/fuzz_matmul.py
    - name: Train MNIST
      run: time PYTHONPATH=. NV=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py | tee beautiful_mnist.txt
-    # TODO: too slow
    - name: Run 10 CIFAR training steps
-      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=1300 NV=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
-    # - name: Run 10 CIFAR training steps w HALF
-    #   run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=240 NV=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_half.txt
-    # - name: Run 10 CIFAR training steps w BF16
-    #   run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=270 NV=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
+      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=270 NV=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
+    - name: Run 10 CIFAR training steps w HALF
+      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=240 NV=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_half.txt
+    - name: Run 10 CIFAR training steps w BF16
+      run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=270 NV=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
    # TODO: too slow
    # - name: Run 10 CIFAR training steps w winograd
    #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=350 NV=1 CAPTURE_PROCESS_REPLAY=0 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_wino.txt
-    # - name: Run full CIFAR training w 1 GPU
-    #   run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_one_gpu.txt
-    # - name: Run full CIFAR training steps w 6 GPUS
-    #   run: time BENCHMARK_LOG=cifar_6gpu CAPTURE_PROCESS_REPLAY=0 NV=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu.txt
+    - name: Run full CIFAR training w 1 GPU
+      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_one_gpu.txt
+    - name: Run full CIFAR training steps w 6 GPUS
+      run: time BENCHMARK_LOG=cifar_6gpu CAPTURE_PROCESS_REPLAY=0 NV=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu.txt
    - name: Run MLPerf resnet eval on training data
      run: time BENCHMARK_LOG=resnet_eval NV=1 MODEL=resnet python3 examples/mlperf/model_eval.py
    #- name: Run 10 MLPerf ResNet50 training steps (1 gpu)
@@ -410,7 +409,7 @@ jobs:
    #    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
    - name: Test speed vs theoretical
-      run: AMD=1 IGNORE_BEAM_CACHE=1 CCACHE=0 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
+      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
      run: |
        AMD=1 AMD_LLVM=0 python3 test/opt/test_tensor_cores.py
@@ -525,18 +524,17 @@ jobs:
      run: test/external/process_replay/reset.py
    - name: Train MNIST
      run: time PYTHONPATH=. AMD=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py | tee beautiful_mnist.txt
-    # TODO: too slow
    - name: Run 10 CIFAR training steps
-      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=2000 AMD=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
-    # - name: Run 10 CIFAR training steps w HALF
-    #   run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=390 AMD=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_half.txt
+      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=330 AMD=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
+    - name: Run 10 CIFAR training steps w HALF
+      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=350 AMD=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_half.txt
    # - name: Run 10 CIFAR training steps w BF16
    #   run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=288 AMD=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
    # TODO: too slow
    # - name: Run 10 CIFAR training steps w winograd
    #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=66 AMD=1 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_wino.txt
-    # - name: Run full CIFAR training w 1 GPU
-    #   run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_one_gpu.txt
+    - name: Run full CIFAR training w 1 GPU
+      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_one_gpu.txt
    #- name: Run full CIFAR training steps w 6 GPUS
    #  run: time BENCHMARK_LOG=cifar_6gpu AMD=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu.txt
    #- name: Run full CIFAR training steps w 6 GPUS (REMOTE)
@@ -625,24 +623,24 @@ jobs:
        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
    - name: reset process replay
      run: test/external/process_replay/reset.py
-    # - name: openpilot compile3 0.9.9 driving_vision
-    #   run: BENCHMARK_LOG=openpilot_0_9_9_vision PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_vision.onnx
-    # - name: openpilot compile3 0.9.9 driving_policy
-    #   run: BENCHMARK_LOG=openpilot_0_9_9_policy PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_policy.onnx
-    # - name: openpilot compile3 0.9.9 dmonitoring
-    #   run: BENCHMARK_LOG=openpilot_0_9_9_dmonitoring PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/dmonitoring_model.onnx
+    - name: openpilot compile3 0.9.9 driving_vision
+      run: BENCHMARK_LOG=openpilot_0_9_9_vision PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_vision.onnx
+    - name: openpilot compile3 0.9.9 driving_policy
+      run: BENCHMARK_LOG=openpilot_0_9_9_policy PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_policy.onnx
+    - name: openpilot compile3 0.9.9 dmonitoring
+      run: BENCHMARK_LOG=openpilot_0_9_9_dmonitoring PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/dmonitoring_model.onnx
    - name: openpilot compile3 0.10.0 driving_policy
-      run: BENCHMARK_LOG=openpilot_0_10_0_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=4 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.10.0/selfdrive/modeld/models/driving_policy.onnx
+      run: BENCHMARK_LOG=openpilot_0_10_0_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=5 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.10.0/selfdrive/modeld/models/driving_policy.onnx
    - name: openpilot compile3 0.10.0 dmonitoring
-      run: BENCHMARK_LOG=openpilot_0_10_0_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=11 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.10.0/selfdrive/modeld/models/dmonitoring_model.onnx
-    - name: DEBUG=2 openpilot compile3 0.10.1 driving_vision
-      run: PYTHONPATH="." DEBUG=2 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
+      run: BENCHMARK_LOG=openpilot_0_10_0_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=13 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.10.0/selfdrive/modeld/models/dmonitoring_model.onnx
    - name: openpilot compile3 0.10.1 driving_vision
-      run: BENCHMARK_LOG=openpilot_0_10_1_vision PYTHONPATH="." ASSERT_MIN_STEP_TIME=17 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
+      # TODO: ASSERT_MIN_STEP_TIME=17
+      run: BENCHMARK_LOG=openpilot_0_10_1_vision PYTHONPATH="." ASSERT_MIN_STEP_TIME=25 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
    - name: openpilot compile3 0.10.1 driving_policy
-      run: BENCHMARK_LOG=openpilot_0_10_1_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=4 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_policy.onnx
+      run: BENCHMARK_LOG=openpilot_0_10_1_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=5 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_policy.onnx
    - name: openpilot compile3 0.10.1 dmonitoring
-      run: BENCHMARK_LOG=openpilot_0_10_1_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=10 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/dmonitoring_model.onnx
+      # TODO: ASSERT_MIN_STEP_TIME=10
+      run: BENCHMARK_LOG=openpilot_0_10_1_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=13 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/dmonitoring_model.onnx
    - name: benchmark MobileNetV2 on DSP
      run: |
        # generate quantized weights
@@ -706,9 +704,8 @@ jobs:
      run: |
        AMD=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyDefaulttoCPUJit
        AMD=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyCPUtoDefaultJit
-    # TODO: too slow
-    # - name: Run full CIFAR training w 1 GPU
-    #   run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee am_train_cifar_one_gpu.txt
+    - name: Run full CIFAR training w 1 GPU
+      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee am_train_cifar_one_gpu.txt
    # TODO: enable
    # - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
    #   run: BENCHMARK_LOG=resnet_10steps AMD=1 MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee am_train_resnet_one_gpu.txt
@@ -770,9 +767,8 @@ jobs:
        NV=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyCPUtoDefaultJit
    - name: Test LLAMA-3
      run: BENCHMARK_LOG=llama3_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --benchmark --temperature 0 | tee nv_llama3_beam.txt
-    # TODO: too slow
-    # - name: Run full CIFAR training w 1 GPU
-    #   run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee nv_train_cifar_one_gpu.txt
+    - name: Run full CIFAR training w 1 GPU
+      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee nv_train_cifar_one_gpu.txt
    #- name: Run 10 MLPerf ResNet50 training steps (1 gpu)
    #  run: BENCHMARK_LOG=resnet_10steps NV=1 MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee nv_train_resnet_one_gpu.txt
    - name: Run 10 MLPerf Bert training steps (1 gpu)
@@ -22,13 +22,13 @@ jobs:
    - name: Run SDXL with new search
      # TODO: GCVM_L2_PROTECTION_FAULT_STATUS with llvm19
      run: |
-        BENCHMARK_LOG=search_sdxl PYTHONPATH=. AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 CCACHE=0 python examples/sdxl.py --noshow --timing --seed 0
+        BENCHMARK_LOG=search_sdxl PYTHONPATH=. AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 DISABLE_COMPILER_CACHE=1 python examples/sdxl.py --noshow --timing --seed 0
    - name: Run SDXL with cached search
      run: |
        BENCHMARK_LOG=search_sdxl_cached PYTHONPATH=. AMD=1 JITBEAM=2 python examples/sdxl.py --noshow --timing --seed 0
    - name: Run winograd cifar with new search
      run: |
-        BENCHMARK_LOG=search_wino_cifar WINO=1 DEFAULT_FLOAT=HALF JITBEAM=4 IGNORE_BEAM_CACHE=1 CCACHE=0 BS=1024 STEPS=500 python examples/hlb_cifar10.py
+        BENCHMARK_LOG=search_wino_cifar WINO=1 DEFAULT_FLOAT=HALF JITBEAM=4 IGNORE_BEAM_CACHE=1 DISABLE_COMPILER_CACHE=1 BS=1024 STEPS=500 python examples/hlb_cifar10.py
    - name: Run winograd cifar with cached search
      run: |
        BENCHMARK_LOG=search_wino_cifar_cached WINO=1 DEFAULT_FLOAT=HALF JITBEAM=4 BS=1024 STEPS=500 python examples/hlb_cifar10.py
@@ -27,4 +27,4 @@ jobs:
      run: |
        rm "~/.cache/tinygrad/cache_mlperf.db" || true
        BENCHMARK_LOG=mlpert_train_resnet LOGMLPERF=0 CACHEDB="~/.cache/tinygrad/cache_mlperf.db" examples/mlperf/training_submission_v5.1/tinycorp/benchmarks/resnet/implementations/tinybox_red/run_and_time.sh
-        rm "~/.cache/tinygrad/cache_mlperf.db"
+        rm "~/.cache/tinygrad/cache_mlperf.db"
@@ -20,11 +20,11 @@ jobs:
    - name: Install dependencies
      run: |
        python -m pip install --upgrade pip
-        pip install setuptools wheel build twine
+        pip install setuptools wheel twine
    - name: Build and publish
      env:
        TWINE_USERNAME: ${{ secrets.PYPI_USERNAME }}
        TWINE_PASSWORD: ${{ secrets.PYPI_PASSWORD }}
      run: |
-        python -m build
+        python setup.py sdist bdist_wheel
        twine upload dist/*
@@ -1,7 +1,10 @@
 name: Unit Tests
 env:
  # increment this when downloads substantially change to avoid the internet
-  CACHE_VERSION: '13'
+  DOWNLOAD_CACHE_VERSION: '12'
+  PYTHON_CACHE_VERSION: '4'
+  APT_CACHE_VERSION: '1'
+  BUILD_CACHE_VERSION: '1'
  CAPTURE_PROCESS_REPLAY: 1
  GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
  PYTHONPATH: ${{ github.workspace }}
@@ -287,8 +290,8 @@ jobs:
        python extra/optimization/extract_dataset.py
        gzip -c /tmp/sops > extra/datasets/sops.gz
        #DEBUG=1 MIN_ASTS=1 python extra/optimization/get_action_space.py
-    - name: Repo line count < 18500 lines
-      run: MAX_LINE_COUNT=18500 python sz.py
+    - name: Repo line count < 18000 lines
+      run: MAX_LINE_COUNT=18000 python sz.py

  spec:
    strategy:
@@ -341,11 +344,10 @@ jobs:
          key: gpu-image
          deps: testing_minimal
          opencl: 'true'
-      - name: Test CL IMAGE=2 ops
+      - name: Test CL IMAGE=2 ops + training
        run: |
          CL=1 IMAGE=2 python -m pytest -n=auto test/test_ops.py --durations=20
-        # TODO: training is broken
-        # CL=1 IMAGE=2 python test/models/test_end2end.py TestEnd2End.test_linear_mnist
+          CL=1 IMAGE=2 python test/models/test_end2end.py TestEnd2End.test_linear_mnist
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

@@ -390,7 +392,7 @@ jobs:
          llvm: 'true'
      - name: Test openpilot model kernel count and gate usage
        run: |
-          ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1397 ALLOWED_GATED_READ_IMAGE=94 FLOAT16=1 CL=1 IMAGE=2 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
+          ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1452 ALLOWED_GATED_READ_IMAGE=122 FLOAT16=1 CL=1 IMAGE=2 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
      - name: Test openpilot CL compile fp16
        run: FLOAT16=1 DEBUGCL=1 CL=1 IMAGE=2 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
      - name: Test openpilot CL compile fp32 (test correctness)
@@ -63,5 +63,3 @@ profile_stats
 *.log
 target
 .mypy_cache
-mutants
-.mutmut-cache
@@ -28,7 +28,7 @@ repos:
        pass_filenames: false
      - id: tests
        name: subset of tests
-        entry: env OMP_NUM_THREADS=1 PYTHONPATH="." python3 -m pytest -n=6 test/test_ops.py test/test_dtype.py test/test_schedule.py test/test_assign.py
+        entry: env OMP_NUM_THREADS=1 PYTHONPATH="." python3 -m pytest -n=8 test/test_ops.py test/test_dtype.py test/test_schedule.py test/test_assign.py
        language: system
        always_run: true
        pass_filenames: false
@@ -21,38 +21,17 @@ tinygrad: For something between [PyTorch](https://github.com/pytorch/pytorch) an

 ---

-tinygrad is an end-to-end deep learning stack:
+Despite tinygrad's size, it is a fully featured deep learning framework.

- **Tensor library** with autograd
- **IR and compiler** that fuse and lower kernels
- **JIT + graph execution**
- **nn / optim / datasets** for real training
+Due to its extreme simplicity, it is the easiest framework to add new accelerators to, with support for both inference and training. If XLA is CISC, tinygrad is RISC.

-It’s inspired by PyTorch (ergonomics), JAX (functional transforms and IR-based AD), and TVM (scheduling and codegen), but stays intentionally tiny and hackable.
+tinygrad is now beta software, we [raised some money](https://geohot.github.io/blog/jekyll/update/2023/05/24/the-tiny-corp-raised-5M.html) to make it good. Someday, we will tape out chips.

---
+## Features

-## How tinygrad compares
+### LLaMA and Stable Diffusion

-**PyTorch**
-
- ✅ Similar: eager `Tensor` API, autograd, `optim`, basic datasets and layers.
- ✅ You can write familiar training loops.
- 🔁 Unlike PyTorch, the entire compiler and IR are visible and hackable.
-
-**JAX**
-
- ✅ IR-based autodiff over primitives (like JAXPR + XLA).
- ✅ Function-level JIT (`TinyJit`) that captures and replays kernels.
- 🔁 Fewer functional transforms (no full `vmap`/`pmap` yet), but far easier to read.
-
-**TVM**
-
- ✅ Multiple lowering passes, scheduling, and BEAM search over kernels.
- ✅ Device “graphs” for batched execution.
- 🔁 tinygrad also ships the **front-end framework** (tensors, nn, optim), not just the compiler.
-
---
+tinygrad can run [LLaMA](/docs/showcase.md#llama) and [Stable Diffusion](/docs/showcase.md#stable-diffusion)!

 ### Laziness

@@ -0,0 +1,564 @@
+#!/bin/bash -e
+
+# setup instructions for clang2py
+if [[ ! $(clang2py -V) ]]; then
+  pushd .
+  cd /tmp
+  sudo apt-get install -y --no-install-recommends clang
+  pip install --upgrade pip setuptools
+  pip install clang==14.0.6
+  git clone https://github.com/nimlgen/ctypeslib.git
+  cd ctypeslib
+  pip install .
+  clang2py -V
+  popd
+fi
+
+BASE=tinygrad/runtime/autogen/
+
+fixup() {
+  sed -i '1s/^/# mypy: ignore-errors\n/' $1
+  sed -i 's/ *$//' $1
+  grep FIXME_STUB $1 || true
+}
+
+patch_dlopen() {
+  path=$1; shift
+  name=$1; shift
+  cat <<EOF | sed -i "/import ctypes.*/r /dev/stdin" $path
+PATHS_TO_TRY = [
+$(for p in "$@"; do echo "  $p,"; done)
+]
+def _try_dlopen_$name():
+  library = ctypes.util.find_library("$name")
+  if library: return ctypes.CDLL(library)
+  for candidate in PATHS_TO_TRY:
+    try: return ctypes.CDLL(candidate)
+    except OSError: pass
+  return None
+EOF
+}
+
+generate_opencl() {
+  clang2py /usr/include/CL/cl.h -o $BASE/opencl.py -l /usr/lib/x86_64-linux-gnu/libOpenCL.so.1 -k cdefstum
+  fixup $BASE/opencl.py
+  # hot patches
+  sed -i "s\import ctypes\import ctypes, ctypes.util\g" $BASE/opencl.py
+  sed -i "s\ctypes.CDLL('/usr/lib/x86_64-linux-gnu/libOpenCL.so.1')\ctypes.CDLL(ctypes.util.find_library('OpenCL'))\g" $BASE/opencl.py
+  python3 -c "import tinygrad.runtime.autogen.opencl"
+}
+
+generate_hip() {
+  clang2py /opt/rocm/include/hip/hip_ext.h /opt/rocm/include/hip/hiprtc.h \
+  /opt/rocm/include/hip/hip_runtime_api.h /opt/rocm/include/hip/driver_types.h \
+  --clang-args="-D__HIP_PLATFORM_AMD__ -I/opt/rocm/include -x c++" -o $BASE/hip.py -l /opt/rocm/lib/libamdhip64.so
+  echo "hipDeviceProp_t = hipDeviceProp_tR0600" >> $BASE/hip.py
+  echo "hipGetDeviceProperties = hipGetDevicePropertiesR0600" >> $BASE/hip.py
+  fixup $BASE/hip.py
+  # we can trust HIP is always at /opt/rocm/lib
+  #sed -i "s\import ctypes\import ctypes, ctypes.util\g" $BASE/hip.py
+  #sed -i "s\ctypes.CDLL('/opt/rocm/lib/libhiprtc.so')\ctypes.CDLL(ctypes.util.find_library('hiprtc'))\g" $BASE/hip.py
+  #sed -i "s\ctypes.CDLL('/opt/rocm/lib/libamdhip64.so')\ctypes.CDLL(ctypes.util.find_library('amdhip64'))\g" $BASE/hip.py
+  sed -i "s\import ctypes\import ctypes, os\g" $BASE/hip.py
+  sed -i "s\'/opt/rocm/\os.getenv('ROCM_PATH', '/opt/rocm/')+'/\g" $BASE/hip.py
+  python3 -c "import tinygrad.runtime.autogen.hip"
+}
+
+generate_comgr() {
+  clang2py /opt/rocm/include/amd_comgr/amd_comgr.h \
+  --clang-args="-D__HIP_PLATFORM_AMD__ -I/opt/rocm/include -x c++" -o $BASE/comgr.py -l /opt/rocm/lib/libamd_comgr.so
+  fixup $BASE/comgr.py
+  sed -i "s\import ctypes\import ctypes, ctypes.util, os\g" $BASE/comgr.py
+  patch_dlopen $BASE/comgr.py amd_comgr "'/opt/rocm/lib/libamd_comgr.so'" "os.getenv('ROCM_PATH', '')+'/lib/libamd_comgr.so'" "'/usr/local/lib/libamd_comgr.dylib'" "'/opt/homebrew/lib/libamd_comgr.dylib'"
+  sed -i "s\ctypes.CDLL('/opt/rocm/lib/libamd_comgr.so')\_try_dlopen_amd_comgr()\g" $BASE/comgr.py
+  python3 -c "import tinygrad.runtime.autogen.comgr"
+}
+
+generate_kfd() {
+  clang2py /usr/include/linux/kfd_ioctl.h -o $BASE/kfd.py -k cdefstum
+
+  fixup $BASE/kfd.py
+  sed -i "s/import ctypes/import ctypes, os/g" $BASE/kfd.py
+  sed -i "s/import fcntl, functools/import functools/g" $BASE/kfd.py
+  sed -i "/import functools/a from tinygrad.runtime.support.hcq import FileIOInterface" $BASE/kfd.py
+  sed -i "s/def _do_ioctl(__idir, __base, __nr, __user_struct, __fd, \*\*kwargs):/def _do_ioctl(__idir, __base, __nr, __user_struct, __fd:FileIOInterface, \*\*kwargs):/g" $BASE/kfd.py
+  sed -i "s/fcntl.ioctl(__fd, (__idir<<30)/__fd.ioctl((__idir<<30)/g" $BASE/kfd.py
+  sed -i "s/!!/not not /g" $BASE/kfd.py
+  python3 -c "import tinygrad.runtime.autogen.kfd"
+}
+
+generate_cuda() {
+  clang2py /usr/include/cuda.h --clang-args="-D__CUDA_API_VERSION_INTERNAL" -o $BASE/cuda.py -l /usr/lib/x86_64-linux-gnu/libcuda.so
+  sed -i "s\import ctypes\import ctypes, ctypes.util\g" $BASE/cuda.py
+  sed -i "s\ctypes.CDLL('/usr/lib/x86_64-linux-gnu/libcuda.so')\ctypes.CDLL(ctypes.util.find_library('cuda'))\g" $BASE/cuda.py
+  fixup $BASE/cuda.py
+  python3 -c "import tinygrad.runtime.autogen.cuda"
+}
+
+generate_nvrtc() {
+  clang2py /usr/local/cuda/include/nvrtc.h /usr/local/cuda/include/nvJitLink.h -o $BASE/nvrtc.py -l /usr/local/cuda/lib64/libnvrtc.so -l /usr/local/cuda/lib64/libnvJitLink.so
+  sed -i "s\import ctypes\import ctypes, ctypes.util\g" $BASE/nvrtc.py
+  sed -i "s\ctypes.CDLL('/usr/local/cuda/lib64/libnvrtc.so')\ctypes.CDLL(ctypes.util.find_library('nvrtc'))\g" $BASE/nvrtc.py
+  sed -i "s\ctypes.CDLL('/usr/local/cuda/lib64/libnvJitLink.so')\ctypes.CDLL(ctypes.util.find_library('nvJitLink'))\g" $BASE/nvrtc.py
+  fixup $BASE/nvrtc.py
+  python3 -c "import tinygrad.runtime.autogen.nvrtc"
+}
+
+generate_nv() {
+  NVKERN_COMMIT_HASH=81fe4fb417c8ac3b9bdcc1d56827d116743892a5
+  NVKERN_SRC=/tmp/open-gpu-kernel-modules-$NVKERN_COMMIT_HASH
+  if [ ! -d "$NVKERN_SRC" ]; then
+    git clone https://github.com/NVIDIA/open-gpu-kernel-modules $NVKERN_SRC
+    pushd .
+    cd $NVKERN_SRC
+    git reset --hard $NVKERN_COMMIT_HASH
+    popd
+  fi
+
+  clang2py -k cdefstum \
+    extra/nv_gpu_driver/clc6c0qmd.h \
+    extra/nv_gpu_driver/clcec0qmd.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/cl0000.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/cl0080.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/cl2080.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/cl2080_notification.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/clc56f.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/clc86f.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/clc96f.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/clc761.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/cl83de.h \
+    $NVKERN_SRC/src/nvidia/generated/g_allclasses.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/clc6c0.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/clcdc0.h \
+    $NVKERN_SRC/kernel-open/nvidia-uvm/clc6b5.h \
+    $NVKERN_SRC/kernel-open/nvidia-uvm/clc9b5.h \
+    $NVKERN_SRC/kernel-open/nvidia-uvm/uvm_ioctl.h \
+    $NVKERN_SRC/kernel-open/nvidia-uvm/uvm_linux_ioctl.h \
+    $NVKERN_SRC/kernel-open/nvidia-uvm/hwref/ampere/ga100/dev_fault.h \
+    $NVKERN_SRC/src/nvidia/arch/nvalloc/unix/include/nv_escape.h \
+    $NVKERN_SRC/src/nvidia/arch/nvalloc/unix/include/nv-ioctl.h \
+    $NVKERN_SRC/src/nvidia/arch/nvalloc/unix/include/nv-ioctl-numbers.h \
+    $NVKERN_SRC/src/nvidia/arch/nvalloc/unix/include/nv-ioctl-numa.h \
+    $NVKERN_SRC/src/nvidia/arch/nvalloc/unix/include/nv-unix-nvos-params-wrappers.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/alloc/alloc_channel.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/nvos.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl/ctrl0000/*.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl/ctrl0080/*.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl/ctrl2080/*.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl/ctrl83de/*.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl/ctrlc36f.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl/ctrlcb33.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl/ctrla06c.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl/ctrl90f1.h \
+    --clang-args="-include $NVKERN_SRC/src/common/sdk/nvidia/inc/nvtypes.h -I$NVKERN_SRC/src/common/inc -I$NVKERN_SRC/kernel-open/nvidia-uvm -I$NVKERN_SRC/kernel-open/common/inc -I$NVKERN_SRC/src/common/sdk/nvidia/inc -I$NVKERN_SRC/src/nvidia/arch/nvalloc/unix/include -I$NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl" \
+    -o $BASE/nv_gpu.py
+  fixup $BASE/nv_gpu.py
+  sed -i "s\(0000000001)\1\g" $BASE/nv_gpu.py
+  sed -i "s\import ctypes\import ctypes, os\g" $BASE/nv_gpu.py
+  sed -i 's/#\?\s\([A-Za-z0-9_]\+\) = MW ( \([0-9]\+\) : \([0-9]\+\) )/\1 = (\2 , \3)/' $BASE/nv_gpu.py # NVC6C0_QMDV03_00 processing
+  sed -i 's/#\sdef NVC6C0_QMD\([A-Za-z0-9_()]\+\):/def NVC6C0_QMD\1:/' $BASE/nv_gpu.py
+  sed -i 's/#\sdef NVCEC0_QMD\([A-Za-z0-9_()]\+\):/def NVCEC0_QMD\1:/' $BASE/nv_gpu.py
+  sed -E -i -n '/^def (NVCEC0_QMDV05_00_RELEASE)(_ENABLE)\(i\):/{p;s//\1'"0"'\2=\1\2(0)\n\1'"1"'\2=\1\2(1)/;H;b};p;${x;s/^\n//;p}' "$BASE/nv_gpu.py"
+  sed -i 's/#\s*return MW(\([0-9i()*+]\+\):\([0-9i()*+]\+\))/    return (\1 , \2)/' $BASE/nv_gpu.py
+  sed -i 's/#\?\s*\(.*\)\s*=\s*\(NV\)\?BIT\(32\)\?\s*(\s*\([0-9]\+\)\s*)/\1 = (1 << \4)/' $BASE/nv_gpu.py # name = BIT(x) -> name = (1 << x)
+  sed -i "s/UVM_\([A-Za-z0-9_]\+\) = \['i', '(', '\([0-9]\+\)', ')'\]/UVM_\1 = \2/" $BASE/nv_gpu.py # UVM_name = ['i', '(', '<num>', ')'] -> UVM_name = <num>
+
+  # Parse status codes
+  sed -n '1i\
+nv_status_codes = {}
+/^NV_STATUS_CODE/ { s/^NV_STATUS_CODE(\([^,]*\), *\([^,]*\), *"\([^"]*\)") *.*$/\1 = \2\nnv_status_codes[\1] = "\3"/; p }' $NVKERN_SRC/src/common/sdk/nvidia/inc/nvstatuscodes.h >> $BASE/nv_gpu.py
+  python3 -c "import tinygrad.runtime.autogen.nv_gpu"
+
+  clang2py -k cdefstum \
+    $NVKERN_SRC/src/nvidia/inc/kernel/gpu/fsp/kern_fsp_cot_payload.h \
+    $NVKERN_SRC/src/nvidia/arch/nvalloc/common/inc/gsp/gspifpub.h \
+    $NVKERN_SRC/src/nvidia/arch/nvalloc/common/inc/gsp/gsp_fw_wpr_meta.h \
+    $NVKERN_SRC/src/nvidia/arch/nvalloc/common/inc/gsp/gsp_fw_sr_meta.h \
+    $NVKERN_SRC/src/nvidia/inc/kernel/gpu/gsp/gsp_init_args.h \
+    $NVKERN_SRC/src/nvidia/inc/kernel/gpu/gsp/gsp_init_args.h \
+    $NVKERN_SRC/src/common/uproc/os/common/include/libos_init_args.h \
+    $NVKERN_SRC/src/nvidia/arch/nvalloc/common/inc/rmRiscvUcode.h \
+    $NVKERN_SRC/src/common/shared/msgq/inc/msgq/msgq_priv.h \
+    $NVKERN_SRC/src/nvidia/inc/kernel/vgpu/rpc_headers.h \
+    $NVKERN_SRC/src/nvidia/inc/kernel/vgpu/rpc_global_enums.h \
+    $NVKERN_SRC/src/nvidia/generated/g_rpc-structures.h \
+    $NVKERN_SRC/src/nvidia/arch/nvalloc/common/inc/fsp/fsp_nvdm_format.h \
+    extra/nv_gpu_driver/g_rpc-message-header.h \
+    extra/nv_gpu_driver/gsp_static_config.h \
+    extra/nv_gpu_driver/vbios.h \
+    --clang-args="-DRPC_MESSAGE_STRUCTURES -DRPC_STRUCTURES -include $NVKERN_SRC/src/common/sdk/nvidia/inc/nvtypes.h -I$NVKERN_SRC/src/nvidia/generated -I$NVKERN_SRC/src/common/inc -I$NVKERN_SRC/src/nvidia/inc -I$NVKERN_SRC/src/nvidia/interface/ -I$NVKERN_SRC/src/nvidia/inc/kernel -I$NVKERN_SRC/src/nvidia/inc/libraries -I$NVKERN_SRC/src/nvidia/arch/nvalloc/common/inc -I$NVKERN_SRC/kernel-open/nvidia-uvm -I$NVKERN_SRC/kernel-open/common/inc -I$NVKERN_SRC/src/common/sdk/nvidia/inc -I$NVKERN_SRC/src/nvidia/arch/nvalloc/unix/include -I$NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl" \
+    -o $BASE/nv/nv.py
+
+  fixup $BASE/nv/nv.py
+  python3 -c "import tinygrad.runtime.autogen.nv.nv"
+}
+
+generate_amd() {
+  # clang2py broken when pass -x c++ to prev headers
+  clang2py -k cdefstum \
+    extra/hip_gpu_driver/sdma_registers.h \
+    extra/hip_gpu_driver/nvd.h \
+    extra/hip_gpu_driver/gc_11_0_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
+
+  fixup $BASE/amd_gpu.py
+  sed -i "s\import ctypes\import ctypes, os\g" $BASE/amd_gpu.py
+  python3 -c "import tinygrad.runtime.autogen.amd_gpu"
+}
+
+generate_hsa() {
+  clang2py \
+    /opt/rocm/include/hsa/hsa.h \
+    /opt/rocm/include/hsa/hsa_ext_amd.h \
+    /opt/rocm/include/hsa/amd_hsa_signal.h \
+    /opt/rocm/include/hsa/amd_hsa_queue.h \
+    /opt/rocm/include/hsa/amd_hsa_kernel_code.h \
+    /opt/rocm/include/hsa/hsa_ext_finalize.h /opt/rocm/include/hsa/hsa_ext_image.h \
+    /opt/rocm/include/hsa/hsa_ven_amd_aqlprofile.h \
+    --clang-args="-I/opt/rocm/include" \
+    -o $BASE/hsa.py -l /opt/rocm/lib/libhsa-runtime64.so
+
+  fixup $BASE/hsa.py
+  sed -i "s\import ctypes\import ctypes, ctypes.util, os\g" $BASE/hsa.py
+  sed -i "s\ctypes.CDLL('/opt/rocm/lib/libhsa-runtime64.so')\ctypes.CDLL(os.getenv('ROCM_PATH')+'/lib/libhsa-runtime64.so' if os.getenv('ROCM_PATH') else ctypes.util.find_library('hsa-runtime64'))\g" $BASE/hsa.py
+  python3 -c "import tinygrad.runtime.autogen.hsa"
+}
+
+generate_io_uring() {
+  clang2py -k cdefstum \
+    /usr/include/liburing.h \
+    /usr/include/linux/io_uring.h \
+    -o $BASE/io_uring.py
+
+  sed -r '/^#define __NR_io_uring/ s/^#define __(NR_io_uring[^ ]+) (.*)$/\1 = \2/; t; d' /usr/include/asm-generic/unistd.h >> $BASE/io_uring.py # io_uring syscalls numbers
+  fixup $BASE/io_uring.py
+}
+
+generate_ib() {
+  clang2py -k cdefstum \
+    /usr/include/infiniband/verbs.h \
+    /usr/include/infiniband/verbs_api.h \
+    /usr/include/infiniband/ib_user_ioctl_verbs.h \
+    /usr/include/rdma/ib_user_verbs.h \
+    -o $BASE/ib.py
+
+  sed -i "s\import ctypes\import ctypes, ctypes.util\g" "$BASE/ib.py"
+  sed -i "s\FIXME_STUB\libibverbs\g" "$BASE/ib.py"
+  sed -i "s\FunctionFactoryStub()\ctypes.CDLL(ctypes.util.find_library('ibverbs'), use_errno=True)\g" "$BASE/ib.py"
+
+  fixup $BASE/ib.py
+}
+
+generate_libc() {
+  clang2py -k cdefstum \
+    $(dpkg -L libc6-dev | grep sys/mman.h) \
+    $(dpkg -L libc6-dev | grep sys/syscall.h) \
+    /usr/include/string.h \
+    /usr/include/elf.h \
+    /usr/include/unistd.h \
+    /usr/include/asm-generic/mman-common.h \
+    -o $BASE/libc.py
+
+  sed -i "s\import ctypes\import ctypes, ctypes.util, os\g" $BASE/libc.py
+  sed -i "s\FIXME_STUB\libc\g" $BASE/libc.py
+  sed -i "s\FunctionFactoryStub()\None if (libc_path := ctypes.util.find_library('c')) is None else ctypes.CDLL(libc_path, use_errno=True)\g" $BASE/libc.py
+
+  fixup $BASE/libc.py
+}
+
+generate_llvm() {
+  INC="$(llvm-config-14 --includedir)"
+  clang2py -k cdefstum \
+    $(find "$INC/llvm-c/" -type f -name '*.h' | sort) \
+    "$INC/llvm/Config/Targets.def" \
+    "$INC/llvm/Config/AsmPrinters.def" \
+    "$INC/llvm/Config/AsmParsers.def" \
+    "$INC/llvm/Config/Disassemblers.def" \
+    --clang-args="$(llvm-config-14 --cflags)" \
+    -o "$BASE/llvm.py"
+
+  sed -i "s\import ctypes\import ctypes, tinygrad.runtime.support.llvm as llvm_support\g" "$BASE/llvm.py"
+  sed -i "s\FIXME_STUB\llvm\g" "$BASE/llvm.py"
+  sed -i "s\FunctionFactoryStub()\ctypes.CDLL(llvm_support.LLVM_PATH)\g" "$BASE/llvm.py"
+
+  fixup "$BASE/llvm.py"
+}
+
+generate_kgsl() {
+  clang2py extra/qcom_gpu_driver/msm_kgsl.h -o $BASE/kgsl.py -k cdefstum
+  fixup $BASE/kgsl.py
+  sed -i "s\import ctypes\import ctypes, os\g" $BASE/kgsl.py
+  sed -nE 's/#define ([A-Za-z0-9_]+)_SHIFT\s*[^\S\r\n]*[0-9]*$/def \1(val): return (val << \1_SHIFT) \& \1_MASK/p' extra/qcom_gpu_driver/msm_kgsl.h >> $BASE/kgsl.py
+  sed -i "s\fcntl.ioctl(__fd, (__idir<<30)\__fd.ioctl((__idir<<30)\g" $BASE/kgsl.py
+  python3 -c "import tinygrad.runtime.autogen.kgsl"
+}
+
+generate_adreno() {
+  clang2py extra/qcom_gpu_driver/a6xx.xml.h -o $BASE/adreno.py -k cestum
+  sed -nE 's/#define ([A-Za-z0-9_]+)__SHIFT\s*[^\S\r\n]*[0-9]*$/def \1(val): return (val << \1__SHIFT) \& \1__MASK/p' extra/qcom_gpu_driver/a6xx.xml.h >> $BASE/adreno.py
+  fixup $BASE/adreno.py
+  sed -i "s\import ctypes\import ctypes, os\g" $BASE/adreno.py
+  python3 -c "import tinygrad.runtime.autogen.adreno"
+}
+
+generate_qcom() {
+  clang2py -k cdefstum \
+    extra/dsp/include/ion.h \
+    extra/dsp/include/msm_ion.h \
+    extra/dsp/include/adsprpc_shared.h \
+    extra/dsp/include/remote_default.h \
+    extra/dsp/include/apps_std.h \
+    -o $BASE/qcom_dsp.py
+
+  fixup $BASE/qcom_dsp.py
+  python3 -c "import tinygrad.runtime.autogen.qcom_dsp"
+}
+
+generate_pci() {
+  clang2py -k cdefstum \
+    /usr/include/linux/pci_regs.h \
+    -o $BASE/pci.py
+  fixup $BASE/pci.py
+}
+
+generate_vfio() {
+  clang2py -k cdefstum \
+    /usr/include/linux/vfio.h \
+    -o $BASE/vfio.py
+  fixup $BASE/vfio.py
+  sed -i "s\import ctypes\import ctypes, os\g" $BASE/vfio.py
+  sed -i "s\import fcntl, functools\import functools" $BASE/vfio.py
+  sed -i "s\import ctypes,os\a from tinygrad.runtime.support import FileIOInterface\g" $BASE/vfio.py
+  sed -i "s\fcntl.ioctl(__fd, (__idir<<30)\return __fd.ioctl((__idir<<30)\g" $BASE/vfio.py
+}
+
+generate_am() {
+  AMKERN_COMMIT_HASH=ceb12c04e2b5b53ec0779362831f5ee40c4921e4
+  AMKERN_SRC=/tmp/ROCK-Kernel-Driver-$AMKERN_COMMIT_HASH
+  if [ ! -d "$AMKERN_SRC" ]; then
+    git clone https://github.com/ROCm/ROCK-Kernel-Driver $AMKERN_SRC --depth 1
+  fi
+  AMKERN_AMD=$AMKERN_SRC/drivers/gpu/drm/amd/
+  AMKERN_INC=$AMKERN_AMD/include/
+
+  clang2py -k cdefstum \
+    extra/amdpci/headers/v11_structs.h \
+    extra/amdpci/headers/v12_structs.h \
+    extra/amdpci/headers/amdgpu_vm.h \
+    extra/amdpci/headers/discovery.h \
+    extra/amdpci/headers/amdgpu_ucode.h \
+    extra/amdpci/headers/psp_gfx_if.h \
+    extra/amdpci/headers/amdgpu_psp.h \
+    extra/amdpci/headers/amdgpu_irq.h \
+    extra/amdpci/headers/amdgpu_doorbell.h \
+    $AMKERN_INC/soc15_ih_clientid.h \
+    --clang-args="-include stdint.h" \
+    -o $BASE/am/am.py
+  fixup $BASE/am/am.py
+  sed -i "s\(int64_t)\ \g" $BASE/am/am.py
+  sed -i "s\AMDGPU_PTE_MTYPE_VG10(2)\AMDGPU_PTE_MTYPE_VG10(0, 2)\g" $BASE/am/am.py # incorrect parsing (TODO: remove when clang2py is gone).
+
+  clang2py -k cdefstum \
+    $AMKERN_AMD/amdkfd/kfd_pm4_headers_ai.h \
+    $AMKERN_AMD/amdgpu/soc15d.h \
+    -o $BASE/am/pm4_soc15.py
+  fixup $BASE/am/pm4_soc15.py
+
+  clang2py -k cdefstum \
+    $AMKERN_AMD/amdkfd/kfd_pm4_headers_ai.h \
+    $AMKERN_AMD/amdgpu/nvd.h \
+    -o $BASE/am/pm4_nv.py
+  fixup $BASE/am/pm4_nv.py
+
+  clang2py -k cdefstum \
+    extra/hip_gpu_driver/sdma_registers.h \
+    $AMKERN_AMD/amdgpu/vega10_sdma_pkt_open.h \
+    --clang-args="-I/opt/rocm/include -x c++" \
+    -o $BASE/am/sdma_4_0_0.py
+  fixup $BASE/am/sdma_4_0_0.py
+
+  clang2py -k cdefstum \
+    extra/hip_gpu_driver/sdma_registers.h \
+    $AMKERN_AMD/amdgpu/navi10_sdma_pkt_open.h \
+    --clang-args="-I/opt/rocm/include -x c++" \
+    -o $BASE/am/sdma_5_0_0.py
+  fixup $BASE/am/sdma_5_0_0.py
+
+  clang2py -k cdefstum \
+    extra/hip_gpu_driver/sdma_registers.h \
+    $AMKERN_AMD/amdgpu/sdma_v6_0_0_pkt_open.h \
+    --clang-args="-I/opt/rocm/include -x c++" \
+    -o $BASE/am/sdma_6_0_0.py
+  fixup $BASE/am/sdma_6_0_0.py
+
+  clang2py -k cdefstum \
+    $AMKERN_AMD/pm/swsmu/inc/pmfw_if/smu_v13_0_0_ppsmc.h \
+    $AMKERN_AMD/pm/swsmu/inc/pmfw_if/smu13_driver_if_v13_0_0.h \
+    extra/amdpci/headers/amdgpu_smu.h \
+    -o $BASE/am/smu_v13_0_0.py
+  fixup $BASE/am/smu_v13_0_0.py
+
+  clang2py -k cdefstum \
+    $AMKERN_AMD/pm/swsmu/inc/pmfw_if/smu_v14_0_0_pmfw.h \
+    $AMKERN_AMD/pm/swsmu/inc/pmfw_if/smu_v14_0_2_ppsmc.h \
+    $AMKERN_AMD/pm/swsmu/inc/pmfw_if/smu14_driver_if_v14_0.h \
+    extra/amdpci/headers/amdgpu_smu.h \
+    --clang-args="-include stdint.h" \
+    -o $BASE/am/smu_v14_0_2.py
+  fixup $BASE/am/smu_v14_0_2.py
+}
+
+generate_sqtt() {
+  clang2py -k cdefstum \
+    extra/sqtt/sqtt.h \
+    -o $BASE/sqtt.py
+  fixup $BASE/sqtt.py
+  sed -i "s\import ctypes\import ctypes, os\g" $BASE/sqtt.py
+  python3 -c "import tinygrad.runtime.autogen.sqtt"
+
+  ROCPROF_COMMIT_HASH=dd0485100971522cc4cd8ae136bdda431061a04d
+  ROCPROF_SRC=/tmp/rocprof-trace-decoder-$ROCPROF_COMMIT_HASH
+  if [ ! -d "$ROCPROF_SRC" ]; then
+    git clone https://github.com/ROCm/rocprof-trace-decoder $ROCPROF_SRC
+    pushd .
+    cd $ROCPROF_SRC
+    git reset --hard $ROCPROF_COMMIT_HASH
+    popd
+  fi
+
+  clang2py -k cdefstum \
+    $ROCPROF_SRC/include/rocprof_trace_decoder.h \
+    $ROCPROF_SRC/include/trace_decoder_instrument.h \
+    $ROCPROF_SRC/include/trace_decoder_types.h \
+    -o extra/sqtt/rocprof/rocprof.py
+  fixup extra/sqtt/rocprof/rocprof.py
+  sed -i '1s/^/# pylint: skip-file\n/' extra/sqtt/rocprof/rocprof.py
+  sed -i "s/import ctypes/import ctypes, ctypes.util/g" extra/sqtt/rocprof/rocprof.py
+  sed -i "s|FunctionFactoryStub()|ctypes.CDLL(ctypes.util.find_library('rocprof-trace-decoder'))|g" extra/sqtt/rocprof/rocprof.py
+}
+
+generate_webgpu() {
+  clang2py extra/webgpu/webgpu.h -o $BASE/webgpu.py
+  fixup $BASE/webgpu.py
+  sed -i "s/FIXME_STUB/webgpu/g" "$BASE/webgpu.py"
+  sed -i "s/FunctionFactoryStub()/ctypes.CDLL(webgpu_support.WEBGPU_PATH)/g" "$BASE/webgpu.py"
+  sed -i "s/import ctypes/import ctypes, tinygrad.runtime.support.webgpu as webgpu_support/g" "$BASE/webgpu.py"
+  python3 -c "import tinygrad.runtime.autogen.webgpu"
+}
+
+generate_libusb() {
+  clang2py -k cdefstum \
+    /usr/include/libusb-1.0/libusb.h \
+    -o $BASE/libusb.py
+
+  fixup $BASE/libusb.py
+  sed -i "s\import ctypes\import ctypes, ctypes.util, os\g" $BASE/libusb.py
+  sed -i "s/FIXME_STUB/libusb/g" "$BASE/libusb.py"
+  sed -i "s/libusb_le16_to_cpu = libusb_cpu_to_le16//g" "$BASE/libusb.py"
+  sed -i "s/FunctionFactoryStub()/None if (lib_path:=os.getenv('LIBUSB_PATH', ctypes.util.find_library('usb-1.0'))) is None else ctypes.CDLL(lib_path)/g" "$BASE/libusb.py"
+  python3 -c "import tinygrad.runtime.autogen.libusb"
+}
+
+generate_mesa() {
+  MESA_TAG="mesa-25.2.4"
+  MESA_SRC=/tmp/mesa-$MESA_TAG
+  TINYMESA_TAG=tinymesa-32dc66c
+  TINYMESA_DIR=/tmp/tinymesa-$MESA_TAG-$TINYMESA_TAG/
+  TINYMESA_SO=$TINYMESA_DIR/libtinymesa_cpu.so
+  if [ ! -d "$MESA_SRC" ]; then
+    git clone --depth 1 --branch $MESA_TAG https://gitlab.freedesktop.org/mesa/mesa.git $MESA_SRC
+    pushd .
+    cd $MESA_SRC
+    git reset --hard $MESA_COMMIT_HASH
+    # clang 14 doesn't support packed enums
+    sed -i "s/enum \w\+ \(\w\+\);$/uint8_t \1;/" $MESA_SRC/src/nouveau/headers/nv_device_info.h
+    sed -i "s/enum \w\+ \(\w\+\);$/uint8_t \1;/" $MESA_SRC/src/nouveau/compiler/nak.h
+    sed -i "s/nir_instr_type \(\w\+\);/uint8_t \1;/" $MESA_SRC/src/compiler/nir/nir.h
+    mkdir -p gen/util/format
+    python3 src/util/format/u_format_table.py src/util/format/u_format.yaml --enums > gen/util/format/u_format_gen.h
+    python3 src/compiler/nir/nir_opcodes_h.py > gen/nir_opcodes.h
+    python3 src/compiler/nir/nir_intrinsics_h.py --outdir gen
+    python3 src/compiler/nir/nir_intrinsics_indices_h.py --outdir gen
+    python3 src/compiler/nir/nir_builder_opcodes_h.py > gen/nir_builder_opcodes.h
+    python3 src/compiler/nir/nir_intrinsics_h.py --outdir gen
+    python3 src/compiler/builtin_types_h.py gen/builtin_types.h
+    popd
+  fi
+
+  if [ ! -d "$TINYMESA_DIR" ]; then
+    mkdir $TINYMESA_DIR
+    curl -L https://github.com/sirhcm/tinymesa/releases/download/$TINYMESA_TAG/libtinymesa_cpu-$MESA_TAG-linux-amd64.so -o $TINYMESA_SO
+  fi
+
+  clang2py -k cdefstu \
+    $MESA_SRC/src/compiler/nir/nir.h \
+    $MESA_SRC/src/compiler/nir/nir_builder.h \
+    $MESA_SRC/src/compiler/nir/nir_shader_compiler_options.h \
+    $MESA_SRC/src/compiler/nir/nir_serialize.h \
+    $MESA_SRC/gen/nir_intrinsics.h \
+    $MESA_SRC/src/nouveau/headers/nv_device_info.h \
+    $MESA_SRC/src/nouveau/compiler/nak.h \
+    $MESA_SRC/src/gallium/auxiliary/gallivm/lp_bld.h \
+    $MESA_SRC/src/gallium/auxiliary/gallivm/lp_bld_passmgr.h \
+    $MESA_SRC/src/gallium/auxiliary/gallivm/lp_bld_misc.h \
+    $MESA_SRC/src/gallium/auxiliary/gallivm/lp_bld_type.h \
+    $MESA_SRC/src/gallium/auxiliary/gallivm/lp_bld_init.h \
+    $MESA_SRC/src/gallium/auxiliary/gallivm/lp_bld_nir.h \
+    $MESA_SRC/src/gallium/auxiliary/gallivm/lp_bld_struct.h \
+    $MESA_SRC/src/gallium/auxiliary/gallivm/lp_bld_jit_types.h \
+    $MESA_SRC/src/gallium/auxiliary/gallivm/lp_bld_flow.h \
+    $MESA_SRC/src/gallium/auxiliary/gallivm/lp_bld_const.h \
+    $MESA_SRC/src/compiler/glsl_types.h \
+    $MESA_SRC/src/util/blob.h \
+    $MESA_SRC/src/util/ralloc.h \
+    --clang-args="-DHAVE_ENDIAN_H -DHAVE_STRUCT_TIMESPEC -DHAVE_PTHREAD -I$MESA_SRC/src -I$MESA_SRC/include -I$MESA_SRC/gen -I$MESA_SRC/src/compiler/nir -I$MESA_SRC/src/gallium/auxiliary -I$MESA_SRC/src/gallium/include -I$(llvm-config-20 --includedir)" \
+    -l $TINYMESA_SO \
+    -o $BASE/mesa.py
+
+  LVP_NIR_OPTIONS=$(./extra/mesa/lvp_nir_options.sh $MESA_SRC)
+
+  fixup $BASE/mesa.py
+  patch_dlopen $BASE/mesa.py tinymesa_cpu "(BASE:=os.getenv('MESA_PATH', f\"/usr{'/local/' if helpers.OSX else '/'}lib\"))+'/libtinymesa_cpu'+(EXT:='.dylib' if helpers.OSX else '.so')" "f'{BASE}/libtinymesa{EXT}'" "'/opt/homebrew/lib/libtinymesa_cpu.dylib'" "'/opt/homebrew/lib/libtinymesa.dylib'"
+  echo "lvp_nir_options = gzip.decompress(base64.b64decode('$LVP_NIR_OPTIONS'))" >> $BASE/mesa.py
+  sed -i "/in_dll/s/.*/try: &\nexcept (AttributeError, ValueError): pass/" $BASE/mesa.py
+  sed -i "s/import ctypes/import ctypes, ctypes.util, os, gzip, base64, subprocess, tinygrad.helpers as helpers/" $BASE/mesa.py
+  sed -i "s/ctypes.CDLL('.\+')/(dll := _try_dlopen_tinymesa_cpu())/" $BASE/mesa.py
+  echo "def __getattr__(nm): raise AttributeError('LLVMpipe requires tinymesa_cpu' if 'tinymesa_cpu' not in dll._name else f'attribute {nm} not found') if dll else FileNotFoundError(f'libtinymesa not found (MESA_PATH={BASE}). See https://github.com/sirhcm/tinymesa ($TINYMESA_TAG, $MESA_TAG)')" >> $BASE/mesa.py
+  sed -i "s/ctypes.glsl_base_type/glsl_base_type/" $BASE/mesa.py
+  # bitfield bug in clang2py
+  sed -i "s/('fp_fast_math', ctypes.c_bool, 9)/('fp_fast_math', ctypes.c_uint32, 9)/" $BASE/mesa.py
+  sed -i "s/('\(\w\+\)', pipe_shader_type, 8)/('\1', ctypes.c_ubyte)/" $BASE/mesa.py
+  sed -i "s/\([0-9]\+\)()/\1/" $BASE/mesa.py
+  sed -i '/struct_nir_builder._pack_ = 1 # source:False/d' "$BASE/mesa.py"
+  python3 -c "import tinygrad.runtime.autogen.mesa"
+}
+
+if [ "$1" == "opencl" ]; then generate_opencl
+elif [ "$1" == "hip" ]; then generate_hip
+elif [ "$1" == "comgr" ]; then generate_comgr
+elif [ "$1" == "cuda" ]; then generate_cuda
+elif [ "$1" == "nvrtc" ]; then generate_nvrtc
+elif [ "$1" == "hsa" ]; then generate_hsa
+elif [ "$1" == "kfd" ]; then generate_kfd
+elif [ "$1" == "nv" ]; then generate_nv
+elif [ "$1" == "amd" ]; then generate_amd
+elif [ "$1" == "am" ]; then generate_am
+elif [ "$1" == "nvdrv" ]; then generate_nvdrv
+elif [ "$1" == "sqtt" ]; then generate_sqtt
+elif [ "$1" == "qcom" ]; then generate_qcom
+elif [ "$1" == "io_uring" ]; then generate_io_uring
+elif [ "$1" == "ib" ]; then generate_ib
+elif [ "$1" == "libc" ]; then generate_libc
+elif [ "$1" == "llvm" ]; then generate_llvm
+elif [ "$1" == "kgsl" ]; then generate_kgsl
+elif [ "$1" == "adreno" ]; then generate_adreno
+elif [ "$1" == "pci" ]; then generate_pci
+elif [ "$1" == "vfio" ]; then generate_vfio
+elif [ "$1" == "webgpu" ]; then generate_webgpu
+elif [ "$1" == "libusb" ]; then generate_libusb
+elif [ "$1" == "mesa" ]; then generate_mesa
+elif [ "$1" == "all" ]; then generate_opencl; generate_hip; generate_comgr; generate_cuda; generate_nvrtc; generate_hsa; generate_kfd; generate_nv; generate_amd; generate_io_uring; generate_libc; generate_am; generate_webgpu; generate_mesa
+else echo "usage: $0 <type>"
+fi
@@ -15,7 +15,7 @@ export IGNORE_JIT_FIRST_BEAM=1
 export BASEDIR="/raid/datasets/wiki"

 # pip install -e ".[mlperf]"
-export LOGMLPERF=${LOGMLPERF:-1}
+export LOGMLPERF=1

 export SEED=$RANDOM
 DATETIME=$(date "+%m%d%H%M")
@@ -1,10 +1,11 @@
 import os, sys, pickle, time, re
 import numpy as np
-if "JIT_BATCH_SIZE" not in os.environ: os.environ["JIT_BATCH_SIZE"] = "0"

 from tinygrad import fetch, Tensor, TinyJit, Context, GlobalCounters, Device, dtypes
 from tinygrad.helpers import DEBUG, getenv
 from tinygrad.engine.realize import CompiledRunner
+
+import onnx
 from tinygrad.nn.onnx import OnnxRunner

 OPENPILOT_MODEL = sys.argv[1] if len(sys.argv) > 1 else "https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/supercombo.onnx"
@@ -39,7 +40,7 @@ def compile(onnx_file):
  np.testing.assert_equal(test_val, ret, "JIT run failed")
  print("jit run validated")

-  # check gated read_image usage
+  # checks from compile2
  kernel_count = 0
  read_image_count = 0
  gated_read_image_count = 0
@@ -95,7 +96,6 @@ def test_vs_compile(run, inputs, test_val=None):
  return val

 def test_vs_onnx(new_inputs, test_val, onnx_file, tol):
-  import onnx
  import onnxruntime as ort
  
  onnx_inputs = {k:v.numpy() for k,v in new_inputs.items()}
@@ -2,8 +2,7 @@ import os, pathlib, argparse
 from examples.llama3 import Tokenizer
 from tabulate import tabulate
 from tinygrad import fetch
-from tinygrad.helpers import flatten, getenv
-from sz import NONCORE_DIRS
+from tinygrad.helpers import flatten

 # llama 3 tokenizer
 tokenizer = Tokenizer(fetch("https://huggingface.co/bofenghuang/Meta-Llama-3-8B/resolve/main/original/tokenizer.model").as_posix())
@@ -11,15 +10,19 @@ tokenizer = Tokenizer(fetch("https://huggingface.co/bofenghuang/Meta-Llama-3-8B/
 def read_code(base_path):
  ret = []
  for path, _, files in os.walk(os.path.join(base_path, "tinygrad")):
-    if not getenv("CORE") and any(path.split("./")[1].startswith(x) for x in NONCORE_DIRS): continue
    for name in files:
      if not name.endswith(".py"): continue
      if 'tinygrad/runtime/autogen' in path.replace('\\', '/'): continue
      fullpath = os.path.join(path, name)
      code = pathlib.Path(fullpath).read_text()
-      ret.append((fullpath.split("tinygrad/", 1)[1], code))
+      ret.append(("### " + fullpath.split("tinygrad/", 1)[1], code))
  return ret

+def write_code_to_file(filename, code_list):
+  """Writes the combined code to a specified file."""
+  with open(filename, 'w') as f:
+    f.write('\n'.join(flatten(code_list)))
+
 if __name__ == "__main__":
  parser = argparse.ArgumentParser(description="Analyze and optionally save tinygrad code.")
  parser.add_argument("--output", help="Output file to write the combined code to.")
@@ -29,11 +32,10 @@ if __name__ == "__main__":

  table = []
  for name,code in ret:
-    table.append([name, len(tokenizer.encode(code))])
+    table.append([name, len(tokenizer.encode(name+"\x00"+code))])
  print(tabulate([["name", "llm tokens"]]+sorted(table, key=lambda x: -x[1]), headers="firstrow"))

-  banner = "#"*40
-  code_str = ''.join([f"{banner}\n# {name}\n{banner}\n\n{code}\n" for name,code in ret])
+  code_str = '\x00'.join(flatten(ret))
  print(f"code has {len(code_str)} chars")
  newline_count = code_str.count('\n')
  print(f"code has {newline_count} newlines")
@@ -42,5 +44,5 @@ if __name__ == "__main__":
  print(f"code has {len(encoded)} tokens")

  if args.output:
-    with open(args.output, 'w') as f: f.write(code_str)
-    print(f"Combined code written to {args.output}")
+    write_code_to_file(args.output, ret)
+    print(f"Combined code written to {args.output}")
@@ -3,7 +3,7 @@
 import sys, base64, multiprocessing, itertools, collections
 from typing import Optional, Union, Literal, List

-from tinygrad import Tensor, TinyJit, Variable, nn, dtypes
+from tinygrad import Tensor, TinyJit, Variable, nn
 from tinygrad.nn.state import torch_load, load_state_dict
 from tinygrad.helpers import getenv, fetch

@@ -244,16 +244,15 @@ def transcribe_waveform(model: Whisper, enc, waveforms, truncate=False):

  log_spec = prep_audio(waveforms, model.batch_size, truncate)
  nsample = model.decoder.max_tokens_to_sample
-  nctx = model.decoder.max_self_attn_cache_len

  def inferloop(ctx: Union[np.ndarray, List[np.ndarray]], encoded_audio):
    pos, next_tokens = 0, ctx
-    for i in range(nsample):
-      next_tokens = model.decoder(Tensor(next_tokens, dtype=dtypes.int32), pos, encoded_audio)[:, -1].argmax(axis=-1).numpy().astype(np.int32).reshape(-1, 1)
+    for i in range((nsample-len(start_tokens))*2):
+      next_tokens = model.decoder(Tensor(next_tokens), pos, encoded_audio)[:, -1].argmax(axis=-1).numpy().astype(np.int32).reshape(-1, 1)
      next_tokens[ctx[:, -1] == eot] = eot
      ctx = np.concatenate((ctx, next_tokens), axis=1)
      pos = ctx.shape[-1] - 1
-      if (next_tokens == eot).all() or pos == nctx: break
+      if (next_tokens == eot).all(): break
    return ctx

  def gettexttoks(line): return [tok for tok in line if tok < eot or tok > enc._special_tokens["<|notimestamps|>"]][-nsample+len(start_tokens):]
@@ -1,34 +0,0 @@
-#!/usr/bin/env python3
-from tinygrad import Tensor, Device, GlobalCounters, Context, dtypes
-from tinygrad.helpers import getenv, colored
-
-SZ = 8_000_000_000
-GPUS = getenv("GPUS", 4) # TODO: expose a way in tinygrad to access this
-
-if __name__ == "__main__":
-  # create tensors
-  tens = [Tensor.ones(SZ, dtype=dtypes.uint8, device=f"{Device.DEFAULT}:{i}").contiguous() for i in range(GPUS)]
-  Tensor.realize(*tens)
-
-  bw = [[0.0]*GPUS for _ in range(GPUS)]
-  for i in range(GPUS):
-    for j in range(GPUS):
-      GlobalCounters.reset()
-      with Context(DEBUG=2):
-        if i == j:
-          # this copy would be optimized out, just add 1
-          (tens[i]+1).realize()
-        else:
-          tens[i].to(f"{Device.DEFAULT}:{j}").realize()
-      t = max(GlobalCounters.time_sum_s, 1e-9)
-      bw[i][j] = SZ / t / 1e9  # GB/s
-
-  def fmt(x):
-    c = "green" if x > 50 else "yellow" if x > 20 else "red"
-    return colored(f"{x:6.1f}", c)
-
-  # header
-  print(" " * 8 + " ".join(f"{'d'+str(j):>6}" for j in range(GPUS)))
-  # rows
-  for i in range(GPUS):
-    print(f"{'s'+str(i):>6} -> " + " ".join(fmt(x) for x in bw[i]))
@@ -1,11 +1,10 @@
 from tinygrad import Tensor, Device, Context, GlobalCounters, dtypes
-from tinygrad.uop.ops import UOp, KernelInfo, sint, AxisType
+from tinygrad.uop.ops import UOp, KernelInfo
 from tinygrad.engine.realize import ExecItem, get_runner
 from tinygrad.dtype import AddrSpace
 from tinygrad.helpers import getenv

 N = 4096
-M = K = N
 run_count = 5

 # ---------------------------
@@ -43,17 +42,26 @@ LANES_PER_WAVE_X = 8
 LANES_PER_WAVE_Y = 4
 ITERS_PER_WAVE_N = WAVE_TILE_N // (LANES_PER_WAVE_X * TN)
 ITERS_PER_WAVE_M = WAVE_TILE_M // (LANES_PER_WAVE_Y * TM)
+N_PER_ITER = WAVE_TILE_N // ITERS_PER_WAVE_N
+M_PER_ITER = WAVE_TILE_M // ITERS_PER_WAVE_M
 assert WAVE_TILE_N % (LANES_PER_WAVE_X * TN) == 0, "WAVE_TILE_N must be divisible by LANES_PER_WAVE_X*TN"
 assert WAVE_TILE_M % (LANES_PER_WAVE_Y * TM) == 0, "WAVE_TILE_M must be divisible by LANES_PER_WAVE_Y*TM"

-def rngs_for_shape(shape:tuple[sint, ...], rng:int, axis_type=AxisType.LOOP): return [UOp.range(s, rng+i, axis_type) for i,s in enumerate(shape)]
-def copy(dest:UOp, src:UOp, rng:int, set=False, upcast=False):
-  assert dest.shape == src.shape
-  rngs = rngs_for_shape(src.shape, rng, AxisType.UPCAST if upcast else AxisType.LOOP)
-  copy = dest[*rngs].store(src[*rngs]).end(*rngs)
-  return dest.after(copy) if set else copy
-
 def hand_spec_kernel3():
+  # ---------------------------
+  # per-thread read mapping
+  # ---------------------------
+  # A: read BK x BN tiles; B: read BN x BK tiles
+  tid = UOp.special(THREADS_PER_BLOCK, "lidx0")
+
+  waveIdx = (tid // WARP_SIZE) % WAVES_IN_BLOCK_X
+  waveIdy = (tid // WARP_SIZE) // WAVES_IN_BLOCK_X
+  assert waveIdy.vmax+1 == WAVES_IN_BLOCK_Y
+
+  idxInWave = (tid % WARP_SIZE) % LANES_PER_WAVE_X
+  idyInWave = (tid % WARP_SIZE) // LANES_PER_WAVE_X
+  assert idyInWave.vmax+1 == LANES_PER_WAVE_Y
+
  # ---------------------------
  # block indices & placeholders
  # ---------------------------
@@ -64,66 +72,66 @@ def hand_spec_kernel3():
  b = UOp.placeholder((N, N), dtypes.float, slot=2)
  c = UOp.placeholder((N, N), dtypes.float, slot=0)

-  # index the output with the globals
-  c = c.reshape(M // BLOCK_M, BLOCK_M, N // BLOCK_N, BLOCK_N)[blockIdx_y, :, blockIdx_x, :]
+  BM_As_stride = (BLOCK_M + 4) if is_kernel5 else BLOCK_M
+  As = UOp.placeholder((BLOCK_K, BM_As_stride), dtypes.float, slot=0, addrspace=AddrSpace.LOCAL).shrink_to((BLOCK_K, BLOCK_M))
+  Bs = UOp.placeholder((BLOCK_K, BLOCK_N), dtypes.float, slot=1, addrspace=AddrSpace.LOCAL)

-  # open the main reduction range
-  k_tile_range = UOp.range(N // BLOCK_K, 0, AxisType.REDUCE)
-  a = a.reshape(M // BLOCK_M, BLOCK_M, N // BLOCK_K, BLOCK_K)[blockIdx_y, :, k_tile_range, :]
-  b = b.reshape(N // BLOCK_K, BLOCK_K, N // BLOCK_N, BLOCK_N)[k_tile_range, :, blockIdx_x, :]
+  A_col = UOp.placeholder((ITERS_PER_WAVE_M, TM), dtypes.float, slot=0, addrspace=AddrSpace.REG)
+  B_row = UOp.placeholder((ITERS_PER_WAVE_N, TN), dtypes.float, slot=1, addrspace=AddrSpace.REG)
+  c_regs = UOp.placeholder((ITERS_PER_WAVE_M, TM, ITERS_PER_WAVE_N, TN), dtypes.float, slot=2, addrspace=AddrSpace.REG)

-  # globals are no longer used, they are already in the indexes
-  del blockIdx_y, blockIdx_x
+  i = UOp.range(c_regs.size, 16)
+  c_regs = c_regs[i].set(0.0, end=i)
+
+  k_tile_range = UOp.range(N // BLOCK_K, 0)

  # ---------------------------
  # GLOBAL -> LOCAL (As, Bs)
  # ---------------------------
-  tid = UOp.special(THREADS_PER_BLOCK, "lidx0")
+  b = b.reshape(N // BLOCK_K, BLOCK_K,
+                N // BLOCK_N, BLOCK_N)
+  i = UOp.range(BLOCK_N * BLOCK_K // THREADS_PER_BLOCK, 1)
+  index_x = tid % BLOCK_N
+  index_y = (tid // BLOCK_N) + (THREADS_PER_BLOCK // BLOCK_N) * i
+  Bs_store = Bs[index_y, index_x].store(b[k_tile_range, index_y, blockIdx_x, index_x]).end(i)

-  # A: read BM x BK tiles (permute on store into locals)
-  BM_As_stride = (BLOCK_M + 4) if is_kernel5 else BLOCK_M
-  As = UOp.placeholder((BLOCK_K, BM_As_stride), dtypes.float, slot=0, addrspace=AddrSpace.LOCAL).shrink_to((BLOCK_K, BLOCK_M))
-  As_store = copy(As.permute((1,0)).reshape(-1, THREADS_PER_BLOCK)[:, tid], a.reshape(-1, THREADS_PER_BLOCK)[:, tid], rng=100)
-
-  # B: read BK x BN tiles
-  Bs = UOp.placeholder((BLOCK_K, BLOCK_N), dtypes.float, slot=1, addrspace=AddrSpace.LOCAL)
-  Bs_store = copy(Bs.reshape(-1, THREADS_PER_BLOCK)[:, tid], b.reshape(-1, THREADS_PER_BLOCK)[:, tid], rng=200)
+  a = a.reshape(N // BLOCK_M, BLOCK_M,
+                N // BLOCK_K, BLOCK_K)
+  i = UOp.range(BLOCK_M * BLOCK_K // THREADS_PER_BLOCK, 2)
+  index_x = tid % BLOCK_K
+  index_y = (tid // BLOCK_K) + (THREADS_PER_BLOCK // BLOCK_K) * i
+  As_store = As[index_x, index_y].store(a[blockIdx_y, index_y, k_tile_range, index_x]).end(i)

  # TODO: can we automate barrier?
  barrier = UOp.barrier(As_store, Bs_store)
-  As, Bs = As.after(barrier), Bs.after(barrier)
+  Bs = Bs.after(barrier)
+  As = As.after(barrier)

  # open inner k range
-  k = UOp.range(BLOCK_K, 3, AxisType.REDUCE)
+  k = UOp.range(BLOCK_K, 3)

  # ---------------------------
  # LOCAL -> REG (per-wave tiles)
  # ---------------------------
-  waveIdx = (tid // WARP_SIZE) % WAVES_IN_BLOCK_X
-  waveIdy = (tid // WARP_SIZE) // WAVES_IN_BLOCK_X
-  assert waveIdy.vmax+1 == WAVES_IN_BLOCK_Y
+  Bs_view = Bs.reshape(BLOCK_K, WAVES_IN_BLOCK_X, ITERS_PER_WAVE_N, LANES_PER_WAVE_X, TN)
+  iterWaveN = UOp.range(ITERS_PER_WAVE_N, 4)
+  i = UOp.range(TN, 5)
+  B_row = B_row[iterWaveN, i].set(Bs_view[k, waveIdx, iterWaveN, idxInWave, i], end=(iterWaveN, i))

-  laneIdx = (tid % WARP_SIZE) % LANES_PER_WAVE_X
-  laneIdy = (tid % WARP_SIZE) // LANES_PER_WAVE_X
-  assert laneIdy.vmax+1 == LANES_PER_WAVE_Y
-
-  A_col = UOp.placeholder((ITERS_PER_WAVE_M, TM), dtypes.float, slot=0, addrspace=AddrSpace.REG)
-  A_col = copy(A_col, As[k, :].reshape(WAVES_IN_BLOCK_Y, ITERS_PER_WAVE_M, LANES_PER_WAVE_Y, TM)[waveIdy, :, laneIdy, :], 300, set=True, upcast=True)
-
-  B_row = UOp.placeholder((ITERS_PER_WAVE_N, TN), dtypes.float, slot=1, addrspace=AddrSpace.REG)
-  B_row = copy(B_row, Bs[k, :].reshape(WAVES_IN_BLOCK_X, ITERS_PER_WAVE_N, LANES_PER_WAVE_X, TN)[waveIdx, :, laneIdx, :], 400, set=True, upcast=True)
+  As_view = As.reshape(BLOCK_K, WAVES_IN_BLOCK_Y, ITERS_PER_WAVE_M, LANES_PER_WAVE_Y, TM)
+  iterWaveM = UOp.range(ITERS_PER_WAVE_M, 6)
+  i = UOp.range(TM, 7)
+  A_col = A_col[iterWaveM, i].set(As_view[k, waveIdy, iterWaveM, idyInWave, i], end=(iterWaveM, i))

  # ---------------------------
  # FMA: c_regs += A_col * B_row
  # ---------------------------
-  c_regs = UOp.placeholder((ITERS_PER_WAVE_M, TM, ITERS_PER_WAVE_N, TN), dtypes.float, slot=2, addrspace=AddrSpace.REG)
-  i = UOp.range(c_regs.size, 16)
-  c_regs = c_regs.after(c_regs.flatten()[i].store(0.0).end(i))
-
-  # TODO: why don't these work as upcast?
-  # why if the ranges merge is it slow?!? (if you change the order on end, they will merge. big slowdown on METAL)
-  iterWaveM, yt, iterWaveN, xt = rngs = rngs_for_shape(c_regs.shape, 500)
-  sink = c_regs[*rngs].store(c_regs.after(k)[*rngs] + A_col[iterWaveM, yt] * B_row[iterWaveN, xt]).end(iterWaveM, iterWaveN, yt, xt)
+  iterWaveM = UOp.range(ITERS_PER_WAVE_M, 8)
+  yt = UOp.range(TM, 9)
+  iterWaveN = UOp.range(ITERS_PER_WAVE_N, 10)
+  xt = UOp.range(TN, 12)
+  c_idx = c_regs.after(k, k_tile_range)[iterWaveM, yt, iterWaveN, xt]
+  sink = c_idx.store(c_idx + A_col[iterWaveM, yt] * B_row[iterWaveN, xt]).end(iterWaveM, iterWaveN, yt, xt)

  # Close k, sync, and close K tiles
  sink = sink.end(k).barrier().end(k_tile_range)
@@ -131,11 +139,15 @@ def hand_spec_kernel3():
  # ---------------------------
  # REG -> GLOBAL (epilogue)
  # ---------------------------
-  c = c.reshape(WAVES_IN_BLOCK_Y, ITERS_PER_WAVE_M, LANES_PER_WAVE_Y, TM,
-                WAVES_IN_BLOCK_X, ITERS_PER_WAVE_N, LANES_PER_WAVE_X, TN)
-  c = c[waveIdy, :, laneIdy, :,
-        waveIdx, :, laneIdx, :]
-  sink = copy(c, c_regs.after(sink), rng=600)
+  c = c.reshape(N//BLOCK_M, WAVES_IN_BLOCK_Y, ITERS_PER_WAVE_M, LANES_PER_WAVE_Y, TM,
+                N//BLOCK_N, WAVES_IN_BLOCK_X, ITERS_PER_WAVE_N, LANES_PER_WAVE_X, TN)
+  iterWaveM = UOp.range(ITERS_PER_WAVE_M, 1000)
+  yt = UOp.range(TM, 1001)
+  iterWaveN = UOp.range(ITERS_PER_WAVE_N, 1002)
+  xt = UOp.range(TN, 1003)
+  c_glbl_idx = c[blockIdx_y, waveIdy, iterWaveM, idyInWave, yt, blockIdx_x, waveIdx, iterWaveN, idxInWave, xt]
+  sink = c_glbl_idx.store(c_regs.after(sink)[iterWaveM, yt, iterWaveN, xt])
+  sink = sink.end(iterWaveM, iterWaveN, yt, xt)

  return sink.sink(arg=KernelInfo(opts_to_apply=())).simplify()

@@ -1,11 +1,17 @@
 import numpy as np, os
 from tinygrad.helpers import getenv, flat_mv
 from tinygrad import dtypes
+from typing import Optional, List, Tuple, cast, Dict, Final, DefaultDict, Self
 from tinygrad.engine.realize import get_program

 # for copied uops
-from tinygrad import dtypes
-from tinygrad.dtype import DTYPES_DICT
+from tinygrad.codegen.opt.kernel import Kernel, KernelOptError
+from tinygrad.uop.ops import UOp, Ops, BinaryOps, UnaryOps, TernaryOps, KernelInfo
+from tinygrad.codegen.opt.search import Opt, OptOps
+from tinygrad import Device, dtypes, Tensor
+from tinygrad.dtype import PtrDType, DType, DTYPES_DICT
+from tinygrad.shape.shapetracker import ShapeTracker
+from tinygrad.shape.view import View

 script_dir = os.path.dirname(os.path.abspath(__file__))

@@ -47,6 +53,12 @@ def randoms():
    nc = nc.astype(np.bfloat16 if DTYPE_IN == dtypes.bfloat16 else np.float16)
  return na, nb, nc

+def ast_to_cuda_prog(compiler, ast, opts):
+  k = Kernel(ast)
+  k.apply_opts(opts)
+  p = get_program(k.ast, k.opts, k.applied_opts)
+  return CUDAProgram(device, p.function_name, compiler.compile(p.src))
+
 if __name__ == "__main__":
  print(f"gemm variation: {GEMM_VARIATION=} {M=} {N=} {K=} {DTYPE_IN=} {DTYPE_OUT=} {DTYPE_ACC=}")
  prog, global_size, local_size = None, None, None
@@ -177,11 +189,11 @@ if __name__ == "__main__":

    tms = []
    na, nb, nc = randoms()
-    cudaalloc._copyin(a, memoryview(bytearray(na)))
-    cudaalloc._copyin(b, memoryview(bytearray(nb)))
+    cudaalloc.copyin(a, bytearray(na))
+    cudaalloc.copyin(b, bytearray(nb))
    for i in range(CNT):
      tms.append(prog(*args, **kwargs))
-    cudaalloc._copyout(flat_mv(nc.data), c)
+    cudaalloc.copyout(flat_mv(nc.data), c)
    comp = na.astype(np.float32) @ nb.astype(np.float32)
    result = nc.reshape(M, N).astype(np.float32)

@@ -37,11 +37,11 @@ WARPGROUP_SIZE = 1
 BLOCK_M = BLOCK_M * WARPGROUP_SIZE

 # TODO: improve the syntax of this. better syntax, faster iteration
-#  -- DONE: add working slice a[gx, :, i] -> shape of the : (aka (16,16,32) becomes (16,))
-#  -- DONE(ish): add argfix to movement (traits shared with Tensor)
+#  -- add working slice a[gx, :, i] -> shape of the : (aka (16,16,32) becomes (16,))
+#  -- add argfix to movement (traits shared with Tensor)
 #  -- fix WMMA to not require all the junk
 #  -- improve syntax for vectorized loads/stores (both with DEVECTORIZE and without)
-#  -- DONE: be able to use CONTRACT on a range
+#  -- be able to use CONTRACT on a range
 #  -- fix upcasted RANGE on an already vectorized buffer
 #  -- improve "all ranges not ended error" / fix the bug with after on ended ranges (if you are after end of range, range is closed)

@@ -1,141 +0,0 @@
-import os
-import numpy as np
-np.set_printoptions(linewidth=1000000)
-os.environ["AMD_LLVM"] = "0"
-
-from tinygrad import Tensor, Context, dtypes, UOp, GlobalCounters
-from tinygrad.helpers import DEBUG, getenv
-from tinygrad.dtype import AddrSpace
-from tinygrad.uop.ops import sint, AxisType, KernelInfo, Ops
-
-WARP_SIZE = 64
-
-# Reg tile sizes (tensor cores)
-TC_M = 16
-TC_N = 16
-TC_K = 32
-
-N,M,K = 4096,4096,4096
-
-# Threadblock tile sizes (block-level tile of C that a block computes)
-BLOCK_M = 64
-BLOCK_N = 64
-BLOCK_K = 64
-
-WARPGROUP_SIZE = 1
-BLOCK_M = BLOCK_M * WARPGROUP_SIZE
-
-TID_SIZE = WARPGROUP_SIZE*WARP_SIZE
-
-def copy(dest:UOp, src:UOp, rng:int, set=False, upcast=()):
-  assert dest.shape == src.shape
-  rngs = [UOp.range(s, rng+i, AxisType.UPCAST if i in upcast else AxisType.LOOP) for i,s in enumerate(src.shape)]
-  copy = dest[*rngs].store(src[*rngs]).end(*rngs)
-  return dest.after(copy) if set else copy
-
-def compute_on_locals(acc:UOp, Asl:UOp, Bsl:UOp, rng:int, afters:tuple[UOp, ...], warpgroup, warp) -> UOp:
-  K_inner_loop = UOp.range(BLOCK_K//TC_K, rng, AxisType.REDUCE)
-
-  # load from locals into registers
-  Ar = UOp.placeholder((BLOCK_M//TC_M//WARPGROUP_SIZE,), dtypes.half.vec(8), slot=1, addrspace=AddrSpace.REG)
-  Br = UOp.placeholder((BLOCK_N//TC_N,), dtypes.half.vec(8), slot=2, addrspace=AddrSpace.REG)
-
-  M_load_loop = UOp.range(BLOCK_M//TC_M//WARPGROUP_SIZE, rng+10)
-  Asl = Asl.reshape(BLOCK_K//TC_K, TC_K, BLOCK_M//TC_M//WARPGROUP_SIZE, WARPGROUP_SIZE, TC_M)
-  load_rng = UOp.range(8, rng+11, axis_type=AxisType.UPCAST)
-  A_in = Asl[K_inner_loop, (warp//16)*8+load_rng, M_load_loop, warpgroup, warp%16].contract(load_rng)
-  Ar = Ar[M_load_loop].set(A_in, end=M_load_loop)
-
-  N_load_loop = UOp.range(BLOCK_N//TC_N, rng+20)
-  Bsl = Bsl.reshape(BLOCK_K//TC_K, TC_K, BLOCK_N//TC_N, TC_N)
-  load_rng = UOp.range(8, rng+21, axis_type=AxisType.UPCAST)
-  B_in = Bsl[K_inner_loop, (warp//16)*8+load_rng, N_load_loop, warp%16].contract(load_rng)
-  Br = Br[N_load_loop].set(B_in, end=N_load_loop)
-
-  M_inner_loop = UOp.range(BLOCK_M//TC_M//WARPGROUP_SIZE, rng+30)
-  N_inner_loop = UOp.range(BLOCK_N//TC_N, rng+31)
-
-  # load values
-  acc_after = acc.after(*afters, M_inner_loop, N_inner_loop, K_inner_loop)
-  acc_load = acc_after[N_inner_loop, M_inner_loop]
-
-  # do WMMA
-  wmma_arg = ('WMMA_16_16_32_half_float', (16, 16, 32), dtypes.half, dtypes.float, 'AMD', 64, ((), (), ((3, 2), (2, 2))), ())
-  out = UOp(Ops.WMMA, dtypes.float.vec(4), (Ar[M_inner_loop], Br[N_inner_loop], acc_load), arg=wmma_arg)
-
-  # store back the acc
-  acc_store = acc[N_inner_loop, M_inner_loop].store(out)
-  return acc_store.end(M_inner_loop, N_inner_loop, K_inner_loop)
-
-def custom_gemm(C:UOp, A:UOp, B:UOp) -> UOp:
-  gx, gy = UOp.special(M//BLOCK_M, "gidx0"), UOp.special(N//BLOCK_N, "gidx1")
-  K_outer_loop = UOp.range(K//BLOCK_K, 0, AxisType.REDUCE)
-
-  # split out the globals into blocks
-  C = C.src[0].cast(dtypes.float.vec(4).ptr(C.ptrdtype.size)).reshape((M//BLOCK_M, BLOCK_M, N//BLOCK_N, BLOCK_N))
-  A = A.reshape((M//BLOCK_M, BLOCK_M, K//BLOCK_K, BLOCK_K))[gx, :, K_outer_loop, :]
-  B = B.reshape((K//BLOCK_K, BLOCK_K, N//BLOCK_N, BLOCK_N))[K_outer_loop, :, gy, :]
-
-  # ---------------------------
-  # GLOBAL -> LOCAL (As, Bs)
-  # ---------------------------
-  tid = UOp.special(TID_SIZE, "lidx0")
-  warpgroup, warp = tid//WARP_SIZE, tid%WARP_SIZE
-
-  A_view = A.reshape(-1, TID_SIZE, 8)
-  B_view = B.reshape(-1, TID_SIZE, 8)
-
-  # A: read BM x BK tiles (permute on store into locals)
-  As = UOp.placeholder((BLOCK_K, BLOCK_M), dtypes.half, slot=0, addrspace=AddrSpace.LOCAL).shrink_to(BLOCK_K, BLOCK_M)
-  As_view = As.reshape(-1, TID_SIZE, 8)
-
-  Bs = UOp.placeholder((BLOCK_K, BLOCK_N+4), dtypes.half, slot=1, addrspace=AddrSpace.LOCAL).shrink_to(BLOCK_K, BLOCK_N)
-  Bs_view = Bs.reshape(-1, TID_SIZE, 8)
-
-  outer_copy = UOp.range(A_view.shape[0], 100, AxisType.UPCAST)
-  inner_copy = UOp.range(A_view.shape[2], 101, AxisType.UPCAST)
-  As_store = As_view[outer_copy, tid, inner_copy].store(A_view[outer_copy, tid, inner_copy])
-  Bs_store = Bs_view[outer_copy, tid, inner_copy].store(B_view[outer_copy, tid, inner_copy])
-
-  if getenv("NOLOAD"):
-    As_store = As[0,0].store(0)
-    Bs_store = Bs[0,0].store(0)
-
-  # TODO: can we automate barrier?
-  barrier = UOp.barrier(UOp.group(As_store, Bs_store).end(outer_copy, inner_copy))
-
-  if getenv("COMPUTE"):
-    As, Bs = As.after(barrier), Bs.after(barrier)
-
-    acc = UOp.placeholder((BLOCK_N//TC_N, BLOCK_M//TC_M//WARPGROUP_SIZE), dtypes.float.vec(4), 0, AddrSpace.REG)
-
-    sink = compute_on_locals(acc, As, Bs, 200, afters=(barrier,), warpgroup=warpgroup, warp=warp)
-    sink = sink.end(K_outer_loop)
-
-    C_view = C[gx, :, gy, :].reshape(BLOCK_M//TC_M//WARPGROUP_SIZE, WARPGROUP_SIZE, TC_M, BLOCK_N//TC_N, TC_N)[:, warpgroup, warp%16, :, (warp//16)*4]
-    sink = copy(C_view, acc.after(sink), rng=300)
-  else:
-    sink = C.after(barrier.end(K_outer_loop))[0,0,0,0].store(As[0,0]+Bs[0,0])
-
-  return sink.sink(arg=KernelInfo(name="custom_gemm", opts_to_apply=())).simplify()
-
-if __name__ == "__main__":
-  a = Tensor.randn(M, K, dtype=dtypes.half)
-  b = Tensor.randn(K, N, dtype=dtypes.half)
-  c = Tensor.empty(M, N, dtype=dtypes.float)
-  with Context(DEBUG=0): Tensor.realize(a,b)
-
-
-  GlobalCounters.reset()
-  with Context(DEBUG=max(2, DEBUG.value), DEVECTORIZE=2):
-    tst = Tensor.custom_kernel(c, a, b, fxn=custom_gemm)[0]
-    tst.realize()
-  print(f"{(N*M*K*2 / GlobalCounters.time_sum_s)*1e-12:.2f} REAL TFLOPS")
-
-
-  with Context(DEBUG=0):
-    ref = a.dot(b, dtype=dtypes.float)
-    ref.realize()
-    #print(ref.numpy())
-    #print(tst.numpy())
-    assert Tensor.isclose(ref, tst, atol=1e-2).all().item(), "matrix not close"
@@ -17,7 +17,7 @@ M = getenv("M", N)
 K = getenv("K", N)
 CNT = getenv("CNT", 10)

-atol, rtol = {dtypes.half:{1e-3, 1e-2}, dtypes.bfloat16:(1e-3, 1e-2), dtypes.fp8e4m3:(1e-1, 1e-1), dtypes.fp8e5m2:(1.0, 5e-1)}.get(dtype_in, (1e-4, 3e-2))
+atol, rtol = {dtypes.bfloat16:(1e-3, 1e-2), dtypes.fp8e4m3:(1e-1, 1e-1), dtypes.fp8e5m2:(1.0, 5e-1)}.get(dtype_in, (1e-4, 3e-2))
 ATOL, RTOL = getenv("ATOL", atol), getenv("RTOL", rtol)

 INT_LOW = getenv("INT_LOW", 0)
@@ -12,7 +12,7 @@ MPS = getenv("MPS", 0)
 if getenv("FP16_ACC"): torch.backends.cuda.matmul.allow_fp16_accumulation = True

 for dtype in [torch.float32, torch.float16, torch.bfloat16]:
-  for N in [256, 512, 1024, 2048, 4096] + ([6144, 8192] if getenv("BIG") else []):
+  for N in [256, 512, 1024, 2048, 4096]:
    FLOPS = N*N*N*2

    b = torch.rand((N,N), dtype=dtype)
@@ -1,16 +0,0 @@
-from tinygrad import Tensor, Device, TinyJit, dtypes
-from tinygrad.helpers import getenv
-
-GPUS = getenv("GPUS", 4) # TODO: expose a way in tinygrad to access this
-N = 6144
-
-@TinyJit
-def many_matmul(A, B):
-  out = A
-  for _ in range(8): out = out@B
-  return out
-
-if __name__ == "__main__":
-  A = Tensor.ones(GPUS, N, N, dtype=dtypes.half).shard(devices=tuple([f"{Device.DEFAULT}:{i}" for i in range(GPUS)]), axis=0).contiguous()
-  B = Tensor.ones(GPUS, N, N, dtype=dtypes.half).shard(devices=tuple([f"{Device.DEFAULT}:{i}" for i in range(GPUS)]), axis=0).contiguous()
-  while 1: many_matmul(A, B)
@@ -51,15 +51,11 @@ def create_report(dev, test, result, stdout, stderr):
  dmesg_output = subprocess.check_output(["sudo", "dmesg", "--ctime", "--color=never"], text=True)
  with open(dmesg_path, "w") as f: f.write(dmesg_output)

-  env_vars = " ".join(f"{k}={v}" for k, v in test.env.items())
-  reproduce_cmd = f"{env_vars} {test.cmd}"
-
  summary_path = os.path.join(report_path, "summary.txt")
  with open(summary_path, "w") as f:
    f.write(f"Test: {test.name()}\n")
    f.write(f"Dev params: {vars(dev)}\n")
    f.write(f"Test params: {vars(test)}\n")
-    f.write(f"Reproduce cmd: {reproduce_cmd}\n")
    f.write(f"Exit Code: {result}\n")

  print(f"Crash report saved to {report_path}")
@@ -19,6 +19,5 @@ trap 'rm -f "$TMP"' EXIT
 EOF
  sed -n '/struct nir_shader_compiler_options/,/^}/{p;/^}/q}' $1/src/gallium/drivers/llvmpipe/lp_screen.c
  echo "int main(void) { write(1, &gallivm_nir_options, sizeof(gallivm_nir_options)); }"
-) | cc -x c -o $TMP - -I$1/src/compiler/nir -I$1/src -I$1/include || exit 1
+) | cc -x c -o $TMP - -I$1/src/compiler/nir -I$1/src -I$1/include && $TMP | gzip | base64 -w0

-printf 'lvp_nir_options = gzip.decompress(base64.b64decode("%s"))' $("$TMP" | gzip | base64 -w0)
@@ -32,7 +32,7 @@ def launchBenchmark(instruction, vgprIndices, dense=True, accum=False, extra="")
  src = src.replace("DIRECTIVE", DIRECTIVE)
  lib = COMPILER.compile(src)
  fxn = AMDProgram(DEV, "matmul", lib)
-  elapsed = min([fxn(global_size=(NUM_WORKGROUPS,1,1), local_size=(WAVE_SIZE*NUM_WAVES,1,1), wait=True) for _ in range(2)])
+  elapsed = fxn(global_size=(NUM_WORKGROUPS,1,1), local_size=(WAVE_SIZE*NUM_WAVES,1,1), wait=True)
  FLOPs = FLOPS_PER_MATMUL * NUM_WAVES * NUM_WORKGROUPS * INTERNAL_LOOP * INSTRUCTIONS_PER_LOOP
  print(f"{instruction:<29} : {FLOPs/elapsed/10**12:.2f} T(FL)OPS")

@@ -3,14 +3,14 @@
    .p2align 8
    .type matmul,@function
 matmul:
-  s_mov_b32 s1, INTERNAL_LOOP
-  s_mov_b32 s2, 0
-  inner_loop:
-    INSTRUCTION
-    s_sub_u32 s1, s1, 1
-    s_cmp_lg_i32 s1, s2
-    s_cbranch_scc1 inner_loop
-  s_endpgm
+    s_mov_b32 s1, INTERNAL_LOOP
+    s_mov_b32 s2, 0
+    inner_loop:
+        INSTRUCTION
+        s_sub_u32 s1, s1, 1
+        s_cmp_lg_i32 s1, s2
+        s_cbranch_scc1 inner_loop
+    s_endpgm

 .rodata
 .p2align 6
@@ -89,20 +89,6 @@ class Attention:
    keys, values = repeat_kv(keys, self.n_rep), repeat_kv(values, self.n_rep)
    xq, keys, values = xq.transpose(1, 2), keys.transpose(1, 2), values.transpose(1, 2)
    attn = xq.scaled_dot_product_attention(keys, values, mask).transpose(1, 2)
-    if getenv("STUB_ATTENTION"):
-      # TODO: do we need mask?
-      from tinygrad.uop.ops import UOp, KernelInfo
-      def fa_custom_forward(attn:UOp, q:UOp, k:UOp, v:UOp) -> UOp:
-        return UOp.sink(arg=KernelInfo(name="fa_custom_forward"))
-      def fa_custom_backward(out_q:UOp, out_k:UOp, out_v:UOp, grad:UOp, q:UOp, k:UOp, v:UOp) -> UOp:
-        return UOp.sink(arg=KernelInfo(name="fa_custom_backward"))
-      def fa_backward(grad:UOp, kernel:UOp) -> tuple[None, UOp, UOp, UOp]:
-        grad_q = Tensor.empty_like(q:=Tensor(kernel.src[1]))
-        grad_k = Tensor.empty_like(k:=Tensor(kernel.src[2]))
-        grad_v = Tensor.empty_like(v:=Tensor(kernel.src[3]))
-        ck = Tensor.custom_kernel(grad_q, grad_k, grad_v, Tensor(grad), q, k, v, fxn=fa_custom_backward)[:3]
-        return (None, ck[0].uop, ck[1].uop, ck[2].uop)
-      attn = Tensor.empty_like(attn).custom_kernel(xq, keys, values, fxn=fa_custom_forward, grad_fxn=fa_backward)[0]
    attn = attn.reshape(bsz, seqlen, -1)
    return self.wo(attn)

@@ -1,53 +0,0 @@
-/*
- * NVIDIA_COPYRIGHT_BEGIN
- *
- * Copyright (c) 2022, NVIDIA CORPORATION.  All rights reserved.
- *
- * NVIDIA CORPORATION and its licensors retain all intellectual property
- * and proprietary rights in and to this software, related documentation
- * and any modifications thereto.  Any use, reproduction, disclosure or
- * distribution of this software and related documentation without an express
- * license agreement from NVIDIA CORPORATION is strictly prohibited.
- *
- * NVIDIA_COPYRIGHT_END
- */
-
-#include <stdint.h>
-#include <stdlib.h>
-
-typedef enum {
-  NVJITLINK_SUCCESS = 0,
-  NVJITLINK_ERROR_UNRECOGNIZED_OPTION,
-  NVJITLINK_ERROR_MISSING_ARCH,
-  NVJITLINK_ERROR_INVALID_INPUT,
-  NVJITLINK_ERROR_PTX_COMPILE,
-  NVJITLINK_ERROR_NVVM_COMPILE,
-  NVJITLINK_ERROR_INTERNAL
-} nvJitLinkResult;
-
-typedef enum {
-  NVJITLINK_INPUT_NONE = 0,
-  NVJITLINK_INPUT_CUBIN = 1,
-  NVJITLINK_INPUT_PTX,
-  NVJITLINK_INPUT_LTOIR,
-  NVJITLINK_INPUT_FATBIN,
-  NVJITLINK_INPUT_OBJECT,
-  NVJITLINK_INPUT_LIBRARY
-} nvJitLinkInputType;
-
-typedef struct nvJitLink* nvJitLinkHandle;
-
-nvJitLinkResult nvJitLinkCreate(nvJitLinkHandle *handle, uint32_t numOptions, const char **options);
-nvJitLinkResult nvJitLinkDestroy(nvJitLinkHandle *handle);
-nvJitLinkResult nvJitLinkAddData(nvJitLinkHandle handle, nvJitLinkInputType inputType, const void *data, size_t size, const char *name);
-nvJitLinkResult nvJitLinkAddFile(nvJitLinkHandle handle, nvJitLinkInputType inputType, const char *fileName);
-nvJitLinkResult nvJitLinkComplete(nvJitLinkHandle handle);
-nvJitLinkResult nvJitLinkGetLinkedCubinSize(nvJitLinkHandle handle, size_t *size);
-nvJitLinkResult nvJitLinkGetLinkedCubin(nvJitLinkHandle handle, void *cubin);
-nvJitLinkResult nvJitLinkGetLinkedPtxSize(nvJitLinkHandle handle, size_t *size);
-nvJitLinkResult nvJitLinkGetLinkedPtx(nvJitLinkHandle handle, char *ptx);
-nvJitLinkResult nvJitLinkGetErrorLogSize(nvJitLinkHandle handle, size_t *size);
-nvJitLinkResult nvJitLinkGetErrorLog(nvJitLinkHandle handle, char *log);
-nvJitLinkResult nvJitLinkGetInfoLogSize(nvJitLinkHandle handle, size_t *size);
-nvJitLinkResult nvJitLinkGetInfoLog(nvJitLinkHandle handle, char *log);
-nvJitLinkResult nvJitLinkVersion(unsigned int *major, unsigned int *minor);
@@ -65,8 +65,6 @@
 #define NVCEC0_QMDV05_00_GRID_HEIGHT_RESUME                        MW(271:256)
 #define NVCEC0_QMDV05_00_GRID_DEPTH_RESUME                         MW(287:272)
 #define NVCEC0_QMDV05_00_RELEASE_ENABLE(i)                         MW((288+(i)*16):(288+(i)*16))
-#define NVCEC0_QMDV05_00_RELEASE0_ENABLE                           NVCEC0_QMDV05_00_RELEASE_ENABLE(0)
-#define NVCEC0_QMDV05_00_RELEASE1_ENABLE                           NVCEC0_QMDV05_00_RELEASE_ENABLE(1)
 #define NVCEC0_QMDV05_00_RELEASE_ENABLE_FALSE                      0x00000000
 #define NVCEC0_QMDV05_00_RELEASE_ENABLE_TRUE                       0x00000001
 #define NVCEC0_QMDV05_00_RELEASE_STRUCTURE_SIZE(i)                 MW((290+(i)*16):(289+(i)*16))
@@ -58,13 +58,13 @@ def install_hook(c_function, python_function):
  return orig_func

 # *** ioctl lib end ***
-from tinygrad.runtime.autogen import nv_570 as nv_gpu
+import tinygrad.runtime.autogen.nv_gpu as nv_gpu
 nvescs = {getattr(nv_gpu, x):x for x in dir(nv_gpu) if x.startswith("NV_ESC")}
 nvcmds = {getattr(nv_gpu, x):(x, getattr(nv_gpu, "struct_"+x+"_PARAMS", getattr(nv_gpu, "struct_"+x.replace("_CMD_", "_")+"_PARAMS", None))) for x in dir(nv_gpu) if \
          x.startswith("NV") and x[6:].startswith("_CTRL_") and isinstance(getattr(nv_gpu, x), int)}

 def get_classes():
-  hdrpy = (pathlib.Path(__file__).parent.parent.parent / "tinygrad/runtime/autogen/nv_570.py").read_text()
+  hdrpy = (pathlib.Path(__file__).parent.parent.parent / "tinygrad/runtime/autogen/nv_gpu.py").read_text()
  clss = re.search(r'NV01_ROOT.*?NV_SEMAPHORE_SURFACE = \(0x000000da\) # macro', hdrpy, re.DOTALL).group()
  pattern = r'([0-9a-zA-Z_]*) = +\((0x[0-9a-fA-F]+)\)'
  matches = re.findall(pattern, clss, re.MULTILINE)
@@ -272,4 +272,4 @@ def compare_launch_state(states, good_states):

  return True, "PASS"

-# IOCTL=1 CUDA=1 CUDA_PTX=1 python3 test/test_ops.py TestOps.test_tiny_add
+# IOCTL=1 CUDA=1 CUDA_PTX=1 python3 test/test_ops.py TestOps.test_tiny_add
@@ -1,134 +0,0 @@
-/*
- * SPDX-FileCopyrightText: Copyright (c) 1993-2021 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a
- * copy of this software and associated documentation files (the "Software"),
- * to deal in the Software without restriction, including without limitation
- * the rights to use, copy, modify, merge, publish, distribute, sublicense,
- * and/or sell copies of the Software, and to permit persons to whom the
- * Software is furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in
- * all copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
- * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
- * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
- * DEALINGS IN THE SOFTWARE.
- */
-
-#ifndef PCIEXPTBL_H
-#define PCIEXPTBL_H
-
-#define NV_BCRT_HASH_INFO_BASE_CODE_TYPE_VBIOS_BASE   0x00
-#define NV_BCRT_HASH_INFO_BASE_CODE_TYPE_VBIOS_EXT    0xE0
-
-//
-// The VBIOS object comes from walking the PCI expansion code block
-// The following structure holds the expansion code format.
-//
-#define PCI_EXP_ROM_SIGNATURE     0xaa55
-#define PCI_EXP_ROM_SIGNATURE_NV  0x4e56 // "VN" in word format
-#define PCI_EXP_ROM_SIGNATURE_NV2 0xbb77
-#define IS_VALID_PCI_ROM_SIG(sig) ((sig == PCI_EXP_ROM_SIGNATURE) ||    \
-                                   (sig == PCI_EXP_ROM_SIGNATURE_NV) || \
-                                   (sig == PCI_EXP_ROM_SIGNATURE_NV2))
-
-#define OFFSETOF_PCI_EXP_ROM_SIG                    0x0
-#define OFFSETOF_PCI_EXP_ROM_NBSI_DATA_OFFSET       0x16
-#define OFFSETOF_PCI_EXP_ROM_PCI_DATA_STRUCT_PTR    0x18
-
-#pragma pack(1)
-typedef struct _PCI_EXP_ROM_STANDARD
-{
-    NvU16       sig;                //  00h: ROM Signature 0xaa55
-    NvU8        reserved [0x16];    //  02h: Reserved (processor architecture unique data)
-    NvU16       pciDataStrucPtr;    //  18h: Pointer to PCI Data Structure
-    NvU32       sizeOfBlock;        //  1Ah: <NBSI-specific appendage>
-} PCI_EXP_ROM_STANDARD, *PPCI_EXP_ROM_STANDARD;
-#pragma pack()
-
-#pragma pack(1)
-typedef struct _PCI_EXP_ROM_NBSI
-{
-    NvU16       sig;                //  00h: ROM Signature 0xaa55
-    NvU8        reserved [0x14];    //  02h: Reserved (processor architecture unique data)
-    NvU16       nbsiDataOffset;     //  16h: Offset from header to NBSI image
-    NvU16       pciDataStrucPtr;    //  18h: Pointer to PCI Data Structure
-    NvU32       sizeOfBlock;        //  1Ah: <NBSI-specific appendage>
-} PCI_EXP_ROM_NBSI, *PPCI_EXP_ROM_NBSI;
-#pragma pack()
-
-typedef union _PCI_EXP_ROM {
-    PCI_EXP_ROM_STANDARD standard;
-    PCI_EXP_ROM_NBSI nbsi;
-} PCI_EXP_ROM, *PPCI_EXP_ROM;
-
-#define PCI_DATA_STRUCT_SIGNATURE     0x52494350 // "PCIR" in dword format
-#define PCI_DATA_STRUCT_SIGNATURE_NV  0x5344504E // "NPDS" in dword format
-#define PCI_DATA_STRUCT_SIGNATURE_NV2 0x53494752 // "RGIS" in dword format
-#define IS_VALID_PCI_DATA_SIG(sig) ((sig == PCI_DATA_STRUCT_SIGNATURE) ||    \
-                                    (sig == PCI_DATA_STRUCT_SIGNATURE_NV) || \
-                                    (sig == PCI_DATA_STRUCT_SIGNATURE_NV2))
-
-#define PCI_LAST_IMAGE NVBIT(7)
-#define PCI_ROM_IMAGE_BLOCK_SIZE 512U
-
-#define OFFSETOF_PCI_DATA_STRUCT_SIG        0x0
-#define OFFSETOF_PCI_DATA_STRUCT_VENDOR_ID  0x4
-#define OFFSETOF_PCI_DATA_STRUCT_LEN        0xa
-#define OFFSETOF_PCI_DATA_STRUCT_CLASS_CODE 0xd
-#define OFFSETOF_PCI_DATA_STRUCT_CODE_TYPE  0x14
-#define OFFSETOF_PCI_DATA_STRUCT_IMAGE_LEN  0x10
-#define OFFSETOF_PCI_DATA_STRUCT_LAST_IMAGE 0x15
-
-#pragma pack(1)
-typedef struct _PCI_DATA_STRUCT
-{
-    NvU32       sig;                //  00h: Signature, the string "PCIR" or NVIDIA's alternate "NPDS"
-    NvU16       vendorID;           //  04h: Vendor Identification
-    NvU16       deviceID;           //  06h: Device Identification
-    NvU16       deviceListPtr;      //  08h: Device List Pointer
-    NvU16       pciDataStructLen;   //  0Ah: PCI Data Structure Length
-    NvU8        pciDataStructRev;   //  0Ch: PCI Data Structure Revision
-    NvU8        classCode[3];       //  0Dh: Class Code
-    NvU16       imageLen;           //  10h: Image Length (units of 512 bytes)
-    NvU16       vendorRomRev;       //  12h: Revision Level of the Vendor's ROM
-    NvU8        codeType;           //  14h: holds NBSI_OBJ_CODE_TYPE (0x70) and others
-    NvU8        lastImage;          //  15h: Last Image Indicator: bit7=1 is lastImage
-    NvU16       maxRunTimeImageLen; //  16h: Maximum Run-time Image Length (units of 512 bytes)
-} PCI_DATA_STRUCT, *PPCI_DATA_STRUCT;
-#pragma pack()
-
-#define NV_PCI_DATA_EXT_SIG 0x4544504E // "NPDE" in dword format
-#define NV_PCI_DATA_EXT_REV_10 0x100      // 1.0
-#define NV_PCI_DATA_EXT_REV_11 0x101      // 1.1
-
-#define OFFSETOF_PCI_DATA_EXT_STRUCT_SIG            0x0
-#define OFFSETOF_PCI_DATA_EXT_STRUCT_LEN            0x6
-#define OFFSETOF_PCI_DATA_EXT_STRUCT_REV            0x4
-#define OFFSETOF_PCI_DATA_EXT_STRUCT_SUBIMAGE_LEN   0x8
-#define OFFSETOF_PCI_DATA_EXT_STRUCT_LAST_IMAGE     0xa
-#define OFFSETOF_PCI_DATA_EXT_STRUCT_FLAGS          0xb
-
-#define PCI_DATA_EXT_STRUCT_FLAGS_CHECKSUM_DISABLED 0x04
-
-#pragma pack(1)
-typedef struct _NV_PCI_DATA_EXT_STRUCT
-{
-    NvU32   signature;          //  00h: Signature, the string "NPDE"
-    NvU16   nvPciDataExtRev;    //  04h: NVIDIA PCI Data Extension Revision
-    NvU16   nvPciDataExtLen;    //  06h: NVIDIA PCI Data Extension Length
-    NvU16   subimageLen;        //  08h: Sub-image Length
-    NvU8    privLastImage;      //  0Ah: Private Last Image Indicator
-    NvU8    flags;              //  0Bh: Private images enabled if bit0=1
-} NV_PCI_DATA_EXT_STRUCT, *PNV_PCI_DATA_EXT_STRUCT;
-#pragma pack()
-
-#endif // PCIEXPTBL_H
-
-
@@ -0,0 +1,68 @@
+import ctypes
+from dataclasses import dataclass
+import tinygrad.runtime.autogen.comgr as comgr
+from tinygrad.runtime.support.compiler_amd import check
+
+@dataclass
+class InstrCtx:
+  pc:int=0
+  inst:str=""
+
+@comgr.amd_comgr_create_disassembly_info.argtypes[2]
+def instr_cb(text, user_data):
+  c = ctypes.cast(user_data, ctypes.POINTER(ctypes.py_object)).contents.value
+  c.inst = ctypes.string_at(text).decode("utf-8","replace").strip()
+  return comgr.AMD_COMGR_STATUS_SUCCESS
+
+# nop callback
+@comgr.amd_comgr_create_disassembly_info.argtypes[3]
+def addr_cb(*args): return comgr.AMD_COMGR_STATUS_SUCCESS
+
+def comgr_get_address_table(lib:bytes) -> dict[int, tuple[str, int]]:
+  check(comgr.amd_comgr_create_data(comgr.AMD_COMGR_DATA_KIND_EXECUTABLE, ctypes.byref(data_src:=comgr.amd_comgr_data_t())))
+  lib_buf = ctypes.create_string_buffer(lib, len(lib))
+  check(comgr.amd_comgr_set_data(data_src, len(lib), lib_buf))
+  check(comgr.amd_comgr_get_data_isa_name(data_src, isa_sz:=ctypes.c_size_t(128), isa:=(ctypes.c_char*isa_sz.value)()))
+
+  @comgr.amd_comgr_create_disassembly_info.argtypes[1]
+  def memory_cb(from_addr, to, size, _):
+    base, buf_len = ctypes.addressof(lib_buf), len(lib_buf)
+    start = int(from_addr) - base
+    if start < 0 or start >= buf_len: return 0
+    ctypes.memmove(to, base + start, n:=min(int(size), buf_len - start))
+    return n
+
+  info_src = comgr.amd_comgr_disassembly_info_t()
+  check(comgr.amd_comgr_create_disassembly_info(ctypes.cast(isa, ctypes.POINTER(ctypes.c_char)), memory_cb, instr_cb, addr_cb, info_src))
+
+  @comgr.amd_comgr_iterate_symbols.argtypes[1]
+  def sym_callback(sym, udata):
+    check(comgr.amd_comgr_symbol_get_info(sym, comgr.AMD_COMGR_SYMBOL_INFO_TYPE, ctypes.byref(sym_type:=ctypes.c_int())))
+    if sym_type.value != comgr.AMD_COMGR_SYMBOL_TYPE_FUNC: return comgr.AMD_COMGR_STATUS_SUCCESS
+    check(comgr.amd_comgr_symbol_get_info(sym, comgr.AMD_COMGR_SYMBOL_INFO_VALUE, ctypes.byref(vaddr:=ctypes.c_uint64())))
+    check(comgr.amd_comgr_symbol_get_info(sym, comgr.AMD_COMGR_SYMBOL_INFO_SIZE, ctypes.byref(size:=ctypes.c_uint64())))
+    check(comgr.amd_comgr_map_elf_virtual_address_to_code_object_offset(data_src, vaddr.value, ctypes.byref(offset:=ctypes.c_uint64()),
+                                                                        ctypes.byref(ctypes.c_uint64()), ctypes.byref(nobits:=ctypes.c_bool())))
+    check(nobits.value)
+    base = ctypes.addressof(lib_buf)
+    pc = base + offset.value
+    end = pc + size.value
+    addr_table = ctypes.cast(udata, ctypes.POINTER(ctypes.py_object)).contents.value
+    instr_ref = ctypes.py_object(ctx:=InstrCtx())
+    instr_ptr = ctypes.cast(ctypes.pointer(instr_ref), ctypes.c_void_p)
+    while pc < end:
+      size_read = ctypes.c_uint64(0)
+      ctx.pc = pc
+      st = comgr.amd_comgr_disassemble_instruction(info_src, ctypes.c_uint64(pc), instr_ptr, ctypes.byref(size_read))
+      if st == comgr.AMD_COMGR_STATUS_SUCCESS and size_read.value:
+        rel = (pc - base) - offset.value
+        addr_table[vaddr.value + rel] = (ctx.inst, int(size_read.value))
+        pc += size_read.value
+      else: # don't inf loop if comgr fails
+        b = ctypes.c_ubyte.from_buffer(lib_buf, pc - base).value
+        addr_table[vaddr.value + (pc - base - offset.value)] = (f"DISASSEMBLER ISSUE 0x{b:02x}", 1)
+        pc += 1
+    return comgr.AMD_COMGR_STATUS_SUCCESS
+  addr_table:dict[int, tuple[str, int]] = {}
+  check(comgr.amd_comgr_iterate_symbols(data_src, sym_callback, ctypes.cast(ctypes.pointer(ctypes.py_object(addr_table)), ctypes.c_void_p)))
+  return addr_table
@@ -4,7 +4,7 @@ import argparse, ctypes, struct, hashlib, pickle, code, typing, functools
 import tinygrad.runtime.autogen.sqtt as sqtt
 from tinygrad.device import ProfileEvent, ProfileDeviceEvent, ProfileProgramEvent
 from tinygrad.runtime.ops_amd import ProfileSQTTEvent
-from tinygrad.helpers import round_up, flatten, all_same, temp
+from tinygrad.helpers import round_up, flatten, all_same
 from dataclasses import dataclass

 CHUNK_CLASSES = {
@@ -154,22 +154,8 @@ class RGP:
      if device not in device_events: raise RuntimeError(f"Device {device} not found in profile, devices in profile: {', '.join(device_events.keys())} ")
      device_event = device_events[device]
    sqtt_events = [x for x in profile if isinstance(x, ProfileSQTTEvent) and x.device == device_event.device]
-    device_props = device_event.props
-    # merge events per SE
-    merged_sqtt_events:dict[int, ProfileSQTTEvent] = {}
-    for ev in sqtt_events:
-      if ev.se not in merged_sqtt_events: merged_sqtt_events[ev.se] = ev
-      else:
-        merged_sqtt_events[ev.se] = ProfileSQTTEvent(
-          device=ev.device,
-          kern=ev.kern,
-          se=ev.se,
-          itrace=merged_sqtt_events[ev.se].itrace or ev.itrace,
-          blob=merged_sqtt_events[ev.se].blob + ev.blob,
-        )
-    sqtt_events = list(merged_sqtt_events.values())
-
    if len(sqtt_events) == 0: raise RuntimeError(f"Device {device_event.device} doesn't contain SQTT data")
+    device_props = sqtt_events[0].props
    gfx_ver = device_props['gfx_target_version'] // 10000
    gfx_iplvl = getattr(sqtt, f"SQTT_GFXIP_LEVEL_GFXIP_{device_props['gfx_target_version']//10000}_{(device_props['gfx_target_version']//100)%100}",
                        getattr(sqtt, f"SQTT_GFXIP_LEVEL_GFXIP_{device_props['gfx_target_version']//10000}", None))
@@ -185,7 +171,9 @@ class RGP:
      magic_number=sqtt.SQTT_FILE_MAGIC_NUMBER,
      version_major=sqtt.SQTT_FILE_VERSION_MAJOR,
      version_minor=sqtt.SQTT_FILE_VERSION_MINOR,
-      flags=sqtt.struct_sqtt_file_header_flags(value=1,),
+      flags=sqtt.struct_sqtt_file_header_flags(
+        _0=sqtt.union_sqtt_file_header_flags_0(value=1),
+      ),
      chunk_offset=ctypes.sizeof(sqtt.struct_sqtt_file_header),
    )
    chunks = [
@@ -208,7 +196,7 @@ class RGP:
        flags=0,
        trace_shader_core_clock=0x93f05080,
        trace_memory_clock=0x4a723a40,
-        device_id={110000: 0x744c, 110003: 0x7480, 120001: 0x7550, 120000: 0x7550}[device_props['gfx_target_version']],
+        device_id={110000: 0x744c, 110003: 0x7480, 120001: 0x7550}[device_props['gfx_target_version']],
        device_revision_id=0xc8,
        vgprs_per_simd=1536,
        sgprs_per_simd=128*16,
@@ -263,7 +251,7 @@ class RGP:
        profiling_mode=sqtt.SQTT_PROFILING_MODE_PRESENT,
        instruction_trace_mode=sqtt.SQTT_INSTRUCTION_TRACE_FULL_FRAME if sqtt_itrace_enabled else sqtt.SQTT_INSTRUCTION_TRACE_DISABLED,
        instruction_trace_data=sqtt.union_sqtt_instruction_trace_data(
-          shader_engine_filter=sqtt.union_sqtt_instruction_trace_data_shader_engine_filter(mask=sqtt_itrace_se_mask),
+          shader_engine_filter=sqtt.struct_sqtt_instruction_trace_data_shader_engine_filter(mask=sqtt_itrace_se_mask),
        ),
      )),
      *flatten([(
@@ -274,11 +262,13 @@ class RGP:
          ),
          shader_engine_index=sqtt_event.se,
          sqtt_version={11: sqtt.SQTT_VERSION_3_2, 12: sqtt.SQTT_VERSION_3_3}.get(gfx_ver),
-          v1=sqtt.struct_sqtt_file_chunk_sqtt_desc_0_v1(
-            instrumentation_spec_version=1,
-            instrumentation_api_version=0,
-            compute_unit_index=0,
-          )
+          _0=sqtt.union_sqtt_file_chunk_sqtt_desc_0(
+            v1=sqtt.struct_sqtt_file_chunk_sqtt_desc_0_v1(
+              instrumentation_spec_version=1,
+              instrumentation_api_version=0,
+              compute_unit_index=0,
+            )
+          ),
        )),
        RGPChunk(sqtt.struct_sqtt_file_chunk_sqtt_data(
          header=sqtt.struct_sqtt_file_chunk_header(
@@ -320,7 +310,7 @@ class RGP:
 if __name__ == '__main__':
  parser = argparse.ArgumentParser(prog='rgptool', description='A tool to create (from pickled tinygrad profile), inspect and modify Radeon GPU Profiler files')
  parser.add_argument('command')
-  parser.add_argument('input', nargs='?', default=temp("profile.pkl", append_user=True))
+  parser.add_argument('input')
  parser.add_argument('-d', '--device')
  parser.add_argument('-o', '--output')
  args = parser.parse_args()
@@ -342,4 +332,3 @@ if __name__ == '__main__':

  if args.output is not None:
    with open(args.output, 'wb+') as fd: fd.write(rgp.to_bytes())
-    print(f"Saved to {args.output}")
@@ -1,101 +1,76 @@
 import ctypes, pathlib, argparse, pickle, re, functools, dataclasses, itertools
-from tinygrad.helpers import temp, unwrap, DEBUG
-from tinygrad.device import ProfileEvent, ProfileDeviceEvent, ProfileProgramEvent
+from extra.sqtt.rocprof import rocprof
+from extra.sqtt.disasm import comgr_get_address_table
+from tinygrad.helpers import temp, DEBUG
+from tinygrad.device import ProfileEvent, ProfileProgramEvent
 from tinygrad.runtime.ops_amd import ProfileSQTTEvent, ProfilePMCEvent
-from tinygrad.runtime.autogen import llvm, rocprof
-from tinygrad.runtime.support.elf import elf_loader

-# to pass NULL to callbacks
-llvm.LLVMCreateDisasmCPUFeatures.argtypes = tuple(llvm.LLVMCreateDisasmCPUFeatures.argtypes[:5]) + (ctypes.c_void_p, ctypes.c_void_p)
-def llvm_disasm(arch:str, lib:bytes) -> dict[int, tuple[str, int]]:
-  llvm.LLVMInitializeAMDGPUTargetInfo()
-  llvm.LLVMInitializeAMDGPUTargetMC()
-  llvm.LLVMInitializeAMDGPUAsmParser()
-  llvm.LLVMInitializeAMDGPUDisassembler()
-  ctx = llvm.LLVMCreateDisasmCPUFeatures("amdgcn-amd-amdhsa".encode(), arch.encode(), "".encode(), None, 0, None, None)
+@dataclasses.dataclass
+class InstInfo:
+  typ:str=""
+  inst:str=""
+  hit:int=0
+  lat:int=0
+  stall:int=0
+  def __str__(self): return f"{self.inst:>20} hits:{self.typ:>6} hits:{self.hit:>6} latency:{self.lat:>6} stall:{self.stall:>6}"

-  image, sections, relocs = elf_loader(lib)
-  text = next((sh.header for sh in sections if sh.name == ".text"), None)
-  off, sz = unwrap(text).sh_addr, unwrap(text).sh_size
-
-  addr_table:dict[int, tuple[str, int]] = {}
-  out = ctypes.create_string_buffer(128)
-  cur_off = off
-  while cur_off < sz + off:
-    view = (ctypes.c_ubyte * ((sz + off) - cur_off)).from_buffer_copy(memoryview(image)[cur_off:])
-    instr_sz = llvm.LLVMDisasmInstruction(ctx, view, ctypes.c_uint64(len(view)), ctypes.c_uint64(0), out, ctypes.c_size_t(128))
-    addr_table[cur_off] = (out.value.decode("utf-8", "replace").strip(), instr_sz)
-    cur_off += instr_sz
-  return addr_table
-
-@dataclasses.dataclass(frozen=True)
-class InstExec:
-  typ:str
-  inst:str
-  stall:int
-  dur:int
-  time:int
-
-@dataclasses.dataclass(frozen=True)
-class WaveExec:
-  wave_id:int
-  cu:int
-  simd:int
-  begin_time:int
-  end_time:int
-  insts:list[InstExec]
+  def on_ev(self, ev):
+    self.hit, self.lat, self.stall = self.hit + 1, self.lat + ev.duration, self.stall + ev.stall

 class _ROCParseCtx:
-  def __init__(self, dev_evs:dict[str, ProfileDeviceEvent], sqtt_evs:list[ProfileSQTTEvent], prog_evs:list[ProfileProgramEvent]):
-    self.dev_evs, self.sqtt_evs, self.prog_evs = dev_evs, iter(sqtt_evs), prog_evs
-    self.disasms:dict[tuple[str, int], tuple[str, int]] = {}
-    self.inst_execs:dict[str, list[WaveExec]] = {}
+  def __init__(self, sqtt_evs:list[ProfileSQTTEvent], prog_evs:list[ProfileProgramEvent]):
+    self.sqtt_evs, self.prog_evs = iter(sqtt_evs), prog_evs
+    self.wave_events, self.disasms, self.addr2prg = {}, {}, {}

    for prog in prog_evs:
-      arch = "gfx%d%x%x" % ((trgt:=unwrap(dev_evs[prog.device].props)['gfx_target_version']) // 10000, (trgt // 100) % 100, trgt % 100)
-      for addr, info in llvm_disasm(arch, unwrap(prog.lib)).items():
-        self.disasms[(prog.name, unwrap(prog.base) + addr)] = info
+      for addr, info in comgr_get_address_table(prog.lib).items():
+        self.disasms[prog.base + addr] = info
+        self.addr2prg[prog.base + addr] = prog

  def next_sqtt(self):
    x = next(self.sqtt_evs, None)
-    self.active_kern = x.kern if x is not None else None
    self.active_se = x.se if x is not None else None
-    self.active_blob = (ctypes.c_ubyte * len(x.blob)).from_buffer_copy(x.blob) if x is not None else None
-    return self.active_blob
+    return x
+
+  def find_program(self, addr): return self.addr2prg[addr]

  def on_occupancy_ev(self, ev):
-    if DEBUG >= 5: print("OCC", ev.time, self.active_se, ev.cu, ev.simd, ev.wave_id, ev.start)
+    if DEBUG >= 4: print("OCC", ev.time, self.active_se, ev.cu, ev.simd, ev.wave_id, ev.start)

  def on_wave_ev(self, ev):
-    if DEBUG >= 5: print("WAVE", ev.wave_id, self.active_se, ev.cu, ev.simd, ev.contexts, ev.begin_time, ev.end_time)
+    if DEBUG >= 4: print("WAVE", ev.wave_id, self.active_se, ev.cu, ev.simd, ev.contexts, ev.begin_time, ev.end_time)

-    inst_execs:list[InstExec] = []
+    asm = {}
    for j in range(ev.instructions_size):
      inst_ev = ev.instructions_array[j]
-      inst_typ = rocprof.enum_rocprofiler_thread_trace_decoder_inst_category_t.get(inst_ev.category)
-      inst_disasm = self.disasms[(unwrap(self.active_kern), unwrap(inst_ev.pc.address))][0]
-      inst_execs.append(InstExec(inst_typ, inst_disasm, inst_ev.stall, inst_ev.duration, inst_ev.time))
+      inst_typ = rocprof.rocprofiler_thread_trace_decoder_inst_category_t__enumvalues[inst_ev.category]
+      asm.setdefault(inst_ev.pc.address, InstInfo(typ=inst_typ, inst=self.disasms[inst_ev.pc.address][0]))
+      asm[inst_ev.pc.address].on_ev(inst_ev)

-    if ev.instructions_size > 0:
-      self.inst_execs.setdefault(unwrap(self.active_kern), []).append(WaveExec(ev.wave_id, ev.cu, ev.simd, ev.begin_time, ev.end_time, inst_execs))
+    self.wave_events[(self.find_program(ev.instructions_array[0].pc.address).name, ev.wave_id, ev.cu, ev.simd)] = asm

-def decode(profile:list[ProfileEvent]) -> _ROCParseCtx:
-  dev_events:dict[str, ProfileDeviceEvent] = {}
+if __name__ == "__main__":
+  parser = argparse.ArgumentParser()
+  parser.add_argument('--profile', type=pathlib.Path, help='Path to profile', default=pathlib.Path(temp("profile.pkl", append_user=True)))
+  args = parser.parse_args()
+
+  with args.profile.open("rb") as f: profile = pickle.load(f)
  sqtt_events:list[ProfileSQTTEvent] = []
+  pmc_events:list[ProfilePMCEvent] = []
  prog_events:list[ProfileProgramEvent] = []
  for e in profile:
-    if isinstance(e, ProfileDeviceEvent): dev_events[e.device] = e
    if isinstance(e, ProfileSQTTEvent): sqtt_events.append(e)
+    if isinstance(e, ProfilePMCEvent): pmc_events.append(e)
    if isinstance(e, ProfileProgramEvent) and e.device.startswith("AMD"): prog_events.append(e)

-  ROCParseCtx = _ROCParseCtx(dev_events, sqtt_events, prog_events)
+  ROCParseCtx = _ROCParseCtx(sqtt_events, prog_events)

  @rocprof.rocprof_trace_decoder_se_data_callback_t
  def copy_cb(buf, buf_size, data_ptr):
-    if (prof_info:=ROCParseCtx.next_sqtt()) is None: return 0
-    buf[0] = ctypes.cast(prof_info, ctypes.POINTER(ctypes.c_ubyte))
-    buf_size[0] = len(prof_info)
-    return len(prof_info)
+    if (prof:=ROCParseCtx.next_sqtt()) is None: return 0
+    buf[0] = ctypes.cast((ctypes.c_ubyte * len(prof.blob)).from_buffer_copy(prof.blob), ctypes.POINTER(ctypes.c_ubyte))
+    buf_size[0] = len(prof.blob)
+    return len(prof.blob)

  @rocprof.rocprof_trace_decoder_trace_callback_t
  def trace_cb(record_type, events_ptr, n, data_ptr):
@@ -105,12 +80,12 @@ def decode(profile:list[ProfileEvent]) -> _ROCParseCtx:
      case rocprof.ROCPROFILER_THREAD_TRACE_DECODER_RECORD_WAVE:
        for ev in (rocprof.rocprofiler_thread_trace_decoder_wave_t * n).from_address(events_ptr): ROCParseCtx.on_wave_ev(ev)
      case _:
-        if DEBUG >= 5: print(rocprof.enum_rocprofiler_thread_trace_decoder_record_type_t.get(record_type), events_ptr, n)
+        if DEBUG >= 2: print(rocprof.rocprofiler_thread_trace_decoder_record_type_t__enumvalues[record_type], events_ptr, n)
    return rocprof.ROCPROFILER_THREAD_TRACE_DECODER_STATUS_SUCCESS

  @rocprof.rocprof_trace_decoder_isa_callback_t
  def isa_cb(instr_ptr, mem_size_ptr, size_ptr, pc, data_ptr):
-    instr, mem_size_ptr[0] = ROCParseCtx.disasms[(unwrap(ROCParseCtx.active_kern), pc.address)]
+    instr, mem_size_ptr[0] = ROCParseCtx.disasms[pc.address]

    # this is the number of bytes to next instruction, set to 0 for end_pgm
    if instr == "s_endpgm": mem_size_ptr[0] = 0
@@ -125,20 +100,10 @@ def decode(profile:list[ProfileEvent]) -> _ROCParseCtx:

  try:
    rocprof.rocprof_trace_decoder_parse_data(copy_cb, trace_cb, isa_cb, None)
-  except AttributeError as e: raise RuntimeError("Failed to find rocprof-trace-decoder. Run sudo ./extra/sqtt/install_sqtt_decoder.py to install") from e
-  return ROCParseCtx
+    print('SQTT:', ROCParseCtx.wave_events.keys())
+  except Exception as e: print("Error in sqtt decoder:", e)

-if __name__ == "__main__":
-  parser = argparse.ArgumentParser()
-  parser.add_argument('--profile', type=pathlib.Path, help='Path to profile', default=pathlib.Path(temp("profile.pkl", append_user=True)))
-  args = parser.parse_args()
-
-  with args.profile.open("rb") as f: profile = pickle.load(f)
-  rctx = decode(profile)
-  print('SQTT:', rctx.inst_execs.keys())
-
-  for ev in profile:
-    if not isinstance(ev, ProfilePMCEvent): continue
+  for ev in pmc_events:
    print(f"PMC Event: dev={ev.device} kern={ev.kern}")
    ptr = 0
    for s in ev.sched:
@@ -12,9 +12,7 @@ if __name__ == "__main__":
    lib = fp.parent/"rocprof-trace-decoder-macos-arm64-0.1.4-Darwin"/"lib"/"librocprof-trace-decoder.dylib"
    os.chmod(fp, 0o755)
    os.system(f"sudo {fp} --prefix={fp.parent} --include-subdir")
-    shutil.copy2(lib, DEST)
  else:
-    lib = DEST/"librocprof-trace-decoder.so"
-    os.system("sudo curl -L https://github.com/ROCm/rocprof-trace-decoder/raw/43bf0fef74a83c3c25badfc5a09c0bd39ed8c6f9/releases/linux_glibc_2_28_x86_64/librocprof-trace-decoder.so -o"+str(lib))
-    os.system("sudo ldconfig")
+    lib = fetch("https://github.com/ROCm/rocprof-trace-decoder/raw/5420409ad0963b2d76450add067b9058493ccbd0/releases/linux_glibc_2_28_x86_64/librocprof-trace-decoder.so", name="librocprof-trace-decoder.so")
+  shutil.copy2(lib, DEST)
  print(f"Installed {lib.name} to", DEST)
@@ -0,0 +1,658 @@
+# pylint: skip-file
+# mypy: ignore-errors
+# -*- coding: utf-8 -*-
+#
+# TARGET arch is: []
+# WORD_SIZE is: 8
+# POINTER_SIZE is: 8
+# LONGDOUBLE_SIZE is: 16
+#
+import ctypes, ctypes.util
+
+
+class AsDictMixin:
+    import sys
+    if sys.version_info >= (3, 14): _layout_ = 'ms'
+    @classmethod
+    def as_dict(cls, self):
+        result = {}
+        if not isinstance(self, AsDictMixin):
+            # not a structure, assume it's already a python object
+            return self
+        if not hasattr(cls, "_fields_"):
+            return result
+        # sys.version_info >= (3, 5)
+        # for (field, *_) in cls._fields_:  # noqa
+        for field_tuple in cls._fields_:  # noqa
+            field = field_tuple[0]
+            if field.startswith('PADDING_'):
+                continue
+            value = getattr(self, field)
+            type_ = type(value)
+            if hasattr(value, "_length_") and hasattr(value, "_type_"):
+                # array
+                if not hasattr(type_, "as_dict"):
+                    value = [v for v in value]
+                else:
+                    type_ = type_._type_
+                    value = [type_.as_dict(v) for v in value]
+            elif hasattr(value, "contents") and hasattr(value, "_type_"):
+                # pointer
+                try:
+                    if not hasattr(type_, "as_dict"):
+                        value = value.contents
+                    else:
+                        type_ = type_._type_
+                        value = type_.as_dict(value.contents)
+                except ValueError:
+                    # nullptr
+                    value = None
+            elif isinstance(value, AsDictMixin):
+                # other structure
+                value = type_.as_dict(value)
+            result[field] = value
+        return result
+
+
+class Structure(ctypes.Structure, AsDictMixin):
+
+    def __init__(self, *args, **kwds):
+        # We don't want to use positional arguments fill PADDING_* fields
+
+        args = dict(zip(self.__class__._field_names_(), args))
+        args.update(kwds)
+        super(Structure, self).__init__(**args)
+
+    @classmethod
+    def _field_names_(cls):
+        if hasattr(cls, '_fields_'):
+            return (f[0] for f in cls._fields_ if not f[0].startswith('PADDING'))
+        else:
+            return ()
+
+    @classmethod
+    def get_type(cls, field):
+        for f in cls._fields_:
+            if f[0] == field:
+                return f[1]
+        return None
+
+    @classmethod
+    def bind(cls, bound_fields):
+        fields = {}
+        for name, type_ in cls._fields_:
+            if hasattr(type_, "restype"):
+                if name in bound_fields:
+                    if bound_fields[name] is None:
+                        fields[name] = type_()
+                    else:
+                        # use a closure to capture the callback from the loop scope
+                        fields[name] = (
+                            type_((lambda callback: lambda *args: callback(*args))(
+                                bound_fields[name]))
+                        )
+                    del bound_fields[name]
+                else:
+                    # default callback implementation (does nothing)
+                    try:
+                        default_ = type_(0).restype().value
+                    except TypeError:
+                        default_ = None
+                    fields[name] = type_((
+                        lambda default_: lambda *args: default_)(default_))
+            else:
+                # not a callback function, use default initialization
+                if name in bound_fields:
+                    fields[name] = bound_fields[name]
+                    del bound_fields[name]
+                else:
+                    fields[name] = type_()
+        if len(bound_fields) != 0:
+            raise ValueError(
+                "Cannot bind the following unknown callback(s) {}.{}".format(
+                    cls.__name__, bound_fields.keys()
+            ))
+        return cls(**fields)
+
+
+class Union(ctypes.Union, AsDictMixin):
+    pass
+
+
+
+c_int128 = ctypes.c_ubyte*16
+c_uint128 = c_int128
+void = None
+if ctypes.sizeof(ctypes.c_longdouble) == 16:
+    c_long_double_t = ctypes.c_longdouble
+else:
+    c_long_double_t = ctypes.c_ubyte*16
+
+def string_cast(char_pointer, encoding='utf-8', errors='strict'):
+    value = ctypes.cast(char_pointer, ctypes.c_char_p).value
+    if value is not None and encoding is not None:
+        value = value.decode(encoding, errors=errors)
+    return value
+
+
+def char_pointer_cast(string, encoding='utf-8'):
+    if encoding is not None:
+        try:
+            string = string.encode(encoding)
+        except AttributeError:
+            # In Python3, bytes has no encode attribute
+            pass
+    string = ctypes.c_char_p(string)
+    return ctypes.cast(string, ctypes.POINTER(ctypes.c_char))
+
+
+
+class FunctionFactoryStub:
+    def __getattr__(self, _):
+      return ctypes.CFUNCTYPE(lambda y:y)
+
+# libraries['FIXME_STUB'] explanation
+# As you did not list (-l libraryname.so) a library that exports this function
+# This is a non-working stub instead.
+# You can either re-run clan2py with -l /path/to/library.so
+# Or manually fix this by comment the ctypes.CDLL loading
+_libraries = {}
+_libraries['FIXME_STUB'] = ctypes.CDLL(ctypes.util.find_library('rocprof-trace-decoder')) #  ctypes.CDLL('FIXME_STUB')
+
+
+
+# values for enumeration 'rocprofiler_thread_trace_decoder_info_t'
+rocprofiler_thread_trace_decoder_info_t__enumvalues = {
+    0: 'ROCPROFILER_THREAD_TRACE_DECODER_INFO_NONE',
+    1: 'ROCPROFILER_THREAD_TRACE_DECODER_INFO_DATA_LOST',
+    2: 'ROCPROFILER_THREAD_TRACE_DECODER_INFO_STITCH_INCOMPLETE',
+    3: 'ROCPROFILER_THREAD_TRACE_DECODER_INFO_WAVE_INCOMPLETE',
+    4: 'ROCPROFILER_THREAD_TRACE_DECODER_INFO_LAST',
+}
+ROCPROFILER_THREAD_TRACE_DECODER_INFO_NONE = 0
+ROCPROFILER_THREAD_TRACE_DECODER_INFO_DATA_LOST = 1
+ROCPROFILER_THREAD_TRACE_DECODER_INFO_STITCH_INCOMPLETE = 2
+ROCPROFILER_THREAD_TRACE_DECODER_INFO_WAVE_INCOMPLETE = 3
+ROCPROFILER_THREAD_TRACE_DECODER_INFO_LAST = 4
+rocprofiler_thread_trace_decoder_info_t = ctypes.c_uint32 # enum
+class struct_rocprofiler_thread_trace_decoder_pc_t(Structure):
+    pass
+
+struct_rocprofiler_thread_trace_decoder_pc_t._pack_ = 1 # source:False
+struct_rocprofiler_thread_trace_decoder_pc_t._fields_ = [
+    ('address', ctypes.c_uint64),
+    ('code_object_id', ctypes.c_uint64),
+]
+
+rocprofiler_thread_trace_decoder_pc_t = struct_rocprofiler_thread_trace_decoder_pc_t
+class struct_rocprofiler_thread_trace_decoder_perfevent_t(Structure):
+    pass
+
+struct_rocprofiler_thread_trace_decoder_perfevent_t._pack_ = 1 # source:False
+struct_rocprofiler_thread_trace_decoder_perfevent_t._fields_ = [
+    ('time', ctypes.c_int64),
+    ('events0', ctypes.c_uint16),
+    ('events1', ctypes.c_uint16),
+    ('events2', ctypes.c_uint16),
+    ('events3', ctypes.c_uint16),
+    ('CU', ctypes.c_ubyte),
+    ('bank', ctypes.c_ubyte),
+    ('PADDING_0', ctypes.c_ubyte * 6),
+]
+
+rocprofiler_thread_trace_decoder_perfevent_t = struct_rocprofiler_thread_trace_decoder_perfevent_t
+class struct_rocprofiler_thread_trace_decoder_occupancy_t(Structure):
+    pass
+
+struct_rocprofiler_thread_trace_decoder_occupancy_t._pack_ = 1 # source:False
+struct_rocprofiler_thread_trace_decoder_occupancy_t._fields_ = [
+    ('pc', rocprofiler_thread_trace_decoder_pc_t),
+    ('time', ctypes.c_uint64),
+    ('reserved', ctypes.c_ubyte),
+    ('cu', ctypes.c_ubyte),
+    ('simd', ctypes.c_ubyte),
+    ('wave_id', ctypes.c_ubyte),
+    ('start', ctypes.c_uint32, 1),
+    ('_rsvd', ctypes.c_uint32, 31),
+]
+
+rocprofiler_thread_trace_decoder_occupancy_t = struct_rocprofiler_thread_trace_decoder_occupancy_t
+
+# values for enumeration 'rocprofiler_thread_trace_decoder_wstate_type_t'
+rocprofiler_thread_trace_decoder_wstate_type_t__enumvalues = {
+    0: 'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_EMPTY',
+    1: 'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_IDLE',
+    2: 'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_EXEC',
+    3: 'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_WAIT',
+    4: 'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_STALL',
+    5: 'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_LAST',
+}
+ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_EMPTY = 0
+ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_IDLE = 1
+ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_EXEC = 2
+ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_WAIT = 3
+ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_STALL = 4
+ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_LAST = 5
+rocprofiler_thread_trace_decoder_wstate_type_t = ctypes.c_uint32 # enum
+class struct_rocprofiler_thread_trace_decoder_wave_state_t(Structure):
+    pass
+
+struct_rocprofiler_thread_trace_decoder_wave_state_t._pack_ = 1 # source:False
+struct_rocprofiler_thread_trace_decoder_wave_state_t._fields_ = [
+    ('type', ctypes.c_int32),
+    ('duration', ctypes.c_int32),
+]
+
+rocprofiler_thread_trace_decoder_wave_state_t = struct_rocprofiler_thread_trace_decoder_wave_state_t
+
+# values for enumeration 'rocprofiler_thread_trace_decoder_inst_category_t'
+rocprofiler_thread_trace_decoder_inst_category_t__enumvalues = {
+    0: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_NONE',
+    1: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_SMEM',
+    2: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_SALU',
+    3: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_VMEM',
+    4: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_FLAT',
+    5: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_LDS',
+    6: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_VALU',
+    7: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_JUMP',
+    8: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_NEXT',
+    9: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_IMMED',
+    10: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_CONTEXT',
+    11: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_MESSAGE',
+    12: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_BVH',
+    13: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_LAST',
+}
+ROCPROFILER_THREAD_TRACE_DECODER_INST_NONE = 0
+ROCPROFILER_THREAD_TRACE_DECODER_INST_SMEM = 1
+ROCPROFILER_THREAD_TRACE_DECODER_INST_SALU = 2
+ROCPROFILER_THREAD_TRACE_DECODER_INST_VMEM = 3
+ROCPROFILER_THREAD_TRACE_DECODER_INST_FLAT = 4
+ROCPROFILER_THREAD_TRACE_DECODER_INST_LDS = 5
+ROCPROFILER_THREAD_TRACE_DECODER_INST_VALU = 6
+ROCPROFILER_THREAD_TRACE_DECODER_INST_JUMP = 7
+ROCPROFILER_THREAD_TRACE_DECODER_INST_NEXT = 8
+ROCPROFILER_THREAD_TRACE_DECODER_INST_IMMED = 9
+ROCPROFILER_THREAD_TRACE_DECODER_INST_CONTEXT = 10
+ROCPROFILER_THREAD_TRACE_DECODER_INST_MESSAGE = 11
+ROCPROFILER_THREAD_TRACE_DECODER_INST_BVH = 12
+ROCPROFILER_THREAD_TRACE_DECODER_INST_LAST = 13
+rocprofiler_thread_trace_decoder_inst_category_t = ctypes.c_uint32 # enum
+class struct_rocprofiler_thread_trace_decoder_inst_t(Structure):
+    pass
+
+struct_rocprofiler_thread_trace_decoder_inst_t._pack_ = 1 # source:False
+struct_rocprofiler_thread_trace_decoder_inst_t._fields_ = [
+    ('category', ctypes.c_uint32, 8),
+    ('stall', ctypes.c_uint32, 24),
+    ('duration', ctypes.c_int32),
+    ('time', ctypes.c_int64),
+    ('pc', rocprofiler_thread_trace_decoder_pc_t),
+]
+
+rocprofiler_thread_trace_decoder_inst_t = struct_rocprofiler_thread_trace_decoder_inst_t
+class struct_rocprofiler_thread_trace_decoder_wave_t(Structure):
+    pass
+
+struct_rocprofiler_thread_trace_decoder_wave_t._pack_ = 1 # source:False
+struct_rocprofiler_thread_trace_decoder_wave_t._fields_ = [
+    ('cu', ctypes.c_ubyte),
+    ('simd', ctypes.c_ubyte),
+    ('wave_id', ctypes.c_ubyte),
+    ('contexts', ctypes.c_ubyte),
+    ('_rsvd1', ctypes.c_uint32),
+    ('_rsvd2', ctypes.c_uint32),
+    ('_rsvd3', ctypes.c_uint32),
+    ('begin_time', ctypes.c_int64),
+    ('end_time', ctypes.c_int64),
+    ('timeline_size', ctypes.c_uint64),
+    ('instructions_size', ctypes.c_uint64),
+    ('timeline_array', ctypes.POINTER(struct_rocprofiler_thread_trace_decoder_wave_state_t)),
+    ('instructions_array', ctypes.POINTER(struct_rocprofiler_thread_trace_decoder_inst_t)),
+]
+
+rocprofiler_thread_trace_decoder_wave_t = struct_rocprofiler_thread_trace_decoder_wave_t
+class struct_rocprofiler_thread_trace_decoder_realtime_t(Structure):
+    pass
+
+struct_rocprofiler_thread_trace_decoder_realtime_t._pack_ = 1 # source:False
+struct_rocprofiler_thread_trace_decoder_realtime_t._fields_ = [
+    ('shader_clock', ctypes.c_int64),
+    ('realtime_clock', ctypes.c_uint64),
+    ('reserved', ctypes.c_uint64),
+]
+
+rocprofiler_thread_trace_decoder_realtime_t = struct_rocprofiler_thread_trace_decoder_realtime_t
+
+# values for enumeration 'rocprofiler_thread_trace_decoder_shaderdata_flags_t'
+rocprofiler_thread_trace_decoder_shaderdata_flags_t__enumvalues = {
+    0: 'ROCPROFILER_THREAD_TRACE_DECODER_SHADERDATA_FLAGS_IMM',
+    1: 'ROCPROFILER_THREAD_TRACE_DECODER_SHADERDATA_FLAGS_PRIV',
+}
+ROCPROFILER_THREAD_TRACE_DECODER_SHADERDATA_FLAGS_IMM = 0
+ROCPROFILER_THREAD_TRACE_DECODER_SHADERDATA_FLAGS_PRIV = 1
+rocprofiler_thread_trace_decoder_shaderdata_flags_t = ctypes.c_uint32 # enum
+class struct_rocprofiler_thread_trace_decoder_shaderdata_t(Structure):
+    pass
+
+struct_rocprofiler_thread_trace_decoder_shaderdata_t._pack_ = 1 # source:False
+struct_rocprofiler_thread_trace_decoder_shaderdata_t._fields_ = [
+    ('time', ctypes.c_int64),
+    ('value', ctypes.c_uint64),
+    ('cu', ctypes.c_ubyte),
+    ('simd', ctypes.c_ubyte),
+    ('wave_id', ctypes.c_ubyte),
+    ('flags', ctypes.c_ubyte),
+    ('reserved', ctypes.c_uint32),
+]
+
+rocprofiler_thread_trace_decoder_shaderdata_t = struct_rocprofiler_thread_trace_decoder_shaderdata_t
+
+# values for enumeration 'rocprofiler_thread_trace_decoder_record_type_t'
+rocprofiler_thread_trace_decoder_record_type_t__enumvalues = {
+    0: 'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_GFXIP',
+    1: 'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_OCCUPANCY',
+    2: 'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_PERFEVENT',
+    3: 'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_WAVE',
+    4: 'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_INFO',
+    5: 'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_DEBUG',
+    6: 'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_SHADERDATA',
+    7: 'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_REALTIME',
+    8: 'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_RT_FREQUENCY',
+    9: 'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_LAST',
+}
+ROCPROFILER_THREAD_TRACE_DECODER_RECORD_GFXIP = 0
+ROCPROFILER_THREAD_TRACE_DECODER_RECORD_OCCUPANCY = 1
+ROCPROFILER_THREAD_TRACE_DECODER_RECORD_PERFEVENT = 2
+ROCPROFILER_THREAD_TRACE_DECODER_RECORD_WAVE = 3
+ROCPROFILER_THREAD_TRACE_DECODER_RECORD_INFO = 4
+ROCPROFILER_THREAD_TRACE_DECODER_RECORD_DEBUG = 5
+ROCPROFILER_THREAD_TRACE_DECODER_RECORD_SHADERDATA = 6
+ROCPROFILER_THREAD_TRACE_DECODER_RECORD_REALTIME = 7
+ROCPROFILER_THREAD_TRACE_DECODER_RECORD_RT_FREQUENCY = 8
+ROCPROFILER_THREAD_TRACE_DECODER_RECORD_LAST = 9
+rocprofiler_thread_trace_decoder_record_type_t = ctypes.c_uint32 # enum
+
+# values for enumeration 'c__EA_rocprofiler_thread_trace_decoder_status_t'
+c__EA_rocprofiler_thread_trace_decoder_status_t__enumvalues = {
+    0: 'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_SUCCESS',
+    1: 'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR',
+    2: 'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR_OUT_OF_RESOURCES',
+    3: 'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR_INVALID_ARGUMENT',
+    4: 'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR_INVALID_SHADER_DATA',
+    5: 'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_LAST',
+}
+ROCPROFILER_THREAD_TRACE_DECODER_STATUS_SUCCESS = 0
+ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR = 1
+ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR_OUT_OF_RESOURCES = 2
+ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR_INVALID_ARGUMENT = 3
+ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR_INVALID_SHADER_DATA = 4
+ROCPROFILER_THREAD_TRACE_DECODER_STATUS_LAST = 5
+c__EA_rocprofiler_thread_trace_decoder_status_t = ctypes.c_uint32 # enum
+rocprofiler_thread_trace_decoder_status_t = c__EA_rocprofiler_thread_trace_decoder_status_t
+rocprofiler_thread_trace_decoder_status_t__enumvalues = c__EA_rocprofiler_thread_trace_decoder_status_t__enumvalues
+rocprof_trace_decoder_trace_callback_t = ctypes.CFUNCTYPE(c__EA_rocprofiler_thread_trace_decoder_status_t, rocprofiler_thread_trace_decoder_record_type_t, ctypes.POINTER(None), ctypes.c_uint64, ctypes.POINTER(None))
+rocprof_trace_decoder_isa_callback_t = ctypes.CFUNCTYPE(c__EA_rocprofiler_thread_trace_decoder_status_t, ctypes.POINTER(ctypes.c_char), ctypes.POINTER(ctypes.c_uint64), ctypes.POINTER(ctypes.c_uint64), struct_rocprofiler_thread_trace_decoder_pc_t, ctypes.POINTER(None))
+rocprof_trace_decoder_se_data_callback_t = ctypes.CFUNCTYPE(ctypes.c_uint64, ctypes.POINTER(ctypes.POINTER(ctypes.c_ubyte)), ctypes.POINTER(ctypes.c_uint64), ctypes.POINTER(None))
+try:
+    rocprof_trace_decoder_parse_data = _libraries['FIXME_STUB'].rocprof_trace_decoder_parse_data
+    rocprof_trace_decoder_parse_data.restype = rocprofiler_thread_trace_decoder_status_t
+    rocprof_trace_decoder_parse_data.argtypes = [rocprof_trace_decoder_se_data_callback_t, rocprof_trace_decoder_trace_callback_t, rocprof_trace_decoder_isa_callback_t, ctypes.POINTER(None)]
+except AttributeError:
+    pass
+try:
+    rocprof_trace_decoder_get_info_string = _libraries['FIXME_STUB'].rocprof_trace_decoder_get_info_string
+    rocprof_trace_decoder_get_info_string.restype = ctypes.POINTER(ctypes.c_char)
+    rocprof_trace_decoder_get_info_string.argtypes = [rocprofiler_thread_trace_decoder_info_t]
+except AttributeError:
+    pass
+try:
+    rocprof_trace_decoder_get_status_string = _libraries['FIXME_STUB'].rocprof_trace_decoder_get_status_string
+    rocprof_trace_decoder_get_status_string.restype = ctypes.POINTER(ctypes.c_char)
+    rocprof_trace_decoder_get_status_string.argtypes = [rocprofiler_thread_trace_decoder_status_t]
+except AttributeError:
+    pass
+rocprofiler_thread_trace_decoder_debug_callback_t = ctypes.CFUNCTYPE(None, ctypes.c_int64, ctypes.POINTER(ctypes.c_char), ctypes.POINTER(ctypes.c_char), ctypes.POINTER(None))
+uint64_t = ctypes.c_uint64
+try:
+    rocprof_trace_decoder_dump_data = _libraries['FIXME_STUB'].rocprof_trace_decoder_dump_data
+    rocprof_trace_decoder_dump_data.restype = rocprofiler_thread_trace_decoder_status_t
+    rocprof_trace_decoder_dump_data.argtypes = [ctypes.POINTER(ctypes.c_char), uint64_t, rocprofiler_thread_trace_decoder_debug_callback_t, ctypes.POINTER(None)]
+except AttributeError:
+    pass
+class union_rocprof_trace_decoder_gfx9_header_t(Union):
+    pass
+
+class struct_rocprof_trace_decoder_gfx9_header_t_0(Structure):
+    pass
+
+struct_rocprof_trace_decoder_gfx9_header_t_0._pack_ = 1 # source:False
+struct_rocprof_trace_decoder_gfx9_header_t_0._fields_ = [
+    ('legacy_version', ctypes.c_uint64, 13),
+    ('gfx9_version2', ctypes.c_uint64, 3),
+    ('DSIMDM', ctypes.c_uint64, 4),
+    ('DCU', ctypes.c_uint64, 5),
+    ('reserved1', ctypes.c_uint64, 1),
+    ('SEID', ctypes.c_uint64, 6),
+    ('reserved2', ctypes.c_uint64, 32),
+]
+
+union_rocprof_trace_decoder_gfx9_header_t._pack_ = 1 # source:False
+union_rocprof_trace_decoder_gfx9_header_t._anonymous_ = ('_0',)
+union_rocprof_trace_decoder_gfx9_header_t._fields_ = [
+    ('_0', struct_rocprof_trace_decoder_gfx9_header_t_0),
+    ('raw', ctypes.c_uint64),
+]
+
+rocprof_trace_decoder_gfx9_header_t = union_rocprof_trace_decoder_gfx9_header_t
+class union_rocprof_trace_decoder_instrument_enable_t(Union):
+    pass
+
+class struct_rocprof_trace_decoder_instrument_enable_t_0(Structure):
+    pass
+
+struct_rocprof_trace_decoder_instrument_enable_t_0._pack_ = 1 # source:False
+struct_rocprof_trace_decoder_instrument_enable_t_0._fields_ = [
+    ('char1', ctypes.c_uint32, 8),
+    ('char2', ctypes.c_uint32, 8),
+    ('char3', ctypes.c_uint32, 8),
+    ('char4', ctypes.c_uint32, 8),
+]
+
+union_rocprof_trace_decoder_instrument_enable_t._pack_ = 1 # source:False
+union_rocprof_trace_decoder_instrument_enable_t._anonymous_ = ('_0',)
+union_rocprof_trace_decoder_instrument_enable_t._fields_ = [
+    ('_0', struct_rocprof_trace_decoder_instrument_enable_t_0),
+    ('u32All', ctypes.c_uint32),
+]
+
+rocprof_trace_decoder_instrument_enable_t = union_rocprof_trace_decoder_instrument_enable_t
+class union_rocprof_trace_decoder_packet_header_t(Union):
+    pass
+
+class struct_rocprof_trace_decoder_packet_header_t_0(Structure):
+    pass
+
+struct_rocprof_trace_decoder_packet_header_t_0._pack_ = 1 # source:False
+struct_rocprof_trace_decoder_packet_header_t_0._fields_ = [
+    ('opcode', ctypes.c_uint32, 8),
+    ('type', ctypes.c_uint32, 4),
+    ('data20', ctypes.c_uint32, 20),
+]
+
+union_rocprof_trace_decoder_packet_header_t._pack_ = 1 # source:False
+union_rocprof_trace_decoder_packet_header_t._anonymous_ = ('_0',)
+union_rocprof_trace_decoder_packet_header_t._fields_ = [
+    ('_0', struct_rocprof_trace_decoder_packet_header_t_0),
+    ('u32All', ctypes.c_uint32),
+]
+
+rocprof_trace_decoder_packet_header_t = union_rocprof_trace_decoder_packet_header_t
+
+# values for enumeration 'rocprof_trace_decoder_packet_opcode_t'
+rocprof_trace_decoder_packet_opcode_t__enumvalues = {
+    4: 'ROCPROF_TRACE_DECODER_PACKET_OPCODE_CODEOBJ',
+    5: 'ROCPROF_TRACE_DECODER_PACKET_OPCODE_RT_TIMESTAMP',
+    6: 'ROCPROF_TRACE_DECODER_PACKET_OPCODE_AGENT_INFO',
+}
+ROCPROF_TRACE_DECODER_PACKET_OPCODE_CODEOBJ = 4
+ROCPROF_TRACE_DECODER_PACKET_OPCODE_RT_TIMESTAMP = 5
+ROCPROF_TRACE_DECODER_PACKET_OPCODE_AGENT_INFO = 6
+rocprof_trace_decoder_packet_opcode_t = ctypes.c_uint32 # enum
+
+# values for enumeration 'rocprof_trace_decoder_agent_info_type_t'
+rocprof_trace_decoder_agent_info_type_t__enumvalues = {
+    0: 'ROCPROF_TRACE_DECODER_AGENT_INFO_TYPE_RT_FREQUENCY_KHZ',
+    1: 'ROCPROF_TRACE_DECODER_AGENT_INFO_TYPE_COUNTER_INTERVAL',
+    2: 'ROCPROF_TRACE_DECODER_AGENT_INFO_TYPE_LAST',
+}
+ROCPROF_TRACE_DECODER_AGENT_INFO_TYPE_RT_FREQUENCY_KHZ = 0
+ROCPROF_TRACE_DECODER_AGENT_INFO_TYPE_COUNTER_INTERVAL = 1
+ROCPROF_TRACE_DECODER_AGENT_INFO_TYPE_LAST = 2
+rocprof_trace_decoder_agent_info_type_t = ctypes.c_uint32 # enum
+class union_rocprof_trace_decoder_codeobj_marker_tail_t(Union):
+    pass
+
+class struct_rocprof_trace_decoder_codeobj_marker_tail_t_0(Structure):
+    pass
+
+struct_rocprof_trace_decoder_codeobj_marker_tail_t_0._pack_ = 1 # source:False
+struct_rocprof_trace_decoder_codeobj_marker_tail_t_0._fields_ = [
+    ('isUnload', ctypes.c_uint32, 1),
+    ('bFromStart', ctypes.c_uint32, 1),
+    ('legacy_id', ctypes.c_uint32, 30),
+]
+
+union_rocprof_trace_decoder_codeobj_marker_tail_t._pack_ = 1 # source:False
+union_rocprof_trace_decoder_codeobj_marker_tail_t._anonymous_ = ('_0',)
+union_rocprof_trace_decoder_codeobj_marker_tail_t._fields_ = [
+    ('_0', struct_rocprof_trace_decoder_codeobj_marker_tail_t_0),
+    ('raw', ctypes.c_uint32),
+]
+
+rocprof_trace_decoder_codeobj_marker_tail_t = union_rocprof_trace_decoder_codeobj_marker_tail_t
+
+# values for enumeration 'rocprof_trace_decoder_codeobj_marker_type_t'
+rocprof_trace_decoder_codeobj_marker_type_t__enumvalues = {
+    0: 'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_TAIL',
+    1: 'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_SIZE_LO',
+    2: 'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ADDR_LO',
+    3: 'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ADDR_HI',
+    4: 'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_SIZE_HI',
+    5: 'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ID_LO',
+    6: 'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ID_HI',
+    7: 'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_LAST',
+}
+ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_TAIL = 0
+ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_SIZE_LO = 1
+ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ADDR_LO = 2
+ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ADDR_HI = 3
+ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_SIZE_HI = 4
+ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ID_LO = 5
+ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ID_HI = 6
+ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_LAST = 7
+rocprof_trace_decoder_codeobj_marker_type_t = ctypes.c_uint32 # enum
+__all__ = \
+    ['ROCPROFILER_THREAD_TRACE_DECODER_INFO_DATA_LOST',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INFO_LAST',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INFO_NONE',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INFO_STITCH_INCOMPLETE',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INFO_WAVE_INCOMPLETE',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INST_BVH',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INST_CONTEXT',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INST_FLAT',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INST_IMMED',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INST_JUMP',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INST_LAST',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INST_LDS',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INST_MESSAGE',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INST_NEXT',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INST_NONE',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INST_SALU',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INST_SMEM',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INST_VALU',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INST_VMEM',
+    'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_DEBUG',
+    'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_GFXIP',
+    'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_INFO',
+    'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_LAST',
+    'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_OCCUPANCY',
+    'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_PERFEVENT',
+    'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_REALTIME',
+    'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_RT_FREQUENCY',
+    'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_SHADERDATA',
+    'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_WAVE',
+    'ROCPROFILER_THREAD_TRACE_DECODER_SHADERDATA_FLAGS_IMM',
+    'ROCPROFILER_THREAD_TRACE_DECODER_SHADERDATA_FLAGS_PRIV',
+    'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR',
+    'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR_INVALID_ARGUMENT',
+    'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR_INVALID_SHADER_DATA',
+    'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR_OUT_OF_RESOURCES',
+    'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_LAST',
+    'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_SUCCESS',
+    'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_EMPTY',
+    'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_EXEC',
+    'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_IDLE',
+    'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_LAST',
+    'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_STALL',
+    'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_WAIT',
+    'ROCPROF_TRACE_DECODER_AGENT_INFO_TYPE_COUNTER_INTERVAL',
+    'ROCPROF_TRACE_DECODER_AGENT_INFO_TYPE_LAST',
+    'ROCPROF_TRACE_DECODER_AGENT_INFO_TYPE_RT_FREQUENCY_KHZ',
+    'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ADDR_HI',
+    'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ADDR_LO',
+    'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ID_HI',
+    'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ID_LO',
+    'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_LAST',
+    'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_SIZE_HI',
+    'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_SIZE_LO',
+    'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_TAIL',
+    'ROCPROF_TRACE_DECODER_PACKET_OPCODE_AGENT_INFO',
+    'ROCPROF_TRACE_DECODER_PACKET_OPCODE_CODEOBJ',
+    'ROCPROF_TRACE_DECODER_PACKET_OPCODE_RT_TIMESTAMP',
+    'c__EA_rocprofiler_thread_trace_decoder_status_t',
+    'rocprof_trace_decoder_agent_info_type_t',
+    'rocprof_trace_decoder_codeobj_marker_tail_t',
+    'rocprof_trace_decoder_codeobj_marker_type_t',
+    'rocprof_trace_decoder_dump_data',
+    'rocprof_trace_decoder_get_info_string',
+    'rocprof_trace_decoder_get_status_string',
+    'rocprof_trace_decoder_gfx9_header_t',
+    'rocprof_trace_decoder_instrument_enable_t',
+    'rocprof_trace_decoder_isa_callback_t',
+    'rocprof_trace_decoder_packet_header_t',
+    'rocprof_trace_decoder_packet_opcode_t',
+    'rocprof_trace_decoder_parse_data',
+    'rocprof_trace_decoder_se_data_callback_t',
+    'rocprof_trace_decoder_trace_callback_t',
+    'rocprofiler_thread_trace_decoder_debug_callback_t',
+    'rocprofiler_thread_trace_decoder_info_t',
+    'rocprofiler_thread_trace_decoder_inst_category_t',
+    'rocprofiler_thread_trace_decoder_inst_t',
+    'rocprofiler_thread_trace_decoder_occupancy_t',
+    'rocprofiler_thread_trace_decoder_pc_t',
+    'rocprofiler_thread_trace_decoder_perfevent_t',
+    'rocprofiler_thread_trace_decoder_realtime_t',
+    'rocprofiler_thread_trace_decoder_record_type_t',
+    'rocprofiler_thread_trace_decoder_shaderdata_flags_t',
+    'rocprofiler_thread_trace_decoder_shaderdata_t',
+    'rocprofiler_thread_trace_decoder_status_t',
+    'rocprofiler_thread_trace_decoder_status_t__enumvalues',
+    'rocprofiler_thread_trace_decoder_wave_state_t',
+    'rocprofiler_thread_trace_decoder_wave_t',
+    'rocprofiler_thread_trace_decoder_wstate_type_t',
+    'struct_rocprof_trace_decoder_codeobj_marker_tail_t_0',
+    'struct_rocprof_trace_decoder_gfx9_header_t_0',
+    'struct_rocprof_trace_decoder_instrument_enable_t_0',
+    'struct_rocprof_trace_decoder_packet_header_t_0',
+    'struct_rocprofiler_thread_trace_decoder_inst_t',
+    'struct_rocprofiler_thread_trace_decoder_occupancy_t',
+    'struct_rocprofiler_thread_trace_decoder_pc_t',
+    'struct_rocprofiler_thread_trace_decoder_perfevent_t',
+    'struct_rocprofiler_thread_trace_decoder_realtime_t',
+    'struct_rocprofiler_thread_trace_decoder_shaderdata_t',
+    'struct_rocprofiler_thread_trace_decoder_wave_state_t',
+    'struct_rocprofiler_thread_trace_decoder_wave_t', 'uint64_t',
+    'union_rocprof_trace_decoder_codeobj_marker_tail_t',
+    'union_rocprof_trace_decoder_gfx9_header_t',
+    'union_rocprof_trace_decoder_instrument_enable_t',
+    'union_rocprof_trace_decoder_packet_header_t']
@@ -1,112 +0,0 @@
-import os
-os.environ["PYTHONPATH"] = "."
-os.environ["SQTT"] = "1"
-if "DEV" not in os.environ: os.environ["DEV"] = "AMD"
-os.environ["VIZ"] = "1"
-os.environ["AMD_LLVM"] = "0"
-
-import unittest
-import sys, contextlib
-from tinygrad import Tensor
-from tinygrad.dtype import dtypes
-from tinygrad.renderer import ProgramSpec
-from tinygrad.uop.ops import UOp, Ops, KernelInfo, AddrSpace
-from tinygrad.engine.realize import CompiledRunner
-from tinygrad.device import Device, ProfileDeviceEvent
-
-from extra.sqtt.roc import decode, WaveExec
-
-dev = Device[os.environ["DEV"]]
-
-def custom(arg:str, s:UOp|None=None) -> UOp: return UOp(Ops.CUSTOM, src=(s,) if s is not None else (), arg=arg)
-
-def asm_kernel(instrs:list[str], l:int=1, g:int=1) -> Tensor:
-  name = sys._getframe(1).f_code.co_name
-  def fxn(_):
-    L = UOp.special(l, "lidx0")
-    G = UOp.special(g, "gidx0")
-    op = custom("asm volatile (")
-    for inst in instrs: op = custom(f'  "{inst}\\n\\t"', op)
-    op = custom(");", op)
-    return UOp.sink(op, L, G, arg=KernelInfo(name=name))
-  k = Tensor.custom_kernel(Tensor.empty(1), fxn=fxn)[0]
-  return k
-
-@contextlib.contextmanager
-def save_sqtt():
-  # clear the old traces
-  dev.profile_events.clear()
-  sqtt:dict[str, list[WaveExec]] = {}
-  yield sqtt
-  # decode sqtt
-  if os.environ["DEV"] == "AMD":
-    rctx = decode(dev.profile_events+[ProfileDeviceEvent("AMD", props=dev.device_props())])
-    assert len(rctx.inst_execs) > 0, "empty sqtt output"
-    sqtt.update(rctx.inst_execs)
-
-class TestTiming(unittest.TestCase):
-  def test_v_add(self):
-    with save_sqtt() as sqtt:
-      asm_kernel([f"v_add_f32 v{10+i} v{10+i+1} {10+i}" for i in range(3)]).realize()
-    wave = list(sqtt.values())[0][:-1]
-    assert all(s.dur == 1 for s in wave)
-    assert all(s.stall == 0 for s in wave)
-
-  def test_chain_v_add_1l(self):
-    with save_sqtt() as sqtt:
-      asm_kernel([
-        "v_add_f32_e32 v1 v0 v0",
-        "v_add_f32_e32 v2 v1 v1",
-      ]).realize()
-    wave = list(sqtt.values())[0][:-1]
-    assert all(s.dur == 1 for s in wave)
-    assert all(s.stall == 0 for s in wave)
-
-  def test_multi_cycle_inst(self):
-    def custom_vrcp(A, B):
-      op = custom("float a = 0.0;")
-      op = custom("float b = (*(data1_1+0));", op)
-      #op = custom('asm volatile("v_mul_f32_e32 %2 %2 %1" : "+v"(a) : "v"(b));', op)
-      op = custom('asm volatile("v_rcp_f32_e32 %2 %1" : "+v"(a) : "v"(b));', op)
-      op = custom('asm volatile("v_add_f32_e64 %1 %1 1.0" : "+v"(a));', op)
-      op = custom("*(data0_1+0) = a;", op)
-      return UOp.sink(op, A, B, arg=KernelInfo(name="custom_vrcp"))
-    out = Tensor([0.]).realize()
-    inp = Tensor([-2.0]).realize()
-    with save_sqtt() as sqtt:
-      Tensor.custom_kernel(out, inp, fxn=custom_vrcp)[0].realize()
-
-    wave = list(sqtt.values())[0][0]
-    for i in range(len(wave.insts)):
-      if wave.insts[i].inst.startswith("global_store"):
-        print(f"store diff {wave.insts[i].time-(wave.insts[i-1].time)}")
-    self.assertEqual(out.item(), 0.5)
-
-  def test_wmma(self):
-    with save_sqtt() as sqtt:
-      asm_kernel([
-        "v_wmma_f32_16x16x16_f16 v[16:23], v[0:7], v[8:15], v[16:23]",
-        "v_add_f32_e32 v0 v16 v0",
-      ], l=32*4).realize()
-    assert len(sqtt) == 2, f"expected two waves, got {len(sqtt)} {list(sqtt.keys())}"
-    wmma = list(sqtt.values())[0][0]
-    self.assertGreater(wmma.dur, 1) # rgp says 32 clocks
-
-  def test_sleep(self):
-    n = 1
-    def sleep_kernel(data0):
-      assert data0.dtype.base == dtypes.ulong
-      op = custom("unsigned long long t0 = __builtin_readcyclecounter();")
-      op = custom(f"__builtin_amdgcn_s_sleep({n});", op)
-      op = custom(f"unsigned long long t1 = __builtin_readcyclecounter();", op)
-      op = custom(f"data0_{data0.size}[0] = t1 - t0;", op)
-      return UOp.sink(data0, op, arg=KernelInfo(name=f"sleep_{n}"))
-    diff_hw_reg = Tensor.empty(1, dtype=dtypes.ulong)
-    diff_hw_reg = Tensor.custom_kernel(diff_hw_reg, fxn=sleep_kernel)[0]
-    with save_sqtt() as sqtt:
-      diff_hw_reg.realize()
-    diff_sqtt = list(sqtt.values())[0][2]
-    self.assertEqual(diff_sqtt.dur, diff_hw_reg.item()-1) # 1 cycle for reading the counter register
-
-if __name__ == "__main__":
-  unittest.main()
@@ -10,7 +10,7 @@ constexpr int ATTN_N = 1024;
 constexpr int ATTN_H = 16;
 constexpr int ATTN_D = 64;

-template<int D> constexpr size_t ROWS = 16*(64/D); // height of each worker tile (rows)
+template<int D> constexpr size_t ROWS = 16*(128/D); // height of each worker tile (rows)
 template<int D, typename T=bf16, typename L=row_l> using qkvo_tile = rt<T, ROWS<D>, D, L>;
 template<int D, typename T=float> using attn_tile = rt<T, ROWS<D>, ROWS<D>>;
 template<int D> using shared_tile = st_bf<ROWS<D>, D>;
@@ -23,7 +23,7 @@ if __name__ == "__main__":
  Tensor.realize(q, k, v, out)

  NUM_WORKERS = 4
-  ROWS = 16 * (64 // D)
+  ROWS = 16 * (128 // D)

  gsz = (N // (ROWS*NUM_WORKERS), H, B)
  for _ in range(5):
@@ -1 +0,0 @@
-WARP_THREADS = 32
@@ -1,277 +0,0 @@
-import math, functools
-from typing import cast, Callable
-from tinygrad import Tensor, Device, Context, GlobalCounters, dtypes
-from tinygrad.uop.ops import AxisType, UOp, KernelInfo, Ops
-from tinygrad.engine.realize import ExecItem, get_runner
-from tinygrad.dtype import AddrSpace, PtrDType
-from tinygrad.helpers import getenv, prod
-
-from extra.thunder.tiny.tk import WARP_THREADS
-from extra.thunder.tiny.tk.tiles import ALL_TILES, GL, ST, RT, RV
-
-class Group:
-  def __init__(self, warps:int, ker):
-    self.warps = warps
-    self.group_threads = warps * WARP_THREADS
-    self.threadIdx_x = ker.threadIdx_x
-    self.ker = ker
-
-  # helpers
-  @property
-  def laneid(self): return self.threadIdx_x % self.group_threads
-  @property
-  def warpid(self): return self.laneid // WARP_THREADS
-  @property
-  def groupid(self): return self.threadIdx_x // self.group_threads
-
-  # ops that only work on a single warp
-
-  clear_rid = 1000
-  def clear(self, reg:ALL_TILES, value:float=0):
-    reg = cast(UOp, reg)
-    assert self.warps == 1
-
-    rngs_for_shape = tuple(UOp.range(dim, Group.clear_rid + i) for i, dim in enumerate(reg.shape))
-    Group.clear_rid += len(reg.shape)
-
-    reg_store = reg[*rngs_for_shape].store(value).end(*rngs_for_shape)
-
-    self.ker.push_store(reg_store, reg)
-    return reg.after(reg_store).reshape(reg.shape)
-
-  def zero(self, reg:ALL_TILES): return self.clear(reg, 0)
-  def neg_inf(self, reg:ALL_TILES): return self.clear(reg, -math.inf)
-
-  copy_rid = 300
-  def copy(self, dst:ALL_TILES, src:ALL_TILES):
-    dst, src = cast(UOp, dst), cast(UOp, src)
-    assert self.warps == 1
-    assert dst.shape == src.shape
-
-    rngs_for_shape = tuple(UOp.range(dim, Group.copy_rid + i) for i, dim in enumerate(dst.shape))
-    Group.copy_rid += len(dst.shape)
-
-    dst_store = dst[*rngs_for_shape].store(src[*rngs_for_shape].cast(dst.dtype.base)).end(*rngs_for_shape)
-
-    self.ker.push_store(dst_store, dst)
-    return dst.after(dst_store).reshape(dst.shape)
-
-  def mma_AB(self, c:UOp|RT, a:UOp|RT, b:UOp|RT, after=True):
-    c, a, b = cast(UOp, c), cast(UOp, a), cast(UOp, b)
-    assert self.warps == 1
-
-    for height in self.ker.range(c.shape[-3], track=False):
-      for width in self.ker.range(c.shape[-2], track=False):
-        for inner in self.ker.range(a.shape[-2], AxisType.REDUCE, track=False):
-          wmma_arg = ("WMMA_8_16_16_bfloat16_float", (8, 16, 16), dtypes.bfloat16, dtypes.float, "CUDA", 32, (((4, 2), (3, 2), (8, 2)), ((4, 2), (3, 2)), ((4, 2), (3, 2))), ())
-
-          a_in = UOp.vectorize(*[a[height, inner, i] for i in range(8)])
-          b_in1 = UOp.vectorize(*([b[inner, width, i] for i in range(2)] + [b[inner, width, 4+i] for i in range(2)]))
-          c_out1 = UOp.vectorize(*[c[height, width, i] for i in range(4)])
-          b_in2 = UOp.vectorize(*([b[inner, width, 2+i] for i in range(2)] + [b[inner, width, 6+i] for i in range(2)]))
-          c_out2 = UOp.vectorize(*[c[height, width, 4+i] for i in range(4)])
-
-          out1 = UOp(Ops.WMMA, dtypes.float32.vec(4), (a_in, b_in1, c_out1), arg=wmma_arg)
-          out2 = UOp(Ops.WMMA, dtypes.float32.vec(4), (a_in, b_in2, c_out2), arg=wmma_arg)
-          c_i = [c[height, width, i].store(out1.gep(i)) for i in range(4)] + [c[height, width, 4+i].store(out2.gep(i)) for i in range(4)]
-          c_store = UOp.group(*c_i).end(height, width, inner)
-
-    self.ker.push_store(c_store, c)
-    return c.after(c_store).reshape(c.shape) if after else c_store
-
-  def mma_ABt(self, c:UOp|RT, a:UOp|RT, b:UOp|RT, after=True):
-    c, a, b = cast(UOp, c), cast(UOp, a), cast(UOp, b)
-    assert self.warps == 1
-
-    for height in self.ker.range(c.shape[-3], track=False):
-      for width in self.ker.range(c.shape[-2], track=False):
-        for inner in self.ker.range(a.shape[-2], AxisType.REDUCE, track=False):
-          wmma_arg = ("WMMA_8_16_16_bfloat16_float", (8, 16, 16), dtypes.bfloat16, dtypes.float, "CUDA", 32, (((4, 2), (3, 2), (8, 2)), ((4, 2), (3, 2)), ((4, 2), (3, 2))), ())
-
-          a_in = UOp.vectorize(*[a[height, inner, i] for i in range(8)])
-          b_in1 = UOp.vectorize(*([b[width, inner, i] for i in range(2)] + [b[width, inner, 4+i] for i in range(2)]))
-          c_out1 = UOp.vectorize(*[c[height, width, i] for i in range(4)])
-          b_in2 = UOp.vectorize(*([b[width, inner, 2+i] for i in range(2)] + [b[width, inner, 6+i] for i in range(2)]))
-          c_out2 = UOp.vectorize(*[c[height, width, 4+i] for i in range(4)])
-
-          out1 = UOp(Ops.WMMA, dtypes.float32.vec(4), (a_in, b_in1, c_out1), arg=wmma_arg)
-          out2 = UOp(Ops.WMMA, dtypes.float32.vec(4), (a_in, b_in2, c_out2), arg=wmma_arg)
-          c_i = [c[height, width, i].store(out1.gep(i)) for i in range(4)] + [c[height, width, 4+i].store(out2.gep(i)) for i in range(4)]
-          c_store = UOp.group(*c_i).end(height, width, inner)
-
-    self.ker.push_store(c_store, c)
-    return c.after(c_store).reshape(c.shape) if after else c_store
-
-  map_rid = 400
-  def map(self, a:ALL_TILES, op:Callable[[UOp], UOp]|Callable[[UOp, tuple], UOp]):
-    a = cast(UOp, a)
-    assert self.warps == 1
-
-    rngs_for_shape = tuple(UOp.range(dim, Group.map_rid + i) for i, dim in enumerate(a.shape))
-    Group.map_rid += len(a.shape)
-
-    if op.__code__.co_argcount == 1:
-      to_store = op(a[*rngs_for_shape])
-    else:
-      to_store = op(a[*rngs_for_shape], rngs_for_shape)
-
-    a_store = a[*rngs_for_shape].store(to_store).end(*rngs_for_shape)
-
-    self.ker.push_store(a_store, a)
-    return a.after(a_store).reshape(a.shape)
-
-  def row_reduce(self, vec:UOp|RV, src:UOp|RT, op:Callable[[UOp, UOp], UOp]):
-    vec, src = cast(UOp, vec), cast(UOp, src)
-    assert self.warps == 1
-
-    red_local = self.ker.alloc((self.group_threads, 2), src.dtype.base, AddrSpace.LOCAL)
-    red_reg = self.ker.alloc((2,), src.dtype.base, AddrSpace.REG)
-
-    for height in self.ker.range(src.shape[-3], track=False):
-      i = UOp.range(red_reg.size, Group.clear_rid)
-      Group.clear_rid += 1
-      red_reg = red_reg.after(height, *[tkr._rng for tkr in self.ker.range_stack])
-      reg_store = red_reg.flatten()[i].store(0.).end(i)
-      red_reg = red_reg.after(reg_store).reshape(red_reg.shape)
-
-      for outer in self.ker.range(2, track=False):
-        for width in self.ker.range(src.shape[-2], AxisType.REDUCE, track=False):
-          for inner in self.ker.range(4, AxisType.REDUCE, track=False):
-            elem_index = inner + 2 * (inner // 2) + outer * 2
-            reg_store = red_reg[outer].store(op(red_reg[outer], src[height, width, elem_index])).end(inner, width, outer)
-            red_reg = red_reg.after(reg_store).reshape(red_reg.shape)
-
-      # store to shared memory
-      for outer in self.ker.range(2, track=False):
-        red_local_store = red_local[self.laneid, outer].store(red_reg[outer]).end(outer)
-        red_local = red_local.after(red_local_store.barrier()).reshape(red_local.shape)
-
-      # reduce from shared memory
-      for outer in self.ker.range(2, track=False):
-        for inner in self.ker.range(3, AxisType.REDUCE, track=False):
-          offset = (self.laneid // 4) * 4 + ((self.laneid + inner + 1) % 4)
-          reg_store = red_reg[outer].store(op(red_reg[outer], red_local[offset, outer])).end(inner, outer)
-          red_reg = red_reg.after(reg_store).reshape(red_reg.shape)
-
-      # reduce with vec
-      for outer in self.ker.range(2, track=False):
-        vec_store = vec[height, 0, outer].store(op(vec[height, 0, outer], red_reg[outer])).end(outer, height)
-
-    self.ker.push_store(vec_store, vec)
-    return vec.after(vec_store).reshape(vec.shape)
-
-  # ops that can work across multiple warps
-
-  LOAD_INNER = 8
-  def load(self, dst:ALL_TILES, src:ALL_TILES, dst_idxs:tuple[UOp|int,...]=(), idxs:tuple[UOp|int,...]=(), axis:int=0, transpose:bool=False):
-    dst, src = cast(UOp, dst), cast(UOp, src)
-    assert isinstance(dst.dtype, PtrDType) and isinstance(src.dtype, PtrDType)
-    dst_dtype, src_dtype = cast(PtrDType, dst.dtype), cast(PtrDType, src.dtype)
-    if dst_dtype.addrspace == AddrSpace.REG and src_dtype.addrspace == AddrSpace.LOCAL:
-      srcf = src.flatten(-2)
-
-      if self.warps % 4 == 0: local_warpid = (self.warpid // 4) + (self.warpid % 4) * (self.warps // 4)
-      else: local_warpid = self.warpid
-      warp_laneid = self.threadIdx_x % WARP_THREADS
-
-      for height in self.ker.range(dst.shape[-3], track=False):
-        for width in self.ker.range(dst.shape[-2], track=False):
-          for inner in self.ker.range(RT.BASE_TILE_NEPT, track=False):
-            if not transpose:
-              row = (local_warpid * dst.shape[-3] + height) * RT.BASE_TILE_ROWS + (warp_laneid // 4)
-              col = width * RT.BASE_TILE_COLS + 2 * (warp_laneid % 4)
-
-              row_offset = ((inner % 4) // 2) * 8
-              col_offset = (inner % 2) + (inner // 4) * 8
-            else:
-              row = (local_warpid * dst.shape[-3] + height) * RT.BASE_TILE_ROWS + 2 * (warp_laneid % 4)
-              col = width * RT.BASE_TILE_COLS + (warp_laneid // 4)
-
-              row_offset = (inner % 2) + (inner // 4) * 8
-              col_offset = ((inner % 4) // 2) * 8
-
-            src_i_last = (row + row_offset) * src.shape[-1] + col + col_offset
-
-            dst_store = dst[*dst_idxs, height, width, inner].store(srcf[*idxs[:-2], src_i_last])
-            dst_store = dst_store.end(height, width, inner)
-    elif dst_dtype.addrspace == AddrSpace.LOCAL and src_dtype.addrspace == AddrSpace.GLOBAL:
-      dstf = dst.flatten(-2)
-
-      srcf = src.flatten()
-      row_stride = prod(src.shape[axis+1:])
-
-      idxs = tuple(idx * dst.shape[-2] if i == axis else idx for i, idx in enumerate(idxs))
-      idxs = tuple(idx * dst.shape[-1] if i == 3 else idx for i, idx in enumerate(idxs))
-      src_i = ((idxs[0] * src.shape[-3] + idxs[1]) * src.shape[-2] + idxs[2]) * src.shape[-1] + idxs[3]
-
-      memcpy_per_row = dst.shape[-1] // Group.LOAD_INNER
-      total_calls = prod(dst.shape[-2:]) // (self.group_threads * Group.LOAD_INNER)
-
-      for outer in self.ker.range(total_calls, track=False):
-        for inner in self.ker.range(Group.LOAD_INNER, track=False):
-          load_idx = outer * self.group_threads + self.laneid
-          row = load_idx // memcpy_per_row
-          col = (load_idx * Group.LOAD_INNER) % dst.shape[-1]
-
-          dst_i = row * dst.shape[-1] + col + inner
-          src_i += row * row_stride + col + inner
-
-          dst_store = dstf[*dst_idxs, dst_i].store(srcf[src_i]).end(outer, inner)
-    else:
-      raise NotImplementedError(f"load from {src_dtype.addrspace} to {dst_dtype.addrspace} not implemented")
-
-    return dst.after(dst_store.barrier()).reshape(dst.shape)
-
-  STORE_INNER = 8
-  def store(self, dst:ALL_TILES, src:ALL_TILES, idxs:tuple[UOp|int,...]=(), src_idxs:tuple[UOp|int,...]=(), axis=0, after=True):
-    dst, src = cast(UOp, dst), cast(UOp, src)
-    assert isinstance(dst.dtype, PtrDType) and isinstance(src.dtype, PtrDType)
-    dst_dtype, src_dtype = cast(PtrDType, dst.dtype), cast(PtrDType, src.dtype)
-    if src_dtype.addrspace == AddrSpace.REG and dst_dtype.addrspace == AddrSpace.LOCAL:
-      dstf = dst.flatten(-2)
-
-      if self.warps % 4 == 0: local_warpid = (self.warpid // 4) + (self.warpid % 4) * (self.warps // 4)
-      else: local_warpid = self.warpid
-      warp_laneid = self.threadIdx_x % WARP_THREADS
-
-      for height in self.ker.range(src.shape[-3], track=False):
-        for width in self.ker.range(src.shape[-2], track=False):
-          for inner in self.ker.range(RT.BASE_TILE_NEPT, track=False):
-            row = (local_warpid * src.shape[-3] + height) * RT.BASE_TILE_ROWS + (warp_laneid // 4)
-            col = width * RT.BASE_TILE_COLS + 2 * (warp_laneid % 4)
-
-            row_offset = ((inner % 4) // 2) * 8
-            col_offset = (inner % 2) + (inner // 4) * 8
-
-            dst_i_last = (row + row_offset) * dst.shape[-1] + col + col_offset
-
-            dst_store = dstf[*idxs[:-2], dst_i_last].store(src[*src_idxs, height, width, inner])
-            dst_store = dst_store.end(height, width, inner)
-    elif src_dtype.addrspace == AddrSpace.LOCAL and dst_dtype.addrspace == AddrSpace.GLOBAL:
-      dstf = dst.flatten()
-      row_stride = prod(dst.shape[axis+1:])
-
-      idxs = tuple(idx * src.shape[-2] if i == axis else idx for i, idx in enumerate(idxs))
-      idxs = tuple(idx * src.shape[-1] if i == 3 else idx for i, idx in enumerate(idxs))
-      dst_i = ((idxs[0] * dst.shape[-3] + idxs[1]) * dst.shape[-2] + idxs[2]) * dst.shape[-1] + idxs[3]
-
-      srcf = src.flatten(-2)
-
-      memcpy_per_row = src.shape[-1] // Group.STORE_INNER
-      total_calls = prod(src.shape[-2:]) // (self.group_threads * Group.STORE_INNER)
-
-      for outer in self.ker.range(total_calls, track=False):
-        for inner in self.ker.range(Group.STORE_INNER, track=False):
-          load_idx = outer * self.group_threads + self.laneid
-          row = load_idx // memcpy_per_row
-          col = (load_idx * Group.STORE_INNER) % src.shape[-1]
-
-          src_i = row * src.shape[-1] + col + inner
-          dst_i += row * row_stride + col + inner
-
-          dst_store = dstf[dst_i].store(srcf[*src_idxs, src_i]).end(outer, inner)
-    else:
-      raise NotImplementedError(f"store from {src_dtype.addrspace} to {dst_dtype.addrspace} not implemented")
-
-    self.ker.push_store(dst_store, dst)
-    return dst.after(dst_store.barrier()).reshape(dst.shape) if after else dst_store
@@ -1,88 +0,0 @@
-from contextlib import AbstractContextManager
-from tinygrad.uop.ops import UOp, KernelInfo, AxisType, AddrSpace
-from extra.thunder.tiny.tk import WARP_THREADS
-from extra.thunder.tiny.tk.group import Group
-from extra.thunder.tiny.tk.tiles import GL, ST, RT, RV
-
-class _tk_range:
-  user_rid = 0
-  def __init__(self, end:int, axis_type:AxisType): self.end, self.axis_type, self.done = end, axis_type, False
-  def __iter__(self): return self
-  def __next__(self):
-    if not self.done:
-      self.done = True
-      _tk_range.user_rid += 1
-      self._rng = UOp.range(self.end, _tk_range.user_rid-1, axis_type=self.axis_type)
-      return self._rng
-    raise StopIteration
-
-class Kernel(AbstractContextManager):
-  def __init__(self, grid_size:tuple[int, int, int], block_size:int):
-    self.blockIdx_x = UOp.special(grid_size[0], "gidx0")
-    self.blockIdx_y = UOp.special(grid_size[1], "gidx1")
-    self.blockIdx_z = UOp.special(grid_size[2], "gidx2")
-    self.threadIdx_x = UOp.special(block_size, "lidx0")
-
-    self.range_stack = []
-    self.store_stack = []
-
-    self.global_slot = 0
-    self.shared_slot = 0
-    self.register_slot = 0
-    self.allocs = {}
-
-  @property
-  def warpid(self): return self.threadIdx_x // WARP_THREADS
-
-  def __enter__(self): return self
-  def __exit__(self, exc_type, exc_value, traceback): pass
-
-  def group(self, size:int): return Group(size, self)
-  @property
-  def warp(self): return self.group(1)
-  @property
-  def warpgroup(self): return self.group(4)
-
-  def range(self, end:int, axis_type:AxisType=AxisType.LOOP, track:bool=True):
-    rng = _tk_range(end, axis_type)
-    if track: self.range_stack.append(rng)
-    return rng
-
-  def alloc(self, shape, dtype, addrspace:AddrSpace, name:str|None=None):
-    match addrspace:
-      case AddrSpace.GLOBAL:
-        slot = self.global_slot
-        self.global_slot += 1
-      case AddrSpace.LOCAL:
-        slot = self.shared_slot
-        self.shared_slot += 1
-      case AddrSpace.REG:
-        slot = self.register_slot
-        self.register_slot += 1
-
-    uop = UOp.placeholder(shape, dtype, slot=slot, addrspace=addrspace)
-
-    if name:
-      if (name, shape) in self.allocs: return self.allocs[(name, shape)]
-      self.allocs[(name, shape)] = uop
-
-    return uop
-
-  def gl(self, shape, dtype): return GL.create(shape, dtype, self)
-  def st(self, shape, dtype): return ST.create(shape, dtype, self)
-  def rt(self, shape, dtype): return RT.create(shape, dtype, self)
-  def rv(self, length, dtype, layout="naive"): return RV.create(length, dtype, layout, self)
-
-  def push_store(self, store:UOp, uop:UOp): self.store_stack.append((store, uop))
-
-  def finish(self):
-    # end all ranges
-    rngs = []
-    while self.range_stack: rngs.append(self.range_stack.pop(0)._rng)
-
-    return self.store_stack.pop()[0]._uop.end(*rngs).sink(arg=KernelInfo(opts_to_apply=())).simplify()
-
-  def endrange(self):
-    last_store = self.store_stack.pop()
-    last_range = self.range_stack.pop()
-    return last_store[1].after(last_store[0].barrier().end(last_range._rng)).reshape(last_store[1].shape)
@@ -1,143 +0,0 @@
-import functools
-from tinygrad.dtype import AddrSpace
-from tinygrad.mixin import MathMixin
-from tinygrad.uop.ops import UOp, Ops
-
-from extra.thunder.tiny.tk import WARP_THREADS
-
-def unwrap(x):
-  if hasattr(x, "_uop"): return x._uop
-  if isinstance(x, (list, tuple)): return type(x)(unwrap(y) for y in x)
-  if isinstance(x, dict): return {k: unwrap(v) for k,v in x.items()}
-  return x
-
-def wrap(x, ker, cls):
-  if isinstance(x, UOp): return cls(x, ker)
-  if isinstance(x, (list, tuple)): return type(x)(wrap(y, ker, cls) for y in x)
-  return x
-
-def autowrap(source_cls, blacklist=None):
-  if blacklist is None:
-    blacklist = {
-      "__init__", "__new__", "__str__", "__del__", "__repr__", "__dict__", "__getattribute__",
-      "__setattr__", "__delattr__", "__weakref__", "__slots__", "__class__",
-      "__reduce__", "__reduce_ex__", "__getstate__", "__setstate__", "__hash__"
-    }
-
-  def decorator(cls):
-    def __getattr__(self, name):
-      uop = object.__getattribute__(self, "_uop")
-      val = getattr(uop, name)
-      if callable(val):
-        @functools.wraps(val)
-        def proxy(*args, **kwargs):
-          return wrap(val(*unwrap(args), **unwrap(kwargs)), self.ker, cls)
-        return proxy
-      if name in UOp.__slots__: return val
-      return wrap(val, self.ker, cls)
-    cls.__getattr__ = __getattr__
-
-    for name in dir(source_cls):
-      if name in blacklist or not name.startswith("__"): continue
-
-      for base in cls.mro():
-        if base is source_cls: break
-        if name in base.__dict__: break
-      else:
-        original = getattr(source_cls, name)
-        if callable(original):
-          def make_proxy(op_name, func):
-            def proxy(self, *args, **kwargs):
-              return wrap(func(self._uop, *unwrap(args), **unwrap(kwargs)), self.ker, cls)
-            return proxy
-          setattr(cls, name, make_proxy(name, original))
-
-    return cls
-  return decorator
-
-class TileMathMixin(MathMixin):
-  def alu(self, op, *src, inner_op=lambda x:x):
-    assert isinstance(self, (RT, RV))
-    if len(src) == 0:
-      if self._uop._shape is None: uop = UOp.alu(self._uop, op)
-      else: uop = self.ker.warp.map(self._uop, lambda x: UOp.alu(x, op))
-    elif len(src) == 1:
-      if self._uop._shape is None: uop = UOp.alu(self._uop, op, inner_op(self._uop.ufix(src[0])))
-      elif isinstance(src[0], (int,float,bool)): uop = self.ker.warp.map(self._uop, lambda x: UOp.alu(x, op, inner_op(x.ufix(src[0]))))
-      elif src[0]._shape is None: uop = UOp.alu(self._uop, op, inner_op(self._uop.ufix(src[0])))
-      else:
-        if isinstance(self, RT) and isinstance(src[0], RV): uop = self.ker.warp.map(self._uop, lambda x, idx: UOp.alu(x, op, inner_op(src[0]._uop[idx[0], 0, (idx[2]%4)//2])))
-        else: uop = self.ker.warp.map(self._uop, lambda x, idx: UOp.alu(x, op, inner_op(src[0]._uop[*idx])))
-    else: raise NotImplementedError
-    return type(self)(uop, self.ker)
-  def const_like(self, b): return b
-
-  # override ops that do compute on the src uop
-  def sub(self, x, reverse=False):
-    return self.ufix(x).alu(Ops.ADD, self, inner_op=lambda y: -y) if reverse else self.alu(Ops.ADD, self.ufix(x), inner_op=lambda y: -y)
-  def div(self, x, reverse=False):
-    return self.ufix(x).alu(Ops.MUL, self, inner_op=lambda y: 1/y) if reverse else self.alu(Ops.MUL, self.ufix(x), inner_op=lambda y: 1/y)
-
-@autowrap(UOp)
-class GL:
-  def __init__(self, uop, ker):
-    self._uop, self.ker = uop, ker
-
-  @classmethod
-  def create(cls, shape, dtype, ker):
-    uop = ker.alloc(shape, dtype, AddrSpace.GLOBAL)
-    return cls(uop, ker)
-
-@autowrap(UOp)
-class ST:
-  def __init__(self, uop, ker):
-    self._uop, self.ker = uop, ker
-
-  @classmethod
-  def create(cls, shape, dtype, ker):
-    uop = ker.alloc(shape, dtype, AddrSpace.LOCAL)
-    return cls(uop, ker)
-
-@autowrap(UOp)
-class RT(TileMathMixin):
-  BASE_TILE_ROWS, BASE_TILE_COLS = 16, 16
-  BASE_TILE_NE = BASE_TILE_ROWS * BASE_TILE_COLS
-  BASE_TILE_NEPT = BASE_TILE_NE // WARP_THREADS
-
-  def __init__(self, uop, ker):
-    self._uop, self.ker = uop, ker
-
-  @classmethod
-  def create(cls, shape, dtype, ker):
-    assert len(shape) == 2
-    assert shape[0] % RT.BASE_TILE_ROWS == 0
-    assert shape[1] % RT.BASE_TILE_COLS == 0
-
-    height = shape[0] // RT.BASE_TILE_ROWS
-    width = shape[1] // RT.BASE_TILE_COLS
-
-    uop = ker.alloc((height, width, RT.BASE_TILE_NEPT), dtype, AddrSpace.REG)
-    return cls(uop, ker)
-
-@autowrap(UOp)
-class RV(TileMathMixin):
-  def __init__(self, uop, ker):
-    self._uop, self.ker = uop, ker
-
-  @classmethod
-  def create(cls, length, dtype, layout, ker):
-    tiles = length // RT.BASE_TILE_ROWS
-
-    match layout:
-      case "naive":
-        inner_dim = 1
-        outer_dim = (tiles + 1) // 2
-      case "ortho":
-        inner_dim = 1
-        outer_dim = tiles
-      case _: raise NotImplementedError(f"rv layout {layout} not implemented")
-
-    uop = ker.alloc((outer_dim, inner_dim, 2), dtype, AddrSpace.REG)
-    return RV(uop, ker)
-
-ALL_TILES = UOp | GL | ST | RT | RV
@@ -1,5 +1,6 @@
 #!/usr/bin/env python3
 import sys, os, zlib, struct, hashlib
+from hexdump import hexdump
 from tinygrad.helpers import DEBUG, getenv, fetch
 from tinygrad.runtime.support.usb import USB3

@@ -119,7 +119,14 @@ extension TinyGPUViewModel: OSSystemExtensionRequestDelegate {

 		os_log("sysex actionForReplacingExtension: %@ %@", existing, ext)

+		// Add appropriate logic here to determine whether to replace the extension
+		// with the new extension. Common things to check for include
+		// testing whether the new extension's version number is newer than
+		// the current version number, or whether the bundleIdentifier is different.
+		// For simplicity, this sample always replaces the current extension
+		// with the new one.
 		replacementAction = .replace
+
 		self.state = .activating
 		return replacementAction
 	}
@@ -7,48 +7,30 @@
 struct TinyGPUDriverUserClient_IVars
 {
 	OSSharedPtr<TinyGPUDriver> provider = nullptr;
-
-	TinyGPUCreateDMAResp *dmas = nullptr;
-	size_t dmaCount = 0;
-	size_t dmaCap = 0;
-
-	int ensureDMACap(size_t need)
-	{
-		// not thread-safe
-		if (need <= dmaCap) return 0;
-
-		size_t newCap = dmaCap ? dmaCap * 2 : 16;
-		while (newCap < need) newCap *= 2;
-
-		auto *newArr = IONewZero(TinyGPUCreateDMAResp, newCap);
-		if (!newArr) return -kIOReturnNoMemory;
-
-		if (dmas && dmaCount) {
-			memcpy(newArr, dmas, dmaCount * sizeof(TinyGPUCreateDMAResp));
-		}
-
-		IOSafeDeleteNULL(dmas, TinyGPUCreateDMAResp, dmaCap);
-		dmas = newArr;
-		dmaCap = newCap;
-		return 0;
-	}
 };

 bool TinyGPUDriverUserClient::init()
 {
-	auto ok = super::init();
-	if (!ok) return false;
+	auto theAnswer = super::init();
+	if (!theAnswer) {
+		return false;
+	}

 	ivars = IONewZero(TinyGPUDriverUserClient_IVars, 1);
-	if (!ivars) return false;
+	if (ivars == nullptr) {
+		return false;
+	}
+
 	return true;
 }

 void TinyGPUDriverUserClient::free()
 {
-	if (ivars) {
-		IOSafeDeleteNULL(ivars, TinyGPUDriverUserClient_IVars, 1);
+	if (ivars != nullptr) {
+		ivars->provider.reset();
 	}
+
+	IOSafeDeleteNULL(ivars, TinyGPUDriverUserClient_IVars, 1);
 	super::free();
 }

@@ -77,22 +59,6 @@ error:

 kern_return_t TinyGPUDriverUserClient::Stop_Impl(IOService* in_provider)
 {
-	// release all DMA allocations for this client
-	if (ivars) {
-		for (size_t i = 0; i < ivars->dmaCount; i++) {
-			auto &d = ivars->dmas[i];
-			if (d.dmaCmd) {
-				d.dmaCmd->CompleteDMA(kIODMACommandCompleteDMANoOptions);
-				d.dmaCmd->release();
-				d.dmaCmd = nullptr;
-			}
-		}
-		ivars->dmaCount = 0;
-		IOSafeDeleteNULL(ivars->dmas, TinyGPUCreateDMAResp, ivars->dmaCap);
-		ivars->dmas = nullptr;
-		ivars->provider.reset();
-	}
-
 	return Stop(in_provider, SUPERDISPATCH);
 }

@@ -136,26 +102,26 @@ kern_return_t TinyGPUDriverUserClient::ExternalMethod(uint64_t selector, IOUserC

 kern_return_t IMPL(TinyGPUDriverUserClient, CopyClientMemoryForType)
 {
-	if (!memory) return kIOReturnBadArgument;
-	if (!ivars->provider.get()) return kIOReturnNotAttached;
+	if (!memory) {
+		return kIOReturnBadArgument;
+	}
+
+	if (ivars->provider.get() == nullptr) {
+		return kIOReturnNotAttached;
+	}

-	// bar handling, type is bar num
 	if (type < 6) {
 		uint32_t bar = (uint32_t)type;
 		return ivars->provider->MapBar(bar, memory);
 	}

-	// dma handling, type is size
-	if (ivars->ensureDMACap(ivars->dmaCount + 1)) {
-		os_log(OS_LOG_DEFAULT, "tinygpu: cannot grow dma array");
-		return kIOReturnNoMemory;
+	// dma page buffer
+	TinyGPUCreateDMAResp buf;
+	kern_return_t err = ivars->provider->CreateDMA(type, &buf);
+	if (err) {
+		return err;
 	}

-	TinyGPUCreateDMAResp buf{};
-	kern_return_t err = ivars->provider->CreateDMA(type, &buf);
-	if (err) return err;
-
-	ivars->dmas[ivars->dmaCount++] = buf;
 	*memory = buf.sharedBuf;
 	return 0;
 }
@@ -1,7 +1,7 @@
 # extra/weekly_commits_table.py
 import os, subprocess, datetime as dt

-NAMES = ["chenyu","George Hotz","nimlgen","qazal","wozeparrot"]
+NAMES = ["chenyu","George Hotz","nimlgen","qazal","Sieds Lykles","wozeparrot"]
 REPO  = os.environ.get("REPO_PATH",".")
 today = dt.date.today()
 days  = [(today - dt.timedelta(i)).strftime("%Y-%m-%d") for i in range(6,-1,-1)]
@@ -40,4 +40,4 @@ for d in days:
 print("** Commits by day (last 7) **")
 print("```")
 print("\n".join([header, rule] + rows))
-print("```")
+print("```")
@@ -0,0 +1,10 @@
+[mypy]
+warn_unused_configs = True
+files = tinygrad
+ignore_missing_imports = True
+check_untyped_defs = True
+explicit_package_bases = True
+warn_unreachable = True
+warn_redundant_casts = True
+# NOTE: had to comment this out to make mypy pass on both CI and OSX
+#warn_unused_ignores = True
@@ -1,232 +0,0 @@
-[project]
-name = "tinygrad"
-version = "0.11.0"
-description = "You like pytorch? You like micrograd? You love tinygrad! <3"
-authors = [{ name = "George Hotz" }]
-
-classifiers = ["Programming Language :: Python :: 3"]
-
-license = 'MIT'
-readme = "README.md"
-requires-python = ">=3.11"
-dependencies = []
-
-[build-system]
-requires = ["setuptools"]
-build-backend = "setuptools.build_meta"
-
-[tool.setuptools]
-include-package-data = true
-packages = [
-  'tinygrad',
-  'tinygrad.apps',
-  'tinygrad.codegen',
-  'tinygrad.codegen.opt',
-  'tinygrad.codegen.late',
-  'tinygrad.engine',
-  'tinygrad.mixin',
-  'tinygrad.nn',
-  'tinygrad.renderer',
-  'tinygrad.runtime',
-  'tinygrad.runtime.autogen',
-  'tinygrad.runtime.autogen.am',
-  'tinygrad.runtime.graph',
-  'tinygrad.runtime.support',
-  'tinygrad.runtime.support.am',
-  'tinygrad.runtime.support.nv',
-  'tinygrad.schedule',
-  'tinygrad.uop',
-  'tinygrad.viz',
-]
-
-[tool.setuptools.package-data]
-tinygrad = ["py.typed"]
-"tinygrad.viz" = ["index.html", "assets/**/*", "js/*"]
-
-
-[project.optional-dependencies]
-arm = ["unicorn"]
-triton = ["triton-nightly>=2.1.0.dev20231014192330"]
-linting = [
-  "pylint",
-  "mypy==1.18.1",
-  "typing-extensions",
-  "pre-commit",
-  "ruff",
-  "numpy",
-  "typeguard",
-]
-# mlperf = [
-#   "mlperf-logging @ git+https://github.com/mlperf/[email protected]",
-# ]
-testing_minimal = [
-  "numpy",
-  "torch==2.9.0",
-  "pytest",
-  "pytest-xdist",
-  "pytest-timeout",
-  "pytest-split",
-  "hypothesis",
-  "z3-solver",
-]
-testing_unit = ["tinygrad[testing_minimal]", "tqdm", "safetensors", "tabulate"]
-testing = [
-  "tinygrad[testing_minimal]",
-  "pillow",
-  "onnx==1.18.0",
-  "onnx2torch",
-  "onnxruntime",
-  "opencv-python",
-  "tabulate",
-  "tqdm",
-  "safetensors",
-  "transformers",
-  "sentencepiece",
-  "tiktoken",
-  "blobfile",
-  "librosa",
-  # librosa needs numba but uv ignores python upper bounds and some numba versions require <python3.10
-  "numba>=0.55",
-  "networkx",
-  "nibabel",
-  "bottle",
-  "ggml-python",
-  "capstone",
-  "pycocotools",
-  "boto3",
-  "pandas",
-  "influxdb3-python",
-]
-docs = [
-  "mkdocs",
-  "mkdocs-material",
-  "mkdocstrings[python]",
-  "markdown-callouts",
-  "markdown-exec[ansi]",
-  "black",
-  "numpy",
-]
-
-
-[tool.mutmut]
-paths_to_mutate = ["tinygrad/"]
-do_not_mutate = [
-  "tinygrad/apps/*",
-  "tinygrad/codegen/*",
-  "tinygrad/engine/*",
-  "tinygrad/nn/*",
-  "tinygrad/renderer/*",
-  "tinygrad/runtime/*",
-  "tinygrad/schedule/*",
-  "tinygrad/uop/*",
-  "tinygrad/viz/*",
-  "tinygrad/device.py",
-  "tinygrad/dtype.py",
-  "tinygrad/gradient.py",
-  "tinygrad/helpers.py",
-  "tinygrad/tensor.py",
-]
-tests_dir = ["test/test_tiny.py", "test/test_ops.py"]
-debug = true
-
-
-[tool.mypy]
-warn_unused_configs = true
-files = ["tinygrad"]
-ignore_missing_imports = true
-check_untyped_defs = true
-explicit_package_bases = true
-warn_unreachable = true
-warn_redundant_casts = true
-# NOTE: had to comment this out to make mypy pass on both CI and OSX
-#warn_unused_ignores = true
-
-[tool.pytest.ini_options]
-norecursedirs = [
-  "extra",
-  ".hypothesis",
-  ".git",
-]
-timeout = 300
-timeout_method = "thread"
-timeout_func_only = true
-testpaths = ["test"]
-
-[tool.ruff]
-preview = true
-target-version = "py311"
-line-length = 150
-indent-width = 2
-exclude = [
-  ".git/",
-  "docs/",
-  "extra/",
-  "test/external/mlperf_resnet",
-  "test/external/mlperf_unet3d",
-]
-
-[tool.ruff.lint]
-select = [
-  "F",  # Pyflakes
-  "W6",
-  "E71",
-  "E72",
-  "E112",   # no-indented-block
-  "E113",   # unexpected-indentation
-  # "E124",
-  "E203",   # whitespace-before-punctuation
-  "E272",   # multiple-spaces-before-keyword
-  "E275",   # missing-whitespace-after-keyword
-  "E303",   # too-many-blank-lines
-  "E304",   # blank-line-after-decorator
-  "E501",   # line-too-long
-  # "E502",
-  "E702",   # multiple-statements-on-one-line-semicolon
-  "E703",   # useless-semicolon
-  "E731",   # lambda-assignment
-  "W191",   # tab-indentation
-  "W291",   # trailing-whitespace
-  "W293",   # blank-line-with-whitespace
-  "UP039",  # unnecessary-class-parentheses
-  "C416",   # unnecessary-comprehension
-  "RET506", # superfluous-else-raise
-  "RET507", # superfluous-else-continue
-  "A",      # builtin-variable-shadowing, builtin-argument-shadowing, builtin-attribute-shadowing
-  "FURB110",# if-exp-instead-of-or-operator
-  "RUF018", # assignment-in-assert
-]
-
-# detect unused imports in examples
-[tool.ruff.lint.per-file-ignores]
-"examples/**/*.py" = [
-  "W6",
-  "E71",
-  "E72",
-  "E112",
-  "E113",
-  "E203",
-  "E272",
-  "E275",
-  "E303",
-  "E304",
-  "E501",
-  "E702",
-  "E703",
-  "E731",
-  "W191",
-  "W291",
-  "W293",
-  "UP039",
-  "C416",
-  "RET506",
-  "RET507",
-  "A",
-  "FURB110",
-  "RUF018",
-  "F541",
-  "F841",
-]
-"tinygrad/runtime/autogen/**/*.py" = ["E501", "F401", "E722", "E731", "F821", "A006"]
-
-[tool.ruff.format]
-exclude = ["*"]
@@ -0,0 +1,6 @@
+[pytest]
+norecursedirs = extra
+timeout = 300
+timeout_method = thread
+timeout_func_only = true
+testpaths = test
@@ -0,0 +1,56 @@
+indent-width = 2
+preview = true
+target-version = "py310"
+
+lint.select = [
+  "F",  # Pyflakes
+  "W6",
+  "E71",
+  "E72",
+  "E112",   # no-indented-block
+  "E113",   # unexpected-indentation
+  # "E124",
+  "E203",   # whitespace-before-punctuation
+  "E272",   # multiple-spaces-before-keyword
+  "E275",   # missing-whitespace-after-keyword
+  "E303",   # too-many-blank-lines
+  "E304",   # blank-line-after-decorator
+  "E501",   # line-too-long
+  # "E502",
+  "E702",   # multiple-statements-on-one-line-semicolon
+  "E703",   # useless-semicolon
+  "E731",   # lambda-assignment
+  "W191",   # tab-indentation
+  "W291",   # trailing-whitespace
+  "W293",   # blank-line-with-whitespace
+  "UP039",  # unnecessary-class-parentheses
+  "C416",   # unnecessary-comprehension
+  "RET506", # superfluous-else-raise
+  "RET507", # superfluous-else-continue
+  "A",      # builtin-variable-shadowing, builtin-argument-shadowing, builtin-attribute-shadowing
+  "FURB110",# if-exp-instead-of-or-operator
+  "RUF018", # assignment-in-assert
+]
+
+line-length = 150
+
+exclude = [
+  ".git/",
+  "docs/",
+  "extra/",
+  "tinygrad/runtime/autogen",
+  "test/external/mlperf_resnet",
+  "test/external/mlperf_unet3d",
+]
+
+# detect unused imports in examples
+[lint.per-file-ignores]
+"examples/**/*.py" = [
+  "W6", "E71", "E72", "E112", "E113", "E203", "E272", "E275",
+  "E303", "E304", "E501", "E702", "E703", "E731", "W191",
+  "W291", "W293", "UP039", "C416", "RET506", "RET507", "A",
+  "FURB110", "RUF018", "F541", "F841"
+]
+
+[format]
+exclude = ["*"]
@@ -0,0 +1,110 @@
+#!/usr/bin/env python3
+
+from pathlib import Path
+from setuptools import setup
+
+directory = Path(__file__).resolve().parent
+with open(directory / 'README.md', encoding='utf-8') as f:
+  long_description = f.read()
+
+testing_minimal = [
+  "numpy",
+  "torch==2.9.0",
+  "pytest",
+  "pytest-xdist",
+  "pytest-timeout",
+  "pytest-split",
+  "hypothesis",
+  "z3-solver",
+]
+
+setup(name='tinygrad',
+      version='0.11.0',
+      description='You like pytorch? You like micrograd? You love tinygrad! <3',
+      author='George Hotz',
+      license='MIT',
+      long_description=long_description,
+      long_description_content_type='text/markdown',
+      packages = [
+        'tinygrad',
+        'tinygrad.apps',
+        'tinygrad.codegen',
+        'tinygrad.codegen.opt',
+        'tinygrad.codegen.late',
+        'tinygrad.engine',
+        'tinygrad.nn',
+        'tinygrad.renderer',
+        'tinygrad.runtime',
+        'tinygrad.runtime.autogen',
+        'tinygrad.runtime.autogen.am',
+        'tinygrad.runtime.autogen.nv',
+        'tinygrad.runtime.graph',
+        'tinygrad.runtime.support',
+        'tinygrad.runtime.support.am',
+        'tinygrad.runtime.support.nv',
+        'tinygrad.schedule',
+        'tinygrad.uop',
+        'tinygrad.viz',
+      ],
+      package_data = {'tinygrad': ['py.typed'], 'tinygrad.viz': ['index.html', 'assets/**/*', 'js/*']},
+      classifiers=[
+        "Programming Language :: Python :: 3",
+        "License :: OSI Approved :: MIT License"
+      ],
+      install_requires=[],
+      python_requires='>=3.11',
+      extras_require={
+        'arm': ["unicorn"],
+        'triton': ["triton-nightly>=2.1.0.dev20231014192330"],
+        'linting': [
+            "pylint",
+            "mypy==1.18.1",
+            "typing-extensions",
+            "pre-commit",
+            "ruff",
+            "numpy",
+            "typeguard",
+        ],
+        #'mlperf': ["mlperf-logging @ git+https://github.com/mlperf/[email protected]"],
+        'testing_minimal': testing_minimal,
+        'testing_unit': testing_minimal + [
+            "tqdm",
+            "safetensors",
+            "tabulate",  # for sz.py
+        ],
+        'testing': testing_minimal + [
+            "pillow",
+            "onnx==1.18.0",
+            "onnx2torch",
+            "onnxruntime",
+            "opencv-python",
+            "tabulate",
+            "tqdm",
+            "safetensors",
+            "transformers",
+            "sentencepiece",
+            "tiktoken",
+            "blobfile",
+            "librosa",
+            "numba>=0.55",  # librosa needs numba but uv ignores python upper bounds and some numba versions require <python3.10
+            "networkx",
+            "nibabel",
+            "bottle",
+            "ggml-python",
+            "capstone",
+            "pycocotools",
+            "boto3",
+            "pandas",
+            "influxdb3-python"
+        ],
+        'docs': [
+            "mkdocs",
+            "mkdocs-material",
+            "mkdocstrings[python]",
+            "markdown-callouts",
+            "markdown-exec[ansi]",
+            "black",
+            "numpy",
+        ],
+      },
+      include_package_data=True)
@@ -1,90 +1,237 @@
-# ruff: noqa: E501 E712 F401
+# ruff: noqa: E501
 from tinygrad import dtypes, Device
-from tinygrad.uop.ops import UOp, AxisType, Ops, KernelInfo
+from tinygrad.uop.ops import UOp, AxisType, Ops
 from tinygrad.codegen import full_rewrite
-from tinygrad.codegen.opt import Opt, OptOps # pylint: disable=unused-import
 from tinygrad.renderer import ProgramSpec
 from tinygrad.engine.realize import CompiledRunner
-from tinygrad.helpers import dedup, getenv
+from tinygrad.helpers import dedup
 from tinygrad.device import Buffer
-from tinygrad.dtype import ImageDType, Invalid
+from tinygrad.dtype import ImageDType

-# PYTHONPATH="." DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
+# PYTHONPATH="." DEBUG=5 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
+# kernel 672
+# faster on d59d4cd, 50% slower with the new linearizer

-def vision_conv_143():
-  c0 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((16, 1024, 4)), (), 0)
-  c2 = UOp.range(32, 3, AxisType.LOOP)
-  c5 = UOp.range(128, 4, AxisType.LOOP)
-  c8 = UOp.range(16, 2, AxisType.LOOP)
-  c16 = UOp.range(7, 0, AxisType.REDUCE)
-  c17 = c8*2+c16
-  c24 = ((c17<3)!=True)&(c17<35)
-  c26 = UOp.range(7, 1, AxisType.REDUCE)
-  c27 = c2*2+c26
-  c32 = ((c27<3)!=True)&(c27<67)
-  c34 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((32, 1024, 4)), (), 1)
-  c38 = c5//2
-  c45 = (c32&c24).where((c27*64+c38+c17*4096+-12480), UOp.const(dtypes.index, Invalid))
-  c48 = (c24&c32).where(c34.index(c45), UOp.const(dtypes.float, 0.0))
-  c49 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((64, 49, 4)), (), 2)
-  c61 = c48*c49.index((c26*4+c5%2+c16*28+c38*196))
-  c63 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(128), (), 3)
-  c65 = c61.reduce(c16, c26, arg=Ops.ADD)+c63.index(c5)
-  c67 = c0.index((c2*128+c5+c8*4096), ptr=True).store(c65).end(c8, c2, c5)
+""" d59d4cd
+c0 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((32, 1024, 4)), arg=0, src=())
+c1 = UOp.range(UOp.const(dtypes.index, 64), 3, AxisType.LOOP)
+c2 = UOp.range(UOp.const(dtypes.index, 64), 4, AxisType.LOOP)
+c3 = UOp.range(UOp.const(dtypes.index, 32), 2, AxisType.LOOP)
+c4 = (((c1*UOp.const(dtypes.index, 64))+c2)+(c3*UOp.const(dtypes.index, 4096)))
+c5 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((32, 1024, 4)), arg=1, src=())
+c6 = c5.index(c4).load()
+c7 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((32, 3072, 4)), arg=2, src=())
+c8 = UOp.range(UOp.const(dtypes.index, 48), 0, AxisType.REDUCE)
+c9 = UOp.range(UOp.const(dtypes.index, 4), 1, AxisType.REDUCE)
+c10 = c7.index(((((c8*UOp.const(dtypes.index, 4))+c9)+(c1*UOp.const(dtypes.index, 192)))+(c3*UOp.const(dtypes.index, 12288)))).load()
+c11 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((16, 192, 4)), arg=3, src=())
+c12 = c11.index(((((c9*UOp.const(dtypes.index, 4))+(c2%UOp.const(dtypes.index, 4)))+(c8*UOp.const(dtypes.index, 16)))+((c2//UOp.const(dtypes.index, 4))*UOp.const(dtypes.index, 768)))).load()
+c13 = UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(64), arg=4, src=())
+c14 = c13.index(c2).load()
+c15 = UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(64), arg=5, src=())
+c16 = c15.index(c2).load()
+c17 = (c6+(((c10*c12.cast(dtypes.float)).cast(dtypes.float).reduce(c8, c9, arg=Ops.ADD)+c14.cast(dtypes.float))*c16.cast(dtypes.float)))
+c18 = c0.index(c4).store(c17, c3, c1, c2)
+ast = c18.sink()
+more upcast axis : [(3, 320, 0, 4)]
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+__kernel void r_512_16_4_4_48_4(write_only image2d_t data0_131072, read_only image2d_t data1_131072, read_only image2d_t data2_393216, read_only image2d_t data3_12288, __global half* data4_64, __global half* data5_64) {
+const sampler_t smp = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAMP | CLK_FILTER_NEAREST;
+  float acc0[16];
+  int idx0 = get_global_id(0); /* 16 */
+  int idx1 = get_global_id(1); /* 512 */
+  int alu0 = (idx1>>4);
+  *(acc0+0) = 0.0f;
+  *(acc0+1) = 0.0f;
+  *(acc0+2) = 0.0f;
+  *(acc0+3) = 0.0f;
+  *(acc0+4) = 0.0f;
+  *(acc0+5) = 0.0f;
+  *(acc0+6) = 0.0f;
+  *(acc0+7) = 0.0f;
+  *(acc0+8) = 0.0f;
+  *(acc0+9) = 0.0f;
+  *(acc0+10) = 0.0f;
+  *(acc0+11) = 0.0f;
+  *(acc0+12) = 0.0f;
+  *(acc0+13) = 0.0f;
+  *(acc0+14) = 0.0f;
+  *(acc0+15) = 0.0f;
+  for (int Ridx0 = 0; Ridx0 < 48; Ridx0++) {
+    int alu17 = ((idx1*192)+Ridx0);
+    int alu18 = (alu17+48);
+    int alu19 = (alu17+96);
+    int alu20 = (alu17+144);
+    int alu21 = (Ridx0<<2);
+    float4 val0 = read_imagef(data3_12288, smp, (int2)(alu21,idx0));
+    float4 val1 = read_imagef(data3_12288, smp, (int2)((alu21+1),idx0));
+    float4 val2 = read_imagef(data3_12288, smp, (int2)((alu21+2),idx0));
+    float4 val3 = read_imagef(data3_12288, smp, (int2)((alu21+3),idx0));
+    float4 val4 = read_imagef(data2_393216, smp, (int2)((alu18-(3072*(((alu18>>10)*43)>>7))),alu0));
+    float4 val5 = read_imagef(data2_393216, smp, (int2)((alu19-(3072*(((alu19>>10)*43)>>7))),alu0));
+    float4 val6 = read_imagef(data2_393216, smp, (int2)((alu20-(3072*(((alu20>>10)*43)>>7))),alu0));
+    float4 val7 = read_imagef(data2_393216, smp, (int2)((alu17-(3072*(((alu17>>10)*43)>>7))),alu0));
+    *(acc0+1) = ((*(acc0+1))+(val4.x*val0.x)+(val4.y*val1.x)+(val4.z*val2.x)+(val4.w*val3.x));
+    *(acc0+5) = ((*(acc0+5))+(val4.x*val0.y)+(val4.y*val1.y)+(val4.z*val2.y)+(val4.w*val3.y));
+    *(acc0+9) = ((*(acc0+9))+(val4.x*val0.z)+(val4.y*val1.z)+(val4.z*val2.z)+(val4.w*val3.z));
+    *(acc0+13) = ((*(acc0+13))+(val4.x*val0.w)+(val4.y*val1.w)+(val4.z*val2.w)+(val4.w*val3.w));
+    *(acc0+2) = ((*(acc0+2))+(val5.x*val0.x)+(val5.y*val1.x)+(val5.z*val2.x)+(val5.w*val3.x));
+    *(acc0+6) = ((*(acc0+6))+(val5.x*val0.y)+(val5.y*val1.y)+(val5.z*val2.y)+(val5.w*val3.y));
+    *(acc0+10) = ((*(acc0+10))+(val5.x*val0.z)+(val5.y*val1.z)+(val5.z*val2.z)+(val5.w*val3.z));
+    *(acc0+14) = ((*(acc0+14))+(val5.x*val0.w)+(val5.y*val1.w)+(val5.z*val2.w)+(val5.w*val3.w));
+    *(acc0+3) = ((*(acc0+3))+(val6.x*val0.x)+(val6.y*val1.x)+(val6.z*val2.x)+(val6.w*val3.x));
+    *(acc0+7) = ((*(acc0+7))+(val6.x*val0.y)+(val6.y*val1.y)+(val6.z*val2.y)+(val6.w*val3.y));
+    *(acc0+11) = ((*(acc0+11))+(val6.x*val0.z)+(val6.y*val1.z)+(val6.z*val2.z)+(val6.w*val3.z));
+    *(acc0+15) = ((*(acc0+15))+(val6.x*val0.w)+(val6.y*val1.w)+(val6.z*val2.w)+(val6.w*val3.w));
+    *(acc0+0) = ((*(acc0+0))+(val7.x*val0.x)+(val7.y*val1.x)+(val7.z*val2.x)+(val7.w*val3.x));
+    *(acc0+4) = ((*(acc0+4))+(val7.x*val0.y)+(val7.y*val1.y)+(val7.z*val2.y)+(val7.w*val3.y));
+    *(acc0+8) = ((*(acc0+8))+(val7.x*val0.z)+(val7.y*val1.z)+(val7.z*val2.z)+(val7.w*val3.z));
+    *(acc0+12) = ((*(acc0+12))+(val7.x*val0.w)+(val7.y*val1.w)+(val7.z*val2.w)+(val7.w*val3.w));
+  }
+  int alu39 = (idx0<<2);
+  half4 val8 = (*((__global half4*)((data4_64+alu39))));
+  half4 val9 = (*((__global half4*)((data5_64+alu39))));
+  int alu40 = (idx0+(idx1<<6));
+  int2 cast0 = (int2)((alu40&1023),alu0);
+  float4 val10 = read_imagef(data1_131072, smp, cast0);
+  int2 cast1 = (int2)(((alu40+16)&1023),alu0);
+  float4 val11 = read_imagef(data1_131072, smp, cast1);
+  int2 cast2 = (int2)(((alu40+32)&1023),alu0);
+  float4 val12 = read_imagef(data1_131072, smp, cast2);
+  int2 cast3 = (int2)(((alu40+48)&1023),alu0);
+  float4 val13 = read_imagef(data1_131072, smp, cast3);
+  float cast4 = ((float)(val8.x));
+  float cast5 = ((float)(val9.x));
+  float cast6 = ((float)(val8.y));
+  float cast7 = ((float)(val9.y));
+  float cast8 = ((float)(val8.z));
+  float cast9 = ((float)(val9.z));
+  float cast10 = ((float)(val8.w));
+  float cast11 = ((float)(val9.w));
+  write_imagef(data0_131072, cast0, (float4)((val10.x+(((*(acc0+0))+cast4)*cast5)),(val10.y+(((*(acc0+4))+cast6)*cast7)),(val10.z+(((*(acc0+8))+cast8)*cast9)),(val10.w+(((*(acc0+12))+cast10)*cast11))));
+  write_imagef(data0_131072, cast1, (float4)((val11.x+(((*(acc0+1))+cast4)*cast5)),(val11.y+(((*(acc0+5))+cast6)*cast7)),(val11.z+(((*(acc0+9))+cast8)*cast9)),(val11.w+(((*(acc0+13))+cast10)*cast11))));
+  write_imagef(data0_131072, cast2, (float4)((val12.x+(((*(acc0+2))+cast4)*cast5)),(val12.y+(((*(acc0+6))+cast6)*cast7)),(val12.z+(((*(acc0+10))+cast8)*cast9)),(val12.w+(((*(acc0+14))+cast10)*cast11))));
+  write_imagef(data0_131072, cast3, (float4)((val13.x+(((*(acc0+3))+cast4)*cast5)),(val13.y+(((*(acc0+7))+cast6)*cast7)),(val13.z+(((*(acc0+11))+cast8)*cast9)),(val13.w+(((*(acc0+15))+cast10)*cast11))));
+}
+*** QCOM     672 r_512_16_4_4_48_4                              arg  6 mem   0.10 GB tm    322.55us/    77.83ms (    157 GFLOPS    4|160    GB/s) ['mul', '__add__', 'conv2d']
+"""

-  opts = None
-  # JITBEAM=2
-  # (Opt(op=OptOps.UPCAST, axis=2, arg=4), Opt(op=OptOps.NOLOCALS, axis=None, arg=None), Opt(op=OptOps.UPCAST, axis=2, arg=2), Opt(op=OptOps.UPCAST, axis=1, arg=4), Opt(op=OptOps.SWAP, axis=1, arg=2))
-  return c67.sink(arg=KernelInfo(name="conv", opts_to_apply=opts))
+""" master 99e76f33a0f4ec84c79c1271dbc955fe6b5a7778
+c0 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((32, 1024, 4)), (), 0)
+c2 = UOp.range(64, 3, AxisType.LOOP)
+c4 = UOp.range(64, 4, AxisType.LOOP)
+c7 = UOp.range(32, 2, AxisType.LOOP)
+c10 = (((c2*64)+c4)+(c7*4096))
+c12 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((32, 1024, 4)), (), 1)
+c14 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((32, 3072, 4)), (), 2)
+c16 = UOp.range(48, 0, AxisType.REDUCE)
+c19 = UOp.range(4, 1, AxisType.REDUCE)
+c28 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((16, 192, 4)), (), 3)
+c40 = (c14.index(((((c16*4)+c19)+(c2*192))+(c7*12288)))*c28.index(((((c19*4)+(c4%4))+(c16*16))+((c4//4)*768))))
+c42 = UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(64), (), 4)
+c46 = UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(64), (), 5)
+c50 = (c12.index(c10)+((c40.reduce(c16, c19, arg=Ops.ADD)+c42.index(c4).cast(dtypes.float))*c46.index(c4).cast(dtypes.float)))
+c52 = c0.index(c10, ptr=True).store(c50).end(c7, c2, c4)
+ast = c52.sink()
+more upcast axis : [(3, 320, 0, 4)]
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+__kernel void r_512_16_4_4_48_4(write_only image2d_t data0_131072, read_only image2d_t data1_131072, read_only image2d_t data2_393216, read_only image2d_t data3_12288, __global half* data4_64, __global half* data5_64) {
+const sampler_t smp = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAMP | CLK_FILTER_NEAREST;
+  float acc0[16];
+  int idx0 = get_global_id(0); /* 16 */
+  int idx1 = get_global_id(1); /* 512 */
+  *(acc0+0) = 0.0f;
+  *(acc0+1) = 0.0f;
+  *(acc0+2) = 0.0f;
+  *(acc0+3) = 0.0f;
+  *(acc0+4) = 0.0f;
+  *(acc0+5) = 0.0f;
+  *(acc0+6) = 0.0f;
+  *(acc0+7) = 0.0f;
+  *(acc0+8) = 0.0f;
+  *(acc0+9) = 0.0f;
+  *(acc0+10) = 0.0f;
+  *(acc0+11) = 0.0f;
+  *(acc0+12) = 0.0f;
+  *(acc0+13) = 0.0f;
+  *(acc0+14) = 0.0f;
+  *(acc0+15) = 0.0f;
+  int alu16 = (idx0<<2);
+  half4 val0 = (*((__global half4*)((data4_64+alu16))));
+  half4 val1 = (*((__global half4*)((data5_64+alu16))));
+  int alu17 = (idx0+(idx1<<6));
+  int alu18 = (idx1>>4);
+  int2 cast0 = (int2)((alu17&1023),alu18);
+  float4 val2 = read_imagef(data1_131072, smp, cast0);
+  int2 cast1 = (int2)(((alu17+16)&1023),alu18);
+  float4 val3 = read_imagef(data1_131072, smp, cast1);
+  int2 cast2 = (int2)(((alu17+32)&1023),alu18);
+  float4 val4 = read_imagef(data1_131072, smp, cast2);
+  int2 cast3 = (int2)(((alu17+48)&1023),alu18);
+  float4 val5 = read_imagef(data1_131072, smp, cast3);
+  for (int Ridx0 = 0; Ridx0 < 48; Ridx0++) {
+    int alu19 = ((idx1*192)+Ridx0);
+    int alu20 = (alu19+48);
+    int alu21 = (alu19+96);
+    int alu22 = (alu19+144);
+    int alu23 = (Ridx0<<2);
+    float4 val6 = read_imagef(data3_12288, smp, (int2)(alu23,idx0));
+    float4 val7 = read_imagef(data3_12288, smp, (int2)((alu23+1),idx0));
+    float4 val8 = read_imagef(data3_12288, smp, (int2)((alu23+2),idx0));
+    float4 val9 = read_imagef(data3_12288, smp, (int2)((alu23+3),idx0));
+    float4 val10 = read_imagef(data2_393216, smp, (int2)((alu20-(3072*(((alu20>>10)*43)>>7))),alu18));
+    *(acc0+1) = ((*(acc0+1))+(val10.x*val6.x)+(val10.y*val7.x)+(val10.z*val8.x)+(val10.w*val9.x));
+    *(acc0+5) = ((*(acc0+5))+(val10.x*val6.y)+(val10.y*val7.y)+(val10.z*val8.y)+(val10.w*val9.y));
+    *(acc0+9) = ((*(acc0+9))+(val10.x*val6.z)+(val10.y*val7.z)+(val10.z*val8.z)+(val10.w*val9.z));
+    *(acc0+13) = ((*(acc0+13))+(val10.x*val6.w)+(val10.y*val7.w)+(val10.z*val8.w)+(val10.w*val9.w));
+    float4 val11 = read_imagef(data2_393216, smp, (int2)((alu21-(3072*(((alu21>>10)*43)>>7))),alu18));
+    *(acc0+2) = ((*(acc0+2))+(val11.x*val6.x)+(val11.y*val7.x)+(val11.z*val8.x)+(val11.w*val9.x));
+    *(acc0+6) = ((*(acc0+6))+(val11.x*val6.y)+(val11.y*val7.y)+(val11.z*val8.y)+(val11.w*val9.y));
+    *(acc0+10) = ((*(acc0+10))+(val11.x*val6.z)+(val11.y*val7.z)+(val11.z*val8.z)+(val11.w*val9.z));
+    *(acc0+14) = ((*(acc0+14))+(val11.x*val6.w)+(val11.y*val7.w)+(val11.z*val8.w)+(val11.w*val9.w));
+    float4 val12 = read_imagef(data2_393216, smp, (int2)((alu22-(3072*(((alu22>>10)*43)>>7))),alu18));
+    *(acc0+3) = ((*(acc0+3))+(val12.x*val6.x)+(val12.y*val7.x)+(val12.z*val8.x)+(val12.w*val9.x));
+    *(acc0+7) = ((*(acc0+7))+(val12.x*val6.y)+(val12.y*val7.y)+(val12.z*val8.y)+(val12.w*val9.y));
+    *(acc0+11) = ((*(acc0+11))+(val12.x*val6.z)+(val12.y*val7.z)+(val12.z*val8.z)+(val12.w*val9.z));
+    *(acc0+15) = ((*(acc0+15))+(val12.x*val6.w)+(val12.y*val7.w)+(val12.z*val8.w)+(val12.w*val9.w));
+    float4 val13 = read_imagef(data2_393216, smp, (int2)((alu19-(3072*(((alu19>>10)*43)>>7))),alu18));
+    *(acc0+0) = ((*(acc0+0))+(val13.x*val6.x)+(val13.y*val7.x)+(val13.z*val8.x)+(val13.w*val9.x));
+    *(acc0+4) = ((*(acc0+4))+(val13.x*val6.y)+(val13.y*val7.y)+(val13.z*val8.y)+(val13.w*val9.y));
+    *(acc0+8) = ((*(acc0+8))+(val13.x*val6.z)+(val13.y*val7.z)+(val13.z*val8.z)+(val13.w*val9.z));
+    *(acc0+12) = ((*(acc0+12))+(val13.x*val6.w)+(val13.y*val7.w)+(val13.z*val8.w)+(val13.w*val9.w));
+  }
+  float cast4 = ((float)(val0.x));
+  float cast5 = ((float)(val1.x));
+  float cast6 = ((float)(val0.y));
+  float cast7 = ((float)(val1.y));
+  float cast8 = ((float)(val0.z));
+  float cast9 = ((float)(val1.z));
+  float cast10 = ((float)(val0.w));
+  float cast11 = ((float)(val1.w));
+  write_imagef(data0_131072, cast0, (float4)((val2.x+(((*(acc0+0))+cast4)*cast5)),(val2.y+(((*(acc0+4))+cast6)*cast7)),(val2.z+(((*(acc0+8))+cast8)*cast9)),(val2.w+(((*(acc0+12))+cast10)*cast11))));
+  write_imagef(data0_131072, cast1, (float4)((val3.x+(((*(acc0+1))+cast4)*cast5)),(val3.y+(((*(acc0+5))+cast6)*cast7)),(val3.z+(((*(acc0+9))+cast8)*cast9)),(val3.w+(((*(acc0+13))+cast10)*cast11))));
+  write_imagef(data0_131072, cast2, (float4)((val4.x+(((*(acc0+2))+cast4)*cast5)),(val4.y+(((*(acc0+6))+cast6)*cast7)),(val4.z+(((*(acc0+10))+cast8)*cast9)),(val4.w+(((*(acc0+14))+cast10)*cast11))));
+  write_imagef(data0_131072, cast3, (float4)((val5.x+(((*(acc0+3))+cast4)*cast5)),(val5.y+(((*(acc0+7))+cast6)*cast7)),(val5.z+(((*(acc0+11))+cast8)*cast9)),(val5.w+(((*(acc0+15))+cast10)*cast11))));
+}
+*** QCOM     672 r_512_16_4_4_48_4                              arg  6 mem   0.10 GB tm    527.97us/    78.94ms (     96 GFLOPS    3|98     GB/s) ['conv2d', 'mul', '__add__']
+"""

-def vision_conv_153():
-  c0 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((8, 1024, 4)), (), 0)
-  c2 = UOp.range(16, 3, AxisType.LOOP)
-  c5 = UOp.range(256, 4, AxisType.LOOP)
-  c8 = UOp.range(8, 2, AxisType.LOOP)
-  c16 = UOp.range(7, 0, AxisType.REDUCE)
-  c17 = c8*2+c16
-  c24 = ((c17<3)!=True)&(c17<19)
-  c26 = UOp.range(7, 1, AxisType.REDUCE)
-  c27 = c2*2+c26
-  c32 = ((c27<3)!=True)&(c27<35)
-  c34 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((16, 1024, 4)), (), 1)
-  c38 = c5//2
-  c45 = (c32&c24).where((c27*128+c38+c17*4096+-12672), UOp.const(dtypes.index, Invalid))
-  c48 = (c24&c32).where(c34.index(c45), UOp.const(dtypes.float, 0.0))
-  c49 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((128, 49, 4)), (), 2)
-  c61 = c48*c49.index((c26*4+c5%2+c16*28+c38*196))
-  c63 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(256), (), 3)
-  c65 = c61.reduce(c16, c26, arg=Ops.ADD)+c63.index(c5)
-  c67 = c0.index((c2*256+c5+c8*4096), ptr=True).store(c65).end(c8, c2, c5)
-
-  opts = None
-  # JITBEAM=2
-  # (Opt(op=OptOps.UPCAST, axis=2, arg=4), Opt(op=OptOps.NOLOCALS, axis=None, arg=None), Opt(op=OptOps.UPCAST, axis=2, arg=2), Opt(op=OptOps.SWAP, axis=1, arg=2))
-  return c67.sink(arg=KernelInfo(name="conv", opts_to_apply=opts))
-
-def dm_conv_172():
-  c0 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((1, 240, 4)), (), 0)
-  c2 = UOp.range(960, 4, AxisType.LOOP)
-  c5 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((8, 384, 4)), (), 1)
-  c7 = UOp.range(32, 0, AxisType.REDUCE)
-  c10 = UOp.range(4, 1, AxisType.REDUCE)
-  c13 = UOp.range(12, 3, AxisType.REDUCE)
-  c18 = UOp.range(8, 2, AxisType.REDUCE)
-  c23 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((240, 128, 4)), (), 2)
-  c35 = c5.index((c7*4+c10+c13*128+c18*1536))*c23.index((c10*4+c2%4+c7*16+c2//4*512))
-  c37 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(960), (), 3)
-  c39 = c35.reduce(c7, c10, arg=Ops.ADD)+c37.index(c2)
-  c50 = (1.0+((c39+0.044708251953125*(c39*(c39*c39)))*-2.3021129851685216).exp2()).reciprocal()*c39
-  c53 = c50.reduce(c18, c13, arg=Ops.ADD)*0.010416666666666666
-  c55 = c0.index(c2, ptr=True).store(c53).end(c2)
-
-  opts = None
-  # JITBEAM=2
-  # (Opt(op=OptOps.UPCAST, axis=0, arg=4), Opt(op=OptOps.GROUPTOP, axis=1, arg=32), Opt(op=OptOps.UNROLL, axis=1, arg=4), Opt(op=OptOps.LOCAL, axis=0, arg=8), Opt(op=OptOps.UNROLL, axis=0, arg=4), Opt(op=OptOps.GROUP, axis=1, arg=0))
-  return c55.sink(arg=KernelInfo(name="conv", opts_to_apply=opts))
-
-ast = {143: vision_conv_143, 153: vision_conv_153, 172: dm_conv_172}[getenv("NUM", 143)]()
+c0 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((32, 1024, 4)), (), 0)
+c2 = UOp.range(64, 3, AxisType.LOOP)
+c4 = UOp.range(64, 4, AxisType.LOOP)
+c7 = UOp.range(32, 2, AxisType.LOOP)
+c10 = (((c2*64)+c4)+(c7*4096))
+c12 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((32, 1024, 4)), (), 1)
+c14 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((32, 3072, 4)), (), 2)
+c16 = UOp.range(48, 0, AxisType.REDUCE)
+c19 = UOp.range(4, 1, AxisType.REDUCE)
+c28 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((16, 192, 4)), (), 3)
+c40 = (c14.index(((((c16*4)+c19)+(c2*192))+(c7*12288)))*c28.index(((((c19*4)+(c4%4))+(c16*16))+((c4//4)*768))))
+c42 = UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(64), (), 4)
+c46 = UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(64), (), 5)
+c50 = (c12.index(c10)+((c40.reduce(c16, c19, arg=Ops.ADD)+c42.index(c4).cast(dtypes.float))*c46.index(c4).cast(dtypes.float)))
+c52 = c0.index(c10, ptr=True).store(c50).end(c7, c2, c4)
+ast = c52.sink()

 compiler = Device.default.compiler
 renderer = Device.default.renderer
@@ -93,13 +240,20 @@ allocator = Device.default.allocator
 uops = full_rewrite(ast, renderer)
 src = renderer.render(uops)

+# NOLOCALS=1 IMAGE=2 DEV=CL
 lib = compiler.compile(src)
-ps = ProgramSpec("conv", src, Device.DEFAULT, ast, uops)
+# r_64_8_16_4_4_48_4
+# NOLOCALS: r_512_16_4_4_48_4
+ps = ProgramSpec("r_512_16_4_4_48_4", src, Device.DEFAULT, ast, uops)
+print(ps.src)
+print(ps.applied_opts)
+# (Opt(op=OptOps.UNROLL, axis=0, arg=4), Opt(op=OptOps.UPCAST, axis=1, arg=4), Opt(op=OptOps.UPCAST, axis=0, arg=4), Opt(op=OptOps.NOLOCALS, axis=None, arg=None))
 cr = CompiledRunner(ps, precompiled=lib)

 gs = sorted(dedup([u for u in ast.toposort() if u.op is Ops.DEFINE_GLOBAL]), key=lambda u: u.arg)
-# print(len(gs))
-# print([g.dtype for g in gs])
+print(len(gs))
+print([g.dtype for g in gs])
+
 bufs = [Buffer(ps.device, g.size, g.dtype if isinstance(g.dtype, ImageDType) else g.dtype._base).ensure_allocated() for g in gs]

 t = cr(bufs, wait=True)
@@ -0,0 +1,39 @@
+import subprocess
+import random
+import time
+from concurrent.futures import ThreadPoolExecutor, as_completed
+
+def run_test(i, full_run=False):
+  print(f"\rRunning iteration {i}...", end=" ", flush=True)
+
+  p = subprocess.Popen(['python3', 'test/test_tiny.py', 'TestTiny.test_plus'],  stdout=subprocess.PIPE, stderr=subprocess.PIPE)
+
+  if not full_run:
+    time.sleep(random.uniform(0, 1200) / 1000)
+    p.kill()
+    _, stderr = p.communicate()
+  else:
+    _, stderr = p.communicate()
+
+  if full_run:
+    stderr_text = stderr.decode()
+    print(stderr_text)
+    assert "Ran 1 test in" in stderr_text and "OK" in stderr_text
+
+max_workers = 4
+with ThreadPoolExecutor(max_workers=max_workers) as executor:
+  futures = []
+  for i in range(1000000):
+    if i % 100 == 0:
+      for future in as_completed(futures):
+        try: future.result()
+        except Exception as e:
+          print(f"\nError in iteration: {e}")
+      futures = []
+
+      run_test(i, True)
+    else:
+      future = executor.submit(run_test, i, False)
+      futures.append(future)
+
+    if len(futures) > max_workers * 2: futures = [f for f in futures if not f.done()]
@@ -1,44 +0,0 @@
-import subprocess
-import random
-import time
-from concurrent.futures import ProcessPoolExecutor, as_completed
-from tinygrad.helpers import getenv
-
-# checks that HCQ drivers can be killed during operation without causing issues
-
-def run_test(i, full_run=False, force_ok=False):
-  print(f"\rRunning iteration {i}...", end=" ", flush=True)
-
-  p = subprocess.Popen(["python3", "test/test_tiny.py", "TestTiny.test_plus"], stdout=subprocess.PIPE, stderr=subprocess.PIPE)
-
-  if not full_run:
-    time.sleep(random.uniform(0, 1200) / 1000.0)
-    p.kill()
-    _, stderr = p.communicate()
-  else:
-    _, stderr = p.communicate()
-    stderr_text = stderr.decode()
-    assert ("Ran 1 test in" in stderr_text and "OK" in stderr_text) or (not force_ok and "Failed to take lock file" in stderr_text), stderr_text
-
-if __name__ == "__main__":
-  max_workers = getenv("MAX_WORKERS", 4)
-  with ProcessPoolExecutor(max_workers=max_workers) as executor:
-    futures = []
-    for i in range(1000000):
-      if i % 100 == 0:
-        # wait for everything we launched so far
-        for f in as_completed(futures):
-          try:
-            f.result()
-          except Exception as e:
-            print(f"\nError in iteration: {e}")
-        futures = []
-
-        # do a full run in the main proc
-        run_test(i, True, force_ok=True)
-      else:
-        futures.append(executor.submit(run_test, i, bool(getenv("FULL_RUN", 0))))
-
-      # keep list small
-      if len(futures) > max_workers * 2:
-        futures = [f for f in futures if not f.done()]
@@ -1,55 +0,0 @@
-import time
-from tinygrad.tensor import Tensor, Device
-
-MODEL_WIDTH = 512
-MODEL_HEIGHT = 256
-MODEL_FRAME_SIZE = MODEL_WIDTH * MODEL_HEIGHT * 3 // 2
-IMG_INPUT_SHAPE = (1, 12, 128, 256)
-
-def tensor_arange(end): return Tensor([float(i) for i in range(end)])
-def tensor_round(tensor:Tensor): return (tensor + 0.5).floor()
-
-h_src, w_src = 1208, 1928
-h_dst, w_dst = MODEL_HEIGHT, MODEL_WIDTH
-x = tensor_arange(w_dst).reshape(1, w_dst).expand(h_dst, w_dst)
-y = tensor_arange(h_dst).reshape(h_dst, 1).expand(h_dst, w_dst)
-ones = Tensor.ones_like(x)
-dst_coords = x.reshape((1,-1)).cat(y.reshape((1,-1))).cat(ones.reshape((1,-1)))
-
-def warp_perspective_tinygrad(src:Tensor, M_inv:Tensor) -> Tensor:
-  src_coords = M_inv @ dst_coords
-  src_coords = src_coords / src_coords[2:3, :]
-
-  x_src = src_coords[0].reshape(h_dst, w_dst)
-  y_src = src_coords[1].reshape(h_dst, w_dst)
-
-  x_nearest = tensor_round(x_src).clip(0, w_src - 1).cast('int')
-  y_nearest = tensor_round(y_src).clip(0, h_src - 1).cast('int')
-
-  # TODO: make 2d indexing fast
-  idx = y_nearest*src.shape[1] + x_nearest
-  dst = src.flatten()[idx]
-  return dst.reshape(h_dst, w_dst)
-
-if __name__ == "__main__":
-  from tinygrad.engine.jit import TinyJit
-  update_img_jit = TinyJit(warp_perspective_tinygrad, prune=True)
-
-  step_times = []
-  for _ in range(10):
-    # regenerate inputs
-    inputs = [Tensor.randn(1928,1208), Tensor.randn(3,3)]
-    Tensor.realize(*inputs)
-    Device.default.synchronize()
-
-    # do the warp
-    st = time.perf_counter()
-    out = update_img_jit(*inputs)
-    mt = time.perf_counter()
-    val = out.contiguous().realize()
-    Device.default.synchronize()
-    et = time.perf_counter()
-
-    # measure the time
-    step_times.append((et-st)*1e3)
-    print(f"enqueue {(mt-st)*1e3:6.2f} ms -- total run {step_times[-1]:6.2f} ms")
@@ -1,20 +0,0 @@
-import os
-if "DEV" not in os.environ: os.environ["DEV"] = "AMD"
-
-import unittest, time
-from tinygrad import Device
-
-class TestOpen(unittest.TestCase):
-  def generate_test_open(n):
-    def test(self):
-      dev = Device[Device.DEFAULT]
-      for i in range(10):
-        dev.allocator.alloc(10 << 20)
-        time.sleep(0.5)
-    test.__name__ = f'test_open_{n}'
-    return test
-
-  for i in range(64): locals()[f'test_open_{i}'] = generate_test_open(i)
-
-if __name__ == '__main__':
-  unittest.main()
@@ -112,6 +112,7 @@ backend_test.exclude('test_dequantizelinear_e5m2_cpu')
 backend_test.exclude('test_dequantizelinear_float4e2m1_cpu')

 # we don't support indexes
+backend_test.exclude('test_nonzero_*')

 # no support for int pow
 backend_test.exclude('test_pow_types_int32_int32_cpu')
@@ -85,11 +85,12 @@ class TestKernelSpeed(unittest.TestCase):
    gbs = mems / tm / 1e9
    self._compare(tm, tflops, gbs, nv_tflops, nv_gbs, amd_tflops, amd_gbs)

+  # NOTE: tiny7 was slower than tiny12
  # TODO: why are convs so slow?!?
  def test_conv_3x3_256_32_32_256_256(self): self._test_conv_3x3(256, 32, 32, 256, 256, nv_tflops=27, amd_tflops=14)

  # theoretical is nv_tflops=165, amd_tflops=123
-  def test_gemm_4096(self): self._test_matmul(4096, nv_tflops=110, amd_tflops=65)
+  def test_gemm_4096(self): self._test_matmul(4096, nv_tflops=115, amd_tflops=65)
  def test_gemm_8192(self): self._test_matmul(8192, nv_tflops=115, amd_tflops=60)

  # theoretical is nv_gbs=1008, amd_gbs=960
@@ -33,7 +33,7 @@ remu = _try_dlopen_remu()
 def create_sdma_packets():
  # TODO: clean up this, if we want to keep it
  structs = {}
-  for name,pkt in [(name,s) for name,s in amd_gpu.__dict__.items() if name.startswith("rocr_AMD_SDMA_PKT_") and name.endswith("_TAG")]:
+  for name,pkt in [(name,s) for name,s in amd_gpu.__dict__.items() if name.startswith("struct_SDMA_PKT_") and name.endswith("_TAG")]:
    names = set()
    fields = []
    for pkt_fields in pkt._fields_:
@@ -47,7 +47,7 @@ def create_sdma_packets():
          # merge together 64-bit fields, otherwise just append them
          if fname.endswith("_63_32") and fields[-1][0].endswith("_31_0"): fields[-1] = tuple([fname[:-6], ctypes.c_ulong, 64])
          else: fields.append(tuple([fname, *union_fields[1:]]))
-    new_name = name[18:-4].lower()
+    new_name = name[16:-4].lower()
    structs[new_name] = init_c_struct_t(tuple(fields))
    assert ctypes.sizeof(structs[new_name]) == ctypes.sizeof(pkt), f"{ctypes.sizeof(structs[new_name])} != {ctypes.sizeof(pkt)}"
  return type("SDMA_PKTS", (object, ), structs)
@@ -164,7 +164,7 @@ def cuStreamWaitEvent(stream: Any, event, flags: int) -> int: return orig_cuda.C
 def cuCtxSynchronize() -> int: return orig_cuda.CUDA_SUCCESS

 def cuGetErrorString(error: int, pStr) -> int:
-  error_str = orig_cuda.enum_cudaError_enum.get(error, "Unknown CUDA error").encode()
+  error_str = orig_cuda.cudaError_enum__enumvalues.get(error, "Unknown CUDA error").encode()
  buf = ctypes.create_string_buffer(error_str)
  # Set the pointer to point to our error string buffer
  pStr._obj.value = ctypes.cast(buf, ctypes.POINTER(ctypes.c_char))
@@ -1,5 +1,5 @@
 import ctypes, mmap, collections, functools, os
-from tinygrad.runtime.autogen import nv_570 as nv_gpu
+import tinygrad.runtime.autogen.nv_gpu as nv_gpu
 from typing import Any
 from tinygrad.helpers import to_mv
 from test.mockgpu.driver import VirtDriver, VirtFileDesc, VirtFile
@@ -153,10 +153,8 @@ class NVDriver(VirtDriver):
                 51059, 51069, 51071, 51632, 51639, 51639, 51706, 52019, 222, 50287, 50273, 50031, 50017] # from ada102
      params.numClasses = len(classes)
      if struct.cmd == nv_gpu.NV0080_CTRL_CMD_GPU_GET_CLASSLIST:
-        if params.classList and params.numClasses > 0:
-          clslist = to_mv(params.classList, params.numClasses * 4).cast('I')
-          for i,c in enumerate(classes): clslist[i] = c
-        else: params.numClasses = len(classes)
+        clslist = to_mv(params.classList, params.numClasses * 4).cast('I')
+        for i,c in enumerate(classes): clslist[i] = c
      else:
        for i,c in enumerate(classes): params.classList[i] = c
    elif struct.cmd == nv_gpu.NV2080_CTRL_CMD_GR_GET_INFO:
@@ -194,9 +192,6 @@ class NVDriver(VirtDriver):
      params.mmuFaultInfoList[0].faultAddress = int(os.environ['MOCKGPU_EMU_FAULTADDR'], base=16)
      params.mmuFaultInfoList[0].faultType = 1
      params.mmuFaultInfoList[0].accessType = 1
-    elif struct.cmd == nv_gpu.NV0000_CTRL_CMD_SYSTEM_GET_BUILD_VERSION_V2:
-      params = nv_gpu.NV0000_CTRL_SYSTEM_GET_BUILD_VERSION_V2_PARAMS.from_address(params_ptr)
-      params.driverVersionBuffer = b"570.00.00\0"
    else: raise RuntimeError(f"Unknown {struct.cmd} to rm_control")
    return 0

@@ -259,4 +254,4 @@ class NVDriver(VirtDriver):
      for gpu in self.gpus.values():
        for q in gpu.queues:
          if q.ctrl.GPGet != q.ctrl.GPPut:
-            any_progress |= q.execute()
+            any_progress |= q.execute()
@@ -1,5 +1,5 @@
 import ctypes, time
-from tinygrad.runtime.autogen import nv_570 as nv_gpu
+import tinygrad.runtime.autogen.nv_gpu as nv_gpu
 from enum import Enum, auto
 from test.mockgpu.gpu import VirtGPU
 from test.mockgpu.helpers import _try_dlopen_gpuocelot
@@ -1,6 +1,5 @@
 import unittest
 from tinygrad import Tensor, UOp, Context
-from tinygrad.dtype import AddrSpace
 from tinygrad.uop.ops import KernelInfo, AxisType

 # **** kernels ****
@@ -9,24 +8,16 @@ def custom_arange_kernel(C:UOp) -> UOp:
  i = UOp.range(C.size, 0)
  return C[i].store(i.cast(C.dtype.base)).end(i).sink(arg=KernelInfo(name=f"custom_arange_{C.size}"))

-def custom_eye_kernel(C:UOp) -> UOp:
-  i = UOp.range(C.shape[0], 0)
-  j = UOp.range(C.shape[1], 1)
-  return C[i, j].store((i.eq(j)).cast(C.dtype.base)).end(i, j).sink(arg=KernelInfo(name=f"custom_eye_{C.size}"))
-
 def custom_add_one_kernel(B:UOp, A:UOp) -> UOp:
-  A,B = A.flatten(), B.flatten()
  assert B.size == A.size
  i = UOp.range(A.size, 0)
  return B[i].store(A[i] + 1).end(i).sink(arg=KernelInfo(name=f"add_one_{A.size}"))

 def custom_elementwise_add_kernel(C:UOp, A:UOp, B:UOp) -> UOp:
-  C,A,B = C.flatten(), A.flatten(), B.flatten()
  i = UOp.range(C.size, 0)
  return C[i].store(A[i]+B[i]).end(i).sink(arg=KernelInfo(name=f"custom_add_kernel_{C.size}")).simplify()

 def custom_elementwise_addmul_kernel(C:UOp, D:UOp, A:UOp, B:UOp) -> UOp:
-  C,D,A,B = C.flatten(), D.flatten(), A.flatten(), B.flatten()
  assert C.size == D.size
  i = UOp.range(C.size, 0)
  store_c = C[i].store(A[i]+B[i])
@@ -48,43 +39,13 @@ def custom_sum(B:UOp, A:UOp) -> UOp:
  return B.sink(arg=KernelInfo(name=f"custom_sum_{A.shape[0]}", opts_to_apply=()))

 def flip_contract_kernel(dest:UOp, src:UOp):
-  i = UOp.range(dest.shape[0], 0)
-  j = UOp.range(dest.shape[1], 1, AxisType.UPCAST)
-  vec = src[i, j].contract(j)
-  store = UOp.group(*[dest[i, k].store(vec.gep(3-k)) for k in range(4)])
+  assert dest.size%4 == 0
+  i = UOp.range(dest.size//4, 0)
+  j = UOp.range(4, 1, AxisType.UPCAST)
+  vec = src[i*4+j].contract(j)
+  store = UOp.group(*[dest[i*4+k].store(vec.gep(3-k)) for k in range(4)])
  return store.end(i).sink(arg=KernelInfo(name=f"flip_contract_{dest.size}", opts_to_apply=()))

-def slice_sum_kernel(dest:UOp, src:UOp):
-  G = UOp.range(src.shape[0], 0)
-  slice_src = src[G, :]
-  reg = UOp.placeholder((1,), dest.dtype.base, 0, addrspace=AddrSpace.REG)
-  reg = reg.after(G)[0].set(0)
-  R = UOp.range(src.shape[1], 1, AxisType.REDUCE)
-  reg = reg[0].set(reg.after(R)[0] + slice_src[R], end=R)
-  ast = dest[G].set(reg[0], end=G)
-  return ast.sink(arg=KernelInfo(name=f"slice_sum_{src.shape[0]}_{src.shape[1]}", opts_to_apply=()))
-
-def simple_qkv_kernel(O:UOp, Q:UOp, K:UOp, V:UOp) -> UOp:
-  # attention without softmax
-  N, d = Q.shape[0], Q.shape[1]
-
-  i = UOp.range(N, 0)  # output row
-  d_out = UOp.range(d, 1)  # output column
-  j = UOp.range(N, 2, axis_type=AxisType.REDUCE)
-
-  k_inner = UOp.range(d, 3, axis_type=AxisType.REDUCE)
-  qk_acc = UOp.placeholder((1,), Q.dtype.base, 0, addrspace=AddrSpace.REG)
-  qk_acc = qk_acc.after(i, j)[0].set(0.0)
-  qk_acc = qk_acc[0].set(qk_acc.after(k_inner)[0] + Q[i, k_inner] * K[j, k_inner], end=k_inner)
-  qk_score = qk_acc[0] / (d ** 0.5)
-
-  out_acc = UOp.placeholder((1,), Q.dtype.base, 1, addrspace=AddrSpace.REG)
-  out_acc = out_acc.after(i, d_out)[0].set(0.0)
-  out_acc = out_acc[0].set(out_acc.after(j)[0] + qk_score * V[j, d_out], end=j)
-
-  store = O[i, d_out].store(out_acc[0])
-  return store.end(d_out).end(i).sink(arg=KernelInfo(name=f"simple_qkv_{N}_{d}", opts_to_apply=()))
-
 # **** backward callbacks ****

 def backward_gemm(gradient:UOp, kernel:UOp) -> tuple[UOp, UOp]:
@@ -130,12 +91,6 @@ class TestCustomKernel(unittest.TestCase):
    tst = tst.custom_kernel(fxn=custom_arange_kernel)[0]
    self.assertTrue((ref == tst).all().item())

-  def test_eye(self):
-    ref = Tensor.eye(1024).contiguous().realize()
-    tst = Tensor.empty_like(ref)
-    tst = tst.custom_kernel(fxn=custom_eye_kernel)[0]
-    self.assertTrue((ref == tst).all().item())
-
  def test_flip_contract(self):
    a = Tensor.randn(10,4)
    b = Tensor.empty_like(a)
@@ -156,12 +111,6 @@ class TestCustomKernel(unittest.TestCase):
    b = Tensor.custom_kernel(tst, a, fxn=custom_sum)[0]
    self.assertEqual(b.item(), 15)

-  def test_slice_sum(self):
-    A = Tensor.randn(16, 16).contiguous()
-    B = Tensor.empty(16)
-    B = Tensor.custom_kernel(B, A, fxn=slice_sum_kernel)[0]
-    self.assertTrue(B.allclose(A.sum(1)))
-
  def test_gemm(self):
    N = 16
    a = Tensor.randn(N, N)
@@ -203,19 +152,5 @@ class TestCustomKernel(unittest.TestCase):
    err = (grad_b - real_grad_b).square().max()
    self.assertLess(err.item(), 1e-6)

-  def test_simple_qkv(self):
-    N, d = 8, 4
-    Q = Tensor.randn(N, d)
-    K = Tensor.randn(N, d)
-    V = Tensor.randn(N, d)
-    O = Tensor.empty(N, d)
-
-    O_custom = Tensor.custom_kernel(O, Q, K, V, fxn=lambda o,q,k,v: simple_qkv_kernel(o,q,k,v))[0]
-    O_ref = ((Q @ K.T) / (d ** 0.5)) @ V
-
-    Tensor.realize(O_custom, O_ref)
-    err = (O_custom - O_ref).square().max()
-    self.assertLess(err.item(), 1e-6)
-
 if __name__ == '__main__':
  unittest.main()
@@ -1,31 +0,0 @@
-import numpy as np
-import unittest
-from tinygrad import Tensor, UOp
-from tinygrad.uop.ops import AxisType
-
-class TestFold(unittest.TestCase):
-  def test_reduce_add(self):
-    a = Tensor.randn(10, 10).realize()
-    a_red = a.sum(axis=1)
-    np.testing.assert_allclose(a_red.numpy(), a.numpy().sum(axis=1), atol=1e-6)
-
-  def test_fold_add(self):
-    a = Tensor.randn(10, 10).realize()
-    init = Tensor.zeros(10, 1).contiguous()
-    a_red = (init+a).fold(init).reshape(10)
-    np.testing.assert_allclose(a_red.numpy(), a.numpy().sum(axis=1), atol=1e-6)
-
-  #@unittest.skip("no outer fold yet")
-  def test_fold_matmul(self):
-    vec = Tensor.randn(1, 10).realize()
-    mats = Tensor.randn(3, 10, 10).realize()
-    np_mats = mats.numpy()
-    np_ref = ((vec.numpy() @ np_mats[0]) @ np_mats[1]) @ np_mats[2]
-
-    i = UOp.range(3, -1, AxisType.OUTER)
-    out = (vec @ mats[i]).contiguous().fold(vec, i)
-
-    np.testing.assert_allclose(out.numpy(), np_ref, atol=1e-6)
-
-if __name__ == '__main__':
-  unittest.main()
@@ -4,7 +4,7 @@ from dataclasses import replace

 from tinygrad.codegen.opt import Opt, OptOps
 from tinygrad.codegen.gpudims import get_grouped_dims
-from tinygrad.uop.ops import UOp, Ops, GroupOp, AxisType, PatternMatcher, graph_rewrite, UPat
+from tinygrad.uop.ops import UOp, Ops, GroupOp
 from tinygrad.device import Device, Buffer, is_dtype_supported
 from tinygrad.tensor import Tensor, _to_np_dtype
 from tinygrad.engine.realize import run_schedule, lower_schedule, CompiledRunner, get_program
@@ -38,22 +38,6 @@ class TestLinearizer(unittest.TestCase):
    np.testing.assert_equal(a.numpy(), ta)
    np.testing.assert_equal(b.numpy(), tb)

-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "broken on ptx")
-  def test_late_bias_load(self):
-    img = Tensor.empty(1, 3, 16, 16)
-    w = Tensor.empty(16, 3, 3, 3)
-    b = Tensor.empty(16)
-    out = img.conv2d(w, b)
-    ast = helper_linearizer_opt(out)
-    uops = get_program(ast, opts=[]).uops
-    # slice at the last loop end
-    uslice = [i for i,u in enumerate(uops) if u.op == Ops.END][-1]
-    # only valid test if outermost range is the reduce
-    if uops[uslice].src[-1].arg[-1] == AxisType.REDUCE:
-      load_types = [u.src[0].dtype for u in uops[uslice+1:] if u.op == Ops.LOAD]
-      # assert that there is a global load after the reduce ends
-      assert any(dt.addrspace == AddrSpace.GLOBAL for dt in load_types)
-
  def _test_no_nested_ranges(self, lins, skip=None):
    for l in lins:
      range_in_acc = flatten([[x for x in u.src if x.op is Ops.RANGE] for u in l.uops if u.op is Ops.DEFINE_REG])
@@ -78,7 +62,6 @@ class TestLinearizer(unittest.TestCase):
    ranges = [i for i,u in enumerate(uops) if u.op is Ops.RANGE]
    assert len(ranges) == 1 # NOTE: it collapses now

-  @unittest.expectedFailure # TODO: investigate
  def test_two_nested_range_alt_indexing(self):
    a = Tensor([2, 2]).realize()
    out = a.reshape(2, 1).pad(((1, 1), (1, 1)), value=2).sum()
@@ -279,8 +262,6 @@ class TestLinearizer(unittest.TestCase):
    _assert_grouped_dims("gidx", (65536,), (16,16,256), False, [16,16,256], False)
    #                              2             ->             3
    _assert_grouped_dims("gidx", (128,128), (16,16,256), False, [16,16,64], False)
-    #                              2             ->             2
-    _assert_grouped_dims("gidx", (65536,2), (65535,65535,65535), False, [32768,4], False)
    # test when the only divisor is the square root of dim
    _assert_grouped_dims("gidx", (121,), (12,12,12), False, [11,11], False)

@@ -305,27 +286,6 @@ class TestLinearizer(unittest.TestCase):
    with self.assertRaises(RuntimeError):
      get_grouped_dims("gidx", (2,3,4,5,6), (16,16,16))

-    # TODO: In the above cases we only test if the shape after reshape is correct, never the indices.
-    # We should check if the returned indices are correct, for all cases.
-    # (65536, 2) -> (32768, 4)
-    dims, expected_limited_dims = (65536,2), (32768, 4)
-    idxs = get_grouped_dims("gidx", dims, (65535,65535,65535))
-    def match_div(): raise RuntimeError("match_div")
-    def match_mod(): raise RuntimeError("match_mod")
-    flat_idx_pattern = UPat(Ops.SPECIAL, arg='gidx0')*expected_limited_dims[1]+UPat(Ops.SPECIAL, arg='gidx1')
-    pm = PatternMatcher([
-      (flat_idx_pattern//dims[1], match_div),
-      (flat_idx_pattern%dims[1], match_mod)
-    ])
-
-    with self.assertRaises(RuntimeError) as error:
-      graph_rewrite(idxs[0], pm)
-    self.assertIn("match_div", str(error.exception))
-
-    with self.assertRaises(RuntimeError) as error:
-      graph_rewrite(idxs[1], pm)
-    self.assertIn("match_mod", str(error.exception))
-
    # # variable too large
    # with self.assertRaises(AssertionError):
    #   get_grouped_dims("gidx", (Variable("start_pos",0,16),3,4), (16,16,16), False,)
@@ -472,8 +432,6 @@ def helper_realized_ast(r:Tensor|list[Tensor]) -> tuple[UOp, list[Buffer]]:
  # now all input buffers in s[-1] should be realized
  # create fresh buffers for the outputs
  bufs = [Buffer(x.device, x.size, x.dtype).allocate() if i < len(s[-1].ast.src) else x for i,x in enumerate(s[-1].bufs)]
-  # ensure buffers are allocated
-  for b in bufs: b.ensure_allocated()
  return s[-1].ast, bufs

 def helper_linearizer_ast(ast:UOp, inputs:list[Tensor], *args, **kwargs):
@@ -596,12 +596,6 @@ class TestMultiTensor(unittest.TestCase):
    # ast are the same on devices
    self.assertEqual(len(set(asts)), 1)

-  def test_flip(self):
-    rng = Tensor.rand((10, 10, 10))
-    t0 = rng.shard(devices_2, axis=1)
-    out = t0.flip(0) + 1
-    self.assertTrue((rng.flip(0)+1).allclose(out.to(rng.device)))
-
  def test_reshape_on_axis(self):
    t0 = Tensor.rand((26, 15, 7)).shard(devices_3, axis=1)

@@ -1551,10 +1551,8 @@ class TestOps(unittest.TestCase):
                                lambda x: Tensor.stack(*x.std_mean(axis=(1,2))))

  def test_std_mean_loaded_nan(self):
-    with warnings.catch_warnings():
-      warnings.filterwarnings("ignore", message="std_mean\\(\\): degrees of freedom is <= 0")
-      helper_test_op([(1,0,3,0,5)], lambda x: torch.stack(torch.std_mean(x, axis=(1,3))),
-                                    lambda x: Tensor.stack(*x.std_mean(axis=(1,3))))
+    helper_test_op([(1,0,3,0,5)], lambda x: torch.stack(torch.std_mean(x, axis=(1,3))),
+                                  lambda x: Tensor.stack(*x.std_mean(axis=(1,3))))
  def test_softmax(self):
    helper_test_op([(45,65)], torch.nn.Softmax(dim=1), Tensor.softmax, atol=1e-7, grad_atol=1e-7)
    helper_test_op([(45)], torch.nn.Softmax(dim=0), Tensor.softmax, atol=1e-7, grad_atol=1e-7)
@@ -2822,13 +2820,13 @@ class TestOps(unittest.TestCase):
  @slow_test
  def test_slice_fancy_indexing_list_indices(self):
    a,b,c,d,e,i,j,k,o,p = self._get_index_randoms()
-    helper_test_op([(2,5,6,5,3,4)], lambda x: x[((0,),)])
-    helper_test_op([(2,5,6,5,3,4)], lambda x: x[(0,),b,c,d,:], lambda x: x[(0,),j,k,o,:])
+    helper_test_op([(2,5,6,5,3,4)], lambda x: x[[[0]]], lambda x: x[[[0]]])
+    helper_test_op([(2,5,6,5,3,4)], lambda x: x[[0],b,c,d,:], lambda x: x[[0],j,k,o,:])
    helper_test_op([(2,5,6,5,3,4)], lambda x: x[[[[0]]],b,c,d,[[1]]], lambda x: x[[[[0]]],j,k,o,[[1]]])
-    helper_test_op([(2,5,6,5,3,4)], lambda x: x[(1,0,-1),b,c,d,:], lambda x: x[(1,0,-1),j,k,o,:])
-    helper_test_op([(2,5,6,5,3,4)], lambda x: x[a,b,c,(1,2,3),...], lambda x: x[i,j,k,(1,2,3),...])
+    helper_test_op([(2,5,6,5,3,4)], lambda x: x[[1,0,-1],b,c,d,:], lambda x: x[[1,0,-1],j,k,o,:])
+    helper_test_op([(2,5,6,5,3,4)], lambda x: x[a,b,c,[1,2,3],...], lambda x: x[i,j,k,[1,2,3],...])
    helper_test_op([(2,5,6,5,3,4)], lambda x: x[a,b,c,[[1],[2],[3]],...], lambda x: x[i,j,k,[[1],[2],[3]],...])
-    helper_test_op([(2,5,6,5,3,4)], lambda x: x[a,(2,1,0),c,(-2,1,0),e], lambda x: x[i,(2,1,0),k,(-2,1,0),p])
+    helper_test_op([(2,5,6,5,3,4)], lambda x: x[a,[2,1,0],c,[-2,1,0],e], lambda x: x[i,[2,1,0],k,[-2,1,0],p])

  @slow_test
  def test_slice_fancy_indexing_tuple_indices(self):
@@ -2843,10 +2841,11 @@ class TestOps(unittest.TestCase):
  @slow_test
  def test_slice_fancy_indexing_list_with_tensors(self):
    a,b,c,d,e,i,j,k,o,p = self._get_index_randoms()
-    helper_test_op([(2,5,6,5,3,4)], lambda x: x[(a,)], lambda x: x[(i,)])
-    helper_test_op([(2,5,6,5,3,4)], lambda x: x[(a,1)], lambda x: x[(i,1)])
-    helper_test_op([(2,5,6,5,3,4)], lambda x: x[(a,(1,1))], lambda x: x[(i,(1,1))])
-    helper_test_op([(2,5,6,5,3,4)], lambda x: x[(a,b,c,d,e)], lambda x: x[(i,j,k,o,p)])
+    helper_test_op([(2,5,6,5,3,4)], lambda x: x[[a]], lambda x: x[[i]])
+    helper_test_op([(2,5,6,5,3,4)], lambda x: x[[a,1]], lambda x: x[[i,1]])
+    helper_test_op([(2,5,6,5,3,4)], lambda x: x[[a,[1,1]]], lambda x: x[[i,[1,1]]])
+    helper_test_op([(2,5,6,5,3,4)], lambda x: x[[a,(1,1)]], lambda x: x[[i,(1,1)]])
+    helper_test_op([(2,5,6,5,3,4)], lambda x: x[[a,b,c,d,e]], lambda x: x[[i,j,k,o,p]])

  def test_slice_fancy_indexing_errors(self):
    a = Tensor.ones(10,11,12)
@@ -300,12 +300,6 @@ class TestRangeify(unittest.TestCase):
    w2 = Tensor.empty(12, 8, 3, 3)
    x.conv2d(w1).conv2d(w2).realize()

-  def test_resnet_conv2d(self):
-    x = Tensor.empty(1, 8, 32, 32)
-    w1 = Tensor.empty(8, 8, 3, 3)
-    w2 = Tensor.empty(8, 8, 1, 1)
-    x.conv2d(w1).conv2d(w2).realize()
-
  def test_xception_conv2d(self):
    # NOTE: this fusion is bad, it's recomputing the inner many times
    x = Tensor.empty(1, 4, 32, 32)
@@ -1042,12 +1042,13 @@ class TestSchedule(unittest.TestCase):
      compare = torch.nn.functional.scaled_dot_product_attention(torch.tensor(q.numpy()),torch.tensor(k.numpy()),torch.tensor(v.numpy()))
      np.testing.assert_allclose(out.numpy(), compare.numpy(), atol=1e-6, rtol=1e-3)

-    out = Tensor.scaled_dot_product_attention(q,k,v)
-    run_schedule(check_schedule(out, 4)) # TODO: should be 1?
-    if getenv("CHECK", 1):
-      import torch
-      compare = torch.nn.functional.scaled_dot_product_attention(torch.tensor(q.numpy()),torch.tensor(k.numpy()),torch.tensor(v.numpy()))
-      np.testing.assert_allclose(out.numpy(), compare.numpy(), atol=1e-6, rtol=1e-3)
+    with Context(FUSE_ATTENTION=1):
+      out = Tensor.scaled_dot_product_attention(q,k,v)
+      run_schedule(check_schedule(out, 4)) # TODO: should be 1?
+      if getenv("CHECK", 1):
+        import torch
+        compare = torch.nn.functional.scaled_dot_product_attention(torch.tensor(q.numpy()),torch.tensor(k.numpy()),torch.tensor(v.numpy()))
+        np.testing.assert_allclose(out.numpy(), compare.numpy(), atol=1e-6, rtol=1e-3)

  def test_ugly_reduceop_pairing(self):
    Tensor.manual_seed(0)
@@ -1573,13 +1574,6 @@ class TestSchedule(unittest.TestCase):
  def test_conv2d(self): _test_conv2d(5 if SPLIT_REDUCEOP else 4)
  def test_conv2d_fused(self): _test_conv2d(5 if SPLIT_REDUCEOP else 4)

-  def test_resnet_conv2d(self):
-    x = Tensor.empty(1, 8, 32, 32)
-    w1 = Tensor.empty(8, 8, 3, 3)
-    w2 = Tensor.empty(8, 8, 1, 1)
-    out = x.conv2d(w1).conv2d(w2)
-    check_schedule(out, 2)
-
  @unittest.skipUnless(is_dtype_supported(dtypes.half) and is_dtype_supported(dtypes.ulong), "need half and ulong")
  def test_conv2d_half(self): _test_conv2d(5 if SPLIT_REDUCEOP else 4, dtype=dtypes.half)
  @unittest.skipUnless(is_dtype_supported(dtypes.half), "need half")
@@ -32,7 +32,7 @@ def run_one_schedule_item(out): lower_schedule_item(get_single_element(out.sched
 class TestFuse(unittest.TestCase):
  def _test_fuse(self, fxn, *args, atol=1e-6, allow_multiple=False, **kwargs):
    GlobalCounters.reset()
-    out_single = fxn(*args, **kwargs)
+    out_single = fxn(*args, **kwargs).fuse()
    if not allow_multiple: run_one_schedule_item(out_single)
    np_single = out_single.numpy()
    GlobalCounters.reset()
@@ -100,7 +100,7 @@ class TestFuse(unittest.TestCase):
    q = (x @ wq).contiguous()
    k = (x @ wk).contiguous()
    v = (x @ wv).contiguous()
-    attn = q.scaled_dot_product_attention(k, v)
+    attn = q.scaled_dot_product_attention(k, v).fuse()
    s = attn.schedule()
    self.assertEqual(len(s), 4) # 3 matmul and 1 attention

@@ -121,7 +121,7 @@ class TestFuse(unittest.TestCase):
  def test_mismatch_reduce(self):
    a = Tensor.ones(16, 10).contiguous().realize()
    b = Tensor.ones(16, 20).contiguous().realize()
-    c = (a.sum(axis=1) + b.sum(axis=1))
+    c = (a.sum(axis=1) + b.sum(axis=1)).fuse()
    self.assertListEqual(c.tolist(), [30]*16)

  @unittest.skipUnless(Device.DEFAULT == "METAL", "METAL TC")
@@ -129,7 +129,7 @@ class TestFuse(unittest.TestCase):
    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))
+    out = (C + (A @ B)).fuse()
    out.realize()

 class TestSoftmaxFusion(unittest.TestCase):
@@ -180,7 +180,7 @@ class TestSoftmaxFusion(unittest.TestCase):

    print("*** auto single kernel softmax ***")
    with Context(NOOPT=1, DEBUG=max(DEBUG.value, 2)):
-      out = self.test.contiguous().softmax(-1)
+      out = self.test.contiguous().softmax(-1).fuse()
      run_one_schedule_item(out)

    np.testing.assert_allclose(sout.numpy(), out.numpy(), atol=3e-7)
@@ -3,7 +3,7 @@ import torch
 import unittest, copy, mmap, random, math, array
 from tinygrad import Tensor, Device, dtypes
 from tinygrad.tensor import _METADATA
-from tinygrad.helpers import Context, getenv, temp, mv_address
+from tinygrad.helpers import getenv, temp, mv_address
 from extra.gradcheck import numerical_jacobian, jacobian, gradcheck
 from hypothesis import given, settings, strategies as strat
 from tinygrad.device import is_dtype_supported
@@ -846,16 +846,6 @@ class TestTensorMetadata(unittest.TestCase):
    #self.assertEqual(len(bw), 1)
    #self.assertEqual(bw[0].name, "sigmoid")

-  def test_tracemeta_0(self):
-    with Context(TRACEMETA=0):
-      x = Tensor.rand(3, requires_grad=True)
-      y = Tensor.rand(3, requires_grad=True)
-      out = (x.relu() * y.sigmoid()).sum()
-      self.assertIsNone(out.uop.metadata)
-      self.assertIsNone(out.uop.src[0].metadata)
-      si = out.schedule()[-1]
-      self.assertEqual(si.metadata, ())
-
 class TestIdxUpcast(unittest.TestCase):
  def _find_op(self, ast: UOp, op: Ops):
    if ast.op is op: return ast
@@ -32,8 +32,8 @@ class TestTiny(unittest.TestCase):
    self.assertListEqual(out.tolist(), [2]*16)

  def test_cat(self):
-    out = Tensor.cat(Tensor.ones(8).contiguous(), Tensor.zeros(8).contiguous())
-    self.assertListEqual(out.tolist(), [1]*8+[0]*8)
+    out = Tensor.cat(Tensor.ones(8).contiguous(), Tensor.ones(8).contiguous())
+    self.assertListEqual(out.tolist(), [1]*16)

  def test_sum(self):
    out = Tensor.ones(256).contiguous().sum()
@@ -62,7 +62,7 @@ class TestTiny(unittest.TestCase):
    out = Tensor.rand(10)
    for x in out.tolist():
      self.assertGreaterEqual(x, 0.0)
-      self.assertLess(x, 1.0)
+      self.assertLessEqual(x, 1.0)

  # *** JIT (for Python speed) ***

@@ -138,7 +138,9 @@ class TestTiny(unittest.TestCase):
      nn.Conv2d(8, 8, 5), Tensor.relu]

    # replace random weights with ones
-    Tensor.realize(*[p.replace(Tensor.ones_like(p).contiguous()) for p in nn.state.get_parameters(layers)])
+    # 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_()
@@ -517,7 +517,7 @@ class TestUOpStr(unittest.TestCase):

 class TestUPatHelpers(unittest.TestCase):
  def test_location(self):
-    self.assertEqual(sym.patterns[-1][0].location[0].replace("\\", "/").split("/")[-1], "math.py")
+    self.assertEqual(sym.patterns[-1][0].location[0].replace("\\", "/").split("/")[-1], "symbolic.py")
    self.assertEqual(shared_spec.patterns[0][0].location[0].replace("\\", "/").split("/")[-1], "spec.py")
    test_upat = UPat(Ops.CONST, dtypes.bool)
    self.assertEqual(test_upat.location[0].split("/")[-1], __file__.replace("\\", "/").split("/")[-1])
@@ -1,443 +0,0 @@
-import unittest, math
-
-from tinygrad import Tensor, Device, dtypes, Context
-from tinygrad.engine.realize import ExecItem, get_runner
-from tinygrad.helpers import CI
-from tinygrad.renderer.ptx import PTXRenderer
-import numpy as np
-
-from extra.thunder.tiny.tk import WARP_THREADS
-from extra.thunder.tiny.tk.kernel import Kernel
-
-@unittest.skipUnless(Device.DEFAULT in ["CUDA", "NV"], "only cuda")
-@unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "no ptx")
-class TestTK(unittest.TestCase):
-  @unittest.skipIf(CI, "no wmma in ci")
-  def test_simple_matmul(self):
-    N = 32
-    BLOCK_SIZE = 16
-    with Kernel((N // BLOCK_SIZE, N // BLOCK_SIZE, 1), WARP_THREADS) as ker:
-      warp = ker.warp
-
-      c = ker.gl((1, 1, N, N), dtypes.float32)
-      a = ker.gl((1, 1, N, N), dtypes.bfloat16)
-      b = ker.gl((1, 1, N, N), dtypes.bfloat16)
-
-      a_smem = ker.st((BLOCK_SIZE, BLOCK_SIZE), dtypes.bfloat16)
-      b_smem = ker.st((BLOCK_SIZE, BLOCK_SIZE), dtypes.bfloat16)
-      c_smem = ker.st((BLOCK_SIZE, BLOCK_SIZE), dtypes.float32)
-
-      a_reg = ker.rt((BLOCK_SIZE, BLOCK_SIZE), dtypes.bfloat16)
-      b_reg = ker.rt((BLOCK_SIZE, BLOCK_SIZE), dtypes.bfloat16)
-      c_reg = ker.rt((BLOCK_SIZE, BLOCK_SIZE), dtypes.float32)
-
-      col, row = ker.blockIdx_x, ker.blockIdx_y
-
-      c_reg = warp.zero(c_reg)
-      for tile in ker.range(N // BLOCK_SIZE):
-        a_smem = warp.load(a_smem, a, (), (0, 0, row, tile), axis=2)
-        b_smem = warp.load(b_smem, b, (), (0, 0, tile, col), axis=2)
-
-        a_reg = warp.load(a_reg, a_smem)
-        b_reg = warp.load(b_reg, b_smem, transpose=True)
-
-        c_reg = warp.mma_AB(c_reg, a_reg, b_reg)
-      c_reg = ker.endrange()
-
-      c_smem = warp.store(c_smem, c_reg)
-      c = warp.store(c, c_smem, (0, 0, row, col), (), axis=2)
-
-      sink = ker.finish()
-
-    with Context(DEBUG=0):
-      a = Tensor.rand(1, 1, N, N, dtype="bfloat16").contiguous()
-      b = Tensor.rand(1, 1, N, N, dtype="bfloat16").contiguous()
-      c = Tensor.empty(1, 1, N, N, dtype="float32")
-      Tensor.realize(a, b, c)
-
-    ei = ExecItem(get_runner(Device.DEFAULT, sink), [t.uop.buffer for t in (c, a, b)])
-    for _ in range(5): ei.run(wait=True)
-    c = c.float()
-
-    ref = a.matmul(b, dtype=dtypes.float32).float()
-
-    np.testing.assert_allclose(c.numpy(), ref.numpy())
-
-  @unittest.skipIf(CI, "no wmma in ci")
-  def test_simple_matmul_transposed(self):
-    N = 32
-    BLOCK_SIZE = 16
-    with Kernel((N // BLOCK_SIZE, N // BLOCK_SIZE, 1), WARP_THREADS) as ker:
-      warp = ker.warp
-
-      c = ker.gl((1, 1, N, N), dtypes.float32)
-      a = ker.gl((1, 1, N, N), dtypes.bfloat16)
-      b = ker.gl((1, 1, N, N), dtypes.bfloat16)
-
-      a_smem = ker.st((BLOCK_SIZE, BLOCK_SIZE), dtypes.bfloat16)
-      b_smem = ker.st((BLOCK_SIZE, BLOCK_SIZE), dtypes.bfloat16)
-      c_smem = ker.st((BLOCK_SIZE, BLOCK_SIZE), dtypes.float32)
-
-      a_reg = ker.rt((BLOCK_SIZE, BLOCK_SIZE), dtypes.bfloat16)
-      b_reg = ker.rt((BLOCK_SIZE, BLOCK_SIZE), dtypes.bfloat16)
-      c_reg = ker.rt((BLOCK_SIZE, BLOCK_SIZE), dtypes.float32)
-
-      col, row = ker.blockIdx_x, ker.blockIdx_y
-
-      c_reg = warp.zero(c_reg)
-      for tile in ker.range(N // BLOCK_SIZE):
-        a_smem = warp.load(a_smem, a, (), (0, 0, row, tile), axis=2)
-        b_smem = warp.load(b_smem, b, (), (0, 0, col, tile), axis=2)
-
-        a_reg = warp.load(a_reg, a_smem)
-        b_reg = warp.load(b_reg, b_smem)
-
-        c_reg = warp.mma_ABt(c_reg, a_reg, b_reg)
-      c_reg = ker.endrange()
-
-      c_smem = warp.store(c_smem, c_reg)
-      c = warp.store(c, c_smem, (0, 0, row, col), (), axis=2)
-
-      sink = ker.finish()
-
-    with Context(DEBUG=0):
-      a = Tensor.rand(1, 1, N, N, dtype="bfloat16").contiguous()
-      b = Tensor.rand(1, 1, N, N, dtype="bfloat16").contiguous()
-      c = Tensor.empty(1, 1, N, N, dtype="float32")
-      Tensor.realize(a, b, c)
-
-    ei = ExecItem(get_runner(Device.DEFAULT, sink), [t.uop.buffer for t in (c, a, b)])
-    for _ in range(5): ei.run(wait=True)
-    c = c.float()
-
-    ref = a.matmul(b.transpose(2, 3), dtype=dtypes.float32).float()
-
-    np.testing.assert_allclose(c.numpy(), ref.numpy())
-
-  def test_load_store(self):
-    N = 32
-    BLOCK_SIZE = 16
-    with Kernel((N // BLOCK_SIZE, N // BLOCK_SIZE, 1), WARP_THREADS) as ker:
-      warp = ker.warp
-
-      b = ker.gl((1, 1, N, N), dtypes.float32)
-      a = ker.gl((1, 1, N, N), dtypes.float32)
-
-      a_smem = ker.st((BLOCK_SIZE, BLOCK_SIZE), dtypes.float32)
-      b_smem = ker.st((BLOCK_SIZE, BLOCK_SIZE), dtypes.float32)
-
-      a_reg = ker.rt((BLOCK_SIZE, BLOCK_SIZE), dtypes.float32)
-      b_reg = ker.rt((BLOCK_SIZE, BLOCK_SIZE), dtypes.float32)
-
-      col, row = ker.blockIdx_x, ker.blockIdx_y
-
-      a_smem = warp.load(a_smem, a, (), (0, 0, row, col), axis=2)
-      a_reg = warp.load(a_reg, a_smem)
-      b_reg = warp.copy(b_reg, a_reg)
-      b_smem = warp.store(b_smem, b_reg)
-      b = warp.store(b, b_smem, (0, 0, row, col), (), axis=2)
-
-      sink = ker.finish()
-
-    with Context(DEBUG=0):
-      a = Tensor.rand(1, 1, N, N, dtype="float32").contiguous()
-      b = Tensor.empty(1, 1, N, N, dtype="float32")
-      Tensor.realize(a, b)
-
-    ei = ExecItem(get_runner(Device.DEFAULT, sink), [t.uop.buffer for t in (b, a)])
-    for _ in range(5): ei.run(wait=True)
-    b = b.float()
-
-    ref = a.float()
-
-    np.testing.assert_allclose(b.numpy(), ref.numpy())
-
-  def test_add(self):
-    N = 32
-    BLOCK_SIZE = 16
-    with Kernel((1, 1, 1), WARP_THREADS) as ker:
-      warp = ker.warp
-
-      b = ker.gl((1, 1, N, N), dtypes.float32)
-      a = ker.gl((1, 1, N, N), dtypes.float32)
-
-      a_smem = ker.st((BLOCK_SIZE, BLOCK_SIZE), dtypes.float32)
-
-      a_reg = ker.rt((BLOCK_SIZE, BLOCK_SIZE), dtypes.float32)
-
-      for tile_row in ker.range(N // BLOCK_SIZE):
-        for tile_col in ker.range(N // BLOCK_SIZE):
-          a_smem = warp.load(a_smem, a, (), (0, 0, tile_row, tile_col), axis=2)
-          a_reg = warp.load(a_reg, a_smem)
-
-          a_reg += 1
-
-          a_smem = warp.store(a_smem, a_reg)
-          b = warp.store(b, a_smem, (0, 0, tile_row, tile_col), (), axis=2)
-
-      sink = ker.finish()
-
-      with Context(DEBUG=0):
-        a = Tensor.rand(1, 1, N, N, dtype="float32").contiguous()
-        b = Tensor.empty(1, 1, N, N, dtype="float32")
-        Tensor.realize(a, b)
-
-      ei = ExecItem(get_runner(Device.DEFAULT, sink), [t.uop.buffer for t in (b, a)])
-      for _ in range(5): ei.run(wait=True)
-      b = b.float()
-
-      ref = a.float() + 1
-
-      np.testing.assert_allclose(b.numpy(), ref.numpy())
-
-  def test_max(self):
-    N = 16
-    BLOCK_SIZE = 16
-    with Kernel((1, 1, 1), WARP_THREADS) as ker:
-      warp = ker.warp
-
-      b = ker.gl((1, 1, N, N), dtypes.float32)
-      a = ker.gl((1, 1, N, N), dtypes.float32)
-
-      a_smem = ker.st((BLOCK_SIZE, BLOCK_SIZE), dtypes.float32)
-      b_smem = ker.st((BLOCK_SIZE, BLOCK_SIZE), dtypes.float32)
-
-      a_reg = ker.rt((BLOCK_SIZE, BLOCK_SIZE), dtypes.float32)
-      b_reg = ker.rt((BLOCK_SIZE, BLOCK_SIZE), dtypes.float32)
-
-      max_reg = ker.rv(BLOCK_SIZE, dtypes.float32, "ortho")
-
-      for tile_row in ker.range(N // BLOCK_SIZE):
-        max_reg = warp.neg_inf(max_reg.after(tile_row))
-
-        for tile_col in ker.range(N // BLOCK_SIZE):
-          a_smem = warp.load(a_smem, a, (), (0, 0, tile_row, tile_col), axis=2)
-          a_reg = warp.load(a_reg, a_smem)
-          max_reg = warp.row_reduce(max_reg, a_reg, lambda a, b: a.maximum(b))
-        max_reg = ker.endrange()
-
-        b_reg = warp.map(b_reg, lambda _, idx: max_reg[idx[0], 0, (idx[2]%4)//2])
-        b_smem = warp.store(b_smem, b_reg)
-
-        for tile_col in ker.range(N // BLOCK_SIZE):
-          b = warp.store(b, b_smem, (0, 0, tile_row, tile_col), (), axis=2)
-
-      sink = ker.finish()
-
-    with Context(DEBUG=0):
-      a = Tensor.rand(1, 1, N, N, dtype="float32").contiguous()
-      b = Tensor.empty(1, 1, N, N, dtype="float32")
-      Tensor.realize(a, b)
-
-    ei = ExecItem(get_runner(Device.DEFAULT, sink), [t.uop.buffer for t in (b, a)])
-    for _ in range(5): ei.run(wait=True)
-    b = b.float()
-
-    ref = a.float().max(axis=3, keepdim=True).expand(a.shape)
-
-    np.testing.assert_allclose(b.numpy(), ref.numpy())
-
-  def test_max_nonsquare(self):
-    N, M = 16, 64
-    BLOCK_N, BLOCK_M = 16, 64
-    with Kernel((1, 1, 1), WARP_THREADS) as ker:
-      warp = ker.warp
-
-      b = ker.gl((1, 1, N, M), dtypes.float32)
-      a = ker.gl((1, 1, N, M), dtypes.float32)
-
-      a_smem = ker.st((BLOCK_N, BLOCK_M), dtypes.float32)
-      b_smem = ker.st((BLOCK_N, BLOCK_M), dtypes.float32)
-
-      a_reg = ker.rt((BLOCK_N, BLOCK_M), dtypes.float32)
-      b_reg = ker.rt((BLOCK_N, BLOCK_M), dtypes.float32)
-
-      max_reg = ker.rv(BLOCK_N, dtypes.float32, "ortho")
-
-      for tile_row in ker.range(N // BLOCK_N):
-        max_reg = warp.neg_inf(max_reg.after(tile_row))
-
-        for tile_col in ker.range(M // BLOCK_M):
-          a_smem = warp.load(a_smem, a, (), (0, 0, tile_row, tile_col), axis=2)
-          a_reg = warp.load(a_reg, a_smem)
-          max_reg = warp.row_reduce(max_reg, a_reg, lambda a, b: a.maximum(b))
-        max_reg = ker.endrange()
-
-        b_reg = warp.map(b_reg, lambda _, idx: max_reg[idx[0], 0, (idx[2]%4)//2])
-        b_smem = warp.store(b_smem, b_reg)
-
-        for tile_col in ker.range(M // BLOCK_M):
-          b = warp.store(b, b_smem, (0, 0, tile_row, tile_col), (), axis=2)
-
-      sink = ker.finish()
-
-    with Context(DEBUG=0):
-      a = Tensor.rand(1, 1, N, M, dtype="float32").contiguous()
-      b = Tensor.empty(1, 1, N, M, dtype="float32")
-      Tensor.realize(a, b)
-
-    ei = ExecItem(get_runner(Device.DEFAULT, sink), [t.uop.buffer for t in (b, a)])
-    for _ in range(5): ei.run(wait=True)
-    b = b.float()
-
-    ref = a.float().max(axis=3, keepdim=True).expand(a.shape)
-
-    np.testing.assert_allclose(b.numpy(), ref.numpy())
-
-  def test_sum(self):
-    N = 32
-    BLOCK_SIZE = 16
-    with Kernel((1, 1, 1), WARP_THREADS) as ker:
-      warp = ker.warp
-
-      b = ker.gl((1, 1, N, N), dtypes.float32)
-      a = ker.gl((1, 1, N, N), dtypes.float32)
-
-      a_smem = ker.st((BLOCK_SIZE, BLOCK_SIZE), dtypes.float32)
-      b_smem = ker.st((BLOCK_SIZE, BLOCK_SIZE), dtypes.float32)
-
-      a_reg = ker.rt((BLOCK_SIZE, BLOCK_SIZE), dtypes.float32)
-      b_reg = ker.rt((BLOCK_SIZE, BLOCK_SIZE), dtypes.float32)
-
-      sum_reg = ker.rv(BLOCK_SIZE, dtypes.float32, "ortho")
-
-      for tile_row in ker.range(N // BLOCK_SIZE):
-        sum_reg = warp.zero(sum_reg.after(tile_row))
-
-        for tile_col in ker.range(N // BLOCK_SIZE):
-          a_smem = warp.load(a_smem, a, (), (0, 0, tile_row, tile_col), axis=2)
-          a_reg = warp.load(a_reg, a_smem)
-          sum_reg = warp.row_reduce(sum_reg, a_reg, lambda a, b: a + b)
-        sum_reg = ker.endrange()
-
-        b_reg = warp.map(b_reg, lambda _, idx: sum_reg[idx[0], 0, (idx[2]%4)//2])
-        b_smem = warp.store(b_smem, b_reg)
-
-        for tile_col in ker.range(N // BLOCK_SIZE):
-          b = warp.store(b, b_smem, (0, 0, tile_row, tile_col), (), axis=2)
-
-      sink = ker.finish()
-
-    with Context(DEBUG=0):
-      a = Tensor.rand(1, 1, N, N, dtype="float32").contiguous()
-      b = Tensor.empty(1, 1, N, N, dtype="float32")
-      Tensor.realize(a, b)
-
-    ei = ExecItem(get_runner(Device.DEFAULT, sink), [t.uop.buffer for t in (b, a)])
-    for _ in range(5): ei.run(wait=True)
-    b = b.float()
-
-    ref = a.float().sum(axis=3, keepdim=True).expand(a.shape)
-
-    np.testing.assert_allclose(b.numpy(), ref.numpy(), atol=1e-5, rtol=1e-5)
-
-  def test_sum_nonsquare(self):
-    N, M = 16, 64
-    BLOCK_N, BLOCK_M = 16, 64
-    with Kernel((1, 1, 1), WARP_THREADS) as ker:
-      warp = ker.warp
-
-      b = ker.gl((1, 1, N, M), dtypes.float32)
-      a = ker.gl((1, 1, N, M), dtypes.float32)
-
-      a_smem = ker.st((BLOCK_N, BLOCK_M), dtypes.float32)
-      b_smem = ker.st((BLOCK_N, BLOCK_M), dtypes.float32)
-
-      a_reg = ker.rt((BLOCK_N, BLOCK_M), dtypes.float32)
-      b_reg = ker.rt((BLOCK_N, BLOCK_M), dtypes.float32)
-
-      sum_reg = ker.rv(BLOCK_N, dtypes.float32, "ortho")
-
-      for tile_row in ker.range(N // BLOCK_N):
-        sum_reg = warp.zero(sum_reg.after(tile_row))
-
-        for tile_col in ker.range(M // BLOCK_M):
-          a_smem = warp.load(a_smem, a, (), (0, 0, tile_row, tile_col), axis=2)
-          a_reg = warp.load(a_reg, a_smem)
-          sum_reg = warp.row_reduce(sum_reg, a_reg, lambda a, b: a + b)
-        sum_reg = ker.endrange()
-
-        b_reg = warp.map(b_reg, lambda _, idx: sum_reg[idx[0], 0, (idx[2]%4)//2])
-        b_smem = warp.store(b_smem, b_reg)
-
-        for tile_col in ker.range(M // BLOCK_M):
-          b = warp.store(b, b_smem, (0, 0, tile_row, tile_col), (), axis=2)
-
-      sink = ker.finish()
-
-    with Context(DEBUG=0):
-      a = Tensor.rand(1, 1, N, M, dtype="float32").contiguous()
-      b = Tensor.empty(1, 1, N, M, dtype="float32")
-      Tensor.realize(a, b)
-
-    ei = ExecItem(get_runner(Device.DEFAULT, sink), [t.uop.buffer for t in (b, a)])
-    for _ in range(5): ei.run(wait=True)
-    b = b.float()
-
-    ref = a.float().sum(axis=3, keepdim=True).expand(a.shape)
-
-    np.testing.assert_allclose(b.numpy(), ref.numpy(), atol=1e-5, rtol=1e-5)
-
-  @unittest.skip("fake range not ended")
-  def test_softmax(self):
-    N = 32
-    BLOCK_SIZE = 16
-    with Kernel((1, 1, 1), WARP_THREADS) as ker:
-      warp = ker.warp
-
-      b = ker.gl((1, 1, BLOCK_SIZE, N), dtypes.float32)
-      a = ker.gl((1, 1, BLOCK_SIZE, N), dtypes.float32)
-
-      a_smem = ker.st((BLOCK_SIZE, BLOCK_SIZE), dtypes.float32)
-
-      a_reg = ker.rt((BLOCK_SIZE, BLOCK_SIZE), dtypes.float32)
-
-      max_vec_last = ker.rv(BLOCK_SIZE, dtypes.float32, "ortho")
-      max_vec = ker.rv(BLOCK_SIZE, dtypes.float32, "ortho")
-      norm_vec = ker.rv(BLOCK_SIZE, dtypes.float32, "ortho")
-
-      max_vec = warp.neg_inf(max_vec)
-      norm_vec = warp.zero(norm_vec)
-
-      for tile_col in ker.range(N // BLOCK_SIZE):
-        a_smem = warp.load(a_smem, a, (), (0, 0, 0, tile_col), axis=2)
-        a_reg = warp.load(a_reg, a_smem)
-
-        a_reg *= 1.0 / math.log(2)
-
-        max_vec_last = warp.copy(max_vec_last.after(tile_col), max_vec)
-        max_vec = warp.row_reduce(max_vec, a_reg, lambda a, b: a.maximum(b))
-        a_reg = (a_reg - max_vec).exp2()
-        max_vec_last = (max_vec_last - max_vec).exp2()
-        norm_vec *= max_vec_last
-        norm_vec = warp.row_reduce(norm_vec, a_reg, lambda a, b: a + b)
-      norm_vec = ker.endrange()
-
-      for tile_col in ker.range(N // BLOCK_SIZE):
-        a_smem = warp.load(a_smem, a, (), (0, 0, 0, tile_col), axis=2)
-        a_reg = warp.load(a_reg, a_smem)
-
-        a_reg *= 1.0 / math.log(2)
-        a_reg = (a_reg - max_vec).exp2()
-        a_reg /= norm_vec
-
-        a_smem = warp.store(a_smem, a_reg)
-        b = warp.store(b, a_smem, (0, 0, 0, tile_col), (), axis=2)
-
-      sink = ker.finish()
-
-    with Context(DEBUG=0):
-      a = Tensor.rand(1, 1, BLOCK_SIZE, N, dtype="float32")
-      b = Tensor.empty(1, 1, BLOCK_SIZE, N, dtype="float32")
-      Tensor.realize(a, b)
-
-    ei = ExecItem(get_runner(Device.DEFAULT, sink), [t.uop.buffer for t in (b, a)])
-    for _ in range(5): ei.run(wait=True)
-    b = b.float()
-
-    ref = a.float().softmax(axis=3)
-
-    np.testing.assert_allclose(b.numpy(), ref.numpy(), atol=1e-5, rtol=1e-5)
-
-if __name__ == "__main__":
-  unittest.main()
@@ -1,85 +0,0 @@
-import ctypes, subprocess, tempfile, unittest
-from tinygrad.helpers import WIN
-from tinygrad.runtime.support.c import Struct
-
-class TestAutogen(unittest.TestCase):
-  def test_packed_struct_sizeof(self):
-    layout = [('a', ctypes.c_char), ('b', ctypes.c_int, 5), ('c', ctypes.c_char)]
-    class X(ctypes.Structure): _fields_, _layout_ = layout, 'gcc-sysv'
-    class Y(ctypes.Structure): _fields_, _pack_, _layout_ = layout, 1, 'ms'
-    class Z(Struct): _packed_, _fields_ = True, layout
-    self.assertNotEqual(ctypes.sizeof(X), 4) # ctypes bug! gcc-13.3.0 says this should have size 4
-    self.assertEqual(ctypes.sizeof(Y), 6)
-    self.assertEqual(ctypes.sizeof(Z), 3)
-    layout = [('a', ctypes.c_int, 31), ('b', ctypes.c_int, 31), ('c', ctypes.c_int, 1), ('d', ctypes.c_int, 1)]
-    class Foo(ctypes.Structure): _fields_, _layout_ = layout, 'gcc-sysv'
-    class Bar(ctypes.Structure): _fields_, _pack_, _layout_ = layout, 1, 'ms'
-    class Baz(Struct): _fields_, _packed_ = layout, True
-    self.assertEqual(ctypes.sizeof(Foo), 12)
-    self.assertEqual(ctypes.sizeof(Bar), 12)
-    self.assertEqual(ctypes.sizeof(Baz), 8)
-
-  @unittest.skipIf(WIN, "doesn't compile on windows")
-  def test_packed_struct_interop(self):
-    class Baz(Struct): pass
-    Baz._packed_ = True
-    Baz._fields_ = [('a', ctypes.c_int, 30), ('b', ctypes.c_int, 30), ('c', ctypes.c_int, 2), ('d', ctypes.c_int, 2)]
-    src = '''
-      struct __attribute__((packed)) baz {
-        int a:30;
-        int b:30;
-        int c:2;
-        int d:2;
-      };
-
-      int test(struct baz x) {
-        return x.a + x.b + x.c + x.d;
-      }
-    '''
-    args = ('-x', 'c', '-fPIC', '-shared')
-    with tempfile.NamedTemporaryFile(suffix=".so") as f:
-      subprocess.check_output(('clang',) + args + ('-', '-o', f.name), input=src.encode('utf-8'))
-      b = Baz(0xAA000, 0x00BB0, 0, 1)
-      test = ctypes.CDLL(f.name).test
-      test.argtypes = [Baz]
-      self.assertEqual(test(b), b.a + b.b + b.c + b.d)
-
-  @unittest.skipIf(WIN, "doesn't compile on windows")
-  def test_packed_structs(self):
-    NvU32 = ctypes.c_uint32
-    NvU64 = ctypes.c_uint64
-    class FWSECLIC_READ_VBIOS_DESC(Struct): pass
-    FWSECLIC_READ_VBIOS_DESC._packed_ = True
-    FWSECLIC_READ_VBIOS_DESC._fields_ = [
-      ('version', NvU32),
-      ('size', NvU32),
-      ('gfwImageOffset', NvU64),
-      ('gfwImageSize', NvU32),
-      ('flags', NvU32),
-    ]
-    class FWSECLIC_FRTS_REGION_DESC(Struct): pass
-    FWSECLIC_FRTS_REGION_DESC._packed_ = True
-    FWSECLIC_FRTS_REGION_DESC._fields_ = [
-      ('version', NvU32),
-      ('size', NvU32),
-      ('frtsRegionOffset4K', NvU32),
-      ('frtsRegionSize', NvU32),
-      ('frtsRegionMediaType', NvU32),
-    ]
-    class FWSECLIC_FRTS_CMD(Struct): pass
-    FWSECLIC_FRTS_CMD._packed_ = True
-    FWSECLIC_FRTS_CMD._fields_ = [
-      ('readVbiosDesc', FWSECLIC_READ_VBIOS_DESC),
-      ('frtsRegionDesc', FWSECLIC_FRTS_REGION_DESC),
-    ]
-    read_vbios_desc = FWSECLIC_READ_VBIOS_DESC(version=0x1, size=ctypes.sizeof(FWSECLIC_READ_VBIOS_DESC), flags=2)
-    frst_reg_desc = FWSECLIC_FRTS_REGION_DESC(version=0x1, size=ctypes.sizeof(FWSECLIC_FRTS_REGION_DESC),
-      frtsRegionOffset4K=0xdead, frtsRegionSize=0x100, frtsRegionMediaType=2)
-    frts_cmd = FWSECLIC_FRTS_CMD(readVbiosDesc=read_vbios_desc, frtsRegionDesc=frst_reg_desc)
-    assert int.from_bytes(frts_cmd, 'little') == 0x2000001000000dead0000001400000001000000020000000000000000000000000000001800000001
-    assert int.from_bytes(frts_cmd.readVbiosDesc, 'little') == int.from_bytes(read_vbios_desc, 'little')
-    assert int.from_bytes(frts_cmd.frtsRegionDesc, 'little') == int.from_bytes(frst_reg_desc, 'little')
-    assert frts_cmd.readVbiosDesc.__class__ is FWSECLIC_READ_VBIOS_DESC
-    assert frts_cmd.frtsRegionDesc.__class__ is FWSECLIC_FRTS_REGION_DESC
-
-if __name__ == "__main__": unittest.main()
@@ -62,7 +62,6 @@ class TestConv(unittest.TestCase):
    np.testing.assert_allclose(r1.numpy(), np.maximum(out.numpy(), 0), atol=1e-5)
    np.testing.assert_allclose(r2.numpy(), np.where(out.numpy() > 0, out.numpy(), (np.exp(out.numpy()) - 1)), atol=1e-5)

-  @unittest.skip("this test is flaky")
  def test_two_overlapping_binops_no_rerun_wino(self):
    with Context(WINO=1):
      x = Tensor.randn(1,4,16,16)
@@ -30,7 +30,7 @@ class TestDevice(unittest.TestCase):

  @unittest.skipIf(WIN and CI, "skipping windows test") # TODO: subproccess causes memory violation?
  def test_env_overwrite_default_compiler(self):
-    expect_failure = "\ntry: assert Device[Device.DEFAULT].compiler is None;\nexcept Exception: pass"
+    expect_failure = "\ntry: assert Device[Device.DEFAULT].compiler is None;\nexcept RuntimeError: pass"

    if Device.DEFAULT == "CPU":
      from tinygrad.runtime.support.compiler_cpu import CPULLVMCompiler, ClangJITCompiler
@@ -81,20 +81,20 @@ class TestCompiler(unittest.TestCase):
  def test_compile_cached(self):
    diskcache_put("key", "123", None) # clear cache
    getenv.cache_clear()
-    with Context(CCACHE=1):
+    with Context(DISABLE_COMPILER_CACHE=0):
      self.assertEqual(MockCompiler("key").compile_cached("123"), str.encode("123"))
      self.assertEqual(diskcache_get("key", "123"), str.encode("123"))

  def test_compile_cached_disabled(self):
    diskcache_put("disabled_key", "123", None) # clear cache
    getenv.cache_clear()
-    with Context(CCACHE=0):
+    with Context(DISABLE_COMPILER_CACHE=1):
      self.assertEqual(MockCompiler("disabled_key").compile_cached("123"), str.encode("123"))
      self.assertIsNone(diskcache_get("disabled_key", "123"))

  def test_device_compile(self):
    getenv.cache_clear()
-    with Context(CCACHE=0):
+    with Context(DISABLE_COMPILER_CACHE=1):
      a = Tensor([0.,1.], device=Device.DEFAULT).realize()
      (a + 1).realize()

@@ -894,7 +894,7 @@ class TestNumpy(unittest.TestCase):
    a = Tensor([[1, 2, 3],
                [4, 5, 6],
                [7, 8, 9]])
-    self.assertIs(a[...], a)
+    self.assertIsNot(a[...], a)
    numpy_testing_assert_equal_helper(a[...], a)
    # `a[...]` was `a` in numpy <1.9.
    #numpy_testing_assert_equal_helper(data_ptr(a[...]), data_ptr(a))
@@ -1037,9 +1037,9 @@ class TestNumpy(unittest.TestCase):
    # Before `...` would return a itself.
    a = Tensor([5])

-    self.assertIs(a, a[()])
-    self.assertIs(a, a[...])
-    self.assertIs(a, a[:])
+    self.assertIsNot(a, a[()])
+    self.assertIsNot(a, a[...])
+    self.assertIsNot(a, a[:])

  def test_broaderrors_indexing(self):
    a = Tensor.zeros(5, 5)
@@ -643,10 +643,6 @@ class TestSymbolic(unittest.TestCase):
    self.helper_test_variable(lidx+(gidx//4)*8+2*(gidx%4), 0, 372, "(lidx+(gidx*2))")
    self.helper_test_variable(lidx+2*(gidx%4)+(gidx//4)*8, 0, 372, "(lidx+(gidx*2))")

-  def test_div_mod_recombine_partial(self):
-    gidx = Variable("gidx", 0, 15)
-    self.helper_test_variable((gidx//2)%4+(gidx//8)*4, 0, 7, "gidx//2")
-
  def test_div_mod_recombine_folded_mod(self):
    a = Variable("a", 0, 2)
    b = Variable("b", 0, 100)
@@ -366,8 +366,8 @@ def load_profile(lst:list[ProfileEvent]) -> dict:
        else: v["events"].append({"event":"free", "ts":ts, "key":key, "arg": {"users":[u("<IIBB") for _ in range(u("<I")[0])]}})
  return {"dur":total_dur, "peak":global_peak, "layout":layout, "markers":markers}

-class TestVizProfiler(BaseTestViz):
-  def test_node(self):
+class TestVizProfiler(unittest.TestCase):
+  def test_perfetto_node(self):
    prof = [ProfileRangeEvent(device='NV', name='E_2', st=decimal.Decimal(1000), en=decimal.Decimal(1010), is_copy=False),
            ProfileDeviceEvent(device='NV', comp_tdiff=decimal.Decimal(-1000), copy_tdiff=decimal.Decimal(-100))]

@@ -381,7 +381,7 @@ class TestVizProfiler(BaseTestViz):
    self.assertEqual(event['dur'], 10)
    assert event['ref'] is None

-  def test_copy_node(self):
+  def test_perfetto_copy_node(self):
    prof = [ProfileRangeEvent(device='NV', name='COPYxx', st=decimal.Decimal(1000), en=decimal.Decimal(1010), is_copy=True),
            ProfileRangeEvent(device='NV:2', name='COPYxx', st=decimal.Decimal(1000), en=decimal.Decimal(1010), is_copy=True),
            ProfileDeviceEvent(device='NV', comp_tdiff=decimal.Decimal(-1000), copy_tdiff=decimal.Decimal(-100)),
@@ -399,7 +399,7 @@ class TestVizProfiler(BaseTestViz):

    self.assertEqual(j["dur"], (event2["st"]+event2["dur"])-event["st"])

-  def test_graph(self):
+  def test_perfetto_graph(self):
    prof = [ProfileDeviceEvent(device='NV', comp_tdiff=decimal.Decimal(-1000), copy_tdiff=decimal.Decimal(-100)),
            ProfileDeviceEvent(device='NV:1', comp_tdiff=decimal.Decimal(-500), copy_tdiff=decimal.Decimal(-50)),
            ProfileGraphEvent(ents=[ProfileGraphEntry(device='NV', name='E_25_4n2', st_id=0, en_id=1, is_copy=False),
@@ -436,12 +436,6 @@ class TestVizProfiler(BaseTestViz):
    sz = len(get_profile(prof))
    self.assertLessEqual(sz/n_events, 26)

-  def test_calltrace(self):
-    def fxn(): return Tensor.empty(10).mul(2).realize()
-    fxn()
-    trace = get_viz_list()[0]["steps"][0]["trace"]
-    assert any(fxn.__code__.co_filename == f and fxn.__code__.co_firstlineno == l for f,l,*_ in trace), str(trace)
-
  # can pack up to 1hr 11 min of trace events
  def test_trace_duration(self):
    dur_mins = 72
@@ -4,7 +4,7 @@ from tinygrad.helpers import DEVECTORIZE, TRANSCENDENTAL, SPEC
 from tinygrad.uop.ops import PatternMatcher, graph_rewrite, UOp, pm_lower_index_dtype, Ops, UPat
 from tinygrad.uop.spec import type_verify, program_spec, kernel_spec
 from tinygrad.renderer import Renderer
-from tinygrad.dtype import dtypes, PtrDType
+from tinygrad.dtype import dtypes
 from tinygrad.helpers import panic

 # import all pattern matchers here
@@ -19,19 +19,13 @@ from tinygrad.codegen.simplify import pm_simplify_ranges, pm_flatten_range, pm_s
 from tinygrad.schedule.rangeify import pm_add_buffers_local, rangeify_codegen, pm_mops
 from tinygrad.codegen.late.linearizer import CFGContext, pm_split_ends, pm_add_control_flow, linearize

-pm_syntactic_sugar = PatternMatcher([
-  # INDEX on ptr INDEX concats them
-  (UPat(Ops.INDEX, name="i1").f(Ops.INDEX, name="i2", allow_any_len=True),
-   lambda i1,i2: i2.replace(src=i1.src+i2.src[1:]) if isinstance(i1.dtype, PtrDType) and not isinstance(i2.dtype, PtrDType) else None),
-])
-
 def full_rewrite_to_sink(sink:UOp, ren:Renderer|None=None, optimize:bool=True) -> UOp:
  if ren is None: ren = Renderer()

  if SPEC: type_verify(sink, kernel_spec)

  # preprocess
-  sink = graph_rewrite(sink, pm_mops+pm_syntactic_sugar, name="early movement ops", bottom_up=True)
+  sink = graph_rewrite(sink, pm_mops, name="early movement ops")

  # first we optimize
  if optimize:
@@ -136,7 +130,7 @@ def full_rewrite(sink:UOp, ren:Renderer|None=None) -> list[UOp]:
  """

  full_sink = full_rewrite_to_sink(sink, ren, optimize=sink.tag is None)
-  assert len(full_sink.ranges) == 0, f"all ranges must end by the sink, {full_sink.ranges}"
+  assert len(full_sink.ranges) == 0, "all ranges must end by the sink"
  lst = line_rewrite(linearize(full_sink), pm_linearize_cleanups)
  if SPEC: type_verify(lst, program_spec)
  return lst
@@ -26,35 +26,28 @@ def _split_dims(dims, max_sizes):
  return tuple(_dims[:2] if _dims[2] == 1 else _dims[0] if _dims[1:3] == [1,1] else _dims)

 def get_grouped_dims(prefix, dims:tuple[sint, ...], max_sizes:tuple[int, ...]|None, reverse=False) -> list[UOp]:
-  if reverse: return get_grouped_dims(prefix, dims[::-1], max_sizes)[::-1]
-  if max_sizes is None: limited = dims
-  else:
-    # try to group first: (a, b, c, d) -> (ab, c, d)
-    limited = grouped if (grouped := _group_dims(dims, max_sizes)) else dims
-    # check if grouping failed
-    if len(limited) > len(max_sizes): raise RuntimeError(f"cannot limit dim {dims=}, {max_sizes=}")
-    # try to split up dims: (a,) -> (b, c)
-    if limited == dims: limited = _split_dims(dims, max_sizes)
-  raw_idxs = [UOp(Ops.SPECIAL, dtypes.index, (sint_to_uop(s),), (f"{prefix}{i}")) for i,s in enumerate(limited)]
+  if reverse: dims = dims[::-1]
+  # try to group first: (a, b, c, d) -> (ab, c, d)
+  limited = (grouped if (grouped := _group_dims(dims, max_sizes)) else dims) if max_sizes is not None else dims
+  # check if grouping failed
+  if max_sizes is not None and len(limited) > len(max_sizes): raise RuntimeError(f"cannot limit dim {dims=}, {max_sizes=}")
+  # try to split up dims: (a,) -> (b, c)
+  if limited == dims: limited = _split_dims(dims, max_sizes) if max_sizes is not None else dims
+  ret = raw_idxs = [UOp(Ops.SPECIAL, dtypes.index, (sint_to_uop(s),), (f"{prefix}{i}")) for i,s in enumerate(limited)]
  if len(limited) < len(dims):
    ret = []
-    if (contraction:=get_contraction(dims, limited)) is None: raise RuntimeError(f"get_contraction should not be None {dims=} {limited=}")
+    if (contraction:=get_contraction(dims, limited)) is None: raise AssertionError(f"get_contraction should not be None {dims=} {limited=}")
    for idx, contraction_group in zip(raw_idxs, contraction):
      for c in contraction_group[:-1]:
        ret.append(idx % dims[c])
        idx //= dims[c]
      ret.append(idx)
-    return ret
-  elif (a:=len(limited)) > (b:=len(dims)):
-    if a == 2 and b == 1: return [raw_idxs[0] * limited[1] + raw_idxs[1]]
-    if a == 3 and b == 1: return [(raw_idxs[0] * limited[1] + raw_idxs[1]) * limited[2] + raw_idxs[2]]
-    if a == 3 and b == 2: return [raw_idxs[0] * limited[1] + raw_idxs[1], raw_idxs[2]]
-  elif limited != dims:
-    # Convert to 1D
-    flat = raw_idxs[0]*limited[1]+raw_idxs[1] if len(dims) == 2 else raw_idxs[0]*(limited[1]*limited[2])+raw_idxs[1]*limited[2]+raw_idxs[2]
-    # Get back original indices from 1D
-    return [flat//dims[1], flat%dims[1]] if len(dims) == 2 else [flat//(dims[2]*dims[1]), (flat//dims[2])%dims[1], flat%dims[2]]
-  return raw_idxs
+  elif len(limited) > len(dims):
+    a, b = len(limited), len(dims)
+    if a == 2 and b == 1: ret = [raw_idxs[0] * limited[1] + raw_idxs[1]]
+    if a == 3 and b == 1: ret = [raw_idxs[0] * (limited[1] * limited[2]) + raw_idxs[1] * limited[2] + raw_idxs[2]]
+    if a == 3 and b == 2: ret = [raw_idxs[0] * limited[1] + raw_idxs[1], raw_idxs[2]]
+  return ret[::-1] if reverse else ret

 def add_gpudims(ctx:Renderer, s:UOp):
  if s.arg is None: return None
@@ -4,7 +4,7 @@ from collections import defaultdict
 from dataclasses import dataclass
 from tinygrad.dtype import dtypes, ImageDType, DType, AddrSpace, Invalid, PtrDType
 from tinygrad.uop.ops import UOp, Ops, UPat, PatternMatcher, graph_rewrite, GroupOp, identity_element
-from tinygrad.uop.symbolic import uop_given_valid, parse_valid, sym, symbolic, invalid_gate
+from tinygrad.uop.symbolic import uop_given_valid, parse_valid, sym, symbolic_flat, invalid_gate
 from tinygrad.helpers import getenv, flatten, AMX, prod
 from tinygrad.renderer import Renderer

@@ -61,7 +61,7 @@ def expand_index(buf:UOp, vec:UOp):
  if getenv("UNSAFE_DISABLE_MASK", 0): vec = vec.get_idx()
  # generate the individual indexes
  midx = graph_rewrite(UOp.sink(*[buf.index(vec.gep(i), ptr=True) for i in range(vec.dtype.count)]),
-                       symbolic+load_store_indexing, name=f"index_buf_{buf.arg}")
+                       symbolic_flat+load_store_indexing, name=f"index_buf_{buf.arg}")
  # extract all the relevant offsets
  offsets_rootsrc: defaultdict[Any, dict[int, list[int]]] = defaultdict(dict)
  for i in range(vec.dtype.count):
@@ -299,7 +299,7 @@ def reduce_to_acc(ctx:ReduceContext, red:UOp):
    ended_ranges = flatten([x.ended_ranges for x in topo if x.op is Ops.END])
    input_ranges = tuple([x for x in topo if x.op is Ops.RANGE and x not in reduce_range and x not in ended_ranges])
    identity = red.const(red.dtype, identity_element(red.arg, red.dtype.scalar()))
-    acc = UOp(Ops.DEFINE_REG, red.dtype.ptr(size=1, addrspace=AddrSpace.REG), arg=ctx.acc_num)
+    acc = UOp(Ops.DEFINE_REG, red.dtype.ptr(size=1, addrspace=AddrSpace.REG), arg=(ctx.acc_num,))
    acc_init = acc.after(*input_ranges).index(UOp.const(dtypes.int, 0)).store(identity) if len(input_ranges) else \
               acc.index(UOp.const(dtypes.int, 0)).store(identity)
    lst = [acc.after(acc_init, *reduce_range).index(UOp.const(dtypes.int, 0))] + lst  # put acc as the first element
@@ -308,19 +308,9 @@ def reduce_to_acc(ctx:ReduceContext, red:UOp):
  if len(reduce_range) == 0: return ret
  return acc.after(acc.index(UOp.const(dtypes.int, 0)).store(ret).end(*reduce_range)).index(UOp.const(dtypes.int, 0))

-def fold_to_store(x:UOp):
-  _, acc, ranges = x.src[0], x.src[1], x.src[2:]
-  assert acc.op is Ops.INDEX
-  buf = acc.src[0]
-  ret = x.substitute({buf: buf.rtag().after(*ranges)}).substitute({buf.rtag(): buf})
-  base, acc, ranges = ret.src[0], ret.src[1], ret.src[2:]
-  return buf.after(acc.store(base).end(*ranges)).index(acc.src[1])
-
 pm_reduce = PatternMatcher([
-  # REDUCE -> DEFINE_ACC+STORE
+  # REDUCE -> DEFINE_ACC+ASSIGN
  (UPat(Ops.REDUCE, name="red"), reduce_to_acc),
-  # FOLD -> STORE
-  (UPat(Ops.FOLD, name="x"), fold_to_store),
  # tensor core built in accumulate
  (UPat(Ops.WMMA, name="wmma") + UPat.var("add"),
    lambda add, wmma: UOp(wmma.op, wmma.dtype, (wmma.src[0], wmma.src[1], wmma.src[2]+add), wmma.arg)),
@@ -1,59 +1,42 @@
 import heapq
-from typing import Any
 from collections import defaultdict
-from tinygrad.uop.ops import PatternMatcher, UOp, Ops, UPat, multirange_str
-from tinygrad.helpers import prod, getenv, TUPLE_ORDER
+from tinygrad.uop.ops import PatternMatcher, UOp, Ops, UPat

-def linearize(sink:UOp) -> list[UOp]:
+def linearize(u:UOp) -> list[UOp]:
  # this is a toposort with priority
-  lst = list(sink.toposort())
+  lst = list(u.toposort())
  consumers: defaultdict[UOp, list[UOp]] = defaultdict(list)
  in_degree:dict[UOp, int] = {}
-  out_degree:dict[UOp, int] = {}
-  priorities:dict[UOp, tuple[int, int, Any]] = {}
+  priorities:dict[UOp, int] = {}

  # get consumers and assign priorities
  # NOTE: this requires the lst be locally toposorted
  for u in reversed(lst):
    for s in u.src: consumers[s].append(u)
    in_degree[u] = len(u.src)
-    out_degree[u] = len(consumers[u])
-
-    # we place UOps with higher run_counts later
-    run_count = prod([int(r.vmax)+1 for r in u.ranges])
-
-    # simple priority override. this is all bottom up now, smaller numbers will be closer to the top
-    extra = None
-    match u.op:
-      # the order and placement of these defines is important
-      case Ops.DEFINE_GLOBAL: priority, extra = -20, u.arg
-      case Ops.DEFINE_VAR: priority, extra = -19, u.arg
-      case Ops.DEFINE_LOCAL: priority = -18
-      case Ops.DEFINE_REG: priority = -17
-      case Ops.CONST: priority = -10  # early consts
-      case Ops.LOAD: priority = -1    # place loads early
-      case Ops.STORE: priority = 1    # place stores late
-      case Ops.RANGE: priority = 5    # placing RANGE is good
-      case Ops.END: priority = -5     # placing END is bad
-      case _: priority = 0            # everything else has priority 0
-    priorities[u] = (run_count, priority, extra)
+    # put loads in the beginning of the block and prevent priority inversion. hack for BARRIER grouping too
+    priority = [0] + [priorities[x] for x in consumers[u]]
+    if u.op is Ops.LOAD: priority.append(-1000)
+    if u.op is Ops.BARRIER: priority.append(-1500)
+    # ranges are scheduled as late as possible so anything that can be outside is
+    # if u.op is Ops.RANGE: priority = [2000]
+    if u.op is Ops.END: priority = [-1000]
+    # move defines and consts to the top
+    if u.op in {Ops.DEFINE_GLOBAL, Ops.DEFINE_LOCAL, Ops.DEFINE_REG, Ops.DEFINE_VAR, Ops.SPECIAL, Ops.CONST}: priority.append(-2000)
+    priorities[u] = min(priority)

  # number the uops in "ideal" order
-  nkey = {u:i for i,u in enumerate(sorted(lst, key=lambda x: priorities[x]+(x.tuplize if TUPLE_ORDER else ())))}
+  nkey = {u:i for i,u in enumerate(sorted(lst, key=lambda x: (priorities[x],)+x.tuplize))}

-  # then force them to be toposorted in as close to the ideal order as possible
-  heap = [(-nkey[sink], sink)]
+  # then force then to be toposorted in as close to the ideal order as possible
+  heapq.heapify(heap:=[(nkey[u],u) for u in lst if in_degree[u] == 0])
  newlst = []
  while heap:
    newlst.append(u:=heapq.heappop(heap)[1])
-    for v in u.src:
-      out_degree[v] -= 1
-      if out_degree[v] == 0: heapq.heappush(heap, (-nkey[v],v))
-  newlst = newlst[::-1]
-
-  if getenv("DEBUG_LINEARIZE"):
-    for i,u in enumerate(newlst):
-      print(f"{i:4d} {str(u.op):20s} {multirange_str(u.ranges, color=True, pad=10)} {priorities[u]}")
+    for v in consumers[u]:
+      in_degree[v] -= 1
+      if in_degree[v] == 0: heapq.heappush(heap, (nkey[v],v))
+  assert len(newlst) == len(lst), f"len mismatch {len(newlst)} != {len(lst)}"
  return newlst

 class CFGContext:
@@ -81,10 +81,10 @@ def hand_coded_optimizations(k:Scheduler) -> Scheduler:
            return k

  # are we grouping? (requires local shape support)
-  if resolve(prod(k.output_shape[i] for i in k.upcastable_dims) <= (240 if NOLOCALS else 2048), False):
-    for axis, sz in itertools.product((0, 1, 2), (16,)):
+  if resolve(prod(k.output_shape[i] for i in k.upcastable_dims) <= 2048, False):
+    for sz in [16]:
      try:
-        k.apply_opt(Opt(OptOps.GROUPTOP, axis, sz))
+        k.apply_opt(Opt(OptOps.GROUPTOP, 0, sz))
        break
      except KernelOptError: pass

@@ -107,7 +107,7 @@ def hand_coded_optimizations(k:Scheduler) -> Scheduler:
  # potentially do more upcasts of non reduce axes based on a heuristic
  is_dsp = k.ren is not None and k.ren.device == "DSP"
  upcasted_axis: set[int] = set()
-  while resolve(prod(k.output_shape[i] for i in k.upcastable_dims) >= 1024) and (k.upcast_size() < 32):
+  while resolve(prod(k.output_shape[i] for i in k.upcastable_dims) >= 1024):
    xb_choices = []
    # consider all upcastable axes with 3 or 4 upcast (128 on the DSP)
    for axis, upcast_amount in itertools.product(k.upcastable_dims, ([128] if not len(upcasted_axis) else []) if is_dsp else [3,4]):
@@ -135,7 +135,8 @@ def hand_coded_optimizations(k:Scheduler) -> Scheduler:
  # if last reduce dim is small(ish), loop unroll the reduce
  # NOTE: this can fail on multireduce with mismatching dimensions, this is okay
  try:
-    if k.unrollable_dims and (k.upcast_size() <= 4 or not k.axes_of(AxisType.UNROLL)) and (k.upcast_size() < 64):
+    upcast_size = prod(k.full_shape[a] for a in k.axes_of(AxisType.UPCAST, AxisType.UNROLL))
+    if k.unrollable_dims and (upcast_size <= 4 or not k.axes_of(AxisType.UNROLL)) and (upcast_size < 64):
      if (s:=k.full_shape[k.unrollable_dims[-1]]) <= 32:
        k.apply_opt(Opt(OptOps.UNROLL, len(k.unrollable_dims)-1, 0))
        # if it's small, upcast a second reduce dimension too
@@ -2,8 +2,7 @@ from __future__ import annotations
 import math, itertools
 from collections import defaultdict
 from typing import cast, Final
-from tinygrad.uop.ops import PatternMatcher, UPat, Ops, UOp, KernelInfo, graph_rewrite, AxisType, ssimplify, GroupOp
-from tinygrad.uop.ops import axis_letters, axis_colors, axis_to_pos
+from tinygrad.uop.ops import PatternMatcher, UPat, Ops, UOp, KernelInfo, graph_rewrite, AxisType, ssimplify, GroupOp, axis_letters, axis_colors
 from tinygrad.device import Buffer
 from tinygrad.dtype import dtypes, ImageDType
 from tinygrad.helpers import colored, BEAM, getenv, DEBUG, to_function_name, NOOPT, argsort, round_up, prod, merge_dicts, get_single_element, flatten
@@ -13,6 +12,10 @@ from tinygrad.renderer import Renderer

 remove_tags = PatternMatcher([(UPat(GroupOp.All, name="x"), lambda x: x.replace(tag=None) if x.tag is not None else None)])

+# NOTE: LOCAL and GROUP_REDUCE have the same priority. the order here matters
+axis_to_pos = {AxisType.LOOP: -1, AxisType.THREAD: 0, AxisType.GLOBAL: 0, AxisType.WARP: 1, AxisType.LOCAL: 2, AxisType.UPCAST: 3,
+               AxisType.GROUP_REDUCE: 2, AxisType.REDUCE: 4, AxisType.UNROLL: 5}
+
 class Scheduler:
  def __init__(self, ast:UOp, ren:Renderer):
    self.ast, self.ren = ast, ren
@@ -102,8 +105,6 @@ class Scheduler:
  def ranges_of(self, *axis_type:AxisType) -> list[UOp]: return [r for r in self.rngs if r.arg[-1] in axis_type]
  def axes_of(self, *axis_type:AxisType) -> list[int]: return [i for i,t in enumerate(self.axis_types) if t in axis_type]

-  def upcast_size(self) -> int: return prod(self.full_shape[a] for a in self.axes_of(AxisType.UPCAST, AxisType.UNROLL))
-
  # copied from kernel.py
  @property
  def upcastable_dims(self) -> list[int]: return [i for i in self.axes_of(AxisType.GLOBAL, AxisType.LOCAL, AxisType.LOOP) \
@@ -216,7 +217,8 @@ class Scheduler:
    return ret

  def _apply_tc_opt(self, use_tensor_cores:int, axis:int, tc_select:int, opt_level:int) -> None|list[UOp]:
-    if not (reduceops := self.reduceops): raise KernelOptError("no reduce ops for TensorCore")
+    reduceops = [x for x in self.ast.toposort() if x.op is Ops.REDUCE]
+    if not len(reduceops): raise KernelOptError("no reduce ops for TensorCore")
    reduceop = reduceops[0]
    if use_tensor_cores and reduceop is not None and reduceop.arg is Ops.ADD:
      mul = reduceop.src[0] if reduceop.src[0].op is not Ops.CAST else reduceop.src[0].src[0]
@@ -310,10 +312,9 @@ class Scheduler:

  # helpers for hand_coded_optimizations
  @property
-  def reduceops(self) -> list[UOp]: return [x for x in self.ast.backward_slice if x.op is Ops.REDUCE]
-  @property
  def reduceop(self) -> UOp|None:
-    if not (red := self.reduceops): return None
+    red = [x for x in self.ast.backward_slice if x.op is Ops.REDUCE]
+    if not len(red): return None
    return UOp(Ops.REDUCE_AXIS, red[0].dtype, red[0].src, (red[0].arg, ()))
  @property
  def bufs(self) -> list[UOp]: return [x for x in self.ast.toposort() if x.op is Ops.INDEX][::-1]
@@ -1,6 +1,6 @@
 import itertools
 from tinygrad.uop.ops import UOp, PatternMatcher, UPat, Ops, graph_rewrite, _substitute, range_start, ImageDType
-from tinygrad.uop.symbolic import symbolic
+from tinygrad.uop.symbolic import symbolic_flat
 from tinygrad.helpers import partition, dedup
 from tinygrad.dtype import dtypes

@@ -28,12 +28,13 @@ def simplify_merge_adjacent(u:UOp) -> UOp|None:
        s0, s1 = r0.src[0], r1.src[0]
        # do the merge
        new_range = r0.replace(src=(s0*s1,))
-        nidx = graph_rewrite(u, _substitute+symbolic+pm_flatten_range, ctx={r0:new_range//s1, r1:new_range%s1},
+        nidx = graph_rewrite(u, _substitute+symbolic_flat+pm_flatten_range, ctx={r0:new_range//s1, r1:new_range%s1},
                             name=f"check_merge_{r0.arg[0]}_{r1.arg[0]}")

        # check if it simplifies
        if count_divmod(nidx) <= count_divmod(u):
          u = nidx
+          continue
  return u

 pm_simplify_ranges = PatternMatcher([
@@ -108,7 +109,7 @@ pm_reduce_collapse = pm_reduce_unparented + PatternMatcher([
    lambda x,y,c,r: y.where(c, 0).reduce(*r.src[1:], arg=Ops.ADD)*x.cast(c.dtype)),
  # MUL casted bool
  ((UPat.var("x") * UPat.var("gate", dtype=dtypes.bool).cast()), lambda x,gate: gate.where(x, 0)),
-])+symbolic
+])+symbolic_flat

 pm_reduce_load_collapse = pm_reduce_collapse + PatternMatcher([
  # lift x+y out of reduce on ne
@@ -4,8 +4,8 @@ from collections import defaultdict
 from typing import Any, Generic, TypeVar, Iterator, Sequence, cast, Generator
 import importlib, inspect, functools, pathlib, os, platform, contextlib, sys, re, atexit, pickle, decimal
 from tinygrad.helpers import CI, OSX, LRU, getenv, diskcache_get, diskcache_put, DEBUG, GlobalCounters, flat_mv, PROFILE, temp, colored, CPU_LLVM
-from tinygrad.helpers import Context, CCACHE, ALLOW_DEVICE_USAGE, MAX_BUFFER_SIZE, cpu_events, ProfileEvent, ProfilePointEvent, dedup
-from tinygrad.helpers import unwrap_class_type, suppress_finalizing, AMD_LLVM, select_first_inited
+from tinygrad.helpers import Context, DISABLE_COMPILER_CACHE, ALLOW_DEVICE_USAGE, MAX_BUFFER_SIZE, cpu_events, ProfileEvent, ProfilePointEvent, dedup
+from tinygrad.helpers import unwrap_class_type, suppress_finalizing, AMD_LLVM
 from tinygrad.dtype import DType, ImageDType, PtrDType, dtypes, _to_np_dtype
 from tinygrad.renderer import Renderer

@@ -54,7 +54,7 @@ atexit.register(lambda: [Device[dn].finalize() for dn in Device._opened_devices]

@dataclass(frozen=True)
 class ProfileDeviceEvent(ProfileEvent):
-  device:str; comp_tdiff:decimal.Decimal=decimal.Decimal(0); copy_tdiff:decimal.Decimal=decimal.Decimal(0); props:dict[str,Any]|None=None # noqa: E702
+  device:str; comp_tdiff:decimal.Decimal=decimal.Decimal(0); copy_tdiff:decimal.Decimal=decimal.Decimal(0) # noqa: E702

@dataclass(frozen=True)
 class ProfileProgramEvent(ProfileEvent): device:str; name:str; lib:bytes|None; base:int|None # noqa: E702
@@ -266,7 +266,7 @@ class LRUAllocator(Allocator, Generic[DeviceType]):
 class CompileError(Exception): pass

 class Compiler:
-  def __init__(self, cachekey:str|None=None): self.cachekey = cachekey if CCACHE else None
+  def __init__(self, cachekey:str|None=None): self.cachekey = None if DISABLE_COMPILER_CACHE else cachekey
  def compile(self, src:str) -> bytes: return src.encode()   # NOTE: empty compiler is the default
  def compile_cached(self, src:str) -> bytes:
    if self.cachekey is None or (lib := diskcache_get(self.cachekey, src)) is None:
@@ -291,8 +291,8 @@ class Compiled:
    if len(enable_comps) > 1: raise RuntimeError(f"{self.device}: multiple compilers set in env {enable_comps}")
    for _, comp_pair in disable_comps: self.compilers.remove(comp_pair)

-    self.renderer, self.compiler = select_first_inited([list(enable_comps)[0][1]] if len(enable_comps) == 1 else self.compilers,
-                                                       f"No compiler for {self.device} is available")
+    try: self.renderer, self.compiler = next(self._get_available_compilers([list(enable_comps)[0][1]] if len(enable_comps) == 1 else self.compilers))
+    except StopIteration as exc: raise RuntimeError(f"no usable compilers for {self.device}") from exc

    if DEBUG >= 1: print(f"{self.device}: using {self.compiler.__class__.__name__}")

@@ -300,6 +300,10 @@ class Compiled:
    compiler_name = f"{unwrap_class_type(c).__name__.upper().removesuffix('COMPILER').removeprefix(devname:=self.device.split(':')[0].upper())}"
    return f"{devname}_{compiler_name if len(compiler_name) > 0 else unwrap_class_type(c).__name__.upper()}"

+  def _get_available_compilers(self, compilers) -> Iterator[tuple[Renderer, Compiler]]:
+    for renderer, compiler in compilers:
+      with contextlib.suppress(Exception): yield renderer(), compiler()
+
  def synchronize(self):
    """
    Synchronize all pending operations on the device.
@@ -339,7 +343,6 @@ def is_dtype_supported(dtype:DType, device:str|None=None) -> bool:
  # PYTHON supports half memoryview in 3.12+ https://github.com/python/cpython/issues/90751
  if dtype == dtypes.half:
    if device == "CL": return not CI and not OSX
-    if device == "QCOM": return False # QCOM compiler is flaky with half
    if device in ["CUDA", "NV"]: return not CI
    if device == "CPU" and CPU_LLVM: return OSX
    if device == "PYTHON": return sys.version_info >= (3, 12)
@@ -3,7 +3,7 @@ import time, pprint, random, itertools, math
 from dataclasses import dataclass, replace, field
 from tinygrad.helpers import all_same, colored, DEBUG, GlobalCounters, ansilen, BEAM, NOOPT, all_int, CAPTURING, Metadata, TRACEMETA, TracingKey
 from tinygrad.helpers import DEVECTORIZE, time_to_str, VALIDATE_WITH_CPU, getenv, cpu_profile, PROFILE, ProfilePointEvent, cpu_events, prod, Context
-from tinygrad.helpers import unwrap, disable_gc
+from tinygrad.helpers import unwrap
 from tinygrad.uop.ops import Ops, PatternMatcher, UOp, UPat, sym_infer, graph_rewrite, print_uops, track_rewrites, KernelInfo, pyrender
 from tinygrad.device import Device, Buffer
 from tinygrad.renderer import Renderer, ProgramSpec, Estimates
@@ -13,7 +13,6 @@ from tinygrad.codegen.opt import Opt

 # **************** Program Creation ****************

-@disable_gc()
@track_rewrites(name=lambda *args,ret,**kwargs: TracingKey(ret.name, (ret.function_name, ret.ast), ret=ret), replay=True)
 def get_program(ast:UOp, renderer:Renderer|None=None, opts:list[Opt]|None=None) -> ProgramSpec:
  """
@@ -29,14 +29,14 @@ pm_gradient = PatternMatcher([
  (UPat(Ops.MUL, name="ret"), lambda ctx, ret: (ret.src[1]*ctx, ret.src[0]*ctx)),
  (UPat(Ops.WHERE, name="ret"), lambda ctx, ret: (None, ret.src[0].where(ctx, ctx.const_like(0)), ret.src[0].where(ctx.const_like(0), ctx))),
  (UPat(Ops.REDUCE_AXIS, name="ret"), reduce_gradient),
-  (UPat(Ops.CONTIGUOUS), lambda ctx: (ctx,)),
+  (UPat((Ops.CONTIGUOUS, Ops.FUSE)), lambda ctx: (ctx,)),
  (UPat(Ops.CONTIGUOUS_BACKWARD), lambda ctx: (ctx.contiguous(),)),
  (UPat(Ops.RESHAPE, name="ret"), lambda ctx, ret: (ctx.reshape(ret.src[0].shape), None)),
  (UPat(Ops.EXPAND, name="ret"), lambda ctx, ret: (ctx.r(Ops.ADD,tuple(i for i,(s,n) in enumerate(zip(ret.src[0].shape, ret.shape)) if s!=n)), None)),
  (UPat(Ops.PAD, name="ret"), lambda ctx, ret: (ctx.shrink(tuple([(p[0], s+p[0]) for s,p in zip(ret.src[0].shape, ret.marg)])), None, None)),
  (UPat(Ops.SHRINK, name="ret"), lambda ctx, ret: (ctx.pad(tuple([(p[0], s-p[1]) for s,p in zip(ret.src[0].shape, ret.marg)])), None, None)),
  (UPat(Ops.PERMUTE, name="ret"), lambda ctx, ret: (ctx.permute(argsort(ret.marg)),)),
-  (UPat(Ops.FLIP, name="ret"), lambda ctx, ret: (ctx.flip([i for i,x in enumerate(ret.marg) if x]),)),
+  (UPat(Ops.FLIP, name="ret"), lambda ctx, ret: (ctx.flip(ret.marg),)),
  (UPat(Ops.MULTI, name="ret"), lambda ctx, ret: ctx.shard(ret.device, ret.axis).src),
  # NOTE: this is only correct when the KERNEL has a single output
  (UPat(Ops.AFTER), lambda ctx: (ctx, ctx)),
@@ -1,5 +1,5 @@
 from __future__ import annotations
-import os, functools, platform, time, re, contextlib, operator, hashlib, pickle, sqlite3, tempfile, pathlib, string, ctypes, sys, gzip, getpass, gc
+import os, functools, platform, time, re, contextlib, operator, hashlib, pickle, sqlite3, tempfile, pathlib, string, ctypes, sys, gzip, getpass
 import urllib.request, subprocess, shutil, math, types, copyreg, inspect, importlib, decimal, itertools
 from dataclasses import dataclass, field
 from typing import ClassVar, Iterable, Any, TypeVar, Callable, Sequence, TypeGuard, Iterator, Generic, Generator, cast, overload
@@ -44,7 +44,8 @@ def fully_flatten(l):
    return flattened
  return [l]
 def fromimport(mod, frm): return getattr(__import__(mod, fromlist=[frm]), frm)
-def _is_balanced(s:str) -> bool: return (d := 0, all((d := d + (c == '(') - (c == ')')) >= 0 for c in s))[1] and d == 0
+def _is_balanced(s:str) -> bool:
+  return (acc:=list(itertools.accumulate([(1 if ch=='(' else -1 if ch==')' else 0) for ch in s])))[-1]==0 and all(x>=0 for x in acc)
 def strip_parens(fst:str) -> str: return fst[1:-1] if fst and fst[0]=='(' and fst[-1] == ')' and _is_balanced(fst[1:-1]) else fst
 def ceildiv(num, amt): return int(ret) if isinstance((ret:=-(num//-amt)), float) else ret
 def round_up(num:int, amt:int) -> int: return (num+amt-1)//amt * amt
@@ -114,13 +115,6 @@ def suppress_finalizing(func):
      if not getattr(sys, 'is_finalizing', lambda: True)(): raise # re-raise if not finalizing
  return wrapper

-def select_first_inited(candidates:Sequence[Callable[...,T]|Sequence[Callable[...,T]]], err_msg: str) -> tuple[T,...]|T:
-  excs = []
-  for typ in candidates:
-    try: return tuple([cast(Callable, t)() for t in typ]) if isinstance(typ, Sequence) else cast(Callable, typ)()
-    except Exception as e: excs.append(e)
-  raise ExceptionGroup(err_msg, excs)
-
 def unwrap_class_type(cls_t): return cls_t.func if isinstance(cls_t, functools.partial) else cls_t

 def pluralize(st:str, cnt:int): return f"{cnt} {st}"+('' if cnt == 1 else 's')
@@ -173,22 +167,20 @@ TRANSCENDENTAL, NOLOCALS = ContextVar("TRANSCENDENTAL", 1), ContextVar("NOLOCALS
 SPLIT_REDUCEOP, NO_MEMORY_PLANNER, RING = ContextVar("SPLIT_REDUCEOP", 1), ContextVar("NO_MEMORY_PLANNER", 0), ContextVar("RING", 1)
 PICKLE_BUFFERS, LRU = ContextVar("PICKLE_BUFFERS", 1), ContextVar("LRU", 1)
 CACHELEVEL, IGNORE_BEAM_CACHE, DEVECTORIZE = ContextVar("CACHELEVEL", 2), ContextVar("IGNORE_BEAM_CACHE", 0), ContextVar("DEVECTORIZE", 1)
+DISABLE_COMPILER_CACHE = ContextVar("DISABLE_COMPILER_CACHE", 0)
 VALIDATE_WITH_CPU, DISABLE_FAST_IDIV = ContextVar("VALIDATE_WITH_CPU", 0), ContextVar("DISABLE_FAST_IDIV", 0)
 CORRECT_DIVMOD_FOLDING, FUSE_OPTIM = ContextVar("CORRECT_DIVMOD_FOLDING", 0), ContextVar("FUSE_OPTIM", 0)
 ALLOW_DEVICE_USAGE, MAX_BUFFER_SIZE = ContextVar("ALLOW_DEVICE_USAGE", 1), ContextVar("MAX_BUFFER_SIZE", 0)
+FUSE_ATTENTION = ContextVar("FUSE_ATTENTION", 0)
 EMULATE = ContextVar("EMULATE", "")
 CPU_COUNT = ContextVar("CPU_COUNT", max(1, len(os.sched_getaffinity(0)) if hasattr(os, "sched_getaffinity") else (os.cpu_count() or 1)))
-CPU_LLVM, CPU_LVP, AMD_LLVM = ContextVar("CPU_LLVM", 0), ContextVar("CPU_LVP", 0), ContextVar("AMD_LLVM", 0)
+CPU_LLVM, CPU_LVP, AMD_LLVM = ContextVar("CPU_LLVM", 0), ContextVar("CPU_LVP", 0), ContextVar("AMD_LLVM", 1)
 VIZ = PROFILE = ContextVar("VIZ", 0)
 SPEC = ContextVar("SPEC", 1)
 # TODO: disable by default due to speed
 IGNORE_OOB = ContextVar("IGNORE_OOB", 1)
 PCONTIG = ContextVar("PCONTIG", 0)  # partial contiguous in rangeify
 DEBUG_RANGEIFY = ContextVar("DEBUG_RANGEIFY", 0)
-# set to 1, this uses tuplize in the linearizer sort order
-TUPLE_ORDER = ContextVar("TUPLE_ORDER", 1)
-# set to 0 to disable the compiler cache
-CCACHE = ContextVar("CCACHE", 1)

@dataclass(frozen=True)
 class Metadata:
@@ -241,29 +233,11 @@ class Profiling(contextlib.ContextDecorator):

 def perf_counter_us() -> decimal.Decimal: return decimal.Decimal(time.perf_counter_ns())/1000

-@functools.cache
-def lines(fn) -> list[str]:
-  try:
-    with open(fn, encoding="utf-8") as f: return f.readlines()
-  except (FileNotFoundError, OSError): return []
-
-def printable(loc:tuple[str, int]) -> str:
-  try: return lines(loc[0])[loc[1]-1].strip()
-  except IndexError: return "<missing>"
-
-def get_stacktrace(frm, max_frames=30) -> tuple[tuple, ...]:
-  ret:list[tuple] = []
-  for i in range(max_frames):
-    if (frm:=frm.f_back) is None: break
-    ret.append(((fc:=frm.f_code).co_filename, frm.f_lineno, fc.co_name, printable((fc.co_filename, frm.f_lineno))))
-  return tuple(ret)
-
@dataclass(frozen=True)
 class TracingKey:
  display_name:str                       # display name of this trace event
  keys:tuple[Any, ...]=()                # optional keys to search for related traces
  ret:Any=None
-  tb:tuple[tuple, ...]|None=field(default_factory=lambda: get_stacktrace(sys._getframe(1)) if VIZ else None)

 class ProfileEvent: pass

@@ -380,12 +354,10 @@ def fetch(url:str, name:pathlib.Path|str|None=None, subdir:str|None=None, gunzip

 # *** Exec helpers

-def system(cmd, **kwargs): return subprocess.check_output(cmd.split(), **kwargs).decode().strip()
-
 def cpu_objdump(lib, objdump_tool='objdump'):
  with tempfile.NamedTemporaryFile(delete=True) as f:
    pathlib.Path(f.name).write_bytes(lib)
-    print(system(f"{objdump_tool} -d {f.name}"))
+    print(subprocess.check_output([objdump_tool, '-d', f.name]).decode('utf-8'))

 def capstone_flatdump(lib: bytes):
  try: import capstone
@@ -416,7 +388,6 @@ def to_mv(ptr:int, sz:int) -> memoryview: return memoryview((ctypes.c_uint8 * sz
 def mv_address(mv): return ctypes.addressof(ctypes.c_char.from_buffer(mv))
 def to_char_p_p(options: list[bytes], to_type=ctypes.c_char):
  return (ctypes.POINTER(to_type) * len(options))(*[ctypes.cast(ctypes.create_string_buffer(o), ctypes.POINTER(to_type)) for o in options])
-def charptr(s:str|bytes): return ctypes.cast(ctypes.c_char_p(s if isinstance(s, bytes) else s.encode()), ctypes.POINTER(ctypes.c_char))
@functools.cache
 def init_c_struct_t(fields: tuple[tuple[str, type[ctypes._SimpleCData]], ...]):
  class CStruct(ctypes.Structure):
@@ -464,13 +435,6 @@ class tqdm(Generic[T]):
 class trange(tqdm):
  def __init__(self, n:int, **kwargs): super().__init__(iterable=range(n), total=n, **kwargs)

-class disable_gc(contextlib.ContextDecorator):
-  def __enter__(self):
-    self._was_enabled = gc.isenabled()
-    if self._was_enabled: gc.disable()
-  def __exit__(self, *exc):
-    if self._was_enabled: gc.enable()
-
 # *** universal support for code object pickling

 def _reconstruct_code(*args): return types.CodeType(*args)
@@ -1,6 +0,0 @@
-from tinygrad.mixin.math import MathMixin
-from tinygrad.mixin.movement import MovementMixin
-
-
-class OpMixin(MathMixin, MovementMixin):
-  pass
@@ -1,279 +0,0 @@
-from typing import Self
-from tinygrad.uop import Ops
-from tinygrad.dtype import dtypes, ConstType
-
-
-class MathMixin:
-  # required to implement
-  def alu(self, op: Ops, *src: Self) -> Self:
-    raise NotImplementedError
-
-  def const_like(self, b: ConstType) -> Self:
-    raise NotImplementedError
-
-  # great functions you get!
-  def ufix(self, x: Self | ConstType) -> Self:
-    return self.const_like(x) if not isinstance(x, MathMixin) else x
-
-  def _binop(self, op: Ops, x: Self | ConstType, reverse: bool) -> Self:
-    return self.ufix(x).alu(op, self) if reverse else self.alu(op, self.ufix(x))
-
-  def logical_not(self):
-    return self.ne(True)
-
-  def neg(self):
-    if (dtype := getattr(self, "dtype")) is None:
-      raise TypeError(f"MathTraits __neg__ requires a dtype, {self=}")
-    return self.logical_not() if dtype.scalar() == dtypes.bool else self * (-1)
-
-  def _check_dtype(self):
-    if (dtype := getattr(self, "dtype")) is not None:
-      if isinstance(dtype, tuple):
-        dtype = dtype[0]
-      if not (dtypes.is_bool(dtype) or dtypes.is_int(dtype)):
-        raise RuntimeError(f"{dtype} is not supported")
-
-  def add(self, x: Self | ConstType, reverse: bool = False):
-    """
-    Adds `self` and `x`.
-    Equivalent to `self + x`.
-    Supports broadcasting to a common shape, type promotion, and integer, float, boolean inputs.
-    ```python exec="true" source="above" session="tensor" result="python"
-    Tensor.manual_seed(42)
-    t = Tensor.randn(4)
-    print(t.numpy())
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(t.add(20).numpy())
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(t.add(Tensor([[2.0], [3.5]])).numpy())
-    ```
-    """
-    return self._binop(Ops.ADD, x, reverse)
-
-  def mul(self, x: Self | ConstType, reverse: bool = False):
-    """
-    Multiplies `self` and `x`.
-    Equivalent to `self * x`.
-    Supports broadcasting to a common shape, type promotion, and integer, float, boolean inputs.
-
-    ```python exec="true" source="above" session="tensor" result="python"
-    Tensor.manual_seed(42)
-    t = Tensor.randn(4)
-    print(t.numpy())
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(t.mul(3).numpy())
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(t.mul(Tensor([[-1.0], [2.0]])).numpy())
-    ```
-    """
-    return self._binop(Ops.MUL, x, reverse)
-
-  def bitwise_and(self, x: Self | ConstType, reverse: bool = False):
-    """
-    Computes the bitwise AND of `self` and `x`.
-    Equivalent to `self & x`.
-    Supports broadcasting to a common shape, type promotion, and integer, boolean inputs.
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(Tensor([2, 5, 255]).bitwise_and(Tensor([3, 14, 16])).numpy())
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(Tensor([True, True, False, False]).bitwise_and(Tensor([True, False, True, False])).numpy())
-    ```
-    """
-    self._check_dtype()
-    return self._binop(Ops.AND, x, reverse)
-
-  def bitwise_or(self, x: Self | ConstType, reverse: bool = False):
-    """
-    Computes the bitwise OR of `self` and `x`.
-    Equivalent to `self | x`.
-    Supports broadcasting to a common shape, type promotion, and integer, boolean inputs.
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(Tensor([2, 5, 255]).bitwise_or(Tensor([4, 4, 4])).numpy())
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(Tensor([True, True, False, False]).bitwise_or(Tensor([True, False, True, False])).numpy())
-    ```
-    """
-    self._check_dtype()
-    return self._binop(Ops.OR, x, reverse)
-
-  def bitwise_xor(self, x: Self | ConstType, reverse: bool = False):
-    """
-    Computes bitwise xor of `self` and `x`.
-    Equivalent to `self ^ x`.
-    Supports broadcasting to a common shape, type promotion, and integer, boolean inputs.
-
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(Tensor([-1, -2, 3]).bitwise_xor(Tensor([1, 0, 3])).numpy())
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(Tensor([True, True, False, False]).bitwise_xor(Tensor([True, False, True, False])).numpy())
-    ```
-    """
-    self._check_dtype()
-    return self._binop(Ops.XOR, x, reverse)
-
-  def idiv(self, x: Self | ConstType, reverse: bool = False):
-    """
-    Divides `self` by `x`.
-    Equivalent to `self // x`.
-    Supports broadcasting to a common shape, type promotion, and integer inputs.
-    `idiv` performs integer division (truncate towards zero).
-
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(Tensor([-4, 7, 5, 4, -7, 8]).idiv(Tensor([2, -3, 8, -2, 3, 5])).numpy())
-    ```
-    """
-    return self._binop(Ops.IDIV, x, reverse)
-
-  def mod(self, x: Self | ConstType, reverse: bool = False):
-    return self._binop(Ops.MOD, x, reverse)
-
-  def sub(self, x: Self | ConstType, reverse: bool = False):
-    return self.ufix(x).alu(Ops.ADD, -self) if reverse else self.alu(Ops.ADD, self.ufix(-x))
-
-  def div(self, x: Self | ConstType, reverse: bool = False):
-    return (self.ufix(x) * self.alu(Ops.RECIPROCAL)) if reverse else (self * self.ufix(x).alu(Ops.RECIPROCAL))
-
-  def __neg__(self):
-    return self.neg()
-
-  def __add__(self, x: Self | ConstType):
-    return self.add(x)
-
-  def __sub__(self, x: Self | ConstType):
-    return self.sub(x)
-
-  def __mul__(self, x: Self | ConstType):
-    return self.mul(x)
-
-  def __truediv__(self, x: Self | ConstType):
-    return self.div(x)
-
-  def __floordiv__(self, x: Self | ConstType):
-    return self.idiv(x)  # TODO: idiv is trunc div, not floordiv
-
-  def __mod__(self, x: Self | ConstType):
-    return self.mod(x)
-
-  def __and__(self, x: Self | ConstType):
-    return self.bitwise_and(x)
-
-  def __or__(self, x: Self | ConstType):
-    return self.bitwise_or(x)
-
-  def __xor__(self, x: Self | ConstType):
-    return self.bitwise_xor(x)
-
-  def __radd__(self, x: Self | ConstType):
-    return self.add(x, True)
-
-  def __rsub__(self, x: Self | ConstType):
-    return self.sub(x, True)
-
-  def __rmul__(self, x: Self | ConstType):
-    return self.mul(x, True)
-
-  def __rtruediv__(self, x: Self | ConstType):
-    return self.div(x, True)
-
-  def __rfloordiv__(self, x: Self | ConstType):
-    return self.idiv(x, True)
-
-  def __rand__(self, x: Self | ConstType):
-    return self.bitwise_and(x, True)
-
-  def __ror__(self, x: Self | ConstType):
-    return self.bitwise_or(x, True)
-
-  def __rxor__(self, x: Self | ConstType):
-    return self.bitwise_xor(x, True)
-
-  def __rmod__(self, x: Self | ConstType):
-    return self.mod(x, True)
-
-  def __lt__(self, x: Self | ConstType):
-    return self.alu(Ops.CMPLT, self.ufix(x))
-
-  def __gt__(self, x: Self | ConstType):
-    return self.ufix(x).alu(Ops.CMPLT, self)
-
-  def __ge__(self, x: Self | ConstType):
-    return (self < x).logical_not()
-
-  def __le__(self, x: Self | ConstType):
-    return (self > x).logical_not()
-
-  def ne(self, x: Self | ConstType):
-    return self.alu(Ops.CMPNE, self.ufix(x))
-
-  def eq(self, x: Self | ConstType):
-    return self.ne(x).logical_not()
-
-  def __ne__(self, x: Self | ConstType):  # type: ignore[override]
-    return self.ne(x)
-
-  # NOTE: __eq__ isn't overridden, and means the same thing as is by default
-
-  def lshift(self, x: Self | int, reverse: bool = False):
-    return self._binop(Ops.SHL, x, reverse)
-
-  def rshift(self, x: Self | int, reverse: bool = False):
-    return self._binop(Ops.SHR, x, reverse)
-
-  def __lshift__(self, x: Self | int):
-    return self.lshift(x)
-
-  def __rshift__(self, x: Self | int):
-    return self.rshift(x)
-
-  def __rlshift__(self, x: Self | int):
-    return self.lshift(x, True)
-
-  def __rrshift__(self, x: Self | int):
-    return self.rshift(x, True)
-
-  def maximum(self, x: Self | ConstType):
-    return self.alu(Ops.MAX, self.ufix(x))
-
-  def minimum(self, x: Self | ConstType):
-    return -(-self).maximum(-x)
-
-  def where(self, x: Self | ConstType, y: Self | ConstType):
-    if isinstance(x, type(self)):
-      return self.alu(Ops.WHERE, x, x.ufix(y))
-    if isinstance(y, type(self)):
-      return self.alu(Ops.WHERE, y.ufix(x), y)
-    raise RuntimeError("where needs at least one UOp arg")
-
-  def threefry(self, seed: Self):
-    return self.alu(Ops.THREEFRY, seed)
-
-  def reciprocal(self):
-    return self.alu(Ops.RECIPROCAL)
-
-  def trunc(self):
-    return self.alu(Ops.TRUNC)
-
-  def sqrt(self):
-    return self.alu(Ops.SQRT)
-
-  def sin(self):
-    return self.alu(Ops.SIN)
-
-  def log2(self):
-    return self.alu(Ops.LOG2)
-
-  def exp2(self):
-    return self.alu(Ops.EXP2)
-
-  def pow(self, x: Self | ConstType):
-    return self.alu(Ops.POW, self.ufix(x))
-
-  def __pow__(self, x: Self | ConstType):
-    return self.pow(x)
@@ -1,376 +0,0 @@
-# mixins add syntactic sugar to Tensor and UOp
-import functools
-from typing import TypeAlias, TYPE_CHECKING, Self
-from tinygrad.uop import Ops
-from tinygrad.helpers import prod, argfix, flatten, dedup, make_tuple, ceildiv
-from tinygrad.uop.ops import resolve, smax
-
-if TYPE_CHECKING:
-  from tinygrad.uop.ops import UOp
-sint: TypeAlias = "UOp | int"
-
-
-def _align_left(*shapes: tuple[sint, ...]) -> tuple[tuple[sint, ...], ...]:
-  # unsqueeze left to make every shape same length
-  max_dim = max(len(shape) for shape in shapes)
-  return tuple((1,) * (max_dim - len(shape)) + shape for shape in shapes)
-
-
-class MovementMixin:
-  # required to implement
-  def _mop(self, op: Ops, arg) -> Self:
-    raise NotImplementedError
-
-  @property
-  def shape(self) -> tuple[sint, ...]:
-    raise NotImplementedError
-
-  # great functions you get!
-  @property
-  def ndim(self) -> int:
-    """
-    Returns the number of dimensions in the tensor.
-
-    ```python exec="true" source="above" session="tensor" result="python"
-    t = Tensor([[1, 2], [3, 4]])
-    print(t.ndim)
-    ```
-    """
-    return len(self.shape)
-
-  def numel(self) -> sint:
-    """
-    Returns the total number of elements in the tensor.
-
-    ```python exec="true" source="above" session="tensor" result="python"
-    t = Tensor([[[1, 2], [3, 4]], [[5, 6], [7, 8]]])
-    print(t.numel())
-    ```
-    """
-    return prod(self.shape)
-
-  def _resolve_dim(self, dim: int, *, extra: bool = False) -> int:
-    total = self.ndim + int(extra)
-    if not -max(1, total) <= dim <= max(1, total) - 1:
-      raise IndexError(f"{dim=} out of range {[-max(1, total), max(1, total) - 1]}")
-    return dim + total if dim < 0 else dim
-
-  def _broadcast_to(self, new_shape: tuple[sint, ...]) -> Self:
-    if self.shape == new_shape:
-      return self
-    if self.ndim > len(new_shape):
-      raise ValueError(f"cannot broadcast tensor to fewer dimensions. shape={self.shape} to {new_shape=}")
-    # first unsqueeze left with 1s https://data-apis.org/array-api/latest/API_specification/broadcasting.html
-    shape, _ = _align_left(self.shape, new_shape)
-    # for each dimension, check either dim is 1, or it does not change
-    if not all(s == ns or s == 1 for s, ns in zip(shape, new_shape)):
-      raise ValueError(f"cannot broadcast {self.shape} to {new_shape=}")
-    reshaped = self.reshape(shape)
-    ret = reshaped._mop(Ops.EXPAND, arg=new_shape)
-    return reshaped if ret.shape == reshaped.shape else ret
-
-  def expand(self, shape, *args) -> Self:
-    """
-    Returns a tensor that is expanded to the shape that is specified.
-    Expand can also increase the number of dimensions that a tensor has.
-
-    Passing a `-1` or `None` to a dimension means that its size will not be changed.
-
-    ```python exec="true" source="above" session="tensor" result="python"
-    t = Tensor([1, 2, 3])
-    print(t.expand(4, -1).numpy())
-    ```
-    """
-    new_shape = tuple(from_ if to == -1 or to is None else to for from_, to in zip(*(_align_left(self.shape, argfix(shape, *args)))))
-    return self._broadcast_to(new_shape)
-
-  def reshape(self, shape, *args) -> Self:
-    """
-    Returns a tensor with the same data as the original tensor but with a different shape.
-    `shape` can be passed as a tuple or as separate arguments.
-
-    ```python exec="true" source="above" session="tensor" result="python"
-    t = Tensor.arange(6)
-    print(t.reshape(2, 3).numpy())
-    ```
-    """
-    # resolve None and args
-    new_shape = tuple([s if s is not None else self.shape[i] for i, s in enumerate(argfix(shape, *args))])
-    # resolve -1
-    if (c := new_shape.count(-1)) > 1:
-      raise RuntimeError(f"only one dimension can be inferred using -1, getting {new_shape}")
-    if c:
-      new_shape = tuple([-prod(self.shape) // prod(new_shape) if s == -1 else s for s in new_shape])
-    if prod(self.shape) != prod(new_shape):
-      raise ValueError(f"size mismatch, can't reshape ({self.shape}) -> ({new_shape})")
-    ret = self._mop(Ops.RESHAPE, arg=new_shape)
-    return self if ret.shape == self.shape else ret
-
-  def shrink(self, arg: tuple[tuple[sint, sint] | None, ...]) -> Self:
-    """
-    Returns a tensor that shrinks the each axis based on input arg.
-    `arg` must have the same length as `self.ndim`.
-    For each axis, it can be `None`, which means no shrink, or a tuple `(start, end)` that works the same as Python slice.
-
-    ```python exec="true" source="above" session="tensor" result="python"
-    t = Tensor.arange(9).reshape(3, 3)
-    print(t.numpy())
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(t.shrink(((None, (1, 3)))).numpy())
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(t.shrink((((0, 2), (0, 2)))).numpy())
-    ```
-    """
-    if self.ndim != len(arg):
-      raise ValueError(f"{self.ndim=} != {len(arg)=}")
-    ret = self._mop(Ops.SHRINK, arg=[x if x is not None else (0, s) for x, s in zip(arg, self.shape)])
-    return self if ret.shape == self.shape else ret
-
-  def permute(self, order, *args) -> Self:
-    """
-    Returns a tensor that is a permutation of the original tensor.
-    The new tensor has the same data as the original tensor but with the dimensions permuted according to the order specified.
-    `order` can be passed as a tuple or as separate arguments.
-
-    ```python exec="true" source="above" session="tensor" result="python"
-    t = Tensor.empty(2, 3, 5)
-    print(t.shape)
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(t.permute(2, 0, 1).shape)
-    ```
-    """
-    order_arg = tuple(self._resolve_dim(x) for x in argfix(order, *args))
-    if sorted(order_arg) != list(range(self.ndim)):
-      raise RuntimeError(f"order is not a valid permutation, getting {order_arg}")
-    return self._mop(Ops.PERMUTE, arg=order_arg) if order_arg != tuple(range(self.ndim)) else self
-
-  def flip(self, axis, *args) -> Self:
-    """
-    Returns a tensor that reverses the order of the original tensor along given `axis`.
-    `axis` can be passed as a tuple or as separate arguments.
-
-    ```python exec="true" source="above" session="tensor" result="python"
-    t = Tensor.arange(6).reshape(2, 3)
-    print(t.numpy())
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(t.flip(0).numpy())
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(t.flip((0, 1)).numpy())
-    ```
-    """
-    axis_arg = tuple(self._resolve_dim(x) for x in argfix(axis, *args))
-    assert all(not isinstance(x, bool) and x >= 0 and x < self.ndim for x in axis_arg), f"flip args must be axis ints {axis_arg}"
-    if len(axis_arg) != len(dedup(axis_arg)):
-      raise RuntimeError(f"dim can appear at most once, getting {axis_arg}")
-    flip_arg = tuple([i in axis_arg for i in range(len(self.shape))])
-    return self._mop(Ops.FLIP, arg=flip_arg) if any(flip_arg) else self
-
-  # **** high level ****
-
-  def shrink_to(self, shape, *args) -> Self:
-    return self.shrink(tuple([None if ns is None else (0, ns) for ns in argfix(shape, *args)]))
-
-  def view(self, shape, *args) -> Self:
-    """`.view` is an alias for `.reshape`."""
-    return self.reshape(shape, *args)
-
-  def squeeze(self, dim: int | None = None) -> Self:
-    """
-    Returns a tensor with specified dimensions of input of size 1 removed.
-    If `dim` is not specified, all dimensions with size 1 are removed.
-
-    ```python exec="true" source="above" session="tensor" result="python"
-    t = Tensor.zeros(2, 1, 2, 1, 2)
-    print(t.squeeze().shape)
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(t.squeeze(0).shape)
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(t.squeeze(1).shape)
-    ```
-    """
-    if dim is None:
-      return self.reshape(tuple(dim for dim in self.shape if dim != 1))
-    dim = self._resolve_dim(dim)
-    return self if not self.ndim or self.shape[dim] != 1 else self.reshape(self.shape[:dim] + self.shape[dim + 1 :])
-
-  def unsqueeze(self, dim: int) -> Self:
-    """
-    Returns a tensor with a new dimension of size 1 inserted at the specified `dim`.
-
-    ```python exec="true" source="above" session="tensor" result="python"
-    t = Tensor([1, 2, 3, 4])
-    print(t.unsqueeze(0).numpy())
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(t.unsqueeze(1).numpy())
-    ```
-    """
-    dim = self._resolve_dim(dim, extra=True)
-    return self.reshape(self.shape[:dim] + (1,) + self.shape[dim:])
-
-  @property
-  def T(self) -> Self:
-    """`.T` is an alias for `.transpose()`."""
-    return self.transpose()
-
-  def transpose(self, dim0=1, dim1=0) -> Self:
-    """
-    Returns a tensor that is a transposed version of the original tensor.
-    The given dimensions `dim0` and `dim1` are swapped.
-
-    ```python exec="true" source="above" session="tensor" result="python"
-    t = Tensor.arange(6).reshape(2, 3)
-    print(t.numpy())
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(t.transpose(0, 1).numpy())
-    ```
-    """
-    order = list(range(self.ndim))
-    order[dim0], order[dim1] = order[dim1], order[dim0]
-    return self.permute(order)
-
-  def flatten(self, start_dim=0, end_dim=-1) -> Self:
-    """
-    Flattens the tensor by reshaping it into a one-dimensional tensor.
-    If `start_dim` or `end_dim` are passed, only dimensions starting with `start_dim` and ending with `end_dim` are flattened.
-
-    ```python exec="true" source="above" session="tensor" result="python"
-    t = Tensor.arange(8).reshape(2, 2, 2)
-    print(t.flatten().numpy())
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(t.flatten(start_dim=1).numpy())
-    ```
-    """
-    start_dim, end_dim = self._resolve_dim(start_dim), self._resolve_dim(end_dim)
-    return self.reshape(self.shape[:start_dim] + (prod(self.shape[start_dim : end_dim + 1]),) + self.shape[end_dim + 1 :])
-
-  def unflatten(self, dim: int, sizes: tuple[int, ...]) -> Self:
-    """
-    Unflattens dimension `dim` of the tensor into multiple dimensions specified by `sizes`. `Tensor.flatten()` is the inverse of this function.
-
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(Tensor.ones(3, 4, 1).unflatten(1, (2, 2)).shape)
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(Tensor.ones(3, 4, 1).unflatten(1, (-1, 2)).shape)
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(Tensor.ones(5, 12, 3).unflatten(-2, (2, 2, 3, 1, 1)).shape)
-    ```
-    """
-    dim = self._resolve_dim(dim)
-    return self.reshape(self.shape[:dim] + sizes + self.shape[dim + 1 :])
-
-  def rearrange(self, formula: str, **sizes) -> Self:
-    """
-    Rearranges input according to formula
-
-    See: https://einops.rocks/api/rearrange/
-
-    ```python exec="true" source="above" session="tensor" result="python"
-    x = Tensor([[1, 2], [3, 4]])
-    print(Tensor.rearrange(x, "batch channel -> (batch channel)").numpy())
-    ```
-    """
-
-    def parse_formula(formula: str):
-      tokens = f" {formula} ".replace("…", "...").replace("(", " ( ").replace(")", " ) ").replace(" ", "  ").replace(" 1 ", " ( ) ").split()
-      lparens, rparens = map(lambda x: [i for i, ch in enumerate(tokens) if ch == x], ("(", ")"))
-      pairs = list(zip(lparens, rparens))
-      assert len(lparens) == len(rparens) and sorted(flatten(pairs)) == flatten(pairs), "bracket mismatch"
-      return [name for name in tokens if name not in ("(", ")")], [(s - 2 * i, e - 1 - 2 * i) for i, (s, e) in enumerate(pairs)]
-
-    assert formula.count("->") == 1, 'need exactly one "->" in formula'
-
-    (lhs, unflatten_dims), (rhs, flatten_dims) = map(parse_formula, formula.split("->"))
-
-    for name in sizes:
-      assert name in lhs, f"axis {name} is not used in transform"
-    assert sorted(lhs) == sorted(rhs) and len(lhs) == len(set(lhs)), f"name mismatch in {formula}"
-    for name in flatten((lhs, rhs)):
-      assert name == "..." or (name.isidentifier() and "_" not in (name[0], name[-1])), f"invalid axis name {name}"
-    assert "..." not in flatten([lhs[s:e] for s, e in unflatten_dims]), f"cannot have collapsed ellipsis (...) in lhs of {formula}"
-    assert lhs.count("...") <= 1, f"too many ellipses in {formula}"
-
-    # resolve ellipsis
-    if "..." in lhs:
-      ell_len = len(self.shape) - len(lhs) + 1 + sum(e - s - 1 for s, e in unflatten_dims)
-    lhs, rhs = map(lambda l: l[: (i := l.index("..."))] + [f"...{j}" for j in range(ell_len)] + l[i + 1 :] if "..." in l else l, (lhs, rhs))
-    unflatten_dims = [(s + (ell_len - 1 if "...0" in lhs[:s] else 0), e + (ell_len - 1 if "...0" in lhs[:e] else 0)) for s, e in unflatten_dims]
-    flatten_dims = [(s + (ell_len - 1 if "...0" in rhs[:s] else 0), e + (ell_len - 1 if "...0" in rhs[:e] else 0)) for s, e in flatten_dims]
-
-    # apply movement ops in order unflatten -> permute -> flatten/unsqueeze
-    t = functools.reduce(lambda x, dims: x.unflatten(dims[0], tuple(sizes.get(lhs[d], -1) for d in range(*dims))), unflatten_dims, self)
-    for i, name in enumerate(lhs):
-      assert (name not in sizes) or sizes[name] == t.shape[i], f"size provided for dimension {name} incorrect"
-    t = t.permute([lhs.index(name) for name in rhs])
-    return functools.reduce(lambda x, dims: x.flatten(dims[0], dims[1] - 1) if dims[0] < dims[1] else x.unsqueeze(dims[0]), reversed(flatten_dims), t)
-
-  # *** movement ops with expand ***
-
-  def repeat_interleave(self, repeats: int, dim: int | None = None) -> Self:
-    """
-    Repeats elements of a tensor.
-
-    ```python exec="true" source="above" session="tensor" result="python"
-    t = Tensor([1, 2, 3])
-    print(t.repeat_interleave(2).numpy())
-    ```
-    """
-    x, dim = (self.flatten(), 0) if dim is None else (self, self._resolve_dim(dim))
-    shp = x.shape
-    x = x.reshape(*shp[: dim + 1], 1, *shp[dim + 1 :])
-    x = x.expand(*shp[: dim + 1], repeats, *shp[dim + 1 :])
-    x = x.reshape(*shp[:dim], shp[dim] * repeats, *shp[dim + 1 :])
-    return x
-
-  def repeat(self, repeats, *args) -> Self:
-    """
-    Repeats tensor number of times along each dimension specified by `repeats`.
-    `repeats` can be passed as a tuple or as separate arguments.
-
-    ```python exec="true" source="above" session="tensor" result="python"
-    t = Tensor([1, 2, 3])
-    print(t.repeat(4, 2).numpy())
-    ```
-    ```python exec="true" source="above" session="tensor" result="python"
-    print(t.repeat(4, 2, 1).shape)
-    ```
-    """
-    repeats = argfix(repeats, *args)
-    base_shape = _align_left(self.shape, repeats)[0]
-    unsqueezed_shape = flatten([[s] if r == 1 else [1, s] for r, s in zip(repeats, base_shape)])
-    expanded_shape = flatten([[s] if r == 1 else [r, s] for r, s in zip(repeats, base_shape)])
-    final_shape = [r * s for r, s in zip(repeats, base_shape)]
-    return self.reshape(unsqueezed_shape).expand(expanded_shape).reshape(final_shape)
-
-  # **** pool level ****
-
-  def _pool(self, k_: tuple[sint, ...], stride: int | tuple[int, ...] = 1, dilation: int | tuple[int, ...] = 1) -> Self:
-    assert len(self.shape) >= len(k_), f"can't pool {self.shape} with {k_}"
-    s_, d_ = make_tuple(stride, len(k_)), make_tuple(dilation, len(k_))
-    assert len(k_) == len(s_) == len(d_), f"stride/dilation mismatch kernel:{k_} stride:{s_} dilation:{d_}"
-    noop, i_ = [None] * (self.ndim - len(k_)), self.shape[-len(k_) :]
-    assert all(resolve(d * (k - 1) + 1 <= i) for k, d, i in zip(k_, d_, i_)), "kernel size cannot be greater than actual input size"
-    o_ = [ceildiv(i - d * (k - 1), s) for i, d, k, s in zip(i_, d_, k_, s_)]
-    # input size scaling factor to make sure shrink for stride is possible
-    f_ = [smax(1, ceildiv(o * s - d, i)) for o, s, i, d in zip(o_, s_, i_, d_)]
-    # repeats such that we don't need padding
-    x = self.repeat([1] * len(noop) + [ceildiv(k * (i * f + d), i) for k, i, d, f in zip(k_, i_, d_, f_)])
-    # handle dilation
-    x = x.shrink_to(noop + [k * (i * f + d) for k, i, d, f in zip(k_, i_, d_, f_)])
-    x = x.reshape(noop + flatten((k, (i * f + d)) for k, i, d, f in zip(k_, i_, d_, f_)))
-    # handle stride
-    x = x.shrink_to(noop + flatten((k, o * s) for k, o, s in zip(k_, o_, s_))).reshape(noop + flatten((k, o, s) for k, o, s in zip(k_, o_, s_)))
-    x = x.shrink_to(noop + flatten((k, o, 1) for k, o in zip(k_, o_))).reshape(noop + flatten((k, o) for k, o in zip(k_, o_)))
-    # permute to move reduce to the end
-    return x.permute(*range(len(noop)), *[len(noop) + i * 2 + 1 for i in range(len(i_))], *[len(noop) + i * 2 for i in range(len(i_))])
@@ -323,7 +323,7 @@ class Embedding:
    if not dtypes.is_int(idx.dtype): raise TypeError(f"Expected integer dtype for index in embedding, got {idx.dtype}")
    big_shp = idx.shape+(self.vocab_sz, self.embed_sz)
    arange, idx, vals = self.arange.expand(big_shp), idx.reshape(idx.shape+(1, 1)).expand(big_shp), self.weight.expand(big_shp)
-    return (arange == idx).where(vals, 0).sum(-2, dtype=vals.dtype)
+    return (arange == idx).mul(vals).sum(-2, dtype=vals.dtype)

 class LSTMCell:
  """
@@ -1124,16 +1124,6 @@ def get_onnx_ops() -> dict[str, types.FunctionType|dict[OpSetId, types.FunctionT
    return output.flatten(start_dim=2) if len(original_input_shape) == 3 else output.permute(0, 2, 1, 3)

  # ***** Indexing Ops *****
-  def NonZero(x:Tensor):
-    mask = (x!=0).flatten()
-    flat_idx = Tensor.arange(mask.numel(), dtype=dtypes.int64, device=x.device).masked_select(mask)
-    if flat_idx.ndim == 0: flat_idx = flat_idx.reshape(1)
-    if x.ndim == 0:
-      return Tensor.zeros((0, flat_idx.shape[0]), dtype=dtypes.int64, device=x.device, requires_grad=False)
-    strides = [prod(int(s) for s in x.shape[i+1:]) if i+1 < x.ndim else 1 for i in range(x.ndim)]
-    coords = [((flat_idx // stride) % int(dim)) for stride, dim in zip(strides, x.shape)]
-    return Tensor.stack(*coords, dim=0)
-
  def ArrayFeatureExtractor(x:Tensor, indices:Tensor): return x[..., indices]

  def Gather(x:Tensor, indices:Tensor, axis:int=0):
@@ -194,10 +194,6 @@ def torch_load(t:Tensor) -> dict[str, Tensor]:
  """
  offsets: dict[str|int, int] = {}
  lens: dict[str|int, int] = {}
-
-  def _rebuild_tensor(storage, storage_offset, size, stride):
-    return _rebuild_tensor_v2(storage, storage_offset, size, stride)
-
  def _rebuild_tensor_v2(storage, storage_offset, size, stride, requires_grad=None, backward_hooks=None, metadata=None):
    #print(storage, storage_offset, size, stride, requires_grad, backward_hooks, metadata)
    lens[storage[2]] = storage[4] * storage[1].itemsize
@@ -224,8 +220,7 @@ def torch_load(t:Tensor) -> dict[str, Tensor]:
  deserialized_objects: dict[str, Any] = {}
  intercept = {"HalfStorage": dtypes.float16, "FloatStorage": dtypes.float32, "BFloat16Storage": dtypes.bfloat16,
               "IntStorage": dtypes.int32, "BoolStorage": dtypes.bool,
-               "LongStorage": dtypes.int64, "_rebuild_tensor": _rebuild_tensor, "_rebuild_tensor_v2": _rebuild_tensor_v2,
-               "FloatTensor": None, "Parameter": Parameter}
+               "LongStorage": dtypes.int64, "_rebuild_tensor_v2": _rebuild_tensor_v2, "FloatTensor": None, "Parameter": Parameter}
  whitelist = {"torch", "collections", "numpy", "_codecs"}  # NOTE: this is not for security, only speed
  class Dummy: pass
  class TorchPickle(pickle.Unpickler):
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
geohot	6ffd33e1e5	fix later	2025-11-02 13:18:41 +08:00
geohot	4198efb8bc	mixin	2025-11-02 13:05:26 +08:00
geohot	13e8914deb	use Self type	2025-11-02 13:00:35 +08:00