works

* fix const padding cpu

* remove comment

.github/actions/process-replay/action.yml

@@ -11,5 +11,5 @@ runs:
         git fetch origin $CURRENT_SHA
         export COMMIT_MESSAGE=$(git show -s --format=%B "$CURRENT_SHA")
         export CURRENT_HEAD=$(git rev-parse HEAD)
-        cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && IGNORE_OOB=1 PYTHONPATH=. python3 process_replay.py
+        cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && CHECK_OOB=0 PYTHONPATH=. python3 process_replay.py
         git checkout $CURRENT_HEAD  # restore to branch

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

@@ -56,7 +56,15 @@ runs:
 
     # **** Caching packages ****
 
+    - name: Cache Python packages (PR)
+      if: github.event_name == 'pull_request'
+      id: restore-venv-pr
+      uses: actions/cache/restore@v4
+      with:
+        path: ${{ github.workspace }}/.venv
+        key: venv-${{ runner.os }}-python-${{ steps.setup-python.outputs.python-version }}-${{ inputs.deps }}-${{ inputs.pydeps }}-${{ env.CACHE_VERSION }}
     - name: Cache Python packages
+      if: github.event_name != 'pull_request'
       id: restore-venv
       uses: actions/cache@v4
       with:
@@ -65,23 +73,23 @@ runs:
 
     # **** Caching downloads ****
 
-    - name: Cache downloads (Linux)
-      if: inputs.key != '' && runner.os == 'Linux'
-      uses: actions/cache@v4
+    - name: Cache downloads (PR)
+      if: inputs.key != '' && github.event_name == 'pull_request'
+      uses: actions/cache/restore@v4
       with:
-        path: ~/.cache/tinygrad/downloads/
+        path: ${{ runner.os == 'Linux' && '~/.cache/tinygrad/downloads/' || '~/Library/Caches/tinygrad/downloads/' }}
         key: downloads-${{ github.job }}-${{ inputs.key }}-${{ env.CACHE_VERSION }}
-    - name: Cache downloads (macOS)
-      if: inputs.key != '' && runner.os == 'macOS'
+    - name: Cache downloads
+      if: inputs.key != '' && github.event_name != 'pull_request'
       uses: actions/cache@v4
       with:
-        path: ~/Library/Caches/tinygrad/downloads/
+        path: ${{ runner.os == 'Linux' && '~/.cache/tinygrad/downloads/' || '~/Library/Caches/tinygrad/downloads/' }}
         key: downloads-${{ github.job }}-${{ inputs.key }}-${{ env.CACHE_VERSION }}
 
     # **** Python deps ****
 
     - name: Install dependencies in venv (with extra)
-      if: inputs.deps != '' && steps.restore-venv.outputs.cache-hit != 'true'
+      if: inputs.deps != '' && steps.restore-venv-pr.outputs.cache-hit != 'true' && steps.restore-venv.outputs.cache-hit != 'true'
       shell: bash
       run: |
         python -m venv .venv
@@ -92,7 +100,7 @@ runs:
         fi
         python -m pip install -e ".[${{ inputs.deps }}]" ${{ inputs.pydeps }} --extra-index-url https://download.pytorch.org/whl/cpu --extra-index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/Triton-Nightly/pypi/simple/
     - name: Install dependencies in venv (without extra)
-      if: inputs.deps == '' && steps.restore-venv.outputs.cache-hit != 'true'
+      if: inputs.deps == '' && steps.restore-venv-pr.outputs.cache-hit != 'true' && steps.restore-venv.outputs.cache-hit != 'true'
       shell: bash
       run: |
         python -m venv .venv
@@ -137,7 +145,7 @@ runs:
       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/6.2 $(lsb_release -cs) main
+        deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/7.1 $(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
 
@@ -182,8 +190,14 @@ runs:
         echo "pkgs=$pkgs" >> "$GITHUB_OUTPUT"
         echo "hash=$(echo -n "$pkgs" | sha256sum | cut -d' ' -f1)" >> "$GITHUB_OUTPUT"
 
+    - name: Cache apt (PR)
+      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.cuda == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true') && github.event_name == 'pull_request'
+      uses: actions/cache/restore@v4
+      with:
+        path: /var/cache/apt/archives/
+        key: ${{ runner.os }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.CACHE_VERSION }}
     - name: Cache apt
-      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.cuda == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true')
+      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.cuda == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true') && github.event_name != 'pull_request'
       uses: actions/cache@v4
       with:
         path: /var/cache/apt/archives/
@@ -219,7 +233,7 @@ runs:
       shell: bash
       run: |
         sudo mkdir -p /usr/local/lib
-        curl -s -H "Authorization: token $GH_TOKEN" curl -s https://api.github.com/repos/nimlgen/amdcomgr_dylib/releases/latest | \
+        curl -s -H "Authorization: token $GH_TOKEN" curl -s https://api.github.com/repos/tinygrad/amdcomgr_dylib/releases/latest | \
           jq -r '.assets[] | select(.name == "libamd_comgr.dylib").browser_download_url' | \
           sudo xargs curl -fL -o /usr/local/lib/libamd_comgr.dylib
         cargo build --release --manifest-path ./extra/remu/Cargo.toml
@@ -239,8 +253,17 @@ runs:
         ln -s /opt/homebrew/opt/[email protected] /opt/homebrew/opt/boost || true
         ln -s /opt/homebrew/opt/boost/lib/libboost_atomic-mt.dylib /opt/homebrew/opt/boost/lib/libboost_atomic.dylib || true
         ln -s /opt/homebrew/opt/boost/lib/libboost_thread-mt.dylib /opt/homebrew/opt/boost/lib/libboost_thread.dylib || true
+    - name: Cache gpuocelot (PR)
+      if: inputs.ocelot == 'true' && github.event_name == 'pull_request'
+      id: cache-build-pr
+      uses: actions/cache/restore@v4
+      env:
+        cache-name: cache-gpuocelot-build-1
+      with:
+        path: ${{ github.workspace }}/gpuocelot/ocelot
+        key: ${{ runner.os }}-gpuocelot-b16039dc940dc6bc4ea0a98380495769ff35ed99-rebuild-${{ env.CACHE_VERSION }}
     - name: Cache gpuocelot
-      if: inputs.ocelot == 'true'
+      if: inputs.ocelot == 'true' && github.event_name != 'pull_request'
       id: cache-build
       uses: actions/cache@v4
       env:
@@ -249,7 +272,7 @@ runs:
         path: ${{ github.workspace }}/gpuocelot/ocelot
         key: ${{ runner.os }}-gpuocelot-b16039dc940dc6bc4ea0a98380495769ff35ed99-rebuild-${{ env.CACHE_VERSION }}
     - name: Clone/compile gpuocelot
-      if: inputs.ocelot == 'true' && steps.cache-build.outputs.cache-hit != 'true'
+      if: inputs.ocelot == 'true' && steps.cache-build-pr.outputs.cache-hit != 'true' && steps.cache-build.outputs.cache-hit != 'true'
       shell: bash
       run: |
         git clone --recurse-submodules https://github.com/gpuocelot/gpuocelot.git ${{ github.workspace }}/gpuocelot

.github/workflows/autogen.yml

@@ -14,10 +14,12 @@ on:
     paths:
     - 'tinygrad/runtime/autogen/**/*'
     - 'tinygrad/runtime/support/autogen.py'
+    - '.github/workflows/autogen.yml'
   workflow_dispatch:
     paths:
     - 'tinygrad/runtime/autogen/**/*'
     - 'tinygrad/runtime/support/autogen.py'
+    - '.github/workflows/autogen.yml'
 
 jobs:
   autogen:
@@ -30,6 +32,7 @@ jobs:
     - name: Setup Environment
       uses: ./.github/actions/setup-tinygrad
       with:
+        key: 'autogen'
         opencl: 'true'
         amd: 'true'
         cuda: 'true'
@@ -38,103 +41,37 @@ jobs:
         mesa: '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
-    - name: Verify OpenCL autogen
+      run: sudo apt-get install -y --no-install-recommends libclang-20-dev llvm-20-dev hip-dev libusb-1.0-0-dev libdrm-dev
+    - name: Regenerate autogen files
       run: |
-        mv tinygrad/runtime/autogen/opencl.py /tmp/opencl.py.bak
+        find tinygrad/runtime/autogen -type f -name "*.py" -not -path "*/amd/*" -not -name "__init__.py" -not -name "comgr.py" -not -name "metal.py" -not -name "iokit.py" -not -name "corefoundation.py" -not -name "libclang.py" -delete
         python3 -c "from tinygrad.runtime.autogen import 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"
-        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
-    - 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
-        diff /tmp/hsa.py.bak tinygrad/runtime/autogen/hsa.py
-        diff /tmp/hip.py.bak tinygrad/runtime/autogen/hip.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 cuda, nvrtc, nvjitlink, nv_570, nv_580, nv"
+        python3 -c "from tinygrad.runtime.autogen import comgr_3, hsa, hip, amd_gpu, sqtt, rocprof, amdgpu_kd, amdgpu_drm"
+        python3 -c "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_v13_0_6, smu_v13_0_12, smu_v14_0_2"
         python3 -c "from tinygrad.runtime.autogen import libc, kfd, io_uring, ib, pci, vfio"
-        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"
-        diff /tmp/webgpu.py.bak tinygrad/runtime/autogen/webgpu.py
-    - name: Verify Qualcomm autogen
-      run: |
-        mv tinygrad/runtime/autogen/kgsl.py /tmp/kgsl.py.bak
-        mv tinygrad/runtime/autogen/qcom_dsp.py /tmp/qcom_dsp.py.bak
         python3 -c "from tinygrad.runtime.autogen import kgsl, qcom_dsp"
-        diff /tmp/kgsl.py.bak tinygrad/runtime/autogen/kgsl.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
-    - name: Verify mesa autogen
-      run: |
-        mv tinygrad/runtime/autogen/mesa.py /tmp/mesa.py.bak
         python3 -c "from tinygrad.runtime.autogen import mesa"
-        diff /tmp/mesa.py.bak tinygrad/runtime/autogen/mesa.py
-    - name: Verify libclang autogen
-      run: |
-        cp tinygrad/runtime/autogen/libclang.py /tmp/libclang.py.bak
+        python3 -c "from tinygrad.runtime.autogen import avcodec"
         REGEN=1 python3 -c "from tinygrad.runtime.autogen import libclang"
-        diff /tmp/libclang.py.bak tinygrad/runtime/autogen/libclang.py
+    - name: Check for differences
+      run: |
+        if ! git diff --quiet; then
+          git diff
+          git diff > autogen-ubuntu.patch
+          echo "Autogen mismatch detected. Patch available at: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}#artifacts"
+          exit 1
+        fi
+    - name: Upload patch artifact
+      if: failure()
+      uses: actions/upload-artifact@v4
+      with:
+        name: autogen-ubuntu-patch
+        path: autogen-ubuntu.patch
+
   autogen-mac:
     name: In-tree Autogen (macos)
     runs-on: macos-14
@@ -145,14 +82,29 @@ jobs:
     - name: Setup Environment
       uses: ./.github/actions/setup-tinygrad
       with:
+        key: 'autogen-mac'
         llvm: 'true'
-    - name: Verify macos autogen
+    - name: Regenerate autogen files
       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)
+        rm tinygrad/runtime/autogen/metal.py tinygrad/runtime/autogen/iokit.py tinygrad/runtime/autogen/corefoundation.py
+        python3 -c "from tinygrad.runtime.autogen import metal, iokit, corefoundation"
+    - name: Check for differences
+      run: |
+        if ! git diff --quiet; then
+          git diff
+          git diff > autogen-macos.patch
+          echo "Autogen mismatch detected. Patch available at: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}#artifacts"
+          exit 1
+        fi
+    - name: Upload patch artifact
+      if: failure()
+      uses: actions/upload-artifact@v4
+      with:
+        name: autogen-macos-patch
+        path: autogen-macos.patch
+
+  autogen-comgr-2:
+    name: In-tree Autogen (comgr 2)
     runs-on: ubuntu-24.04
     timeout-minutes: 15
     steps:
@@ -160,17 +112,32 @@ jobs:
       uses: actions/checkout@v4
     - name: Setup Environment
       uses: ./.github/actions/setup-tinygrad
+      with:
+        key: 'autogen-comgr'
     - 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
+        deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/6.2 $(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
+    - name: Regenerate autogen files
       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
+        rm tinygrad/runtime/autogen/comgr.py
+        python3 -c "from tinygrad.runtime.autogen import comgr"
+    - name: Check for differences
+      run: |
+        if ! git diff --quiet; then
+          git diff
+          git diff > autogen-comgr2.patch
+          echo "Autogen mismatch detected. Patch available at: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}#artifacts"
+          exit 1
+        fi
+    - name: Upload patch artifact
+      if: failure()
+      uses: actions/upload-artifact@v4
+      with:
+        name: autogen-comgr2-patch
+        path: autogen-comgr2.patch

.github/workflows/benchmark.yml

@@ -16,6 +16,43 @@ on:
   workflow_dispatch:
 
 jobs:
+  # the goal of this test is to replicate a normal person on a laptop running the test
+  # no process replay, no benchmarks, no CI, just a normal laptop person
+  # the 3 minute timeout should not be raised
+  testmacpytest:
+    name: Mac pytest
+    env:
+      CI: ""
+      CAPTURE_PROCESS_REPLAY: "0"
+    runs-on: [self-hosted, macOS]
+    timeout-minutes: 3
+    defaults:
+      run:
+        shell: bash -e -o pipefail {0}
+    if: github.repository_owner == 'tinygrad'
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v4
+    # brew install uv
+    - name: setup python environment
+      run: |
+        rm -rf /tmp/tinygrad_pytest_ci
+        uv venv /tmp/tinygrad_pytest_ci
+        source /tmp/tinygrad_pytest_ci/bin/activate
+        uv pip install .[testing]
+    - name: setup staging db
+      run: |
+        echo "CACHEDB=/tmp/pytest-db-ci.db" >> $GITHUB_ENV
+        rm -f /tmp/pytest-db-ci*
+    - name: Run pytest -nauto
+      run: |
+        source /tmp/tinygrad_pytest_ci/bin/activate
+        pytest -nauto --durations=20
+    - name: openpilot compile3 0.10.1 driving_vision
+      run: FLOAT16=1 CL=1 IMAGE=2 python3.11 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
+    - name: IMAGE=1 openpilot compile3 0.10.1 driving_vision
+      run: FLOAT16=1 CL=1 IMAGE=1 python3.11 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
+
   testmacbenchmark:
     name: Mac Benchmark
     env:
@@ -49,19 +86,19 @@ jobs:
     - name: Print macOS version
       run: sw_vers
     - name: Run Stable Diffusion
-      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
+      run: BENCHMARK_LOG=stable_diffusion JIT=1 ASSERT_MIN_STEP_TIME=720 python3.11 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing
     - name: Run Stable Diffusion without fp16
-      run: BENCHMARK_LOG=stable_diffusion_fp32 JIT=1 ASSERT_MIN_STEP_TIME=720 python3.11 examples/stable_diffusion.py --seed 0 --noshow --timing | tee sd_no_fp16.txt
+      run: BENCHMARK_LOG=stable_diffusion_fp32 JIT=1 ASSERT_MIN_STEP_TIME=720 python3.11 examples/stable_diffusion.py --seed 0 --noshow --timing
     - name: Run Stable Diffusion v2
       # TODO: very slow step time
-      run: BENCHMARK_LOG=stable_diffusion_v2 JIT=1 ASSERT_MIN_STEP_TIME=4500 python3.11 examples/sdv2.py --fp16 --seed 0 --noshow --timing | tee sdv2.txt
+      run: BENCHMARK_LOG=stable_diffusion_v2 JIT=1 ASSERT_MIN_STEP_TIME=4500 python3.11 examples/sdv2.py --fp16 --seed 0 --noshow --timing
     # process replay can't capture this, the graph is too large
     - name: Run SDXL
-      run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=5000 CAPTURE_PROCESS_REPLAY=0 JIT=1 python3.11 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
+      run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=5000 CAPTURE_PROCESS_REPLAY=0 JIT=1 python3.11 examples/sdxl.py --seed 0 --noshow --timing
     - name: Run model inference benchmark
       run: METAL=1 NOCLANG=1 python3.11 test/external/external_model_benchmark.py
     - name: Test speed vs torch
-      run: BIG=2 MPS=1 python3.11 test/speed/external_test_speed_v_torch.py | tee torch_speed.txt
+      run: BIG=2 MPS=1 python3.11 test/speed/external_test_speed_v_torch.py
     - name: Test tensor cores
       run: METAL=1 python3.11 test/opt/test_tensor_cores.py
     - name: Test AMX tensor cores
@@ -71,84 +108,59 @@ jobs:
         DEBUG=2 CPU=1 CPU_LLVM=0 AMX=1 python3.11 test/opt/test_gen_float4.py TestFloat4.test_float4_multidim_amx TestFloat4.test_float4_multidim_unaligned_load_amx
         DEBUG=2 CPU=1 CPU_LLVM=1 AMX=1 python3.11 test/opt/test_gen_float4.py TestFloat4.test_float4_multidim_amx TestFloat4.test_float4_multidim_unaligned_load_amx
     - name: Run Tensor Core GEMM (float)
-      run: DEBUG=2 SHOULD_USE_TC=1 python3.11 extra/gemm/simple_matmul.py | tee matmul.txt
+      run: DEBUG=2 SHOULD_USE_TC=1 python3.11 extra/gemm/simple_matmul.py
     - name: Run Tensor Core GEMM (half)
-      run: DEBUG=2 SHOULD_USE_TC=1 HALF=1 python3.11 extra/gemm/simple_matmul.py | tee matmul_half.txt
+      run: DEBUG=2 SHOULD_USE_TC=1 HALF=1 python3.11 extra/gemm/simple_matmul.py
     - name: Run Tensor Core GEMM (bfloat16)
-      run: DEBUG=2 SHOULD_USE_TC=1 BFLOAT16=1 python3.11 extra/gemm/simple_matmul.py | tee matmul_bfloat16.txt
+      run: DEBUG=2 SHOULD_USE_TC=1 BFLOAT16=1 python3.11 extra/gemm/simple_matmul.py
     - name: Fuzz Padded Tensor Core GEMM
       run: METAL=1 M_START=6 M_STOP=10 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=6 K_STOP=24 K_STEP=1 TC_OPT=2 DEBUG=2 python3.11 ./extra/gemm/fuzz_matmul.py
     - name: Run LLaMA
       run: |
-        BENCHMARK_LOG=llama_nojit JIT=0 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_unjitted.txt
-        BENCHMARK_LOG=llama JIT=1 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_jitted.txt
+        BENCHMARK_LOG=llama_nojit JIT=0 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
+        BENCHMARK_LOG=llama JIT=1 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
     - name: Run LLaMA with BEAM
-      run: BENCHMARK_LOG=llama_beam JITBEAM=2 IGNORE_BEAM_CACHE=1 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_beam.txt
+      run: BENCHMARK_LOG=llama_beam JITBEAM=2 IGNORE_BEAM_CACHE=1 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
     - name: Run quantized LLaMA
       run: |
-        BENCHMARK_LOG=llama_int8 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing --quantize int8 | tee llama_int8.txt
-        BENCHMARK_LOG=llama_nf4 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing --quantize nf4 | tee llama_nf4.txt
+        BENCHMARK_LOG=llama_int8 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing --quantize int8
+        BENCHMARK_LOG=llama_nf4 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing --quantize nf4
     - name: Run quantized LLaMA3
       run: |
-        BENCHMARK_LOG=llama3_int8 python3.11 examples/llama3.py --size 8B --temperature 0 --benchmark --quantize int8 | tee llama3_int8.txt
-        BENCHMARK_LOG=llama3_nf4 python3.11 examples/llama3.py --size 8B --temperature 0 --benchmark --quantize nf4 | tee llama3_nf4.txt
+        BENCHMARK_LOG=llama3_int8 python3.11 examples/llama3.py --size 8B --temperature 0 --benchmark --quantize int8
+        BENCHMARK_LOG=llama3_nf4 python3.11 examples/llama3.py --size 8B --temperature 0 --benchmark --quantize nf4
     #- name: Run LLaMA 7B on 4 (virtual) GPUs
-    #  run: python3.11 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_four_gpu.txt
+    #  run: python3.11 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing
     - name: Run GPT2
       run: |
-        BENCHMARK_LOG=gpt2_nojit JIT=0 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
-        BENCHMARK_LOG=gpt2 JIT=1 ASSERT_MIN_STEP_TIME=13 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
+        BENCHMARK_LOG=gpt2_nojit JIT=0 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
+        BENCHMARK_LOG=gpt2 JIT=1 ASSERT_MIN_STEP_TIME=13 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
     - name: Run GPT2 w HALF
-      run: BENCHMARK_LOG=gpt2_half HALF=1 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
+      run: BENCHMARK_LOG=gpt2_half HALF=1 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing
     - name: Run GPT2 w HALF/BEAM
-      run: BENCHMARK_LOG=gpt2_half_beam HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
+      run: BENCHMARK_LOG=gpt2_half_beam HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing
     - name: Run OLMoE
       run: BENCHMARK_LOG=olmoe python3.11 examples/olmoe.py
     - name: Train MNIST
-      run: time PYTHONPATH=. TARGET_EVAL_ACC_PCT=96.0 python3.11 examples/beautiful_mnist.py | tee beautiful_mnist.txt
+      run: time PYTHONPATH=. TARGET_EVAL_ACC_PCT=96.0 python3.11 examples/beautiful_mnist.py
 
     # NOTE: this is failing in CI. it is not failing on my machine and I don't really have a way to debug it
     # the error is "RuntimeError: Internal Error (0000000e:Internal Error)"
     #- name: Run 10 CIFAR training steps
-    #  run: BENCHMARK_LOG=cifar_10steps JIT=1 ASSERT_MIN_STEP_TIME=3000 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar.txt
+    #  run: BENCHMARK_LOG=cifar_10steps JIT=1 ASSERT_MIN_STEP_TIME=3000 STEPS=10 python3.11 examples/hlb_cifar10.py
     #- name: Run 10 CIFAR training steps w HALF
-    #  run: BENCHMARK_LOG=cifar_10steps_half JIT=2 ASSERT_MIN_STEP_TIME=3000 STEPS=10 DEFAULT_FLOAT=HALF python3.11 examples/hlb_cifar10.py | tee train_cifar_half.txt
+    #  run: BENCHMARK_LOG=cifar_10steps_half JIT=2 ASSERT_MIN_STEP_TIME=3000 STEPS=10 DEFAULT_FLOAT=HALF python3.11 examples/hlb_cifar10.py
 
     #- name: Run 10 CIFAR training steps w BF16
-    #  run: STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3.11 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
+    #  run: STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3.11 examples/hlb_cifar10.py
     # TODO: too slow
     # - name: Run 10 CIFAR training steps w winograd
-    #   run: BENCHMARK_LOG=cifar_10steps_wino JIT=1 ASSERT_MIN_STEP_TIME=150 WINO=1 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar_wino.txt
+    #   run: BENCHMARK_LOG=cifar_10steps_wino JIT=1 ASSERT_MIN_STEP_TIME=150 WINO=1 STEPS=10 python3.11 examples/hlb_cifar10.py
     - uses: actions/upload-artifact@v4
       with:
         name: Speed (Mac)
         path: |
           onnx_inference_speed.csv
-          torch_speed.txt
-          llama_unjitted.txt
-          llama_jitted.txt
-          llama_beam.txt
-          llama_int8.txt
-          llama_nf4.txt
-          llama3_int8.txt
-          llama3_nf4.txt
-          llama_four_gpu.txt
-          gpt2_unjitted.txt
-          gpt2_jitted.txt
-          gpt2_half.txt
-          gpt2_half_beam.txt
-          matmul.txt
-          matmul_half.txt
-          matmul_bfloat16.txt
-          sd.txt
-          sd_no_fp16.txt
-          sdv2.txt
-          sdxl.txt
-          beautiful_mnist.txt
-          train_cifar.txt
-          train_cifar_half.txt
-          train_cifar_bf16.txt
-          train_cifar_wino.txt
     - name: Run process replay tests
       run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3.11 process_replay.py
 
@@ -170,6 +182,10 @@ jobs:
       run: |
         echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
         rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
+    - name: Kill stale pids
+      run: |
+        PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
+        PYTHONPATH=. ./extra/hcq/hcq_smi.py nv kill_pids
     - name: UsbGPU boot time
       run: sudo -E PYTHONPATH=. DEBUG=2 AM_RESET=1 AMD=1 AMD_IFACE=USB time python3.11 test/test_tiny.py TestTiny.test_plus
     - name: UsbGPU tiny tests
@@ -215,7 +231,7 @@ jobs:
     - name: Run model inference benchmark
       run: NV=1 CAPTURE_PROCESS_REPLAY=0 NOCLANG=1 python3 test/external/external_model_benchmark.py
     - name: Test speed vs torch
-      run: NV=1 CAPTURE_PROCESS_REPLAY=0 HALF=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py | tee torch_speed.txt
+      run: NV=1 CAPTURE_PROCESS_REPLAY=0 HALF=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py
     - 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
     - name: Test benchmark allreduce
@@ -226,79 +242,58 @@ jobs:
         NV=1 NV_PTX=1 ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
     - name: Run Tensor Core GEMM (CUDA)
       run: |
-        CUDA=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul.txt
-        CUDA=1 SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_bfloat16.txt
-        CUDA=1 SHOULD_USE_TC=1 ALLOW_TF32=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py | tee matmul_tf32.txt
-        CUDA=1 SHOULD_USE_TC=1 FP8E4M3=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_fp8.txt
+        CUDA=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+        CUDA=1 SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+        CUDA=1 SHOULD_USE_TC=1 ALLOW_TF32=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
+        CUDA=1 SHOULD_USE_TC=1 FP8E4M3=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
     - name: Run Tensor Core GEMM (PTX)
-      run: NV=1 NV_PTX=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_ptx.txt
+      run: NV=1 NV_PTX=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
     - name: Run Tensor Core GEMM (NV)
-      run: NV=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_nv.txt
+      run: NV=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
     - name: Test NV=1
       run: DEBUG=2 NV=1 python -m pytest -rA test/test_tiny.py
     - name: Test CUDA=1
       run: DEBUG=2 CUDA=1 python -m pytest -rA test/test_tiny.py
     - name: Run Stable Diffusion
-      run: BENCHMARK_LOG=stable_diffusion NV=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
+      run: BENCHMARK_LOG=stable_diffusion NV=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing
     # TODO: too slow
     # - name: Run SDXL
-    #   run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=2000 CAPTURE_PROCESS_REPLAY=0 NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
+    #   run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=2000 CAPTURE_PROCESS_REPLAY=0 NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/sdxl.py --seed 0 --noshow --timing
     - name: Run LLaMA
       run: |
-        BENCHMARK_LOG=llama_nojit NV=1 JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_unjitted.txt
-        BENCHMARK_LOG=llama NV=1 JIT=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_jitted.txt
+        BENCHMARK_LOG=llama_nojit NV=1 JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
+        BENCHMARK_LOG=llama NV=1 JIT=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
     - name: Run LLaMA with BEAM
-      run: BENCHMARK_LOG=llama_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_beam.txt
+      run: BENCHMARK_LOG=llama_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
     # - name: Run LLaMA 7B on 4 GPUs
-    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_four_gpu.txt
+    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing
     # - name: Run LLaMA 7B on 6 GPUs
-    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_six_gpu.txt
+    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
     - name: Run LLaMA-3 8B BEAM
-      run: BENCHMARK_LOG=llama3_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_beam.txt
+      run: BENCHMARK_LOG=llama3_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
     - name: Run LLaMA-3 8B on 4 GPUs with BEAM
-      run: BENCHMARK_LOG=llama3_beam_4gpu NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_four_gpu.txt
+      run: BENCHMARK_LOG=llama3_beam_4gpu NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
     - name: Run quantized LLaMA3
-      run: BENCHMARK_LOG=llama3_fp8 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --temperature 0 --benchmark --quantize fp8 | tee llama3_fp8.txt
+      run: BENCHMARK_LOG=llama3_fp8 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --temperature 0 --benchmark --quantize fp8
     # - name: Run LLaMA-3 8B on 6 GPUs
-    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 6 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_six_gpu.txt
+    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 6 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
     # - name: Run LLaMA-2 70B
-    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 MAX_CONTEXT=256 python3 examples/llama.py --gen 2 --size 70B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_2_70B.txt
+    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 MAX_CONTEXT=256 python3 examples/llama.py --gen 2 --size 70B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
     - name: Run Mixtral 8x7B
-      run: time BENCHMARK_LOG=mixtral NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/mixtral.py --temperature 0 --count 10 --timing | tee mixtral.txt
+      run: time BENCHMARK_LOG=mixtral NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/mixtral.py --temperature 0 --count 10 --timing
     - name: Run GPT2
       run: |
-        BENCHMARK_LOG=gpt2_nojit NV=1 JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
-        BENCHMARK_LOG=gpt2 NV=1 JIT=1 ASSERT_MIN_STEP_TIME=4 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
+        BENCHMARK_LOG=gpt2_nojit NV=1 JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
+        BENCHMARK_LOG=gpt2 NV=1 JIT=1 ASSERT_MIN_STEP_TIME=4 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
     - name: Run GPT2 w HALF
-      run: BENCHMARK_LOG=gpt2_half NV=1 HALF=1 ASSERT_MIN_STEP_TIME=6 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
+      run: BENCHMARK_LOG=gpt2_half NV=1 HALF=1 ASSERT_MIN_STEP_TIME=6 python3 examples/gpt2.py --count 10 --temperature 0 --timing
     - name: Run GPT2 w HALF/BEAM
-      run: BENCHMARK_LOG=gpt2_half_beam NV=1 HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
+      run: BENCHMARK_LOG=gpt2_half_beam NV=1 HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing
     - uses: actions/upload-artifact@v4
       with:
         name: Speed (NVIDIA)
         path: |
           onnx_inference_speed.csv
-          torch_speed.txt
-          matmul.txt
-          matmul_bfloat16.txt
-          matmul_tf32.txt
-          matmul_ptx.txt
-          matmul_nv.txt
-          sd.txt
-          sdxl.txt
-          llama_unjitted.txt
-          llama_jitted.txt
-          llama_beam.txt
-          llama3_beam.txt
-          llama3_four_gpu.txt
-          llama3_six_gpu.txt
-          llama3_fp8.txt
-          llama_2_70B.txt
-          mixtral.txt
-          gpt2_unjitted.txt
-          gpt2_jitted.txt
-          gpt2_half.txt
-          gpt2_half_beam.txt
     - name: Run process replay tests
       run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py
 
@@ -337,44 +332,30 @@ jobs:
     # - name: Fuzz Padded Tensor Core GEMM (PTX)
     #   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: HEVC Decode Benchmark
-      run: VALIDATE=1 MAX_FRAMES=100 NV=1 PYTHONPATH=. python3 extra/hevc/decode.py
+      run: VALIDATE=1 MAX_FRAMES=100 JITBEAM=1 NV=1 PYTHONPATH=. python3 extra/hevc/decode.py
     - name: Train MNIST
-      run: time PYTHONPATH=. NV=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py | tee beautiful_mnist.txt
+      run: time PYTHONPATH=. NV=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py
     - name: Run 10 CIFAR training steps
-      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=120 NV=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
+      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=120 NV=1 STEPS=10 python3 examples/hlb_cifar10.py
     - name: Run 10 CIFAR training steps w HALF
-      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=110 NV=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_half.txt
+      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=120 NV=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
     - name: Run 10 CIFAR training steps w BF16
-      run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=120 NV=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
+      run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=120 NV=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py
     # - 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
+    #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=350 NV=1 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
     - 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
+      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
     - 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
+      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
     - 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)
-      run: BENCHMARK_LOG=resnet_10steps NV=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet_one_gpu.txt
+      run: BENCHMARK_LOG=resnet_10steps NV=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
     - name: Run 10 MLPerf ResNet50 training steps (6 gpu)
-      run: BENCHMARK_LOG=resnet_10steps_6gpu NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet.txt
+      run: BENCHMARK_LOG=resnet_10steps_6gpu NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py
     - name: Run 10 MLPerf Bert training steps (6 gpu)
       # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps_6gpu NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=72 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py | tee train_bert.txt
-    - uses: actions/upload-artifact@v4
-      with:
-        name: Speed (NVIDIA Training)
-        path: |
-          beautiful_mnist.txt
-          train_cifar.txt
-          train_cifar_half.txt
-          train_cifar_bf16.txt
-          train_cifar_wino.txt
-          train_cifar_one_gpu.txt
-          train_cifar_six_gpu.txt
-          train_resnet.txt
-          train_resnet_one_gpu.txt
-          train_bert.txt
+      run: BENCHMARK_LOG=bert_10steps_6gpu NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=72 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
     - name: Run process replay tests
       run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py
 
@@ -389,10 +370,12 @@ jobs:
     steps:
     - name: Checkout Code
       uses: actions/checkout@v4
-    - name: Remove amdgpu
-      run: sudo rmmod amdgpu || true
-    - name: Cleanup running AM processes
-      run: python extra/amdpci/am_smi.py --pids --kill
+    - name: Setcap to python
+      run: ./extra/amdpci/setup_python_cap.sh
+    - name: Remove amd modules
+      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd rmmod
+    - name: Kill stale pids
+      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
     #- name: Insert amdgpu
     #  run: sudo modprobe amdgpu
     - name: Symlink models and datasets
@@ -426,7 +409,7 @@ jobs:
     #- name: Test speed vs torch
     #  run: |
     #    python3 -c "import torch; print(torch.__version__)"
-    #    LD_PRELOAD="/opt/rocm/lib/libhsa-runtime64.so" HSA=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py | tee torch_speed.txt
+    #    LD_PRELOAD="/opt/rocm/lib/libhsa-runtime64.so" HSA=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py
     - 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
     - name: Test tensor cores AMD_LLVM=0
@@ -437,7 +420,7 @@ jobs:
     - name: Run Tensor Core GEMM (AMD)
       run: |
         AMD=1 SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
-        AMD=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py | tee matmul_amd.txt
+        AMD=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
     - name: Test AMD=1
       run: DEBUG=2 AMD=1 python -m pytest -rA test/test_tiny.py
     #- name: Test HIP=1
@@ -452,61 +435,39 @@ jobs:
     - name: Test AM warm start time
       run: time AMD=1 python3 test/test_tiny.py TestTiny.test_plus
     - name: Run Stable Diffusion
-      run: BENCHMARK_LOG=stable_diffusion ASSERT_MIN_STEP_TIME=550 AMD=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
+      run: BENCHMARK_LOG=stable_diffusion ASSERT_MIN_STEP_TIME=550 AMD=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing
     - name: Run SDXL
-      run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=3200 CAPTURE_PROCESS_REPLAY=0 AMD=1 python3 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
+      run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=3200 CAPTURE_PROCESS_REPLAY=0 AMD=1 python3 examples/sdxl.py --seed 0 --noshow --timing
     - name: Run LLaMA 7B
       run: |
-        BENCHMARK_LOG=llama_nojit AMD=1 JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_unjitted.txt
-        BENCHMARK_LOG=llama AMD=1 JIT=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_jitted.txt
+        BENCHMARK_LOG=llama_nojit AMD=1 JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
+        BENCHMARK_LOG=llama AMD=1 JIT=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
     - name: Run LLaMA 7B with BEAM
-      run: BENCHMARK_LOG=llama_beam AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_beam.txt
+      run: BENCHMARK_LOG=llama_beam AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
     # - name: Run LLaMA 7B on 4 GPUs
-    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_four_gpu.txt
+    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing
     # - name: Run LLaMA 7B on 6 GPUs
-    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_six_gpu.txt
+    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
     - name: Run LLaMA-3 8B BEAM
-      run: BENCHMARK_LOG=llama3_beam AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_beam.txt
+      run: BENCHMARK_LOG=llama3_beam AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
     - name: Run LLaMA-3 8B on 4 GPUs with BEAM
-      run: BENCHMARK_LOG=llama3_beam_4gpu AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_four_gpu.txt
+      run: BENCHMARK_LOG=llama3_beam_4gpu AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
     # - name: Run LLaMA-3 8B on 6 GPUs
-    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 6 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_six_gpu.txt
+    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 6 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
     #- name: Restore amdgpu
     #  run: sudo modprobe amdgpu
     # - name: Run LLaMA-2 70B
-    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 2 --size 70B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_2_70B.txt
+    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 2 --size 70B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
     - name: Run Mixtral 8x7B
-      run: time BENCHMARK_LOG=mixtral AMD=1 python3 examples/mixtral.py --temperature 0 --count 10 --timing | tee mixtral.txt
+      run: time BENCHMARK_LOG=mixtral AMD=1 python3 examples/mixtral.py --temperature 0 --count 10 --timing
     - name: Run GPT2
       run: |
-        BENCHMARK_LOG=gpt2_nojit AMD=1 JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
-        BENCHMARK_LOG=gpt2 AMD=1 JIT=1 ASSERT_MIN_STEP_TIME=5 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
+        BENCHMARK_LOG=gpt2_nojit AMD=1 JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
+        BENCHMARK_LOG=gpt2 AMD=1 JIT=1 ASSERT_MIN_STEP_TIME=5 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
     - name: Run GPT2 w HALF
-      run: BENCHMARK_LOG=gpt2_half AMD=1 HALF=1 ASSERT_MIN_STEP_TIME=5 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
+      run: BENCHMARK_LOG=gpt2_half AMD=1 HALF=1 ASSERT_MIN_STEP_TIME=5 python3 examples/gpt2.py --count 10 --temperature 0 --timing
     - name: Run GPT2 w HALF/BEAM
-      run: BENCHMARK_LOG=gpt2_half_beam AMD=1 HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
-    - uses: actions/upload-artifact@v4
-      with:
-        name: Speed (AMD)
-        path: |
-          onnx_inference_speed.csv
-          torch_speed.txt
-          llama_unjitted.txt
-          llama_jitted.txt
-          llama_beam.txt
-          llama3_beam.txt
-          llama3_four_gpu.txt
-          llama3_six_gpu.txt
-          llama_2_70B.txt
-          gpt2_unjitted.txt
-          gpt2_jitted.txt
-          gpt2_half.txt
-          gpt2_half_beam.txt
-          matmul.txt
-          matmul_amd.txt
-          sd.txt
-          sdxl.txt
-          mixtral.txt
+      run: BENCHMARK_LOG=gpt2_half_beam AMD=1 HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing
     - name: Run process replay tests
       run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py
 
@@ -521,10 +482,12 @@ jobs:
     steps:
     - name: Checkout Code
       uses: actions/checkout@v4
-    - name: Remove amdgpu
-      run: sudo rmmod amdgpu || true
-    - name: Cleanup running AM processes
-      run: python extra/amdpci/am_smi.py --pids --kill
+    - name: Setcap to python
+      run: ./extra/amdpci/setup_python_cap.sh
+    - name: Remove amd modules
+      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd rmmod
+    - name: Kill stale pids
+      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
     - name: Symlink models and datasets
       run: |
         mkdir -p weights
@@ -543,31 +506,23 @@ jobs:
     - name: reset process replay
       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
+      run: time PYTHONPATH=. AMD=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py
     - name: Run 10 CIFAR training steps
-      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=200 AMD=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
+      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=200 AMD=1 STEPS=10 python3 examples/hlb_cifar10.py
     - name: Run 10 CIFAR training steps w HALF
-      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=200 AMD=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_half.txt
+      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=230 AMD=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
     # - 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
+    #   run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=288 AMD=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py
     # 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
+    #   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
     - 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
+      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
     - 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
-    - uses: actions/upload-artifact@v4
-      with:
-        name: Speed (AMD Training)
-        path: |
-          beautiful_mnist.txt
-          train_cifar.txt
-          train_cifar_half.txt
-          train_cifar_bf16.txt
-          train_cifar_wino.txt
-          train_cifar_one_gpu.txt
-          train_cifar_six_gpu.txt
+      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
+    # TODO: broken on some of the machines
+    #- name: Test full tinyfs load
+    #  run: TINYFS_ENDPOINT=10.0.52.11:6767 PYTHONPATH=. python extra/tinyfs/fetch_file.py --hash d734f5e3be9f1e9d863bfaa4fc6c1ef2 --len 175866113 --dest mapping.json --check
     - name: Run process replay tests
       run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py
 
@@ -582,10 +537,12 @@ jobs:
     steps:
     - name: Checkout Code
       uses: actions/checkout@v4
-    - name: Remove amdgpu
-      run: sudo rmmod amdgpu || true
-    - name: Cleanup running AM processes
-      run: python extra/amdpci/am_smi.py --pids --kill
+    - name: Setcap to python
+      run: ./extra/amdpci/setup_python_cap.sh
+    - name: Remove amd modules
+      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd rmmod
+    - name: Kill stale pids
+      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
     - name: Symlink models and datasets
       run: |
         mkdir -p weights
@@ -606,19 +563,12 @@ jobs:
     - name: Run MLPerf resnet eval
       run: time BENCHMARK_LOG=resnet_eval AMD=1 MODEL=resnet python3 examples/mlperf/model_eval.py
     - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-      run: BENCHMARK_LOG=resnet_10steps AMD=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet_one_gpu.txt
+      run: BENCHMARK_LOG=resnet_10steps AMD=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
     - name: Run 10 MLPerf ResNet50 training steps (6 gpu)
-      run: BENCHMARK_LOG=resnet_10steps_6gpu AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet.txt
+      run: BENCHMARK_LOG=resnet_10steps_6gpu AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py
     - name: Run 10 MLPerf Bert training steps (6 gpu)
       # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps_6gpu AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=72 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py | tee train_bert.txt
-    - uses: actions/upload-artifact@v4
-      with:
-        name: Speed (AMD MLPerf)
-        path: |
-          train_resnet.txt
-          train_resnet_one_gpu.txt
-          train_bert.txt
+      run: BENCHMARK_LOG=bert_10steps_6gpu AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=72 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
     - name: Run process replay tests
       run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py
 
@@ -641,19 +591,21 @@ jobs:
     - name: reset process replay
       run: test/external/process_replay/reset.py
     - 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=3 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
     - name: DEBUG=2 IMAGE=1 openpilot compile3 0.10.1 driving_vision
       run: PYTHONPATH="." DEBUG=2 DEV=QCOM FLOAT16=1 IMAGE=1 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: IMAGE=1 openpilot compile3 0.10.1 driving_vision
+      run: BENCHMARK_LOG=image_1_openpilot_0_10_1_vision PYTHONPATH="." DEV=QCOM FLOAT16=1 IMAGE=1 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_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
     - 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=3 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
+      run: BENCHMARK_LOG=openpilot_0_10_1_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/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/dmonitoring_model.onnx
     - name: benchmark MobileNetV2 on DSP
       run: |
         # generate quantized weights
@@ -665,6 +617,27 @@ jobs:
     - name: Run process replay tests
       run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py
 
+  testcommausbgpubenchmark:
+    name: UsbGPU Benchmark (comma)
+    runs-on: [self-hosted, Linux, comma4]
+    timeout-minutes: 20
+    defaults:
+      run:
+        shell: bash -e -o pipefail {0}
+    if: github.repository_owner == 'tinygrad'
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v4
+    - name: setup staging db
+      if: github.ref == 'refs/heads/update_benchmark_staging'
+      run: |
+        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
+        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
+    - name: openpilot compile3 0.10.1 driving_vision
+      run: BENCHMARK_LOG=usbgpu_openpilot_0_10_1_vision PYTHONPATH="." DEV=AMD AMD_LLVM=1 AMD_IFACE=USB ASSERT_MIN_STEP_TIME=50 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
+    - name: openpilot load_pickle 0.10.1 driving_vision
+      run: BENCHMARK_LOG=usbgpu_openpilot_0_10_1_vision_load_pickle PYTHONPATH="." DEV=AMD AMD_IFACE=USB ASSERT_MIN_LOAD_TIME=15 python3 examples/openpilot/load_pickle.py
+
   testreddriverbenchmark:
     name: AM Benchmark
     runs-on: [self-hosted, Linux, tinyboxrandom]
@@ -676,10 +649,12 @@ jobs:
     steps:
     - name: Checkout Code
       uses: actions/checkout@v4
+    - name: Setcap to python
+      run: ./extra/amdpci/setup_python_cap.sh
     - name: Remove amd modules
-      run: ./extra/hcq/hcq_smi.py amd rmmod
+      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd rmmod
     - name: Kill stale pids
-      run: ./extra/hcq/hcq_smi.py amd kill_pids
+      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
     - name: Symlink models and datasets
       run: |
         mkdir -p weights
@@ -708,7 +683,7 @@ jobs:
     #     AMD=1 AMD_LLVM=1 python3 test/test_linearizer.py test/opt/test_tensor_cores.py
     #     AMD=1 SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
     - name: Run Tensor Core GEMM (AMD)
-      run: AMD=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py | tee am_matmul_amd.txt
+      run: AMD=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
     - name: Test AMD=1
       run: DEBUG=2 AMD=1 python -m pytest -rA test/test_tiny.py
     - name: Test DISK copy time
@@ -718,20 +693,12 @@ jobs:
         AMD=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyDefaulttoCPUJit
         AMD=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyCPUtoDefaultJit
     - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee am_train_cifar_one_gpu.txt
+      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
     # - 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
+    #   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
     - name: Run 10 MLPerf Bert training steps (1 gpu)
       # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py | tee am_train_bert_one_gpu.txt
-    - uses: actions/upload-artifact@v4
-      with:
-        name: Speed (AM Driver)
-        path: |
-          am_matmul_amd.txt
-          am_train_cifar_one_gpu.txt
-          am_train_resnet_one_gpu.txt
-          am_train_bert_one_gpu.txt
+      run: BENCHMARK_LOG=bert_10steps AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
     - name: Run process replay tests
       run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py
 
@@ -746,10 +713,12 @@ jobs:
     steps:
     - name: Checkout Code
       uses: actions/checkout@v4
+    - name: Setcap to python
+      run: ./extra/amdpci/setup_python_cap.sh
     - name: Remove nv modules
-      run: ./extra/hcq/hcq_smi.py nv rmmod
+      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py nv rmmod
     - name: Kill stale pids
-      run: ./extra/hcq/hcq_smi.py nv kill_pids
+      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py nv kill_pids
     - name: Symlink models and datasets
       run: |
         mkdir -p weights
@@ -778,21 +747,13 @@ jobs:
         NV=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyDefaulttoCPUJit
         NV=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyCPUtoDefaultJit
     - name: Test LLAMA-3
-      run: BENCHMARK_LOG=llama3_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --benchmark --temperature 0 | tee nv_llama3_beam.txt
+      run: BENCHMARK_LOG=llama3_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --benchmark --temperature 0
     - 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
+      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
     - 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
+      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
     - name: Run 10 MLPerf Bert training steps (1 gpu)
       # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py | tee nv_train_bert_one_gpu.txt
-    - uses: actions/upload-artifact@v4
-      with:
-        name: Speed (NV Driver)
-        path: |
-          nv_llama3_beam.txt
-          nv_train_cifar_one_gpu.txt
-          nv_train_resnet_one_gpu.txt
-          nv_train_bert_one_gpu.txt
+      run: BENCHMARK_LOG=bert_10steps NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
     - name: Run process replay tests
       run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py

.github/workflows/test.yml

@@ -1,10 +1,11 @@
 name: Unit Tests
 env:
   # increment this when downloads substantially change to avoid the internet
-  CACHE_VERSION: '15'
+  CACHE_VERSION: '17'
   CAPTURE_PROCESS_REPLAY: 1
   GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
   PYTHONPATH: ${{ github.workspace }}
+  CHECK_OOB: 1
 
 on:
   push:
@@ -25,17 +26,19 @@ jobs:
       uses: ./.github/actions/setup-tinygrad
       with:
         key: llvm-speed
-        deps: testing_minimal
+        deps: testing_unit
         llvm: 'true'
     - name: Speed Test
-      run: CPU=1 CPU_LLVM=1 python3 test/speed/external_test_speed_v_torch.py
+      run: CPU=1 CPU_LLVM=1 THREADS=0 python3 test/speed/external_test_speed_v_torch.py
     - name: Speed Test (BEAM=2)
-      run: BEAM=2 CPU=1 CPU_LLVM=1 python3 test/speed/external_test_speed_v_torch.py
+      run: BEAM=2 CPU=1 CPU_LLVM=1 THREADS=0 python3 test/speed/external_test_speed_v_torch.py
 
   docs:
     name: Docs
     runs-on: ubuntu-22.04
     timeout-minutes: 10
+    env:
+      CHECK_OOB: 0
     steps:
     - name: Checkout Code
       uses: actions/checkout@v4
@@ -95,27 +98,23 @@ jobs:
       uses: ./.github/actions/setup-tinygrad
       with:
         key: torch-backend-pillow-torchvision-et-pt
-        deps: testing_minimal
+        deps: testing_unit
         pydeps: "pillow torchvision expecttest"
         llvm: 'true'
     - name: Install ninja
       run: |
         sudo apt update || true
         sudo apt install -y --no-install-recommends ninja-build
-    - name: Lint with ruff
-      run: |
-        pip3 install --upgrade --force-reinstall ruff==0.11.0
-        python3 -m ruff check extra/torch_backend/backend.py
     - name: Test one op
-      run: FORWARD_ONLY=1 TINY_BACKEND=1 python3 test/test_ops.py TestOps.test_add
+      run: FORWARD_ONLY=1 TINY_BACKEND=1 python3 test/test_tiny.py TestTiny.test_plus
     - name: Test ResNet-18
       run: DEBUG=2 python3 extra/torch_backend/example.py
-    - name: My (custom) tests
-      run: python3 extra/torch_backend/test.py
+    - name: custom tests
+      run: python3 -m pytest -n auto extra/torch_backend/test.py --durations=20
     - name: Test one op in torch tests
       run: DEBUG=2 python3 extra/torch_backend/torch_tests.py TestTinyBackendPRIVATEUSE1.test_unary_log_tiny_float32
     - name: Test Ops with TINY_BACKEND
-      run: CPU=1 CPU_LLVM=1 LLVMOPT=0 TINY_BACKEND=1 python3 -m pytest -n auto test/test_ops.py --durations=20
+      run: CPU=1 CPU_LLVM=1 LLVMOPT=0 TINY_BACKEND=1 python3 -m pytest -n auto test/backend/test_ops.py --durations=20
     - name: Test in-place operations on views
       run: TORCH_DEBUG=1 python3 extra/torch_backend/test_inplace.py
     - name: Test multi-gpu
@@ -135,7 +134,7 @@ jobs:
       uses: ./.github/actions/setup-tinygrad
       with:
         key: torch-backend-pillow-torchvision-et-pt
-        deps: testing_minimal
+        deps: testing_unit
         llvm: 'true'
     - name: Install ninja
       run: |
@@ -157,27 +156,27 @@ jobs:
       uses: ./.github/actions/setup-tinygrad
       with:
         key: be-minimal
-        deps: testing_minimal
+        deps: testing_unit
     - name: Test dtype with Python emulator
-      run: DEBUG=1 PYTHON=1 python3 -m pytest -n=auto test/test_dtype.py test/test_dtype_alu.py
+      run: DEBUG=1 PYTHON=1 python3 -m pytest -n=auto test/backend/test_dtype.py test/backend/test_dtype_alu.py
     - name: Test ops with Python emulator
-      run: DEBUG=2 SKIP_SLOW_TEST=1 PYTHON=1 python3 -m pytest -n=auto test/test_ops.py --durations=20
+      run: DEBUG=2 SKIP_SLOW_TEST=1 PYTHON=1 python3 -m pytest -n=auto test/backend/test_ops.py --durations=20
     - name: Test uops with Python emulator
-      run: PYTHON=1 python3 -m pytest test/test_uops.py --durations=20
+      run: PYTHON=1 python3 -m pytest test/backend/test_uops.py --durations=20
     - name: Test symbolic with Python emulator
-      run: PYTHON=1 python3 test/test_symbolic_ops.py
+      run: PYTHON=1 python3 test/backend/test_symbolic_ops.py
     - name: test_renderer_failures with Python emulator
-      run: PYTHON=1 python3 -m pytest -rA test/test_renderer_failures.py::TestRendererFailures
+      run: PYTHON=1 python3 -m pytest -rA test/backend/test_renderer_failures.py::TestRendererFailures
     - name: Test IMAGE=2 support
       run: |
-        IMAGE=2 PYTHON=1 python3 test/test_ops.py TestOps.test_gemm
-        IMAGE=2 PYTHON=1 python3 test/test_ops.py TestOps.test_simple_conv2d
+        IMAGE=2 PYTHON=1 python3 test/backend/test_ops.py TestOps.test_gemm
+        IMAGE=2 PYTHON=1 python3 test/backend/test_ops.py TestOps.test_simple_conv2d
     - name: Test emulated METAL tensor cores
       run: |
-        DEBUG=2 EMULATE=METAL FORWARD_ONLY=1 PYTHON=1 python3 test/test_ops.py TestOps.test_big_gemm
+        DEBUG=2 EMULATE=METAL FORWARD_ONLY=1 PYTHON=1 python3 test/backend/test_ops.py TestOps.test_big_gemm
         DEBUG=2 EMULATE=METAL FORWARD_ONLY=1 PYTHON=1 python3 test/opt/test_tensor_cores.py
     - name: Test emulated AMX tensor cores
-      run: DEBUG=2 AMX=1 EMULATE=AMX FORWARD_ONLY=1 PYTHON=1 python3 test/test_ops.py TestOps.test_gemm
+      run: DEBUG=2 AMX=1 EMULATE=AMX FORWARD_ONLY=1 PYTHON=1 python3 test/backend/test_ops.py TestOps.test_gemm
     - name: Test emulated AMD tensor cores
       run: |
         DEBUG=2 EMULATE=AMD FORWARD_ONLY=1 PYTHON=1 N=16 HALF=1 ACC_HALF=0 python3 ./extra/gemm/simple_matmul.py
@@ -198,9 +197,9 @@ jobs:
         DEBUG=2 EMULATE=AMD_RDNA4 FORWARD_ONLY=1 PYTHON=1 python3 test/opt/test_tensor_cores.py
     - name: Test emulated CUDA tensor cores
       run: |
-        DEBUG=2 EMULATE=CUDA FORWARD_ONLY=1 PYTHON=1 python3 test/test_ops.py TestOps.test_gemm_fp16
-        DEBUG=2 EMULATE=CUDA ALLOW_TF32=1 FORWARD_ONLY=1 PYTHON=1 python3 test/test_ops.py TestOps.test_gemm
-        DEBUG=2 EMULATE=CUDA_SM75 FORWARD_ONLY=1 PYTHON=1 python3 test/test_ops.py TestOps.test_gemm_fp16
+        DEBUG=2 EMULATE=CUDA FORWARD_ONLY=1 PYTHON=1 python3 test/backend/test_ops.py TestOps.test_gemm_fp16
+        DEBUG=2 EMULATE=CUDA ALLOW_TF32=1 FORWARD_ONLY=1 PYTHON=1 python3 test/backend/test_ops.py TestOps.test_gemm
+        DEBUG=2 EMULATE=CUDA_SM75 FORWARD_ONLY=1 PYTHON=1 python3 test/backend/test_ops.py TestOps.test_gemm_fp16
         DEBUG=2 EMULATE=CUDA_SM89 ALLOW_TF32=1 FORWARD_ONLY=1 PYTHON=1 python3 test/opt/test_tensor_cores.py
     - name: Test emulated INTEL OpenCL tensor cores
       run: DEBUG=2 EMULATE=INTEL FORWARD_ONLY=1 PYTHON=1 HALF=1 N=64 python3 ./extra/gemm/simple_matmul.py
@@ -208,18 +207,17 @@ jobs:
       run: DEBUG=2 AMX=1 EMULATE=AMX FORWARD_ONLY=1 PYTHON=1 python3 test/opt/test_tensor_cores.py
     - name: Test device flop counts
       run: |
-        DEBUG=2 EMULATE=METAL PYTHON=1 python3 ./test/test_uops_stats.py TestUOpsStatsMatmulHalf
-        DEBUG=2 EMULATE=AMD PYTHON=1 python3 ./test/test_uops_stats.py TestUOpsStatsMatmulHalf
-        DEBUG=2 EMULATE=CUDA PYTHON=1 python3 ./test/test_uops_stats.py TestUOpsStatsMatmulHalf
-        DEBUG=2 EMULATE=INTEL PYTHON=1 python3 ./test/test_uops_stats.py TestUOpsStatsMatmulHalf
-        DEBUG=2 AMX=1 EMULATE=AMX PYTHON=1 python3 ./test/test_uops_stats.py TestUOpsStats.test_simple_matmul
+        DEBUG=2 EMULATE=METAL PYTHON=1 python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
+        DEBUG=2 EMULATE=AMD PYTHON=1 python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
+        DEBUG=2 EMULATE=CUDA PYTHON=1 python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
+        DEBUG=2 EMULATE=INTEL PYTHON=1 python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
+        DEBUG=2 AMX=1 EMULATE=AMX PYTHON=1 python3 ./test/null/test_uops_stats.py TestUOpsStats.test_simple_matmul
 
   linter:
     name: Linters
     runs-on: ubuntu-latest
     timeout-minutes: 10
 
-    # TODO: run the pre-commit hook to replace a lot of this
     steps:
     - name: Checkout Code
       uses: actions/checkout@v4
@@ -231,18 +229,51 @@ jobs:
         deps: linting
     - name: Lint bad-indentation and trailing-whitespace with pylint
       run: python -m pylint --disable=all -e W0311 -e C0303 --jobs=0 --indent-string='  ' --recursive=y .
-    - name: Lint with ruff
+    - name: Run pre-commit linting hooks
+      run: SKIP=tiny,tests,example pre-commit run --all-files
+    - name: Lint additional files with ruff
       run: |
-        pip3 install --upgrade --force-reinstall ruff==0.11.0
-        python3 -m ruff check .
         python3 -m ruff check examples/mlperf/ --ignore E501
         python3 -m ruff check extra/thunder/tiny/ --ignore E501 --ignore F841 --ignore E722
-    - name: Run mypy
+        python3 -m ruff check extra/torch_backend/backend.py
+    - name: Run mypy with lineprecision report
       run: |
-        python -m mypy --strict-equality --lineprecision-report .
+        python -m mypy --lineprecision-report .
+        grep -v autogen lineprecision.txt | awk 'NR>2 {lines+=$2; precise+=$3; imprecise+=$4; any+=$5; empty+=$6} END {t=lines-empty; printf "TOTAL: %d lines, %d precise (%.1f%%), %d imprecise (%.1f%%), %d any (%.1f%%)\n", t, precise, 100*precise/t, imprecise, 100*imprecise/t, any, 100*any/t}'
         cat lineprecision.txt
     - name: Run TYPED=1
-      run: TYPED=1 python -c "import tinygrad"
+      run: CHECK_OOB=0 DEV=CPU TYPED=1 python test/test_tiny.py
+
+  nulltest:
+    name: Null Tests
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v4
+    - name: Setup Environment
+      uses: ./.github/actions/setup-tinygrad
+      with:
+        key: unittest-13
+        pydeps: "pillow ftfy regex pre-commit"
+        deps: testing_unit
+        llvm: 'true'
+        amd: 'true'
+    - name: Run NULL backend tests
+      run: NULL=1 python -m pytest -n=auto test/null/ --durations=20
+    - name: Run targetted tests on NULL backend
+      run: NULL=1 python3 -m unittest test.backend.test_multitensor.TestMultiTensor.test_data_parallel_resnet_train_step
+    # TODO: too slow
+    # - name: Run SDXL on NULL backend
+    #   run: NULL=1 DEBUG=1 python3 examples/sdxl.py --seed 0 --noshow --timing --fakeweights
+    - name: Run Clip tests for SD MLPerf on NULL backend
+      run: NULL=1 python -m pytest -n=auto test/external/mlperf_stable_diffusion/external_test_models.py::TestOpenClip --durations=20
+    - name: Run AMD emulated BERT training on NULL backend
+      run: EMULATE=AMD_RDNA4 NULL=1 NULL_ALLOW_COPYOUT=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
+    # TODO: support fake weights
+    #- name: Run LLaMA 7B on 4 fake devices
+    #  run: NULL=1 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 3 --temperature 0 --timing
 
   unittest:
     name: Unit Tests
@@ -255,27 +286,19 @@ jobs:
     - name: Setup Environment
       uses: ./.github/actions/setup-tinygrad
       with:
-        key: unittest-12
-        pydeps: "pillow numpy ftfy regex"
+        key: unittest-13
+        pydeps: "pillow ftfy regex pre-commit"
         deps: testing_unit
+        llvm: 'true'
+        amd: 'true'
+    - name: Run pre-commit test hooks
+      run: SKIP=ruff,mypy pre-commit run --all-files
     - name: Check Device.DEFAULT
       run: python -c "from tinygrad import Device; assert Device.DEFAULT == 'CPU', Device.DEFAULT"
     - name: Run unit tests
       run: |
-        CPU=1 python test/unit/test_device.py TestRunAsModule.test_module_runs
-        CPU=1 python -m pytest -n=auto test/unit/ --durations=20 --deselect=test/unit/test_device.py::TestRunAsModule::test_module_runs
-    - name: Run targetted tests on NULL backend
-      run: NULL=1 python3 -m unittest test.test_multitensor.TestMultiTensor.test_data_parallel_resnet_train_step test/device/test_null.py
-    # TODO: too slow
-    # - name: Run SDXL on NULL backend
-    #   run: NULL=1 DEBUG=1 python3 examples/sdxl.py --seed 0 --noshow --timing --fakeweights
-    - name: Run Clip tests for SD MLPerf on NULL backend
-      run: NULL=1 python -m pytest -n=auto test/external/mlperf_stable_diffusion/external_test_models.py::TestOpenClip --durations=20
-    - name: Run AMD emulated BERT training on NULL backend
-      run: EMULATE=AMD_RDNA4 NULL=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
-    # TODO: support fake weights
-    #- name: Run LLaMA 7B on 4 fake devices
-    #  run: NULL=1 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 3 --temperature 0 --timing
+        CPU=1 python test/null/test_device.py TestRunAsModule.test_module_runs
+        CPU=1 python -m pytest -n=auto test/unit/ --durations=20
     - name: Run GC tests
       run: python test/external/external_uop_gc.py
     - name: External Benchmark Schedule
@@ -289,8 +312,8 @@ 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 < 20000 lines
-      run: MAX_LINE_COUNT=20000 python sz.py
+    - name: Repo line count < 24000 lines
+      run: MAX_LINE_COUNT=24000 python sz.py
 
   spec:
     strategy:
@@ -310,7 +333,7 @@ jobs:
         deps: testing_unit
         python-version: '3.14'
     - name: Test SPEC=2
-      run: IGNORE_OOB=0 SPEC=2 PYTHONPATH="." pytest --maxfail=10 -n auto --durations=30 --ignore=test/models --ignore test/test_custom_kernel.py --ignore test/unit/test_hashing.py --timeout 60 -k "not test_setitem_big" --splits 2 --group ${{ matrix.group }}
+      run: SPEC=2 pytest --maxfail=10 -n auto --durations=30 test/unit test/backend test/opt --ignore test/backend/test_custom_kernel.py --ignore test/unit/test_hashing.py --timeout 60 -k "not test_setitem_big" --splits 2 --group ${{ matrix.group }}
 
   fuzzing:
     name: Fuzzing
@@ -344,11 +367,11 @@ jobs:
         uses: ./.github/actions/setup-tinygrad
         with:
           key: gpu-image
-          deps: testing_minimal
+          deps: testing_unit
           opencl: 'true'
       - name: Test CL IMAGE=2 ops
         run: |
-          CL=1 IMAGE=2 python -m pytest -n=auto test/test_ops.py --durations=20
+          CL=1 IMAGE=2 python -m pytest -n=auto test/backend/test_ops.py --durations=20
         # TODO: training is broken
         # CL=1 IMAGE=2 python test/models/test_end2end.py TestEnd2End.test_linear_mnist
       - name: Run process replay tests
@@ -365,14 +388,14 @@ jobs:
         uses: ./.github/actions/setup-tinygrad
         with:
           key: gen-dataset
-          deps: testing_minimal
+          deps: testing
           opencl: 'true'
       - name: Generate Dataset
         run: CL=1 extra/optimization/generate_dataset.sh
       - name: Run Kernel Count Test
         run: CL=1 python -m pytest -n=auto test/external/external_test_opt.py
       - name: Run fused optimizer tests
-        run: CL=1 FUSE_OPTIM=1 python -m pytest -n=auto test/models/test_mnist.py test/test_optim.py -k "not muon"
+        run: CL=1 FUSE_OPTIM=1 python -m pytest -n=auto test/models/test_mnist.py test/backend/test_optim.py -k "not muon"
       - name: Upload artifact
         uses: actions/upload-artifact@v4
         with:
@@ -420,7 +443,7 @@ jobs:
         with:
           key: onnxoptc
           deps: testing
-          python-version: '3.11'
+          python-version: '3.12'
           llvm: 'true'
       - name: Test ONNX (CPU)
         run: CPU=1 CPU_LLVM=0 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
@@ -431,7 +454,7 @@ jobs:
       - name: Test Additional ONNX Ops (CPU)
         run: CPU=1 CPU_LLVM=0 python3 test/external/external_test_onnx_ops.py
       - name: Test Quantize ONNX
-        run: CPU=1 CPU_LLVM=0 python3 test/test_quantize_onnx.py
+        run: CPU=1 CPU_LLVM=0 python3 test/backend/test_quantize_onnx.py
       - name: Run process replay tests
         uses: ./.github/actions/process-replay
 
@@ -448,7 +471,7 @@ jobs:
           key: onnxoptl
           deps: testing
           pydeps: "tensorflow==2.19"
-          python-version: '3.11'
+          python-version: '3.12'
           opencl: 'true'
       - name: Test ONNX (CL)
         run: CL=1 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
@@ -461,11 +484,11 @@ jobs:
       - name: Test MLPerf stuff
         run: CL=1 python -m pytest -n=auto test/external/external_test_optim.py test/external/external_test_losses.py test/external/external_test_metrics.py test/external/external_test_datasets.py --durations=20
       - name: NULL=1 beautiful_mnist_multigpu
-        run: NULL=1 python examples/beautiful_mnist_multigpu.py
+        run: NULL=1 NULL_ALLOW_COPYOUT=1 python examples/beautiful_mnist_multigpu.py
       - name: Test Bert training
-        run: NULL=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=24 GPUS=4 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
+        run: NULL=1 NULL_ALLOW_COPYOUT=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=24 GPUS=4 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
       - name: Test llama 3 training
-        run: NULL=1 SAMPLES=300 BS=8 SEQLEN=512 GRADIENT_ACC_STEPS=1 FAKEDATA=1 DEFAULT_FLOAT=bfloat16 OPTIM_DTYPE=bfloat16 LLAMA3_SIZE=1B MODEL=llama3 python3 examples/mlperf/model_train.py
+        run: NULL=1 NULL_ALLOW_COPYOUT=1 SAMPLES=300 BS=8 SEQLEN=512 GRADIENT_ACC_STEPS=1 FAKEDATA=1 DEFAULT_FLOAT=bfloat16 OPTIM_DTYPE=bfloat16 LLAMA3_SIZE=1B MODEL=llama3 python3 examples/mlperf/model_train.py
       - name: Run process replay tests
         uses: ./.github/actions/process-replay
 
@@ -473,6 +496,8 @@ jobs:
     name: Test LLM
     runs-on: ubuntu-24.04
     timeout-minutes: 15
+    env:
+      CHECK_OOB: 0
     steps:
       - name: Checkout Code
         uses: actions/checkout@v4
@@ -520,7 +545,7 @@ jobs:
       with:
         key: metal
         deps: testing
-        python-version: '3.11'
+        python-version: '3.12'
     - name: Test models (Metal)
       run: METAL=1 python -m pytest -n=auto test/models --durations=20
     - name: Test LLaMA compile speed
@@ -539,15 +564,15 @@ jobs:
       uses: ./.github/actions/setup-tinygrad
       with:
         key: devectorize-minimal
-        deps: testing_minimal
+        deps: testing_unit
         pydeps: "pillow"
         llvm: "true"
     - name: Test LLVM=1 DEVECTORIZE=0
-      run: CPU=1 CPU_LLVM=1 DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/test_ops.py
+      run: CPU=1 CPU_LLVM=1 DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/backend/test_ops.py
     - name: Test LLVM=1 DEVECTORIZE=0 for model
       run: CPU=1 CPU_LLVM=1 DEVECTORIZE=0 python3 test/models/test_efficientnet.py
     - name: Test CPU=1 DEVECTORIZE=0
-      run: CPU=1 CPU_LLVM=0 DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/test_ops.py
+      run: CPU=1 CPU_LLVM=0 DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/backend/test_ops.py
 
   testdsp:
     name: Linux (DSP)
@@ -560,8 +585,8 @@ jobs:
       uses: ./.github/actions/setup-tinygrad
       with:
         key: dsp-minimal
-        deps: testing_minimal
-        pydeps: "onnx==1.18.0 onnxruntime pillow"
+        deps: testing_unit
+        pydeps: "onnx==1.18.0 onnxruntime ml_dtypes"
         llvm: "true"
     - name: Set up Docker Buildx
       uses: docker/setup-buildx-action@v3
@@ -573,15 +598,15 @@ jobs:
         load: true
         tags: qemu-hexagon:latest
         cache-from: type=gha
-        cache-to: type=gha,mode=min
+        cache-to: ${{ github.event_name != 'pull_request' && 'type=gha,mode=min' || '' }}
     - name: Set MOCKDSP env
       run: printf "MOCKDSP=1" >> $GITHUB_ENV
     - name: Run test_tiny on DSP
       run: DEBUG=2 DSP=1 python test/test_tiny.py
     - name: Test transcendentals
-      run: CC=clang-20 DEBUG=2 DSP=1 python test/test_transcendental.py TestTranscendentalVectorized
+      run: CC=clang-20 DEBUG=2 DSP=1 python test/backend/test_transcendental.py TestTranscendentalVectorized
     - name: Test quantize onnx
-      run: DEBUG=2 DSP=1 python3 test/test_quantize_onnx.py
+      run: DEBUG=2 DSP=1 python3 test/backend/test_quantize_onnx.py
 
   testwebgpu:
     name: Linux (WebGPU)
@@ -594,34 +619,106 @@ jobs:
       uses: ./.github/actions/setup-tinygrad
       with:
         key: webgpu-minimal
-        deps: testing_minimal
-        python-version: '3.11'
+        deps: testing_unit
+        python-version: '3.12'
         webgpu: 'true'
     - name: Check Device.DEFAULT (WEBGPU) and print some source
       run: |
         WEBGPU=1 python -c "from tinygrad import Device; assert Device.DEFAULT == 'WEBGPU', Device.DEFAULT"
-        WEBGPU=1 DEBUG=4 FORWARD_ONLY=1 python3 test/test_ops.py TestOps.test_add
+        WEBGPU=1 DEBUG=4 FORWARD_ONLY=1 python3 test/test_tiny.py TestTiny.test_plus
     - name: Run selected webgpu tests
       run: |
-          WEBGPU=1 WEBGPU_BACKEND="WGPUBackendType_Vulkan" python3 -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit \
-          --ignore=test/test_copy_speed.py --ignore=test/test_rearrange_einops.py \
-          --ignore=test/test_fuzz_shape_ops.py --durations=20
+          WEBGPU=1 WEBGPU_BACKEND="WGPUBackendType_Vulkan" python3 -m pytest -n=auto test/backend --durations=20
     - name: Run process replay tests
       uses: ./.github/actions/process-replay
 
+  testamdasm:
+    name: AMD ASM IDE
+    runs-on: ubuntu-24.04
+    timeout-minutes: 20
+    env:
+      AMD: 1
+      PYTHON_REMU: 1
+      MOCKGPU: 1
+    steps:
+      - name: Checkout Code
+        uses: actions/checkout@v4
+      - name: Setup Environment
+        uses: ./.github/actions/setup-tinygrad
+        with:
+          key: rdna3-emu
+          deps: testing_unit
+          amd: 'true'
+          python-version: '3.14'
+      - name: Verify AMD autogen is up to date
+        run: |
+          python -m tinygrad.renderer.amd.generate
+          git diff --exit-code tinygrad/runtime/autogen/amd/
+      - name: Install LLVM 21
+        run: |
+          wget -qO- https://apt.llvm.org/llvm-snapshot.gpg.key | sudo tee /etc/apt/trusted.gpg.d/apt.llvm.org.asc
+          echo "deb http://apt.llvm.org/$(lsb_release -cs)/ llvm-toolchain-$(lsb_release -cs)-21 main" | sudo tee /etc/apt/sources.list.d/llvm.list
+          sudo apt-get update
+          sudo apt-get install llvm-21 llvm-21-tools cloc
+      - name: Install rocprof-trace-decoder
+        run: sudo PYTHONPATH="." ./extra/sqtt/install_rocprof_decoder.py
+      - name: Run AMD renderer tests
+        run: AMD_LLVM=0 python -m pytest -n=auto test/amd/ --durations 20
+      - name: Run AMD renderer tests (AMD_LLVM=1)
+        run: AMD_LLVM=1 python -m pytest -n=auto test/amd/ --durations 20
+      - name: Run SQTT profiling tests
+        run: PROFILE=1 SQTT=1 python3 -m pytest -n=auto test/amd/test_sqtt_profiler.py
+      - name: Run AMD emulated tests on NULL backend
+        env:
+          AMD: 0
+        run: |
+          PYTHONPATH=. NULL=1 EMULATE=AMD python extra/mmapeak/mmapeak.py
+          PYTHONPATH=. NULL=1 EMULATE=AMD_CDNA4 python3 -m pytest -n=auto test/testextra/test_tk.py test/backend/test_asm_gemm.py
+      - name: Run ASM matmul on MOCKGPU
+        run: PYTHONPATH="." AMD=1 MOCKGPU=1 N=256 python3 extra/gemm/amd_asm_matmul.py
+      - name: Run LLVM test
+        run: AMD_LLVM=1 python test/device/test_amd_llvm.py
+
+  testmockam:
+    name: Linux (am)
+    runs-on: ubuntu-24.04
+    timeout-minutes: 15
+    env:
+      AMD: 1
+      MOCKGPU: 1
+      AMD_IFACE: PCI
+    steps:
+      - name: Checkout Code
+        uses: actions/checkout@v4
+      - name: Setup Environment
+        uses: ./.github/actions/setup-tinygrad
+        with:
+          key: mockam
+          deps: testing_unit
+          amd: 'true'
+      - name: Run test_tiny on MOCKAM
+        run: python test/test_tiny.py
+      - name: Run test_tiny on MOCKAM USB
+        run: AMD_IFACE=USB python test/test_tiny.py
+      - name: Run test_hcq on MOCKAM
+        run: python -m pytest test/device/test_hcq.py
+
   testamd:
     strategy:
       fail-fast: false
       matrix:
         backend: [amd, amdllvm]
+        arch: [rdna3, rdna4]
+        #arch: [rdna3, rdna4, cdna4]
 
-    name: Linux (${{ matrix.backend }})
+    name: Linux (${{ matrix.backend }} ${{ matrix.arch }})
     runs-on: ubuntu-22.04
-    timeout-minutes: 20
+    timeout-minutes: 15
     env:
       AMD: 1
       MOCKGPU: 1
-      FORWARD_ONLY: 1
+      MOCKGPU_ARCH: ${{ matrix.arch }}
+      SKIP_SLOW_TEST: 1
       AMD_LLVM: ${{ matrix.backend == 'amdllvm' && '1' || matrix.backend != 'amdllvm' && '0' }}
     steps:
       - name: Checkout Code
@@ -630,61 +727,20 @@ jobs:
         uses: ./.github/actions/setup-tinygrad
         with:
           key: ${{ matrix.backend }}-minimal
-          deps: testing_minimal
+          deps: testing_unit
           amd: 'true'
           llvm: ${{ matrix.backend == 'amdllvm' && 'true' }}
       - name: Check Device.DEFAULT and print some source
         run: |
           python3 -c "from tinygrad import Device; assert Device.DEFAULT in ['AMD'], Device.DEFAULT"
-          DEBUG=5 FORWARD_ONLY=1 python3 test/test_ops.py TestOps.test_add
-      - name: Run LLVM test
-        if: matrix.backend=='amdllvm'
-        run: python test/device/test_amd_llvm.py
+          DEBUG=5 FORWARD_ONLY=1 python3 test/test_tiny.py TestTiny.test_plus
       - name: Run pytest (amd)
-        run: python -m pytest -n=auto test/test_ops.py test/test_dtype.py test/test_dtype_alu.py test/test_linearizer.py test/test_randomness.py test/test_jit.py test/test_graph.py test/test_multitensor.py test/device/test_hcq.py test/testextra/test_cfg_viz.py --durations=20
-      - name: Run pytest (amd)
-        run: python -m pytest test/external/external_test_am.py --durations=20
+        run: python -m pytest -n=auto test/backend/test_ops.py test/backend/test_dtype.py test/backend/test_dtype_alu.py test/backend/test_linearizer.py test/backend/test_randomness.py test/backend/test_jit.py test/backend/test_graph.py test/backend/test_multitensor.py test/device/test_hcq.py test/testextra/test_cfg_viz.py test/external/external_test_am.py --durations=20
       - name: Run TRANSCENDENTAL math
-        run: TRANSCENDENTAL=2 python -m pytest -n=auto test/test_ops.py::TestOps::test_sin test/test_ops.py::TestOps::test_cos test/test_ops.py::TestOps::test_tan test/test_ops.py::TestOps::test_exp test/test_ops.py::TestOps::test_log --durations=20
-      - name: Run TestOps.test_add with SQTT
-        run: |
-          VIZ=1 PMC=1 DEBUG=5 python3 test/test_ops.py TestOps.test_add
-          VIZ=1 SQTT=1 DEBUG=5 python3 test/test_ops.py TestOps.test_add
-          extra/sqtt/rgptool.py create "/tmp/profile.pkl.$USER" -o /tmp/gpu0.rgp
+        run: TRANSCENDENTAL=2 python -m pytest -n=auto test/backend/test_ops.py::TestOps::test_sin test/backend/test_ops.py::TestOps::test_cos test/backend/test_ops.py::TestOps::test_tan test/backend/test_ops.py::TestOps::test_exp test/backend/test_ops.py::TestOps::test_log --durations=20
       - name: Run process replay tests
         uses: ./.github/actions/process-replay
 
-  testrdna3:
-    name: AMD ASM IDE
-    runs-on: ubuntu-24.04
-    timeout-minutes: 10
-    steps:
-      - name: Checkout Code
-        uses: actions/checkout@v4
-      - name: Setup Environment
-        uses: ./.github/actions/setup-tinygrad
-        with:
-          key: rdna3-emu
-          deps: testing_minimal
-          amd: 'true'
-      - name: Install LLVM 21
-        run: |
-          wget -qO- https://apt.llvm.org/llvm-snapshot.gpg.key | sudo tee /etc/apt/trusted.gpg.d/apt.llvm.org.asc
-          echo "deb http://apt.llvm.org/$(lsb_release -cs)/ llvm-toolchain-$(lsb_release -cs)-21 main" | sudo tee /etc/apt/sources.list.d/llvm.list
-          sudo apt-get update
-          sudo apt-get install llvm-21 llvm-21-tools cloc
-      - name: RDNA3 Line Count
-        run: cloc --by-file extra/assembly/amd/*.py
-      - name: Run RDNA3 emulator tests
-        run: python -m pytest -n=auto extra/assembly/amd/ --durations 20
-      - name: Install pdfplumber
-        run: pip install pdfplumber
-      - name: Verify AMD autogen is up to date
-        run: |
-          python -m extra.assembly.amd.dsl --arch all
-          python -m extra.assembly.amd.pcode --arch all
-          git diff --exit-code extra/assembly/amd/autogen/
-
   testnvidia:
     strategy:
       fail-fast: false
@@ -704,7 +760,7 @@ jobs:
         uses: ./.github/actions/setup-tinygrad
         with:
           key: ${{ matrix.backend }}-minimal
-          deps: testing_minimal
+          deps: testing_unit
           cuda: 'true'
           ocelot: 'true'
       - name: Set env
@@ -712,10 +768,12 @@ jobs:
       - name: Check Device.DEFAULT and print some source
         run: |
           python3 -c "from tinygrad import Device; assert Device.DEFAULT in ['CUDA','NV'], Device.DEFAULT"
-          DEBUG=5 FORWARD_ONLY=1 python3 test/test_ops.py TestOps.test_add
+          DEBUG=5 FORWARD_ONLY=1 python3 test/test_tiny.py TestTiny.test_plus
       - name: Run pytest (cuda)
         # skip multitensor because it's slow
-        run: python -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit --ignore test/test_gc.py --ignore test/test_multitensor.py --durations=20
+        run: python -m pytest -n=auto test/backend --ignore test/backend/test_multitensor.py --durations=20
+      - name: Run TestOps.test_add with PMA
+        run: VIZ=-1 PMA=1 DEBUG=5 python3 test/backend/test_ops.py TestOps.test_add
       - name: Run process replay tests
         uses: ./.github/actions/process-replay
 
@@ -735,7 +793,7 @@ jobs:
         uses: ./.github/actions/setup-tinygrad
         with:
           key: ${{ matrix.backend }}-minimal
-          deps: testing_minimal
+          deps: testing_unit
           opencl: ${{ matrix.backend == 'opencl' && 'true' }}
           llvm: ${{ matrix.backend == 'llvm' || matrix.backend == 'lvp' }}
           mesa: ${{ matrix.backend == 'lvp' && 'true' }}
@@ -744,11 +802,11 @@ jobs:
       - name: Check Device.DEFAULT and print some source
         run: |
           python3 -c "from tinygrad import Device; assert Device.DEFAULT in ['CPU','CL'], Device.DEFAULT"
-          DEBUG=5 FORWARD_ONLY=1 python3 test/test_ops.py TestOps.test_add
+          DEBUG=5 FORWARD_ONLY=1 python3 test/test_tiny.py TestTiny.test_plus
       - name: Run pytest (${{ matrix.backend }})
-        run: python -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit --durations=20
+        run: python -m pytest -n=auto test/backend --durations=20
       - name: Run TRANSCENDENTAL math
-        run: TRANSCENDENTAL=2 python -m pytest -n=auto test/test_ops.py::TestOps::test_sin test/test_ops.py::TestOps::test_cos test/test_ops.py::TestOps::test_tan test/test_ops.py::TestOps::test_exp test/test_ops.py::TestOps::test_log --durations=20
+        run: TRANSCENDENTAL=2 python -m pytest -n=auto test/backend/test_ops.py::TestOps::test_sin test/backend/test_ops.py::TestOps::test_cos test/backend/test_ops.py::TestOps::test_tan test/backend/test_ops.py::TestOps::test_exp test/backend/test_ops.py::TestOps::test_log --durations=20
       - name: Run process replay tests
         uses: ./.github/actions/process-replay
 
@@ -766,25 +824,29 @@ jobs:
       with:
         key: metal
         deps: testing
-        python-version: '3.11'
+        python-version: '3.12'
         amd: 'true'
         cuda: 'true'
         ocelot: 'true'
         llvm: 'true'
     - name: Run unit tests
       run: METAL=1 python -m pytest -n=auto test/unit/ --durations=20
+    - name: Run NULL backend tests
+      run: NULL=1 python -m pytest -n=auto test/null/ --durations=20
     - name: Run ONNX
       run: METAL=1 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
     - name: Test tensor core ops (fake)
-      run: METAL=1 DEBUG=3 TC=2 python test/test_ops.py TestOps.test_gemm
+      run: METAL=1 DEBUG=3 TC=2 python test/backend/test_ops.py TestOps.test_gemm
     - name: Test tensor core ops (real)
-      run: METAL=1 DEBUG=3 python test/test_ops.py TestOps.test_big_gemm
+      run: METAL=1 DEBUG=3 python test/backend/test_ops.py TestOps.test_big_gemm
     - name: Test Beam Search
       run: METAL=1 IGNORE_BEAM_CACHE=1 python3 -m pytest extra/optimization/test_beam_search.py
+    - name: Test Device Specific
+      run: METAL=1 python3 -m pytest test/device/test_metal.py
     #- name: Fuzz Test linearizer
     #  run: METAL=1 DEPTH=4 FUZZ_N=50 FUZZ_MAX_SIZE=1000000 python test/external/fuzz_linearizer.py
     - name: Run TRANSCENDENTAL math
-      run: METAL=1 TRANSCENDENTAL=2 python -m pytest -n=auto test/test_ops.py::TestOps::test_sin test/test_ops.py::TestOps::test_cos test/test_ops.py::TestOps::test_tan test/test_ops.py::TestOps::test_exp test/test_ops.py::TestOps::test_log --durations=20
+      run: METAL=1 TRANSCENDENTAL=2 python -m pytest -n=auto test/backend/test_ops.py::TestOps::test_sin test/backend/test_ops.py::TestOps::test_cos test/backend/test_ops.py::TestOps::test_tan test/backend/test_ops.py::TestOps::test_exp test/backend/test_ops.py::TestOps::test_log --durations=20
     - name: Run pytest (amd)
       env:
         MOCKGPU: 1
@@ -807,6 +869,8 @@ jobs:
         NV_PTX: 1
         NV: 1
         FORWARD_ONLY: 1
+        # TODO: failing due to library loading error
+        CAPTURE_PROCESS_REPLAY: 0
       run: |
         python3 -m pytest -n=auto test/device/test_hcq.py test/test_tiny.py --durations=20
     - name: Run process replay tests
@@ -825,14 +889,14 @@ jobs:
         key: osx-webgpu
         deps: testing
         webgpu: 'true'
-    - name: Test infinity math in WGSL
-      run: WEBGPU=1 python -m pytest -n=auto test/test_renderer_failures.py::TestWGSLFailures::test_multiply_infinity --durations=20
     - name: Build WEBGPU Efficientnet
       run: WEBGPU=1 WEBGPU_BACKEND="WGPUBackendType_Metal" python3 -m examples.compile_efficientnet
-    - name: Clean npm cache
-      run: npm cache clean --force
-    - name: Install Puppeteer
-      run: npm install puppeteer
+    - name: Run selected webgpu tests
+      run: WEBGPU=1 WEBGPU_BACKEND="WGPUBackendType_Metal" python3 -m pytest -n=auto test/backend --durations=20
+    #- name: Clean npm cache
+    #  run: npm cache clean --force
+    #- name: Install Puppeteer
+    #  run: npm install puppeteer
     # this is also flaky
     #- name: Run WEBGPU Efficientnet
     #  run: node test/web/test_webgpu.js
@@ -864,8 +928,7 @@ jobs:
         uses: ./.github/actions/setup-tinygrad
         with:
           key: macos-${{ matrix.backend }}-minimal
-          deps: testing_minimal
-          pydeps: "capstone"
+          deps: testing_unit
           llvm: ${{ matrix.backend == 'llvm' || matrix.backend == 'lvp' }}
           mesa: ${{ matrix.backend == 'lvp' && 'true' }}
       - name: Set env
@@ -875,7 +938,7 @@ jobs:
           python -c "from tinygrad import Device; assert Device.DEFAULT == {'LLVM':'CPU','LVP':'CPU'}.get(x:='${{ matrix.backend }}'.upper(), x), Device.DEFAULT"
           DEBUG=4 python3 test/test_tiny.py TestTiny.test_plus
       - name: Run pytest (${{ matrix.backend }})
-        run: python3 -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit --durations=20
+        run: python3 -m pytest -n=auto test/backend --durations=20
       - name: Run process replay tests
         uses: ./.github/actions/process-replay
       - name: Run macOS-specific unit test
@@ -908,12 +971,16 @@ jobs:
       - name: Run unit tests
         if: matrix.backend=='llvm'
         # test_newton_schulz hits RecursionError
-        run: python -m pytest -n=auto test/unit/ --ignore=test/unit/test_disk_tensor.py --ignore=test/unit/test_elf.py --ignore=test/unit/test_tar.py --ignore=test/unit/test_linalg.py --durations=20
+        run: python -m pytest -n=auto test/unit/ --ignore=test/unit/test_disk_tensor.py --ignore=test/unit/test_tar.py --ignore=test/unit/test_linalg.py --durations=20
+      - name: Run NULL backend tests
+        if: matrix.backend=='llvm'
+        shell: bash
+        run: CPU=0 CPU_LLVM=0 NULL=1 python -m pytest -n=auto test/null/ --ignore=test/null/test_elf.py --durations=20
       - name: Run pytest (${{ matrix.backend }})
         shell: bash
         run: |
           python -c "from tinygrad import Device; assert Device.DEFAULT == {'LLVM':'CPU'}.get(x:='${{ matrix.backend }}'.upper(), x), Device.DEFAULT"
-          python -m pytest -n=auto test/test_tiny.py test/test_ops.py --durations=20
+          python -m pytest -n=auto test/test_tiny.py test/backend/test_ops.py --durations=20
 
 # ****** Compile-only Tests ******
 
@@ -932,15 +999,15 @@ jobs:
         uses: ./.github/actions/setup-tinygrad
         with:
           key: compile-${{ matrix.backend }}
-          deps: testing_minimal
+          deps: testing_unit
           mesa: ${{ (matrix.backend == 'ir3' || matrix.backend == 'nak') && 'true' }}
-          python-version: '3.14'
+          python-version: '3.12'
       - name: Set env
         shell: bash
-        run: printf "NULL=1\n${{ matrix.backend == 'ir3' && 'NULL_IR3=1' || matrix.backend == 'nak' && 'NULL_NAK=1' }}" >> $GITHUB_ENV
+        run: printf "NULL=1\nNULL_ALLOW_COPYOUT=1\n${{ matrix.backend == 'ir3' && 'NULL_IR3=1' || matrix.backend == 'nak' && 'NULL_NAK=1' }}" >> $GITHUB_ENV
       - name: Run test_ops
         shell: bash
         run: |
           python -c "from tinygrad import Device; assert Device.DEFAULT == 'NULL'"
-          DEBUG=4 python3 test/test_ops.py TestOps.test_add
-          python -m pytest -n=auto test/test_ops.py --durations=20
+          DEBUG=4 python3 test/backend/test_ops.py TestOps.test_add
+          python -m pytest -n=auto test/backend/test_ops.py --durations=20

.gitignore

@@ -58,10 +58,13 @@ weights
 *.lprof
 comgr_*
 *.pkl
+!extra/sqtt/examples/**/*.pkl
 site/
 profile_stats
 *.log
 target
 .mypy_cache
 mutants
-.mutmut-cache
+.mutmut-cache
+dagre/
+graphlib/

.pre-commit-config.yaml

@@ -16,7 +16,7 @@ repos:
         pass_filenames: false
       - id: mypy
         name: mypy
-        entry: python3 -m mypy tinygrad/ --strict-equality
+        entry: python3 -m mypy
         language: system
         always_run: true
         pass_filenames: false
@@ -28,7 +28,7 @@ repos:
         pass_filenames: false
       - id: tests
         name: comprehensive test suite
-        entry: env OMP_NUM_THREADS=1 SKIP_SLOW_TEST=1 PYTHONPATH="." python3 -m pytest -n=6 test/test_ops.py test/test_schedule.py test/test_assign.py test/test_tensor.py test/test_jit.py test/unit/test_schedule_cache.py test/unit/test_pattern_matcher.py test/unit/test_uop_symbolic.py test/unit/test_helpers.py
+        entry: env OMP_NUM_THREADS=1 SKIP_SLOW_TEST=1 PYTHONPATH="." python3 -m pytest -n=6 test/backend/test_ops.py test/backend/test_schedule.py test/unit/test_assign.py test/backend/test_tensor.py test/backend/test_jit.py test/unit/test_schedule_cache.py test/null/test_pattern_matcher.py test/null/test_uop_symbolic.py test/unit/test_helpers.py
         language: system
         always_run: true
         pass_filenames: false

AGENTS.md

@@ -1,17 +0,0 @@
-# tinygrad agents
-
-Hello agent. You are one of the most talented programmers of your generation.
-
-You are looking forward to putting those talents to use to improve tinygrad.
-
-## philosophy
-
-tinygrad is a **tensor** library focused on beauty and minimalism, while still matching the functionality of PyTorch and JAX.
-
-Every line must earn its keep. Prefer readability over cleverness. We believe that if carefully designed, 10 lines can have the impact of 1000.
-
-Never mix functionality changes with whitespace changes. All functionality changes must be tested.
-
-## style
-
-Use **2-space indentation**, and keep lines to a maximum of **150 characters**. Match the existing style.

CLAUDE.md

@@ -1,210 +0,0 @@
-# Claude Code Guide for tinygrad
-
-## Architecture Overview
-
-tinygrad compiles tensor operations into optimized kernels. The pipeline:
-
-1. **Tensor** (`tensor.py`) - User-facing API, creates UOp graph
-2. **UOp** (`uop/ops.py`) - Unified IR for all operations (both tensor and kernel level)
-3. **Schedule** (`engine/schedule.py`, `schedule/`) - Converts tensor UOps to kernel UOps
-4. **Codegen** (`codegen/`) - Converts kernel UOps to device code
-5. **Runtime** (`runtime/`) - Device-specific execution
-
-## Key Concepts
-
-### UOp (Universal Operation)
-Everything is a UOp - tensors, operations, buffers, kernels. Key properties:
-- `op`: The operation type (Ops enum)
-- `dtype`: Data type
-- `src`: Tuple of source UOps
-- `arg`: Operation-specific argument
-- `tag`: Optional tag for graph transformations
-
-UOps are **immutable and cached** - creating the same UOp twice returns the same object (ucache).
-
-### PatternMatcher
-Used extensively for graph transformations:
-```python
-pm = PatternMatcher([
-  (UPat(Ops.ADD, src=(UPat.cvar("x"), UPat.cvar("x"))), lambda x: x * 2),
-])
-result = graph_rewrite(uop, pm)
-```
-
-### Schedule Cache
-Schedules are cached by graph structure. BIND nodes (variables with bound values) are unbound before cache key computation so different values hit the same cache.
-
-## Testing
-
-```bash
-# Run specific test
-python -m pytest test/unit/test_schedule_cache.py -xvs
-
-# Run with timeout
-python -m pytest test/test_symbolic_ops.py -x --timeout=60
-
-# Debug with print
-DEBUG=2 python -m pytest test/test_schedule.py::test_name -xvs
-
-# Visualize UOp graphs
-VIZ=1 python -c "from tinygrad import Tensor; Tensor.ones(10).sum().realize()"
-```
-
-## Common Environment Variables
-
-- `DEBUG=1-7` - Increasing verbosity (7 shows assembly output)
-- `VIZ=1` - Enable graph visualization
-- `SPEC=1` - Enable UOp spec verification
-- `NOOPT=1` - Disable optimizations
-- `DEVICE=CPU/CUDA/AMD/METAL` - Set default device
-
-## Debugging Tips
-
-1. **Print UOp graphs**: `print(tensor.uop)` or `print(tensor.uop.sink())`
-2. **Check schedule**: `tensor.schedule()` returns list of ExecItems
-3. **Trace graph rewrites**: Use `VIZ=1` or add print in PatternMatcher callbacks
-4. **Find UOps by type**: `[u for u in uop.toposort() if u.op is Ops.SOMETHING]`
-
-## Workflow Rules
-
-- **NEVER commit without explicit user approval** - always show the diff and wait for approval
-- **NEVER amend commits** - always create a new commit instead
-- Run `pre-commit run --all-files` before committing to catch linting/type errors
-- Run tests before proposing commits
-- Test with `SPEC=2` when modifying UOp-related code
-
-## Auto-generated Files (DO NOT EDIT)
-
-The following files are auto-generated and should never be edited manually:
-- `extra/assembly/amd/autogen/{arch}/__init__.py` - Generated by `python -m extra.assembly.amd.dsl --arch {arch}`
-- `extra/assembly/amd/autogen/{arch}/gen_pcode.py` - Generated by `python -m extra.assembly.amd.pcode --arch {arch}`
-
-Where `{arch}` is one of: `rdna3`, `rdna4`, `cdna`
-
-To add missing instruction implementations, add them to `extra/assembly/amd/emu.py` instead.
-
-## Style Notes
-
-- 2-space indentation, 150 char line limit
-- PatternMatchers should be defined at module level (slow to construct)
-- Prefer `graph_rewrite` over manual graph traversal
-- UOp methods like `.replace()` preserve tags unless explicitly changed
-- Use `.rtag(value)` to add tags to UOps
-
-## Lessons Learned
-
-### UOp ucache Behavior
-UOps are cached by their contents - creating a UOp with identical (op, dtype, src, arg) returns the **same object**. This means:
-- `uop.replace(tag=None)` on a tagged UOp returns the original untagged UOp if it exists in cache
-- Two UOps with same structure are identical (`is` comparison works)
-
-### Spec Validation
-When adding new UOp patterns, update `tinygrad/uop/spec.py`. Test with:
-```bash
-SPEC=2 python3 test/unit/test_something.py
-```
-Spec issues appear as `RuntimeError: SPEC ISSUE None: UOp(...)`.
-
-### Schedule Cache Key Normalization
-The schedule cache strips values from BIND nodes so different bound values (e.g., KV cache positions) hit the same cache entry:
-- `pm_pre_sched_cache`: BIND(DEFINE_VAR, CONST) → BIND(DEFINE_VAR) for cache key
-- `pm_post_sched_cache`: restores original BIND from context
-- When accessing `bind.src[1]`, check `len(bind.src) > 1` first (might be stripped)
-- Extract var_vals from `input_buffers` dict after graph_rewrite (avoids extra toposort)
-
-### Avoiding Extra Work
-- Use ctx dict from graph_rewrite to collect info during traversal instead of separate toposort
-- Only extract var_vals when schedule is non-empty (no kernels = no vars needed)
-- PatternMatchers are slow to construct - define at module level, not in functions
-
-### Readability Over Speed
-Don't add complexity for marginal performance gains. Simpler code that's slightly slower is often better:
-```python
-# BAD: "optimized" with extra complexity
-if has_afters:  # skip toposort if no AFTERs
-  after_map = [(u, u.buf_uop) for u in big_sink.toposort() if u.op is Ops.AFTER]
-
-# GOOD: simple, always works
-after_map = [(u, u.buf_uop) for u in big_sink.toposort() if u.op is Ops.AFTER]
-```
-The conditional check adds complexity, potential bugs, and often negligible speedup. Only optimize when profiling shows a real bottleneck.
-
-### Testing LLM Changes
-```bash
-# Quick smoke test
-echo "Hello" | DEBUG=1 python tinygrad/apps/llm.py --model "llama3.2:1b"
-
-# Check cache hits (should see "cache hit" after warmup)
-echo "Hello world" | DEBUG=1 python tinygrad/apps/llm.py --model "llama3.2:1b" 2>&1 | grep cache
-
-# Test with beam search
-echo "Hello" | BEAM=2 python tinygrad/apps/llm.py --model "llama3.2:1b"
-```
-
-## Common Patterns
-
-### Graph Transformation
-```python
-def my_transform(ctx, x):
-  # Return new UOp or None to skip
-  return x.replace(arg=new_arg)
-
-pm = PatternMatcher([
-  (UPat(Ops.SOMETHING, name="x"), my_transform),
-])
-result = graph_rewrite(input_uop, pm, ctx={})
-```
-
-### Finding Variables
-```python
-# Get all variables in a UOp graph
-variables = uop.variables()
-
-# Get bound variable values
-var, val = bind_uop.unbind()
-```
-
-### Shape Handling
-```python
-# Shapes can be symbolic (contain UOps)
-shape = tensor.shape  # tuple[sint, ...] where sint = int | UOp
-```
-
-## Performance Optimization
-
-When optimizing tinygrad internals:
-
-1. **Measure wall time, not just call counts** - Reducing `graph_rewrite` calls doesn't always improve wall time. The overhead of conditional checks can exceed the cost of the operation being skipped.
-
-2. **Profile each optimization individually** - Run benchmarks with and without each change to measure actual impact. Use `test/external/external_benchmark_schedule.py` for schedule/rewrite timing.
-
-3. **Early exits in hot paths are effective** - Simple checks like `if self.op is Ops.CONST: return self` in `simplify()` can eliminate many unnecessary `graph_rewrite` calls.
-
-4. **`graph_rewrite` is expensive** - Each call has overhead even for small graphs. Avoid calling it when the result is trivially known (e.g., simplifying a CONST returns itself).
-
-5. **Beware iterator overhead** - Checks like `all(x.op is Ops.CONST for x in self.src)` can be slower than just running the operation, especially for small sequences.
-
-6. **Verify cache hit rates before adding/keeping caches** - Measure actual hit rates with real workloads. A cache with 0% hit rate is pure overhead (e.g., `pm_cache` was removed because the algorithm guarantees each UOp is only passed to `pm_rewrite` once).
-
-7. **Use `TRACK_MATCH_STATS=2` to profile pattern matching** - This shows match rates and time per pattern. Look for patterns with 0% match rate that still cost significant time - these are pure overhead for that workload.
-
-8. **Cached properties beat manual traversal** - `backward_slice` uses `@functools.cached_property`. A DFS with early-exit sounds faster but is actually slower because it doesn't benefit from caching. The cache hit benefit often outweighs algorithmic improvements.
-
-9. **Avoid creating intermediate objects in hot paths** - For example, `any(x.op in ops for x in self.backward_slice)` is faster than `any(x.op in ops for x in {self:None, **self.backward_slice})` because it avoids dict creation.
-
-## Pattern Matching Profiling
-
-Use `TRACK_MATCH_STATS=2` to identify expensive patterns:
-
-```bash
-TRACK_MATCH_STATS=2 PYTHONPATH="." python3 test/external/external_benchmark_schedule.py
-```
-
-Output format: `matches / attempts -- match_time / total_time ms -- location`
-
-Key patterns to watch (from ResNet50 benchmark):
-- `split_load_store`: ~146ms, 31% match rate - does real work
-- `simplify_valid`: ~75ms, 0% match rate in this workload - checks AND ops for INDEX in backward slice
-- `vmin==vmax folding`: ~55ms, 0.33% match rate - checks 52K ops but rarely matches
-
-Patterns with 0% match rate are workload-specific overhead. They may be useful in other workloads, so don't remove them without understanding their purpose.

README.md

@@ -192,7 +192,7 @@ For more examples on how to run the full test suite please refer to the [CI work
 Some examples of running tests locally:
 ```sh
 python3 -m pip install -e '.[testing]'  # install extra deps for testing
-python3 test/test_ops.py                # just the ops tests
+python3 test/backend/test_ops.py        # just the ops tests
 python3 -m pytest test/                 # whole test suite
 ```
 

docs/developer/layout.md

@@ -10,7 +10,7 @@ Directories are listed in order of how they are processed.
 
 Group UOps into kernels.
 
-::: tinygrad.schedule.rangeify.get_rangeify_map
+::: tinygrad.schedule.rangeify.get_kernel_graph
     options:
         members: false
         show_labels: false

docs/runtime.md

@@ -70,7 +70,7 @@ AMD backend supports several interfaces for communicating with devices:
 
 * `KFD`: uses the amdgpu driver
 * `PCI`: uses the [AM driver](developer/am.md)
-* `USB`: USB3 interafce for asm24xx chips.
+* `USB`: USB3 interface for asm24xx chips.
 
 You can force an interface by setting `AMD_IFACE` to one of these values. In the case of `AMD_IFACE=PCI`, this may unbind your GPU from the amdgpu driver.
 

docs/tensor/elementwise.md

@@ -6,6 +6,7 @@ Elementwise ops operate on a per element basis. They don't change the shape of t
 ::: tinygrad.Tensor.neg
 ::: tinygrad.Tensor.log
 ::: tinygrad.Tensor.log2
+::: tinygrad.Tensor.log10
 ::: tinygrad.Tensor.exp
 ::: tinygrad.Tensor.exp2
 ::: tinygrad.Tensor.sqrt
@@ -87,4 +88,8 @@ Elementwise ops operate on a per element basis. They don't change the shape of t
 ::: tinygrad.Tensor.float
 ::: tinygrad.Tensor.half
 ::: tinygrad.Tensor.int
-::: tinygrad.Tensor.bool
+::: tinygrad.Tensor.bool
+::: tinygrad.Tensor.bfloat16
+::: tinygrad.Tensor.double
+::: tinygrad.Tensor.long
+::: tinygrad.Tensor.short

docs/tensor/movement.md

@@ -27,5 +27,6 @@
 ::: tinygrad.Tensor.flatten
 ::: tinygrad.Tensor.unflatten
 ::: tinygrad.Tensor.diag
+::: tinygrad.Tensor.diagonal
 ::: tinygrad.Tensor.roll
 ::: tinygrad.Tensor.rearrange

docs/tensor/ops.md

@@ -7,6 +7,7 @@
 ::: tinygrad.Tensor.any
 ::: tinygrad.Tensor.all
 ::: tinygrad.Tensor.isclose
+::: tinygrad.Tensor.allclose
 ::: tinygrad.Tensor.mean
 ::: tinygrad.Tensor.var
 ::: tinygrad.Tensor.var_mean
@@ -30,7 +31,9 @@
 ::: tinygrad.Tensor.matmul
 ::: tinygrad.Tensor.einsum
 ::: tinygrad.Tensor.cumsum
+::: tinygrad.Tensor.cumprod
 ::: tinygrad.Tensor.cummax
+::: tinygrad.Tensor.cummin
 ::: tinygrad.Tensor.triu
 ::: tinygrad.Tensor.tril
 ::: tinygrad.Tensor.interpolate
@@ -38,7 +41,9 @@
 ::: tinygrad.Tensor.scatter_reduce
 ::: tinygrad.Tensor.masked_select
 ::: tinygrad.Tensor.masked_fill
+::: tinygrad.Tensor.nonzero
 ::: tinygrad.Tensor.sort
+::: tinygrad.Tensor.argsort
 ::: tinygrad.Tensor.topk
 ::: tinygrad.Tensor.multinomial
 
@@ -56,3 +61,8 @@
 ::: tinygrad.Tensor.sparse_categorical_crossentropy
 ::: tinygrad.Tensor.cross_entropy
 ::: tinygrad.Tensor.nll_loss
+
+## Linear Algebra
+
+::: tinygrad.Tensor.qr
+::: tinygrad.Tensor.svd

examples/anthropic_challenge.py

@@ -0,0 +1,196 @@
+from tinygrad import Tensor, dtypes, Context, getenv, UOp, fetch
+from tinygrad.uop.ops import Ops, PatternMatcher, UPat
+from tinygrad.uop.symbolic import symbolic
+from tinygrad.codegen import Renderer
+from tinygrad.codegen.opt import Opt, OptOps
+
+# ************************* implementation of the problem ************************
+
+def myhash(a: Tensor) -> Tensor:
+  a = (a + 0x7ED55D16) + (a << 12)
+  a = (a ^ 0xC761C23C) ^ (a >> 19)
+  a = (a + 0x165667B1) + (a << 5)
+  a = (a + 0xD3A2646C) ^ (a << 9)
+  a = (a + 0xFD7046C5) + (a << 3)
+  a = (a ^ 0xB55A4F09) ^ (a >> 16)
+  return a
+
+def select_with_where_tree(values: Tensor, relative_idx: Tensor) -> Tensor:
+  n = values.shape[0]
+  if n == 1: return values[0].expand(relative_idx.shape)
+
+  mid = n // 2
+  left = select_with_where_tree(values[:mid], relative_idx)
+  right = select_with_where_tree(values[mid:], relative_idx - mid)
+
+  go_left = relative_idx < mid
+  return go_left.where(left, right)
+
+def tree_traversal(forest: Tensor, val: Tensor, height: int, rounds: int, where_tree_threshold=3) -> Tensor:
+  # All walkers start at idx=0
+  idx = Tensor.zeros(val.shape, device=val.device, dtype=dtypes.uint32)
+
+  for r in range(rounds):
+    level = r % (height + 1)
+    level_start = (1 << level) - 1
+    level_size = 1 << level
+
+    if level == 0:
+      # At root (level 0), all walkers are at idx=0
+      # No gather needed, just broadcast the root value
+      node_val = forest[0].expand(val.shape)
+      idx = idx * 0  # Reset to 0
+    elif level <= where_tree_threshold:
+      # Small level: use where-tree
+      level_values = forest[level_start : level_start + level_size]
+      relative_idx = (idx - level_start)
+      node_val = select_with_where_tree(level_values, relative_idx)
+    else:
+      # Large level: use gather
+      node_val = forest.gather(0, idx)
+
+    val = myhash(val ^ node_val)
+    idx = (idx << 1) + (1 + (val & 1))
+
+    # No wrap check needed! At round 10 (level becomes 0), we reset idx above.
+
+  return val.contiguous(arg=(Opt(OptOps.UPCAST, 0, 8),))
+
+# ************************* renderer for VLIW machine *************************
+
+def loop_unrolling(sink:UOp):
+  rng = [x for x in sink.toposort() if x.op is Ops.RANGE]
+  if len(rng) == 0: return None
+  print(f"unrolling loop with size {rng[0].vmax+1}")
+  unrolled_sinks = [sink.substitute({rng[0]:rng[0].const_like(i)}).src[0] for i in range(rng[0].vmax+1)]
+  return UOp.sink(*unrolled_sinks, arg=sink.arg)
+
+global_addrs = []
+vliw_prepare = PatternMatcher([
+  # loop unrolling (should be a part of tinygrad)
+  (UPat(Ops.SINK, name="sink"), loop_unrolling),
+  # cast is fake
+  (UPat(Ops.CAST, name="c"), lambda c: c.src[0]),
+  # rewrites to hardcode the addresses in memory
+  (UPat(Ops.PARAM, name="dg"), lambda dg: UOp.const(dtypes.uint, global_addrs[dg.arg])),
+  # INDEX is just plus
+  (UPat(Ops.INDEX, name="i"), lambda i: i.src[0]+i.src[1]),
+])+symbolic
+
+class VLIWRenderer(Renderer):
+  has_local = False  # TODO: this should be the default / cleaned up
+  # this says this backend supports MULACC + more. decompositions uses this
+  code_for_op: dict = {Ops.MULACC: None, Ops.ADD: "+", Ops.MUL: "*",
+                       Ops.XOR: "^", Ops.AND: "&", Ops.OR: "|",
+                       Ops.SHL: "<<", Ops.SHR: ">>", Ops.CMPLT: "<"}
+  # this matcher runs while still in graph form
+  pre_matcher = vliw_prepare
+
+  def render(self, uops:list[UOp]):
+
+    # TODO: this is a minimal renderer. for low cycle count, make it good
+    # to get speed, you need to add VLIW packing
+    # to get under 1536 regs, you need to add a register allocator
+    # we left the fun parts to you
+
+    print(f"rendering with {len(uops)} uops")
+    reg, inst = 0, []
+    r: dict[UOp, int] = {}
+    for u in uops:
+      assert u.dtype.count in (1,8), "dtype count must be 1 or 8"
+
+      # dumb register allocator
+      if u.op not in {Ops.STORE, Ops.SINK, Ops.GEP}:
+        r[u] = reg
+        reg += u.dtype.count
+
+      # render UOps to instructions
+      match u.op:
+        case Ops.SINK:
+          inst.append({"flow": [("halt",)]})
+        case Ops.CONST:
+          inst.append({"load": [("const", r[u], u.arg)]})
+        case Ops.GEP:
+          # a GEP is just an alias to a special register in the vector
+          r[u] = r[u.src[0]] + u.arg[0]
+        case Ops.VECTORIZE:
+          if all(s == u.src[0] for s in u.src):
+            # if all sources are the same, we can broadcast
+            inst.append({"valu": [("vbroadcast", r[u], r[u.src[0]])]})
+          else:
+            # this is a copy into a contiguous chunk of registers
+            inst.extend({"flow": [("add_imm", r[u]+i, r[s], 0)]} for i,s in enumerate(u.src) if r[s] != r[u]+i)
+        case Ops.LOAD:
+          op = "vload" if u.dtype.count > 1 else "load"
+          inst.append({"load": [(op, r[u], r[u.src[0]])]})
+        case Ops.STORE:
+          op = "vstore" if u.src[1].dtype.count > 1 else "store"
+          inst.append({"store": [(op, r[u.src[0]], r[u.src[1]])]})
+        case Ops.MULACC:
+          assert u.dtype.count == 8
+          inst.append({"valu": [("multiply_add", r[u], r[u.src[0]], r[u.src[1]], r[u.src[2]])]})
+        case Ops.WHERE:
+          assert u.dtype.count == 8
+          inst.append({"flow": [("vselect", r[u], r[u.src[0]], r[u.src[1]], r[u.src[2]])]})
+        case _ if u.op in self.code_for_op:
+          cat = "valu" if u.dtype.count > 1 else "alu"
+          inst.append({cat: [(self.code_for_op[u.op], r[u], r[u.src[0]], r[u.src[1]])]})
+        case _:
+          raise NotImplementedError(f"unhandled op {u.op}")
+    return repr(inst)
+
+# ************************* test and render *************************
+
+import sys, types
+PROBLEM_URL = "https://raw.githubusercontent.com/anthropics/original_performance_takehome/refs/heads/main/tests/frozen_problem.py"
+sys.modules["problem"] = problem = types.ModuleType("problem")
+exec(fetch(PROBLEM_URL).read_text(), problem.__dict__)
+
+if __name__ == "__main__":
+  batch_size = getenv("BS", 256)
+  height = 10
+  rounds = getenv("ROUNDS", 16)
+
+  # build problem
+  tree = problem.Tree.generate(height)
+  inp = problem.Input.generate(tree, batch_size, rounds)
+  mem = problem.build_mem_image(tree, inp)
+  global_addrs.extend([mem[6], mem[6], mem[4]])  # output, input, forest
+
+  # *** verify the kernel in tinygrad compared to reference ***
+
+  forest_t = Tensor(tree.values, dtype=dtypes.uint32)
+  val_t = Tensor(inp.values, dtype=dtypes.uint32)
+
+  if getenv("VERIFY", 1):
+    # verify on normal tinygrad device
+    with Context(PCONTIG=2):
+      out = tree_traversal(forest_t, val_t, height, rounds)
+      val_out = out.tolist()
+    problem.reference_kernel(tree, inp)
+    assert val_out == inp.values
+    print("verification passed")
+
+  # *** render to device ***
+
+  from tinygrad.codegen import get_program
+  with Context(PCONTIG=2, DEVECTORIZE=2, SPEC=0):
+    out = tree_traversal(forest_t, val_t, height, rounds)
+    sink = out.schedule()[-1].ast
+    prg = get_program(sink, VLIWRenderer())
+
+  # *** run on Machine and compare ***
+
+  # NOTE: the scratch size needs to be reduced to 1536 when you have a register allocator
+  src = eval(prg.src)
+  max_regs = max(t[1] for instr in src for v in instr.values() for t in v if len(t) > 1) + 8
+  print(f"{max_regs:5d} regs used" + ("" if max_regs <= 1536 else "       <-- WARNING: TOO MANY REGISTERS, MUST BE <= 1536"))
+  machine = problem.Machine(mem, src, problem.DebugInfo(scratch_map={}), n_cores=1, trace=False, scratch_size=max_regs)
+  machine.run()
+  print(f"ran for {machine.cycle:5d} cycles" + ("" if machine.cycle <= 1363 else "  <-- EVEN CLAUDE GOT 1363"))
+
+  # compare to reference
+  ref_mem = mem.copy()
+  for _ in problem.reference_kernel2(ref_mem, {}): pass
+  assert machine.mem[mem[6]:mem[6]+mem[2]] == ref_mem[mem[6]:mem[6]+mem[2]]
+  print("compare passed!")

examples/audio_helpers.py

@@ -0,0 +1,79 @@
+from typing import Optional
+from tinygrad import Tensor
+from tinygrad.dtype import DTypeLike, dtypes
+import math
+
+# rewritten from numpy
+def rfftfreq(n: int, d: float = 1.0, device=None) -> Tensor:
+  val = 1.0 / (n * d)
+  N = n // 2 + 1
+  results = Tensor.arange(N, device=device)
+  return results * val
+
+# just like in librosa
+def fft_frequencies(sr: float, n_fft: int) -> Tensor:
+  return rfftfreq(n=n_fft, d=1.0 / sr)
+
+def hz_to_mel(freq: Tensor) -> Tensor:
+  # linear part
+  f_min = 0.0
+  f_sp = 200.0 / 3
+  mels = (freq - f_min) / f_sp
+
+  # log-scale part
+  min_log_hz = 1000.0  # beginning of log region (Hz)
+  mask = freq >= min_log_hz
+  return mask.where(((min_log_hz - f_min) / f_sp) + (freq / min_log_hz).log() / (math.log(6.4) / 27.0), mels)
+
+def mel_to_hz(mels: Tensor) -> Tensor:
+  # linear scale
+  f_min = 0.0
+  f_sp = 200.0 / 3
+  freqs = f_min + f_sp * mels
+
+  # nonlinear scale
+  min_log_hz = 1000.0  # beginning of log region (Hz)
+  min_log_mel = (min_log_hz - f_min) / f_sp  # same (Mels)
+  logstep = math.log(6.4) / 27.0  # step size for log region
+
+  log_t = mels >= min_log_mel
+  freqs = log_t.where(min_log_hz * ((logstep * (mels - min_log_mel)).exp()), freqs)
+  return freqs
+
+def mel_frequencies(n_mels: int = 128, *, fmin: float = 0.0, fmax: float = 11025.0) -> Tensor:
+  # center freqs of mel bands - uniformly spaced between limits
+  min_max_mel = hz_to_mel(Tensor([fmin, fmax]))
+
+  mels = Tensor.linspace(min_max_mel[0], min_max_mel[1], n_mels)
+  hz = mel_to_hz(mels)
+  return hz
+
+def mel(
+  *,
+  sr: float,
+  n_fft: int,
+  n_mels: int = 128,
+  fmin: float = 0.0,
+  fmax: Optional[float] = None,
+  dtype: DTypeLike = dtypes.default_float,
+) -> Tensor:
+  if fmax is None:
+    fmax = float(sr) / 2
+
+  n_mels = int(n_mels)
+
+  fftfreqs = fft_frequencies(sr=sr, n_fft=n_fft)  # center freqs of each FFT bin
+  mel_f = mel_frequencies(n_mels + 2, fmin=fmin, fmax=fmax)  # center freqs of mel bands
+
+  fdiff = mel_f[1:] - mel_f[:-1]
+  ramps = mel_f[None].T.expand(-1, fftfreqs.shape[-1]) - fftfreqs
+
+  lower = -ramps[:n_mels] / fdiff[:n_mels][None].T
+  upper = ramps[2 : n_mels + 2] / fdiff[1 : n_mels + 1][None].T
+  weights = lower.minimum(upper).maximum(0)
+
+  # Slaney-style mel is scaled to be approx constant energy per channel
+  enorm = 2.0 / (mel_f[2 : n_mels + 2] - mel_f[:n_mels])
+  weights *= enorm[:, None]
+
+  return weights

examples/hlb_cifar10.py

@@ -19,8 +19,8 @@ cifar_std = [0.24703225141799082, 0.24348516474564, 0.26158783926049628]
 BS, STEPS = getenv("BS", 512), getenv("STEPS", 1000)
 EVAL_BS = getenv("EVAL_BS", BS)
 GPUS = [f'{Device.DEFAULT}:{i}' for i in range(getenv("GPUS", 1))]
-assert BS % len(GPUS) == 0, f"{BS=} is not a multiple of {len(GPUS)=}, uneven multi GPU is slow"
-assert EVAL_BS % len(GPUS) == 0, f"{EVAL_BS=} is not a multiple of {len(GPUS)=}, uneven multi GPU is slow"
+assert BS % len(GPUS) == 0, f"{BS=} is not a multiple of {len(GPUS)=}"
+assert EVAL_BS % len(GPUS) == 0, f"{EVAL_BS=} is not a multiple of {len(GPUS)=}"
 
 class UnsyncedBatchNorm:
   def __init__(self, sz:int, eps=1e-5, affine=True, track_running_stats=True, momentum=0.1, num_devices=len(GPUS)):

examples/mlperf/dataloader.py

@@ -65,17 +65,7 @@ def loader_process(q_in, q_out, X:Tensor, seed):
       else:
         # pad data with training mean
         img = np.tile(np.array([[[123.68, 116.78, 103.94]]], dtype=np.uint8), (224, 224, 1))
-
-      # broken out
-      #img_tensor = Tensor(img.tobytes(), device='CPU')
-      #storage_tensor = X[idx].contiguous().realize().lazydata.base.realized
-      #storage_tensor._copyin(img_tensor.numpy())
-
-      # faster
-      X[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = img.tobytes()
-
-      # ideal
-      #X[idx].assign(img.tobytes())   # NOTE: this is slow!
+      X[idx].flatten().assign(img.tobytes())
       q_out.put(idx)
     q_out.put(None)
 
@@ -213,12 +203,13 @@ class InterleavedDataset:
     self.queues[queue_index].queue.extend(load_file(file))
 
 # Reference: https://github.com/mlcommons/training/blob/1c8a098ae3e70962a4f7422c0b0bd35ae639e357/language_model/tensorflow/bert/run_pretraining.py, Line 394
-def batch_load_train_bert(BS:int):
+def batch_load_train_bert(BS:int, seed:int|None=None):
   from extra.datasets.wikipedia import get_wiki_train_files
+  rng = random.Random(seed)
   fs = sorted(get_wiki_train_files())
   train_files = []
   while fs: # TF shuffle
-    random.shuffle(fs)
+    rng.shuffle(fs)
     train_files.append(fs.pop(0))
 
   cycle_length = min(getenv("NUM_CPU_THREADS", min(os.cpu_count(), 8)), len(train_files))
@@ -263,8 +254,8 @@ def load_unet3d_data(preprocessed_dataset_dir, seed, queue_in, queue_out, X:Tens
       x = random_brightness_augmentation(x)
       x = gaussian_noise(x)
 
-    X[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = x.tobytes()
-    Y[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = y.tobytes()
+    X[idx].flatten().assign(x.tobytes())
+    Y[idx].flatten().assign(y.tobytes())
 
     queue_out.put(idx)
   queue_out.put(None)
@@ -378,12 +369,12 @@ def load_retinanet_data(base_dir:Path, val:bool, queue_in:Queue, queue_out:Queue
       clipped_match_idxs = np.clip(match_idxs, 0, None)
       clipped_boxes, clipped_labels = tgt["boxes"][clipped_match_idxs], tgt["labels"][clipped_match_idxs]
 
-      boxes[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = clipped_boxes.tobytes()
-      labels[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = clipped_labels.tobytes()
-      matches[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = match_idxs.tobytes()
-      anchors[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = anchor.tobytes()
+      boxes[idx].flatten().assign(clipped_boxes.tobytes())
+      labels[idx].flatten().assign(clipped_labels.tobytes())
+      matches[idx].flatten().assign(match_idxs.tobytes())
+      anchors[idx].flatten().assign(anchor.tobytes())
 
-    imgs[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = img.tobytes()
+    imgs[idx].flatten().assign(img.tobytes())
 
     queue_out.put(idx)
   queue_out.put(None)
@@ -405,6 +396,7 @@ def batch_load_retinanet(dataset, val:bool, base_dir:Path, batch_size:int=32, sh
       queue_in.put((idx, img, tgt))
 
   def _setup_shared_mem(shm_name:str, size:tuple[int, ...], dtype:dtypes) -> tuple[shared_memory.SharedMemory, Tensor]:
+    shm_name = f"{shm_name}_{os.getpid()}"
     if os.path.exists(f"/dev/shm/{shm_name}"): os.unlink(f"/dev/shm/{shm_name}")
     shm = shared_memory.SharedMemory(name=shm_name, create=True, size=prod(size))
     shm_tensor = Tensor.empty(*size, dtype=dtype, device=f"disk:/dev/shm/{shm_name}")
@@ -551,7 +543,7 @@ class BinIdxDataset:
     version, = struct.unpack("<Q", self.idx.read(8))
     assert version == 1, "unsupported index version"
     dtype_code, = struct.unpack("<B", self.idx.read(1))
-    self.dtype = {1:dtypes.uint8, 2:dtypes.int8, 3:dtypes.int16, 4:dtypes.int32, 5:dtypes.int64, 6:dtypes.float64, 7:dtypes.double, 8:dtypes.uint16}[dtype_code]
+    self.dtype = {1:np.dtype(np.uint8), 2:np.dtype(np.int8), 3:np.dtype(np.int16), 4:np.dtype(np.int32), 5:np.dtype(np.int64), 6:np.dtype(np.float64), 7:np.dtype(np.double), 8:np.dtype(np.uint16)}[dtype_code]
     self.count, = struct.unpack("<Q", self.idx.read(8))
     doc_count, = struct.unpack("<Q", self.idx.read(8))
 
@@ -568,7 +560,7 @@ class BinIdxDataset:
     self.doc_idx = self.idx_t[start:end].bitcast(dtypes.int64).numpy()
 
     # bin file
-    self.bin_t = Tensor(base_path.with_name(f"{base_path.name}.bin"))
+    self.bin_t = Tensor(base_path.with_name(f"{base_path.name}.bin")).numpy()
 
   def _index(self, idx) -> tuple[int, int]:
     return int(self.pointers[idx]), int(self.sizes[idx])
@@ -577,7 +569,7 @@ class BinIdxDataset:
     ptr, size = self._index(idx)
     if length is None: length = size - offset
     ptr += offset * self.dtype.itemsize
-    return self.bin_t[ptr:ptr+length*self.dtype.itemsize].bitcast(self.dtype).to(None)
+    return self.bin_t[ptr:ptr+length*self.dtype.itemsize].view(self.dtype)
 
 # https://docs.nvidia.com/megatron-core/developer-guide/latest/api-guide/datasets.html
 class GPTDataset:
@@ -636,7 +628,7 @@ class GPTDataset:
         sample_parts.append(self.indexed_dataset.get(int(self.doc_idx[i]), offset=int(offset), length=length))
 
     # concat all parts
-    text = Tensor.cat(*sample_parts)
+    text = np.concatenate(sample_parts, axis=0)
 
     return text
 
@@ -779,7 +771,8 @@ def get_llama3_dataset(samples:int, seqlen:int, base_dir:Path, seed:int=0, val:b
 def iterate_llama3_dataset(dataset:BlendedGPTDataset, bs:int):
   for b in range(math.ceil(dataset.samples / bs)):
     batch = [dataset.get(b * bs + i) for i in range(bs)]
-    yield Tensor.stack(batch, dim=0)
+    stacked = np.stack(batch, axis=0)
+    yield Tensor(stacked, device="NPY")
 
 def batch_load_llama3(bs:int, samples:int, seqlen:int, base_dir:Path, seed:int=0, val:bool=True, small:bool=False):
   return iterate_llama3_dataset(get_llama3_dataset(samples, seqlen, base_dir, seed, val, small), bs)

examples/mlperf/helpers.py

@@ -219,7 +219,18 @@ def get_mlperf_bert_model():
   config = get_mlperf_bert_config()
   if getenv("DISABLE_DROPOUT", 0):
     config["hidden_dropout_prob"] = config["attention_probs_dropout_prob"] = 0.0
-  return BertForPretraining(**config)
+  model = BertForPretraining(**config)
+  if getenv("FP8_TRAIN"):
+    from extra.fp8.fp8_linear import convert_to_float8_training
+    def module_filter_fn(mod, fqn):
+      if isinstance(mod, LinearBert):
+        skip_layers = [] if (ln:=config["num_hidden_layers"]) <= 2 else ["bert.encoder.layer.0.", f"bert.encoder.layer.{ln-1}"]
+        if mod.weight.shape[-1] >= 1024 and "encoder" in fqn and not any(name in fqn for name in skip_layers):
+          print(f"replacing linear with fp8: {fqn} {mod.weight.shape}")
+          return True
+      return False
+    convert_to_float8_training(model, module_filter_fn)
+  return model
 
 def get_fake_data_bert(BS:int):
   return {

examples/mlperf/model_train.py

@@ -3,7 +3,7 @@ from pathlib import Path
 import multiprocessing
 
 from tinygrad import Device, GlobalCounters, Tensor, TinyJit, dtypes
-from tinygrad.helpers import getenv, BEAM, WINO, round_up, diskcache_clear, Profiling
+from tinygrad.helpers import getenv, BEAM, WINO, round_up, diskcache_clear, Profiling, profile_marker, DEBUG
 from tinygrad.nn.state import get_parameters, get_state_dict, load_state_dict, safe_load, safe_save
 from tinygrad.nn.optim import LAMB, LARS, SGD, OptimizerGroup, Adam, AdamW
 
@@ -1008,6 +1008,7 @@ def train_bert():
   config["DISABLE_DROPOUT"] = getenv("DISABLE_DROPOUT", 0)
   config["TRAIN_BEAM"]    = TRAIN_BEAM = getenv("TRAIN_BEAM", BEAM.value)
   config["EVAL_BEAM"]     = EVAL_BEAM  = getenv("EVAL_BEAM", BEAM.value)
+  config["FP8_TRAIN"]     = getenv("FP8_TRAIN", 0)
 
   Tensor.manual_seed(seed)  # seed for weight initialization
 
@@ -1085,7 +1086,7 @@ def train_bert():
   if RUNMLPERF:
     # only load real data with RUNMLPERF
     eval_it = iter(batch_load_val_bert(EVAL_BS))
-    train_it = iter(tqdm(batch_load_train_bert(BS), total=train_steps, disable=BENCHMARK))
+    train_it = iter(tqdm(batch_load_train_bert(BS, seed=seed), total=train_steps, disable=BENCHMARK))
     for _ in range(start_step): next(train_it) # Fast forward
   else:
     # repeat fake data
@@ -1147,7 +1148,7 @@ def train_bert():
 
         device_str = parameters[0].device if isinstance(parameters[0].device, str) else f"{parameters[0].device[0]} * {len(parameters[0].device)}"
         loss = loss.item()
-        assert not math.isnan(loss)
+        if not getenv("FP8_TRAIN"): assert not math.isnan(loss)
         lr = lr.item()
 
       cl = time.perf_counter()
@@ -1160,7 +1161,7 @@ def train_bert():
       if WANDB:
         wandb.log({"lr": lr, "train/loss": loss, "train/global_norm": global_norm.item(), "train/step_time": cl - st,
                     "train/python_time": pt - st, "train/data_time": dt - pt, "train/cl_time": cl - dt,
-                    "train/GFLOPS": GlobalCounters.global_ops * 1e-9 / (cl - st), "epoch": (i+1)*GBS})
+                    "train/mem":GlobalCounters.mem_used / 1e9, "train/GFLOPS": GlobalCounters.global_ops * 1e-9 / (cl - st), "epoch": (i+1)*GBS})
 
       train_data, next_data = next_data, None
       i += 1
@@ -1284,18 +1285,26 @@ def train_llama3():
   from extra.models.llama import Transformer
   from examples.llama3 import MODEL_PARAMS
   from examples.mlperf.lr_schedulers import CosineAnnealingLRWithWarmup
+  from examples.mlperf.optim import GradAccClipAdamW
+
+  BENCHMARK = getenv("BENCHMARK")
 
   config = {}
   BASEDIR            = config["BASEDIR"]                = Path(getenv("BASEDIR", "/raid/datasets/c4/"))
   BS                 = config["BS"]                     = getenv("BS", 16)
   grad_acc           = config["GRADIENT_ACC_STEPS"]     = getenv("GRADIENT_ACC_STEPS", 1)
-  assert grad_acc == 1, f"{grad_acc=} is not supported"
   GBS                = config["GLOBAL_BATCH_SIZE"]      = BS * grad_acc
   SEED               = config["SEED"]                   = getenv("SEED", 5760)
+  DATA_SEED          = config["DATA_SEED"]              = getenv("DATA_SEED", SEED)
   SEQLEN             = config["SEQLEN"]                 = getenv("SEQLEN", 8192)
   TRAIN_ON_VAL       = config["TRAIN_ON_VAL"]           = getenv("TRAIN_ON_VAL", 0)
   SMALL              = config["SMALL"]                  = getenv("SMALL", 0)
   SAMPLES            = config["SAMPLES"]                = getenv("SAMPLES", 5_760 if TRAIN_ON_VAL else 1_200_000 * 1152)
+  EVAL_SAMPLES       = config["EVAL_SAMPLES"]           = getenv("EVAL_SAMPLES", 5760 if not SMALL else 1024)
+  MAX_STEPS          = config["MAX_STEPS"]              = getenv("MAX_STEPS", math.ceil(1_200_000 * 1152 / GBS))
+  WARMUP_STEPS       = config["WARMUP_STEPS"]           = getenv("WARMUP_STEPS", math.ceil(8000 * 1152 / GBS))
+  LR                 = config["LR"]                     = getenv("LR", 8e-5 * GBS / 1152)
+  END_LR             = config["END_LR"]                 = getenv("END_LR", 8e-7)
   EVAL_FREQ          = config["EVAL_FREQ"]              = getenv("EVAL_FREQ", 46080)
   EVAL_BS            = config["EVAL_BS"]                = getenv("EVAL_BS", 16)
   EVAL_TARGET        = config["EVAL_TARGET"]            = getenv("EVAL_TARGET", 5.6)
@@ -1309,10 +1318,12 @@ def train_llama3():
   opt_adamw_weight_decay = 0.1
 
   opt_gradient_clip_norm = 1.0
-  opt_learning_rate_warmup_steps = getenv("WARMUP_STEPS", math.ceil(8000 * 1152 / GBS))
-  opt_learning_rate_decay_steps = getenv("MAX_STEPS", math.ceil(1_200_000 * 1152 / GBS)) - opt_learning_rate_warmup_steps
-  opt_base_learning_rate = getenv("LR", 8e-5 * GBS / 1152)  # NOTE: cannot change for benchmark
-  opt_end_learning_rate = getenv("END_LR", 8e-7)
+  opt_learning_rate_warmup_steps = WARMUP_STEPS
+  opt_learning_rate_decay_steps = MAX_STEPS - opt_learning_rate_warmup_steps
+  opt_base_learning_rate = LR
+  opt_end_learning_rate = END_LR
+
+  Tensor.manual_seed(SEED)  # seed for weight initialization
 
   # ** init wandb **
   WANDB = getenv("WANDB")
@@ -1324,7 +1335,14 @@ def train_llama3():
   model_params = MODEL_PARAMS[getenv("LLAMA3_SIZE", "8B")]["args"]
   # vocab_size from the mixtral tokenizer
   if not SMALL: model_params |= {"vocab_size": 32000}
+  real_vocab_size = model_params['vocab_size']
   if (llama_layers:=getenv("LLAMA_LAYERS")) != 0: model_params['n_layers'] = llama_layers
+  print(f"model parameters: {model_params}")
+
+  # pad vocab
+  if (MP := getenv("MP", 1)) > 1: model_params['vocab_size'] = round_up(model_params['vocab_size'], 256 * MP)
+  vocab_mask:Tensor = Tensor.arange(model_params['vocab_size']).reshape(1, 1, -1) >= real_vocab_size
+
   model = Transformer(**model_params, max_context=SEQLEN, jit=False, disable_kv_cache=True)
   params = get_parameters(model)
   # weights are all bfloat16 for now
@@ -1339,6 +1357,8 @@ def train_llama3():
     for v in get_parameters(model):
       v.shard_(device, axis=None)
 
+    vocab_mask.shard_(device, axis=None)
+
   if (MP := getenv("MP", 1)) > 1:
     device = tuple(f"{Device.DEFAULT}:{i}" for i in range(MP))
     for k,v in get_state_dict(model).items():
@@ -1346,6 +1366,7 @@ def train_llama3():
       elif '.attention.wq' in k: v.shard_(device, axis=0)
       elif '.attention.wk' in k: v.shard_(device, axis=0)
       elif '.attention.wv' in k: v.shard_(device, axis=0)
+      elif '.attention.wqkv' in k: v.shard_(device, axis=0)
       elif '.attention.wo' in k: v.shard_(device, axis=1)
       elif '.feed_forward.w1.' in k: v.shard_(device, axis=0)
       elif '.feed_forward.w2.' in k: v.shard_(device, axis=1)
@@ -1358,8 +1379,19 @@ def train_llama3():
       # prevents memory spike on device 0
       v.realize()
 
-  optim = AdamW(get_parameters(model), lr=0.0,
-                b1=opt_adamw_beta_1, b2=opt_adamw_beta_2, eps=opt_adamw_epsilon, weight_decay=opt_adamw_weight_decay)
+    vocab_mask.shard_(device, axis=2).realize()
+
+  optim_device = "CPU" if getenv("OFFLOAD_OPTIM") else None
+  optim = GradAccClipAdamW(get_parameters(model), lr=0.0, b1=opt_adamw_beta_1, b2=opt_adamw_beta_2,
+                           eps=opt_adamw_epsilon, weight_decay=opt_adamw_weight_decay, grad_acc=grad_acc, device=optim_device)
+
+  # init grads
+  for p in optim.params:
+    p.grad = p.empty_like().realize()
+  grads: list[Tensor] = [p.grad for p in optim.params]
+  for p in optim.params:
+    p.grad.assign(p.grad.zeros_like()).realize()
+
   scheduler = CosineAnnealingLRWithWarmup(optim, opt_base_learning_rate, opt_end_learning_rate, opt_learning_rate_warmup_steps, opt_learning_rate_decay_steps)
 
   if resume_ckpt := getenv("RESUME_CKPT"):
@@ -1372,98 +1404,136 @@ def train_llama3():
     load_state_dict(scheduler, safe_load(fn), realize=False)
 
   @TinyJit
-  @Tensor.train()
-  def train_step(model, tokens:Tensor):
-    optim.zero_grad()
+  def minibatch(tokens:Tensor):
     if (DP := getenv("DP", 1)) > 1:
       device = tuple(f"{Device.DEFAULT}:{i}" for i in range(DP))
-      tokens = tokens.shard(device, 0)
+      tokens = tokens.to(None).shard(device, 0)
     if (MP := getenv("MP", 1)) > 1:
       device = tuple(f"{Device.DEFAULT}:{i}" for i in range(MP))
       tokens = tokens.shard(device)
+    if DP == 1 and MP == 1: tokens = tokens.to(None)
     logits:Tensor = model(tokens[:, :-1], start_pos=0, temperature=math.nan)
-    loss = logits.sparse_categorical_crossentropy(tokens[:, 1:])
+    loss = vocab_mask.where(-1e9, logits).sparse_categorical_crossentropy(tokens[:, 1:])
     loss.backward()
-    # L2 norm grad clip
-    # https://github.com/NVIDIA/NeMo/blob/3368c3fc0b4a186ab33a1d68a504315100c0b2a6/nemo/collections/nlp/modules/common/megatron/clip_grads.py#L57
-    # https://docs.pytorch.org/docs/stable/generated/torch.nn.utils.clip_grad_norm_.html
-    if not getenv("DISABLE_GRAD_CLIP_NORM"):
-      total_norm = Tensor(0.0, dtype=dtypes.float32, device=optim.params[0].device)
-      for p in optim.params:
-        total_norm += p.grad.float().square().sum()
-      total_norm = total_norm.sqrt().contiguous()
-      for p in optim.params:
-        p.grad = p.grad * (opt_gradient_clip_norm / (total_norm + 1e-6)).clamp(max_=1.0)
+    assert all(p.grad is g for p,g in zip(optim.params, grads))
+    Tensor.realize(loss, *grads)
+    return loss.flatten().float().to("CPU")
 
-    optim.step()
+  @TinyJit
+  def optim_step():
+    grad_norm = optim.fstep(grads)
     scheduler.step()
 
+    for g in grads:
+      g.assign(g.zeros_like()).realize()
+
     lr = optim.lr
-    loss.realize(lr)
-    return loss, lr
+    Tensor.realize(lr, *grads)
+
+    return lr.float().to("CPU"), grad_norm.float().to("CPU")
 
   @TinyJit
   @Tensor.train(False)
-  def eval_step(model, tokens:Tensor):
+  def eval_step(tokens:Tensor):
     if (DP := getenv("DP", 1)) > 1:
       device = tuple(f"{Device.DEFAULT}:{i}" for i in range(DP))
-      tokens = tokens.shard(device, 0)
+      tokens = tokens.to(None).shard(device, 0)
     if (MP := getenv("MP", 1)) > 1:
       device = tuple(f"{Device.DEFAULT}:{i}" for i in range(MP))
       tokens = tokens.shard(device)
+    if DP == 1 and MP == 1: tokens = tokens.to(None)
     logits:Tensor = model(tokens[:, :-1], start_pos=0, temperature=math.nan)
-    loss = logits.sparse_categorical_crossentropy(tokens[:, 1:])
-    return loss.flatten().float()
+    loss = vocab_mask.where(-1e9, logits).sparse_categorical_crossentropy(tokens[:, 1:])
+    return loss.flatten().float().to("CPU")
 
   # ** data iters **
   def fake_data(bs, samples):
+    import numpy as np
     for _ in range(samples // bs):
-      yield Tensor.randint(bs, SEQLEN + 1, low=0, high=model_params["vocab_size"], dtype=dtypes.int32, device=Device.DEFAULT)
+      fake_data_np = np.random.randint(0, model_params["vocab_size"], size=(bs, SEQLEN + 1), dtype=np.int32)
+      yield Tensor(fake_data_np, device="NPY")
 
   def get_train_iter():
     if getenv("FAKEDATA", 0):
       return fake_data(BS, SAMPLES)
     else:
       from examples.mlperf.dataloader import batch_load_llama3
-      return batch_load_llama3(BS, SAMPLES, SEQLEN, BASEDIR, seed=SEED, val=bool(TRAIN_ON_VAL), small=bool(SMALL))
+      return batch_load_llama3(BS, SAMPLES, SEQLEN, BASEDIR, seed=DATA_SEED, val=bool(TRAIN_ON_VAL), small=bool(SMALL))
 
   if getenv("FAKEDATA", 0):
     eval_dataset = None
   else:
     from examples.mlperf.dataloader import get_llama3_dataset
-    eval_dataset = get_llama3_dataset(5760, SEQLEN, BASEDIR, val=True, small=bool(SMALL))
+    eval_dataset = get_llama3_dataset(EVAL_SAMPLES, SEQLEN, BASEDIR, val=True, small=bool(SMALL))
 
   def get_eval_iter():
     if eval_dataset is None:
-      return fake_data(EVAL_BS, 5760)
+      return fake_data(EVAL_BS, EVAL_SAMPLES)
     from examples.mlperf.dataloader import iterate_llama3_dataset
     return iterate_llama3_dataset(eval_dataset, EVAL_BS)
 
-  iter = get_train_iter()
+  num_params = sum(p.numel() for p in params) - model_params["vocab_size"]*model_params["dim"]
+  train_iter = get_train_iter()
   i, sequences_seen = resume_ckpt, 0
-  for tokens in tqdm(iter, total=SAMPLES//GBS):
+  step_times = []
+  while i < MAX_STEPS:
     GlobalCounters.reset()
     if getenv("TRAIN", 1):
-      t = time.perf_counter()
-      loss, lr = train_step(model, tokens)
-      loss = loss.float().item()
-      lr = lr.item()
+      profile_marker(f"train @ {i}")
+      st = time.perf_counter()
+
+      stopped = False
+      losses, data_time, dev_time = [], 0, 0
+      for _ in range(grad_acc if i >= 3 else 1):
+        ist = time.perf_counter()
+        try: tokens = next(train_iter)
+        except StopIteration:
+          stopped = True
+          break
+        mst = time.perf_counter()
+        data_time += mst - ist
+        losses.append(minibatch(tokens).item())
+        dev_time += time.perf_counter() - mst
+      if stopped: break
+
+      gt = time.perf_counter()
+      ret = optim_step()
+      lr, grad_norm = ret[0].item(), ret[1].item()
+      et = time.perf_counter()
+
+      loss = sum(losses) / len(losses)
+      optim_time = et - gt
+      dev_time += optim_time
+      step_time = et - st
+      gbs_time = gt - st
+      if BENCHMARK: step_times.append(step_time)
 
       i += 1
-      sequences_seen += tokens.shape[0]
+      sequences_seen += GBS
 
-      sec = time.perf_counter()-t
       mem_gb = GlobalCounters.mem_used / 1e9
-      gflops = GlobalCounters.global_ops / 1e9 / sec
+      gflops = GlobalCounters.global_ops / 1e9 / dev_time
+      mfu = ((6 * num_params * SEQLEN * GBS) / (dev_time * max(getenv("DP", 1), getenv("MP", 1)) * 2.3e15)) * 100
       tqdm.write(
-        f"{i:5} {sec:.2f} s run, {loss:.4f} loss, {lr:.12f} LR, {mem_gb:.2f} GB used, {gflops:9.2f} GFLOPS")
-
-      if (fname:=getenv("LOSS_FILE", "")):
-        with open(fname, "a") as f:
-          f.write(f"{i} {loss:.4f} {lr:.12f} {mem_gb:.2f}\n")
+          f"{i:5} {step_time:.3f} s step, {gbs_time:.3f} s gbs, {optim_time:.3f} s optim, {data_time:.3f} s data, {loss:.4f} loss, " \
+          f"{lr:.12f} LR, {grad_norm:.6f} grad_norm, {mem_gb:.2f} GB used, {gflops:9.2f} GFLOPS, {mfu:5.2f}% MFU")
+      if DEBUG >= 1: tqdm.write("  mem per device: " + ', '.join(f"{dev}: {mem/1e9:.2f} GB" for dev, mem in sorted(GlobalCounters.mem_used_per_device.items())))
 
       if WANDB:
-        wandb.log({"lr": lr, "train/loss": loss, "train/step_time": sec, "train/GFLOPS": gflops, "train/sequences_seen": sequences_seen})
+        wandb.log({
+          "train/loss": loss,
+          "train/lr": lr,
+          "train/grad_norm": grad_norm,
+          "train/step_time": step_time,
+          "train/gbs_time": gbs_time,
+          "train/optim_time": optim_time,
+          "train/dev_time": dev_time,
+          "train/data_time": data_time,
+          "train/mem": mem_gb,
+          "train/GFLOPS": gflops,
+          "train/MFU": mfu,
+          "train/sequences_seen": sequences_seen
+        })
 
       if (ckpt_freq := getenv("CKPT")) and (i % ckpt_freq == 0 and (i != 1 or ckpt_freq == 1)):
         tqdm.write("saving checkpoint")
@@ -1475,16 +1545,29 @@ def train_llama3():
         fn = f"{ckpt_dir}/llama3_{i}_optim.safe"
         safe_save(get_state_dict(scheduler), fn)
 
-    if sequences_seen % EVAL_FREQ == 0 and (i != 1 or EVAL_FREQ == 1):
+      if i == BENCHMARK:
+        median_step_time = sorted(step_times)[(BENCHMARK + 1) // 2]
+        estimated_total_minutes = int(median_step_time * (SAMPLES // GBS) / 60)
+        print(f"Estimated training time: {estimated_total_minutes // 60}h{estimated_total_minutes % 60}m")
+        print(f"epoch global_ops: {GlobalCounters.global_ops:_}, "
+              f"epoch global_mem: {GlobalCounters.global_mem:_}")
+
+    if (sequences_seen % EVAL_FREQ == 0 and (i != 1 or EVAL_FREQ == 1)) or (BENCHMARK and i == BENCHMARK):
+      if EVAL_BS == 0: return
       tqdm.write(f"evaluating after {sequences_seen} sequences")
+      profile_marker(f"eval @ {i}")
 
       # run eval
       eval_losses = []
       eval_iter = get_eval_iter()
-      tqdm.write(f"evaluating {5760//EVAL_BS} batches of {EVAL_BS} sequences")
+      tqdm.write(f"evaluating {EVAL_SAMPLES//EVAL_BS} batches of {EVAL_BS} sequences")
+
+      for j,tokens in tqdm(enumerate(eval_iter), total=EVAL_SAMPLES//EVAL_BS):
+        eval_losses += eval_step(tokens).tolist()
+
+        if BENCHMARK and (j+1) == min(BENCHMARK, EVAL_SAMPLES//EVAL_BS):
+          return
 
-      for tokens in tqdm(eval_iter, total=5760//EVAL_BS):
-        eval_losses += eval_step(model, tokens).tolist()
       log_perplexity = Tensor(eval_losses).mean().float().item()
 
       tqdm.write(f"eval log perplexity: {log_perplexity:.4f}")
@@ -1570,7 +1653,7 @@ def train_stable_diffusion():
     loss, out_lr = loss.detach().to("CPU"), optimizer.lr.to("CPU")
     Tensor.realize(loss, out_lr)
     return loss, out_lr
-    
+
   # checkpointing takes ~9 minutes without this, and ~1 minute with this
   @TinyJit
   def ckpt_to_cpu():
@@ -1609,7 +1692,7 @@ def train_stable_diffusion():
     if i == 3:
       for _ in range(3): ckpt_to_cpu() # do this at the beginning of run to prevent OOM surprises when checkpointing
       print("BEAM COMPLETE", flush=True) # allows wrapper script to detect BEAM search completion and retry if it failed
-      
+
     total_train_time = time.perf_counter() - train_start_time
     if WANDB:
       wandb.log({"train/loss": loss_item, "train/lr": lr_item, "train/loop_time_prev": loop_time, "train/dl_time": dl_time, "train/step": i,

examples/mlperf/optim.py

@@ -0,0 +1,56 @@
+from tinygrad.tensor import Tensor
+from tinygrad.dtype import dtypes
+from tinygrad.nn.optim import Optimizer
+from tinygrad.helpers import FUSE_OPTIM
+
+class GradAccClipAdamW(Optimizer):
+  def __init__(self, params:list[Tensor], lr=0.001, b1=0.9, b2=0.999, eps=1e-6, weight_decay=0.0, grad_acc=1, clip_norm=1.0, device=None, fused=FUSE_OPTIM):
+    super().__init__(params, lr, device, fused)
+    self.b1, self.b2, self.eps, self.wd = b1, b2, eps, weight_decay
+    self.b1_t, self.b2_t = (Tensor.ones((1,), dtype=dtypes.float32, device=self.device, requires_grad=False) for _ in [b1, b2])
+    self.m = self._new_optim_param()
+    self.v = self._new_optim_param()
+    self.grad_acc, self.clip_norm = grad_acc, clip_norm
+
+  def fstep(self, grads:list[Tensor]):
+    if self.fused:
+      out, extra = self._step([], grads)
+      updates = [out[0][self.pos_params[i]:self.pos_params[i+1]].reshape(tt.shape) for i, tt in enumerate(self.params)]
+    else:
+      updates, extra = self._step([], grads)
+    for i, tt in enumerate(self.params): tt.assign(self._apply_update(tt, updates[i]))
+    to_realize = extra+self.params+self.buffers
+
+    Tensor.realize(*to_realize)
+    return extra[-1]
+
+  def _step(self, params:list[Tensor], grads:list[Tensor]) -> tuple[list[Tensor], list[Tensor]]:
+    for i in range(len(grads)):
+      if grads[i].device != self.m[i].device: grads[i].assign(grads[i].to(self.m[i].device))
+
+    if self.fused:
+      grads[0].assign(grads[0] / self.grad_acc)
+      total_norm = grads[0].float().square().sum().sqrt()
+      grads[0].assign((grads[0] * (self.clip_norm / (total_norm + 1e-6)).clamp(max_=1.0)).cast(grads[0].dtype))
+    else:
+      for i in range(len(grads)):
+        grads[i].assign(grads[i] / self.grad_acc).realize()
+      total_norm = Tensor.stack(*[g.float().square().sum() for g in grads]).sum().sqrt().contiguous().realize()
+      for i in range(len(grads)):
+        grads[i].assign((grads[i] * (self.clip_norm / (total_norm + 1e-6)).clamp(max_=1.0)).cast(grads[i].dtype)).realize()
+
+    ret = []
+    self.b1_t *= self.b1
+    self.b2_t *= self.b2
+    for i, g in enumerate(grads):
+      self.m[i].assign((self.b1 * self.m[i] + (1.0 - self.b1) * g).cast(self.m[i].dtype))
+      self.v[i].assign((self.b2 * self.v[i] + (1.0 - self.b2) * (g * g)).cast(self.v[i].dtype))
+      m_hat = self.m[i] / (1.0 - self.b1_t)
+      v_hat = self.v[i] / (1.0 - self.b2_t)
+      up = m_hat / (v_hat.sqrt() + self.eps)
+      ret.append((self.lr * up).cast(g.dtype))
+    return ret, [self.b1_t, self.b2_t] + self.m + self.v + [total_norm]
+
+  def _apply_update(self, t:Tensor, up:Tensor) -> Tensor:
+    up = up.shard_like(t) + self.lr.to(t.device) * self.wd * t.detach()
+    return t.detach() - up.cast(t.dtype)

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/bert/implementations/tinybox_1xMI300X/dev_beam.sh

@@ -4,7 +4,7 @@ export PYTHONPATH="." AMD=1
 export MODEL="bert"
 export DEFAULT_FLOAT="HALF" GPUS=1 BS=128 EVAL_BS=128
 
-export IGNORE_OOB=1
+export CHECK_OOB=0
 
 export BEAM=3 BEAM_UOPS_MAX=4000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
 export IGNORE_JIT_FIRST_BEAM=1

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/bert/implementations/tinybox_8xMI300X/dev_beam.sh

@@ -5,7 +5,7 @@ export MODEL="bert"
 export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
 export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
 
-export IGNORE_OOB=1
+export CHECK_OOB=0
 export REWRITE_STACK_LIMIT=500000
 
 export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/bert/implementations/tinybox_8xMI300X/dev_run.sh

@@ -8,7 +8,7 @@ export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
 export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
 export TRAIN_STEPS=3900
 
-export IGNORE_OOB=1
+export CHECK_OOB=0
 export REWRITE_STACK_LIMIT=500000
 
 export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/bert/implementations/tinybox_8xMI300X/run_and_time.sh

@@ -11,7 +11,7 @@ export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
 export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
 export TRAIN_STEPS=3900
 
-export IGNORE_OOB=1
+export CHECK_OOB=0
 export REWRITE_STACK_LIMIT=500000
 
 export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/bert/implementations/tinybox_8xMI350X/dev_run_fa.sh

@@ -0,0 +1,20 @@
+#!/bin/bash
+
+export PYTHONPATH="." AMD=1 DEBUG=0 JIT=1 FLASH_ATTENTION=1
+export MODEL="bert"
+export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
+
+# similar to https://github.com/mlcommons/training_results_v3.1/blob/d06288b2bd675a9d88e0e6181f5bb5626b71ec19/Quanta_Cloud_Technology/results/D54U-3U/bert/result_1.txt#L54
+export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
+export TRAIN_STEPS=3900
+
+export CHECK_OOB=0
+export REWRITE_STACK_LIMIT=5000000
+
+export BEAM=0 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
+export IGNORE_JIT_FIRST_BEAM=1 FREE_INTERMEDIATE=0
+export BASEDIR="/raid/datasets/wiki"
+
+export WANDB=1 PARALLEL=0
+
+RUNMLPERF=1 python3 examples/mlperf/model_train.py

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/bert/implementations/tinybox_8xMI350X/dev_run_fp8.sh

@@ -0,0 +1,24 @@
+#!/bin/bash
+
+export PYTHONPATH="." AMD=1
+export MODEL="bert"
+export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
+
+# similar to https://github.com/mlcommons/training_results_v3.1/blob/d06288b2bd675a9d88e0e6181f5bb5626b71ec19/Quanta_Cloud_Technology/results/D54U-3U/bert/result_1.txt#L54
+export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
+export TRAIN_STEPS=3900
+
+export CHECK_OOB=0
+export REWRITE_STACK_LIMIT=5000000
+
+export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
+export IGNORE_JIT_FIRST_BEAM=1 FREE_INTERMEDIATE=0
+export BASEDIR="/raid/datasets/wiki"
+export BEAM_TIMEOUT_SEC=15
+export FP8_TRAIN=1
+# search
+IGNORE_BEAM_CACHE=1 BENCHMARK=10 BERT_LAYERS=2 RUNMLPERF=0 python3 examples/mlperf/model_train.py
+
+export WANDB=1 PARALLEL=0
+
+RUNMLPERF=1 python3 examples/mlperf/model_train.py

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/bert/implementations/tinybox_8xMI350X/run_and_time.sh

@@ -11,7 +11,7 @@ export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
 export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
 export TRAIN_STEPS=3900
 
-export IGNORE_OOB=1
+export CHECK_OOB=0
 export REWRITE_STACK_LIMIT=5000000
 
 export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/bert/implementations/tinybox_green/dev_beam.sh

@@ -4,7 +4,7 @@ export PYTHONPATH="." NV=1
 export MODEL="bert"
 export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=72 EVAL_BS=72
 
-export IGNORE_OOB=1
+export CHECK_OOB=0
 export REWRITE_STACK_LIMIT=500000
 
 export BEAM=8 BEAM_UOPS_MAX=10000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/bert/implementations/tinybox_green/dev_run.sh

@@ -4,7 +4,7 @@ export PYTHONPATH="." NV=1
 export MODEL="bert"
 export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=72 EVAL_BS=72
 
-export IGNORE_OOB=1
+export CHECK_OOB=0
 export REWRITE_STACK_LIMIT=500000
 
 export BEAM=8 BEAM_UOPS_MAX=10000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/bert/implementations/tinybox_green/run_and_time.sh

@@ -7,7 +7,7 @@ export MODEL="bert"
 export SUBMISSION_PLATFORM="tinybox_green"
 export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=72 EVAL_BS=72
 
-export IGNORE_OOB=1
+export CHECK_OOB=0
 export REWRITE_STACK_LIMIT=500000
 
 export BEAM=8 BEAM_UOPS_MAX=10000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/bert/implementations/tinybox_red/dev_beam.sh

@@ -4,7 +4,7 @@ export PYTHONPATH="." AMD=1
 export MODEL="bert"
 export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
 
-export IGNORE_OOB=1
+export CHECK_OOB=0
 export REWRITE_STACK_LIMIT=500000
 
 export BEAM=5 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/bert/implementations/tinybox_red/dev_run.sh

@@ -4,7 +4,7 @@ export PYTHONPATH="." AMD=1
 export MODEL="bert"
 export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
 
-export IGNORE_OOB=1
+export CHECK_OOB=0
 export REWRITE_STACK_LIMIT=500000
 
 export BEAM=5 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/bert/implementations/tinybox_red/run_and_time.sh

@@ -7,7 +7,7 @@ export MODEL="bert"
 export SUBMISSION_PLATFORM="tinybox_red"
 export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
 
-export IGNORE_OOB=1
+export CHECK_OOB=0
 export REWRITE_STACK_LIMIT=500000
 
 export BEAM=5 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_beam.sh

@@ -0,0 +1,41 @@
+#!/usr/bin/env bash
+
+export PYTHONPATH="."
+export DEV=${DEV:-AMD}
+export EMULATE="AMD_CDNA4"
+export CHECK_OOB=0
+export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
+
+export DEBUG=${DEBUG:-2}
+export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
+export ALL2ALL=${ALL2ALL:-1}
+export USE_ATOMICS=${USE_ATOMICS:-1}
+export ASM_GEMM=${ASM_GEMM:-1}
+export WQKV=${WQKV:-0}
+
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
+export DP=${DP:-8} BS=${BS:-8} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
+export GBS=$((BS * GRADIENT_ACC_STEPS))
+
+export MODEL="llama3"
+export BASEDIR="/raid/datasets/c4-8b/"
+export SMALL=1
+export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
+export EVAL_TARGET=3.3 EVAL_FREQ=12288
+export LR="4e-4" END_LR="4e-5" WARMUP_SAMPLES=256 MAX_STEPS=1200000
+export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
+export SAMPLES=$((MAX_STEPS * GBS))
+export SEQLEN=${SEQLEN:-8192}
+
+export SEED=${SEED:-5760}
+export DATA_SEED=${DATA_SEED:-5760}
+
+export JITBEAM=${JITBEAM:-3}
+export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
+
+export FAKEDATA=1 BENCHMARK=10
+if [ -z "$FULL_LAYERS" ]; then
+  export LLAMA_LAYERS=2
+fi
+
+python3 examples/mlperf/model_train.py

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_run.sh

@@ -0,0 +1,37 @@
+#!/usr/bin/env bash
+
+export PYTHONPATH="."
+export DEV=${DEV:-AMD}
+export EMULATE="AMD_CDNA4"
+export CHECK_OOB=0
+export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
+export DEVICE_IN_FUNCTION_BUG=1
+
+export DEBUG=${DEBUG:-0}
+export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
+export ALL2ALL=${ALL2ALL:-1}
+export USE_ATOMICS=${USE_ATOMICS:-1}
+export ASM_GEMM=${ASM_GEMM:-1}
+export WQKV=${WQKV:-0}
+
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
+export DP=${DP:-8} BS=${BS:-8} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
+export GBS=$((BS * GRADIENT_ACC_STEPS))
+
+export MODEL="llama3"
+export BASEDIR="/raid/datasets/c4-8b/"
+export SMALL=1
+export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
+export EVAL_TARGET=3.3 EVAL_FREQ=12288
+export LR="4e-4" END_LR="4e-5" WARMUP_SAMPLES=256 MAX_STEPS=1200000
+export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
+export SAMPLES=$((MAX_STEPS * GBS))
+export SEQLEN=${SEQLEN:-8192}
+
+export SEED=${SEED:-$RANDOM}
+export DATA_SEED=${DATA_SEED:-5760}
+
+export JITBEAM=${JITBEAM:-3}
+export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
+
+python3 examples/mlperf/model_train.py

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/profile.sh

@@ -0,0 +1,6 @@
+#!/bin/bash
+export BENCHMARK=5
+export EVAL_BS=0
+export VIZ=${VIZ:--1}
+examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_run.sh
+extra/viz/cli.py --profile --device "AMD" --top 20

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/startup_walltime.sh

@@ -0,0 +1,10 @@
+#!/bin/bash
+export BENCHMARK=5
+export EVAL_BS=0
+export FAKEDATA=1
+export NULL_ALLOW_COPYOUT=1
+export HIP_VISIBLE_DEVICES=""
+export DEV=NULL
+export JITBEAM=0
+export LLAMA_LAYERS=${LLAMA_LAYERS:-"2"}
+time examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_run.sh

examples/openpilot/load_pickle.py

@@ -0,0 +1,16 @@
+import sys, pickle
+from extra.bench_log import WallTimeEvent, BenchEvent
+from tinygrad.helpers import getenv
+
+PKL = sys.argv[1] if len(sys.argv) > 1 else "/tmp/openpilot.pkl"
+
+load_times = []
+
+for _ in range(10):
+  with WallTimeEvent(BenchEvent.STEP) as wte: pickle.load(open(PKL, 'rb'))
+  load_times.append(wte.time)
+  print(f"pickle load: {wte.time:6.2f} s")
+
+if (assert_time:=getenv("ASSERT_MIN_LOAD_TIME")):
+  min_time = min(load_times)
+  assert min_time < assert_time, f"Speed regression, expected min load time of < {assert_time} s but took: {min_time} s"

examples/stable_diffusion.py

@@ -6,7 +6,6 @@ import argparse, time
 from collections import namedtuple
 from typing import Dict, Any
 
-from PIL import Image
 import numpy as np
 from tinygrad import Device, GlobalCounters, dtypes, Tensor, TinyJit
 from tinygrad.helpers import Timing, Context, getenv, fetch, colored, tqdm, flatten, profile_marker
@@ -336,6 +335,7 @@ if __name__ == "__main__":
   print(x.shape)
 
   profile_marker("save image")
+  from PIL import Image
   im = Image.fromarray(x.numpy())
   print(f"saving {args.out}")
   im.save(args.out)

examples/test_pkl_imagenet.py

@@ -7,7 +7,7 @@ if __name__ == "__main__":
   with open(fetch(sys.argv[1]), "rb") as f:
     run_onnx_jit = pickle.load(f)
   input_name = run_onnx_jit.captured.expected_names[0]
-  device = run_onnx_jit.captured.expected_st_vars_dtype_device[0][-1]
+  device = run_onnx_jit.captured.expected_input_info[0][-1]
   print(f"input goes into {input_name=} on {device=}")
   hit = 0
   for i,(img,y) in enumerate(imagenet_dataloader(cnt=getenv("CNT", 100))):

examples/tinychat/tinychat-browser/compile.py

@@ -48,7 +48,7 @@ def prepare_browser_chunks(model):
         weight_metadata = metadata.get(name, default)
         weight_metadata["parts"][part_num] = {"file": i, "file_start_pos": cursor, "size": size}
         metadata[name] = weight_metadata
-        data = bytes(state_dict[name].uop.base.realized.as_buffer())
+        data = bytes(state_dict[name].uop.base.realized.as_memoryview())
         data = data if not offsets else data[offsets[0]:offsets[1]]
         writer.write(data)
         cursor += size

examples/webgpu/stable_diffusion/compile.py

@@ -93,7 +93,7 @@ if __name__ == "__main__":
     forward: Any = None
 
   sub_steps = [
-    Step(name = "textModel", input = [Tensor.randn(1, 77)], forward = model.cond_stage_model.transformer.text_model),
+    Step(name = "textModel", input = [Tensor.randint(1, 77, low=0, high=49408, dtype=dtypes.int32)], forward = model.cond_stage_model.transformer.text_model),
     Step(name = "diffusor", input = [Tensor.randn(1, 77, 768), Tensor.randn(1, 77, 768), Tensor.randn(1,4,64,64), Tensor.rand(1), Tensor.randn(1), Tensor.randn(1), Tensor.randn(1)], forward = model),
     Step(name = "decoder", input = [Tensor.randn(1,4,64,64)], forward = model.decode),
     Step(name = "f16tof32", input = [Tensor.randn(2097120, dtype=dtypes.uint32)], forward = u32_to_f16)

examples/whisper.py

@@ -7,6 +7,7 @@ from tinygrad import Tensor, TinyJit, Variable, nn, dtypes
 from tinygrad.nn.state import torch_load, load_state_dict
 from tinygrad.helpers import getenv, fetch
 
+from examples.audio_helpers import mel
 import numpy as np
 import librosa
 
@@ -159,7 +160,7 @@ def prep_audio(waveforms: List[np.ndarray], batch_size: int, truncate=False) ->
 
   stft = librosa.stft(waveforms, n_fft=N_FFT, hop_length=HOP_LENGTH, window='hann', dtype=np.csingle)
   magnitudes = np.absolute(stft[..., :-1]) ** 2
-  mel_spec = librosa.filters.mel(sr=RATE, n_fft=N_FFT, n_mels=N_MELS) @ magnitudes
+  mel_spec = mel(sr=RATE, n_fft=N_FFT, n_mels=N_MELS).numpy() @ magnitudes
 
   log_spec = np.log10(np.clip(mel_spec, 1e-10, None))
   log_spec = np.maximum(log_spec, log_spec.max((1,2), keepdims=True) - 8.0)

extra/amdpci/am_smi.py

@@ -92,7 +92,7 @@ class SMICtx:
     self.prev_terminal_width = 0
     self.prev_terminal_height = 0
 
-    remove_parts = ["Advanced Micro Devices, Inc. [AMD/ATI]", "VGA compatible controller:"]
+    remove_parts = ["Advanced Micro Devices, Inc. [AMD/ATI]", "VGA compatible controller:", "Processing accelerators:"]
     lspci = subprocess.check_output(["lspci"]).decode("utf-8").splitlines()
     self.lspci = {l.split()[0]: l.split(" ", 1)[1] for l in lspci}
     for k,v in self.lspci.items():
@@ -153,8 +153,8 @@ class SMICtx:
     tables = {}
     for dev in self.devs:
       match dev.ip_ver[am.MP1_HWIP]:
-        case (13,0,6): table_t = dev.smu.smu_mod.MetricsTableX_t
-        case (13,0,12): table_t = dev.smu.smu_mod.MetricsTableV2_t
+        case (13,0,6): table_t = dev.smu.smu_mod.MetricsTableV0_t
+        case (13,0,12): table_t = dev.smu.smu_mod.MetricsTable_t
         case _: table_t = dev.smu.smu_mod.SmuMetricsExternal_t
       tables[dev] = dev.smu.read_table(table_t, dev.smu.smu_mod.SMU_TABLE_SMU_METRICS) if dev.pci_state == "D0" else None
     return tables
@@ -165,17 +165,17 @@ class SMICtx:
 
   def get_gfx_activity(self, dev, metrics):
     match dev.ip_ver[am.MP1_HWIP]:
-      case (13,0,6): return max(0, min(100, self._smuq10_round(metrics.SocketGfxBusy)))
+      case (13,0,6)|(13,0,12): return max(0, min(100, self._smuq10_round(metrics.SocketGfxBusy)))
       case _: return metrics.SmuMetrics.AverageGfxActivity
 
   def get_mem_activity(self, dev, metrics):
     match dev.ip_ver[am.MP1_HWIP]:
-      case (13,0,6): return max(0, min(100, self._smuq10_round(metrics.DramBandwidthUtilization)))
+      case (13,0,6)|(13,0,12): return max(0, min(100, self._smuq10_round(metrics.DramBandwidthUtilization)))
       case _: return metrics.SmuMetrics.AverageUclkActivity
 
   def get_temps(self, dev, metrics, compact=False):
     match dev.ip_ver[am.MP1_HWIP]:
-      case (13,0,6):
+      case (13,0,6)|(13,0,12):
         temps = {
           "Hotspot": self._smuq10_round(metrics.MaxSocketTemperature),
           "HBM": self._smuq10_round(metrics.MaxHbmTemperature),
@@ -191,7 +191,7 @@ class SMICtx:
 
   def get_voltage(self, dev, metrics, compact=False):
     match dev.ip_ver[am.MP1_HWIP]:
-      case (13,0,6): return {}
+      case (13,0,6)|(13,0,12): return {}
       case _:
         voltage_keys = [(k, name) for k, name in dev.smu.smu_mod.SVI_PLANE_e.items()
                         if k < dev.smu.smu_mod.SVI_PLANE_COUNT and metrics.SmuMetrics.AvgVoltage[k] != 0]
@@ -205,38 +205,37 @@ class SMICtx:
   def get_gfx_freq(self, dev, metrics):
     if metrics is None: return 0
     match dev.ip_ver[am.MP1_HWIP]:
-      case (13,0,6): return self._smuq10_round(metrics.GfxclkFrequency[0])
+      case (13,0,6)|(13,0,12): return self._smuq10_round(metrics.GfxclkFrequency[0])
       case _:
         return metrics.SmuMetrics.AverageGfxclkFrequencyPostDs if self.get_gfx_activity(dev, metrics) <= self.get_busy_threshold(dev) else \
                metrics.SmuMetrics.AverageGfxclkFrequencyPreDs
 
   def get_mem_freq(self, dev, metrics):
     match dev.ip_ver[am.MP1_HWIP]:
-      case (13,0,6): return self._smuq10_round(metrics.UclkFrequency)
+      case (13,0,6)|(13,0,12): return self._smuq10_round(metrics.UclkFrequency)
       case _:
         return metrics.SmuMetrics.AverageMemclkFrequencyPostDs if self.get_mem_activity(dev, metrics) <= self.get_busy_threshold(dev) else \
                metrics.SmuMetrics.AverageMemclkFrequencyPreDs
 
   def get_fckl_freq(self, dev, metrics):
     match dev.ip_ver[am.MP1_HWIP]:
-      case (13,0,6): return self._smuq10_round(metrics.FclkFrequency)
+      case (13,0,6)|(13,0,12): return self._smuq10_round(metrics.FclkFrequency)
       case _:
         return metrics.SmuMetrics.AverageFclkFrequencyPostDs if self.get_mem_activity(dev, metrics) <= self.get_busy_threshold(dev) else \
                metrics.SmuMetrics.AverageFclkFrequencyPreDs
 
   def get_fan_rpm_pwm(self, dev, metrics):
     match dev.ip_ver[am.MP1_HWIP]:
-      case (13,0,6): return None, None
+      case (13,0,6)|(13,0,12): return None, None
       case _: return metrics.SmuMetrics.AvgFanRpm, metrics.SmuMetrics.AvgFanPwm
 
   def get_power(self, dev, metrics):
     match dev.ip_ver[am.MP1_HWIP]:
       case (13,0,6): return self._smuq10_round(metrics.SocketPower), self._smuq10_round(metrics.MaxSocketPowerLimit)
+      case (13,0,12): return self._smuq10_round(metrics.SocketPower), self._smuq10_round(metrics.SocketPowerLimit)
       case _: return metrics.SmuMetrics.AverageSocketPower, metrics.SmuMetrics.dGPU_W_MAX
 
   def get_mem_usage(self, dev):
-    return 0
-
     usage = 0
     pt_stack = [dev.mm.root_page_table]
     while len(pt_stack) > 0:
@@ -245,8 +244,8 @@ class SMICtx:
         entry = pt.entries[i]
 
         if (entry & am.AMDGPU_PTE_VALID) == 0: continue
-        if pt.lv!=am.AMDGPU_VM_PTB and not dev.gmc.is_pte_huge_page(pt.lv, entry):
-          pt_stack.append(AMPageTableEntry(dev, entry & 0x0000FFFFFFFFF000, lv=pt.lv+1))
+        if pt.lv < am.AMDGPU_VM_PDB0 and not dev.gmc.is_pte_huge_page(pt.lv, entry):
+          pt_stack.append(AMPageTableEntry(dev, dev.xgmi2paddr(entry & 0x0000FFFFFFFFF000), lv=pt.lv+1))
           continue
         if (entry & am.AMDGPU_PTE_SYSTEM) != 0: continue
         usage += (1 << ((9 * (3-pt.lv)) + 12))
@@ -280,7 +279,7 @@ class SMICtx:
       device_line = [f"{bold(dev.pcibus)} {trim(self.lspci[dev.pcibus[5:]], col_size - 20)}"] + [pad("", col_size)]
       activity_line = [f"GFX Activity {draw_bar(self.get_gfx_activity(dev, metrics) / 100, activity_line_width)}"] \
                     + [f"MEM Activity {draw_bar(self.get_mem_activity(dev, metrics) / 100, activity_line_width)}"] \
-                    + [f"MEM Usage    {draw_bar((mem_used / mem_total) / 100, activity_line_width, opt_text=mem_fmt)}"] \
+                    + [f"MEM Usage    {draw_bar(mem_used / mem_total, activity_line_width, opt_text=mem_fmt)}"] \
 
       temps_data, temps_data_compact = self.get_temps(dev, metrics), self.get_temps(dev, metrics, compact=True)
       temps_table = ["=== Temps (°C) ==="] + [f"{name:<16}: {color_temp(val)}" for name, val in temps_data.items()]

extra/amdpci/hive_reset.py

@@ -1,12 +1,18 @@
 #!/usr/bin/env python3
 
+import os
 from tinygrad.helpers import Context
 from tinygrad.runtime.support.system import System, PCIDevice, PCIDevImplBase
+from tinygrad.runtime.support.hcq import FileIOInterface
 from tinygrad.runtime.support.am.amdev import AMDev
 
 if __name__ == "__main__":
   gpus = System.pci_scan_bus(0x1002, [(0xffff, [0x74a1, 0x75a0])])
-  pcidevs = [PCIDevice(f"reset:{gpu}", gpu, bars=[0, 2, 5]) for gpu in gpus]
+  for gpu in gpus:
+    drv_path = f"/sys/bus/pci/devices/{gpu}/driver"
+    if FileIOInterface.exists(drv_path) and os.path.basename(os.readlink(drv_path)) == "amdgpu":
+      raise RuntimeError(f"amdgpu is bound to {gpu}. Stopping...")
+  pcidevs = [PCIDevice("AM", gpu, bars=[0, 2, 5]) for gpu in gpus]
   amdevs = []
   with Context(DEBUG=2):
     for pcidev in pcidevs:

extra/amdpci/proclogs.py

@@ -19,8 +19,9 @@ amdev = importlib.import_module("tinygrad.runtime.support.am.amdev")
 amdev.AMDev = AMDFake
 from tinygrad.runtime.ops_amd import PCIIface
 
-def parse_amdgpu_logs(log_content, register_names=None, *, only_xcc0: bool = False):
+def parse_amdgpu_logs(log_content, register_names=None, register_objects=None, *, only_xcc0: bool = False):
   register_map = register_names or {}
+  register_objs = register_objects or {}
 
   def replace_register(match):
     reg = match.group(1)
@@ -37,6 +38,28 @@ def parse_amdgpu_logs(log_content, register_names=None, *, only_xcc0: bool = Fal
   # remove timing prefix
   processed_log = re.sub(r'^\[\s*\d+(?:\.\d+)?\]\s*', '', processed_log, flags=re.MULTILINE)
 
+  # decode register values into field dicts
+  def decode_value(match):
+    reg_name = match.group(1)
+    xcc_part = match.group(2)  # "xcc=0 " or ""
+    val_str = match.group(3)
+    val = int(val_str, 16)
+
+    reg_obj = register_objs.get(reg_name)
+    if reg_obj is not None and reg_obj.fields:
+      fields = reg_obj.decode(val)
+      # show raw for unaccounted bits
+      accounted = 0
+      for name, (start, end) in reg_obj.fields.items():
+        accounted |= (((1 << (end - start + 1)) - 1) << start)
+      unaccounted = val & ~accounted
+      parts = {k: v for k, v in fields.items() if v != 0}
+      if unaccounted: parts['_raw_unaccounted'] = hex(unaccounted)
+      return f"register {reg_name}, {xcc_part}with value {val_str} {parts}"
+    return match.group(0)
+
+  processed_log = re.sub(r'register (reg\w+), ((?:xcc=\d+ )?)with value (0x[0-9a-fA-F]+)', decode_value, processed_log)
+
   # keep only xcc=0 lines (but keep lines with no xcc at all)
   if only_xcc0:
     kept = []
@@ -50,16 +73,18 @@ def main():
   only_xcc0 = bool(getenv("ONLY_XCC0", 0))
 
   reg_names = {}
+  reg_objs = {}
   dev = PCIIface(None, 0)
   for x, y in dev.dev_impl.__dict__.items():
     if isinstance(y, AMRegister):
       for xcc, addr in y.addr.items():
         reg_names[addr] = f"{x}, xcc={xcc}"
+      reg_objs[x] = y
 
   with open(sys.argv[1], 'r') as f:
     log_content = f.read()
 
-  processed_log = parse_amdgpu_logs(log_content, reg_names, only_xcc0=only_xcc0)
+  processed_log = parse_amdgpu_logs(log_content, reg_names, reg_objs, only_xcc0=only_xcc0)
 
   with open(sys.argv[2], 'w') as f:
     f.write(processed_log)

extra/assembly/amd/asm.py

@@ -1,605 +0,0 @@
-# RDNA3 assembler and disassembler
-from __future__ import annotations
-import re
-from extra.assembly.amd.dsl import Inst, RawImm, Reg, SGPR, VGPR, TTMP, s, v, ttmp, _RegFactory, FLOAT_ENC, SRC_FIELDS, unwrap
-
-# Decoding helpers
-SPECIAL_GPRS = {106: "vcc_lo", 107: "vcc_hi", 124: "null", 125: "m0", 126: "exec_lo", 127: "exec_hi", 253: "scc"}
-SPECIAL_DEC = {**SPECIAL_GPRS, **{v: str(k) for k, v in FLOAT_ENC.items()}}
-SPECIAL_PAIRS = {106: "vcc", 126: "exec"}  # Special register pairs (for 64-bit ops)
-# GFX11 hwreg names (IDs 16-17 are TBA - not supported, IDs 18-19 are PERF_SNAPSHOT)
-HWREG_NAMES = {1: 'HW_REG_MODE', 2: 'HW_REG_STATUS', 3: 'HW_REG_TRAPSTS', 4: 'HW_REG_HW_ID', 5: 'HW_REG_GPR_ALLOC',
-               6: 'HW_REG_LDS_ALLOC', 7: 'HW_REG_IB_STS', 15: 'HW_REG_SH_MEM_BASES', 18: 'HW_REG_PERF_SNAPSHOT_PC_LO',
-               19: 'HW_REG_PERF_SNAPSHOT_PC_HI', 20: 'HW_REG_FLAT_SCR_LO', 21: 'HW_REG_FLAT_SCR_HI',
-               22: 'HW_REG_XNACK_MASK', 23: 'HW_REG_HW_ID1', 24: 'HW_REG_HW_ID2', 25: 'HW_REG_POPS_PACKER', 28: 'HW_REG_IB_STS2'}
-HWREG_IDS = {v.lower(): k for k, v in HWREG_NAMES.items()}  # Reverse map for assembler
-MSG_NAMES = {128: 'MSG_RTN_GET_DOORBELL', 129: 'MSG_RTN_GET_DDID', 130: 'MSG_RTN_GET_TMA',
-             131: 'MSG_RTN_GET_REALTIME', 132: 'MSG_RTN_SAVE_WAVE', 133: 'MSG_RTN_GET_TBA'}
-_16BIT_TYPES = ('f16', 'i16', 'u16', 'b16')
-def _is_16bit(s: str) -> bool: return any(s.endswith(x) for x in _16BIT_TYPES)
-
-def decode_src(val: int) -> str:
-  if val <= 105: return f"s{val}"
-  if val in SPECIAL_DEC: return SPECIAL_DEC[val]
-  if 108 <= val <= 123: return f"ttmp{val - 108}"
-  if 128 <= val <= 192: return str(val - 128)
-  if 193 <= val <= 208: return str(-(val - 192))
-  if 256 <= val <= 511: return f"v{val - 256}"
-  return "lit" if val == 255 else f"?{val}"
-
-def _reg(prefix: str, base: int, cnt: int = 1) -> str: return f"{prefix}{base}" if cnt == 1 else f"{prefix}[{base}:{base+cnt-1}]"
-def _sreg(base: int, cnt: int = 1) -> str: return _reg("s", base, cnt)
-def _vreg(base: int, cnt: int = 1) -> str: return _reg("v", base, cnt)
-
-def _fmt_sdst(v: int, cnt: int = 1) -> str:
-  """Format SGPR destination with special register names."""
-  if v == 124: return "null"
-  if 108 <= v <= 123: return _reg("ttmp", v - 108, cnt)
-  if cnt > 1 and v in SPECIAL_PAIRS: return SPECIAL_PAIRS[v]
-  if cnt > 1: return _sreg(v, cnt)
-  return {126: "exec_lo", 127: "exec_hi", 106: "vcc_lo", 107: "vcc_hi", 125: "m0"}.get(v, f"s{v}")
-
-def _fmt_ssrc(v: int, cnt: int = 1) -> str:
-  """Format SGPR source with special register names and pairs."""
-  if cnt == 2:
-    if v in SPECIAL_PAIRS: return SPECIAL_PAIRS[v]
-    if v <= 105: return _sreg(v, 2)
-    if 108 <= v <= 123: return _reg("ttmp", v - 108, 2)
-  return decode_src(v)
-
-def _fmt_src_n(v: int, cnt: int) -> str:
-  """Format source with given register count (1, 2, or 4)."""
-  if cnt == 1: return decode_src(v)
-  if v >= 256: return _vreg(v - 256, cnt)
-  if v <= 105: return _sreg(v, cnt)
-  if cnt == 2 and v in SPECIAL_PAIRS: return SPECIAL_PAIRS[v]
-  if 108 <= v <= 123: return _reg("ttmp", v - 108, cnt)
-  return decode_src(v)
-
-def _fmt_src64(v: int) -> str:
-  """Format 64-bit source (VGPR pair, SGPR pair, or special pair)."""
-  return _fmt_src_n(v, 2)
-
-def _parse_sop_sizes(op_name: str) -> tuple[int, ...]:
-  """Parse dst and src sizes from SOP instruction name. Returns (dst_cnt, src0_cnt) or (dst_cnt, src0_cnt, src1_cnt)."""
-  if op_name in ('s_bitset0_b64', 's_bitset1_b64'): return (2, 1)
-  if op_name in ('s_lshl_b64', 's_lshr_b64', 's_ashr_i64', 's_bfe_u64', 's_bfe_i64'): return (2, 2, 1)
-  if op_name in ('s_bfm_b64',): return (2, 1, 1)
-  # SOPC: s_bitcmp0_b64, s_bitcmp1_b64 - 64-bit src0, 32-bit src1 (bit index)
-  if op_name in ('s_bitcmp0_b64', 's_bitcmp1_b64'): return (1, 2, 1)
-  if m := re.search(r'_(b|i|u)(32|64)_(b|i|u)(32|64)$', op_name):
-    return (2 if m.group(2) == '64' else 1, 2 if m.group(4) == '64' else 1)
-  if m := re.search(r'_(b|i|u)(32|64)$', op_name):
-    sz = 2 if m.group(2) == '64' else 1
-    return (sz, sz)
-  return (1, 1)
-
-# Waitcnt helpers (RDNA3 format: bits 15:10=vmcnt, bits 9:4=lgkmcnt, bits 3:0=expcnt)
-def waitcnt(vmcnt: int = 0x3f, expcnt: int = 0x7, lgkmcnt: int = 0x3f) -> int:
-  return (expcnt & 0x7) | ((lgkmcnt & 0x3f) << 4) | ((vmcnt & 0x3f) << 10)
-def decode_waitcnt(val: int) -> tuple[int, int, int]:
-  return (val >> 10) & 0x3f, val & 0xf, (val >> 4) & 0x3f  # vmcnt, expcnt, lgkmcnt
-
-# VOP3SD opcodes (shared encoding with VOP3 but different field layout)
-# Note: opcodes 0-255 are VOPC promoted to VOP3 - never treat as VOP3SD
-VOP3SD_OPCODES = {288, 289, 290, 764, 765, 766, 767, 768, 769, 770}
-
-# Disassembler
-def disasm(inst: Inst) -> str:
-  op_val = unwrap(inst._values.get('op', 0))
-  cls_name = inst.__class__.__name__
-  # VOP3 and VOP3SD share encoding - check opcode to determine which
-  is_vop3sd = cls_name == 'VOP3' and op_val in VOP3SD_OPCODES
-  try:
-    from extra.assembly.amd.autogen import rdna3 as autogen
-    if is_vop3sd:
-      op_name = autogen.VOP3SDOp(op_val).name.lower()
-    else:
-      op_name = getattr(autogen, f"{cls_name}Op")(op_val).name.lower() if hasattr(autogen, f"{cls_name}Op") else f"op_{op_val}"
-  except (ValueError, KeyError): op_name = f"op_{op_val}"
-  def fmt_src(v): return f"0x{inst._literal:x}" if v == 255 and inst._literal is not None else decode_src(v)
-
-  # VOP1
-  if cls_name == 'VOP1':
-    vdst, src0 = unwrap(inst._values['vdst']), unwrap(inst._values['src0'])
-    if op_name == 'v_nop': return 'v_nop'
-    if op_name == 'v_pipeflush': return 'v_pipeflush'
-    parts = op_name.split('_')
-    is_16bit_dst = any(p in _16BIT_TYPES for p in parts[-2:-1]) or (len(parts) >= 2 and parts[-1] in _16BIT_TYPES and 'cvt' not in op_name)
-    is_16bit_src = parts[-1] in _16BIT_TYPES and 'sat_pk' not in op_name
-    _F64_OPS = ('v_ceil_f64', 'v_floor_f64', 'v_fract_f64', 'v_frexp_mant_f64', 'v_rcp_f64', 'v_rndne_f64', 'v_rsq_f64', 'v_sqrt_f64', 'v_trunc_f64')
-    is_f64_dst = op_name in _F64_OPS or op_name in ('v_cvt_f64_f32', 'v_cvt_f64_i32', 'v_cvt_f64_u32')
-    is_f64_src = op_name in _F64_OPS or op_name in ('v_cvt_f32_f64', 'v_cvt_i32_f64', 'v_cvt_u32_f64', 'v_frexp_exp_i32_f64')
-    if op_name == 'v_readfirstlane_b32':
-      return f"v_readfirstlane_b32 {decode_src(vdst)}, v{src0 - 256 if src0 >= 256 else src0}"
-    dst_str = _vreg(vdst, 2) if is_f64_dst else f"v{vdst & 0x7f}.{'h' if vdst >= 128 else 'l'}" if is_16bit_dst else f"v{vdst}"
-    src_str = _fmt_src64(src0) if is_f64_src else f"v{(src0 - 256) & 0x7f}.{'h' if src0 >= 384 else 'l'}" if is_16bit_src and src0 >= 256 else fmt_src(src0)
-    return f"{op_name}_e32 {dst_str}, {src_str}"
-
-  # VOP2
-  if cls_name == 'VOP2':
-    vdst, src0_raw, vsrc1 = unwrap(inst._values['vdst']), unwrap(inst._values['src0']), unwrap(inst._values['vsrc1'])
-    suffix = "" if op_name == "v_dot2acc_f32_f16" else "_e32"
-    is_16bit_op = ('_f16' in op_name or '_i16' in op_name or '_u16' in op_name) and '_f32' not in op_name and '_i32' not in op_name and 'pk_' not in op_name
-    if is_16bit_op:
-      dst_str = f"v{vdst & 0x7f}.{'h' if vdst >= 128 else 'l'}"
-      src0_str = f"v{(src0_raw - 256) & 0x7f}.{'h' if src0_raw >= 384 else 'l'}" if src0_raw >= 256 else fmt_src(src0_raw)
-      vsrc1_str = f"v{vsrc1 & 0x7f}.{'h' if vsrc1 >= 128 else 'l'}"
-    else:
-      dst_str, src0_str, vsrc1_str = f"v{vdst}", fmt_src(src0_raw), f"v{vsrc1}"
-    return f"{op_name}{suffix} {dst_str}, {src0_str}, {vsrc1_str}" + (", vcc_lo" if op_name == "v_cndmask_b32" else "")
-
-  # VOPC
-  if cls_name == 'VOPC':
-    src0, vsrc1 = unwrap(inst._values['src0']), unwrap(inst._values['vsrc1'])
-    is_64bit = any(x in op_name for x in ('f64', 'i64', 'u64'))
-    is_64bit_vsrc1 = is_64bit and 'class' not in op_name
-    is_16bit = any(x in op_name for x in ('_f16', '_i16', '_u16')) and 'f32' not in op_name
-    is_cmpx = op_name.startswith('v_cmpx')  # VOPCX writes to exec, no vcc destination
-    src0_str = _fmt_src64(src0) if is_64bit else f"v{(src0 - 256) & 0x7f}.{'h' if src0 >= 384 else 'l'}" if is_16bit and src0 >= 256 else fmt_src(src0)
-    vsrc1_str = _vreg(vsrc1, 2) if is_64bit_vsrc1 else f"v{vsrc1 & 0x7f}.{'h' if vsrc1 >= 128 else 'l'}" if is_16bit else f"v{vsrc1}"
-    return f"{op_name}_e32 {src0_str}, {vsrc1_str}" if is_cmpx else f"{op_name}_e32 vcc_lo, {src0_str}, {vsrc1_str}"
-
-  # SOPP
-  if cls_name == 'SOPP':
-    simm16 = unwrap(inst._values.get('simm16', 0))
-    # No-operand instructions (simm16 is ignored)
-    no_imm_ops = ('s_endpgm', 's_barrier', 's_wakeup', 's_icache_inv', 's_ttracedata', 's_ttracedata_imm',
-                  's_wait_idle', 's_endpgm_saved', 's_code_end', 's_endpgm_ordered_ps_done')
-    if op_name in no_imm_ops: return op_name
-    if op_name == 's_waitcnt':
-      vmcnt, expcnt, lgkmcnt = decode_waitcnt(simm16)
-      parts = []
-      if vmcnt != 0x3f: parts.append(f"vmcnt({vmcnt})")
-      if expcnt != 0x7: parts.append(f"expcnt({expcnt})")
-      if lgkmcnt != 0x3f: parts.append(f"lgkmcnt({lgkmcnt})")
-      return f"s_waitcnt {' '.join(parts)}" if parts else "s_waitcnt 0"
-    if op_name == 's_delay_alu':
-      dep_names = ['VALU_DEP_1','VALU_DEP_2','VALU_DEP_3','VALU_DEP_4','TRANS32_DEP_1','TRANS32_DEP_2','TRANS32_DEP_3','FMA_ACCUM_CYCLE_1','SALU_CYCLE_1','SALU_CYCLE_2','SALU_CYCLE_3']
-      skip_names = ['SAME','NEXT','SKIP_1','SKIP_2','SKIP_3','SKIP_4']
-      id0, skip, id1 = simm16 & 0xf, (simm16 >> 4) & 0x7, (simm16 >> 7) & 0xf
-      def dep_name(v): return dep_names[v-1] if 0 < v <= len(dep_names) else str(v)
-      parts = [f"instid0({dep_name(id0)})"] if id0 else []
-      if skip: parts.append(f"instskip({skip_names[skip]})")
-      if id1: parts.append(f"instid1({dep_name(id1)})")
-      return f"s_delay_alu {' | '.join(p for p in parts if p)}" if parts else "s_delay_alu 0"
-    if op_name.startswith('s_cbranch') or op_name.startswith('s_branch'):
-      return f"{op_name} {simm16}"
-    # Most SOPP ops require immediate (s_nop, s_setkill, s_sethalt, s_sleep, s_setprio, s_sendmsg*, etc.)
-    return f"{op_name} 0x{simm16:x}"
-
-  # SMEM
-  if cls_name == 'SMEM':
-    if op_name in ('s_gl1_inv', 's_dcache_inv'): return op_name
-    sdata, sbase, soffset, offset = unwrap(inst._values['sdata']), unwrap(inst._values['sbase']), unwrap(inst._values['soffset']), unwrap(inst._values.get('offset', 0))
-    glc, dlc = unwrap(inst._values.get('glc', 0)), unwrap(inst._values.get('dlc', 0))
-    # Format offset: "soffset offset:X" if both, "0x{offset:x}" if only imm, or decode_src(soffset)
-    off_str = f"{decode_src(soffset)} offset:0x{offset:x}" if offset and soffset != 124 else f"0x{offset:x}" if offset else decode_src(soffset)
-    sbase_idx, sbase_cnt = sbase * 2, 4 if (8 <= op_val <= 12 or op_name == 's_atc_probe_buffer') else 2
-    sbase_str = _fmt_ssrc(sbase_idx, sbase_cnt) if sbase_cnt == 2 else _sreg(sbase_idx, sbase_cnt) if sbase_idx <= 105 else _reg("ttmp", sbase_idx - 108, sbase_cnt)
-    if op_name in ('s_atc_probe', 's_atc_probe_buffer'): return f"{op_name} {sdata}, {sbase_str}, {off_str}"
-    width = {0:1, 1:2, 2:4, 3:8, 4:16, 8:1, 9:2, 10:4, 11:8, 12:16}.get(op_val, 1)
-    mods = [m for m in ["glc" if glc else "", "dlc" if dlc else ""] if m]
-    return f"{op_name} {_fmt_sdst(sdata, width)}, {sbase_str}, {off_str}" + (" " + " ".join(mods) if mods else "")
-
-  # FLAT
-  if cls_name == 'FLAT':
-    vdst, addr, data, saddr, offset, seg = [unwrap(inst._values.get(f, 0)) for f in ['vdst', 'addr', 'data', 'saddr', 'offset', 'seg']]
-    instr = f"{['flat', 'scratch', 'global'][seg] if seg < 3 else 'flat'}_{op_name.split('_', 1)[1] if '_' in op_name else op_name}"
-    width = {'b32':1, 'b64':2, 'b96':3, 'b128':4, 'u8':1, 'i8':1, 'u16':1, 'i16':1}.get(op_name.split('_')[-1], 1)
-    addr_str = _vreg(addr, 2) if saddr == 0x7F else _vreg(addr)
-    saddr_str = "" if saddr == 0x7F else f", {_sreg(saddr, 2)}" if saddr < 106 else ", off" if saddr == 124 else f", {decode_src(saddr)}"
-    off_str = f" offset:{offset}" if offset else ""
-    vdata_str = _vreg(data if 'store' in op_name else vdst, width)
-    return f"{instr} {addr_str}, {vdata_str}{saddr_str}{off_str}" if 'store' in op_name else f"{instr} {vdata_str}, {addr_str}{saddr_str}{off_str}"
-
-  # VOP3: vector ops with modifiers (can be 1, 2, or 3 sources depending on opcode range)
-  if cls_name == 'VOP3':
-    # Handle VOP3SD opcodes (same encoding, different field layout)
-    if is_vop3sd:
-      vdst = unwrap(inst._values.get('vdst', 0))
-      # VOP3SD: sdst is at bits [14:8], but VOP3 decodes opsel at [14:11], abs at [10:8], clmp at [15]
-      # We need to reconstruct sdst from these fields
-      opsel_raw = unwrap(inst._values.get('opsel', 0))
-      abs_raw = unwrap(inst._values.get('abs', 0))
-      clmp_raw = unwrap(inst._values.get('clmp', 0))
-      sdst = (clmp_raw << 7) | (opsel_raw << 3) | abs_raw
-      src0, src1, src2 = [unwrap(inst._values.get(f, 0)) for f in ('src0', 'src1', 'src2')]
-      neg = unwrap(inst._values.get('neg', 0))
-      omod = unwrap(inst._values.get('omod', 0))
-      omod_str = {1: " mul:2", 2: " mul:4", 3: " div:2"}.get(omod, "")
-      is_f64 = 'f64' in op_name
-      # v_mad_i64_i32/v_mad_u64_u32: 64-bit dst and src2, 32-bit src0/src1
-      is_mad64 = 'mad_i64_i32' in op_name or 'mad_u64_u32' in op_name
-      def fmt_sd_src(v, neg_bit, is_64bit=False):
-        s = _fmt_src64(v) if (is_64bit or is_f64) else fmt_src(v)
-        return f"-{s}" if neg_bit else s
-      src0_str, src1_str = fmt_sd_src(src0, neg & 1), fmt_sd_src(src1, neg & 2)
-      src2_str = fmt_sd_src(src2, neg & 4, is_mad64)
-      dst_str = _vreg(vdst, 2) if (is_f64 or is_mad64) else f"v{vdst}"
-      sdst_str = _fmt_sdst(sdst, 1)
-      # v_add_co_u32, v_sub_co_u32, v_subrev_co_u32, v_add_co_ci_u32, etc. only use 2 sources
-      if op_name in ('v_add_co_u32', 'v_sub_co_u32', 'v_subrev_co_u32', 'v_add_co_ci_u32', 'v_sub_co_ci_u32', 'v_subrev_co_ci_u32'):
-        return f"{op_name} {dst_str}, {sdst_str}, {src0_str}, {src1_str}"
-      # v_div_scale uses 3 sources
-      return f"{op_name} {dst_str}, {sdst_str}, {src0_str}, {src1_str}, {src2_str}" + omod_str
-
-    vdst = unwrap(inst._values.get('vdst', 0))
-    src0, src1, src2 = [unwrap(inst._values.get(f, 0)) for f in ('src0', 'src1', 'src2')]
-    neg, abs_, clmp = unwrap(inst._values.get('neg', 0)), unwrap(inst._values.get('abs', 0)), unwrap(inst._values.get('clmp', 0))
-    opsel = unwrap(inst._values.get('opsel', 0))
-    # Check if 64-bit op (needs register pairs)
-    is_f64 = 'f64' in op_name or 'i64' in op_name or 'u64' in op_name or 'b64' in op_name
-    # v_cmp_class_* has 64-bit src0 but 32-bit src1 (class mask)
-    is_class = 'class' in op_name
-    # Shift ops: v_*rev_*64 have 32-bit shift amount (src0), 64-bit value (src1)
-    is_shift64 = 'rev' in op_name and '64' in op_name and op_name.startswith('v_')
-    # v_ldexp_f64: 64-bit src0 (mantissa), 32-bit src1 (exponent)
-    is_ldexp64 = op_name == 'v_ldexp_f64'
-    # v_trig_preop_f64: 64-bit dst/src0, 32-bit src1 (exponent/scale)
-    is_trig_preop = op_name == 'v_trig_preop_f64'
-    # v_readlane_b32: destination is SGPR (despite vdst field)
-    is_readlane = op_name == 'v_readlane_b32'
-    # SAD/QSAD/MQSAD instructions have mixed sizes
-    # v_qsad_pk_u16_u8, v_mqsad_pk_u16_u8: 64-bit dst/src0/src2, 32-bit src1
-    # v_mqsad_u32_u8: 128-bit (4 reg) dst/src2, 64-bit src0, 32-bit src1
-    is_sad64 = any(x in op_name for x in ('qsad_pk', 'mqsad_pk'))
-    is_mqsad_u32 = 'mqsad_u32' in op_name
-    # Detect 16-bit and 64-bit operand sizes for various instruction patterns
-    if 'cvt_pk' in op_name:
-      is_f16_dst, is_f16_src, is_f16_src2 = False, op_name.endswith('16'), False
-    elif m := re.match(r'v_(?:cvt|frexp_exp)_([a-z0-9_]+)_([a-z0-9]+)', op_name):
-      dst_type, src_type = m.group(1), m.group(2)
-      is_f16_dst, is_f16_src, is_f16_src2 = _is_16bit(dst_type), _is_16bit(src_type), _is_16bit(src_type)
-      is_f64_dst, is_f64_src, is_f64 = '64' in dst_type, '64' in src_type, False
-    elif re.match(r'v_mad_[iu]32_[iu]16', op_name):
-      is_f16_dst, is_f16_src, is_f16_src2 = False, True, False  # 32-bit dst, 16-bit src0/src1, 32-bit src2
-    elif 'pack_b32' in op_name:
-      is_f16_dst, is_f16_src, is_f16_src2 = False, True, True  # 32-bit dst, 16-bit sources
-    else:
-      is_16bit_op = any(x in op_name for x in _16BIT_TYPES) and not any(x in op_name for x in ('dot2', 'pk_', 'sad', 'msad', 'qsad', 'mqsad'))
-      is_f16_dst = is_f16_src = is_f16_src2 = is_16bit_op
-    # Check if any opsel bit is set (any operand uses .h) - if so, we need explicit .l for low-half
-    any_hi = opsel != 0
-    def fmt_vop3_src(v, neg_bit, abs_bit, hi_bit=False, reg_cnt=1, is_16=False):
-      s = _fmt_src_n(v, reg_cnt) if reg_cnt > 1 else f"v{v - 256}.h" if is_16 and v >= 256 and hi_bit else f"v{v - 256}.l" if is_16 and v >= 256 and any_hi else fmt_src(v)
-      if abs_bit: s = f"|{s}|"
-      return f"-{s}" if neg_bit else s
-    # Determine register count for each source (check for cvt-specific 64-bit flags first)
-    is_src0_64 = locals().get('is_f64_src', is_f64 and not is_shift64) or is_sad64 or is_mqsad_u32
-    is_src1_64 = is_f64 and not is_class and not is_ldexp64 and not is_trig_preop
-    src0_cnt = 2 if is_src0_64 else 1
-    src1_cnt = 2 if is_src1_64 else 1
-    src2_cnt = 4 if is_mqsad_u32 else 2 if (is_f64 or is_sad64) else 1
-    src0_str = fmt_vop3_src(src0, neg & 1, abs_ & 1, opsel & 1, src0_cnt, is_f16_src)
-    src1_str = fmt_vop3_src(src1, neg & 2, abs_ & 2, opsel & 2, src1_cnt, is_f16_src)
-    src2_str = fmt_vop3_src(src2, neg & 4, abs_ & 4, opsel & 4, src2_cnt, is_f16_src2)
-    # Format destination - for 16-bit ops, use .h/.l suffix; readlane uses SGPR dest
-    is_dst_64 = locals().get('is_f64_dst', is_f64) or is_sad64
-    dst_cnt = 4 if is_mqsad_u32 else 2 if is_dst_64 else 1
-    if is_readlane:
-      dst_str = _fmt_sdst(vdst, 1)
-    elif dst_cnt > 1:
-      dst_str = _vreg(vdst, dst_cnt)
-    elif is_f16_dst:
-      dst_str = f"v{vdst}.h" if (opsel & 8) else f"v{vdst}.l" if any_hi else f"v{vdst}"
-    else:
-      dst_str = f"v{vdst}"
-    clamp_str = " clamp" if clmp else ""
-    omod = unwrap(inst._values.get('omod', 0))
-    omod_str = {1: " mul:2", 2: " mul:4", 3: " div:2"}.get(omod, "")
-    # op_sel for non-VGPR sources (when opsel bits are set but source is not a VGPR)
-    # For 16-bit ops with VGPR sources, opsel is encoded in .h/.l suffix
-    # For non-VGPR sources or non-16-bit ops, we need explicit op_sel
-    has_nonvgpr_opsel = (src0 < 256 and (opsel & 1)) or (src1 < 256 and (opsel & 2)) or (src2 < 256 and (opsel & 4))
-    need_opsel = has_nonvgpr_opsel or (opsel and not is_f16_src)
-    # Helper to format opsel string based on source count
-    def fmt_opsel(num_src):
-      if not need_opsel: return ""
-      # When dst is .h (for 16-bit ops) and non-VGPR sources have opsel, use all 1s
-      if is_f16_dst and (opsel & 8):  # dst is .h
-        return f" op_sel:[1,1,1{',1' if num_src == 3 else ''}]"
-      # Otherwise output actual opsel values
-      if num_src == 3:
-        return f" op_sel:[{opsel & 1},{(opsel >> 1) & 1},{(opsel >> 2) & 1},{(opsel >> 3) & 1}]"
-      return f" op_sel:[{opsel & 1},{(opsel >> 1) & 1},{(opsel >> 2) & 1}]"
-    # Determine number of sources based on opcode range:
-    # 0-255: VOPC promoted (comparison, 2 src, sdst)
-    # 256-383: VOP2 promoted (2 src)
-    # 384-511: VOP1 promoted (1 src)
-    # 512+: Native VOP3 (2 or 3 src depending on instruction)
-    if op_val < 256:  # VOPC promoted
-      # VOPCX (v_cmpx_*) writes to exec, no explicit destination
-      if op_name.startswith('v_cmpx'):
-        return f"{op_name}_e64 {src0_str}, {src1_str}"
-      return f"{op_name}_e64 {_fmt_sdst(vdst, 1)}, {src0_str}, {src1_str}"
-    elif op_val < 384:  # VOP2 promoted
-      # v_cndmask_b32 in VOP3 format has 3 sources (src2 is mask selector)
-      if 'cndmask' in op_name:
-        return f"{op_name}_e64 {dst_str}, {src0_str}, {src1_str}, {src2_str}" + fmt_opsel(3) + clamp_str + omod_str
-      return f"{op_name}_e64 {dst_str}, {src0_str}, {src1_str}" + fmt_opsel(2) + clamp_str + omod_str
-    elif op_val < 512:  # VOP1 promoted
-      if op_name in ('v_nop', 'v_pipeflush'): return f"{op_name}_e64"
-      return f"{op_name}_e64 {dst_str}, {src0_str}" + fmt_opsel(1) + clamp_str + omod_str
-    else:  # Native VOP3 - determine 2 vs 3 sources based on instruction name
-      # 3-source ops: fma, mad, min3, max3, med3, div_fixup, div_fmas, sad, msad, qsad, mqsad, lerp, alignbit/byte, cubeid/sc/tc/ma, bfe, bfi, perm_b32, permlane, cndmask
-      # Note: v_writelane_b32 is 2-src (src0, src1 with vdst as 3rd operand - read-modify-write)
-      is_3src = any(x in op_name for x in ('fma', 'mad', 'min3', 'max3', 'med3', 'div_fix', 'div_fmas', 'sad', 'lerp', 'align', 'cube',
-                                            'bfe', 'bfi', 'perm_b32', 'permlane', 'cndmask', 'xor3', 'or3', 'add3', 'lshl_or', 'and_or', 'lshl_add',
-                                            'add_lshl', 'xad', 'maxmin', 'minmax', 'dot2', 'cvt_pk_u8', 'mullit'))
-      if is_3src:
-        return f"{op_name} {dst_str}, {src0_str}, {src1_str}, {src2_str}" + fmt_opsel(3) + clamp_str + omod_str
-      return f"{op_name} {dst_str}, {src0_str}, {src1_str}" + fmt_opsel(2) + clamp_str + omod_str
-
-  # VOP3SD: 3-source with scalar destination (v_div_scale_*, v_add_co_u32, v_mad_*64_*32, etc.)
-  if cls_name == 'VOP3SD':
-    vdst, sdst = unwrap(inst._values.get('vdst', 0)), unwrap(inst._values.get('sdst', 0))
-    src0, src1, src2 = [unwrap(inst._values.get(f, 0)) for f in ('src0', 'src1', 'src2')]
-    neg, omod, clmp = unwrap(inst._values.get('neg', 0)), unwrap(inst._values.get('omod', 0)), unwrap(inst._values.get('clmp', 0))
-    is_f64, is_mad64 = 'f64' in op_name, 'mad_i64_i32' in op_name or 'mad_u64_u32' in op_name
-    def fmt_neg(v, neg_bit, is_64=False): return f"-{_fmt_src64(v) if (is_64 or is_f64) else fmt_src(v)}" if neg_bit else _fmt_src64(v) if (is_64 or is_f64) else fmt_src(v)
-    srcs = [fmt_neg(src0, neg & 1), fmt_neg(src1, neg & 2), fmt_neg(src2, neg & 4, is_mad64)]
-    dst_str, sdst_str = _vreg(vdst, 2) if (is_f64 or is_mad64) else f"v{vdst}", _fmt_sdst(sdst, 1)
-    clamp_str, omod_str = " clamp" if clmp else "", {1: " mul:2", 2: " mul:4", 3: " div:2"}.get(omod, "")
-    is_2src = op_name in ('v_add_co_u32', 'v_sub_co_u32', 'v_subrev_co_u32')
-    suffix = "_e64" if op_name.startswith('v_') and 'co_' in op_name else ""
-    return f"{op_name}{suffix} {dst_str}, {sdst_str}, {', '.join(srcs[:2] if is_2src else srcs)}" + clamp_str + omod_str
-
-  # VOPD: dual-issue instructions
-  if cls_name == 'VOPD':
-    from extra.assembly.amd.autogen import rdna3 as autogen
-    opx, opy, vdstx, vdsty_enc = [unwrap(inst._values.get(f, 0)) for f in ('opx', 'opy', 'vdstx', 'vdsty')]
-    srcx0, vsrcx1, srcy0, vsrcy1 = [unwrap(inst._values.get(f, 0)) for f in ('srcx0', 'vsrcx1', 'srcy0', 'vsrcy1')]
-    vdsty = (vdsty_enc << 1) | ((vdstx & 1) ^ 1)  # Decode vdsty
-    def fmt_vopd(op, vdst, src0, vsrc1):
-      try: name = autogen.VOPDOp(op).name.lower()
-      except (ValueError, KeyError): name = f"op_{op}"
-      return f"{name} v{vdst}, {fmt_src(src0)}" if 'mov' in name else f"{name} v{vdst}, {fmt_src(src0)}, v{vsrc1}"
-    return f"{fmt_vopd(opx, vdstx, srcx0, vsrcx1)} :: {fmt_vopd(opy, vdsty, srcy0, vsrcy1)}"
-
-  # VOP3P: packed vector ops
-  if cls_name == 'VOP3P':
-    vdst, clmp = unwrap(inst._values.get('vdst', 0)), unwrap(inst._values.get('clmp', 0))
-    src0, src1, src2 = [unwrap(inst._values.get(f, 0)) for f in ('src0', 'src1', 'src2')]
-    neg, neg_hi = unwrap(inst._values.get('neg', 0)), unwrap(inst._values.get('neg_hi', 0))
-    opsel, opsel_hi, opsel_hi2 = unwrap(inst._values.get('opsel', 0)), unwrap(inst._values.get('opsel_hi', 0)), unwrap(inst._values.get('opsel_hi2', 0))
-    is_wmma, is_3src = 'wmma' in op_name, any(x in op_name for x in ('fma', 'mad', 'dot', 'wmma'))
-    def fmt_bits(name, val, n): return f"{name}:[{','.join(str((val >> i) & 1) for i in range(n))}]"
-    # WMMA: f16/bf16 use 8-reg sources, iu8 uses 4-reg, iu4 uses 2-reg; all have 8-reg dst
-    if is_wmma:
-      src_cnt = 2 if 'iu4' in op_name else 4 if 'iu8' in op_name else 8
-      src0_str, src1_str, src2_str = _fmt_src_n(src0, src_cnt), _fmt_src_n(src1, src_cnt), _fmt_src_n(src2, 8)
-      dst_str = _vreg(vdst, 8)
-    else:
-      src0_str, src1_str, src2_str = _fmt_src_n(src0, 1), _fmt_src_n(src1, 1), _fmt_src_n(src2, 1)
-      dst_str = f"v{vdst}"
-    n = 3 if is_3src else 2
-    full_opsel_hi = opsel_hi | (opsel_hi2 << 2)
-    mods = [fmt_bits("op_sel", opsel, n)] if opsel else []
-    if full_opsel_hi != (0b111 if is_3src else 0b11): mods.append(fmt_bits("op_sel_hi", full_opsel_hi, n))
-    if neg: mods.append(fmt_bits("neg_lo", neg, n))
-    if neg_hi: mods.append(fmt_bits("neg_hi", neg_hi, n))
-    if clmp: mods.append("clamp")
-    mod_str = " " + " ".join(mods) if mods else ""
-    return f"{op_name} {dst_str}, {src0_str}, {src1_str}, {src2_str}{mod_str}" if is_3src else f"{op_name} {dst_str}, {src0_str}, {src1_str}{mod_str}"
-
-  # VINTERP: interpolation instructions
-  if cls_name == 'VINTERP':
-    vdst = unwrap(inst._values.get('vdst', 0))
-    src0, src1, src2 = [unwrap(inst._values.get(f, 0)) for f in ('src0', 'src1', 'src2')]
-    neg, waitexp, clmp = unwrap(inst._values.get('neg', 0)), unwrap(inst._values.get('waitexp', 0)), unwrap(inst._values.get('clmp', 0))
-    def fmt_neg_vi(v, neg_bit): return f"-{v}" if neg_bit else v
-    srcs = [fmt_neg_vi(f"v{s - 256}" if s >= 256 else fmt_src(s), neg & (1 << i)) for i, s in enumerate([src0, src1, src2])]
-    mods = [m for m in [f"wait_exp:{waitexp}" if waitexp else "", "clamp" if clmp else ""] if m]
-    return f"{op_name} v{vdst}, {', '.join(srcs)}" + (" " + " ".join(mods) if mods else "")
-
-  # MUBUF/MTBUF helpers
-  def _buf_vaddr(vaddr, offen, idxen): return _vreg(vaddr, 2) if offen and idxen else f"v{vaddr}" if offen or idxen else "off"
-  def _buf_srsrc(srsrc): srsrc_base = srsrc * 4; return _reg("ttmp", srsrc_base - 108, 4) if 108 <= srsrc_base <= 123 else _sreg(srsrc_base, 4)
-
-  # MUBUF: buffer load/store
-  if cls_name == 'MUBUF':
-    vdata, vaddr, srsrc, soffset = [unwrap(inst._values.get(f, 0)) for f in ('vdata', 'vaddr', 'srsrc', 'soffset')]
-    offset, offen, idxen = unwrap(inst._values.get('offset', 0)), unwrap(inst._values.get('offen', 0)), unwrap(inst._values.get('idxen', 0))
-    glc, dlc, slc, tfe = [unwrap(inst._values.get(f, 0)) for f in ('glc', 'dlc', 'slc', 'tfe')]
-    if op_name in ('buffer_gl0_inv', 'buffer_gl1_inv'): return op_name
-    # Determine data width from op name
-    if 'd16' in op_name: width = 2 if any(x in op_name for x in ('xyz', 'xyzw')) else 1
-    elif 'atomic' in op_name:
-      base_width = 2 if any(x in op_name for x in ('b64', 'u64', 'i64')) else 1
-      width = base_width * 2 if 'cmpswap' in op_name else base_width
-    else: width = {'b32':1, 'b64':2, 'b96':3, 'b128':4, 'b16':1, 'x':1, 'xy':2, 'xyz':3, 'xyzw':4}.get(op_name.split('_')[-1], 1)
-    if tfe: width += 1
-    mods = [m for m in ["offen" if offen else "", "idxen" if idxen else "", f"offset:{offset}" if offset else "",
-                        "glc" if glc else "", "dlc" if dlc else "", "slc" if slc else "", "tfe" if tfe else ""] if m]
-    return f"{op_name} {_vreg(vdata, width)}, {_buf_vaddr(vaddr, offen, idxen)}, {_buf_srsrc(srsrc)}, {decode_src(soffset)}" + (" " + " ".join(mods) if mods else "")
-
-  # MTBUF: typed buffer load/store
-  if cls_name == 'MTBUF':
-    vdata, vaddr, srsrc, soffset = [unwrap(inst._values.get(f, 0)) for f in ('vdata', 'vaddr', 'srsrc', 'soffset')]
-    offset, tbuf_fmt, offen, idxen = [unwrap(inst._values.get(f, 0)) for f in ('offset', 'format', 'offen', 'idxen')]
-    glc, dlc, slc = [unwrap(inst._values.get(f, 0)) for f in ('glc', 'dlc', 'slc')]
-    mods = [f"format:{tbuf_fmt}"] + [m for m in ["idxen" if idxen else "", "offen" if offen else "", f"offset:{offset}" if offset else "",
-                                                  "glc" if glc else "", "dlc" if dlc else "", "slc" if slc else ""] if m]
-    width = 2 if 'd16' in op_name and any(x in op_name for x in ('xyz', 'xyzw')) else 1 if 'd16' in op_name else {'x':1, 'xy':2, 'xyz':3, 'xyzw':4}.get(op_name.split('_')[-1], 1)
-    return f"{op_name} {_vreg(vdata, width)}, {_buf_vaddr(vaddr, offen, idxen)}, {_buf_srsrc(srsrc)}, {decode_src(soffset)} {' '.join(mods)}"
-
-  # SOP1/SOP2/SOPC/SOPK
-  if cls_name in ('SOP1', 'SOP2', 'SOPC', 'SOPK'):
-    sizes = _parse_sop_sizes(op_name)
-    dst_cnt, src0_cnt = sizes[0], sizes[1]
-    src1_cnt = sizes[2] if len(sizes) > 2 else src0_cnt
-    if cls_name == 'SOP1':
-      sdst, ssrc0 = unwrap(inst._values.get('sdst', 0)), unwrap(inst._values.get('ssrc0', 0))
-      if op_name == 's_getpc_b64': return f"{op_name} {_fmt_sdst(sdst, 2)}"
-      if op_name in ('s_setpc_b64', 's_rfe_b64'): return f"{op_name} {_fmt_ssrc(ssrc0, 2)}"
-      if op_name == 's_swappc_b64': return f"{op_name} {_fmt_sdst(sdst, 2)}, {_fmt_ssrc(ssrc0, 2)}"
-      if op_name in ('s_sendmsg_rtn_b32', 's_sendmsg_rtn_b64'):
-        return f"{op_name} {_fmt_sdst(sdst, 2 if 'b64' in op_name else 1)}, sendmsg({MSG_NAMES.get(ssrc0, str(ssrc0))})"
-      ssrc0_str = fmt_src(ssrc0) if src0_cnt == 1 else _fmt_ssrc(ssrc0, src0_cnt)
-      return f"{op_name} {_fmt_sdst(sdst, dst_cnt)}, {ssrc0_str}"
-    if cls_name == 'SOP2':
-      sdst, ssrc0, ssrc1 = [unwrap(inst._values.get(f, 0)) for f in ('sdst', 'ssrc0', 'ssrc1')]
-      ssrc0_str = fmt_src(ssrc0) if ssrc0 == 255 else _fmt_ssrc(ssrc0, src0_cnt)
-      ssrc1_str = fmt_src(ssrc1) if ssrc1 == 255 else _fmt_ssrc(ssrc1, src1_cnt)
-      return f"{op_name} {_fmt_sdst(sdst, dst_cnt)}, {ssrc0_str}, {ssrc1_str}"
-    if cls_name == 'SOPC':
-      return f"{op_name} {_fmt_ssrc(unwrap(inst._values.get('ssrc0', 0)), src0_cnt)}, {_fmt_ssrc(unwrap(inst._values.get('ssrc1', 0)), src1_cnt)}"
-    if cls_name == 'SOPK':
-      sdst, simm16 = unwrap(inst._values.get('sdst', 0)), unwrap(inst._values.get('simm16', 0))
-      if op_name == 's_version': return f"{op_name} 0x{simm16:x}"
-      if op_name in ('s_setreg_b32', 's_getreg_b32'):
-        hwreg_id, hwreg_offset, hwreg_size = simm16 & 0x3f, (simm16 >> 6) & 0x1f, ((simm16 >> 11) & 0x1f) + 1
-        hwreg_str = f"0x{simm16:x}" if hwreg_id in (16, 17) else f"hwreg({HWREG_NAMES.get(hwreg_id, str(hwreg_id))}, {hwreg_offset}, {hwreg_size})"
-        return f"{op_name} {hwreg_str}, {_fmt_sdst(sdst, 1)}" if op_name == 's_setreg_b32' else f"{op_name} {_fmt_sdst(sdst, 1)}, {hwreg_str}"
-      return f"{op_name} {_fmt_sdst(sdst, dst_cnt)}, 0x{simm16:x}"
-
-  # Generic fallback
-  def fmt_field(n, v):
-    v = unwrap(v)
-    if n in SRC_FIELDS: return fmt_src(v) if v != 255 else "0xff"
-    if n in ('sdst', 'vdst'): return f"{'s' if n == 'sdst' else 'v'}{v}"
-    return f"v{v}" if n == 'vsrc1' else f"0x{v:x}" if n == 'simm16' else str(v)
-  ops = [fmt_field(n, inst._values.get(n, 0)) for n in inst._fields if n not in ('encoding', 'op')]
-  return f"{op_name} {', '.join(ops)}" if ops else op_name
-
-# Assembler
-SPECIAL_REGS = {'vcc_lo': RawImm(106), 'vcc_hi': RawImm(107), 'null': RawImm(124), 'off': RawImm(124), 'm0': RawImm(125), 'exec_lo': RawImm(126), 'exec_hi': RawImm(127), 'scc': RawImm(253)}
-FLOAT_CONSTS = {'0.5': 0.5, '-0.5': -0.5, '1.0': 1.0, '-1.0': -1.0, '2.0': 2.0, '-2.0': -2.0, '4.0': 4.0, '-4.0': -4.0}
-REG_MAP: dict[str, _RegFactory] = {'s': s, 'v': v, 't': ttmp, 'ttmp': ttmp}
-
-def parse_operand(op: str) -> tuple:
-  op = op.strip().lower()
-  neg = op.startswith('-') and not op[1:2].isdigit(); op = op[1:] if neg else op
-  abs_ = op.startswith('|') and op.endswith('|') or op.startswith('abs(') and op.endswith(')')
-  op = op[1:-1] if op.startswith('|') else op[4:-1] if op.startswith('abs(') else op
-  hi_half = op.endswith('.h')
-  op = re.sub(r'\.[lh]$', '', op)
-  if op in FLOAT_CONSTS: return (FLOAT_CONSTS[op], neg, abs_, hi_half)
-  if re.match(r'^-?\d+$', op): return (int(op), neg, abs_, hi_half)
-  if m := re.match(r'^-?0x([0-9a-f]+)$', op):
-    v = -int(m.group(1), 16) if op.startswith('-') else int(m.group(1), 16)
-    return (v, neg, abs_, hi_half)
-  if op in SPECIAL_REGS: return (SPECIAL_REGS[op], neg, abs_, hi_half)
-  if op == 'lit': return (RawImm(255), neg, abs_, hi_half)  # literal marker (actual value comes from literal word)
-  if m := re.match(r'^([svt](?:tmp)?)\[(\d+):(\d+)\]$', op): return (REG_MAP[m.group(1)][int(m.group(2)):int(m.group(3))], neg, abs_, hi_half)
-  if m := re.match(r'^([svt](?:tmp)?)(\d+)$', op):
-    reg = REG_MAP[m.group(1)][int(m.group(2))]
-    reg.hi = hi_half
-    return (reg, neg, abs_, hi_half)
-  # hwreg(name, offset, size) or hwreg(name) -> simm16 encoding
-  if m := re.match(r'^hwreg\((\w+)(?:,\s*(\d+),\s*(\d+))?\)$', op):
-    name_str = m.group(1).lower()
-    hwreg_id = HWREG_IDS.get(name_str, int(name_str) if name_str.isdigit() else None)
-    if hwreg_id is None: raise ValueError(f"unknown hwreg name: {name_str}")
-    offset, size = int(m.group(2)) if m.group(2) else 0, int(m.group(3)) if m.group(3) else 32
-    return (((size - 1) << 11) | (offset << 6) | hwreg_id, neg, abs_, hi_half)
-  raise ValueError(f"cannot parse operand: {op}")
-
-SMEM_OPS = {'s_load_b32', 's_load_b64', 's_load_b128', 's_load_b256', 's_load_b512',
-            's_buffer_load_b32', 's_buffer_load_b64', 's_buffer_load_b128', 's_buffer_load_b256', 's_buffer_load_b512'}
-SOP1_SRC_ONLY = {'s_setpc_b64', 's_rfe_b64'}
-SOP1_MSG_IMM = {'s_sendmsg_rtn_b32', 's_sendmsg_rtn_b64'}
-SOPK_IMM_ONLY = {'s_version'}
-SOPK_IMM_FIRST = {'s_setreg_b32'}
-SOPK_UNSUPPORTED = {'s_setreg_imm32_b32'}
-
-def asm(text: str) -> Inst:
-  from extra.assembly.amd.autogen import rdna3 as autogen
-  text = text.strip()
-  clamp = 'clamp' in text.lower()
-  if clamp: text = re.sub(r'\s+clamp\s*$', '', text, flags=re.I)
-  modifiers = {}
-  if m := re.search(r'\s+wait_exp:(\d+)', text, re.I): modifiers['waitexp'] = int(m.group(1)); text = text[:m.start()] + text[m.end():]
-  parts = text.replace(',', ' ').split()
-  if not parts: raise ValueError("empty instruction")
-  mnemonic, op_str = parts[0].lower(), text[len(parts[0]):].strip()
-  # Handle s_waitcnt specially before operand parsing
-  if mnemonic == 's_waitcnt':
-    vmcnt, expcnt, lgkmcnt = 0x3f, 0x7, 0x3f
-    for part in op_str.replace(',', ' ').split():
-      if m := re.match(r'vmcnt\((\d+)\)', part): vmcnt = int(m.group(1))
-      elif m := re.match(r'expcnt\((\d+)\)', part): expcnt = int(m.group(1))
-      elif m := re.match(r'lgkmcnt\((\d+)\)', part): lgkmcnt = int(m.group(1))
-      elif re.match(r'^0x[0-9a-f]+$|^\d+$', part): return autogen.s_waitcnt(simm16=int(part, 0))
-    return autogen.s_waitcnt(simm16=waitcnt(vmcnt, expcnt, lgkmcnt))
-  # Handle VOPD dual-issue instructions: opx dst, src :: opy dst, src
-  if '::' in text:
-    x_part, y_part = text.split('::')
-    x_parts, y_parts = x_part.strip().replace(',', ' ').split(), y_part.strip().replace(',', ' ').split()
-    opx_name, opy_name = x_parts[0].upper(), y_parts[0].upper()
-    opx, opy = autogen.VOPDOp[opx_name], autogen.VOPDOp[opy_name]
-    x_ops, y_ops = [parse_operand(p)[0] for p in x_parts[1:]], [parse_operand(p)[0] for p in y_parts[1:]]
-    vdstx, srcx0 = x_ops[0], x_ops[1] if len(x_ops) > 1 else 0
-    vsrcx1 = x_ops[2] if len(x_ops) > 2 else VGPR(0)
-    vdsty, srcy0 = y_ops[0], y_ops[1] if len(y_ops) > 1 else 0
-    vsrcy1 = y_ops[2] if len(y_ops) > 2 else VGPR(0)
-    # Handle fmaak/fmamk literals (4th operand on x or y side)
-    lit = None
-    if 'fmaak' in opx_name.lower() and len(x_ops) > 3: lit = unwrap(x_ops[3])
-    elif 'fmamk' in opx_name.lower() and len(x_ops) > 3: lit, vsrcx1 = unwrap(x_ops[2]), x_ops[3]
-    elif 'fmaak' in opy_name.lower() and len(y_ops) > 3: lit = unwrap(y_ops[3])
-    elif 'fmamk' in opy_name.lower() and len(y_ops) > 3: lit, vsrcy1 = unwrap(y_ops[2]), y_ops[3]
-    return autogen.VOPD(opx, opy, vdstx=vdstx, vdsty=vdsty, srcx0=srcx0, vsrcx1=vsrcx1, srcy0=srcy0, vsrcy1=vsrcy1, literal=lit)
-  operands, current, depth, in_pipe = [], "", 0, False
-  for ch in op_str:
-    if ch in '[(': depth += 1
-    elif ch in '])': depth -= 1
-    elif ch == '|': in_pipe = not in_pipe
-    if ch == ',' and depth == 0 and not in_pipe: operands.append(current.strip()); current = ""
-    else: current += ch
-  if current.strip(): operands.append(current.strip())
-  parsed = [parse_operand(op) for op in operands]
-  values = [p[0] for p in parsed]
-  neg_bits = sum((1 << (i-1)) for i, p in enumerate(parsed) if i > 0 and p[1])
-  abs_bits = sum((1 << (i-1)) for i, p in enumerate(parsed) if i > 0 and p[2])
-  opsel_bits = (8 if len(parsed) > 0 and parsed[0][3] else 0) | sum((1 << i) for i, p in enumerate(parsed[1:4]) if p[3])
-  lit = None
-  if mnemonic in ('v_fmaak_f32', 'v_fmaak_f16') and len(values) == 4: lit, values = unwrap(values[3]), values[:3]
-  elif mnemonic in ('v_fmamk_f32', 'v_fmamk_f16') and len(values) == 4: lit, values = unwrap(values[2]), [values[0], values[1], values[3]]
-  vcc_ops = {'v_add_co_ci_u32', 'v_sub_co_ci_u32', 'v_subrev_co_ci_u32', 'v_add_co_u32', 'v_sub_co_u32', 'v_subrev_co_u32'}
-  if mnemonic.replace('_e32', '') in vcc_ops and len(values) >= 5: values = [values[0], values[2], values[3]]
-  # v_cmp_*_e32: strip implicit vcc_lo dest. v_cmp_*_e64: keep vdst (vcc_lo encodes to 106)
-  if mnemonic.startswith('v_cmp') and not mnemonic.endswith('_e64') and len(values) >= 3 and operands[0].strip().lower() in ('vcc_lo', 'vcc_hi', 'vcc'):
-    values = values[1:]
-  # CMPX instructions with _e64 suffix: prepend implicit EXEC_LO destination (vdst=126)
-  if 'cmpx' in mnemonic and mnemonic.endswith('_e64') and len(values) == 2:
-    values = [VGPR(126, 1)] + values
-    # Recalculate modifiers: parsed[0]=src0, parsed[1]=src1 (no vdst in user input)
-    neg_bits = sum((1 << i) for i, p in enumerate(parsed[:3]) if p[1])
-    abs_bits = sum((1 << i) for i, p in enumerate(parsed[:3]) if p[2])
-    opsel_bits = sum((1 << i) for i, p in enumerate(parsed[:2]) if p[3])
-  vop3sd_ops = {'v_div_scale_f32', 'v_div_scale_f64'}
-  if mnemonic in vop3sd_ops and len(parsed) >= 5:
-    neg_bits = sum((1 << i) for i, p in enumerate(parsed[2:5]) if p[1])
-    abs_bits = sum((1 << i) for i, p in enumerate(parsed[2:5]) if p[2])
-  if mnemonic in SOPK_UNSUPPORTED: raise ValueError(f"unsupported instruction: {mnemonic}")
-  elif mnemonic in SOP1_SRC_ONLY:
-    return getattr(autogen, mnemonic)(ssrc0=values[0])
-  elif mnemonic in SOP1_MSG_IMM:
-    return getattr(autogen, mnemonic)(sdst=values[0], ssrc0=RawImm(unwrap(values[1])))
-  elif mnemonic in SOPK_IMM_ONLY:
-    return getattr(autogen, mnemonic)(simm16=values[0])
-  elif mnemonic in SOPK_IMM_FIRST:
-    return getattr(autogen, mnemonic)(simm16=values[0], sdst=values[1])
-  elif mnemonic in SMEM_OPS and len(operands) >= 3 and re.match(r'^-?[0-9]|^-?0x', operands[2].strip().lower()):
-    return getattr(autogen, mnemonic)(sdata=values[0], sbase=values[1], offset=values[2], soffset=RawImm(124))
-  elif mnemonic.startswith('buffer_') and len(operands) >= 2 and operands[1].strip().lower() == 'off':
-    return getattr(autogen, mnemonic)(vdata=values[0], vaddr=0, srsrc=values[2], soffset=RawImm(unwrap(values[3])) if len(values) > 3 else RawImm(0))
-  elif (mnemonic.startswith('flat_load') or mnemonic.startswith('global_load') or mnemonic.startswith('scratch_load')) and len(values) >= 3:
-    offset = int(m.group(1)) if (m := re.search(r'offset:(-?\d+)', op_str)) else 0
-    return getattr(autogen, mnemonic)(vdst=values[0], addr=values[1], saddr=values[2], offset=offset)
-  elif (mnemonic.startswith('flat_store') or mnemonic.startswith('global_store') or mnemonic.startswith('scratch_store')) and len(values) >= 3:
-    offset = int(m.group(1)) if (m := re.search(r'offset:(-?\d+)', op_str)) else 0
-    return getattr(autogen, mnemonic)(addr=values[0], data=values[1], saddr=values[2], offset=offset)
-  for suffix in (['_e32', ''] if not (neg_bits or abs_bits or clamp) else ['', '_e32']):
-    if hasattr(autogen, name := mnemonic.replace('.', '_') + suffix):
-      use_opsel = 'opsel' in getattr(autogen, name).func._fields
-      vals = [type(v)(v.idx, v.count, False) if isinstance(v, Reg) and v.hi and use_opsel else v for v in values]
-      inst = getattr(autogen, name)(*vals, literal=lit, **modifiers)
-      if neg_bits and 'neg' in inst._fields: inst._values['neg'] = neg_bits
-      if opsel_bits and use_opsel: inst._values['opsel'] = opsel_bits
-      if abs_bits and 'abs' in inst._fields: inst._values['abs'] = abs_bits
-      if clamp and 'clmp' in inst._fields: inst._values['clmp'] = 1
-      return inst
-  raise ValueError(f"unknown instruction: {mnemonic}")

extra/assembly/amd/dsl.py

@@ -1,577 +0,0 @@
-# library for RDNA3 assembly DSL
-# mypy: ignore-errors
-from __future__ import annotations
-from enum import IntEnum
-from typing import overload, Annotated, TypeVar, Generic
-
-# Bit field DSL
-class BitField:
-  def __init__(self, hi: int, lo: int, name: str | None = None): self.hi, self.lo, self.name = hi, lo, name
-  def __set_name__(self, owner, name): self.name, self._owner = name, owner
-  def __eq__(self, val: int) -> tuple[BitField, int]: return (self, val)  # type: ignore
-  def mask(self) -> int: return (1 << (self.hi - self.lo + 1)) - 1
-  @property
-  def marker(self) -> type | None:
-    # Get marker from Annotated type hint if present
-    import typing
-    if hasattr(self, '_owner') and self.name:
-      hints = typing.get_type_hints(self._owner, include_extras=True)
-      if self.name in hints:
-        hint = hints[self.name]
-        if typing.get_origin(hint) is Annotated:
-          args = typing.get_args(hint)
-          return args[1] if len(args) > 1 else None
-    return None
-  @overload
-  def __get__(self, obj: None, objtype: type) -> BitField: ...
-  @overload
-  def __get__(self, obj: object, objtype: type | None = None) -> int: ...
-  def __get__(self, obj, objtype=None):
-    if obj is None: return self
-    val = unwrap(obj._values.get(self.name, 0))
-    # Convert to IntEnum if marker is an IntEnum subclass
-    if self.marker and isinstance(self.marker, type) and issubclass(self.marker, IntEnum):
-      try: return self.marker(val)
-      except ValueError: pass
-    return val
-
-class _Bits:
-  def __getitem__(self, key) -> BitField: return BitField(key.start, key.stop) if isinstance(key, slice) else BitField(key, key)
-bits = _Bits()
-
-# Register types
-class Reg:
-  def __init__(self, idx: int, count: int = 1, hi: bool = False, neg: bool = False): self.idx, self.count, self.hi, self.neg = idx, count, hi, neg
-  def __repr__(self): return f"{self.__class__.__name__.lower()[0]}[{self.idx}]" if self.count == 1 else f"{self.__class__.__name__.lower()[0]}[{self.idx}:{self.idx + self.count}]"
-  def __neg__(self): return self.__class__(self.idx, self.count, self.hi, neg=not self.neg)
-
-T = TypeVar('T', bound=Reg)
-class _RegFactory(Generic[T]):
-  def __init__(self, cls: type[T], name: str): self._cls, self._name = cls, name
-  @overload
-  def __getitem__(self, key: int) -> Reg: ...
-  @overload
-  def __getitem__(self, key: slice) -> Reg: ...
-  def __getitem__(self, key: int | slice) -> Reg:
-    return self._cls(key.start, key.stop - key.start + 1) if isinstance(key, slice) else self._cls(key)
-  def __repr__(self): return f"<{self._name} factory>"
-
-class SGPR(Reg): pass
-class VGPR(Reg): pass
-class TTMP(Reg): pass
-s: _RegFactory[SGPR] = _RegFactory(SGPR, "SGPR")
-v: _RegFactory[VGPR] = _RegFactory(VGPR, "VGPR")
-ttmp: _RegFactory[TTMP] = _RegFactory(TTMP, "TTMP")
-
-# Field type markers (runtime classes for validation)
-class _SSrc: pass
-class _Src: pass
-class _Imm: pass
-class _SImm: pass
-class _VDSTYEnc: pass  # VOPD vdsty: encoded = actual >> 1, actual = (encoded << 1) | ((vdstx & 1) ^ 1)
-class _SGPRField: pass
-class _VGPRField: pass
-
-# Type aliases for annotations - tells mypy it's a BitField while preserving marker info
-SSrc = Annotated[BitField, _SSrc]
-Src = Annotated[BitField, _Src]
-Imm = Annotated[BitField, _Imm]
-SImm = Annotated[BitField, _SImm]
-VDSTYEnc = Annotated[BitField, _VDSTYEnc]
-SGPRField = Annotated[BitField, _SGPRField]
-VGPRField = Annotated[BitField, _VGPRField]
-class RawImm:
-  def __init__(self, val: int): self.val = val
-  def __repr__(self): return f"RawImm({self.val})"
-  def __eq__(self, other): return isinstance(other, RawImm) and self.val == other.val
-
-def unwrap(val) -> int:
-  return val.val if isinstance(val, RawImm) else val.value if hasattr(val, 'value') else val.idx if hasattr(val, 'idx') else val
-
-# Encoding helpers
-FLOAT_ENC = {0.5: 240, -0.5: 241, 1.0: 242, -1.0: 243, 2.0: 244, -2.0: 245, 4.0: 246, -4.0: 247}
-SRC_FIELDS = {'src0', 'src1', 'src2', 'ssrc0', 'ssrc1', 'soffset', 'srcx0', 'srcy0'}
-RAW_FIELDS = {'vdata', 'vdst', 'vaddr', 'addr', 'data', 'data0', 'data1', 'sdst', 'sdata'}
-
-def _encode_reg(val) -> int:
-  if isinstance(val, TTMP): return 108 + val.idx
-  return val.idx | (0x80 if val.hi else 0)
-
-def encode_src(val) -> int:
-  if isinstance(val, VGPR): return 256 + _encode_reg(val)
-  if isinstance(val, Reg): return _encode_reg(val)
-  if hasattr(val, 'value'): return val.value
-  if isinstance(val, float): return 128 if val == 0.0 else FLOAT_ENC.get(val, 255)
-  return 128 + val if isinstance(val, int) and 0 <= val <= 64 else 192 + (-val) if isinstance(val, int) and -16 <= val <= -1 else 255
-
-# Instruction base class
-class Inst:
-  _fields: dict[str, BitField]
-  _encoding: tuple[BitField, int] | None = None
-  _defaults: dict[str, int] = {}
-  _values: dict[str, int | RawImm]
-  _words: int  # size in 32-bit words, set by decode_program
-  _literal: int | None
-
-  def __init_subclass__(cls, **kwargs):
-    super().__init_subclass__(**kwargs)
-    cls._fields = {n: v[0] if isinstance(v, tuple) else v for n, v in cls.__dict__.items() if isinstance(v, BitField) or (isinstance(v, tuple) and len(v) == 2 and isinstance(v[0], BitField))}
-    if 'encoding' in cls._fields and isinstance(cls.__dict__.get('encoding'), tuple): cls._encoding = cls.__dict__['encoding']
-
-  def __init__(self, *args, literal: int | None = None, **kwargs):
-    self._values, self._literal = dict(self._defaults), literal
-    # Map positional args to field names
-    field_names = [n for n in self._fields if n != 'encoding']
-    orig_args = dict(zip(field_names, args))
-    orig_args.update(kwargs)
-    self._values.update(orig_args)
-    # Validate register counts for SMEM instructions (before encoding)
-    if self.__class__.__name__ == 'SMEM':
-      op_val = orig_args.get(field_names[0]) if args else orig_args.get('op')
-      if op_val is not None:
-        if hasattr(op_val, 'value'): op_val = op_val.value
-        expected_cnt = {0:1, 1:2, 2:4, 3:8, 4:16, 8:1, 9:2, 10:4, 11:8, 12:16}.get(op_val)
-        sdata_val = orig_args.get('sdata')
-        if expected_cnt is not None and isinstance(sdata_val, Reg) and sdata_val.count != expected_cnt:
-          raise ValueError(f"SMEM op {op_val} expects {expected_cnt} registers, got {sdata_val.count}")
-    # Validate register counts for SOP1 instructions (b32 = 1 reg, b64 = 2 regs)
-    if self.__class__.__name__ == 'SOP1':
-      op_val = orig_args.get(field_names[0]) if args else orig_args.get('op')
-      if op_val is not None and hasattr(op_val, 'name'):
-        expected = 2 if op_val.name.endswith('_B64') else 1
-        sdst_val, ssrc0_val = orig_args.get('sdst'), orig_args.get('ssrc0')
-        if isinstance(sdst_val, Reg) and sdst_val.count != expected:
-          raise ValueError(f"SOP1 {op_val.name} expects {expected} destination register(s), got {sdst_val.count}")
-        if isinstance(ssrc0_val, Reg) and ssrc0_val.count != expected:
-          raise ValueError(f"SOP1 {op_val.name} expects {expected} source register(s), got {ssrc0_val.count}")
-    # Type check and encode values
-    for name, val in list(self._values.items()):
-      if name == 'encoding': continue
-      # For RawImm, only process RAW_FIELDS to unwrap to int
-      if isinstance(val, RawImm):
-        if name in RAW_FIELDS: self._values[name] = val.val
-        continue
-      field = self._fields.get(name)
-      marker = field.marker if field else None
-      # Type validation
-      if marker is _SGPRField:
-        if isinstance(val, VGPR): raise TypeError(f"field '{name}' requires SGPR, got VGPR")
-        if not isinstance(val, (SGPR, TTMP, int, RawImm)): raise TypeError(f"field '{name}' requires SGPR, got {type(val).__name__}")
-      if marker is _VGPRField:
-        if not isinstance(val, VGPR): raise TypeError(f"field '{name}' requires VGPR, got {type(val).__name__}")
-      if marker is _SSrc and isinstance(val, VGPR): raise TypeError(f"field '{name}' requires scalar source, got VGPR")
-      # Encode source fields as RawImm for consistent disassembly
-      if name in SRC_FIELDS:
-        encoded = encode_src(val)
-        self._values[name] = RawImm(encoded)
-        # Handle negation modifier for VOP3 instructions
-        if isinstance(val, Reg) and val.neg and 'neg' in self._fields:
-          neg_bit = {'src0': 1, 'src1': 2, 'src2': 4}.get(name, 0)
-          cur_neg = self._values.get('neg', 0)
-          self._values['neg'] = (cur_neg.val if isinstance(cur_neg, RawImm) else cur_neg) | neg_bit
-        # Track literal value if needed (encoded as 255)
-        # For 64-bit ops, store literal in high 32 bits (to match from_bytes decoding and to_bytes encoding)
-        if encoded == 255 and self._literal is None and isinstance(val, int) and not isinstance(val, IntEnum):
-          self._literal = (val << 32) if self._is_64bit_op() else val
-        elif encoded == 255 and self._literal is None and isinstance(val, float):
-          import struct
-          lit32 = struct.unpack('<I', struct.pack('<f', val))[0]
-          self._literal = (lit32 << 32) if self._is_64bit_op() else lit32
-      # Encode raw register fields for consistent repr
-      elif name in RAW_FIELDS:
-        if isinstance(val, Reg): self._values[name] = _encode_reg(val)
-        elif hasattr(val, 'value'): self._values[name] = val.value  # IntEnum like SrcEnum.NULL
-      # Encode sbase (divided by 2) and srsrc/ssamp (divided by 4)
-      elif name == 'sbase' and isinstance(val, Reg):
-        self._values[name] = val.idx // 2
-      elif name in {'srsrc', 'ssamp'} and isinstance(val, Reg):
-        self._values[name] = val.idx // 4
-      # VOPD vdsty: encode as actual >> 1 (constraint: vdsty parity must be opposite of vdstx)
-      elif marker is _VDSTYEnc and isinstance(val, VGPR):
-        self._values[name] = val.idx >> 1
-
-  def _encode_field(self, name: str, val) -> int:
-    if isinstance(val, RawImm): return val.val
-    if name in {'srsrc', 'ssamp'}: return val.idx // 4 if isinstance(val, Reg) else val
-    if name == 'sbase': return val.idx // 2 if isinstance(val, Reg) else val
-    if name in RAW_FIELDS: return _encode_reg(val) if isinstance(val, Reg) else val
-    if isinstance(val, Reg) or name in SRC_FIELDS: return encode_src(val)
-    return val.value if hasattr(val, 'value') else val
-
-  def to_int(self) -> int:
-    word = (self._encoding[1] & self._encoding[0].mask()) << self._encoding[0].lo if self._encoding else 0
-    for n, bf in self._fields.items():
-      if n != 'encoding' and n in self._values: word |= (self._encode_field(n, self._values[n]) & bf.mask()) << bf.lo
-    return word
-
-  def _get_literal(self) -> int | None:
-    for n in SRC_FIELDS:
-      if n in self._values and not isinstance(v := self._values[n], RawImm) and isinstance(v, int) and not isinstance(v, IntEnum) and not (0 <= v <= 64 or -16 <= v <= -1): return v
-    return None
-
-  def _is_64bit_op(self) -> bool:
-    """Check if this instruction uses 64-bit operands (and thus 64-bit literals).
-    Exception: V_LDEXP_F64 has 32-bit integer src1, so its literal is 32-bit."""
-    op = self._values.get('op')
-    if op is None: return False
-    # op may be an enum (from __init__) or an int (from from_int)
-    op_name = op.name if hasattr(op, 'name') else None
-    if op_name is None and self.__class__.__name__ == 'VOP3':
-      from extra.assembly.amd.autogen.rdna3 import VOP3Op
-      try: op_name = VOP3Op(op).name
-      except ValueError: pass
-    if op_name is None: return False
-    # V_LDEXP_F64 has 32-bit integer exponent in src1, so literal is 32-bit
-    if op_name == 'V_LDEXP_F64': return False
-    return op_name.endswith(('_F64', '_B64', '_I64', '_U64'))
-
-  def to_bytes(self) -> bytes:
-    result = self.to_int().to_bytes(self._size(), 'little')
-    lit = self._get_literal() or getattr(self, '_literal', None)
-    if lit is None: return result
-    # For 64-bit ops, literal is stored in high 32 bits internally, but encoded as 4 bytes
-    lit32 = (lit >> 32) if self._is_64bit_op() else lit
-    return result + (lit32 & 0xffffffff).to_bytes(4, 'little')
-
-  @classmethod
-  def _size(cls) -> int: return 4 if issubclass(cls, Inst32) else 8
-  def size(self) -> int:
-    # Literal is always 4 bytes in the binary (for 64-bit ops, it's in high 32 bits)
-    return self._size() + (4 if self._literal is not None else 0)
-
-  @classmethod
-  def from_int(cls, word: int):
-    inst = object.__new__(cls)
-    inst._values = {n: RawImm(v) if n in SRC_FIELDS else v for n, bf in cls._fields.items() if n != 'encoding' for v in [(word >> bf.lo) & bf.mask()]}
-    inst._literal = None
-    return inst
-
-  @classmethod
-  def from_bytes(cls, data: bytes):
-    inst = cls.from_int(int.from_bytes(data[:cls._size()], 'little'))
-    op_val = inst._values.get('op', 0)
-    has_literal = cls.__name__ == 'VOP2' and op_val in (44, 45, 55, 56)
-    has_literal = has_literal or (cls.__name__ == 'SOP2' and op_val in (69, 70))
-    for n in SRC_FIELDS:
-      if n in inst._values and isinstance(inst._values[n], RawImm) and inst._values[n].val == 255: has_literal = True
-    if has_literal:
-      # For 64-bit ops, the literal is 32 bits placed in the HIGH 32 bits of the 64-bit value
-      # (low 32 bits are zero). This is how AMD hardware interprets 32-bit literals for 64-bit ops.
-      if len(data) >= cls._size() + 4:
-        lit32 = int.from_bytes(data[cls._size():cls._size()+4], 'little')
-        inst._literal = (lit32 << 32) if inst._is_64bit_op() else lit32
-    return inst
-
-  def __repr__(self):
-    # Use _fields order and exclude fields that are 0/default (for consistent repr after roundtrip)
-    def is_zero(v): return (isinstance(v, int) and v == 0) or (isinstance(v, VGPR) and v.idx == 0 and v.count == 1)
-    items = [(k, self._values[k]) for k in self._fields if k in self._values and k != 'encoding'
-             and not (is_zero(self._values[k]) and k not in {'op'})]
-    lit = f", literal={hex(self._literal)}" if self._literal is not None else ""
-    return f"{self.__class__.__name__}({', '.join(f'{k}={v}' for k, v in items)}{lit})"
-
-  def __eq__(self, other):
-    if not isinstance(other, Inst): return NotImplemented
-    return self.__class__ == other.__class__ and self._values == other._values and self._literal == other._literal
-
-  def __hash__(self): return hash((self.__class__.__name__, tuple(sorted((k, repr(v)) for k, v in self._values.items())), self._literal))
-
-  def disasm(self) -> str:
-    from extra.assembly.amd.asm import disasm
-    return disasm(self)
-
-class Inst32(Inst): pass
-class Inst64(Inst): pass
-
-# ═══════════════════════════════════════════════════════════════════════════════
-# CODE GENERATION: generates autogen/__init__.py by parsing AMD ISA PDFs
-# Supports both RDNA3.5 and CDNA4 instruction set PDFs - auto-detects format
-# ═══════════════════════════════════════════════════════════════════════════════
-
-PDF_URLS = {
-  "rdna3": "https://docs.amd.com/api/khub/documents/UVVZM22UN7tMUeiW_4ShTQ/content",  # RDNA3.5
-  "rdna4": "https://docs.amd.com/api/khub/documents/uQpkEvk3pv~kfAb2x~j4uw/content",
-  "cdna": ["https://www.amd.com/content/dam/amd/en/documents/instinct-tech-docs/instruction-set-architectures/amd-instinct-mi300-cdna3-instruction-set-architecture.pdf",
-           "https://www.amd.com/content/dam/amd/en/documents/instinct-tech-docs/instruction-set-architectures/amd-instinct-cdna4-instruction-set-architecture.pdf"],
-}
-FIELD_TYPES = {'SSRC0': 'SSrc', 'SSRC1': 'SSrc', 'SOFFSET': 'SSrc', 'SADDR': 'SSrc', 'SRC0': 'Src', 'SRC1': 'Src', 'SRC2': 'Src',
-  'SDST': 'SGPRField', 'SBASE': 'SGPRField', 'SDATA': 'SGPRField', 'SRSRC': 'SGPRField', 'VDST': 'VGPRField', 'VSRC1': 'VGPRField', 'VDATA': 'VGPRField',
-  'VADDR': 'VGPRField', 'ADDR': 'VGPRField', 'DATA': 'VGPRField', 'DATA0': 'VGPRField', 'DATA1': 'VGPRField', 'SIMM16': 'SImm', 'OFFSET': 'Imm',
-  'OPX': 'VOPDOp', 'OPY': 'VOPDOp', 'SRCX0': 'Src', 'SRCY0': 'Src', 'VSRCX1': 'VGPRField', 'VSRCY1': 'VGPRField', 'VDSTX': 'VGPRField', 'VDSTY': 'VDSTYEnc'}
-FIELD_ORDER = {
-  'SOP2': ['op', 'sdst', 'ssrc0', 'ssrc1'], 'SOP1': ['op', 'sdst', 'ssrc0'], 'SOPC': ['op', 'ssrc0', 'ssrc1'],
-  'SOPK': ['op', 'sdst', 'simm16'], 'SOPP': ['op', 'simm16'], 'VOP1': ['op', 'vdst', 'src0'], 'VOPC': ['op', 'src0', 'vsrc1'],
-  'VOP2': ['op', 'vdst', 'src0', 'vsrc1'], 'VOP3SD': ['op', 'vdst', 'sdst', 'src0', 'src1', 'src2', 'clmp'],
-  'SMEM': ['op', 'sdata', 'sbase', 'soffset', 'offset', 'glc', 'dlc'], 'DS': ['op', 'vdst', 'addr', 'data0', 'data1'],
-  'VOP3': ['op', 'vdst', 'src0', 'src1', 'src2', 'omod', 'neg', 'abs', 'clmp', 'opsel'],
-  'VOP3P': ['op', 'vdst', 'src0', 'src1', 'src2', 'neg', 'neg_hi', 'opsel', 'opsel_hi', 'clmp'],
-  'FLAT': ['op', 'vdst', 'addr', 'data', 'saddr', 'offset', 'seg', 'dlc', 'glc', 'slc'],
-  'MUBUF': ['op', 'vdata', 'vaddr', 'srsrc', 'soffset', 'offset', 'offen', 'idxen', 'glc', 'dlc', 'slc', 'tfe'],
-  'MTBUF': ['op', 'vdata', 'vaddr', 'srsrc', 'soffset', 'offset', 'format', 'offen', 'idxen', 'glc', 'dlc', 'slc', 'tfe'],
-  'MIMG': ['op', 'vdata', 'vaddr', 'srsrc', 'ssamp', 'dmask', 'dim', 'unrm', 'dlc', 'glc', 'slc'],
-  'EXP': ['en', 'target', 'vsrc0', 'vsrc1', 'vsrc2', 'vsrc3', 'done', 'row'],
-  'VINTERP': ['op', 'vdst', 'src0', 'src1', 'src2', 'waitexp', 'clmp', 'opsel', 'neg'],
-  'VOPD': ['opx', 'opy', 'vdstx', 'vdsty', 'srcx0', 'vsrcx1', 'srcy0', 'vsrcy1'],
-  'LDSDIR': ['op', 'vdst', 'attr', 'attr_chan', 'wait_va']}
-SRC_EXTRAS = {233: 'DPP8', 234: 'DPP8FI', 250: 'DPP16', 251: 'VCCZ', 252: 'EXECZ', 254: 'LDS_DIRECT'}
-FLOAT_MAP = {'0.5': 'POS_HALF', '-0.5': 'NEG_HALF', '1.0': 'POS_ONE', '-1.0': 'NEG_ONE', '2.0': 'POS_TWO', '-2.0': 'NEG_TWO',
-  '4.0': 'POS_FOUR', '-4.0': 'NEG_FOUR', '1/(2*PI)': 'INV_2PI', '0': 'ZERO'}
-
-def _parse_bits(s: str) -> tuple[int, int] | None:
-  import re
-  return (int(m.group(1)), int(m.group(2) or m.group(1))) if (m := re.match(r'\[(\d+)(?::(\d+))?\]', s)) else None
-
-def _parse_fields_table(table: list, fmt: str, enums: set[str]) -> list[tuple]:
-  import re
-  fields = []
-  for row in table[1:]:
-    if not row or not row[0]: continue
-    name, bits_str = row[0].split('\n')[0].strip(), (row[1] or '').split('\n')[0].strip()
-    if not (bits := _parse_bits(bits_str)): continue
-    enc_val, hi, lo = None, bits[0], bits[1]
-    if name == 'ENCODING' and row[2]:
-      # Handle both RDNA3 ('bXX) and CDNA4 (Must be: XX) encoding formats
-      if m := re.search(r"(?:'b|Must be:\s*)([01_]+)", row[2]):
-        enc_bits = m.group(1).replace('_', '')
-        enc_val = int(enc_bits, 2)
-        declared_width, actual_width = hi - lo + 1, len(enc_bits)
-        if actual_width > declared_width: lo = hi - actual_width + 1
-    ftype = f"{fmt}Op" if name == 'OP' and f"{fmt}Op" in enums else FIELD_TYPES.get(name.upper())
-    fields.append((name, hi, lo, enc_val, ftype))
-  return fields
-
-def _parse_single_pdf(url: str) -> dict:
-  """Parse a single PDF and return raw data (formats, enums, src_enum, doc_name, is_cdna)."""
-  import re, pdfplumber
-  from tinygrad.helpers import fetch
-
-  pdf = pdfplumber.open(fetch(url))
-
-  # Auto-detect document type from first page
-  first_page_text = pdf.pages[0].extract_text() or ''
-  is_cdna4 = 'CDNA4' in first_page_text or 'CDNA 4' in first_page_text
-  is_cdna3 = 'CDNA3' in first_page_text or 'CDNA 3' in first_page_text or 'MI300' in first_page_text
-  is_cdna = is_cdna3 or is_cdna4
-  is_rdna4 = 'RDNA4' in first_page_text or 'RDNA 4' in first_page_text
-  is_rdna35 = 'RDNA3.5' in first_page_text or 'RDNA 3.5' in first_page_text  # Check 3.5 before 3
-  is_rdna3 = not is_rdna35 and ('RDNA3' in first_page_text or 'RDNA 3' in first_page_text)
-  doc_name = "CDNA4" if is_cdna4 else "CDNA3" if is_cdna3 else "RDNA4" if is_rdna4 else "RDNA3.5" if is_rdna35 else "RDNA3" if is_rdna3 else "Unknown"
-
-  # Find the "Microcode Formats" section - search for SOP2 format definition
-  microcode_start = None
-  total_pages = len(pdf.pages)
-  # Search from likely locations (formats are typically 20-95% through the document - RDNA3 has them at ~25%)
-  for i in range(int(total_pages * 0.2), total_pages):
-    text = pdf.pages[i].extract_text() or ''
-    # Look for "X.Y.Z. SOP2" section header or "Chapter X. Microcode Formats"
-    if re.search(r'\d+\.\d+\.\d+\.\s+SOP2\b', text) or re.search(r'Chapter \d+\.\s+Microcode Formats', text):
-      microcode_start = i
-      break
-  if microcode_start is None: microcode_start = int(total_pages * 0.9)
-
-  pages = pdf.pages[microcode_start:microcode_start + 50]
-  page_texts = [p.extract_text() or '' for p in pages]
-  page_tables = [[t.extract() for t in p.find_tables()] for p in pages]
-  full_text = '\n'.join(page_texts)
-
-  # parse SSRC encoding from first page with VCC_LO
-  src_enum = dict(SRC_EXTRAS)
-  for text in page_texts[:10]:
-    if 'SSRC0' in text and 'VCC_LO' in text:
-      for m in re.finditer(r'^(\d+)\s+(\S+)', text, re.M):
-        val, name = int(m.group(1)), m.group(2).rstrip('.:')
-        if name in FLOAT_MAP: src_enum[val] = FLOAT_MAP[name]
-        elif re.match(r'^[A-Z][A-Z0-9_]*$', name): src_enum[val] = name
-      break
-
-  # parse opcode tables
-  enums: dict[str, dict[int, str]] = {}
-  for m in re.finditer(r'Table \d+\. (\w+) Opcodes(.*?)(?=Table \d+\.|\n\d+\.\d+\.\d+\.\s+\w+\s*\nDescription|$)', full_text, re.S):
-    if ops := {int(x.group(1)): x.group(2) for x in re.finditer(r'(\d+)\s+([A-Z][A-Z0-9_]+)', m.group(2))}:
-      enums[m.group(1) + "Op"] = ops
-  if vopd_m := re.search(r'Table \d+\. VOPD Y-Opcodes\n(.*?)(?=Table \d+\.|15\.\d)', full_text, re.S):
-    if ops := {int(x.group(1)): x.group(2) for x in re.finditer(r'(\d+)\s+(V_DUAL_\w+)', vopd_m.group(1))}:
-      enums["VOPDOp"] = ops
-  enum_names = set(enums.keys())
-
-  def is_fields_table(t) -> bool: return t and len(t) > 1 and t[0] and 'Field' in str(t[0][0] or '')
-  def has_encoding(fields) -> bool: return any(f[0] == 'ENCODING' for f in fields)
-  def has_header_before_fields(text) -> bool:
-    return (pos := text.find('Field Name')) != -1 and bool(re.search(r'\d+\.\d+\.\d+\.\s+\w+\s*\n', text[:pos]))
-
-  # find format headers with their page indices
-  format_headers = []
-  for i, text in enumerate(page_texts):
-    for m in re.finditer(r'\d+\.\d+\.\d+\.\s+(\w+)\s*\n?Description', text): format_headers.append((m.group(1), i, m.start()))
-    for m in re.finditer(r'\d+\.\d+\.\d+\.\s+(\w+)\s*\n', text):
-      fmt_name = m.group(1)
-      if is_cdna and fmt_name.isupper() and len(fmt_name) >= 2:
-        format_headers.append((fmt_name, i, m.start()))
-      elif m.start() > len(text) - 200 and 'Description' not in text[m.end():] and i + 1 < len(page_texts):
-        next_text = page_texts[i + 1].lstrip()
-        if next_text.startswith('Description') or (next_text.startswith('"RDNA') and 'Description' in next_text[:200]):
-          format_headers.append((fmt_name, i, m.start()))
-
-  # parse instruction formats
-  formats: dict[str, list] = {}
-  for fmt_name, page_idx, header_pos in format_headers:
-    if fmt_name in formats: continue
-    text, tables = page_texts[page_idx], page_tables[page_idx]
-    field_pos = text.find('Field Name', header_pos)
-
-    fields = None
-    for offset in range(3):
-      if page_idx + offset >= len(pages): break
-      if offset > 0 and has_header_before_fields(page_texts[page_idx + offset]): break
-      for t in page_tables[page_idx + offset] if offset > 0 or field_pos > header_pos else []:
-        if is_fields_table(t) and (f := _parse_fields_table(t, fmt_name, enum_names)) and has_encoding(f):
-          fields = f
-          break
-      if fields: break
-
-    if not fields and field_pos > header_pos:
-      for t in tables:
-        if is_fields_table(t) and (f := _parse_fields_table(t, fmt_name, enum_names)):
-          fields = f
-          break
-
-    if not fields: continue
-    field_names = {f[0] for f in fields}
-
-    for pg_offset in range(1, 3):
-      if page_idx + pg_offset >= len(pages) or has_header_before_fields(page_texts[page_idx + pg_offset]): break
-      for t in page_tables[page_idx + pg_offset]:
-        if is_fields_table(t) and (extra := _parse_fields_table(t, fmt_name, enum_names)) and not has_encoding(extra):
-          for ef in extra:
-            if ef[0] not in field_names:
-              fields.append(ef)
-              field_names.add(ef[0])
-          break
-    formats[fmt_name] = fields
-
-  # fix known PDF errors
-  if 'SMEM' in formats:
-    formats['SMEM'] = [(n, 13 if n == 'DLC' else 14 if n == 'GLC' else h, 13 if n == 'DLC' else 14 if n == 'GLC' else l, e, t)
-                       for n, h, l, e, t in formats['SMEM']]
-
-  return {"formats": formats, "enums": enums, "src_enum": src_enum, "doc_name": doc_name, "is_cdna": is_cdna}
-
-def _merge_results(results: list[dict]) -> dict:
-  """Merge multiple PDF parse results into a superset. Asserts if any conflicts."""
-  merged = {"formats": {}, "enums": {}, "src_enum": dict(SRC_EXTRAS), "doc_names": [], "is_cdna": False}
-  for r in results:
-    merged["doc_names"].append(r["doc_name"])
-    merged["is_cdna"] = merged["is_cdna"] or r["is_cdna"]
-    # Merge src_enum (union, assert no conflicts)
-    for val, name in r["src_enum"].items():
-      if val in merged["src_enum"]:
-        assert merged["src_enum"][val] == name, f"SrcEnum conflict: {val} = {merged['src_enum'][val]} vs {name}"
-      else:
-        merged["src_enum"][val] = name
-    # Merge enums (union of ops per enum, assert no conflicts)
-    for enum_name, ops in r["enums"].items():
-      if enum_name not in merged["enums"]: merged["enums"][enum_name] = {}
-      for val, name in ops.items():
-        if val in merged["enums"][enum_name]:
-          assert merged["enums"][enum_name][val] == name, f"{enum_name} conflict: {val} = {merged['enums'][enum_name][val]} vs {name}"
-        else:
-          merged["enums"][enum_name][val] = name
-    # Merge formats (union of fields, assert no bit position conflicts for same field name)
-    for fmt_name, fields in r["formats"].items():
-      if fmt_name not in merged["formats"]:
-        merged["formats"][fmt_name] = list(fields)
-      else:
-        existing = {f[0]: (f[1], f[2]) for f in merged["formats"][fmt_name]}  # name -> (hi, lo)
-        for f in fields:
-          name, hi, lo = f[0], f[1], f[2]
-          if name in existing:
-            assert existing[name] == (hi, lo), f"Format {fmt_name} field {name} conflict: bits {existing[name]} vs ({hi}, {lo})"
-          else:
-            merged["formats"][fmt_name].append(f)
-  return merged
-
-def generate(output_path: str | None = None, arch: str = "rdna3") -> dict:
-  """Generate instruction definitions from AMD ISA PDF(s). Returns dict with formats for testing."""
-  urls = PDF_URLS[arch]
-  if isinstance(urls, str): urls = [urls]
-
-  # Parse all PDFs and merge
-  results = [_parse_single_pdf(url) for url in urls]
-  if len(results) == 1:
-    merged = results[0]
-    doc_name = merged["doc_name"]
-  else:
-    merged = _merge_results(results)
-    doc_name = "+".join(merged["doc_names"])
-
-  formats, enums, src_enum = merged["formats"], merged["enums"], merged["src_enum"]
-
-  # generate output
-  def enum_lines(name, items):
-    return [f"class {name}(IntEnum):"] + [f"  {n} = {v}" for v, n in sorted(items.items())] + [""]
-  def field_key(f): return order.index(f[0].lower()) if f[0].lower() in order else 1000
-  lines = [f"# autogenerated from AMD {doc_name} ISA PDF by dsl.py - do not edit", "from enum import IntEnum",
-           "from typing import Annotated",
-           "from extra.assembly.amd.dsl import bits, BitField, Inst32, Inst64, SGPR, VGPR, TTMP as TTMP, s as s, v as v, ttmp as ttmp, SSrc, Src, SImm, Imm, VDSTYEnc, SGPRField, VGPRField",
-           "import functools", ""]
-  lines += enum_lines("SrcEnum", src_enum) + sum([enum_lines(n, ops) for n, ops in sorted(enums.items())], [])
-  # Format-specific field defaults (verified against LLVM test vectors)
-  format_defaults = {'VOP3P': {'opsel_hi': 3, 'opsel_hi2': 1}}
-  lines.append("# instruction formats")
-  for fmt_name, fields in sorted(formats.items()):
-    base = "Inst64" if max(f[1] for f in fields) > 31 or fmt_name == 'VOP3SD' else "Inst32"
-    order = FIELD_ORDER.get(fmt_name, [])
-    lines.append(f"class {fmt_name}({base}):")
-    if enc := next((f for f in fields if f[0] == 'ENCODING'), None):
-      enc_str = f"bits[{enc[1]}:{enc[2]}] == 0b{enc[3]:b}" if enc[1] != enc[2] else f"bits[{enc[1]}] == {enc[3]}"
-      lines.append(f"  encoding = {enc_str}")
-    if defaults := format_defaults.get(fmt_name):
-      lines.append(f"  _defaults = {defaults}")
-    for name, hi, lo, _, ftype in sorted([f for f in fields if f[0] != 'ENCODING'], key=field_key):
-      if ftype and ftype.endswith('Op'):
-        ann = f":Annotated[BitField, {ftype}]"
-      else:
-        ann = f":{ftype}" if ftype else ""
-      lines.append(f"  {name.lower()}{ann} = bits[{hi}]" if hi == lo else f"  {name.lower()}{ann} = bits[{hi}:{lo}]")
-    lines.append("")
-  lines.append("# instruction helpers")
-  for cls_name, ops in sorted(enums.items()):
-    fmt = cls_name[:-2]
-    for op_val, name in sorted(ops.items()):
-      seg = {"GLOBAL": ", seg=2", "SCRATCH": ", seg=2"}.get(fmt, "")
-      tgt = {"GLOBAL": "FLAT, GLOBALOp", "SCRATCH": "FLAT, SCRATCHOp"}.get(fmt, f"{fmt}, {cls_name}")
-      if fmt in formats or fmt in ("GLOBAL", "SCRATCH"):
-        if fmt in ("VOP1", "VOP2", "VOPC"):
-          suffix = "_e32"
-        elif fmt == "VOP3" and op_val < 512:
-          suffix = "_e64"
-        else:
-          suffix = ""
-        if name in ('V_FMAMK_F32', 'V_FMAMK_F16'):
-          lines.append(f"def {name.lower()}{suffix}(vdst, src0, K, vsrc1): return {fmt}({cls_name}.{name}, vdst, src0, vsrc1, literal=K)")
-        elif name in ('V_FMAAK_F32', 'V_FMAAK_F16'):
-          lines.append(f"def {name.lower()}{suffix}(vdst, src0, vsrc1, K): return {fmt}({cls_name}.{name}, vdst, src0, vsrc1, literal=K)")
-        else:
-          lines.append(f"{name.lower()}{suffix} = functools.partial({tgt}.{name}{seg})")
-  skip_exports = {'DPP8', 'DPP16'}
-  src_names = {name for _, name in src_enum.items()}
-  lines += [""] + [f"{name} = SrcEnum.{name}" for _, name in sorted(src_enum.items()) if name not in skip_exports]
-  if "NULL" in src_names: lines.append("OFF = NULL\n")
-
-  if output_path is not None:
-    import pathlib
-    pathlib.Path(output_path).write_text('\n'.join(lines))
-  return {"formats": formats, "enums": enums, "src_enum": src_enum}
-
-if __name__ == "__main__":
-  import argparse
-  parser = argparse.ArgumentParser(description="Generate instruction definitions from AMD ISA PDF")
-  parser.add_argument("--arch", choices=list(PDF_URLS.keys()) + ["all"], default="rdna3", help="Target architecture (default: rdna3)")
-  args = parser.parse_args()
-  if args.arch == "all":
-    for arch in PDF_URLS.keys():
-      result = generate(f"extra/assembly/amd/autogen/{arch}/__init__.py", arch=arch)
-      print(f"{arch}: generated SrcEnum ({len(result['src_enum'])}) + {len(result['enums'])} opcode enums + {len(result['formats'])} format classes")
-  else:
-    result = generate(f"extra/assembly/amd/autogen/{args.arch}/__init__.py", arch=args.arch)
-    print(f"generated SrcEnum ({len(result['src_enum'])}) + {len(result['enums'])} opcode enums + {len(result['formats'])} format classes")

extra/assembly/amd/emu.py

@@ -1,718 +0,0 @@
-# RDNA3 emulator - executes compiled pseudocode from AMD ISA PDF
-# mypy: ignore-errors
-from __future__ import annotations
-import ctypes, os
-from extra.assembly.amd.dsl import Inst, RawImm
-from extra.assembly.amd.pcode import _f32, _i32, _sext, _f16, _i16, _f64, _i64
-from extra.assembly.amd.autogen.rdna3.gen_pcode import get_compiled_functions
-from extra.assembly.amd.autogen.rdna3 import (
-  SOP1, SOP2, SOPC, SOPK, SOPP, SMEM, VOP1, VOP2, VOP3, VOP3SD, VOP3P, VOPC, DS, FLAT, VOPD, SrcEnum,
-  SOP1Op, SOP2Op, SOPCOp, SOPKOp, SOPPOp, SMEMOp, VOP1Op, VOP2Op, VOP3Op, VOP3SDOp, VOP3POp, VOPCOp, DSOp, FLATOp, GLOBALOp, VOPDOp
-)
-
-Program = dict[int, Inst]
-WAVE_SIZE, SGPR_COUNT, VGPR_COUNT = 32, 128, 256
-VCC_LO, VCC_HI, NULL, EXEC_LO, EXEC_HI, SCC = SrcEnum.VCC_LO, SrcEnum.VCC_HI, SrcEnum.NULL, SrcEnum.EXEC_LO, SrcEnum.EXEC_HI, SrcEnum.SCC
-
-# VOP3 ops that use 64-bit operands (and thus 64-bit literals when src is 255)
-# Exception: V_LDEXP_F64 has 32-bit integer src1, so literal should NOT be 64-bit when src1=255
-_VOP3_64BIT_OPS = {op.value for op in VOP3Op if op.name.endswith(('_F64', '_B64', '_I64', '_U64'))}
-# Ops where src1 is 32-bit (exponent/shift amount) even though the op name suggests 64-bit
-_VOP3_64BIT_OPS_32BIT_SRC1 = {VOP3Op.V_LDEXP_F64.value}
-# Ops with 16-bit types in name (for source/dest handling)
-_VOP3_16BIT_OPS = {op for op in VOP3Op if any(s in op.name for s in ('_F16', '_B16', '_I16', '_U16'))}
-_VOP1_16BIT_OPS = {op for op in VOP1Op if any(s in op.name for s in ('_F16', '_B16', '_I16', '_U16'))}
-_VOP2_16BIT_OPS = {op for op in VOP2Op if any(s in op.name for s in ('_F16', '_B16', '_I16', '_U16'))}
-# CVT ops with 32/64-bit source (despite 16-bit in name)
-_CVT_32_64_SRC_OPS = {op for op in VOP3Op if op.name.startswith('V_CVT_') and op.name.endswith(('_F32', '_I32', '_U32', '_F64', '_I64', '_U64'))} | \
-                     {op for op in VOP1Op if op.name.startswith('V_CVT_') and op.name.endswith(('_F32', '_I32', '_U32', '_F64', '_I64', '_U64'))}
-# 16-bit dst ops (PACK has 32-bit dst despite F16 in name)
-_VOP3_16BIT_DST_OPS = {op for op in _VOP3_16BIT_OPS if 'PACK' not in op.name}
-_VOP1_16BIT_DST_OPS = {op for op in _VOP1_16BIT_OPS if 'PACK' not in op.name}
-
-# Inline constants for src operands 128-254. Build tables for f32, f16, and f64 formats.
-import struct as _struct
-_FLOAT_CONSTS = {SrcEnum.POS_HALF: 0.5, SrcEnum.NEG_HALF: -0.5, SrcEnum.POS_ONE: 1.0, SrcEnum.NEG_ONE: -1.0,
-                 SrcEnum.POS_TWO: 2.0, SrcEnum.NEG_TWO: -2.0, SrcEnum.POS_FOUR: 4.0, SrcEnum.NEG_FOUR: -4.0, SrcEnum.INV_2PI: 0.15915494309189535}
-def _build_inline_consts(neg_mask, float_to_bits):
-  tbl = list(range(65)) + [((-i) & neg_mask) for i in range(1, 17)] + [0] * (127 - 81)
-  for k, v in _FLOAT_CONSTS.items(): tbl[k - 128] = float_to_bits(v)
-  return tbl
-_INLINE_CONSTS = _build_inline_consts(0xffffffff, lambda f: _struct.unpack('<I', _struct.pack('<f', f))[0])
-_INLINE_CONSTS_F16 = _build_inline_consts(0xffff, lambda f: _struct.unpack('<H', _struct.pack('<e', f))[0])
-_INLINE_CONSTS_F64 = _build_inline_consts(0xffffffffffffffff, lambda f: _struct.unpack('<Q', _struct.pack('<d', f))[0])
-
-# Memory access
-_valid_mem_ranges: list[tuple[int, int]] = []
-def set_valid_mem_ranges(ranges: set[tuple[int, int]]) -> None: _valid_mem_ranges.clear(); _valid_mem_ranges.extend(ranges)
-def _mem_valid(addr: int, size: int) -> bool:
-  for s, z in _valid_mem_ranges:
-    if s <= addr and addr + size <= s + z: return True
-  return not _valid_mem_ranges
-def _ctypes_at(addr: int, size: int): return (ctypes.c_uint8 if size == 1 else ctypes.c_uint16 if size == 2 else ctypes.c_uint32).from_address(addr)
-def mem_read(addr: int, size: int) -> int: return _ctypes_at(addr, size).value if _mem_valid(addr, size) else 0
-def mem_write(addr: int, size: int, val: int) -> None:
-  if _mem_valid(addr, size): _ctypes_at(addr, size).value = val
-
-# Memory op tables (not pseudocode - these are format descriptions)
-def _mem_ops(ops, suffix_map):
-  return {getattr(e, f"{p}_{s}"): v for e in ops for s, v in suffix_map.items() for p in [e.__name__.replace("Op", "")]}
-_LOAD_MAP = {'LOAD_B32': (1,4,0), 'LOAD_B64': (2,4,0), 'LOAD_B96': (3,4,0), 'LOAD_B128': (4,4,0), 'LOAD_U8': (1,1,0), 'LOAD_I8': (1,1,1), 'LOAD_U16': (1,2,0), 'LOAD_I16': (1,2,1)}
-_STORE_MAP = {'STORE_B32': (1,4), 'STORE_B64': (2,4), 'STORE_B96': (3,4), 'STORE_B128': (4,4), 'STORE_B8': (1,1), 'STORE_B16': (1,2)}
-FLAT_LOAD, FLAT_STORE = _mem_ops([GLOBALOp, FLATOp], _LOAD_MAP), _mem_ops([GLOBALOp, FLATOp], _STORE_MAP)
-# D16 ops: load/store 16-bit to lower or upper half of VGPR. Format: (size, sign, hi) where hi=1 means upper 16 bits
-_D16_LOAD_MAP = {'LOAD_D16_U8': (1,0,0), 'LOAD_D16_I8': (1,1,0), 'LOAD_D16_B16': (2,0,0),
-                 'LOAD_D16_HI_U8': (1,0,1), 'LOAD_D16_HI_I8': (1,1,1), 'LOAD_D16_HI_B16': (2,0,1)}
-_D16_STORE_MAP = {'STORE_D16_HI_B8': (1,1), 'STORE_D16_HI_B16': (2,1)}  # (size, hi)
-FLAT_D16_LOAD = _mem_ops([GLOBALOp, FLATOp], _D16_LOAD_MAP)
-FLAT_D16_STORE = _mem_ops([GLOBALOp, FLATOp], _D16_STORE_MAP)
-DS_LOAD = {DSOp.DS_LOAD_B32: (1,4,0), DSOp.DS_LOAD_B64: (2,4,0), DSOp.DS_LOAD_B128: (4,4,0), DSOp.DS_LOAD_U8: (1,1,0), DSOp.DS_LOAD_I8: (1,1,1), DSOp.DS_LOAD_U16: (1,2,0), DSOp.DS_LOAD_I16: (1,2,1)}
-DS_STORE = {DSOp.DS_STORE_B32: (1,4), DSOp.DS_STORE_B64: (2,4), DSOp.DS_STORE_B128: (4,4), DSOp.DS_STORE_B8: (1,1), DSOp.DS_STORE_B16: (1,2)}
-SMEM_LOAD = {SMEMOp.S_LOAD_B32: 1, SMEMOp.S_LOAD_B64: 2, SMEMOp.S_LOAD_B128: 4, SMEMOp.S_LOAD_B256: 8, SMEMOp.S_LOAD_B512: 16}
-
-# VOPD op -> VOP3 op mapping (VOPD is dual-issue of VOP1/VOP2 ops, use VOP3 enums for pseudocode lookup)
-_VOPD_TO_VOP = {
-  VOPDOp.V_DUAL_FMAC_F32: VOP3Op.V_FMAC_F32, VOPDOp.V_DUAL_FMAAK_F32: VOP2Op.V_FMAAK_F32, VOPDOp.V_DUAL_FMAMK_F32: VOP2Op.V_FMAMK_F32,
-  VOPDOp.V_DUAL_MUL_F32: VOP3Op.V_MUL_F32, VOPDOp.V_DUAL_ADD_F32: VOP3Op.V_ADD_F32, VOPDOp.V_DUAL_SUB_F32: VOP3Op.V_SUB_F32,
-  VOPDOp.V_DUAL_SUBREV_F32: VOP3Op.V_SUBREV_F32, VOPDOp.V_DUAL_MUL_DX9_ZERO_F32: VOP3Op.V_MUL_DX9_ZERO_F32,
-  VOPDOp.V_DUAL_MOV_B32: VOP3Op.V_MOV_B32, VOPDOp.V_DUAL_CNDMASK_B32: VOP3Op.V_CNDMASK_B32,
-  VOPDOp.V_DUAL_MAX_F32: VOP3Op.V_MAX_F32, VOPDOp.V_DUAL_MIN_F32: VOP3Op.V_MIN_F32,
-  VOPDOp.V_DUAL_ADD_NC_U32: VOP3Op.V_ADD_NC_U32, VOPDOp.V_DUAL_LSHLREV_B32: VOP3Op.V_LSHLREV_B32, VOPDOp.V_DUAL_AND_B32: VOP3Op.V_AND_B32,
-}
-
-# Compiled pseudocode functions (lazy loaded)
-_COMPILED: dict | None = None
-
-def _get_compiled() -> dict:
-  global _COMPILED
-  if _COMPILED is None: _COMPILED = get_compiled_functions()
-  return _COMPILED
-
-class WaveState:
-  __slots__ = ('sgpr', 'vgpr', 'scc', 'pc', 'literal', '_pend_sgpr')
-  def __init__(self):
-    self.sgpr, self.vgpr = [0] * SGPR_COUNT, [[0] * VGPR_COUNT for _ in range(WAVE_SIZE)]
-    self.sgpr[EXEC_LO], self.scc, self.pc, self.literal, self._pend_sgpr = 0xffffffff, 0, 0, 0, {}
-
-  @property
-  def vcc(self) -> int: return self.sgpr[VCC_LO] | (self.sgpr[VCC_HI] << 32)
-  @vcc.setter
-  def vcc(self, v: int): self.sgpr[VCC_LO], self.sgpr[VCC_HI] = v & 0xffffffff, (v >> 32) & 0xffffffff
-  @property
-  def exec_mask(self) -> int: return self.sgpr[EXEC_LO] | (self.sgpr[EXEC_HI] << 32)
-  @exec_mask.setter
-  def exec_mask(self, v: int): self.sgpr[EXEC_LO], self.sgpr[EXEC_HI] = v & 0xffffffff, (v >> 32) & 0xffffffff
-
-  def rsgpr(self, i: int) -> int: return 0 if i == NULL else self.scc if i == SCC else self.sgpr[i] if i < SGPR_COUNT else 0
-  def wsgpr(self, i: int, v: int):
-    if i < SGPR_COUNT and i != NULL: self.sgpr[i] = v & 0xffffffff
-  def rsgpr64(self, i: int) -> int: return self.rsgpr(i) | (self.rsgpr(i+1) << 32)
-  def wsgpr64(self, i: int, v: int): self.wsgpr(i, v & 0xffffffff); self.wsgpr(i+1, (v >> 32) & 0xffffffff)
-
-  def rsrc(self, v: int, lane: int) -> int:
-    if v < SGPR_COUNT: return self.sgpr[v]
-    if v == SCC: return self.scc
-    if v < 255: return _INLINE_CONSTS[v - 128]
-    if v == 255: return self.literal
-    return self.vgpr[lane][v - 256] if v <= 511 else 0
-
-  def rsrc_f16(self, v: int, lane: int) -> int:
-    """Read source operand for VOP3P packed f16 operations. Uses f16 inline constants."""
-    if v < SGPR_COUNT: return self.sgpr[v]
-    if v == SCC: return self.scc
-    if v < 255: return _INLINE_CONSTS_F16[v - 128]
-    if v == 255: return self.literal
-    return self.vgpr[lane][v - 256] if v <= 511 else 0
-
-  def rsrc64(self, v: int, lane: int) -> int:
-    """Read 64-bit source operand. For inline constants, returns 64-bit representation."""
-    # Inline constants 128-254 need special handling for 64-bit ops
-    if 128 <= v < 255: return _INLINE_CONSTS_F64[v - 128]
-    if v == 255: return self.literal  # 32-bit literal, caller handles extension
-    return self.rsrc(v, lane) | ((self.rsrc(v+1, lane) if v < VCC_LO or 256 <= v <= 511 else 0) << 32)
-
-  def pend_sgpr_lane(self, reg: int, lane: int, val: int):
-    if reg not in self._pend_sgpr: self._pend_sgpr[reg] = 0
-    if val: self._pend_sgpr[reg] |= (1 << lane)
-  def commit_pends(self):
-    for reg, val in self._pend_sgpr.items(): self.sgpr[reg] = val
-    self._pend_sgpr.clear()
-
-# Instruction decode
-def decode_format(word: int) -> tuple[type[Inst] | None, bool]:
-  hi2 = (word >> 30) & 0x3
-  if hi2 == 0b11:
-    enc = (word >> 26) & 0xf
-    if enc == 0b1101: return SMEM, True
-    if enc == 0b0101:
-      op = (word >> 16) & 0x3ff
-      return (VOP3SD, True) if op in (288, 289, 290, 764, 765, 766, 767, 768, 769, 770) else (VOP3, True)
-    return {0b0011: (VOP3P, True), 0b0110: (DS, True), 0b0111: (FLAT, True), 0b0010: (VOPD, True)}.get(enc, (None, True))
-  if hi2 == 0b10:
-    enc = (word >> 23) & 0x7f
-    return {0b1111101: (SOP1, False), 0b1111110: (SOPC, False), 0b1111111: (SOPP, False)}.get(enc, (SOPK, False) if ((word >> 28) & 0xf) == 0b1011 else (SOP2, False))
-  enc = (word >> 25) & 0x7f
-  return (VOPC, False) if enc == 0b0111110 else (VOP1, False) if enc == 0b0111111 else (VOP2, False)
-
-def _unwrap(v) -> int: return v.val if isinstance(v, RawImm) else v.value if hasattr(v, 'value') else v
-
-def decode_program(data: bytes) -> Program:
-  result: Program = {}
-  i = 0
-  while i < len(data):
-    word = int.from_bytes(data[i:i+4], 'little')
-    inst_class, is_64 = decode_format(word)
-    if inst_class is None: i += 4; continue
-    base_size = 8 if is_64 else 4
-    # Pass enough data for potential 64-bit literal (base + 8 bytes max)
-    inst = inst_class.from_bytes(data[i:i+base_size+8])
-    for name, val in inst._values.items(): setattr(inst, name, _unwrap(val))
-    # from_bytes already handles literal reading - only need fallback for cases it doesn't handle
-    if inst._literal is None:
-      has_literal = any(getattr(inst, fld, None) == 255 for fld in ('src0', 'src1', 'src2', 'ssrc0', 'ssrc1', 'srcx0', 'srcy0'))
-      if inst_class == VOP2 and inst.op in (44, 45, 55, 56): has_literal = True
-      if inst_class == VOPD and (inst.opx in (1, 2) or inst.opy in (1, 2)): has_literal = True
-      if inst_class == SOP2 and inst.op in (69, 70): has_literal = True
-      if has_literal:
-        # For 64-bit ops, the 32-bit literal is placed in HIGH 32 bits (low 32 bits = 0)
-        # Exception: some ops have mixed src sizes (e.g., V_LDEXP_F64 has 32-bit src1)
-        op_val = inst._values.get('op')
-        if hasattr(op_val, 'value'): op_val = op_val.value
-        is_64bit = inst_class is VOP3 and op_val in _VOP3_64BIT_OPS
-        # Don't treat literal as 64-bit if the op has 32-bit src1 and src1 is the literal
-        if is_64bit and op_val in _VOP3_64BIT_OPS_32BIT_SRC1 and getattr(inst, 'src1', None) == 255:
-          is_64bit = False
-        lit32 = int.from_bytes(data[i+base_size:i+base_size+4], 'little')
-        inst._literal = (lit32 << 32) if is_64bit else lit32
-    inst._words = inst.size() // 4
-    result[i // 4] = inst
-    i += inst._words * 4
-  return result
-
-# ═══════════════════════════════════════════════════════════════════════════════
-# EXECUTION - All ALU ops use pseudocode from PDF
-# ═══════════════════════════════════════════════════════════════════════════════
-
-def exec_scalar(st: WaveState, inst: Inst) -> int:
-  """Execute scalar instruction. Returns PC delta or negative for special cases."""
-  compiled = _get_compiled()
-  inst_type = type(inst)
-
-  # SOPP: control flow (not ALU)
-  if inst_type is SOPP:
-    op = inst.op
-    if op == SOPPOp.S_ENDPGM: return -1
-    if op == SOPPOp.S_BARRIER: return -2
-    if op == SOPPOp.S_BRANCH: return _sext(inst.simm16, 16)
-    if op == SOPPOp.S_CBRANCH_SCC0: return _sext(inst.simm16, 16) if st.scc == 0 else 0
-    if op == SOPPOp.S_CBRANCH_SCC1: return _sext(inst.simm16, 16) if st.scc == 1 else 0
-    if op == SOPPOp.S_CBRANCH_VCCZ: return _sext(inst.simm16, 16) if (st.vcc & 0xffffffff) == 0 else 0
-    if op == SOPPOp.S_CBRANCH_VCCNZ: return _sext(inst.simm16, 16) if (st.vcc & 0xffffffff) != 0 else 0
-    if op == SOPPOp.S_CBRANCH_EXECZ: return _sext(inst.simm16, 16) if st.exec_mask == 0 else 0
-    if op == SOPPOp.S_CBRANCH_EXECNZ: return _sext(inst.simm16, 16) if st.exec_mask != 0 else 0
-    # Valid SOPP range is 0-61 (max defined opcode); anything above is invalid
-    if op > 61: raise NotImplementedError(f"Invalid SOPP opcode {op}")
-    return 0  # waits, hints, nops
-
-  # SMEM: memory loads (not ALU)
-  if inst_type is SMEM:
-    addr = st.rsgpr64(inst.sbase * 2) + _sext(inst.offset, 21)
-    if inst.soffset not in (NULL, 0x7f): addr += st.rsrc(inst.soffset, 0)
-    if (cnt := SMEM_LOAD.get(inst.op)) is None: raise NotImplementedError(f"SMEM op {inst.op}")
-    for i in range(cnt): st.wsgpr(inst.sdata + i, mem_read((addr + i * 4) & 0xffffffffffffffff, 4))
-    return 0
-
-  # SOP1: special handling for ops not in pseudocode
-  if inst_type is SOP1:
-    op = SOP1Op(inst.op)
-    # S_GETPC_B64: Get program counter (PC is stored as byte offset, convert from words)
-    if op == SOP1Op.S_GETPC_B64:
-      pc_bytes = st.pc * 4  # PC is in words, convert to bytes
-      st.wsgpr64(inst.sdst, pc_bytes)
-      return 0
-    # S_SETPC_B64: Set program counter to source value (indirect jump)
-    # Returns delta such that st.pc + inst_words + delta = target_words
-    if op == SOP1Op.S_SETPC_B64:
-      target_bytes = st.rsrc64(inst.ssrc0, 0)
-      target_words = target_bytes // 4
-      inst_words = 1  # SOP1 is always 1 word
-      return target_words - st.pc - inst_words
-
-  # Get op enum and lookup compiled function
-  if inst_type is SOP1: op_cls, ssrc0, sdst = SOP1Op, inst.ssrc0, inst.sdst
-  elif inst_type is SOP2: op_cls, ssrc0, sdst = SOP2Op, inst.ssrc0, inst.sdst
-  elif inst_type is SOPC: op_cls, ssrc0, sdst = SOPCOp, inst.ssrc0, None
-  elif inst_type is SOPK: op_cls, ssrc0, sdst = SOPKOp, inst.sdst, inst.sdst  # sdst is both src and dst
-  else: raise NotImplementedError(f"Unknown scalar type {inst_type}")
-
-  op = op_cls(inst.op)
-  fn = compiled.get(op_cls, {}).get(op)
-  if fn is None: raise NotImplementedError(f"{op.name} not in pseudocode")
-
-  # Build context - handle 64-bit ops that need 64-bit source reads
-  # 64-bit source ops: name ends with _B64, _I64, _U64 or contains _U64, _I64 before last underscore
-  is_64bit_s0 = op.name.endswith(('_B64', '_I64', '_U64')) or '_U64_' in op.name or '_I64_' in op.name
-  is_64bit_s0s1 = op_cls is SOPCOp and op in (SOPCOp.S_CMP_EQ_U64, SOPCOp.S_CMP_LG_U64)
-  s0 = st.rsrc64(ssrc0, 0) if is_64bit_s0 or is_64bit_s0s1 else (st.rsrc(ssrc0, 0) if inst_type != SOPK else st.rsgpr(inst.sdst))
-  is_64bit_sop2 = is_64bit_s0 and inst_type is SOP2
-  s1 = st.rsrc64(inst.ssrc1, 0) if (is_64bit_sop2 or is_64bit_s0s1) else (st.rsrc(inst.ssrc1, 0) if inst_type in (SOP2, SOPC) else inst.simm16 if inst_type is SOPK else 0)
-  d0 = st.rsgpr64(sdst) if (is_64bit_s0 or is_64bit_s0s1) and sdst is not None else (st.rsgpr(sdst) if sdst is not None else 0)
-  exec_mask = st.exec_mask
-  literal = inst.simm16 if inst_type is SOPK else st.literal
-
-  # Execute compiled function
-  result = fn(s0, s1, 0, d0, st.scc, st.vcc, 0, exec_mask, literal, None, {})
-
-  # Apply results
-  if sdst is not None:
-    if result.get('d0_64'):
-      st.wsgpr64(sdst, result['d0'])
-    else:
-      st.wsgpr(sdst, result['d0'])
-  if 'scc' in result: st.scc = result['scc']
-  if 'exec' in result: st.exec_mask = result['exec']
-  if 'pc_delta' in result: return result['pc_delta']
-  return 0
-
-def exec_vector(st: WaveState, inst: Inst, lane: int, lds: bytearray | None = None) -> None:
-  """Execute vector instruction for one lane."""
-  compiled = _get_compiled()
-  inst_type, V = type(inst), st.vgpr[lane]
-
-  # Memory ops (not ALU pseudocode)
-  if inst_type is FLAT:
-    op, addr_reg, data_reg, vdst, offset, saddr = inst.op, inst.addr, inst.data, inst.vdst, _sext(inst.offset, 13), inst.saddr
-    addr = V[addr_reg] | (V[addr_reg+1] << 32)
-    addr = (st.rsgpr64(saddr) + V[addr_reg] + offset) & 0xffffffffffffffff if saddr not in (NULL, 0x7f) else (addr + offset) & 0xffffffffffffffff
-    if op in FLAT_LOAD:
-      cnt, sz, sign = FLAT_LOAD[op]
-      for i in range(cnt): val = mem_read(addr + i * sz, sz); V[vdst + i] = _sext(val, sz * 8) & 0xffffffff if sign else val
-    elif op in FLAT_STORE:
-      cnt, sz = FLAT_STORE[op]
-      for i in range(cnt): mem_write(addr + i * sz, sz, V[data_reg + i] & ((1 << (sz * 8)) - 1))
-    elif op in FLAT_D16_LOAD:
-      sz, sign, hi = FLAT_D16_LOAD[op]
-      val = mem_read(addr, sz)
-      if sign: val = _sext(val, sz * 8) & 0xffff
-      if hi: V[vdst] = (V[vdst] & 0xffff) | (val << 16)  # upper 16 bits
-      else: V[vdst] = (V[vdst] & 0xffff0000) | (val & 0xffff)  # lower 16 bits
-    elif op in FLAT_D16_STORE:
-      sz, hi = FLAT_D16_STORE[op]
-      val = (V[data_reg] >> 16) & 0xffff if hi else V[data_reg] & 0xffff
-      mem_write(addr, sz, val & ((1 << (sz * 8)) - 1))
-    else: raise NotImplementedError(f"FLAT op {op}")
-    return
-
-  if inst_type is DS:
-    op, addr, vdst = inst.op, (V[inst.addr] + inst.offset0) & 0xffff, inst.vdst
-    if op in DS_LOAD:
-      cnt, sz, sign = DS_LOAD[op]
-      for i in range(cnt): val = int.from_bytes(lds[addr+i*sz:addr+i*sz+sz], 'little'); V[vdst + i] = _sext(val, sz * 8) & 0xffffffff if sign else val
-    elif op in DS_STORE:
-      cnt, sz = DS_STORE[op]
-      for i in range(cnt): lds[addr+i*sz:addr+i*sz+sz] = (V[inst.data0 + i] & ((1 << (sz * 8)) - 1)).to_bytes(sz, 'little')
-    else: raise NotImplementedError(f"DS op {op}")
-    return
-
-  # VOPD: dual-issue, execute two ops using VOP2/VOP3 compiled functions
-  # Both ops execute simultaneously using pre-instruction values, so read all inputs first
-  if inst_type is VOPD:
-    vdsty = (inst.vdsty << 1) | ((inst.vdstx & 1) ^ 1)
-    # Read all source operands BEFORE any writes (dual-issue semantics)
-    sx0, sx1 = st.rsrc(inst.srcx0, lane), V[inst.vsrcx1]
-    sy0, sy1 = st.rsrc(inst.srcy0, lane), V[inst.vsrcy1]
-    dx0, dy0 = V[inst.vdstx], V[vdsty]
-    # Execute X op
-    res_x = None
-    if (op_x := _VOPD_TO_VOP.get(inst.opx)):
-      if (fn_x := compiled.get(type(op_x), {}).get(op_x)):
-        res_x = fn_x(sx0, sx1, 0, dx0, st.scc, st.vcc, lane, st.exec_mask, st.literal, None, {})
-    # Execute Y op
-    res_y = None
-    if (op_y := _VOPD_TO_VOP.get(inst.opy)):
-      if (fn_y := compiled.get(type(op_y), {}).get(op_y)):
-        res_y = fn_y(sy0, sy1, 0, dy0, st.scc, st.vcc, lane, st.exec_mask, st.literal, None, {})
-    # Write results after both ops complete
-    if res_x is not None: V[inst.vdstx] = res_x['d0']
-    if res_y is not None: V[vdsty] = res_y['d0']
-    return
-
-  # VOP3SD: has extra scalar dest for carry output
-  if inst_type is VOP3SD:
-    op = VOP3SDOp(inst.op)
-    fn = compiled.get(VOP3SDOp, {}).get(op)
-    if fn is None: raise NotImplementedError(f"{op.name} not in pseudocode")
-    s0, s1, s2 = st.rsrc(inst.src0, lane), st.rsrc(inst.src1, lane), st.rsrc(inst.src2, lane)
-    # For 64-bit src2 ops (V_MAD_U64_U32, V_MAD_I64_I32), read from consecutive registers
-    mad64_ops = (VOP3SDOp.V_MAD_U64_U32, VOP3SDOp.V_MAD_I64_I32)
-    if op in mad64_ops:
-      if inst.src2 >= 256:  # VGPR
-        s2 = V[inst.src2 - 256] | (V[inst.src2 - 256 + 1] << 32)
-      else:  # SGPR - read 64-bit from consecutive SGPRs
-        s2 = st.rsgpr64(inst.src2)
-    d0 = V[inst.vdst]
-    # For carry-in operations (V_*_CO_CI_*), src2 register contains the carry bitmask (not VCC).
-    # The pseudocode uses VCC but in VOP3SD encoding, the actual carry source is inst.src2.
-    # We pass the src2 register value as 'vcc' to the interpreter so it reads the correct carry.
-    carry_ops = (VOP3SDOp.V_ADD_CO_CI_U32, VOP3SDOp.V_SUB_CO_CI_U32, VOP3SDOp.V_SUBREV_CO_CI_U32)
-    vcc_for_exec = st.rsgpr64(inst.src2) if op in carry_ops else st.vcc
-    result = fn(s0, s1, s2, d0, st.scc, vcc_for_exec, lane, st.exec_mask, st.literal, None, {})
-    # Write result - handle 64-bit destinations
-    if result.get('d0_64'):
-      V[inst.vdst] = result['d0'] & 0xffffffff
-      V[inst.vdst + 1] = (result['d0'] >> 32) & 0xffffffff
-    else:
-      V[inst.vdst] = result['d0'] & 0xffffffff
-    if result.get('vcc_lane') is not None:
-      st.pend_sgpr_lane(inst.sdst, lane, result['vcc_lane'])
-    return
-
-
-
-  # Get op enum and sources (None means "no source" for that operand)
-  if inst_type is VOP1:
-    if inst.op == VOP1Op.V_NOP: return
-    op_cls, op, src0, src1, src2, vdst = VOP1Op, VOP1Op(inst.op), inst.src0, None, None, inst.vdst
-  elif inst_type is VOP2:
-    op_cls, op, src0, src1, src2, vdst = VOP2Op, VOP2Op(inst.op), inst.src0, inst.vsrc1 + 256, None, inst.vdst
-  elif inst_type is VOP3:
-    # VOP3 ops 0-255 are VOPC comparisons encoded as VOP3 (use VOPCOp pseudocode)
-    if inst.op < 256:
-      op_cls, op, src0, src1, src2, vdst = VOPCOp, VOPCOp(inst.op), inst.src0, inst.src1, None, inst.vdst
-    else:
-      op_cls, op, src0, src1, src2, vdst = VOP3Op, VOP3Op(inst.op), inst.src0, inst.src1, inst.src2, inst.vdst
-      # V_PERM_B32: byte permutation - not in pseudocode PDF, implement directly
-      # D0[byte_i] = selector[byte_i] < 8 ? {src1, src0}[selector[byte_i]] : (selector[byte_i] >= 0xD ? 0xFF : 0x00)
-      if op == VOP3Op.V_PERM_B32:
-        s0, s1, s2 = st.rsrc(inst.src0, lane), st.rsrc(inst.src1, lane), st.rsrc(inst.src2, lane)
-        # Combine src0 and src1 into 8-byte value: src0 is bytes 0-3, src1 is bytes 4-7
-        combined = (s0 & 0xffffffff) | ((s1 & 0xffffffff) << 32)
-        result = 0
-        for i in range(4):  # 4 result bytes
-          sel = (s2 >> (i * 8)) & 0xff  # byte selector for this position
-          if sel <= 7: result |= (((combined >> (sel * 8)) & 0xff) << (i * 8))  # select byte from combined
-          elif sel >= 0xd: result |= (0xff << (i * 8))  # 0xD-0xF: constant 0xFF
-          # else 0x8-0xC: constant 0x00 (already 0)
-        V[vdst] = result & 0xffffffff
-        return
-  elif inst_type is VOPC:
-    op_cls, op, src0, src1, src2, vdst = VOPCOp, VOPCOp(inst.op), inst.src0, inst.vsrc1 + 256, None, VCC_LO
-  elif inst_type is VOP3P:
-    # VOP3P: Packed 16-bit operations using compiled functions
-    op = VOP3POp(inst.op)
-    # WMMA: wave-level matrix multiply-accumulate (special handling - needs cross-lane access)
-    if op in (VOP3POp.V_WMMA_F32_16X16X16_F16, VOP3POp.V_WMMA_F32_16X16X16_BF16, VOP3POp.V_WMMA_F16_16X16X16_F16):
-      if lane == 0:  # Only execute once per wave, write results for all lanes
-        exec_wmma(st, inst, op)
-      return
-    # V_FMA_MIX: Mixed precision FMA - inputs can be f16 or f32 controlled by opsel
-    if op in (VOP3POp.V_FMA_MIX_F32, VOP3POp.V_FMA_MIXLO_F16, VOP3POp.V_FMA_MIXHI_F16):
-      opsel = getattr(inst, 'opsel', 0)
-      opsel_hi = getattr(inst, 'opsel_hi', 0)
-      neg = getattr(inst, 'neg', 0)
-      neg_hi = getattr(inst, 'neg_hi', 0)
-      vdst = inst.vdst
-      # Read raw 32-bit values - for V_FMA_MIX, sources can be either f32 or f16
-      s0_raw = st.rsrc(inst.src0, lane)
-      s1_raw = st.rsrc(inst.src1, lane)
-      s2_raw = st.rsrc(inst.src2, lane) if inst.src2 is not None else 0
-      # opsel[i]=0: use as f32, opsel[i]=1: use hi f16 as f32
-      # For src0: opsel[0], for src1: opsel[1], for src2: opsel[2]
-      if opsel & 1: s0 = _f16((s0_raw >> 16) & 0xffff)  # hi f16 -> f32
-      else: s0 = _f32(s0_raw)  # use as f32
-      if opsel & 2: s1 = _f16((s1_raw >> 16) & 0xffff)
-      else: s1 = _f32(s1_raw)
-      if opsel & 4: s2 = _f16((s2_raw >> 16) & 0xffff)
-      else: s2 = _f32(s2_raw)
-      # Apply neg modifiers (for f32 values)
-      if neg & 1: s0 = -s0
-      if neg & 2: s1 = -s1
-      if neg & 4: s2 = -s2
-      # Compute FMA: d = s0 * s1 + s2
-      result = s0 * s1 + s2
-      V = st.vgpr[lane]
-      if op == VOP3POp.V_FMA_MIX_F32:
-        V[vdst] = _i32(result)
-      elif op == VOP3POp.V_FMA_MIXLO_F16:
-        lo = _i16(result) & 0xffff
-        V[vdst] = (V[vdst] & 0xffff0000) | lo
-      else:  # V_FMA_MIXHI_F16
-        hi = _i16(result) & 0xffff
-        V[vdst] = (V[vdst] & 0x0000ffff) | (hi << 16)
-      return
-    # Use rsrc_f16 for VOP3P to get correct f16 inline constants
-    s0_raw = st.rsrc_f16(inst.src0, lane)
-    s1_raw = st.rsrc_f16(inst.src1, lane)
-    s2_raw = st.rsrc_f16(inst.src2, lane) if inst.src2 is not None else 0
-    # Handle opsel (which 16-bit halves to use for each source)
-    opsel = getattr(inst, 'opsel', 0)
-    opsel_hi = getattr(inst, 'opsel_hi', 3)  # Default: use hi for hi result
-    opsel_hi2 = getattr(inst, 'opsel_hi2', 1)  # Default for src2
-    # Handle neg modifiers for VOP3P
-    # neg applies to lo result inputs, neg_hi applies to hi result inputs
-    neg = getattr(inst, 'neg', 0)
-    neg_hi = getattr(inst, 'neg_hi', 0)
-    # Build "virtual" sources with halves arranged for pseudocode: lo half goes to [15:0], hi half goes to [31:16]
-    # opsel bit 0/1/2 selects which half of src0/1/2 goes to the LO result
-    # opsel_hi bit 0/1 selects which half of src0/1 goes to the HI result
-    s0_lo = (s0_raw >> 16) & 0xffff if (opsel & 1) else s0_raw & 0xffff
-    s1_lo = (s1_raw >> 16) & 0xffff if (opsel & 2) else s1_raw & 0xffff
-    s2_lo = (s2_raw >> 16) & 0xffff if (opsel & 4) else s2_raw & 0xffff
-    s0_hi = (s0_raw >> 16) & 0xffff if (opsel_hi & 1) else s0_raw & 0xffff
-    s1_hi = (s1_raw >> 16) & 0xffff if (opsel_hi & 2) else s1_raw & 0xffff
-    s2_hi = (s2_raw >> 16) & 0xffff if opsel_hi2 else s2_raw & 0xffff
-    # Apply neg to lo result inputs (toggle f16 sign bit)
-    if neg & 1: s0_lo ^= 0x8000
-    if neg & 2: s1_lo ^= 0x8000
-    if neg & 4: s2_lo ^= 0x8000
-    # Apply neg_hi to hi result inputs
-    if neg_hi & 1: s0_hi ^= 0x8000
-    if neg_hi & 2: s1_hi ^= 0x8000
-    if neg_hi & 4: s2_hi ^= 0x8000
-    # Pack into format expected by pseudocode: [31:16] = hi input, [15:0] = lo input
-    s0 = (s0_hi << 16) | s0_lo
-    s1 = (s1_hi << 16) | s1_lo
-    s2 = (s2_hi << 16) | s2_lo
-    op_cls, vdst = VOP3POp, inst.vdst
-    fn = compiled.get(op_cls, {}).get(op)
-    if fn is None: raise NotImplementedError(f"{op.name} not in pseudocode")
-    result = fn(s0, s1, s2, 0, st.scc, st.vcc, lane, st.exec_mask, st.literal, None, {})
-    st.vgpr[lane][vdst] = result['d0'] & 0xffffffff
-    return
-  else: raise NotImplementedError(f"Unknown vector type {inst_type}")
-
-  fn = compiled.get(op_cls, {}).get(op)
-  if fn is None: raise NotImplementedError(f"{op.name} not in pseudocode")
-
-  # Read sources (with VOP3 modifiers if applicable)
-  neg, abs_ = (getattr(inst, 'neg', 0), getattr(inst, 'abs', 0)) if inst_type is VOP3 else (0, 0)
-  opsel = getattr(inst, 'opsel', 0) if inst_type is VOP3 else 0
-  def mod_src(val: int, idx: int) -> int:
-    if (abs_ >> idx) & 1: val = _i32(abs(_f32(val)))
-    if (neg >> idx) & 1: val = _i32(-_f32(val))
-    return val
-  def mod_src64(val: int, idx: int) -> int:
-    if (abs_ >> idx) & 1: val = _i64(abs(_f64(val)))
-    if (neg >> idx) & 1: val = _i64(-_f64(val))
-    return val
-
-  # Determine if sources are 64-bit based on instruction type
-  # For 64-bit shift ops: src0 is 32-bit (shift amount), src1 is 64-bit (value to shift)
-  # For most other _B64/_I64/_U64/_F64 ops: all sources are 64-bit
-  is_64bit_op = op.name.endswith(('_B64', '_I64', '_U64', '_F64'))
-  # V_LDEXP_F64: src0 is 64-bit float, src1 is 32-bit integer exponent
-  is_ldexp_64 = op in (VOP3Op.V_LDEXP_F64,)
-  is_shift_64 = op in (VOP3Op.V_LSHLREV_B64, VOP3Op.V_LSHRREV_B64, VOP3Op.V_ASHRREV_I64)
-  # 16-bit source ops: use precomputed sets instead of string checks
-  has_16bit_type = op in _VOP3_16BIT_OPS or op in _VOP1_16BIT_OPS or op in _VOP2_16BIT_OPS
-  is_16bit_src = op_cls is VOP3Op and op in _VOP3_16BIT_OPS and op not in _CVT_32_64_SRC_OPS
-  # VOP2 16-bit ops use f16 inline constants for src0 (vsrc1 is always a VGPR, no inline constants)
-  is_vop2_16bit = op_cls is VOP2Op and op in _VOP2_16BIT_OPS
-
-  if is_shift_64:
-    s0 = mod_src(st.rsrc(src0, lane), 0)  # shift amount is 32-bit
-    s1 = st.rsrc64(src1, lane) if src1 is not None else 0  # value to shift is 64-bit
-    s2 = mod_src(st.rsrc(src2, lane), 2) if src2 is not None else 0
-  elif is_ldexp_64:
-    s0 = mod_src64(st.rsrc64(src0, lane), 0)  # mantissa is 64-bit float
-    s1 = mod_src(st.rsrc(src1, lane), 1) if src1 is not None else 0  # exponent is 32-bit int
-    s2 = mod_src(st.rsrc(src2, lane), 2) if src2 is not None else 0
-  elif is_64bit_op:
-    # 64-bit ops: apply neg/abs modifiers using f64 interpretation for float ops
-    s0 = mod_src64(st.rsrc64(src0, lane), 0)
-    s1 = mod_src64(st.rsrc64(src1, lane), 1) if src1 is not None else 0
-    s2 = mod_src64(st.rsrc64(src2, lane), 2) if src2 is not None else 0
-  elif is_16bit_src:
-    # For 16-bit source ops, opsel bits select which half to use
-    s0_raw = mod_src(st.rsrc(src0, lane), 0)
-    s1_raw = mod_src(st.rsrc(src1, lane), 1) if src1 is not None else 0
-    s2_raw = mod_src(st.rsrc(src2, lane), 2) if src2 is not None else 0
-    # opsel[0] selects hi(1) or lo(0) for src0, opsel[1] for src1, opsel[2] for src2
-    s0 = ((s0_raw >> 16) & 0xffff) if (opsel & 1) else (s0_raw & 0xffff)
-    s1 = ((s1_raw >> 16) & 0xffff) if (opsel & 2) else (s1_raw & 0xffff)
-    s2 = ((s2_raw >> 16) & 0xffff) if (opsel & 4) else (s2_raw & 0xffff)
-  elif is_vop2_16bit:
-    # VOP2 16-bit ops: src0 can use f16 inline constants, vsrc1 is always a VGPR (no inline constants)
-    s0 = mod_src(st.rsrc_f16(src0, lane), 0)
-    s1 = mod_src(st.rsrc(src1, lane), 1) if src1 is not None else 0
-    s2 = mod_src(st.rsrc(src2, lane), 2) if src2 is not None else 0
-  else:
-    s0 = mod_src(st.rsrc(src0, lane), 0)
-    s1 = mod_src(st.rsrc(src1, lane), 1) if src1 is not None else 0
-    s2 = mod_src(st.rsrc(src2, lane), 2) if src2 is not None else 0
-  d0 = V[vdst] if not is_64bit_op else (V[vdst] | (V[vdst + 1] << 32))
-
-  # V_CNDMASK_B32: VOP3 encoding uses src2 as mask (not VCC); VOP2 uses VCC implicitly
-  # Pass the correct mask as vcc to the function so pseudocode VCC.u64[laneId] works correctly
-  vcc_for_fn = st.rsgpr64(src2) if op in (VOP3Op.V_CNDMASK_B32,) and inst_type is VOP3 and src2 is not None and src2 < 256 else st.vcc
-
-  # Execute compiled function - pass src0_idx and vdst_idx for lane instructions
-  # For VGPR access: src0 index is the VGPR number (src0 - 256 if VGPR, else src0 for SGPR)
-  src0_idx = (src0 - 256) if src0 is not None and src0 >= 256 else (src0 if src0 is not None else 0)
-  result = fn(s0, s1, s2, d0, st.scc, vcc_for_fn, lane, st.exec_mask, st.literal, st.vgpr, {}, src0_idx, vdst)
-
-  # Apply results
-  if 'vgpr_write' in result:
-    # Lane instruction wrote to VGPR: (lane, vgpr_idx, value)
-    wr_lane, wr_idx, wr_val = result['vgpr_write']
-    st.vgpr[wr_lane][wr_idx] = wr_val
-  if 'vcc_lane' in result:
-    # VOP2 carry instructions (V_ADD_CO_CI_U32, V_SUB_CO_CI_U32, V_SUBREV_CO_CI_U32) write carry to VCC implicitly
-    # VOPC and VOP3-encoded VOPC write to vdst (which is VCC_LO for VOPC, inst.sdst for VOP3)
-    vcc_dst = VCC_LO if op_cls is VOP2Op and op in (VOP2Op.V_ADD_CO_CI_U32, VOP2Op.V_SUB_CO_CI_U32, VOP2Op.V_SUBREV_CO_CI_U32) else vdst
-    st.pend_sgpr_lane(vcc_dst, lane, result['vcc_lane'])
-  if 'exec_lane' in result:
-    # V_CMPX instructions write to EXEC per-lane
-    st.pend_sgpr_lane(EXEC_LO, lane, result['exec_lane'])
-  if 'd0' in result and op_cls not in (VOPCOp,) and 'vgpr_write' not in result:
-    # V_READFIRSTLANE_B32 and V_READLANE_B32 write to SGPR, not VGPR
-    # V_WRITELANE_B32 uses vgpr_write for cross-lane writes, don't overwrite with d0
-    writes_to_sgpr = op in (VOP1Op.V_READFIRSTLANE_B32,) or \
-                     (op_cls is VOP3Op and op in (VOP3Op.V_READFIRSTLANE_B32, VOP3Op.V_READLANE_B32))
-    # Check for 16-bit destination ops (opsel[3] controls hi/lo write)
-    is_16bit_dst = op in _VOP3_16BIT_DST_OPS or op in _VOP1_16BIT_DST_OPS
-    if writes_to_sgpr:
-      st.wsgpr(vdst, result['d0'] & 0xffffffff)
-    elif result.get('d0_64') or is_64bit_op:
-      V[vdst] = result['d0'] & 0xffffffff
-      V[vdst + 1] = (result['d0'] >> 32) & 0xffffffff
-    elif is_16bit_dst and inst_type is VOP3:
-      # VOP3 16-bit ops: opsel[3] (bit 3 of opsel field) controls hi/lo destination
-      if opsel & 8:  # opsel[3] = 1: write to high 16 bits
-        V[vdst] = (V[vdst] & 0x0000ffff) | ((result['d0'] & 0xffff) << 16)
-      else:  # opsel[3] = 0: write to low 16 bits
-        V[vdst] = (V[vdst] & 0xffff0000) | (result['d0'] & 0xffff)
-    else:
-      V[vdst] = result['d0'] & 0xffffffff
-
-# ═══════════════════════════════════════════════════════════════════════════════
-# WMMA (Wave Matrix Multiply-Accumulate)
-# ═══════════════════════════════════════════════════════════════════════════════
-
-def exec_wmma(st: WaveState, inst, op: VOP3POp) -> None:
-  """Execute WMMA instruction - 16x16x16 matrix multiply across the wave."""
-  src0, src1, src2, vdst = inst.src0, inst.src1, inst.src2, inst.vdst
-  # Read matrix A (16x16 f16/bf16) from lanes 0-15, VGPRs src0 to src0+7 (2 f16 per VGPR = 16 values per lane)
-  # Layout: A[row][k] where row = lane (0-15), k comes from 8 VGPRs × 2 halves
-  mat_a = []
-  for lane in range(16):
-    for reg in range(8):
-      val = st.vgpr[lane][src0 - 256 + reg] if src0 >= 256 else st.rsgpr(src0 + reg)
-      mat_a.append(_f16(val & 0xffff))
-      mat_a.append(_f16((val >> 16) & 0xffff))
-  # Read matrix B (16x16 f16/bf16) - same layout, B[col][k] where col comes from lane
-  mat_b = []
-  for lane in range(16):
-    for reg in range(8):
-      val = st.vgpr[lane][src1 - 256 + reg] if src1 >= 256 else st.rsgpr(src1 + reg)
-      mat_b.append(_f16(val & 0xffff))
-      mat_b.append(_f16((val >> 16) & 0xffff))
-
-  # Read matrix C (16x16 f32) from lanes 0-31, VGPRs src2 to src2+7
-  # Layout: element i is at lane (i % 32), VGPR (i // 32) + src2
-  mat_c = []
-  for i in range(256):
-    lane, reg = i % 32, i // 32
-    val = st.vgpr[lane][src2 - 256 + reg] if src2 >= 256 else st.rsgpr(src2 + reg)
-    mat_c.append(_f32(val))
-
-  # Compute D = A × B + C (16x16 matrix multiply)
-  mat_d = [0.0] * 256
-  for row in range(16):
-    for col in range(16):
-      acc = 0.0
-      for k in range(16):
-        a_val = mat_a[row * 16 + k]
-        b_val = mat_b[col * 16 + k]
-        acc += a_val * b_val
-      mat_d[row * 16 + col] = acc + mat_c[row * 16 + col]
-
-  # Write result matrix D back - same layout as C
-  if op == VOP3POp.V_WMMA_F16_16X16X16_F16:
-    # Output is f16, pack 2 values per VGPR
-    for i in range(0, 256, 2):
-      lane, reg = (i // 2) % 32, (i // 2) // 32
-      lo = _i16(mat_d[i]) & 0xffff
-      hi = _i16(mat_d[i + 1]) & 0xffff
-      st.vgpr[lane][vdst + reg] = (hi << 16) | lo
-  else:
-    # Output is f32
-    for i in range(256):
-      lane, reg = i % 32, i // 32
-      st.vgpr[lane][vdst + reg] = _i32(mat_d[i])
-
-# ═══════════════════════════════════════════════════════════════════════════════
-# MAIN EXECUTION LOOP
-# ═══════════════════════════════════════════════════════════════════════════════
-
-SCALAR_TYPES = {SOP1, SOP2, SOPC, SOPK, SOPP, SMEM}
-VECTOR_TYPES = {VOP1, VOP2, VOP3, VOP3SD, VOPC, FLAT, DS, VOPD, VOP3P}
-
-def step_wave(program: Program, st: WaveState, lds: bytearray, n_lanes: int) -> int:
-  inst = program.get(st.pc)
-  if inst is None: return 1
-  inst_words, st.literal, inst_type = inst._words, getattr(inst, '_literal', None) or 0, type(inst)
-
-  if inst_type in SCALAR_TYPES:
-    delta = exec_scalar(st, inst)
-    if delta == -1: return -1  # endpgm
-    if delta == -2: st.pc += inst_words; return -2  # barrier
-    st.pc += inst_words + delta
-  else:
-    # V_READFIRSTLANE_B32 and V_READLANE_B32 write to SGPR, so they should only execute once per wave (lane 0)
-    is_readlane = (inst_type is VOP1 and inst.op == VOP1Op.V_READFIRSTLANE_B32) or \
-                  (inst_type is VOP3 and inst.op in (VOP3Op.V_READFIRSTLANE_B32, VOP3Op.V_READLANE_B32))
-    if is_readlane:
-      exec_vector(st, inst, 0, lds)  # Execute once with lane 0
-    else:
-      exec_mask = st.exec_mask
-      for lane in range(n_lanes):
-        if exec_mask & (1 << lane): exec_vector(st, inst, lane, lds)
-    st.commit_pends()
-    st.pc += inst_words
-  return 0
-
-def exec_wave(program: Program, st: WaveState, lds: bytearray, n_lanes: int) -> int:
-  while st.pc in program:
-    result = step_wave(program, st, lds, n_lanes)
-    if result == -1: return 0
-    if result == -2: return -2
-  return 0
-
-def exec_workgroup(program: Program, workgroup_id: tuple[int, int, int], local_size: tuple[int, int, int], args_ptr: int,
-                   wg_id_sgpr_base: int, wg_id_enables: tuple[bool, bool, bool]) -> None:
-  lx, ly, lz = local_size
-  total_threads, lds = lx * ly * lz, bytearray(65536)
-  waves: list[tuple[WaveState, int, int]] = []
-  for wave_start in range(0, total_threads, WAVE_SIZE):
-    n_lanes, st = min(WAVE_SIZE, total_threads - wave_start), WaveState()
-    st.exec_mask = (1 << n_lanes) - 1
-    st.wsgpr64(0, args_ptr)
-    gx, gy, gz = workgroup_id
-    # Set workgroup IDs in SGPRs based on USER_SGPR_COUNT and enable flags from COMPUTE_PGM_RSRC2
-    sgpr_idx = wg_id_sgpr_base
-    if wg_id_enables[0]: st.sgpr[sgpr_idx] = gx; sgpr_idx += 1
-    if wg_id_enables[1]: st.sgpr[sgpr_idx] = gy; sgpr_idx += 1
-    if wg_id_enables[2]: st.sgpr[sgpr_idx] = gz
-    for i in range(n_lanes):
-      tid = wave_start + i
-      st.vgpr[i][0] = tid if local_size == (lx, 1, 1) else ((tid // (lx * ly)) << 20) | (((tid // lx) % ly) << 10) | (tid % lx)
-    waves.append((st, n_lanes, wave_start))
-  has_barrier = any(isinstance(inst, SOPP) and inst.op == SOPPOp.S_BARRIER for inst in program.values())
-  for _ in range(2 if has_barrier else 1):
-    for st, n_lanes, _ in waves: exec_wave(program, st, lds, n_lanes)
-
-def run_asm(lib: int, lib_sz: int, gx: int, gy: int, gz: int, lx: int, ly: int, lz: int, args_ptr: int, rsrc2: int = 0x19c) -> int:
-  data = (ctypes.c_char * lib_sz).from_address(lib).raw
-  program = decode_program(data)
-  if not program: return -1
-  # Parse COMPUTE_PGM_RSRC2 for SGPR layout
-  user_sgpr_count = (rsrc2 >> 1) & 0x1f
-  enable_wg_id_x = bool((rsrc2 >> 7) & 1)
-  enable_wg_id_y = bool((rsrc2 >> 8) & 1)
-  enable_wg_id_z = bool((rsrc2 >> 9) & 1)
-  wg_id_enables = (enable_wg_id_x, enable_wg_id_y, enable_wg_id_z)
-  for gidz in range(gz):
-    for gidy in range(gy):
-      for gidx in range(gx): exec_workgroup(program, (gidx, gidy, gidz), (lx, ly, lz), args_ptr, user_sgpr_count, wg_id_enables)
-  return 0

extra/assembly/amd/test/bench_emu.py

@@ -1,294 +0,0 @@
-#!/usr/bin/env python3
-"""Benchmark comparing Python vs Rust RDNA3 emulators on synthetic and real tinygrad kernels."""
-import ctypes, time, os, struct, cProfile, pstats, io
-from pathlib import Path
-from typing import Callable
-
-# Set AMD=1 before importing tinygrad
-os.environ["AMD"] = "1"
-
-from extra.assembly.amd.emu import run_asm as python_run_asm, set_valid_mem_ranges, decode_program, step_wave, WaveState, WAVE_SIZE
-
-REMU_PATH = Path(__file__).parents[3] / "remu/target/release/libremu.so"
-if not REMU_PATH.exists():
-  REMU_PATH = Path(__file__).parents[3] / "remu/target/release/libremu.dylib"
-
-def get_rust_remu():
-  """Load the Rust libremu shared library."""
-  if not REMU_PATH.exists(): return None
-  remu = ctypes.CDLL(str(REMU_PATH))
-  remu.run_asm.restype = ctypes.c_int32
-  remu.run_asm.argtypes = [ctypes.c_void_p, ctypes.c_uint32, ctypes.c_uint32, ctypes.c_uint32, ctypes.c_uint32,
-                           ctypes.c_uint32, ctypes.c_uint32, ctypes.c_uint32, ctypes.c_void_p]
-  return remu
-
-def count_instructions(kernel: bytes) -> int:
-  """Count instructions in a kernel."""
-  return len(decode_program(kernel))
-
-def setup_buffers(buf_sizes: list[int], init_data: dict[int, bytes] | None = None):
-  """Allocate buffers and return args pointer + valid ranges."""
-  if init_data is None: init_data = {}
-  buffers = []
-  for i, size in enumerate(buf_sizes):
-    padded = ((size + 15) // 16) * 16 + 16
-    data = init_data.get(i, b'\x00' * padded)
-    data_list = list(data) + [0] * (padded - len(data))
-    buf = (ctypes.c_uint8 * padded)(*data_list[:padded])
-    buffers.append(buf)
-  args = (ctypes.c_uint64 * len(buffers))(*[ctypes.addressof(b) for b in buffers])
-  args_ptr = ctypes.addressof(args)
-  ranges = {(ctypes.addressof(b), len(b)) for b in buffers}
-  ranges.add((args_ptr, ctypes.sizeof(args)))
-  return buffers, args, args_ptr, ranges
-
-def benchmark_emulator(name: str, run_fn, kernel: bytes, global_size, local_size, args_ptr, iterations: int = 5):
-  """Benchmark an emulator and return average time."""
-  gx, gy, gz = global_size
-  lx, ly, lz = local_size
-  kernel_buf = (ctypes.c_char * len(kernel)).from_buffer_copy(kernel)
-  lib_ptr = ctypes.addressof(kernel_buf)
-
-  # Warmup
-  run_fn(lib_ptr, len(kernel), gx, gy, gz, lx, ly, lz, args_ptr)
-
-  # Timed runs
-  times = []
-  for _ in range(iterations):
-    start = time.perf_counter()
-    result = run_fn(lib_ptr, len(kernel), gx, gy, gz, lx, ly, lz, args_ptr)
-    end = time.perf_counter()
-    if result != 0:
-      print(f"  {name} returned error: {result}")
-      return None
-    times.append(end - start)
-
-  return sum(times) / len(times)
-
-def create_synthetic_kernel(n_ops: int) -> bytes:
-  """Create a synthetic kernel with n_ops vector operations."""
-  instructions = []
-  # VOP2 instructions: v_add_f32, v_mul_f32, v_max_f32, v_min_f32
-  ops = [
-    (0b0000011 << 25) | (1 << 17) | (0 << 9) | 256,  # v_add_f32 v0, v0, v1
-    (0b0001000 << 25) | (1 << 17) | (0 << 9) | 256,  # v_mul_f32 v0, v0, v1
-    (0b0010000 << 25) | (1 << 17) | (0 << 9) | 256,  # v_max_f32 v0, v0, v1
-    (0b0001111 << 25) | (1 << 17) | (0 << 9) | 256,  # v_min_f32 v0, v0, v1
-  ]
-  for i in range(n_ops):
-    instructions.append(ops[i % len(ops)])
-  # S_ENDPGM
-  instructions.append((0b101111111 << 23) | (48 << 16) | 0)
-  return b''.join(struct.pack('<I', inst) for inst in instructions)
-
-def get_tinygrad_kernel(op_name: str) -> tuple[bytes, tuple, tuple, list[int], dict[int, bytes]] | None:
-  """Get a real tinygrad kernel by operation name. Returns (code, global_size, local_size, buf_sizes, buf_data)."""
-  try:
-    from tinygrad import Tensor
-    from tinygrad.runtime.support.elf import elf_loader
-    import numpy as np
-    np.random.seed(42)
-
-    ops = {
-      "add": lambda: Tensor.empty(1024) + Tensor.empty(1024),
-      "mul": lambda: Tensor.empty(1024) * Tensor.empty(1024),
-      "matmul_small": lambda: Tensor.empty(16, 16) @ Tensor.empty(16, 16),
-      "matmul_medium": lambda: Tensor.empty(64, 64) @ Tensor.empty(64, 64),
-      "reduce_sum": lambda: Tensor.empty(4096).sum(),
-      "reduce_max": lambda: Tensor.empty(4096).max(),
-      "softmax": lambda: Tensor.empty(256).softmax(),
-      "layernorm": lambda: Tensor.empty(32, 64).layernorm(),
-      "conv2d": lambda: Tensor.empty(1, 4, 16, 16).conv2d(Tensor.empty(4, 4, 3, 3)),
-      "gelu": lambda: Tensor.empty(1024).gelu(),
-      "exp": lambda: Tensor.empty(1024).exp(),
-      "sin": lambda: Tensor.empty(1024).sin(),
-    }
-
-    if op_name not in ops: return None
-    out = ops[op_name]()
-    sched = out.schedule()
-
-    for ei in sched:
-      lowered = ei.lower()
-      if ei.ast.op.name == 'SINK' and lowered.prg and lowered.prg.p.lib:
-        lib = bytes(lowered.prg.p.lib)
-        _, sections, _ = elf_loader(lib)
-        for sec in sections:
-          if sec.name == '.text':
-            buf_sizes = [b.nbytes for b in lowered.bufs]
-            # Get initial data from numpy arrays if available
-            buf_data = {}
-            for i, buf in enumerate(lowered.bufs):
-              if hasattr(buf, 'base') and buf.base is not None and hasattr(buf.base, '_buf'):
-                try: buf_data[i] = bytes(buf.base._buf)
-                except: pass
-            return (bytes(sec.content), tuple(lowered.prg.p.global_size), tuple(lowered.prg.p.local_size), buf_sizes, buf_data)
-    return None
-  except Exception as e:
-    print(f"  Error getting kernel: {e}")
-    return None
-
-def profile_python_emu(kernel: bytes, global_size, local_size, args_ptr, n_runs: int = 1):
-  """Profile the Python emulator to find bottlenecks."""
-  gx, gy, gz = global_size
-  lx, ly, lz = local_size
-  kernel_buf = (ctypes.c_char * len(kernel)).from_buffer_copy(kernel)
-  lib_ptr = ctypes.addressof(kernel_buf)
-
-  pr = cProfile.Profile()
-  pr.enable()
-  for _ in range(n_runs):
-    python_run_asm(lib_ptr, len(kernel), gx, gy, gz, lx, ly, lz, args_ptr)
-  pr.disable()
-
-  s = io.StringIO()
-  ps = pstats.Stats(pr, stream=s).sort_stats('cumulative')
-  ps.print_stats(20)
-  return s.getvalue()
-
-def measure_step_rate(kernel: bytes, n_steps: int = 10000) -> float:
-  """Measure raw step_wave() performance (steps per second)."""
-  program = decode_program(kernel)
-  if not program: return 0.0
-
-  st = WaveState()
-  st.exec_mask = 0xffffffff
-  lds = bytearray(65536)
-  n_lanes = 32
-
-  # Reset PC for each measurement
-  start = time.perf_counter()
-  for _ in range(n_steps):
-    st.pc = 0
-    while st.pc in program:
-      result = step_wave(program, st, lds, n_lanes)
-      if result == -1: break
-  elapsed = time.perf_counter() - start
-  return n_steps / elapsed if elapsed > 0 else 0
-
-# Test configurations
-SYNTHETIC_TESTS = [
-  ("synthetic_10ops", 10, (1, 1, 1), (32, 1, 1)),
-  ("synthetic_100ops", 100, (1, 1, 1), (32, 1, 1)),
-  ("synthetic_500ops", 500, (1, 1, 1), (32, 1, 1)),
-  ("synthetic_100ops_4wg", 100, (4, 1, 1), (32, 1, 1)),
-  ("synthetic_100ops_16wg", 100, (16, 1, 1), (32, 1, 1)),
-]
-
-TINYGRAD_TESTS = ["add", "mul", "reduce_sum", "softmax", "exp", "gelu", "matmul_small"]
-
-def main():
-  import argparse
-  parser = argparse.ArgumentParser(description="Benchmark RDNA3 emulators")
-  parser.add_argument("--profile", action="store_true", help="Profile Python emulator")
-  parser.add_argument("--synthetic-only", action="store_true", help="Only run synthetic tests")
-  parser.add_argument("--tinygrad-only", action="store_true", help="Only run tinygrad tests")
-  parser.add_argument("--iterations", type=int, default=3, help="Number of iterations per benchmark")
-  args = parser.parse_args()
-
-  rust_remu = get_rust_remu()
-  if rust_remu is None:
-    print("Rust libremu not found. Build with: cargo build --release --manifest-path extra/remu/Cargo.toml")
-    print("Running Python-only benchmarks...\n")
-
-  print("=" * 90)
-  print("RDNA3 Emulator Benchmark: Python vs Rust")
-  print("=" * 90)
-
-  results = []
-
-  # Synthetic workloads
-  if not args.tinygrad_only:
-    print("\n[SYNTHETIC WORKLOADS]")
-    print("-" * 90)
-
-    for name, n_ops, global_size, local_size in SYNTHETIC_TESTS:
-      kernel = create_synthetic_kernel(n_ops)
-      n_insts = count_instructions(kernel)
-      n_workgroups = global_size[0] * global_size[1] * global_size[2]
-      n_threads = local_size[0] * local_size[1] * local_size[2]
-      total_work = n_insts * n_workgroups * n_threads
-
-      print(f"\n{name}: {n_insts} insts × {n_workgroups} WGs × {n_threads} threads = {total_work:,} ops")
-
-      buf_sizes = [4096]
-      buffers, args_arr, args_ptr, ranges = setup_buffers(buf_sizes)
-      set_valid_mem_ranges(ranges)
-
-      # Benchmark
-      py_time = benchmark_emulator("Python", python_run_asm, kernel, global_size, local_size, args_ptr, args.iterations)
-      rust_time = benchmark_emulator("Rust", rust_remu.run_asm, kernel, global_size, local_size, args_ptr, args.iterations) if rust_remu else None
-
-      if py_time:
-        py_rate = total_work / py_time / 1e6
-        print(f"  Python: {py_time*1000:8.3f} ms  ({py_rate:7.2f} M ops/s)")
-      if rust_time:
-        rust_rate = total_work / rust_time / 1e6
-        speedup = py_time / rust_time if py_time else 0
-        print(f"  Rust:   {rust_time*1000:8.3f} ms  ({rust_rate:7.2f} M ops/s)  [{speedup:.1f}x faster]")
-
-      results.append(("synthetic", name, n_insts, n_workgroups, py_time, rust_time))
-
-  # Tinygrad kernels
-  if not args.synthetic_only:
-    print("\n[TINYGRAD KERNELS]")
-    print("-" * 90)
-
-    for op_name in TINYGRAD_TESTS:
-      print(f"\n{op_name}:", end=" ", flush=True)
-      kernel_info = get_tinygrad_kernel(op_name)
-      if kernel_info is None:
-        print("failed to compile")
-        continue
-
-      kernel, global_size, local_size, buf_sizes, buf_data = kernel_info
-      n_insts = count_instructions(kernel)
-      n_workgroups = global_size[0] * global_size[1] * global_size[2]
-      n_threads = local_size[0] * local_size[1] * local_size[2]
-      total_work = n_insts * n_workgroups * n_threads
-
-      print(f"{n_insts} insts × {n_workgroups} WGs × {n_threads} threads = {total_work:,} ops")
-
-      buffers, args_arr, args_ptr, ranges = setup_buffers(buf_sizes, buf_data)
-      set_valid_mem_ranges(ranges)
-
-      py_time = benchmark_emulator("Python", python_run_asm, kernel, global_size, local_size, args_ptr, args.iterations)
-      rust_time = benchmark_emulator("Rust", rust_remu.run_asm, kernel, global_size, local_size, args_ptr, args.iterations) if rust_remu else None
-
-      if py_time:
-        py_rate = total_work / py_time / 1e6
-        print(f"  Python: {py_time*1000:8.3f} ms  ({py_rate:7.2f} M ops/s)")
-      if rust_time:
-        rust_rate = total_work / rust_time / 1e6
-        speedup = py_time / rust_time if py_time else 0
-        print(f"  Rust:   {rust_time*1000:8.3f} ms  ({rust_rate:7.2f} M ops/s)  [{speedup:.1f}x faster]")
-
-      results.append(("tinygrad", op_name, n_insts, n_workgroups, py_time, rust_time))
-
-      # Optional profiling
-      if args.profile and py_time:
-        print("\n  [PROFILE - Top 10 functions]")
-        profile_output = profile_python_emu(kernel, global_size, local_size, args_ptr)
-        for line in profile_output.split('\n')[5:15]:
-          if line.strip(): print(f"    {line}")
-
-  # Summary table
-  print("\n" + "=" * 90)
-  print("SUMMARY")
-  print("=" * 90)
-  print(f"{'Type':<10} {'Name':<25} {'Insts':<8} {'WGs':<6} {'Python (ms)':<14} {'Rust (ms)':<14} {'Speedup':<10}")
-  print("-" * 90)
-
-  for test_type, name, n_insts, n_wgs, py_time, rust_time in results:
-    py_ms = f"{py_time*1000:.3f}" if py_time else "error"
-    if rust_time:
-      rust_ms = f"{rust_time*1000:.3f}"
-      speedup = f"{py_time/rust_time:.1f}x" if py_time else "N/A"
-    else:
-      rust_ms, speedup = "N/A", "N/A"
-    print(f"{test_type:<10} {name:<25} {n_insts:<8} {n_wgs:<6} {py_ms:<14} {rust_ms:<14} {speedup:<10}")
-
-
-
-if __name__ == "__main__":
-  main()

extra/assembly/amd/test/external_test_usability.py

@@ -1,196 +0,0 @@
-# Usability tests for the RDNA3 ASM DSL
-# These tests demonstrate how the DSL *should* work for a good user experience
-# Currently many of these tests fail - they document desired behavior
-
-import unittest
-from extra.assembly.amd.autogen.rdna3 import *
-from extra.assembly.amd.dsl import Inst, RawImm, SGPR, VGPR
-
-class TestRegisterSliceSyntax(unittest.TestCase):
-  """
-  Issue: Register slice syntax should use AMD assembly convention (inclusive end).
-
-  In AMD assembly, s[4:7] means registers s4, s5, s6, s7 (4 registers, inclusive).
-  The DSL should match this convention so that:
-  - s[4:7] gives 4 registers
-  - Disassembler output can be copied directly back into DSL code
-
-  Fix: Change _RegFactory.__getitem__ to use inclusive end:
-    key.stop - key.start + 1  (instead of key.stop - key.start)
-  """
-  def test_register_slice_count(self):
-    # s[4:7] should give 4 registers: s4, s5, s6, s7 (AMD convention, inclusive)
-    reg = s[4:7]
-    self.assertEqual(reg.count, 4, "s[4:7] should give 4 registers (s4, s5, s6, s7)")
-
-  def test_register_slice_roundtrip(self):
-    # Round-trip: DSL -> disasm -> DSL should preserve register count
-    reg = s[4:7]  # 4 registers in AMD convention
-    inst = s_load_b128(reg, s[0:1], NULL, 0)
-    disasm = inst.disasm()
-    # Disasm shows s[4:7] - user should be able to copy this back
-    self.assertIn("s[4:7]", disasm)
-    # And s[4:7] in DSL should give the same 4 registers
-    reg_from_disasm = s[4:7]
-    self.assertEqual(reg_from_disasm.count, 4, "s[4:7] from disasm should give 4 registers")
-
-
-class TestReprReadability(unittest.TestCase):
-  """
-  Issue: repr() leaks internal RawImm type and omits zero-valued fields.
-
-  When you create v_mov_b32_e32(v[0], v[1]), the repr shows:
-    VOP1(op=1, src0=RawImm(257))
-
-  Problems:
-  1. vdst=v[0] is omitted because 0 is treated as "default"
-  2. src0 shows RawImm(257) instead of v[1]
-  3. User sees encoded values (257 = 256 + 1) instead of register names
-
-  Expected repr: VOP1(op=1, vdst=v[0], src0=v[1])
-  """
-  def test_repr_shows_registers_not_raw_imm(self):
-    inst = v_mov_b32_e32(v[0], v[1])
-    # Should show v[1], not RawImm(257)
-    self.assertNotIn("RawImm", repr(inst), "repr should not expose RawImm internal type")
-    self.assertIn("v[1]", repr(inst), "repr should show register name")
-
-  def test_repr_includes_zero_dst(self):
-    inst = v_mov_b32_e32(v[0], v[1])
-    # v[0] is a valid destination register, should be shown
-    self.assertIn("vdst", repr(inst), "repr should include vdst even when 0")
-
-  def test_repr_roundtrip(self):
-    # repr should produce something that can be eval'd back
-    inst = v_mov_b32_e32(v[0], v[1])
-    # This would require repr to output valid Python, e.g.:
-    # "VOP1(op=VOP1Op.V_MOV_B32, vdst=v[0], src0=v[1])"
-    r = repr(inst)
-    # At minimum, it should be human-readable
-    self.assertIn("v[", r, "repr should show register syntax")
-
-
-class TestInstructionEquality(unittest.TestCase):
-  """
-  Issue: No __eq__ method - instruction comparison requires repr() workaround.
-
-  Two identical instructions should compare equal with ==, but currently:
-    inst1 == inst2 returns False
-
-  The test_handwritten.py works around this with:
-    self.assertEqual(repr(self.inst), repr(reasm))
-  """
-  def test_identical_instructions_equal(self):
-    inst1 = v_mov_b32_e32(v[0], v[1])
-    inst2 = v_mov_b32_e32(v[0], v[1])
-    self.assertEqual(inst1, inst2, "identical instructions should be equal")
-
-  def test_different_instructions_not_equal(self):
-    inst1 = v_mov_b32_e32(v[0], v[1])
-    inst2 = v_mov_b32_e32(v[0], v[2])
-    self.assertNotEqual(inst1, inst2, "different instructions should not be equal")
-
-
-class TestVOPDHelperSignature(unittest.TestCase):
-  """
-  Issue: VOPD helper functions have confusing semantics.
-
-  v_dual_mul_f32 is defined as:
-    v_dual_mul_f32 = functools.partial(VOPD, VOPDOp.V_DUAL_MUL_F32)
-
-  This binds VOPDOp.V_DUAL_MUL_F32 to the FIRST positional arg of VOPD.__init__,
-  which is 'opx'. So v_dual_mul_f32 sets the X operation.
-
-  But then test_dual_mul in test_handwritten.py does:
-    v_dual_mul_f32(VOPDOp.V_DUAL_MUL_F32, vdstx=v[0], ...)
-
-  This passes V_DUAL_MUL_F32 as the SECOND positional arg (opy), making both
-  X and Y operations the same. This is confusing because:
-  1. The function name suggests it handles the X operation
-  2. But you still pass an opcode as the first arg (which becomes opy)
-
-  Expected: Either make the helper fully specify both ops, or make the
-  signature clearer about what the positional arg means.
-  """
-  def test_vopd_helper_opy_should_be_required(self):
-    # Using only keyword args "works" but opy silently defaults to 0
-    inst = v_dual_mul_f32(vdstx=v[0], vdsty=v[1], srcx0=v[2], vsrcx1=v[3], srcy0=v[4], vsrcy1=v[5])
-    self.assertEqual(inst.opx, VOPDOp.V_DUAL_MUL_F32)
-    # Bug: opy defaults to 0 (V_DUAL_FMAC_F32) silently - should require explicit opy
-    # This test documents the bug - it should fail once fixed
-    self.assertNotEqual(inst.opy, VOPDOp.V_DUAL_FMAC_F32, "opy should not silently default to FMAC")
-
-  def test_vopd_helper_positional_arg_is_opy(self):
-    # The first positional arg after the partial becomes opy, not a second opx
-    inst = v_dual_mul_f32(VOPDOp.V_DUAL_MOV_B32, vdstx=v[0], vdsty=v[1], srcx0=v[2], vsrcx1=v[3], srcy0=v[4], vsrcy1=v[5])
-    self.assertEqual(inst.opx, VOPDOp.V_DUAL_MUL_F32)  # From partial
-    self.assertEqual(inst.opy, VOPDOp.V_DUAL_MOV_B32)  # From first positional arg
-
-
-class TestFieldAccessPreservesType(unittest.TestCase):
-  """
-  Issue: Field access loses type information.
-
-  After creating an instruction, accessing fields returns encoded int values:
-    inst = v_mov_b32_e32(v[0], v[1])
-    inst.vdst  # returns 0, not VGPR(0)
-
-  This makes it impossible to round-trip register types through field access.
-  """
-  def test_vdst_returns_register(self):
-    inst = v_mov_b32_e32(v[5], v[1])
-    vdst = inst.vdst
-    # Should return a VGPR, not an int
-    self.assertIsInstance(vdst, (VGPR, int), "vdst should return VGPR or at least be usable")
-    # Ideally: self.assertIsInstance(vdst, VGPR)
-
-  def test_src_returns_register_for_vgpr_source(self):
-    inst = v_mov_b32_e32(v[0], v[1])
-    # src0 is encoded as 257 (256 + 1 for v1)
-    # Ideally it should decode back to v[1]
-    src0_raw = inst._values.get('src0')
-    # Currently returns RawImm(257), should return VGPR(1) or similar
-    self.assertNotIsInstance(src0_raw, RawImm, "source should not be RawImm internally")
-
-
-class TestArgumentDiscoverability(unittest.TestCase):
-  """
-  Issue: No clear signature for positional arguments.
-
-  inspect.signature(s_load_b128) shows: (*args, literal=None, **kwargs)
-
-  Users have no way to know the argument order without reading source code.
-  The order is implicitly defined by the class field definition order.
-
-  Possible fixes:
-  1. Add explicit parameter names to functools.partial
-  2. Generate type stubs with proper signatures
-  3. Add docstrings listing the expected arguments
-  """
-  def test_signature_has_named_params(self):
-    import inspect
-    sig = inspect.signature(s_load_b128)
-    params = list(sig.parameters.keys())
-    # Currently: ['args', 'literal', 'kwargs'] (from *args, literal=None, **kwargs)
-    # Expected: something like ['sdata', 'sbase', 'soffset', 'offset', 'literal']
-    self.assertIn('sdata', params, "signature should show field names")
-
-
-class TestSpecialConstants(unittest.TestCase):
-  """
-  Issue: NULL and other constants are IntEnum values that might be confusing.
-
-  NULL = SrcEnum.NULL = 124, but users might expect NULL to be a special object
-  that clearly represents "no register" rather than a magic number.
-  """
-  def test_null_has_clear_repr(self):
-    # NULL should have a clear string representation
-    self.assertIn("NULL", str(NULL) or repr(NULL), "NULL should be clearly identifiable")
-
-  def test_null_is_distinguishable_from_int(self):
-    # NULL should be distinguishable from the raw integer 124
-    self.assertNotEqual(type(NULL), int, "NULL should not be plain int")
-
-
-if __name__ == "__main__":
-  unittest.main()

extra/assembly/amd/test/helpers.py

@@ -1,24 +0,0 @@
-"""Shared test helpers for RDNA3 tests."""
-import shutil
-from dataclasses import dataclass
-
-@dataclass
-class KernelInfo:
-  code: bytes
-  global_size: tuple[int, int, int]
-  local_size: tuple[int, int, int]
-  buf_idxs: list[int]  # indices into shared buffer pool
-  buf_sizes: list[int]  # sizes for each buffer index
-
-# LLVM tool detection (shared across test files)
-def get_llvm_mc():
-  """Find llvm-mc executable, preferring newer versions."""
-  for p in ['llvm-mc', 'llvm-mc-21', 'llvm-mc-20']:
-    if shutil.which(p): return p
-  raise FileNotFoundError("llvm-mc not found")
-
-def get_llvm_objdump():
-  """Find llvm-objdump executable, preferring newer versions."""
-  for p in ['llvm-objdump', 'llvm-objdump-21', 'llvm-objdump-20']:
-    if shutil.which(p): return p
-  raise FileNotFoundError("llvm-objdump not found")

extra/assembly/amd/test/test_formats.py

@@ -1,332 +0,0 @@
-#!/usr/bin/env python3
-"""Test MUBUF, MTBUF, MIMG, EXP, DS formats against LLVM."""
-import unittest
-from extra.assembly.amd.autogen.rdna3 import *
-from extra.assembly.amd.dsl import encode_src
-
-class TestMUBUF(unittest.TestCase):
-  """Test MUBUF (buffer) instructions."""
-
-  def test_buffer_load_b32_basic(self):
-    # buffer_load_b32 v5, off, s[8:11], s3 offset:4095
-    # GFX11: encoding: [0xff,0x0f,0x50,0xe0,0x00,0x05,0x02,0x03]
-    inst = buffer_load_b32(vdata=v[5], vaddr=v[0], srsrc=s[8:12], soffset=s[3], offset=4095)
-    self.assertEqual(inst.to_bytes(), bytes([0xff,0x0f,0x50,0xe0,0x00,0x05,0x02,0x03]))
-
-  def test_buffer_load_b32_idxen(self):
-    # buffer_load_b32 v5, v0, s[8:11], s3 idxen offset:4095
-    # GFX11: encoding: [0xff,0x0f,0x50,0xe0,0x00,0x05,0x82,0x03]
-    inst = buffer_load_b32(vdata=v[5], vaddr=v[0], srsrc=s[8:12], soffset=s[3], offset=4095, idxen=1)
-    self.assertEqual(inst.to_bytes(), bytes([0xff,0x0f,0x50,0xe0,0x00,0x05,0x82,0x03]))
-
-  def test_buffer_load_b32_offen(self):
-    # buffer_load_b32 v5, v0, s[8:11], s3 offen offset:4095
-    # GFX11: encoding: [0xff,0x0f,0x50,0xe0,0x00,0x05,0x42,0x03]
-    inst = buffer_load_b32(vdata=v[5], vaddr=v[0], srsrc=s[8:12], soffset=s[3], offset=4095, offen=1)
-    self.assertEqual(inst.to_bytes(), bytes([0xff,0x0f,0x50,0xe0,0x00,0x05,0x42,0x03]))
-
-  def test_buffer_load_b32_glc(self):
-    # buffer_load_b32 v5, off, s[8:11], s3 offset:4095 glc
-    # GFX11: encoding: [0xff,0x4f,0x50,0xe0,0x00,0x05,0x02,0x03]
-    inst = buffer_load_b32(vdata=v[5], vaddr=v[0], srsrc=s[8:12], soffset=s[3], offset=4095, glc=1)
-    self.assertEqual(inst.to_bytes(), bytes([0xff,0x4f,0x50,0xe0,0x00,0x05,0x02,0x03]))
-
-  def test_buffer_load_b32_slc(self):
-    # buffer_load_b32 v5, off, s[8:11], s3 offset:4095 slc
-    # GFX11: encoding: [0xff,0x1f,0x50,0xe0,0x00,0x05,0x02,0x03]
-    inst = buffer_load_b32(vdata=v[5], vaddr=v[0], srsrc=s[8:12], soffset=s[3], offset=4095, slc=1)
-    self.assertEqual(inst.to_bytes(), bytes([0xff,0x1f,0x50,0xe0,0x00,0x05,0x02,0x03]))
-
-  def test_buffer_load_b32_dlc(self):
-    # buffer_load_b32 v5, off, s[8:11], s3 offset:4095 dlc
-    # GFX11: encoding: [0xff,0x2f,0x50,0xe0,0x00,0x05,0x02,0x03]
-    inst = buffer_load_b32(vdata=v[5], vaddr=v[0], srsrc=s[8:12], soffset=s[3], offset=4095, dlc=1)
-    self.assertEqual(inst.to_bytes(), bytes([0xff,0x2f,0x50,0xe0,0x00,0x05,0x02,0x03]))
-
-  def test_buffer_load_b32_all_flags(self):
-    # buffer_load_b32 v5, off, s[8:11], s3 offset:4095 glc slc dlc
-    # GFX11: encoding: [0xff,0x7f,0x50,0xe0,0x00,0x05,0x02,0x03]
-    inst = buffer_load_b32(vdata=v[5], vaddr=v[0], srsrc=s[8:12], soffset=s[3], offset=4095, glc=1, slc=1, dlc=1)
-    self.assertEqual(inst.to_bytes(), bytes([0xff,0x7f,0x50,0xe0,0x00,0x05,0x02,0x03]))
-
-  def test_buffer_store_b32(self):
-    # buffer_store_b32 v1, off, s[12:15], s4 offset:4095
-    # GFX11: encoding: [0xff,0x0f,0x68,0xe0,0x00,0x01,0x03,0x04]
-    inst = buffer_store_b32(vdata=v[1], vaddr=v[0], srsrc=s[12:16], soffset=s[4], offset=4095)
-    self.assertEqual(inst.to_bytes(), bytes([0xff,0x0f,0x68,0xe0,0x00,0x01,0x03,0x04]))
-
-  def test_buffer_load_b64(self):
-    # buffer_load_b64 v[5:6], off, s[8:11], s3 offset:4095
-    # GFX11: encoding: [0xff,0x0f,0x54,0xe0,0x00,0x05,0x02,0x03]
-    inst = buffer_load_b64(vdata=v[5:7], vaddr=v[0], srsrc=s[8:12], soffset=s[3], offset=4095)
-    self.assertEqual(inst.to_bytes(), bytes([0xff,0x0f,0x54,0xe0,0x00,0x05,0x02,0x03]))
-
-  def test_buffer_load_soffset_m0(self):
-    # buffer_load_b32 v5, off, s[8:11], m0 offset:4095
-    # GFX11: encoding: [0xff,0x0f,0x50,0xe0,0x00,0x05,0x02,0x7d]
-    inst = buffer_load_b32(vdata=v[5], vaddr=v[0], srsrc=s[8:12], soffset=M0, offset=4095)
-    self.assertEqual(inst.to_bytes(), bytes([0xff,0x0f,0x50,0xe0,0x00,0x05,0x02,0x7d]))
-
-  def test_buffer_load_soffset_inline_const(self):
-    # buffer_load_b32 v5, off, s[8:11], 0 offset:4095
-    # GFX11: encoding: [0xff,0x0f,0x50,0xe0,0x00,0x05,0x02,0x80]
-    inst = buffer_load_b32(vdata=v[5], vaddr=v[0], srsrc=s[8:12], soffset=0, offset=4095)
-    self.assertEqual(inst.to_bytes(), bytes([0xff,0x0f,0x50,0xe0,0x00,0x05,0x02,0x80]))
-
-  def test_buffer_disasm_roundtrip(self):
-    inst = buffer_load_b32(vdata=v[5], vaddr=v[0], srsrc=s[8:12], soffset=s[3], offset=4095, glc=1)
-    decoded = MUBUF.from_bytes(inst.to_bytes())
-    self.assertEqual(decoded.to_bytes(), inst.to_bytes())
-
-
-class TestMTBUF(unittest.TestCase):
-  """Test MTBUF (typed buffer) instructions."""
-
-  def test_tbuffer_load_format_x(self):
-    # tbuffer_load_format_x v5, off, s[8:11], s3 format:[BUF_FMT_32_FLOAT] offset:4095
-    # BUF_FMT_32_FLOAT = 22
-    # GFX11: encoding: [0xff,0x0f,0xb0,0xe8,0x00,0x05,0x02,0x03]
-    inst = tbuffer_load_format_x(vdata=v[5], vaddr=v[0], srsrc=s[8:12], soffset=s[3], offset=4095, format=22)
-    self.assertEqual(inst.to_bytes(), bytes([0xff,0x0f,0xb0,0xe8,0x00,0x05,0x02,0x03]))
-
-  def test_tbuffer_store_format_x(self):
-    # tbuffer_store_format_x v5, off, s[8:11], s3 format:[BUF_FMT_32_FLOAT] offset:4095
-    # BUF_FMT_32_FLOAT = 22
-    # GFX11: encoding: [0xff,0x0f,0xb2,0xe8,0x00,0x05,0x02,0x03]
-    inst = tbuffer_store_format_x(vdata=v[5], vaddr=v[0], srsrc=s[8:12], soffset=s[3], offset=4095, format=22)
-    self.assertEqual(inst.to_bytes(), bytes([0xff,0x0f,0xb2,0xe8,0x00,0x05,0x02,0x03]))
-
-  def test_tbuffer_load_format_xy(self):
-    # tbuffer_load_format_xy v[5:6], off, s[8:11], s3 format:[BUF_FMT_32_32_FLOAT] offset:4095
-    # BUF_FMT_32_32_FLOAT = 50
-    # GFX11: encoding: [0xff,0x8f,0x90,0xe9,0x00,0x05,0x02,0x03]
-    inst = tbuffer_load_format_xy(vdata=v[5:7], vaddr=v[0], srsrc=s[8:12], soffset=s[3], offset=4095, format=50)
-    self.assertEqual(inst.to_bytes(), bytes([0xff,0x8f,0x90,0xe9,0x00,0x05,0x02,0x03]))
-
-
-class TestMIMG(unittest.TestCase):
-  """Test MIMG (image) instructions."""
-
-  def test_image_load_2d(self):
-    # image_load v[0:3], v[4:5], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D
-    # GFX11: encoding: [0x04,0x0f,0x00,0xf0,0x04,0x00,0x00,0x00]
-    inst = image_load(vdata=v[0:4], vaddr=v[4:6], srsrc=s[0:8], dmask=0xf, dim=1)  # dim=1 is SQ_RSRC_IMG_2D
-    self.assertEqual(inst.to_bytes(), bytes([0x04,0x0f,0x00,0xf0,0x04,0x00,0x00,0x00]))
-
-  def test_image_store_2d(self):
-    # image_store v[0:3], v[4:5], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D
-    # GFX11: encoding: [0x04,0x0f,0x18,0xf0,0x04,0x00,0x00,0x00]
-    inst = image_store(vdata=v[0:4], vaddr=v[4:6], srsrc=s[0:8], dmask=0xf, dim=1)
-    self.assertEqual(inst.to_bytes(), bytes([0x04,0x0f,0x18,0xf0,0x04,0x00,0x00,0x00]))
-
-  def test_image_load_1d(self):
-    # image_load v[0:3], v4, s[0:7] dmask:0xf dim:SQ_RSRC_IMG_1D
-    # GFX11: encoding: [0x00,0x0f,0x00,0xf0,0x04,0x00,0x00,0x00]
-    inst = image_load(vdata=v[0:4], vaddr=v[4], srsrc=s[0:8], dmask=0xf, dim=0)  # dim=0 is SQ_RSRC_IMG_1D
-    self.assertEqual(inst.to_bytes(), bytes([0x00,0x0f,0x00,0xf0,0x04,0x00,0x00,0x00]))
-
-  def test_image_sample(self):
-    # image_sample v[0:3], v[4:5], s[0:7], s[8:11] dmask:0xf dim:SQ_RSRC_IMG_2D
-    # GFX11: encoding: [0x04,0x0f,0x6c,0xf0,0x04,0x00,0x00,0x08]
-    inst = image_sample(vdata=v[0:4], vaddr=v[4:6], srsrc=s[0:8], ssamp=s[8:12], dmask=0xf, dim=1)
-    self.assertEqual(inst.to_bytes(), bytes([0x04,0x0f,0x6c,0xf0,0x04,0x00,0x00,0x08]))
-
-  def test_image_load_d16(self):
-    # image_load v[0:1], v[4:5], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D d16
-    # GFX11: encoding: [0x04,0x0f,0x02,0xf0,0x04,0x00,0x00,0x00]
-    inst = image_load(vdata=v[0:2], vaddr=v[4:6], srsrc=s[0:8], dmask=0xf, dim=1, d16=1)
-    self.assertEqual(inst.to_bytes(), bytes([0x04,0x0f,0x02,0xf0,0x04,0x00,0x00,0x00]))
-
-
-class TestEXP(unittest.TestCase):
-  """Test EXP (export) instructions."""
-
-  def test_exp_mrt0(self):
-    # exp mrt0 v0, v1, v2, v3
-    # GFX11: encoding: [0x0f,0x00,0x00,0xf8,0x00,0x01,0x02,0x03]
-    inst = EXP(en=0xf, target=0, vsrc0=v[0], vsrc1=v[1], vsrc2=v[2], vsrc3=v[3])
-    self.assertEqual(inst.to_bytes(), bytes([0x0f,0x00,0x00,0xf8,0x00,0x01,0x02,0x03]))
-
-  def test_exp_mrtz(self):
-    # exp mrtz v4, v3, v2, v1
-    # GFX11: encoding: [0x8f,0x00,0x00,0xf8,0x04,0x03,0x02,0x01]
-    inst = EXP(en=0xf, target=8, vsrc0=v[4], vsrc1=v[3], vsrc2=v[2], vsrc3=v[1])
-    self.assertEqual(inst.to_bytes(), bytes([0x8f,0x00,0x00,0xf8,0x04,0x03,0x02,0x01]))
-
-  def test_exp_mrtz_done(self):
-    # exp mrtz v4, v3, v2, v1 done
-    # GFX11: encoding: [0x8f,0x08,0x00,0xf8,0x04,0x03,0x02,0x01]
-    inst = EXP(en=0xf, target=8, vsrc0=v[4], vsrc1=v[3], vsrc2=v[2], vsrc3=v[3], done=1)
-    self.assertEqual(inst.to_bytes(), bytes([0x8f,0x08,0x00,0xf8,0x04,0x03,0x02,0x03]))
-
-  def test_exp_partial_mask(self):
-    # exp mrt0 v0, v1, off, off (en=0x3, only first two components)
-    # GFX11: encoding: [0x03,0x00,0x00,0xf8,0x00,0x01,0x00,0x00]
-    inst = EXP(en=0x3, target=0, vsrc0=v[0], vsrc1=v[1], vsrc2=v[0], vsrc3=v[0])
-    self.assertEqual(inst.to_bytes(), bytes([0x03,0x00,0x00,0xf8,0x00,0x01,0x00,0x00]))
-
-  def test_exp_row_en(self):
-    # exp mrtz v4, v3, v2, v1 row_en
-    # GFX11: encoding: [0x8f,0x20,0x00,0xf8,0x04,0x03,0x02,0x01]
-    inst = EXP(en=0xf, target=8, vsrc0=v[4], vsrc1=v[3], vsrc2=v[2], vsrc3=v[1], row=1)
-    self.assertEqual(inst.to_bytes(), bytes([0x8f,0x20,0x00,0xf8,0x04,0x03,0x02,0x01]))
-
-
-class TestDS(unittest.TestCase):
-  """Test DS (data share / LDS) instructions."""
-
-  def test_ds_store_b32(self):
-    # ds_store_b32 v0, v1
-    # GFX11: encoding: [0x00,0x00,0x34,0xd8,0x00,0x01,0x00,0x00]
-    inst = ds_store_b32(addr=v[0], data0=v[1])
-    self.assertEqual(inst.to_bytes(), bytes([0x00,0x00,0x34,0xd8,0x00,0x01,0x00,0x00]))
-
-  def test_ds_load_b32(self):
-    # ds_load_b32 v0, v1
-    # GFX11: encoding: [0x00,0x00,0xd8,0xd8,0x01,0x00,0x00,0x00]
-    inst = ds_load_b32(vdst=v[0], addr=v[1])
-    self.assertEqual(inst.to_bytes(), bytes([0x00,0x00,0xd8,0xd8,0x01,0x00,0x00,0x00]))
-
-  def test_ds_store_b32_offset(self):
-    # ds_store_b32 v0, v1 offset:64
-    # GFX11: encoding: [0x40,0x00,0x34,0xd8,0x00,0x01,0x00,0x00]
-    inst = ds_store_b32(addr=v[0], data0=v[1], offset0=64)
-    self.assertEqual(inst.to_bytes(), bytes([0x40,0x00,0x34,0xd8,0x00,0x01,0x00,0x00]))
-
-  def test_ds_load_b64(self):
-    # ds_load_b64 v[0:1], v2
-    # GFX11: encoding: [0x00,0x00,0xd8,0xd9,0x02,0x00,0x00,0x00]
-    inst = ds_load_b64(vdst=v[0:2], addr=v[2])
-    self.assertEqual(inst.to_bytes(), bytes([0x00,0x00,0xd8,0xd9,0x02,0x00,0x00,0x00]))
-
-  def test_ds_add_u32(self):
-    # ds_add_u32 v0, v1
-    # GFX11: encoding: [0x00,0x00,0x00,0xd8,0x00,0x01,0x00,0x00]
-    inst = ds_add_u32(addr=v[0], data0=v[1])
-    self.assertEqual(inst.to_bytes(), bytes([0x00,0x00,0x00,0xd8,0x00,0x01,0x00,0x00]))
-
-  def test_ds_store_b32_gds(self):
-    # ds_store_b32 v0, v1 gds
-    # GFX11: encoding: [0x00,0x00,0x36,0xd8,0x00,0x01,0x00,0x00]
-    inst = ds_store_b32(addr=v[0], data0=v[1], gds=1)
-    self.assertEqual(inst.to_bytes(), bytes([0x00,0x00,0x36,0xd8,0x00,0x01,0x00,0x00]))
-
-
-class TestVOP3(unittest.TestCase):
-  """Test VOP3 (3-operand vector) instructions."""
-
-  def test_v_fma_f32(self):
-    # v_fma_f32 v0, v1, v2, v3
-    # GFX11: encoding: [0x00,0x00,0x13,0xd6,0x01,0x05,0x0e,0x04]
-    inst = v_fma_f32(vdst=v[0], src0=v[1], src1=v[2], src2=v[3])
-    self.assertEqual(inst.to_bytes(), bytes([0x00,0x00,0x13,0xd6,0x01,0x05,0x0e,0x04]))
-
-  def test_v_mad_f32(self):
-    # v_fmac_f32_e64 v0, v1, v2 (fmac is fma with implicit dst as src2)
-    # Use v_fma_f32 with vdst == src2
-    inst = v_fma_f32(vdst=v[0], src0=v[1], src1=v[2], src2=v[0])
-    self.assertEqual(inst.to_bytes()[:4], bytes([0x00,0x00,0x13,0xd6]))
-
-  def test_v_add3_u32(self):
-    # v_add3_u32 v0, v1, v2, v3
-    # GFX11: encoding: [0x00,0x00,0x55,0xd6,0x01,0x05,0x0e,0x04]
-    inst = v_add3_u32(vdst=v[0], src0=v[1], src1=v[2], src2=v[3])
-    self.assertEqual(inst.to_bytes(), bytes([0x00,0x00,0x55,0xd6,0x01,0x05,0x0e,0x04]))
-
-
-class TestFLAT(unittest.TestCase):
-  """Test FLAT/GLOBAL/SCRATCH memory instructions."""
-
-  def test_global_load_b32(self):
-    # global_load_b32 v0, v[1:2], off (seg=2 for global)
-    # GFX11: encoding: [0x00,0x00,0x52,0xdc,0x01,0x00,0x7c,0x00]
-    inst = global_load_b32(vdst=v[0], addr=v[1:3], saddr=OFF)
-    self.assertEqual(inst.to_bytes(), bytes([0x00,0x00,0x52,0xdc,0x01,0x00,0x7c,0x00]))
-
-  def test_global_store_b32(self):
-    # global_store_b32 v[0:1], v2, off (seg=2 for global)
-    # GFX11: encoding: [0x00,0x00,0x6a,0xdc,0x00,0x02,0x7c,0x00]
-    inst = global_store_b32(addr=v[0:2], data=v[2], saddr=OFF)
-    self.assertEqual(inst.to_bytes(), bytes([0x00,0x00,0x6a,0xdc,0x00,0x02,0x7c,0x00]))
-
-  def test_global_load_b32_saddr(self):
-    # global_load_b32 v0, v1, s[0:1] (seg=2 for global)
-    # GFX11: encoding: [0x00,0x00,0x52,0xdc,0x01,0x00,0x00,0x00]
-    inst = global_load_b32(vdst=v[0], addr=v[1], saddr=s[0:2])
-    self.assertEqual(inst.to_bytes(), bytes([0x00,0x00,0x52,0xdc,0x01,0x00,0x00,0x00]))
-
-  def test_global_load_b32_offset(self):
-    # global_load_b32 v0, v[1:2], off offset:256 (seg=2 for global)
-    # GFX11: encoding: [0x00,0x01,0x52,0xdc,0x01,0x00,0x7c,0x00]
-    inst = global_load_b32(vdst=v[0], addr=v[1:3], saddr=OFF, offset=256)
-    self.assertEqual(inst.to_bytes(), bytes([0x00,0x01,0x52,0xdc,0x01,0x00,0x7c,0x00]))
-
-  def test_global_load_b64(self):
-    # global_load_b64 v[0:1], v[2:3], off (seg=2 for global)
-    # GFX11: encoding: [0x00,0x00,0x56,0xdc,0x02,0x00,0x7c,0x00]
-    inst = global_load_b64(vdst=v[0:2], addr=v[2:4], saddr=OFF)
-    self.assertEqual(inst.to_bytes(), bytes([0x00,0x00,0x56,0xdc,0x02,0x00,0x7c,0x00]))
-
-
-class TestSMEM(unittest.TestCase):
-  """Test SMEM (scalar memory) instructions - regression tests for glc/dlc bit positions."""
-
-  def test_smem_dlc_bit_position(self):
-    # s_load_b32 s5, s[2:3], s0 dlc - tests that DLC is at bit 13 (not bit 14)
-    # GFX11: encoding: [0x41,0x21,0x00,0xf4,0x00,0x00,0x00,0x00]
-    inst = s_load_b32(sdata=s[5], sbase=s[2], soffset=s[0], dlc=1)
-    self.assertEqual(inst.to_bytes(), bytes([0x41,0x21,0x00,0xf4,0x00,0x00,0x00,0x00]))
-
-  def test_smem_glc_bit_position(self):
-    # s_load_b32 s5, s[2:3], s0 glc - tests that GLC is at bit 14 (not bit 16)
-    # GFX11: encoding: [0x41,0x41,0x00,0xf4,0x00,0x00,0x00,0x00]
-    inst = s_load_b32(sdata=s[5], sbase=s[2], soffset=s[0], glc=1)
-    self.assertEqual(inst.to_bytes(), bytes([0x41,0x41,0x00,0xf4,0x00,0x00,0x00,0x00]))
-
-  def test_smem_glc_dlc_combined(self):
-    # s_load_b32 s5, s[2:3], s0 glc dlc - tests both flags together
-    # GFX11: encoding: [0x41,0x61,0x00,0xf4,0x00,0x00,0x00,0x00]
-    inst = s_load_b32(sdata=s[5], sbase=s[2], soffset=s[0], glc=1, dlc=1)
-    self.assertEqual(inst.to_bytes(), bytes([0x41,0x61,0x00,0xf4,0x00,0x00,0x00,0x00]))
-
-  def test_smem_disasm_roundtrip_dlc(self):
-    # Test that disassembly/reassembly preserves DLC bit correctly
-    data = bytes([0x41,0x21,0x00,0xf4,0x00,0x00,0x00,0x00])
-    decoded = SMEM.from_bytes(data)
-    self.assertEqual(decoded.to_bytes(), data)
-
-  def test_smem_disasm_roundtrip_glc_dlc(self):
-    # Test that disassembly/reassembly preserves GLC+DLC bits correctly
-    data = bytes([0x41,0x61,0x00,0xf4,0x00,0x00,0x00,0x00])
-    decoded = SMEM.from_bytes(data)
-    self.assertEqual(decoded.to_bytes(), data)
-
-
-class TestVOP3Literal(unittest.TestCase):
-  """Test VOP3 literal handling - regression tests for Inst64 literal encoding."""
-
-  def test_vop3_with_literal(self):
-    # v_add3_u32 v5, vcc_hi, 0xaf123456, v255
-    # GFX11: encoding: [0x05,0x00,0x55,0xd6,0x6b,0xfe,0xfd,0x07,0x56,0x34,0x12,0xaf]
-    from extra.assembly.amd.dsl import RawImm
-    inst = VOP3(VOP3Op.V_ADD3_U32, vdst=v[5], src0=RawImm(107), src1=0xaf123456, src2=v[255])
-    expected = bytes([0x05,0x00,0x55,0xd6,0x6b,0xfe,0xfd,0x07,0x56,0x34,0x12,0xaf])
-    self.assertEqual(inst.to_bytes(), expected)
-
-  def test_vop3_literal_null_operand(self):
-    # v_add3_u32 v5, null, exec_lo, 0xaf123456
-    # GFX11: encoding: [0x05,0x00,0x55,0xd6,0x7c,0xfc,0xfc,0x03,0x56,0x34,0x12,0xaf]
-    from extra.assembly.amd.dsl import RawImm
-    inst = VOP3(VOP3Op.V_ADD3_U32, vdst=v[5], src0=NULL, src1=RawImm(126), src2=0xaf123456)
-    expected = bytes([0x05,0x00,0x55,0xd6,0x7c,0xfc,0xfc,0x03,0x56,0x34,0x12,0xaf])
-    self.assertEqual(inst.to_bytes(), expected)
-
-  def test_vop3p_with_literal(self):
-    # Test VOP3P literal encoding (also uses Inst64)
-    from extra.assembly.amd.dsl import RawImm
-    inst = VOP3P(VOP3POp.V_PK_ADD_F16, vdst=v[5], src0=RawImm(240), src1=0x12345678, src2=v[0])
-    self.assertEqual(len(inst.to_bytes()), 12)  # 8 bytes + 4 byte literal
-
-
-if __name__ == "__main__":
-  unittest.main()

extra/assembly/amd/test/test_integration.py

@@ -1,330 +0,0 @@
-#!/usr/bin/env python3
-"""Integration test: round-trip RDNA3 assembly through AMD toolchain."""
-import unittest, re, io, sys, subprocess
-from extra.assembly.amd.autogen.rdna3 import *
-from extra.assembly.amd.asm import waitcnt, asm
-from extra.assembly.amd.test.helpers import get_llvm_mc
-
-def disassemble(lib: bytes, arch: str = "gfx1100") -> str:
-  """Disassemble ELF binary using tinygrad's compiler, return raw output."""
-  from tinygrad.runtime.support.compiler_amd import HIPCompiler
-  old_stdout = sys.stdout
-  sys.stdout = io.StringIO()
-  HIPCompiler(arch).disassemble(lib)
-  output = sys.stdout.getvalue()
-  sys.stdout = old_stdout
-  return output
-
-def parse_disassembly(raw: str) -> list[str]:
-  """Parse disassembly output to list of instruction mnemonics."""
-  lines = []
-  for line in raw.splitlines():
-    if line.startswith('\t'):
-      instr = line.split('//')[0].strip()
-      if instr: lines.append(instr)
-  return lines
-
-def assemble_and_disassemble(instructions: list, arch: str = "gfx1100") -> list[str]:
-  """Assemble instructions with our DSL, then disassemble with AMD toolchain."""
-  from tinygrad.runtime.support.compiler_amd import HIPCompiler
-
-  # Generate bytes from our DSL
-  code_bytes = b''.join(inst.to_bytes() for inst in instructions)
-
-  # Wrap in minimal ELF-compatible assembly with .byte directives
-  byte_str = ', '.join(f'0x{b:02x}' for b in code_bytes)
-  asm_src = f".text\n.globl test\n.p2align 8\n.type test,@function\ntest:\n.byte {byte_str}\n"
-
-  # Assemble with AMD COMGR and disassemble
-  lib = HIPCompiler(arch).compile(asm_src)
-  return parse_disassembly(disassemble(lib, arch))
-
-class TestIntegration(unittest.TestCase):
-  """Test our assembler output matches LLVM disassembly."""
-
-  def test_simple_sop1(self):
-    """Test SOP1 instructions round-trip."""
-    instructions = [
-      s_mov_b32(s[0], s[1]),
-      s_mov_b32(s[2], 0),
-      s_not_b32(s[3], s[4]),
-    ]
-    disasm = assemble_and_disassemble(instructions)
-    self.assertIn('s_mov_b32', disasm[0])
-    self.assertIn('s_mov_b32', disasm[1])
-    self.assertIn('s_not_b32', disasm[2])
-
-  def test_simple_sop2(self):
-    """Test SOP2 instructions round-trip."""
-    instructions = [
-      s_add_u32(s[0], s[1], s[2]),
-      s_sub_u32(s[3], s[4], 10),
-      s_and_b32(s[5], s[6], s[7]),
-    ]
-    disasm = assemble_and_disassemble(instructions)
-    self.assertIn('s_add_u32', disasm[0])
-    self.assertIn('s_sub_u32', disasm[1])
-    self.assertIn('s_and_b32', disasm[2])
-
-  def test_simple_vop2(self):
-    """Test VOP2 instructions round-trip."""
-    instructions = [
-      v_add_f32_e32(v[0], v[1], v[2]),
-      v_mul_f32_e32(v[3], 1.0, v[4]),  # 1.0 is inline constant
-      v_and_b32_e32(v[5], 10, v[6]),  # small inline constant
-    ]
-    disasm = assemble_and_disassemble(instructions)
-    self.assertIn('v_add_f32', disasm[0])
-    self.assertIn('v_mul_f32', disasm[1])
-
-  def test_control_flow(self):
-    """Test control flow instructions."""
-    instructions = [
-      s_waitcnt(simm16=waitcnt(lgkmcnt=0)),
-      s_endpgm(),
-    ]
-    disasm = assemble_and_disassemble(instructions)
-    self.assertIn('s_waitcnt', disasm[0])
-    self.assertIn('s_endpgm', disasm[1])
-
-  def test_memory_ops(self):
-    """Test memory instructions."""
-    instructions = [
-      s_load_b32(s[0], s[0:2], NULL),
-      s_waitcnt(simm16=waitcnt(lgkmcnt=0)),
-      global_store_b32(addr=v[0:2], data=v[2], saddr=OFF),
-      s_endpgm(),
-    ]
-    disasm = assemble_and_disassemble(instructions)
-    self.assertIn('s_load_b32', disasm[0])
-    self.assertIn('s_waitcnt', disasm[1])
-    self.assertIn('global_store_b32', disasm[2])
-
-  def test_full_kernel(self):
-    """Test a complete kernel similar to tinygrad output."""
-    # Simple kernel: load value, add 1, store back
-    instructions = [
-      # Get thread ID
-      v_mov_b32_e32(v[0], s[0]),  # base addr low
-      v_mov_b32_e32(v[1], s[1]),  # base addr high
-      # Load value
-      global_load_b32(vdst=v[2], addr=v[0:2], saddr=OFF),
-      s_waitcnt(simm16=waitcnt(vmcnt=0)),
-      # Add 1.0
-      v_add_f32_e32(v[2], 1.0, v[2]),
-      # Store result
-      global_store_b32(addr=v[0:2], data=v[2], saddr=OFF),
-      s_endpgm(),
-    ]
-    disasm = assemble_and_disassemble(instructions)
-    # Verify key instructions are present
-    self.assertTrue(any('global_load' in d for d in disasm))
-    self.assertTrue(any('v_add_f32' in d for d in disasm))
-    self.assertTrue(any('global_store' in d for d in disasm))
-    self.assertTrue(any('s_endpgm' in d for d in disasm))
-
-  def test_bytes_roundtrip(self):
-    """Test that our bytes match what AMD assembler produces."""
-    from tinygrad.runtime.support.compiler_amd import HIPCompiler
-
-    # Simple instruction
-    inst = s_mov_b32(s[0], s[1])
-    our_bytes = inst.to_bytes()
-
-    # Assemble same instruction with AMD toolchain
-    asm_src = ".text\n.globl test\n.p2align 8\n.type test,@function\ntest:\ns_mov_b32 s0, s1\n"
-    compiler = HIPCompiler("gfx1100")
-    lib = compiler.compile(asm_src)
-    raw = disassemble(lib)
-
-    for line in raw.splitlines():
-      if 's_mov_b32' in line and '//' in line:
-        # Extract hex bytes from comment: "// 000000001300: BE800001"
-        comment = line.split('//')[1].strip()
-        hex_str = comment.split(':')[1].strip()
-        # Convert big-endian hex string to little-endian bytes
-        amd_bytes = bytes.fromhex(hex_str)[::-1]  # reverse for little-endian
-        self.assertEqual(our_bytes, amd_bytes, f"Bytes mismatch: ours={our_bytes.hex()} AMD={amd_bytes.hex()}")
-        return
-    self.fail("Could not find s_mov_b32 in disassembly")
-
-class TestAsm(unittest.TestCase):
-  """Test asm() string parsing."""
-
-  def test_asm_basic(self):
-    """Test basic instruction parsing."""
-    inst = asm('s_mov_b32 s0, s1')
-    self.assertEqual(inst.to_bytes(), s_mov_b32(s[0], s[1]).to_bytes())
-
-  def test_asm_with_immediates(self):
-    """Test parsing with immediate values."""
-    inst = asm('s_add_u32 s0, s1, 10')
-    self.assertEqual(inst.to_bytes(), s_add_u32(s[0], s[1], 10).to_bytes())
-
-  def test_asm_float_const(self):
-    """Test parsing float constants."""
-    inst = asm('v_mul_f32_e32 v0, 1.0, v1')
-    self.assertEqual(inst.to_bytes(), v_mul_f32_e32(v[0], 1.0, v[1]).to_bytes())
-
-  def test_asm_hex_immediate(self):
-    """Test parsing hex immediates."""
-    inst = asm('s_waitcnt 0xfc07')
-    self.assertEqual(inst.to_bytes(), s_waitcnt(simm16=0xfc07).to_bytes())
-
-  def test_asm_special_regs(self):
-    """Test parsing special registers."""
-    inst = asm('s_mov_b32 s0, vcc_lo')
-    self.assertEqual(inst.to_bytes(), s_mov_b32(s[0], VCC_LO).to_bytes())
-
-  def test_asm_register_range(self):
-    """Test parsing register ranges."""
-    inst = asm('s_load_b128 s[4:7], s[0:1], null')
-    self.assertEqual(inst.to_bytes(), s_load_b128(s[4:7], s[0:1], NULL).to_bytes())
-
-  def test_asm_matches_llvm(self):
-    """Test asm() output matches LLVM assembler."""
-    from tinygrad.runtime.support.compiler_amd import HIPCompiler
-    compiler = HIPCompiler('gfx1100')
-
-    def get_llvm_bytes(instr: str) -> bytes:
-      src = f'.text\n.globl test\n.p2align 8\n.type test,@function\ntest:\n{instr}\n'
-      lib = compiler.compile(src)
-      raw = disassemble(lib)
-      for line in raw.splitlines():
-        if instr.split()[0] in line and '//' in line:
-          hex_str = line.split('//')[1].strip().split(':')[1].strip()
-          return bytes.fromhex(hex_str)[::-1]
-      return b''
-
-    tests = ['s_mov_b32 s0, s1', 's_endpgm', 'v_add_f32_e32 v0, v1, v2']
-    for t in tests:
-      self.assertEqual(asm(t).to_bytes(), get_llvm_bytes(t), f"mismatch for: {t}")
-
-  def test_asm_vop3_modifiers(self):
-    """Test asm() with VOP3 modifiers (neg, abs, clamp)."""
-    def get_llvm_encoding(instr: str) -> str:
-      result = subprocess.run([get_llvm_mc(), '-triple=amdgcn', '-mcpu=gfx1100', '-show-encoding'],
-                              input=instr, capture_output=True, text=True)
-      if m := re.search(r'encoding:\s*\[(.*?)\]', result.stdout):
-        return m.group(1).replace('0x','').replace(',','').replace(' ','')
-      return ''
-
-    tests = [
-      'v_fma_f32 v0, -v1, v2, v3',       # neg on src0
-      'v_fma_f32 v0, v1, |v2|, v3',      # abs on src1
-      'v_fma_f32 v0, v1, v2, v3 clamp',  # clamp
-      'v_fma_f32 v0, -v1, |v2|, v3 clamp',  # all modifiers
-      'v_fma_f32 v0, -|v1|, v2, v3',     # neg+abs on same operand
-    ]
-    for t in tests:
-      our_hex = asm(t).to_bytes().hex()
-      llvm_hex = get_llvm_encoding(t)
-      self.assertEqual(our_hex, llvm_hex, f"mismatch for: {t}")
-
-class TestTinygradIntegration(unittest.TestCase):
-  """Test that we can parse disassembled tinygrad kernels."""
-
-  def test_simple_add_kernel(self):
-    """Generate a simple add kernel from tinygrad and verify disassembly."""
-    from tinygrad import Tensor
-    from tinygrad.codegen import get_program
-    from tinygrad.renderer.cstyle import AMDHIPRenderer
-    from tinygrad.runtime.support.compiler_amd import HIPCompiler
-    from tinygrad.uop.ops import Ops
-
-    # Create a computation that generates a real kernel
-    a = Tensor([1.0, 2.0, 3.0, 4.0]).realize()
-    b = Tensor([5.0, 6.0, 7.0, 8.0]).realize()
-    c = a + b
-
-    # Get schedule and find SINK
-    schedule = c.schedule()
-    sink_items = [si for si in schedule if si.ast.op == Ops.SINK]
-    self.assertTrue(len(sink_items) > 0, "No SINK in schedule")
-
-    # Generate program
-    renderer = AMDHIPRenderer('gfx1100')
-    prg = get_program(sink_items[0].ast, renderer)
-    self.assertIsNotNone(prg.src)
-
-    # Compile and disassemble
-    compiler = HIPCompiler('gfx1100')
-    lib = compiler.compile(prg.src)
-    raw_disasm = disassemble(lib)
-    instrs = parse_disassembly(raw_disasm)
-
-    # Verify we got some instructions
-    self.assertTrue(len(instrs) > 0, "No instructions in disassembly")
-    # Should have an endpgm
-    self.assertTrue(any('s_endpgm' in i for i in instrs), "Missing s_endpgm")
-
-  def test_matmul_kernel(self):
-    """Generate a matmul kernel and verify disassembly has expected patterns."""
-    from tinygrad import Tensor
-    from tinygrad.codegen import get_program
-    from tinygrad.renderer.cstyle import AMDHIPRenderer
-    from tinygrad.runtime.support.compiler_amd import HIPCompiler
-    from tinygrad.uop.ops import Ops
-
-    # Create a small matmul
-    a = Tensor.rand(4, 4).realize()
-    b = Tensor.rand(4, 4).realize()
-    c = a @ b
-
-    # Get schedule
-    schedule = c.schedule()
-    sink_items = [si for si in schedule if si.ast.op == Ops.SINK]
-    self.assertTrue(len(sink_items) > 0)
-
-    # Generate and compile
-    renderer = AMDHIPRenderer('gfx1100')
-    prg = get_program(sink_items[0].ast, renderer)
-    compiler = HIPCompiler('gfx1100')
-    lib = compiler.compile(prg.src)
-    raw_disasm = disassemble(lib)
-    instrs = parse_disassembly(raw_disasm)
-
-    # Matmul should have multiply and add instructions
-    has_mul = any('mul' in i.lower() for i in instrs)
-    has_add = any('add' in i.lower() for i in instrs)
-    self.assertTrue(has_mul or has_add, "Matmul should have mul/add ops")
-
-  def test_disasm_to_bytes_roundtrip(self):
-    """Parse disassembled instructions and verify we can re-encode some of them."""
-    from tinygrad import Tensor
-    from tinygrad.codegen import get_program
-    from tinygrad.renderer.cstyle import AMDHIPRenderer
-    from tinygrad.runtime.support.compiler_amd import HIPCompiler
-    from tinygrad.uop.ops import Ops
-
-    # Simple kernel
-    a = Tensor([1.0, 2.0, 3.0, 4.0]).realize()
-    b = (a * 2.0)
-
-    schedule = b.schedule()
-    sink_items = [si for si in schedule if si.ast.op == Ops.SINK]
-    if not sink_items: return  # skip if no kernel
-
-    renderer = AMDHIPRenderer('gfx1100')
-    prg = get_program(sink_items[0].ast, renderer)
-    compiler = HIPCompiler('gfx1100')
-    lib = compiler.compile(prg.src)
-    raw_disasm = disassemble(lib)
-
-    # Find s_endpgm and verify we can encode it
-    for line in raw_disasm.splitlines():
-      if 's_endpgm' in line and '//' in line:
-        # Extract bytes from comment
-        comment = line.split('//')[1].strip()
-        hex_str = comment.split(':')[1].strip()
-        amd_bytes = bytes.fromhex(hex_str)[::-1]
-
-        # Our encoding
-        our_inst = s_endpgm()
-        our_bytes = our_inst.to_bytes()
-
-        self.assertEqual(our_bytes, amd_bytes, f"s_endpgm mismatch: ours={our_bytes.hex()} AMD={amd_bytes.hex()}")
-        return
-
-if __name__ == "__main__":
-  unittest.main()

extra/assembly/amd/test/test_llvm.py

@@ -1,195 +0,0 @@
-#!/usr/bin/env python3
-"""Test RDNA3 assembler/disassembler against LLVM test vectors."""
-import unittest, re, subprocess
-from tinygrad.helpers import fetch
-from extra.assembly.amd.autogen.rdna3 import *
-from extra.assembly.amd.asm import asm
-from extra.assembly.amd.test.helpers import get_llvm_mc
-
-LLVM_BASE = "https://raw.githubusercontent.com/llvm/llvm-project/main/llvm/test/MC/AMDGPU"
-
-# Format info: (filename, format_class, op_enum)
-LLVM_TEST_FILES = {
-  # Scalar ALU
-  'sop1': ('gfx11_asm_sop1.s', SOP1, SOP1Op),
-  'sop2': ('gfx11_asm_sop2.s', SOP2, SOP2Op),
-  'sopp': ('gfx11_asm_sopp.s', SOPP, SOPPOp),
-  'sopk': ('gfx11_asm_sopk.s', SOPK, SOPKOp),
-  'sopc': ('gfx11_asm_sopc.s', SOPC, SOPCOp),
-  # Vector ALU
-  'vop1': ('gfx11_asm_vop1.s', VOP1, VOP1Op),
-  'vop2': ('gfx11_asm_vop2.s', VOP2, VOP2Op),
-  'vopc': ('gfx11_asm_vopc.s', VOPC, VOPCOp),
-  'vop3': ('gfx11_asm_vop3.s', VOP3, VOP3Op),
-  'vop3p': ('gfx11_asm_vop3p.s', VOP3P, VOP3POp),
-  'vop3sd': ('gfx11_asm_vop3.s', VOP3SD, VOP3SDOp),  # VOP3SD shares file with VOP3
-  'vinterp': ('gfx11_asm_vinterp.s', VINTERP, VINTERPOp),
-  'vopd': ('gfx11_asm_vopd.s', VOPD, VOPDOp),
-  'vopcx': ('gfx11_asm_vopcx.s', VOPC, VOPCOp),  # VOPCX uses VOPC format
-  # VOP3 promotions (VOP1/VOP2/VOPC promoted to VOP3 encoding)
-  'vop3_from_vop1': ('gfx11_asm_vop3_from_vop1.s', VOP3, VOP3Op),
-  'vop3_from_vop2': ('gfx11_asm_vop3_from_vop2.s', VOP3, VOP3Op),
-  'vop3_from_vopc': ('gfx11_asm_vop3_from_vopc.s', VOP3, VOP3Op),
-  'vop3_from_vopcx': ('gfx11_asm_vop3_from_vopcx.s', VOP3, VOP3Op),
-  # Memory
-  'ds': ('gfx11_asm_ds.s', DS, DSOp),
-  'smem': ('gfx11_asm_smem.s', SMEM, SMEMOp),
-  'flat': ('gfx11_asm_flat.s', FLAT, FLATOp),
-  'mubuf': ('gfx11_asm_mubuf.s', MUBUF, MUBUFOp),
-  'mtbuf': ('gfx11_asm_mtbuf.s', MTBUF, MTBUFOp),
-  'mimg': ('gfx11_asm_mimg.s', MIMG, MIMGOp),
-  # WMMA (matrix multiply)
-  'wmma': ('gfx11_asm_wmma.s', VOP3P, VOP3POp),
-  # Additional features
-  'vop3_features': ('gfx11_asm_vop3_features.s', VOP3, VOP3Op),
-  'vop3p_features': ('gfx11_asm_vop3p_features.s', VOP3P, VOP3POp),
-  'vopd_features': ('gfx11_asm_vopd_features.s', VOPD, VOPDOp),
-  # Alias files (alternative mnemonics)
-  'vop3_alias': ('gfx11_asm_vop3_alias.s', VOP3, VOP3Op),
-  'vop3p_alias': ('gfx11_asm_vop3p_alias.s', VOP3P, VOP3POp),
-  'vopc_alias': ('gfx11_asm_vopc_alias.s', VOPC, VOPCOp),
-  'vopcx_alias': ('gfx11_asm_vopcx_alias.s', VOPC, VOPCOp),
-  'vinterp_alias': ('gfx11_asm_vinterp_alias.s', VINTERP, VINTERPOp),
-  'smem_alias': ('gfx11_asm_smem_alias.s', SMEM, SMEMOp),
-  'mubuf_alias': ('gfx11_asm_mubuf_alias.s', MUBUF, MUBUFOp),
-  'mtbuf_alias': ('gfx11_asm_mtbuf_alias.s', MTBUF, MTBUFOp),
-}
-
-def parse_llvm_tests(text: str) -> list[tuple[str, bytes]]:
-  """Parse LLVM test format into (asm, expected_bytes) pairs."""
-  tests, lines = [], text.split('\n')
-  for i, line in enumerate(lines):
-    line = line.strip()
-    if not line or line.startswith(('//', '.', ';')): continue
-    asm_text = line.split('//')[0].strip()
-    if not asm_text: continue
-    for j in range(i, min(i + 3, len(lines))):
-      # Match GFX11, W32, or W64 encodings (all valid for gfx11)
-      if m := re.search(r'(?:GFX11|W32|W64)[^:]*:.*?encoding:\s*\[(.*?)\]', lines[j]):
-        hex_bytes = m.group(1).replace('0x', '').replace(',', '').replace(' ', '')
-        if hex_bytes:
-          try: tests.append((asm_text, bytes.fromhex(hex_bytes)))
-          except ValueError: pass
-        break
-  return tests
-
-def try_assemble(text: str):
-  """Try to assemble instruction text, return bytes or None on failure."""
-  try: return asm(text).to_bytes()
-  except: return None
-
-def compile_asm_batch(instrs: list[str]) -> list[bytes]:
-  """Compile multiple instructions with a single llvm-mc call."""
-  if not instrs: return []
-  asm_text = ".text\n" + "\n".join(instrs) + "\n"
-  result = subprocess.run(
-    [get_llvm_mc(), '-triple=amdgcn', '-mcpu=gfx1100', '-mattr=+real-true16,+wavefrontsize32', '-show-encoding'],
-    input=asm_text, capture_output=True, text=True, timeout=30)
-  if result.returncode != 0: raise RuntimeError(f"llvm-mc batch failed: {result.stderr.strip()}")
-  # Parse all encodings from output
-  results = []
-  for line in result.stdout.split('\n'):
-    if 'encoding:' not in line: continue
-    enc = line.split('encoding:')[1].strip()
-    if enc.startswith('[') and enc.endswith(']'):
-      results.append(bytes.fromhex(enc[1:-1].replace('0x', '').replace(',', '').replace(' ', '')))
-  if len(results) != len(instrs): raise RuntimeError(f"expected {len(instrs)} encodings, got {len(results)}")
-  return results
-
-class TestLLVM(unittest.TestCase):
-  """Test assembler and disassembler against all LLVM test vectors."""
-  tests: dict[str, list[tuple[str, bytes]]] = {}
-
-  @classmethod
-  def setUpClass(cls):
-    for name, (filename, _, _) in LLVM_TEST_FILES.items():
-      try:
-        data = fetch(f"{LLVM_BASE}/{filename}").read_bytes()
-        cls.tests[name] = parse_llvm_tests(data.decode('utf-8', errors='ignore'))
-      except Exception as e:
-        print(f"Warning: couldn't fetch {filename}: {e}")
-        cls.tests[name] = []
-
-# Generate test methods dynamically for each format
-def _make_asm_test(name):
-  def test(self):
-    passed, failed, skipped = 0, 0, 0
-    for asm_text, expected in self.tests.get(name, []):
-      result = try_assemble(asm_text)
-      if result is None: skipped += 1
-      elif result == expected: passed += 1
-      else: failed += 1
-    print(f"{name.upper()} asm: {passed} passed, {failed} failed, {skipped} skipped")
-    self.assertEqual(failed, 0)
-  return test
-
-def _make_disasm_test(name):
-  def test(self):
-    _, fmt_cls, op_enum = LLVM_TEST_FILES[name]
-    # VOP3SD opcodes that share encoding with VOP3 (only for vop3sd test, not vopc promotions)
-    vop3sd_opcodes = {288, 289, 290, 764, 765, 766, 767, 768, 769, 770}
-    is_vopc_promotion = name in ('vop3_from_vopc', 'vop3_from_vopcx')
-    undocumented = {'smem': {34, 35}, 'sopk': {22, 23}, 'sopp': {8, 58, 59}}
-
-    # First pass: decode all instructions and collect disasm strings
-    to_test: list[tuple[str, bytes, str | None, str | None]] = []  # (asm_text, data, disasm_str, error)
-    skipped = 0
-    for asm_text, data in self.tests.get(name, []):
-      if len(data) > fmt_cls._size(): continue
-      temp_inst = fmt_cls.from_bytes(data)
-      temp_op = temp_inst._values.get('op', 0)
-      temp_op = temp_op.val if hasattr(temp_op, 'val') else temp_op
-      if temp_op in undocumented.get(name, set()): skipped += 1; continue
-      if name == 'sopp':
-        simm16 = temp_inst._values.get('simm16', 0)
-        simm16 = simm16.val if hasattr(simm16, 'val') else simm16
-        sopp_no_imm = {48, 54, 53, 55, 60, 61, 62}
-        if temp_op in sopp_no_imm and simm16 != 0: skipped += 1; continue
-      try:
-        if fmt_cls.__name__ in ('VOP3', 'VOP3SD'):
-          temp = VOP3.from_bytes(data)
-          op_val = temp._values.get('op', 0)
-          op_val = op_val.val if hasattr(op_val, 'val') else op_val
-          is_vop3sd = (op_val in vop3sd_opcodes) and not is_vopc_promotion
-          decoded = VOP3SD.from_bytes(data) if is_vop3sd else VOP3.from_bytes(data)
-          if is_vop3sd: VOP3SDOp(op_val)
-          else: VOP3Op(op_val)
-        else:
-          decoded = fmt_cls.from_bytes(data)
-          op_val = decoded._values.get('op', 0)
-          op_val = op_val.val if hasattr(op_val, 'val') else op_val
-          op_enum(op_val)
-        if decoded.to_bytes()[:len(data)] != data:
-          to_test.append((asm_text, data, None, "decode roundtrip failed"))
-          continue
-        to_test.append((asm_text, data, decoded.disasm(), None))
-      except Exception as e:
-        to_test.append((asm_text, data, None, f"exception: {e}"))
-
-    # Batch compile all disasm strings with single llvm-mc call
-    disasm_strs = [(i, t[2]) for i, t in enumerate(to_test) if t[2] is not None]
-    llvm_results = compile_asm_batch([s for _, s in disasm_strs]) if disasm_strs else []
-    llvm_map = {i: llvm_results[j] for j, (i, _) in enumerate(disasm_strs)}
-
-    # Match results back
-    passed, failed = 0, 0
-    failures: list[str] = []
-    for idx, (asm_text, data, disasm_str, error) in enumerate(to_test):
-      if error:
-        failed += 1; failures.append(f"{error} for {data.hex()}")
-      elif disasm_str is not None and idx in llvm_map:
-        llvm_bytes = llvm_map[idx]
-        if llvm_bytes is not None and llvm_bytes == data: passed += 1
-        elif llvm_bytes is not None: failed += 1; failures.append(f"'{disasm_str}': expected={data.hex()} got={llvm_bytes.hex()}")
-
-    print(f"{name.upper()} disasm: {passed} passed, {failed} failed" + (f", {skipped} skipped" if skipped else ""))
-    if failures[:10]: print("  " + "\n  ".join(failures[:10]))
-    self.assertEqual(failed, 0)
-  return test
-
-for name in LLVM_TEST_FILES:
-  setattr(TestLLVM, f'test_{name}_asm', _make_asm_test(name))
-  setattr(TestLLVM, f'test_{name}_disasm', _make_disasm_test(name))
-
-if __name__ == "__main__":
-  unittest.main()

extra/assembly/amd/test/test_pcode.py

@@ -1,269 +0,0 @@
-#!/usr/bin/env python3
-"""Tests for the RDNA3 pseudocode DSL."""
-import unittest
-from extra.assembly.amd.pcode import Reg, TypedView, SliceProxy, ExecContext, compile_pseudocode, _expr, MASK32, MASK64, _f32, _i32, _f16, _i16, f32_to_f16, _isnan
-from extra.assembly.amd.autogen.rdna3.gen_pcode import _VOP3SDOp_V_DIV_SCALE_F32, _VOPCOp_V_CMP_CLASS_F32
-
-class TestReg(unittest.TestCase):
-  def test_u32_read(self):
-    r = Reg(0xDEADBEEF)
-    self.assertEqual(int(r.u32), 0xDEADBEEF)
-
-  def test_u32_write(self):
-    r = Reg(0)
-    r.u32 = 0x12345678
-    self.assertEqual(r._val, 0x12345678)
-
-  def test_f32_read(self):
-    r = Reg(0x40400000)  # 3.0f
-    self.assertAlmostEqual(float(r.f32), 3.0)
-
-  def test_f32_write(self):
-    r = Reg(0)
-    r.f32 = 3.0
-    self.assertEqual(r._val, 0x40400000)
-
-  def test_i32_signed(self):
-    r = Reg(0xFFFFFFFF)  # -1 as signed
-    self.assertEqual(int(r.i32), -1)
-
-  def test_u64(self):
-    r = Reg(0xDEADBEEFCAFEBABE)
-    self.assertEqual(int(r.u64), 0xDEADBEEFCAFEBABE)
-
-  def test_f64(self):
-    r = Reg(0x4008000000000000)  # 3.0 as f64
-    self.assertAlmostEqual(float(r.f64), 3.0)
-
-class TestTypedView(unittest.TestCase):
-  def test_bit_slice(self):
-    r = Reg(0xDEADBEEF)
-    # Slices return SliceProxy which supports .u32, .u16 etc (matching pseudocode like S1.u32[1:0].u32)
-    self.assertEqual(r.u32[7:0].u32, 0xEF)
-    self.assertEqual(r.u32[15:8].u32, 0xBE)
-    self.assertEqual(r.u32[23:16].u32, 0xAD)
-    self.assertEqual(r.u32[31:24].u32, 0xDE)
-    # Also works with int() for arithmetic
-    self.assertEqual(int(r.u32[7:0]), 0xEF)
-
-  def test_single_bit_read(self):
-    r = Reg(0b11010101)
-    self.assertEqual(r.u32[0], 1)
-    self.assertEqual(r.u32[1], 0)
-    self.assertEqual(r.u32[2], 1)
-    self.assertEqual(r.u32[3], 0)
-
-  def test_single_bit_write(self):
-    r = Reg(0)
-    r.u32[5] = 1
-    r.u32[3] = 1
-    self.assertEqual(r._val, 0b00101000)
-
-  def test_nested_bit_access(self):
-    # S0.u32[S1.u32[4:0]] - access bit at position from another register
-    s0 = Reg(0b11010101)
-    s1 = Reg(3)
-    bit_pos = s1.u32[4:0]  # SliceProxy, int value = 3
-    bit_val = s0.u32[int(bit_pos)]  # bit 3 of s0 = 0
-    self.assertEqual(int(bit_pos), 3)
-    self.assertEqual(bit_val, 0)
-
-  def test_arithmetic(self):
-    r1 = Reg(0x40400000)  # 3.0f
-    r2 = Reg(0x40800000)  # 4.0f
-    result = r1.f32 + r2.f32
-    self.assertAlmostEqual(result, 7.0)
-
-  def test_comparison(self):
-    r1 = Reg(5)
-    r2 = Reg(3)
-    self.assertTrue(r1.u32 > r2.u32)
-    self.assertFalse(r1.u32 < r2.u32)
-    self.assertTrue(r1.u32 != r2.u32)
-
-class TestSliceProxy(unittest.TestCase):
-  def test_slice_read(self):
-    r = Reg(0x56781234)
-    self.assertEqual(r[15:0].u16, 0x1234)
-    self.assertEqual(r[31:16].u16, 0x5678)
-
-  def test_slice_write(self):
-    r = Reg(0)
-    r[15:0].u16 = 0x1234
-    r[31:16].u16 = 0x5678
-    self.assertEqual(r._val, 0x56781234)
-
-  def test_slice_f16(self):
-    r = Reg(0)
-    r[15:0].f16 = 3.0
-    self.assertAlmostEqual(_f16(r._val & 0xffff), 3.0, places=2)
-
-class TestCompiler(unittest.TestCase):
-  def test_ternary(self):
-    result = _expr("a > b ? 1 : 0")
-    self.assertIn("if", result)
-    self.assertIn("else", result)
-
-  def test_type_prefix_strip(self):
-    self.assertEqual(_expr("1'0U"), "0")
-    self.assertEqual(_expr("32'1"), "1")
-    self.assertEqual(_expr("16'0xFFFF"), "0xFFFF")
-
-  def test_suffix_strip(self):
-    self.assertEqual(_expr("0ULL"), "0")
-    self.assertEqual(_expr("1LL"), "1")
-    self.assertEqual(_expr("5U"), "5")
-    self.assertEqual(_expr("3.14F"), "3.14")
-
-  def test_boolean_ops(self):
-    self.assertIn("and", _expr("a && b"))
-    self.assertIn("or", _expr("a || b"))
-    self.assertIn("!=", _expr("a <> b"))
-
-  def test_pack16(self):
-    result = _expr("{ a, b }")
-    self.assertIn("_pack", result)
-
-  def test_type_cast_strip(self):
-    self.assertEqual(_expr("64'U(x)"), "(x)")
-    self.assertEqual(_expr("32'I(y)"), "(y)")
-
-class TestExecContext(unittest.TestCase):
-  def test_float_add(self):
-    ctx = ExecContext(s0=0x40400000, s1=0x40800000)  # 3.0f, 4.0f
-    ctx.D0.f32 = ctx.S0.f32 + ctx.S1.f32
-    self.assertAlmostEqual(_f32(ctx.D0._val), 7.0)
-
-  def test_float_mul(self):
-    ctx = ExecContext(s0=0x40400000, s1=0x40800000)  # 3.0f, 4.0f
-    ctx.run("D0.f32 = S0.f32 * S1.f32")
-    self.assertAlmostEqual(_f32(ctx.D0._val), 12.0)
-
-  def test_scc_comparison(self):
-    ctx = ExecContext(s0=42, s1=42)
-    ctx.run("SCC = S0.u32 == S1.u32")
-    self.assertEqual(ctx.SCC._val, 1)
-
-  def test_scc_comparison_false(self):
-    ctx = ExecContext(s0=42, s1=43)
-    ctx.run("SCC = S0.u32 == S1.u32")
-    self.assertEqual(ctx.SCC._val, 0)
-
-  def test_ternary(self):
-    code = compile_pseudocode("D0.u32 = S0.u32 > S1.u32 ? 1'1U : 1'0U")
-    ctx = ExecContext(s0=5, s1=3)
-    ctx.run(code)
-    self.assertEqual(ctx.D0._val, 1)
-
-  def test_pack(self):
-    code = compile_pseudocode("D0 = { S1[15:0].u16, S0[15:0].u16 }")
-    ctx = ExecContext(s0=0x1234, s1=0x5678)
-    ctx.run(code)
-    self.assertEqual(ctx.D0._val, 0x56781234)
-
-  def test_tmp_with_typed_access(self):
-    code = compile_pseudocode("""tmp = S0.u32 + S1.u32
-D0.u32 = tmp.u32""")
-    ctx = ExecContext(s0=100, s1=200)
-    ctx.run(code)
-    self.assertEqual(ctx.D0._val, 300)
-
-  def test_s_add_u32_pattern(self):
-    # Real pseudocode pattern from S_ADD_U32
-    code = compile_pseudocode("""tmp = 64'U(S0.u32) + 64'U(S1.u32)
-SCC = tmp >= 0x100000000ULL ? 1'1U : 1'0U
-D0.u32 = tmp.u32""")
-    # Test overflow case
-    ctx = ExecContext(s0=0xFFFFFFFF, s1=0x00000001)
-    ctx.run(code)
-    self.assertEqual(ctx.D0._val, 0)  # Wraps to 0
-    self.assertEqual(ctx.SCC._val, 1)  # Carry set
-
-  def test_s_add_u32_no_overflow(self):
-    code = compile_pseudocode("""tmp = 64'U(S0.u32) + 64'U(S1.u32)
-SCC = tmp >= 0x100000000ULL ? 1'1U : 1'0U
-D0.u32 = tmp.u32""")
-    ctx = ExecContext(s0=100, s1=200)
-    ctx.run(code)
-    self.assertEqual(ctx.D0._val, 300)
-    self.assertEqual(ctx.SCC._val, 0)  # No carry
-
-  def test_vcc_lane_read(self):
-    ctx = ExecContext(vcc=0b1010, lane=1)
-    # Lane 1 is set
-    self.assertEqual(ctx.VCC.u64[1], 1)
-    self.assertEqual(ctx.VCC.u64[2], 0)
-
-  def test_vcc_lane_write(self):
-    ctx = ExecContext(vcc=0, lane=0)
-    ctx.VCC.u64[3] = 1
-    ctx.VCC.u64[1] = 1
-    self.assertEqual(ctx.VCC._val, 0b1010)
-
-  def test_for_loop(self):
-    # CTZ pattern - find first set bit
-    code = compile_pseudocode("""tmp = -1
-for i in 0 : 31 do
-  if S0.u32[i] == 1 then
-    tmp = i
-D0.i32 = tmp""")
-    ctx = ExecContext(s0=0b1000)  # Bit 3 is set
-    ctx.run(code)
-    self.assertEqual(ctx.D0._val & MASK32, 3)
-
-  def test_result_dict(self):
-    ctx = ExecContext(s0=5, s1=3)
-    ctx.D0.u32 = 42
-    ctx.SCC._val = 1
-    result = ctx.result()
-    self.assertEqual(result['d0'], 42)
-    self.assertEqual(result['scc'], 1)
-
-class TestPseudocodeRegressions(unittest.TestCase):
-  """Regression tests for pseudocode instruction emulation bugs."""
-
-  def test_v_div_scale_f32_vcc_always_returned(self):
-    """V_DIV_SCALE_F32 must always return vcc_lane, even when VCC=0 (no scaling needed).
-    Bug: when VCC._val == vcc (both 0), vcc_lane wasn't returned, so VCC bits weren't written.
-    This caused division to produce wrong results for multiple lanes."""
-    # Normal case: 1.0 / 3.0, no scaling needed, VCC should be 0
-    s0 = 0x3f800000  # 1.0
-    s1 = 0x40400000  # 3.0
-    s2 = 0x3f800000  # 1.0 (numerator)
-    result = _VOP3SDOp_V_DIV_SCALE_F32(s0, s1, s2, 0, 0, 0, 0, 0xffffffff, 0, None, {})
-    # Must always have vcc_lane in result
-    self.assertIn('vcc_lane', result, "V_DIV_SCALE_F32 must always return vcc_lane")
-    self.assertEqual(result['vcc_lane'], 0, "vcc_lane should be 0 when no scaling needed")
-
-  def test_v_cmp_class_f32_detects_quiet_nan(self):
-    """V_CMP_CLASS_F32 must correctly identify quiet NaN vs signaling NaN.
-    Bug: isQuietNAN and isSignalNAN both used math.isnan which can't distinguish them."""
-    quiet_nan = 0x7fc00000   # quiet NaN: exponent=255, bit22=1
-    signal_nan = 0x7f800001  # signaling NaN: exponent=255, bit22=0
-    # Test quiet NaN detection (bit 1 in mask)
-    s1_quiet = 0b0000000010  # bit 1 = quiet NaN
-    result = _VOPCOp_V_CMP_CLASS_F32(quiet_nan, s1_quiet, 0, 0, 0, 0, 0, 0xffffffff, 0, None, {})
-    self.assertEqual(result['vcc_lane'], 1, "Should detect quiet NaN with quiet NaN mask")
-    # Test signaling NaN detection (bit 0 in mask)
-    s1_signal = 0b0000000001  # bit 0 = signaling NaN
-    result = _VOPCOp_V_CMP_CLASS_F32(signal_nan, s1_signal, 0, 0, 0, 0, 0, 0xffffffff, 0, None, {})
-    self.assertEqual(result['vcc_lane'], 1, "Should detect signaling NaN with signaling NaN mask")
-    # Test that quiet NaN doesn't match signaling NaN mask
-    result = _VOPCOp_V_CMP_CLASS_F32(quiet_nan, s1_signal, 0, 0, 0, 0, 0, 0xffffffff, 0, None, {})
-    self.assertEqual(result['vcc_lane'], 0, "Quiet NaN should not match signaling NaN mask")
-    # Test that signaling NaN doesn't match quiet NaN mask
-    result = _VOPCOp_V_CMP_CLASS_F32(signal_nan, s1_quiet, 0, 0, 0, 0, 0, 0xffffffff, 0, None, {})
-    self.assertEqual(result['vcc_lane'], 0, "Signaling NaN should not match quiet NaN mask")
-
-  def test_isnan_with_typed_view(self):
-    """_isnan must work with TypedView objects, not just Python floats.
-    Bug: _isnan checked isinstance(x, float) which returned False for TypedView."""
-    nan_reg = Reg(0x7fc00000)  # quiet NaN
-    normal_reg = Reg(0x3f800000)  # 1.0
-    inf_reg = Reg(0x7f800000)  # +inf
-    self.assertTrue(_isnan(nan_reg.f32), "_isnan should return True for NaN TypedView")
-    self.assertFalse(_isnan(normal_reg.f32), "_isnan should return False for normal TypedView")
-    self.assertFalse(_isnan(inf_reg.f32), "_isnan should return False for inf TypedView")
-
-if __name__ == '__main__':
-  unittest.main()

extra/assembly/amd/test/test_pdf_parser.py

@@ -1,153 +0,0 @@
-#!/usr/bin/env python3
-"""Test that PDF parser correctly extracts format fields."""
-import unittest, os
-from extra.assembly.amd.autogen.rdna3 import (
-  SOP1, SOP2, SOPK, SOPP, VOP1, VOP2, VOP3SD, VOPC, FLAT, VOPD,
-  SOP1Op, SOP2Op, VOP1Op, VOP3Op
-)
-
-# expected formats with key fields and whether they have ENCODING
-EXPECTED_FORMATS = {
-  'DPP16': (['SRC0', 'DPP_CTRL', 'BANK_MASK', 'ROW_MASK'], False),
-  'DPP8': (['SRC0', 'LANE_SEL0', 'LANE_SEL7'], False),
-  'DS': (['OP', 'ADDR', 'DATA0', 'DATA1', 'VDST'], True),
-  'EXP': (['EN', 'TARGET', 'VSRC0', 'VSRC1', 'VSRC2', 'VSRC3'], True),
-  'FLAT': (['OP', 'ADDR', 'DATA', 'SADDR', 'VDST', 'OFFSET'], True),
-  'LDSDIR': (['VDST', 'OP'], True),
-  'MIMG': (['OP', 'VADDR', 'VDATA', 'SRSRC', 'DMASK'], True),
-  'MTBUF': (['OP', 'VADDR', 'VDATA', 'SRSRC', 'FORMAT', 'SOFFSET'], True),
-  'MUBUF': (['OP', 'VADDR', 'VDATA', 'SRSRC', 'SOFFSET'], True),
-  'SMEM': (['OP', 'SBASE', 'SDATA', 'OFFSET', 'SOFFSET'], True),
-  'SOP1': (['OP', 'SDST', 'SSRC0'], True),
-  'SOP2': (['OP', 'SDST', 'SSRC0', 'SSRC1'], True),
-  'SOPC': (['OP', 'SSRC0', 'SSRC1'], True),
-  'SOPK': (['OP', 'SDST', 'SIMM16'], True),
-  'SOPP': (['OP', 'SIMM16'], True),
-  'VINTERP': (['OP', 'VDST', 'SRC0', 'SRC1', 'SRC2'], True),
-  'VOP1': (['OP', 'VDST', 'SRC0'], True),
-  'VOP2': (['OP', 'VDST', 'SRC0', 'VSRC1'], True),
-  'VOP3': (['OP', 'VDST', 'SRC0', 'SRC1', 'SRC2'], True),
-  'VOP3P': (['OP', 'VDST', 'SRC0', 'SRC1', 'SRC2'], True),
-  'VOP3SD': (['OP', 'VDST', 'SDST', 'SRC0', 'SRC1', 'SRC2'], True),
-  'VOPC': (['OP', 'SRC0', 'VSRC1'], True),
-  'VOPD': (['OPX', 'OPY', 'SRCX0', 'SRCY0', 'VDSTX', 'VDSTY'], True),
-}
-
-# Skip PDF parsing tests by default - only run with TEST_PDF_PARSER=1
-# These are slow (~5s) and only needed when regenerating autogen/
-@unittest.skipUnless(os.environ.get("TEST_PDF_PARSER"), "set TEST_PDF_PARSER=1 to run PDF parser tests")
-class TestPDFParserGenerate(unittest.TestCase):
-  """Test the PDF parser by running generate() and checking results."""
-
-  def test_pdf_parser(self):
-    """Single test that validates all PDF parser outputs."""
-    from extra.assembly.amd.dsl import generate
-    result = generate()
-
-    # test_all_formats_present
-    for fmt_name in EXPECTED_FORMATS:
-      self.assertIn(fmt_name, result["formats"], f"missing format {fmt_name}")
-
-    # test_format_count
-    self.assertEqual(len(result["formats"]), 23)
-
-    # test_no_duplicate_fields
-    for fmt_name, fields in result["formats"].items():
-      field_names = [f[0] for f in fields]
-      self.assertEqual(len(field_names), len(set(field_names)), f"{fmt_name} has duplicate fields: {field_names}")
-
-    # test_expected_fields
-    for fmt_name, (expected_fields, has_encoding) in EXPECTED_FORMATS.items():
-      fields = {f[0] for f in result["formats"].get(fmt_name, [])}
-      for field in expected_fields:
-        self.assertIn(field, fields, f"{fmt_name} missing {field}")
-      if has_encoding:
-        self.assertIn("ENCODING", fields, f"{fmt_name} should have ENCODING")
-      else:
-        self.assertNotIn("ENCODING", fields, f"{fmt_name} should not have ENCODING")
-
-    # test_vopd_no_dpp16_fields
-    vopd_fields = {f[0] for f in result["formats"].get("VOPD", [])}
-    for field in ['DPP_CTRL', 'BANK_MASK', 'ROW_MASK']:
-      self.assertNotIn(field, vopd_fields, f"VOPD should not have {field}")
-
-    # test_dpp16_no_vinterp_fields
-    dpp16_fields = {f[0] for f in result["formats"].get("DPP16", [])}
-    for field in ['VDST', 'WAITEXP']:
-      self.assertNotIn(field, dpp16_fields, f"DPP16 should not have {field}")
-
-    # test_sopp_no_smem_fields
-    sopp_fields = {f[0] for f in result["formats"].get("SOPP", [])}
-    for field in ['SBASE', 'SDATA']:
-      self.assertNotIn(field, sopp_fields, f"SOPP should not have {field}")
-
-class TestPDFParser(unittest.TestCase):
-  """Verify format classes have correct fields from PDF parsing."""
-
-  def test_sop2_fields(self):
-    """SOP2 should have op, sdst, ssrc0, ssrc1."""
-    for field in ['op', 'sdst', 'ssrc0', 'ssrc1']:
-      self.assertIn(field, SOP2._fields)
-    self.assertEqual(SOP2._fields['op'].hi, 29)
-    self.assertEqual(SOP2._fields['op'].lo, 23)
-
-  def test_sop1_fields(self):
-    """SOP1 should have op, sdst, ssrc0 with correct bit positions."""
-    for field in ['op', 'sdst', 'ssrc0']:
-      self.assertIn(field, SOP1._fields)
-    self.assertNotIn('simm16', SOP1._fields)
-    self.assertEqual(SOP1._fields['ssrc0'].hi, 7)
-    self.assertEqual(SOP1._fields['ssrc0'].lo, 0)
-    assert SOP1._encoding is not None
-    self.assertEqual(SOP1._encoding[0].hi, 31)
-    self.assertEqual(SOP1._encoding[1], 0b101111101)
-
-  def test_vop3sd_fields(self):
-    """VOP3SD should have all fields including src0/src1/src2 from page continuation."""
-    for field in ['op', 'vdst', 'sdst', 'src0', 'src1', 'src2']:
-      self.assertIn(field, VOP3SD._fields)
-    self.assertEqual(VOP3SD._fields['src0'].hi, 40)
-    self.assertEqual(VOP3SD._fields['src0'].lo, 32)
-    self.assertEqual(VOP3SD._size(), 8)
-
-  def test_flat_has_vdst(self):
-    """FLAT should have vdst field."""
-    self.assertIn('vdst', FLAT._fields)
-    self.assertEqual(FLAT._fields['vdst'].hi, 63)
-    self.assertEqual(FLAT._fields['vdst'].lo, 56)
-
-  def test_encoding_bits(self):
-    """Verify encoding bits are correct for major formats."""
-    tests = [
-      (SOP2, 31, 30, 0b10),
-      (SOPK, 31, 28, 0b1011),
-      (SOPP, 31, 23, 0b101111111),
-      (VOP1, 31, 25, 0b0111111),
-      (VOP2, 31, 31, 0b0),
-      (VOPC, 31, 25, 0b0111110),
-      (FLAT, 31, 26, 0b110111),
-    ]
-    for cls, hi, lo, val in tests:
-      assert cls._encoding is not None
-      self.assertEqual(cls._encoding[0].hi, hi, f"{cls.__name__} encoding hi")
-      self.assertEqual(cls._encoding[0].lo, lo, f"{cls.__name__} encoding lo")
-      self.assertEqual(cls._encoding[1], val, f"{cls.__name__} encoding val")
-
-  def test_opcode_enums_exist(self):
-    """Verify opcode enums are generated with expected counts."""
-    self.assertGreater(len(SOP1Op), 50)
-    self.assertGreater(len(SOP2Op), 50)
-    self.assertGreater(len(VOP1Op), 50)
-    self.assertGreater(len(VOP3Op), 200)
-
-  def test_vopd_no_duplicate_fields(self):
-    """VOPD should not have duplicate fields and should not include DPP16 fields."""
-    field_names = list(VOPD._fields.keys())
-    self.assertEqual(len(field_names), len(set(field_names)))
-    for field in ['srcx0', 'srcy0', 'opx', 'opy']:
-      self.assertIn(field, VOPD._fields)
-    for field in ['dpp_ctrl', 'bank_mask', 'row_mask']:
-      self.assertNotIn(field, VOPD._fields)
-
-if __name__ == "__main__":
-  unittest.main()

extra/assembly/amd/test/test_rdna3_asm.py

@@ -1,95 +0,0 @@
-#!/usr/bin/env python3
-import unittest, subprocess
-from extra.assembly.amd.autogen.rdna3 import *
-from extra.assembly.amd.test.helpers import get_llvm_mc
-
-def llvm_assemble(asm: str) -> bytes:
-  """Assemble using llvm-mc and return bytes."""
-  result = subprocess.run(
-    [get_llvm_mc(), "-triple=amdgcn", "-mcpu=gfx1100", "-show-encoding"],
-    input=asm, capture_output=True, text=True
-  )
-  out = b''
-  for line in result.stdout.split('\n'):
-    if 'encoding:' in line:
-      enc = line.split('encoding:')[1].strip()
-      enc = enc.strip('[]').replace('0x', '').replace(',', '')
-      out += bytes.fromhex(enc)
-  if not out: raise ValueError(f"no encoding found: {result.stdout} {result.stderr}")
-  return out
-
-class TestRDNA3Asm(unittest.TestCase):
-  def test_full_program(self):
-    """Test the full program from rdna3fun.py matches llvm-mc output."""
-    program = [
-      v_bfe_u32(v[1], v[0], 10, 10),
-      s_load_b128(s[4:7], s[0:1], NULL),
-      v_and_b32_e32(v[0], 0x3FF, v[0]),
-      s_mulk_i32(s[3], 0x87),
-      v_mad_u64_u32(v[1:2], NULL, s[2], 3, v[1:2]),
-      v_mul_u32_u24_e32(v[0], 45, v[0]),
-      v_ashrrev_i32_e32(v[2], 31, v[1]),
-      v_add3_u32(v[0], v[0], s[3], v[1]),
-      v_lshlrev_b64(v[2:3], 2, v[1:2]),
-      v_ashrrev_i32_e32(v[1], 31, v[0]),
-      v_lshlrev_b64(v[0:1], 2, v[0:1]),
-      s_waitcnt(0xfc07),  # lgkmcnt(0)
-      v_add_co_u32(v[2], VCC_LO, s[6], v[2]),
-      v_add_co_ci_u32_e32(v[3], s[7], v[3]),
-      v_add_co_u32(v[0], VCC_LO, s[4], v[0]),
-      global_load_b32(vdst=v[2], addr=v[2], saddr=OFF),
-      v_add_co_ci_u32_e32(v[1], s[5], v[1]),
-      s_waitcnt(0x03f7),  # vmcnt(0)
-      global_store_b32(addr=v[0], data=v[2], saddr=OFF),
-      s_endpgm(),
-    ]
-
-    asm = """
-v_bfe_u32 v1, v0, 10, 10
-s_load_b128 s[4:7], s[0:1], null
-v_and_b32_e32 v0, 0x3FF, v0
-s_mulk_i32 s3, 0x87
-v_mad_u64_u32 v[1:2], null, s2, 3, v[1:2]
-v_mul_u32_u24_e32 v0, 45, v0
-v_ashrrev_i32_e32 v2, 31, v1
-v_add3_u32 v0, v0, s3, v1
-v_lshlrev_b64 v[2:3], 2, v[1:2]
-v_ashrrev_i32_e32 v1, 31, v0
-v_lshlrev_b64 v[0:1], 2, v[0:1]
-s_waitcnt lgkmcnt(0)
-v_add_co_u32 v2, vcc_lo, s6, v2
-v_add_co_ci_u32_e32 v3, vcc_lo, s7, v3, vcc_lo
-v_add_co_u32 v0, vcc_lo, s4, v0
-global_load_b32 v2, v[2:3], off
-v_add_co_ci_u32_e32 v1, vcc_lo, s5, v1, vcc_lo
-s_waitcnt vmcnt(0)
-global_store_b32 v[0:1], v2, off
-s_endpgm
-"""
-    expected = llvm_assemble(asm)
-    for inst,rt in zip(program, asm.strip().split("\n")): print(f"{inst.disasm():50s} {rt}")
-    actual = b''.join(inst.to_bytes() for inst in program)
-    self.assertEqual(actual, expected)
-
-  def test_sop2_s_add_u32(self):
-    inst = SOP2(SOP2Op.S_ADD_U32, s[3], s[0], s[1])
-    expected = llvm_assemble("s_add_u32 s3, s0, s1")
-    self.assertEqual(inst.to_bytes(), expected)
-
-  def test_vop2_v_and_b32_inline_const(self):
-    inst = v_and_b32_e32(v[0], 10, v[0])
-    expected = llvm_assemble("v_and_b32_e32 v0, 10, v0")
-    self.assertEqual(inst.to_bytes(), expected)
-
-  def test_sopp_s_endpgm(self):
-    inst = s_endpgm()
-    expected = llvm_assemble("s_endpgm")
-    self.assertEqual(inst.to_bytes(), expected)
-
-  def test_sop1_s_mov_b32(self):
-    inst = s_mov_b32(s[0], s[1])
-    expected = llvm_assemble("s_mov_b32 s0, s1")
-    self.assertEqual(inst.to_bytes(), expected)
-
-if __name__ == "__main__":
-  unittest.main()

extra/bench_log.py

@@ -34,7 +34,8 @@ class WallTimeEvent:
     self.start = time.monotonic()
     return self
   def __exit__(self, *_):
-    _events[self.event]["wall"].append(time.monotonic() - self.start)
+    self.time = time.monotonic() - self.start
+    _events[self.event]["wall"].append(self.time)
     return False
 
 class KernelTimeEvent:

extra/export_model.py

@@ -67,12 +67,11 @@ def export_model_clang(functions:Dict[str,str], statements:Dict[str,Tuple[str,in
   forward_args = ",".join(f"{dtype}{'*' if name not in symbolic_vars.values() else ''} {name}" for name,dtype,_ in (outputs+inputs if wasm else inputs+outputs))
 
   if not wasm:
-    thread_id = 0 # NOTE: export does not support threading, thread_id is always 0
     for name,cl in bufs_to_save.items():
       weight = ''.join(["\\x%02X"%x for x in bytes(to_mv(cl._buf.va_addr, cl._buf.size))])
       cprog.append(f"unsigned char {name}_data[] = \"{weight}\";")
     cprog += [f"{dtype_map[dtype]} {name}[{len}];" if name not in bufs_to_save else f"{dtype_map[dtype]} *{name} = ({dtype_map[dtype]} *){name}_data;" for name,(len,dtype,_key) in bufs.items() if name not in input_names+output_names]
-    cprog += [f"void net({forward_args}) {{"] + [f"{name}({', '.join(args)}, {thread_id});" for (name, args, _global_size, _local_size) in statements] + ["}"]
+    cprog += [f"void net({forward_args}) {{"] + [f"{name}({', '.join(args)});" for (name, args, _global_size, _local_size) in statements] + ["}"]
     return '\n'.join(headers + cprog)
   else:
     if bufs_to_save:

extra/fp8/fp8_linear.py

@@ -0,0 +1,101 @@
+from typing import Callable, Any
+from tinygrad import Tensor, dtypes, nn, UOp
+from tinygrad.uop.ops import KernelInfo, AxisType, Ops
+
+def quantize_to_fp8(x: Tensor, dtype=dtypes.fp8e4m3):
+  fp8_min = -448.0 if dtype == dtypes.fp8e4m3 else -57344.0
+  fp8_max = 448.0 if dtype == dtypes.fp8e4m3 else 57344.0
+  x_abs_max = x.abs().max().detach()
+  scale = fp8_max / (x_abs_max + 1e-8)
+  x_scaled = x * scale
+  x_det = x_scaled.detach()
+  x_clamped = x_det.clamp(fp8_min, fp8_max)
+  x_clamped_ste = x_scaled + (x_clamped - x_det)
+  res = x_clamped_ste.cast(dtype)
+  return res, scale.float().reciprocal()
+
+def custom_matmul(output: UOp, inp: UOp, weight: UOp) -> UOp:
+  SEQ = inp.shape[1]
+  OUT = weight.shape[0]
+  IN = weight.shape[-1]
+  seq_idx = UOp.range(SEQ, 2, AxisType.LOOP)
+  out_idx = UOp.range(OUT, 3, AxisType.LOOP)
+  batch_idx = UOp.range(output.size//SEQ//OUT, 1, AxisType.LOOP)
+  reduce_idx = UOp.range(IN, 0, AxisType.REDUCE)
+  product = (inp.index((seq_idx*IN+reduce_idx+batch_idx*IN*SEQ)) * weight.index((out_idx*IN+reduce_idx))).cast(dtypes.float)
+  reduced = product.reduce(reduce_idx, arg=Ops.ADD)
+  store_op = output.index((seq_idx*OUT+out_idx+batch_idx*OUT*SEQ), ptr=True).store(reduced).end(batch_idx, seq_idx, out_idx)
+  return store_op.sink(arg=KernelInfo(name=f"fp8_matmul_{inp.shape}x{weight.shape}"))
+
+def custom_matmul_backward(gradient: UOp, kernel: UOp) -> tuple[UOp, UOp]:
+  _, input_uop, weight_uop = kernel.src[1:]
+  input_tensor = Tensor(input_uop, device=input_uop.device)
+  grad_tensor = Tensor(gradient, device=gradient.device)
+  weight_tensor = Tensor(weight_uop, device=weight_uop.device)
+  grad_quantized, scale = quantize_to_fp8(grad_tensor)
+  scale_scalar = scale.reshape(())
+  grad_weight = Tensor.einsum("bso,bsi->oi", grad_quantized, input_tensor, dtype=dtypes.float)
+  grad_weight = grad_weight * scale_scalar
+  grad_2d = grad_quantized.reshape(grad_tensor.shape[0] * grad_tensor.shape[1], grad_tensor.shape[-1])
+  grad_input = (grad_2d.dot(weight_tensor, dtype=dtypes.float)).contiguous().reshape(input_tensor.shape) * scale
+  return (None, grad_input.uop, grad_weight.uop)
+
+class FP8Linear:
+  def __init__(self, in_features:int, out_features:int, bias:bool=True):
+    self.weight = Tensor.empty(out_features, in_features, dtype=dtypes.float32)
+    self.bias = Tensor.empty(out_features, dtype=dtypes.float32) if bias else None
+
+  def __call__(self, x: Tensor) -> Tensor:
+    original_ndim = len(x.shape)
+    if original_ndim == 2: x = x.reshape(x.shape[0], 1, x.shape[1])
+    batch, seq, _ = x.shape
+    w_fp8, w_scale = quantize_to_fp8(self.weight)
+    x_fp8, x_scale = quantize_to_fp8(x)
+    GPUS = self.weight.device
+    if isinstance(GPUS, tuple) and len(GPUS) > 1:
+      y = Tensor(Tensor.empty((batch//len(GPUS), seq, self.weight.shape[0]), dtype=dtypes.float, device=GPUS).uop.multi(0), device=GPUS)
+    else:
+      y = Tensor.empty((batch, seq, self.weight.shape[0]), dtype=dtypes.float)
+    y = Tensor.custom_kernel(y, x_fp8, w_fp8, fxn=custom_matmul, grad_fxn=custom_matmul_backward)[0]
+    y = y * w_scale * x_scale
+    if self.bias is not None: y = y + self.bias
+    if original_ndim == 2: y = y.reshape(batch, self.weight.shape[0])
+    return y.cast(x.dtype)
+
+def _replace_linear(layer: nn.Linear):
+  fp8_linear = FP8Linear(layer.weight.shape[1], layer.weight.shape[0], layer.bias is not None)
+  fp8_linear.weight = layer.weight
+  if layer.bias is not None: fp8_linear.bias = layer.bias
+  return fp8_linear
+
+def _swap_linear_with_fp8(model, module_filter_fn:Callable[[Any, str],bool]|None=None, fqn:str="", parent:Any|None=None,
+                          attr_name:str="", visited:set|None=None):
+  if visited is None: visited = set()
+  if id(model) in visited: return
+  visited.add(id(model))
+  if isinstance(model, (str, int, float, bool, type(None), Tensor, UOp)): return
+  elif isinstance(model, nn.Linear):
+    if module_filter_fn is not None and not module_filter_fn(model, fqn): return
+    fp8_linear = _replace_linear(model)
+    if parent is not None and attr_name:
+      setattr(parent, attr_name, fp8_linear)
+  elif isinstance(model, list):
+    for i, item in enumerate(model):
+      child_fqn = f"{fqn}.{i}" if fqn else str(i)
+      if isinstance(item, nn.Linear) and (module_filter_fn is None or module_filter_fn(item, child_fqn)): model[i] = _replace_linear(item)
+      else: _swap_linear_with_fp8(item, module_filter_fn, child_fqn, None, "", visited)
+  elif isinstance(model, dict):
+    for key, item in list(model.items()):
+      child_fqn = f"{fqn}.{key}" if fqn else str(key)
+      if isinstance(item, nn.Linear) and (module_filter_fn is None or module_filter_fn(item, child_fqn)): model[key] = _replace_linear(item)
+      else: _swap_linear_with_fp8(item, module_filter_fn, child_fqn, None, "", visited)
+  elif hasattr(model, "__dict__"):
+    for attr_key in list(vars(model).keys()):
+      try: attr = getattr(model, attr_key)
+      except Exception: continue
+      child_fqn = f"{fqn}.{attr_key}" if fqn else attr_key
+      _swap_linear_with_fp8(attr, module_filter_fn, child_fqn, model, attr_key, visited)
+
+def convert_to_float8_training(model, module_filter_fn:Callable[[Any,str],bool]|None=None):
+  _swap_linear_with_fp8(model, module_filter_fn, "", None, "")
+  return model

extra/gemm/amd_asm_matmul.py

@@ -0,0 +1,496 @@
+# RDNA3 128x128 tiled GEMM kernel - DSL version
+# Computes C = A @ B for NxN float32 matrices using 128x128 tiles
+#
+# Architecture: RDNA3 (gfx1100)
+# Tile size: 128x128 (each workgroup computes one tile of C)
+# Workgroup: 128 threads (arranged as 32x4 for coalesced memory access)
+# Inner loop: 8 iterations per K-block, processing 8 columns of A and 8 rows of B
+#
+# Accumulators: 128 vgprs (v[2-129])
+
+import numpy as np
+from tinygrad import Tensor, Device, Context, GlobalCounters
+from tinygrad.uop.ops import UOp, Ops, KernelInfo
+from tinygrad.helpers import getenv, colored
+from tinygrad.dtype import dtypes, AddrSpace
+from tinygrad.engine.realize import Estimates
+from tinygrad.renderer.amd.dsl import s, v, VCC_LO, NULL
+from tinygrad.runtime.autogen.amd.rdna3.ins import *
+
+# =============================================================================
+# Kernel constants
+# =============================================================================
+LDS_SIZE = 8320       # Local data share size in bytes
+LDS_A_STRIDE = 0x210  # LDS stride for A tile (528 bytes)
+LDS_B_STRIDE = 0x200  # LDS stride for B tile (512 bytes)
+LDS_BASE_OFFSET = 0x1080  # Base LDS offset for tiles
+ADDR_MASK = 0x3fffff80    # Address alignment mask
+
+# =============================================================================
+# Named register assignments (VGPRs)
+# =============================================================================
+V_LANE_ID = 0             # lane_id set on startup
+# Use tile gaps (v146-159) for named regs to minimize max VGPR
+V_LANE_ID_MOD8 = 146      # lane_id & 7
+V_LANE_MOD8_X4 = 147      # (lane_id & 7) << 2
+V_LANE_DIV8_X4 = 150      # ((lane_id >> 3) & 3) << 2
+V_LDS_B_BASE = 151        # LDS B-tile base address for inner loop
+V_LDS_A_BASE = 154        # LDS A-tile base address for inner loop
+V_GLOBAL_A_ADDR = 155     # global memory A prefetch address
+V_GLOBAL_B_ADDR = 158     # global memory B prefetch address
+V_LDS_A_ADDR = 159        # single base register for A stores
+V_LDS_B_ADDR = 162        # single base register for B stores
+
+# LDS tile register destinations - SEPARATE from DATA to avoid overlap
+# A on banks 2-3, B on banks 0-1 to avoid bank conflicts in VOPD
+V_A_TILE_REGS = [130, 134, 138, 142]  # A tile: banks 2,2,2,2 (130%4=2, etc.)
+V_B_TILE_REGS = [132, 136, 140, 144, 148, 152, 156, 160]  # B tile: banks 0,0,0,0,0,0,0,0
+
+# =============================================================================
+# Named register assignments (SGPRs)
+# =============================================================================
+S_OUT_PTR = (0, 1)        # output C matrix base pointer
+S_WORKGROUP_X = 2         # workgroup_id_x (system SGPR, follows user SGPRs)
+S_WORKGROUP_Y = 3         # workgroup_id_y (system SGPR)
+S_DIM_N = 4               # matrix dimension N
+S_LOOP_BOUND = 7          # K-8 (loop termination bound)
+S_LOOP_CTR = 12           # loop counter (increments by 8)
+S_PREFETCH_FLAG = 13      # prefetch condition flag / row stride in epilogue
+S_TILE_X = 14             # workgroup_x << 7
+S_TILE_Y = 15             # workgroup_y << 7
+# Kernarg load destinations
+S_KERNARG_A = (20, 21)    # A pointer from kernarg
+S_KERNARG_B = (22, 23)    # B pointer from kernarg
+# Prefetch base pointers (8 pairs each, B: N*4 bytes apart, A: N*64 bytes apart)
+S_PREFETCH_B = 24         # s[24:39] - 8 B tile pointers
+S_PREFETCH_A = 40         # s[40:55] - 8 A tile pointers
+
+# =============================================================================
+# Data tables
+# =============================================================================
+
+# Accumulator grid: ACC_GRID[a_idx][b_idx] = vgpr for C[a,b]
+# a_idx: which A value (0-7), b_idx: which B value (0-15)
+# Scattered due to VOPD bank constraints (vdst_x % 4 != vdst_y % 4)
+# Range is from v2 - v129
+ACC_GRID = [
+  [  5,  3,  9,  8,   37, 35, 41, 40,   69, 67, 73, 72,  101, 99,105,104],  # a0
+  [  4,  2,  7,  6,   36, 34, 39, 38,   68, 66, 71, 70,  100, 98,103,102],  # a1
+  [ 17, 16, 13, 11,   49, 48, 45, 43,   81, 80, 77, 75,  113,112,109,107],  # a2
+  [ 15, 14, 12, 10,   47, 46, 44, 42,   79, 78, 76, 74,  111,110,108,106],  # a3
+  [ 21, 19, 25, 24,   53, 51, 57, 56,   85, 83, 89, 88,  117,115,121,120],  # a4
+  [ 20, 18, 23, 22,   52, 50, 55, 54,   84, 82, 87, 86,  116,114,123,122],  # a5
+  [125,128, 29, 27,   33, 32, 61, 59,   65, 64, 93, 91,   97, 96,129,127],  # a6
+  [119,118, 28, 26,   31, 30, 60, 58,   63, 62, 92, 90,   95, 94,124,126],  # a7
+]
+
+# Optimized (a_pair, b_pair) iteration order for better GPU scheduling
+# Interleaves A and B pairs to maximize instruction-level parallelism
+FMAC_PAIR_ORDER = [
+  (0,0),(0,1),(1,1),(1,0), (2,0),(2,1),(3,1),(3,2), (0,2),(0,3),(1,3),(1,2), (2,2),(2,3),(3,3),(3,4),
+  (0,4),(0,5),(1,5),(1,4), (2,4),(2,5),(3,5),(3,6), (0,6),(0,7),(1,7),(1,6), (2,6),(2,7),(3,7),(3,0),
+]
+
+def derive_fmac_pattern(acc_grid, a_tile_regs=None, b_tile_regs=None):
+  """Generate 64 dual FMAC ops from accumulator grid with optimized iteration order."""
+  pattern = []
+  for idx, (a_pair, b_pair) in enumerate(FMAC_PAIR_ORDER):
+    a_even, a_odd = a_pair * 2, a_pair * 2 + 1
+    b_even, b_odd = b_pair * 2, b_pair * 2 + 1
+    a_base, b_base = a_tile_regs[a_pair], b_tile_regs[b_pair]
+    # Op 1: normal order -> C[a_even, b_even] + C[a_odd, b_odd]
+    pattern.append((acc_grid[a_even][b_even], acc_grid[a_odd][b_odd],
+                   a_base, b_base, a_base+1, b_base+1))
+    # Op 2: alternate swapping A vs B to vary register banks
+    if idx % 2 == 0:  # swap B
+      pattern.append((acc_grid[a_even][b_odd], acc_grid[a_odd][b_even],
+                     a_base, b_base+1, a_base+1, b_base))
+    else:  # swap A
+      pattern.append((acc_grid[a_odd][b_even], acc_grid[a_even][b_odd],
+                     a_base+1, b_base, a_base, b_base+1))
+  return pattern
+
+# Derived: 64 dual FMAC operations
+FMAC_PATTERN = derive_fmac_pattern(ACC_GRID, V_A_TILE_REGS, V_B_TILE_REGS)
+
+def derive_permute_swaps(acc_grid, out_regs):
+  """Derive swap sequence to permute accumulators from FMAC layout to output order.
+
+  After FMAC loop: acc_grid[a][b] holds C[a,b]
+  Output order: for row_half in 0,1; col_group in 0-3; row_in_group in 0-3; b_off in 0-3
+    -> need C[row_half*4 + row_in_group, col_group*4 + b_off] in specified reg order
+  """
+  def target_ab(i):
+    row_half, col_group = i // 64, (i // 16) % 4
+    row_in_group, b_off = (i // 4) % 4, i % 4
+    return (row_half * 4 + row_in_group, col_group * 4 + b_off)
+
+  reg_contents = {acc_grid[a][b]: (a, b) for a in range(8) for b in range(16)}
+  ab_location = {ab: r for r, ab in reg_contents.items()}
+
+  swaps = []
+  for i in range(128):
+    target_reg, needed_ab = out_regs[i], target_ab(i)
+    current_reg = ab_location[needed_ab]
+    if current_reg != target_reg:
+      swaps.append((current_reg, target_reg))
+      ab_at_target = reg_contents.get(target_reg)
+      reg_contents[target_reg], ab_location[needed_ab] = needed_ab, target_reg
+      if ab_at_target is not None:
+        reg_contents[current_reg], ab_location[ab_at_target] = ab_at_target, current_reg
+  return swaps
+
+# Derived: swap sequence to arrange accumulators for output
+# Each group of 4 registers is ascending for direct global_store_b128
+OUT_REGS = [r for i in range(32) for r in range(126 - i*4, 130 - i*4)]
+PERMUTE_SWAPS = derive_permute_swaps(ACC_GRID, OUT_REGS)
+
+# =============================================================================
+# LDS tile staging registers
+# =============================================================================
+# DATA regs receive contiguous global prefetch, then write to LDS
+# TILE regs receive scattered LDS loads (ds_load_b64 pairs), then feed FMACs
+# Contiguous layout with mod4=[3,0,1,2,3,0,1,2] for bank conflict avoidance
+V_LDS_A_DATA = [163, 164, 165, 166, 167, 168, 169, 170]
+V_LDS_B_DATA = [171, 172, 173, 174, 175, 176, 177, 178]
+
+# Initial tile prefetch: (vdst, saddr_lo) - load into A data regs using B prefetch pointers (s[24:31])
+INIT_PREFETCH = [(V_LDS_A_DATA[i], S_PREFETCH_B+2*i) for i in range(4)]
+
+# Global memory prefetch schedule: (vdst1, vdst2, addr_vreg, saddr_lo1, saddr_lo2)
+# First 2 pairs from B prefetch pointers (s[32:39]), next 4 pairs from A prefetch pointers (s[40:55])
+PREFETCH_LOADS = [(V_LDS_A_DATA[4+2*i], V_LDS_A_DATA[4+2*i+1], V_GLOBAL_B_ADDR, S_PREFETCH_B+8+4*i, S_PREFETCH_B+10+4*i) for i in range(2)] + \
+                 [(V_LDS_B_DATA[2*(i-2)], V_LDS_B_DATA[2*(i-2)+1], V_GLOBAL_A_ADDR, S_PREFETCH_A+4*(i-2), S_PREFETCH_A+2+4*(i-2)) for i in range(2, 6)]
+
+# =============================================================================
+# Kernel class
+# =============================================================================
+
+class Kernel:
+  def __init__(self, arch='gfx1100'): self.instructions, self.labels, self.pos, self.arch = [], {}, 0, arch
+  def label(self, name): self.labels[name] = self.pos
+
+  def emit(self, inst, target=None):
+    self.instructions.append(inst)
+    inst._target, inst._pos = target, self.pos
+    self.pos += inst.size()
+    return inst
+
+  def waitcnt(self, lgkm=None, vm=None):
+    """Wait for memory operations. lgkm=N waits until N lgkm ops remain, vm=N waits until N vmem ops remain."""
+    vmcnt, lgkmcnt, expcnt = vm if vm is not None else 63, lgkm if lgkm is not None else 63, 7
+    waitcnt = (expcnt & 0x7) | ((lgkmcnt & 0x3f) << 4) | ((vmcnt & 0x3f) << 10)
+    self.emit(s_waitcnt(simm16=waitcnt))
+
+  def finalize(self):
+    """Patch branch offsets and return the finalized instruction list."""
+    for inst in self.instructions:
+      if inst._target is None: continue
+      offset_dwords = (self.labels[inst._target] - inst._pos - inst.size()) // 4
+      if not -32768 <= offset_dwords <= 32767: raise ValueError(f"branch to '{inst._target}' offset {offset_dwords} exceeds simm16 range")
+      inst.simm16 = offset_dwords
+    return self.instructions
+
+
+# =============================================================================
+# Kernel builder
+# =============================================================================
+
+def build_kernel(N, arch='gfx1100'):
+  assert N % 128 == 0, f"N must be a multiple of 128 (tile size), got {N}"
+  assert N >= 256, f"N must be >= 256 (prefetch pipeline requires at least 2 K-blocks), got {N}"
+  k = Kernel(arch)
+
+  # ===========================================================================
+  # PROLOGUE: Load kernel arguments, compute tile coordinates and addresses
+  # ===========================================================================
+  k.emit(s_load_b128(sdata=s[S_KERNARG_A[0]:S_KERNARG_B[1]], sbase=s[0:1], offset=0x0, soffset=NULL))
+  k.emit(s_load_b64(sdata=s[S_OUT_PTR[0]:S_OUT_PTR[1]], sbase=s[0:1], offset=0x10, soffset=NULL))
+  k.emit(s_mov_b32(s[S_DIM_N], N))
+  k.emit(s_mov_b32(s[S_LOOP_CTR], 0))  # used by LDS swizzle, always 0 for valid workgroups
+  k.emit(s_lshl_b32(s[S_TILE_X], s[S_WORKGROUP_X], 7))
+  k.emit(s_lshl_b32(s[S_TILE_Y], s[S_WORKGROUP_Y], 7))
+
+  # Lane-derived values
+  k.emit(v_and_b32_e32(v[V_LANE_ID_MOD8], 7, v[V_LANE_ID]))
+  k.emit(v_lshrrev_b32_e32(v[4], 3, v[V_LANE_ID]))
+  k.emit(v_or_b32_e32(v[1], s[S_TILE_X], v[V_LANE_ID]))
+  k.emit(v_or_b32_e32(v[22], s[S_TILE_Y], v[4]))
+  k.emit(v_lshlrev_b32_e32(v[V_LANE_MOD8_X4], 2, v[V_LANE_ID_MOD8]))
+  k.waitcnt(lgkm=0)
+
+  # Compute 8 A and B matrix tile base pointers for prefetch
+  k.emit(s_mov_b64(s[S_PREFETCH_B:S_PREFETCH_B+1], s[S_KERNARG_B[0]:S_KERNARG_B[1]]))  # B[0]: no offset
+  for i in range(1, 8):  # B: each pointer 1 row of B apart (N*4 bytes)
+    k.emit(s_add_u32(s[S_PREFETCH_B+i*2], s[S_KERNARG_B[0]], i * N * 4))
+    k.emit(s_addc_u32(s[S_PREFETCH_B+i*2+1], s[S_KERNARG_B[1]], 0))
+  k.emit(s_mov_b64(s[S_PREFETCH_A:S_PREFETCH_A+1], s[S_KERNARG_A[0]:S_KERNARG_A[1]]))  # A[0]: no offset
+  for i in range(1, 8):  # A: each pointer 16 rows of A apart (16*N*4 bytes)
+    k.emit(s_add_u32(s[S_PREFETCH_A+i*2], s[S_KERNARG_A[0]], i * N * 64))
+    k.emit(s_addc_u32(s[S_PREFETCH_A+i*2+1], s[S_KERNARG_A[1]], 0))
+
+  # Global prefetch addresses: B = (tile_x + lane_id) * 4, A = (tile_y*N + (lane_id/8)*N + lane_id%8) * 4
+  k.emit(v_add_nc_u32_e32(v[V_GLOBAL_B_ADDR], s[S_TILE_X], v[V_LANE_ID]))
+  k.emit(v_lshlrev_b32_e32(v[V_GLOBAL_B_ADDR], 2, v[V_GLOBAL_B_ADDR]))
+  k.emit(s_mul_i32(s[19], s[S_TILE_Y], N))
+  k.emit(v_mul_lo_u32(v[V_GLOBAL_A_ADDR], v[4], N))                          # (lane_id/8)*N
+  k.emit(v_add_nc_u32_e32(v[V_GLOBAL_A_ADDR], v[V_LANE_ID_MOD8], v[V_GLOBAL_A_ADDR]))  # + lane_id%8
+  k.emit(v_add_nc_u32_e32(v[V_GLOBAL_A_ADDR], s[19], v[V_GLOBAL_A_ADDR]))
+  k.emit(v_lshlrev_b32_e32(v[V_GLOBAL_A_ADDR], 2, v[V_GLOBAL_A_ADDR]))
+
+  # Do initial loads
+  for vdst, saddr_lo in INIT_PREFETCH:
+    k.emit(global_load_b32(vdst=v[vdst], addr=v[V_GLOBAL_B_ADDR], saddr=s[saddr_lo:saddr_lo+1]))
+  for iter in range(6):
+    vdst1, vdst2, addr, slo1, slo2 = PREFETCH_LOADS[iter]
+    k.emit(global_load_b32(vdst=v[vdst1], addr=v[addr], saddr=s[slo1:slo1+1]))
+    k.emit(global_load_b32(vdst=v[vdst2], addr=v[addr], saddr=s[slo2:slo2+1]))
+
+  # ===========================================================================
+  # LDS store address computation (bank-conflict-avoiding swizzle)
+  # ===========================================================================
+  # This section computes LDS store addresses with a swizzle pattern to avoid bank conflicts.
+  # The swizzle ensures that threads in the same wavefront write to different LDS banks.
+  # Formula: swizzled_addr = base + (lane_id & 7) * LDS_A_STRIDE + swizzle_offset
+  # where swizzle_offset depends on (lane_id >> 3) to distribute across banks.
+  k.emit(v_add_nc_u32_e32(v[9], s[S_LOOP_CTR], v[22]))  # row 0 base
+  k.emit(v_and_b32_e32(v[9], ADDR_MASK, v[9]))
+  k.emit(v_sub_nc_u32_e32(v[9], v[22], v[9]))  # row 0 swizzle offset
+  k.emit(v_lshlrev_b32_e32(v[9], 2, v[9]))  # * 4
+  k.emit(v_mad_u32_u24(v[V_LDS_B_ADDR], LDS_A_STRIDE, v[V_LANE_ID_MOD8], v[9]))
+
+  # For V_LDS_A_BASE and epilogue
+  k.emit(v_bfe_u32(v[2], v[V_LANE_ID], 3, 2))  # v[2] = (lane_id >> 3) & 3
+  k.emit(v_lshlrev_b32_e32(v[V_LANE_DIV8_X4], 2, v[2]))
+
+  # Compute LDS load/store base addresses for inner loop
+  k.emit(v_lshlrev_b32_e32(v[2], 4, v[2]))
+  k.emit(v_and_b32_e32(v[3], 0x7F, v[1]))  # simplified from 3 lines
+  k.emit(v_lshl_or_b32(v[V_LDS_B_BASE], v[V_LANE_ID_MOD8], 4, LDS_BASE_OFFSET))
+  k.emit(v_lshl_add_u32(v[V_LDS_A_ADDR], v[3], 2, LDS_BASE_OFFSET))
+  k.emit(v_lshlrev_b32_e32(v[3], 2, v[V_LANE_ID]))
+  k.emit(v_and_or_b32(v[V_LDS_A_BASE], 0x180, v[3], v[2]))
+
+  # Do initial stores
+  k.waitcnt(vm=0)
+  for i in range(4):  # A tile: 8 values via 4 stride64 stores
+    k.emit(ds_store_2addr_stride64_b32(addr=v[V_LDS_A_ADDR], data0=v[V_LDS_A_DATA[i*2]], data1=v[V_LDS_A_DATA[i*2+1]], offset0=i*4, offset1=i*4+2))
+  for i in range(8):  # B tile: 8 values via 8 scalar stores with 64-byte spacing
+    offset = i * 64
+    k.emit(ds_store_b32(addr=v[V_LDS_B_ADDR], data0=v[V_LDS_B_DATA[i]], offset0=offset & 0xFF, offset1=offset >> 8))
+
+  # Zero all 128 accumulators using VOPD dual moves (64 instructions instead of 128)
+  for i in range(0, len(OUT_REGS), 2):
+    k.emit(VOPD(VOPDOp.V_DUAL_MOV_B32, VOPDOp.V_DUAL_MOV_B32, vdstx=v[OUT_REGS[i]], vdsty=v[OUT_REGS[i+1]], srcx0=0, srcy0=0))
+  k.emit(s_add_i32(s[S_LOOP_BOUND], s[S_DIM_N], -8))
+
+  # S_LOOP_CTR is already 0 from prologue initialization
+  k.emit(s_branch(), target='LOOP_ENTRY')
+
+  # ===========================================================================
+  # MAIN GEMM LOOP
+  # ===========================================================================
+
+  NO_DS, NO_GLOBAL = getenv("NO_DS", 0), getenv("NO_GLOBAL", 0)
+
+  k.label('LOOP_INC')
+  k.emit(s_add_i32(s[S_LOOP_CTR], s[S_LOOP_CTR], 8))
+  k.emit(s_cmp_ge_i32(s[S_LOOP_CTR], s[S_DIM_N]))
+  k.emit(s_cbranch_scc1(), target='EPILOGUE')
+
+  k.label('LOOP_ENTRY')
+  k.emit(s_cmp_lt_i32(s[S_LOOP_CTR], s[S_LOOP_BOUND]))
+  k.emit(s_cselect_b32(s[S_PREFETCH_FLAG], -1, 0))  # s_cselect doesn't modify SCC
+  k.emit(s_cbranch_scc0(), target='SKIP_PREFETCH')  # branch if loop_ctr >= loop_bound
+
+  if not NO_GLOBAL:
+    # Advance prefetch pointers (VGPR)
+    #k.emit(v_add_nc_u32_e32(v[V_GLOBAL_B_ADDR], N * 32, v[V_GLOBAL_B_ADDR]))
+    #k.emit(v_add_nc_u32_e32(v[V_GLOBAL_A_ADDR], 0x20, v[V_GLOBAL_A_ADDR]))
+
+    # Advance prefetch pointers (64-bit adds): B advances 8 rows (8*N*4 bytes), A advances 8 cols (8*4 bytes)
+    k.emit(s_clause(simm16=31))
+    for i in range(8):
+      k.emit(s_add_u32(s[S_PREFETCH_B+i*2], s[S_PREFETCH_B+i*2], N * 32))
+      k.emit(s_addc_u32(s[S_PREFETCH_B+i*2+1], s[S_PREFETCH_B+i*2+1], 0))
+    for i in range(8):
+      k.emit(s_add_u32(s[S_PREFETCH_A+i*2], s[S_PREFETCH_A+i*2], 0x20))
+      k.emit(s_addc_u32(s[S_PREFETCH_A+i*2+1], s[S_PREFETCH_A+i*2+1], 0))
+
+    # do the fetch
+    for vdst, saddr_lo in INIT_PREFETCH:
+      k.emit(global_load_b32(vdst=v[vdst], addr=v[V_GLOBAL_B_ADDR], saddr=s[saddr_lo:saddr_lo+1]))
+
+  k.label('SKIP_PREFETCH')
+
+  # wait for local stores to finish (either initial or loop)
+  # then sync the warp so it's safe to load local
+  k.waitcnt(lgkm=0)
+  k.emit(s_barrier())
+
+  # 8 inner loop iterations
+  for iter in range(8):
+    # Load A tile (4 pairs) and B tile (8 pairs) from LDS
+    if not NO_DS:
+      k.emit(s_clause(simm16=len(V_A_TILE_REGS) + len(V_B_TILE_REGS) - 1))  # 12 loads total: 4 A + 8 B
+      # A tile: 4 ds_load_b64
+      for i, vdst in enumerate(V_A_TILE_REGS):
+        a_off = (i & 1) * 8 + (i >> 1) * 64 + iter * LDS_A_STRIDE
+        k.emit(ds_load_b64(vdst=v[vdst:vdst+1], addr=v[V_LDS_A_BASE], offset0=a_off & 0xFF, offset1=a_off >> 8))
+      # B tile: 8 ds_load_b64
+      for i, vdst in enumerate(V_B_TILE_REGS):
+        b_off = (i & 1) * 8 + (i & 2) * 64 + (i >> 2) * 256 + iter * LDS_B_STRIDE
+        k.emit(ds_load_b64(vdst=v[vdst:vdst+1], addr=v[V_LDS_B_BASE], offset0=b_off & 0xFF, offset1=b_off >> 8))
+
+    # Issue global prefetch (first 6 iterations only)
+    if iter < 6 and not NO_GLOBAL:
+      vdst1, vdst2, addr, slo1, slo2 = PREFETCH_LOADS[iter]
+      k.emit(global_load_b32(vdst=v[vdst1], addr=v[addr], saddr=s[slo1:slo1+1]))
+      k.emit(global_load_b32(vdst=v[vdst2], addr=v[addr], saddr=s[slo2:slo2+1]))
+
+    # 64 dual FMACs
+    k.waitcnt(lgkm=0)
+    k.emit(s_clause(simm16=len(FMAC_PATTERN)-1))
+    for i, (vdst_x, vdst_y, ax, bx, ay, by) in enumerate(FMAC_PATTERN):
+      k.emit(VOPD(VOPDOp.V_DUAL_FMAC_F32, VOPDOp.V_DUAL_FMAC_F32,
+                  vdstx=v[vdst_x], vdsty=v[vdst_y], srcx0=v[ax], vsrcx1=v[bx], srcy0=v[ay], vsrcy1=v[by]))
+
+  # wait for all global loads to finish
+  # then sync the warp so it's safe to store local
+  k.waitcnt(vm=0)
+  k.emit(s_barrier())
+
+  # Store prefetched data to LDS
+  # NOTE: Register naming reflects LDS tile organization, not source matrix:
+  #   V_LDS_A_DATA (v155-162) holds data that goes to LDS A-tile region
+  #   V_LDS_B_DATA (v163-170) holds data that goes to LDS B-tile region
+  # The data sources are swapped: A-tile receives B matrix rows, B-tile receives A matrix columns
+  if not NO_DS:
+    for i in range(4):  # A tile: 8 values via 4 stride64 stores
+      k.emit(ds_store_2addr_stride64_b32(addr=v[V_LDS_A_ADDR], data0=v[V_LDS_A_DATA[i*2]], data1=v[V_LDS_A_DATA[i*2+1]], offset0=i*4, offset1=i*4+2))
+    for i in range(8):  # B tile: 8 values via 8 scalar stores with 64-byte spacing
+      offset = i * 64
+      k.emit(ds_store_b32(addr=v[V_LDS_B_ADDR], data0=v[V_LDS_B_DATA[i]], offset0=offset & 0xFF, offset1=offset >> 8))
+
+  k.emit(s_branch(), target='LOOP_INC')
+
+  # ===========================================================================
+  # EPILOGUE: Permute and store results
+  # ===========================================================================
+  k.label('EPILOGUE')
+
+  # Rearrange accumulators from FMAC layout to contiguous output order
+  for a, b in PERMUTE_SWAPS:
+    k.emit(v_swap_b32_e32(v[a], v[b]))
+
+  # Compute output base coordinates
+  # v[130] = col_base = tile_x + (lane_id & 7) * 4
+  # v[131] = row_base = tile_y + (lane_id & 0x60) + ((lane_id >> 3) & 3) * 4
+  # v[132] = 0 (for 64-bit address high part)
+  k.emit(v_add_nc_u32_e32(v[130], s[S_TILE_X], v[V_LANE_MOD8_X4]))
+  k.emit(v_and_b32_e32(v[131], 0x60, v[V_LANE_ID]))
+  k.emit(v_add_nc_u32_e32(v[131], s[S_TILE_Y], v[131]))
+  k.emit(v_add_nc_u32_e32(v[131], v[V_LANE_DIV8_X4], v[131]))
+  k.emit(v_mov_b32_e32(v[132], 0))
+
+  # Precompute row offsets: v[133-136] for rows 0-3, v[137-140] for rows 16-19
+  for base, row_off in [(133, 0), (137, 16)]:
+    if row_off: k.emit(v_add_nc_u32_e32(v[141], row_off, v[131]))
+    k.emit(v_mul_lo_u32(v[base], v[141] if row_off else v[131], s[S_DIM_N]))
+    for j in range(3): k.emit(v_add_nc_u32_e32(v[base + 1 + j], s[S_DIM_N], v[base + j]))
+
+  # s[S_PREFETCH_FLAG] = row stride in bytes (N * 4)
+  k.emit(s_lshl_b32(s[S_PREFETCH_FLAG], s[S_DIM_N], 2))
+
+  # Store 128 output values as 32 groups of 4 (128-bit stores)
+  # Layout: 2 row halves (0-3, 16-19) x 4 col groups x 4 rows = 32 stores of 4 floats
+  for i, (row_half, col_off, row_in_group) in enumerate([(rh, co, ri)
+      for rh in range(2) for co in [0, 32, 64, 96] for ri in range(4)]):
+    row = row_half * 16 + row_in_group
+    src = OUT_REGS[i*4]  # first reg of ascending group of 4
+
+    if row_in_group == 0:
+      # First row of group: compute full address
+      if col_off == 0: k.emit(v_mov_b32_e32(v[141], v[130]))
+      else: k.emit(v_add_nc_u32_e32(v[141], col_off, v[130]))
+      row_base = 133 + row if row < 4 else 137 + row - 16
+      k.emit(v_add_nc_u32_e32(v[141], v[row_base], v[141]))
+      k.emit(v_lshlrev_b32_e32(v[141], 2, v[141]))
+      k.emit(v_add_co_u32(v[141], VCC_LO, s[S_OUT_PTR[0]], v[141]))
+      k.emit(v_add_co_ci_u32_e32(v[142], s[S_OUT_PTR[1]], v[132]))
+    else:
+      # Subsequent rows: add stride
+      k.emit(v_add_co_u32(v[141], VCC_LO, s[S_PREFETCH_FLAG], v[141]))
+      k.emit(v_add_co_ci_u32_e32(v[142], v[142], v[132]))
+
+    k.emit(global_store_b128(addr=v[141:142], data=v[src:src+3], saddr=NULL))
+
+  k.emit(s_sendmsg(simm16=3))  # DEALLOC_VGPRS
+  k.emit(s_endpgm())
+
+  return k.finalize()
+
+# =============================================================================
+# Test harness
+# =============================================================================
+
+N = getenv("N", 4096)
+BLOCK_M, BLOCK_N = 128, 128
+THREADS = 128
+
+def test_matmul():
+  dev = Device[Device.DEFAULT]
+  print(f"Device arch: {dev.renderer.arch}")
+
+  insts = build_kernel(N, dev.renderer.arch)
+
+  rng = np.random.default_rng(42)
+  a = Tensor(rng.random((N, N), dtype=np.float32) - 0.5)
+  b = Tensor(rng.random((N, N), dtype=np.float32) - 0.5)
+  c = Tensor.empty(N, N)
+  Tensor.realize(a, b, c)
+
+  grid, local = (N // BLOCK_N, N // BLOCK_M, 1), (THREADS, 1, 1)
+  print(f"Grid: {grid}, Local: {local}")
+
+  dname:str = Device.DEFAULT
+  def asm_kernel(A:UOp, B:UOp, C:UOp) -> UOp:
+    gidxs = [UOp.special(n, f"gidx{i}") for i,n in enumerate(grid)]
+    lidxs = [UOp.special(n, f"lidx{i}") for i,n in enumerate(local)]
+    lds = UOp(Ops.DEFINE_LOCAL, dtypes.uint8.ptr(size=max(LDS_SIZE, 65536//getenv("LIMIT_OCC", 65536)), addrspace=AddrSpace.LOCAL), (), 'lds')
+    sink = UOp.sink(A.base, B.base, C.base, lds, *gidxs, *lidxs, arg=KernelInfo(name=colored("kernel", "cyan"),
+                                                                                  estimates=Estimates(ops=N*N*N*2, mem=N*N*4*3)))
+    return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
+  c = Tensor.custom_kernel(a, b, c, fxn=asm_kernel)[2]
+  ei = c.schedule()[0].lower()
+
+  ets = []
+  with Context(DEBUG=2):
+    for _ in range(getenv("CNT", 5)): ets.append(ei.run(wait=True))
+  print(f"REAL TFLOPS {N * N * N * 2 / min(ets) * 1e-12:.2f}")
+
+  if getenv("VERIFY", 1):
+    GlobalCounters.reset()
+    with Context(DEBUG=2): tc = (a @ b).realize()
+    with Context(DEBUG=0): err = (c - tc).square().mean().item()
+    print(f"mean squared error {err}")
+    if err != err or err > 1e-06:
+      c_np, tc_np = c.numpy(), tc.numpy()
+      for bi in range(N // 128):
+        for bj in range(N // 128):
+          blk_c = c_np[bi*128:(bi+1)*128, bj*128:(bj+1)*128]
+          blk_ref = tc_np[bi*128:(bi+1)*128, bj*128:(bj+1)*128]
+          blk_diff = blk_c - blk_ref
+          zero_rows = [i for i in range(128) if np.all(np.abs(blk_c[i,:]) < 1e-10)]
+          nz_rows = [i for i in range(128) if i not in zero_rows]
+          nz_mse = float(np.mean(blk_diff[nz_rows,:]**2)) if nz_rows else 0
+          print(f"Block ({bi},{bj}): zero_rows={zero_rows}, nz_rows_mse={nz_mse:.2e}")
+          # show first few non-zero row comparisons
+          if nz_rows and nz_mse > 1e-6:
+            for r in nz_rows[:3]:
+              print(f"  row {r} asm[0:8]:  {blk_c[r,:8]}")
+              print(f"  row {r} ref[0:8]:  {blk_ref[r,:8]}")
+      raise RuntimeError("matmul is wrong!")
+
+if __name__ == "__main__":
+  test_matmul()

extra/gemm/amd_uop_matmul.py

@@ -140,11 +140,11 @@ def hand_spec_kernel3():
 
   return sink.sink(arg=KernelInfo(opts_to_apply=())).simplify()
 
-def test_matmul(sink:UOp, N=N):
+def test_matmul(sink:UOp, dtype=dtypes.float32, N=N):
   rng = np.random.default_rng()
-  a = Tensor(rng.random((N, N), dtype=np.float32)-0.5)
-  b = Tensor(rng.random((N, N), dtype=np.float32)-0.5)
-  hc = Tensor.empty(N, N)
+  a = Tensor(rng.random((N, N), dtype=np.float32)-0.5, dtype=dtype)
+  b = Tensor(rng.random((N, N), dtype=np.float32)-0.5, dtype=dtype)
+  hc = Tensor.empty(N, N, dtype=dtype)
   Tensor.realize(a, b, hc)
 
   ei = ExecItem(sink, [t.uop.buffer for t in [hc, a, b]], prg=get_runner(Device.DEFAULT, sink))

extra/gemm/asm/cdna/gemm.py

@@ -0,0 +1,122 @@
+import atexit, functools
+from tinygrad import Tensor, Device, dtypes
+from tinygrad.dtype import AddrSpace
+from tinygrad.uop.ops import UOp, Ops, KernelInfo, AxisType
+from tinygrad.renderer import Estimates
+from tinygrad.helpers import getenv, all_same, DEBUG
+from extra.gemm.asm.cdna.asm import build_kernel, TILE_M, TILE_N, TILE_K, NUM_WG
+
+# ** CDNA4 assembly gemm
+
+WORKGROUP_SIZE = 256
+
+@functools.cache
+def custom_asm_gemm(C:UOp, A:UOp, B:UOp, dname:str) -> UOp:
+  batch, M, K = A.shape
+  K2, N = B.shape[(1 if B.ndim == 3 else 0):]
+  assert K == K2
+  lidx = UOp.special(WORKGROUP_SIZE, "lidx0")
+  gidx = UOp.special(NUM_WG, "gidx0")
+  insts = build_kernel(batch, M, N, K, A.dtype.base)
+  lds = UOp(Ops.DEFINE_LOCAL, dtypes.uint8.ptr(size=133_120, addrspace=AddrSpace.LOCAL), (), 'lds')
+  sink = UOp.sink(C.base, A.base, B.base, lds, lidx, gidx,
+                  arg=KernelInfo(name=f"gemm_{batch}_{M}_{N}_{K}", estimates=Estimates(ops=2*batch*M*N*K, mem=(batch*M*K + K*N + batch*M*N)*2)))
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname),
+                                UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
+
+counters = {"used":0, "todos":[]}
+def todo(msg:str) -> bool: counters["todos"].append(msg); return False
+def _asm_gemm_report():
+  print(f'asm_gemm: {counters["used"]} used, {len(counters["todos"])} not used')
+  if DEBUG >= 2 and counters["todos"]:
+    from collections import Counter
+    for msg, cnt in Counter(counters["todos"]).most_common(): print(f'  {cnt:3d}x {msg}')
+atexit.register(_asm_gemm_report)
+
+def can_use_asm_gemm(a:Tensor, b:Tensor) -> bool:
+  if a.dtype != b.dtype: return todo(f"dtypes must match {a.dtype} != {b.dtype}")
+  if a.dtype not in {dtypes.bfloat16, dtypes.float16}: return todo(f"only bfloat16/float16, got {a.dtype}")
+  batch, M, K = (1, *a.shape) if a.ndim == 2 else a.shape
+  N = b.shape[1]
+  if isinstance(a.device, tuple):
+    if a.ndim == 2 and a.uop.axis == 0 and b.uop.axis is None: M //= len(a.device)
+    elif a.ndim == 2 and a.uop.axis == 1 and b.uop.axis == 0: K //= len(a.device)
+    elif a.ndim == 2 and a.uop.axis is None and b.uop.axis == 1: N //= len(a.device)
+    elif a.ndim == 3 and a.uop.axis == 0 and b.uop.axis is None: batch //= len(a.device)
+    elif a.ndim == 3 and a.uop.axis is None and b.uop.axis == 1: N //= len(a.device)
+    elif a.ndim == 3 and a.uop.axis == 2 and b.uop.axis == 0: K //= len(a.device)
+    else: return todo(f"sharding mismatch a.ndim={a.ndim} a.uop.axis={a.uop.axis} b.uop.axis={b.uop.axis}")
+    dname = a.device[0]
+  else: dname = a.device
+  arch = getattr(Device[dname].renderer, "arch", "")
+  if batch not in {1, 2}: return todo(f"GEMM batch size {batch}")
+  if (M % TILE_M != 0 or N % TILE_N != 0 or K % TILE_K != 0) and arch == "gfx950":
+    return todo(f"GEMM shape ({M},{N},{K}) not a multiple of ({TILE_M},{TILE_N},{TILE_K})")
+  return True
+
+# ** UOp gemm to test Tensor.custom_kernel multi and backward correctness on non cdna4
+# note: this can be removed after we have GEMM on mixins
+
+def custom_uop_gemm(C:UOp, A:UOp, B:UOp) -> UOp:
+  M, K = A.shape[0]*A.shape[1], A.shape[2]
+  K2, N = B.shape[(1 if B.ndim == 3 else 0):]
+  assert K == K2
+  m = UOp.range(M, 1, AxisType.LOOP)
+  n = UOp.range(N, 2, AxisType.LOOP)
+  k = UOp.range(K, 0, AxisType.REDUCE)
+  mul = (A.index((m*UOp.const(dtypes.index, K)+k))*B.index((k*UOp.const(dtypes.index, N)+n))).cast(dtypes.float32)
+  red = mul.reduce(k, arg=Ops.ADD, dtype=dtypes.float32).cast(C.dtype.base)
+  store = C.index((m*UOp.const(dtypes.index, N)+n), ptr=True).store(red).end(m, n)
+  return store.sink(arg=KernelInfo(name=f'uop_gemm_{M}_{N}_{K}'))
+
+# ** backward gemm, might use the asm gemm
+
+def custom_gemm_bw(gradient:UOp, kernel:UOp):
+  out, a, b = kernel.src[1:]
+  assert all_same([gradient.device, a.device, b.device, out.device])
+  a_t, b_t, g_t = Tensor(a, device=a.device), Tensor(b, device=a.device), Tensor(gradient, device=a.device)
+  # TODO: this needs to be cleaned up and done properly, the batch dim of grad and a multi need to align
+  g_t = g_t[:a.shape[0]]
+  grad_a = (g_t @ b_t.T).uop
+  grad_b = (a_t.permute(2, 0, 1).reshape(a_t.shape[2], -1) @ g_t.reshape(-1, g_t.shape[-1])).uop
+  return (None, grad_a, grad_b)
+
+# ** main gemm function
+
+def asm_gemm(a:Tensor, b:Tensor) -> Tensor:
+  assert can_use_asm_gemm(a, b), f"{counters['todos'][-1]}"
+  counters["used"] += 1
+  unfold_batch = a.ndim == 3 and isinstance(a.device, tuple) and a.uop.axis == 2 and b.uop.axis == 0
+  if unfold_batch:
+    orig_batch = a.shape[0]
+    a = a.reshape(a.shape[0]*a.shape[1], a.shape[2])
+  squeeze = a.ndim == 2
+  if squeeze: a = a.unsqueeze(0)
+
+  batch, M, K = a.shape
+  N = b.shape[1]
+  is_multi = isinstance(a.device, tuple)
+  if (k_sharded:=is_multi and a.uop.axis == 2): K //= len(a.device)
+  if (m_sharded:=is_multi and a.uop.axis == 1): M //= len(a.device)
+  n_sharded = is_multi and b.uop.axis == 1
+
+  if is_multi:
+    if n_sharded:
+      out = Tensor(Tensor.empty(batch, M, N//len(a.device), dtype=a.dtype, device=a.device).uop.multi(2), device=a.device)
+    elif m_sharded:
+      out = Tensor(Tensor.empty(batch, M, N, dtype=a.dtype, device=a.device).uop.multi(1), device=a.device)
+    else:
+      out = Tensor(Tensor.empty(batch//len(a.device) if a.uop.axis==0 else batch, M, N, dtype=a.dtype, device=a.device).uop.multi(0), device=a.device)
+  else:
+    out = Tensor.empty(batch, M, N, dtype=a.dtype, device=a.device)
+
+  renderer = Device[a.device[0] if is_multi else a.device].renderer
+  dname, arch = renderer.device, getattr(renderer, "arch", "")
+  if arch.startswith("gfx950") and getenv("USE_ASM", 1):
+    out = Tensor.custom_kernel(out, a, b, fxn=functools.partial(custom_asm_gemm, dname=dname), grad_fxn=custom_gemm_bw)[0]
+  else:
+    out = Tensor.custom_kernel(out, a, b, fxn=custom_uop_gemm, grad_fxn=custom_gemm_bw)[0]
+  if k_sharded: out = out.sum(0)
+  out = out.squeeze(0) if squeeze else out
+  if unfold_batch: out = out.reshape(orig_batch, -1, out.shape[-1])
+  return out

extra/gemm/asm/rdna3/template.s

@@ -0,0 +1,76 @@
+.text
+.section	.text.
+.global	gemm
+.p2align	8
+.type	gemm,@function
+
+gemm:
+INSTRUCTIONS
+
+.section .rodata,"a",@progbits
+.p2align 6, 0x0
+.amdhsa_kernel gemm
+  # basic memory requirements
+  .amdhsa_group_segment_fixed_size 30336
+  .amdhsa_private_segment_fixed_size 0
+  .amdhsa_kernarg_size 32
+  # register usage (RSRC1)
+  .amdhsa_next_free_vgpr 256
+  .amdhsa_next_free_sgpr 100
+  # workgroup / workitem IDs (RSRC2)
+  .amdhsa_system_sgpr_workgroup_id_x 1
+  .amdhsa_system_sgpr_workgroup_id_y 1
+  .amdhsa_system_sgpr_workgroup_id_z 1
+  # user SGPRs: kernarg ptr in s[0:1]
+  .amdhsa_user_sgpr_kernarg_segment_ptr 1
+  .amdhsa_user_sgpr_count 2
+  # gfx10+ / gfx11 specifics (RSRC1[29..31])
+  .amdhsa_wavefront_size32 1
+  .amdhsa_workgroup_processor_mode 1
+  .amdhsa_memory_ordered 1
+  .amdhsa_forward_progress 1
+  # misc for gfx11
+  .amdhsa_dx10_clamp 1
+  .amdhsa_ieee_mode 1
+  .amdhsa_uses_dynamic_stack 0
+.end_amdhsa_kernel
+
+.amdgpu_metadata
+---
+amdhsa.kernels:
+  - .args:
+      - .address_space:  generic
+        .name:           C
+        .offset:         0
+        .size:           8
+        .value_kind:     global_buffer
+        .value_type:     f16
+      - .address_space:  generic
+        .name:           A
+        .offset:         8
+        .size:           8
+        .value_kind:     global_buffer
+        .value_type:     f16
+      - .address_space:  generic
+        .name:           B
+        .offset:         16
+        .size:           8
+        .value_kind:     global_buffer
+        .value_type:     f16
+    .group_segment_fixed_size: 30336
+    .kernarg_segment_align: 8
+    .kernarg_segment_size: 32
+    .max_flat_workgroup_size: 128
+    .name:           gemm
+    .private_segment_fixed_size: 0
+    .sgpr_count:     70
+    .sgpr_spill_count: 0
+    .symbol:         gemm.kd
+    .vgpr_count:     256
+    .vgpr_spill_count: 0
+    .wavefront_size: 32
+amdhsa.version:
+  - 1
+  - 1
+...
+.end_amdgpu_metadata

extra/gemm/asm/rdna3/test.py

@@ -0,0 +1,30 @@
+import math, pathlib
+
+from tinygrad import Device, dtypes
+from tinygrad.uop.ops import UOp, Ops, KernelInfo
+
+from extra.gemm.amd_uop_matmul import test_matmul
+
+N = 4096
+TN = 96
+THREADS_PER_WG = 128
+NUM_WG = math.ceil(N / TN) * math.ceil(N / TN)
+
+dname:str = Device.DEFAULT
+template:str = (pathlib.Path(__file__).parent/"template.s").read_text()
+
+def asm_kernel() -> UOp:
+  lidx = UOp.special(THREADS_PER_WG, "lidx0")
+  gidx = UOp.special(NUM_WG, "gidx0")
+
+  a = UOp.placeholder((N*N,), dtypes.half, slot=1)
+  b = UOp.placeholder((N*N,), dtypes.half, slot=2)
+  c = UOp.placeholder((N*N,), dtypes.half, slot=0)
+
+  src = template.replace("INSTRUCTIONS", (pathlib.Path(__file__).parent/"gemm.s").read_text())
+
+  sink = UOp.sink(a, b, c, lidx, gidx, arg=KernelInfo(name="gemm"))
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=src)))
+
+if __name__ == "__main__":
+  test_matmul(asm_kernel(), dtype=dtypes.half, N=N)

extra/gemm/asm/test.py

@@ -1,72 +0,0 @@
-# Run assembly on the AMD runtime and check correctness
-# VIZ=2 to profile
-import pathlib
-from tinygrad import Tensor, Device, dtypes
-from tinygrad.engine.realize import ExecItem, CompiledRunner
-from tinygrad.renderer import ProgramSpec
-from tinygrad.uop.ops import track_rewrites, UOp
-from tinygrad.helpers import TracingKey, getenv
-
-fp = pathlib.Path(__file__).parent/"gemm.s"
-
-N = getenv("N", 8192)
-THREADS_PER_WG = 256
-NUM_WG = N//THREADS_PER_WG * N//THREADS_PER_WG
-
-assert N % THREADS_PER_WG == 0, "N must be divisible by THREADS_PER_WG"
-
-# ** generate inputs on CPU
-
-scale = 10.0
-
-import torch
-torch.manual_seed(0)
-A = (torch.randn(N, N, dtype=torch.float32, device="cpu") / scale).to(torch.bfloat16).contiguous()
-B = (torch.randn(N, N, dtype=torch.float32, device="cpu") / scale).to(torch.bfloat16).contiguous()
-Bt = B.t().contiguous() # transpose B for the baseline gemm
-C_torch = A@Bt
-
-# ** copy buffers to AMD
-
-# input creation and validation run on the copy engine for simpler tracing
-
-def from_torch(t:torch.Tensor) -> Tensor:
-  return Tensor.from_blob(t.data_ptr(), t.shape, dtype=dtypes.bfloat16, device="cpu").to(Device.DEFAULT).realize()
-
-C_tiny = Tensor.matmul(from_torch(A), from_torch(Bt), dtype=dtypes.float32).cast(dtypes.bfloat16)
-C_asm = Tensor.empty_like(C_tiny)
-C_asm.uop.buffer.allocate()
-
-# ** run gemms
-
-# baseline tinygrad
-sched = C_tiny.schedule()
-assert len(sched) == 1
-eis:list[ExecItem] = [sched[-1].lower()]
-ast = sched[-1].ast
-
-# assembly gemm
-@track_rewrites(name=lambda ret: TracingKey(ret.name, (ret.function_name,), ret))
-def get_asm_prg() -> ProgramSpec:
-  src = fp.read_text()
-  lib = Device[Device.DEFAULT].compiler.compile(src)
-  return ProgramSpec("gemm", src, Device.DEFAULT, ast, lib=lib, global_size=[NUM_WG, 1, 1], local_size=[THREADS_PER_WG, 1, 1],
-                     globals=[0, 1, 2], vars=[UOp.variable("SZ", 256, 8192), UOp.variable("NUM_WG", 1, 1024)])
-eis.append(ExecItem(ast, [C_asm.uop.buffer, from_torch(B).uop.buffer, from_torch(A).uop.buffer], fixedvars={"SZ":N, "NUM_WG":NUM_WG},
-                    prg=CompiledRunner(get_asm_prg())))
-
-for ei in eis:
-  et = ei.run(wait=True)
-  print(f"{(N*N*N*2 / et)*1e-12:.2f} REAL TFLOPS")
-
-# ** correctness
-
-import ctypes
-
-def torch_bf16(t:Tensor) -> torch.tensor:
-  asm_out = t.to("cpu").realize().uop.buffer._buf
-  buf = (ctypes.c_uint16*C_asm.uop.size).from_address(asm_out.va_addr)
-  return torch.frombuffer(buf, dtype=torch.bfloat16, count=C_asm.uop.size).reshape(C_asm.shape)
-
-assert torch.allclose(torch_bf16(C_asm), C_torch, rtol=1e-2, atol=1e-3)
-assert torch.allclose(torch_bf16(C_tiny), C_torch, rtol=1e-2, atol=1e-3)

extra/gemm/asm/unpack_kd.py

@@ -1,12 +1,12 @@
-# unpack the complete kernel descriptor of an amdgpu ELF of for gfx950
+# unpack the complete kernel descriptor of an amdgpu ELF
 # https://rocm.docs.amd.com/projects/llvm-project/en/latest/LLVM/llvm/html/AMDGPUUsage.html#code-object-v3-kernel-descriptor
-import struct, pathlib
+import struct, pathlib, sys
 from tinygrad.runtime.support.elf import elf_loader
 
 def bits(x, lo, hi): return (x >> lo) & ((1 << (hi - lo + 1)) - 1)
 def assert_zero(x, lo, hi): assert bits(x, lo, hi) == 0
 
-with open(fp:=pathlib.Path(__file__).parent/"lib", "rb") as f:
+with open(sys.argv[1], "rb") as f:
   lib = f.read()
 
 image, sections, relocs = elf_loader(lib)
@@ -49,7 +49,7 @@ print("COMPUTE_PGM_RSRC3: 0x%08x" % pgm_rsrc3)
 print("COMPUTE_PGM_RSRC1: 0x%08x" % pgm_rsrc1)
 print("COMPUTE_PGM_RSRC2: 0x%08x" % pgm_rsrc2)
 
-# rsrc 3
+# rsrc 3 (gfx950)
 
 accum_offset_raw = bits(pgm_rsrc3, 0, 5)
 assert_zero(pgm_rsrc3, 6, 15)
@@ -169,10 +169,10 @@ assert_zero(desc, 458, 459)
 uses_dynamic_stack = bits(desc, 459, 460)
 print("DESC.USES_DYNAMIC_STACK:", uses_dynamic_stack)
 
+# gfx950 only
 assert_zero(desc, 460, 463)
 kernarg_preload_spec_length = bits(desc, 464, 470)
 print("DESC.KERNARG_PRELOAD_SPEC_LENGTH:", kernarg_preload_spec_length)
-
 kernarg_preload_spec_offset = bits(desc, 471, 479)
 print("DESC.KERNARG_PRELOAD_SPEC_OFFSET:", kernarg_preload_spec_offset)
 

extra/gemm/intel_xmx.py

@@ -37,7 +37,7 @@ b.copyin(row.data)
 c.copyin(mat.data)
 ret = prog(a._buf, b._buf, c._buf, global_size=[1,1,1], local_size=[8,1,1], wait=True)
 print(ret)
-out = np.frombuffer(a.as_buffer(), np.float32)
+out = np.frombuffer(a.as_memoryview(), np.float32)
 real = row.astype(np.float32)@mat.T.astype(np.float32)
 print("out:", out)
 print("real", real)

extra/gemm/triton_nv_matmul.py

@@ -98,10 +98,10 @@ if __name__ == "__main__":
   # check correctness
   if getenv("VERIFY"):
     from tinygrad.engine.realize import run_schedule
-    triton_buf = np.frombuffer(si.bufs[0].as_buffer(), np.float16).reshape(M,N)
+    triton_buf = np.frombuffer(si.bufs[0].as_memoryview(), np.float16).reshape(M,N)
     print(triton_buf)
     run_schedule(sched)
-    tinygrad_buf = np.frombuffer(si.bufs[0].as_buffer(), np.float16).reshape(M,N)
+    tinygrad_buf = np.frombuffer(si.bufs[0].as_memoryview(), np.float16).reshape(M,N)
     print(tinygrad_buf)
     np.testing.assert_allclose(triton_buf, tinygrad_buf)
     print("correct!")

extra/hcq/hcq_smi.py

@@ -1,14 +1,15 @@
 #!/usr/bin/env python3
 
 import argparse, glob, os, time, subprocess, sys
+from tinygrad.helpers import temp
 
 def scan_devs_based_on_lock(prefix:str, args) -> list[str]:
   target_dev = args.pci_bus if 'pci_bus' in args.__dir__() else ""
 
   devs = []
-  for dev in glob.glob(f'/tmp/{prefix}_*.lock'):
-    dev_id = dev[8:-5]
-    if os.path.exists(f"/sys/bus/pci/devices/{dev_id}") and dev_id.startswith(target_dev): devs.append(dev_id)
+  for dev in glob.glob(temp(f'{prefix}_*.lock')):
+    dev_id = dev.split('/')[-1][len(prefix)+1:-5]
+    if dev_id.startswith(target_dev): devs.append(dev_id)
   return devs
 
 def _do_reset_device(pci_bus): os.system(f"sudo sh -c 'echo 1 > /sys/bus/pci/devices/{pci_bus}/reset'")
@@ -53,16 +54,7 @@ def cmd_show_pids(args):
 
   for dev in devs:
     try:
-      pid = subprocess.check_output(['sudo', 'lsof', f'/tmp/{prefix}_{dev}.lock']).decode('utf-8').strip().split('\n')[1].split()[1]
-      print(f"{dev}: {pid}")
-    except subprocess.CalledProcessError: print(f"{dev}: No processes found using this device")
-
-def cmd_kill_pids(args):
-  devs = scan_devs_based_on_lock(prefix:={"amd":"am", "nv":"nv"}[args.backend], args)
-
-  for dev in devs:
-    try:
-      pid = subprocess.check_output(['sudo', 'lsof', f'/tmp/{prefix}_{dev}.lock']).decode('utf-8').strip().split('\n')[1].split()[1]
+      pid = subprocess.check_output(['sudo', 'lsof', temp(f'{prefix}_{dev}.lock')]).decode('utf-8').strip().split('\n')[1].split()[1]
       print(f"{dev}: {pid}")
     except subprocess.CalledProcessError: print(f"{dev}: No processes found using this device")
 
@@ -74,7 +66,7 @@ def cmd_kill_pids(args):
       if i > 0: time.sleep(0.2)
 
       try:
-        try: pid = subprocess.check_output(['sudo', 'lsof', f'/tmp/{prefix}_{dev}.lock']).decode('utf-8').strip().split('\n')[1].split()[1]
+        try: pid = subprocess.check_output(['sudo', 'lsof', temp(f'{prefix}_{dev}.lock')]).decode('utf-8').strip().split('\n')[1].split()[1]
         except subprocess.CalledProcessError: break
 
         print(f"Killing process {pid} (which uses {dev})")

extra/hevc/decode.py

@@ -1,8 +1,42 @@
-import argparse, os, hashlib
-from tinygrad.helpers import getenv, DEBUG, round_up, Timing, tqdm, fetch
+import argparse, os, hashlib, functools
+from typing import Iterator, Callable
+from tinygrad.helpers import getenv, DEBUG, round_up, Timing, tqdm, fetch, ceildiv
 from extra.hevc.hevc import parse_hevc_file_headers, untile_nv12, to_bgr, nv_gpu
 from tinygrad import Tensor, dtypes, Device, Variable, TinyJit
 
+# rounds up hevc input data to 32 bytes, so more optimal kernels can be generated
+HEVC_ROUNDUP = getenv("DATA_ROUNDUP", 32)
+
+@functools.cache
+def _hevc_jitted_decoder(out_image_size:tuple[int, int], max_hist:int, inplace:bool):
+  def hevc_decode_frame(pos:Variable, hevc_tensor:Tensor, offset:Variable, sz:Variable, opaque:Tensor, i:Variable, *hist:Tensor, outbuf:Tensor|None=None):
+    x = hevc_tensor[offset:offset+sz*HEVC_ROUNDUP].decode_hevc_frame(pos, out_image_size, opaque[i], hist)
+    if outbuf is not None: outbuf.assign(x).realize()
+    return x.realize()
+  return TinyJit(hevc_decode_frame)
+
+def hevc_decode(hevc_tensor:Tensor, opaque:Tensor, frame_info:list, luma_h:int, luma_w:int,
+                history:list[Tensor]|None=None, preallocated_outputs:list[Tensor]|None=None, warmup=False) -> Iterator[Tensor]:
+  out_image_size = luma_h + (luma_h + 1) // 2, round_up(luma_w, 64)
+  max_hist = max((hs for _, _, _, hs, _ in frame_info), default=0)
+
+  v_pos = Variable("pos", 0, max_hist + 1)
+  v_offset = Variable("offset", 0, hevc_tensor.numel()-1)
+  v_sz = Variable("sz", 1, ceildiv(hevc_tensor.numel(), HEVC_ROUNDUP))
+  v_i = Variable("i", 0, len(frame_info)-1)
+
+  decode_jit = _hevc_jitted_decoder(out_image_size, max_hist, preallocated_outputs is not None)
+  history = history or [Tensor.empty(*out_image_size, dtype=dtypes.uint8, device="NV").contiguous().realize() for _ in range(max_hist)]
+  assert len(history) == max_hist, f"history length {len(history)} does not match max_hist {max_hist}"
+
+  for i, (offset, sz, frame_pos, _, is_hist) in enumerate(frame_info):
+    history = history[-max_hist:] if max_hist > 0 else []
+    img = decode_jit(v_pos.bind(frame_pos), hevc_tensor, v_offset.bind(offset), v_sz.bind(ceildiv(sz, HEVC_ROUNDUP)),
+                     opaque, v_i.bind(i), *history, outbuf=preallocated_outputs[i] if preallocated_outputs else None)
+    res = preallocated_outputs[i] if preallocated_outputs else img.clone().realize()
+    if is_hist: history.append(res)
+    yield res
+
 if __name__ == "__main__":
   parser = argparse.ArgumentParser()
   parser.add_argument("--input_file", type=str, default="")
@@ -19,49 +53,32 @@ if __name__ == "__main__":
   dat_hash = hashlib.md5(dat).hexdigest()
 
   with Timing("prep infos: "):
-    dat_nv = hevc_tensor.to("NV")
     opaque, frame_info, w, h, luma_w, luma_h, chroma_off = parse_hevc_file_headers(dat)
 
   frame_info = frame_info[:getenv("MAX_FRAMES", len(frame_info))]
 
   # move all needed data to gpu
-  #all_slices = []
   with Timing("copy to gpu: "):
     opaque_nv = opaque.to("NV").contiguous().realize()
     hevc_tensor = hevc_tensor.to("NV")
 
   out_image_size = luma_h + (luma_h + 1) // 2, round_up(luma_w, 64)
-  max_hist = max(history_sz for _, _, _, history_sz, _ in frame_info)
 
-  # define variables
-  v_pos = Variable("pos", 0, max_hist + 1)
-  v_offset = Variable("offset", 0, hevc_tensor.numel()-1)
-  v_sz = Variable("sz", 0, hevc_tensor.numel())
-  v_i = Variable("i", 0, len(frame_info)-1)
+  # preallocate output/hist buffers
+  max_hist = max((hs for _, _, _, hs, _ in frame_info), default=0)
+  hist = [Tensor.empty(*out_image_size, dtype=dtypes.uint8, device="NV").contiguous().realize() for _ in range(max_hist)]
+  out_images = [Tensor.zeros(*out_image_size, dtype=dtypes.uint8, device="NV").contiguous().realize() for _ in range(len(frame_info))]
 
-  @TinyJit
-  def decode_jit(pos:Variable, src:Tensor, data:Tensor, *hist:Tensor):
-    return src.decode_hevc_frame(pos, out_image_size, data, hist).realize()
+  # warmup decode
+  _ = list(hevc_decode(hevc_tensor, opaque_nv, frame_info[:3], luma_h, luma_w, history=hist, preallocated_outputs=out_images))
+  Device.default.synchronize()
 
-  # warm up
-  history = [Tensor.empty(*out_image_size, dtype=dtypes.uint8, device="NV") for _ in range(max_hist)]
-  for i in range(3):
-    hevc_frame = hevc_tensor.shrink((((bound_offset:=v_offset.bind(frame_info[0][0])), bound_offset+v_sz.bind(frame_info[0][1])),))
-    decode_jit(v_pos.bind(0), hevc_frame, opaque_nv[v_i.bind(0)], *history)
-
-  out_images = []
+  # decode all frames using the iterator
   with Timing("decoding whole file: ", on_exit=(lambda et: f", {len(frame_info)} frames, {len(frame_info)/(et/1e9):.2f} fps")):
-    for i, (offset, sz, frame_pos, history_sz, is_hist) in enumerate(frame_info):
-      history = history[-max_hist:] if max_hist > 0 else []
-      # TODO: this shrink should work as a slice
-      hevc_frame = hevc_tensor.shrink((((bound_offset:=v_offset.bind(offset)), bound_offset+v_sz.bind(sz)),))
-
-      outimg = decode_jit(v_pos.bind(frame_pos), hevc_frame, opaque_nv[v_i.bind(i)], *history).clone()
-      out_images.append(outimg)
-      if is_hist: history.append(outimg)
-
+    images = list(hevc_decode(hevc_tensor, opaque_nv, frame_info, luma_h, luma_w, history=hist, preallocated_outputs=out_images))
     Device.default.synchronize()
 
+  # validation
   if getenv("VALIDATE", 0):
     import pickle
     if dat_hash == "b813bfdbec194fd17fdf0e3ceb8cea1c":
@@ -70,7 +87,7 @@ if __name__ == "__main__":
     else: decoded_frames = pickle.load(open(f"extra/hevc/decoded_frames_{dat_hash}.pkl", "rb"))
   else: import cv2
 
-  for i, img in tqdm(enumerate(out_images)):
+  for i, img in tqdm(enumerate(images)):
     if getenv("VALIDATE", 0):
       if i < len(decoded_frames) and len(decoded_frames[i]) > 0:
         img = untile_nv12(img, h, w, luma_w, chroma_off).realize()

extra/hip_gpu_driver/hip_ioctl.py

@@ -13,7 +13,7 @@ def get_struct(argp, stype):
 
 def format_struct(s):
   sdats = []
-  for field_name, field_type in s._fields_:
+  for field_name, *_ in s._real_fields_:
     dat = getattr(s, field_name)
     if isinstance(dat, int): sdats.append(f"{field_name}:0x{dat:X}")
     else: sdats.append(f"{field_name}:{dat}")
@@ -46,6 +46,22 @@ def install_hook(c_function, python_function):
 # *** ioctl lib end ***
 
 import tinygrad.runtime.autogen.kfd as kfd_ioctl
+import tinygrad.runtime.autogen.hsa as hsa
+
+def print_aql_queue(read_pointer_address):
+  rptr_offset = getattr(hsa.amd_queue_v2_t, 'read_dispatch_id').offset
+  queue_base = read_pointer_address - rptr_offset
+  queue = hsa.amd_queue_v2_t.from_address(queue_base)
+  print(f"  AQL Queue @ 0x{queue_base:X}:")
+  for field_name, *_ in hsa.amd_queue_v2_t._real_fields_:
+    val = getattr(queue, field_name)
+    if isinstance(val, int): print(f"    {field_name}: 0x{val:X}")
+    elif hasattr(val, '_length_'):
+      arr_vals = [f"{format_struct(v)}" if hasattr(v, '_real_fields_') else f"{v:#X}" for v in val]
+      print(f"    {field_name}: [{', '.join(arr_vals)}]")
+    elif hasattr(val, '_real_fields_'): print(f"    {field_name}: {format_struct(val)}")
+    else: print(f"    {field_name}: {val}")
+
 def ioctls_from_header():
   hdr = (pathlib.Path(__file__).parent / "kfd_ioctl.h").read_text().replace("\\\n", "")
   pattern = r'#define\s+(AMDKFD_IOC_[A-Z0-9_]+)\s+AMDKFD_IOW?R?\((0x[0-9a-fA-F]+),\s+struct\s([A-Za-z0-9_]+)\)'
@@ -67,6 +83,7 @@ def ioctl(fd, request, argp):
     if name == "AMDKFD_IOC_SVM":
       out = ctypes.cast(s.attrs, ctypes.POINTER(kfd_ioctl.struct_kfd_ioctl_svm_attribute))
       for i in range(s.nattr): print(f"{i}: {kfd_ioctl.enum_kfd_ioctl_svm_attr_type.get(out[i].type):40s}: {out[i].value:#x}")
+    if name == "AMDKFD_IOC_CREATE_QUEUE" and s.queue_type == kfd_ioctl.KFD_IOC_QUEUE_TYPE_COMPUTE_AQL: print_aql_queue(s.read_pointer_address)
   else:
     print(f"{(st-start)*1000:7.2f} ms +{et*1000.:7.2f} ms : ioctl",
           f"{idir=} {size=} {itype=} {nr=} {fd=} {ret=}", os.readlink(f"/proc/self/fd/{fd}") if fd >= 0 else "")

extra/hip_large_kernel.py

@@ -18,7 +18,7 @@ prg = dev.runtime("write_ones", mbin)
 prg(buf0._buf, global_size=(1,65537,1), local_size=(1,1,1), wait=True)
 
 import numpy as np
-def to_np(buf): return np.frombuffer(buf.as_buffer().cast(buf.dtype.base.fmt), dtype=_to_np_dtype(buf.dtype.base))
+def to_np(buf): return np.frombuffer(buf.as_memoryview().cast(buf.dtype.base.fmt), dtype=_to_np_dtype(buf.dtype.base))
 
 big = to_np(buf0)
 print(big)

extra/huggingface_onnx/huggingface_manager.py

@@ -8,14 +8,14 @@ from tinygrad.helpers import _ensure_downloads_dir
 DOWNLOADS_DIR = _ensure_downloads_dir() / "models"
 from tinygrad.helpers import tqdm
 
-def snapshot_download_with_retry(*, repo_id: str, allow_patterns: list[str]|tuple[str, ...]|None=None, cache_dir: str|Path|None=None,
+def snapshot_download_with_retry(*, repo_id: str, allow_patterns: list[str]|tuple[str, ...]|None=None, local_dir: str|Path|None=None,
                                  tries: int=2, **kwargs) -> Path:
   for attempt in range(tries):
     try:
       return Path(snapshot_download(
         repo_id=repo_id,
         allow_patterns=allow_patterns,
-        cache_dir=str(cache_dir) if cache_dir is not None else None,
+        local_dir=str(local_dir) if local_dir is not None else None,
         **kwargs
       ))
     except Exception as e:
@@ -144,14 +144,14 @@ class HuggingFaceONNXManager:
         root_path = snapshot_download_with_retry(
           repo_id=model_id,
           allow_patterns=allow_patterns,
-          cache_dir=str(self.models_dir)
+          local_dir=str(self.models_dir / model_id)
         )
 
         # Download config files (usually small)
         snapshot_download_with_retry(
           repo_id=model_id,
           allow_patterns=["*config.json"],
-          cache_dir=str(self.models_dir)
+          local_dir=str(self.models_dir / model_id)
         )
 
         model_data["download_path"] = str(root_path)

extra/huggingface_onnx/run_models.py

@@ -88,8 +88,8 @@ if __name__ == "__main__":
       # repo id
       # validates all onnx models inside repo
       repo_id = "/".join(path)
-      root_path = snapshot_download_with_retry(repo_id=repo_id, allow_patterns=["*.onnx", "*.onnx_data"], cache_dir=DOWNLOADS_DIR)
-      snapshot_download_with_retry(repo_id=repo_id, allow_patterns=["*config.json"], cache_dir=DOWNLOADS_DIR)
+      root_path = snapshot_download_with_retry(repo_id=repo_id, allow_patterns=["*.onnx", "*.onnx_data"], local_dir=DOWNLOADS_DIR / repo_id)
+      snapshot_download_with_retry(repo_id=repo_id, allow_patterns=["*config.json"], local_dir=DOWNLOADS_DIR / repo_id)
       config = get_config(root_path)
       for onnx_model in root_path.rglob("*.onnx"):
         rtol, atol = get_tolerances(onnx_model.name)
@@ -101,8 +101,8 @@ if __name__ == "__main__":
       onnx_model = path[-1]
       assert path[-1].endswith(".onnx")
       repo_id, relative_path = "/".join(path[:2]), "/".join(path[2:])
-      root_path = snapshot_download_with_retry(repo_id=repo_id, allow_patterns=[relative_path], cache_dir=DOWNLOADS_DIR)
-      snapshot_download_with_retry(repo_id=repo_id, allow_patterns=["*config.json"], cache_dir=DOWNLOADS_DIR)
+      root_path = snapshot_download_with_retry(repo_id=repo_id, allow_patterns=[relative_path], local_dir=DOWNLOADS_DIR / repo_id)
+      snapshot_download_with_retry(repo_id=repo_id, allow_patterns=["*config.json"], local_dir=DOWNLOADS_DIR / repo_id)
       config = get_config(root_path)
       rtol, atol = get_tolerances(onnx_model)
       print(f"validating {relative_path} with truncate={args.truncate}, {rtol=}, {atol=}")

extra/mmapeak/mmapeak.py

@@ -1,99 +1,105 @@
-import os, pathlib
+import os
 
 # TODO: there is a timing bug without this
 os.environ["AMD_AQL"] = "1"
 
-from tinygrad.device import Device
-from tinygrad.runtime.ops_amd import AMDProgram, HIPCompiler
+from tinygrad import Tensor, Device
+from tinygrad.helpers import getenv
+from tinygrad.uop.ops import UOp, Ops, KernelInfo
+from tinygrad.renderer import Estimates
+from tinygrad.renderer.amd.dsl import Reg, Inst, s, v
 
 NUM_WORKGROUPS = 96
 WAVE_SIZE = 32
-NUM_WAVES = 2
+NUM_WAVES = 4
 FLOPS_PER_MATMUL = 16*16*16*2
-INTERNAL_LOOP = 1_000_00
+INTERNAL_LOOP = getenv("LOOP", 10_000)
 INSTRUCTIONS_PER_LOOP = 200
-DIRECTIVE = ".amdhsa_wavefront_size32 1"
 
-assemblyTemplate = (pathlib.Path(__file__).parent / "template.s").read_text()
+def repeat(insts:list[Inst], n:int, counter_sreg:Reg) -> list[Inst]:
+  insts_bytes = b"".join([inst.to_bytes() for inst in insts])
+  sub_inst, cmp_inst = s_sub_u32(counter_sreg, counter_sreg, 1), s_cmp_lg_i32(counter_sreg, 0)
+  loop_sz = len(insts_bytes) + sub_inst.size() + cmp_inst.size()
+  branch_inst = s_cbranch_scc1(simm16=-((loop_sz // 4) + 1) & 0xFFFF)
+  return [s_mov_b32(counter_sreg, n)] + insts + [sub_inst, cmp_inst, branch_inst, s_endpgm()]
 
-def launchBenchmark(instruction, vgprIndices, dense=True, accum=False, extra=""):
+def launchBenchmark(instruction, vgprIndices, dense=True, accum=False, **kwargs):
   if accum:
-    instructions = "{} a[0:{}], v[{}:{}], v[{}:{}], 1{}\n".format(instruction, vgprIndices[0],
-                                                                vgprIndices[1], vgprIndices[2],
-                                                                vgprIndices[1], vgprIndices[2], extra)
+    inst = instruction(v[0:vgprIndices[0]], v[vgprIndices[1]:vgprIndices[2]], v[vgprIndices[1]:vgprIndices[2]], 1, acc_cd=1, **kwargs)
   elif dense:
-    instructions = "{} v[0:{}], v[{}:{}], v[{}:{}], 1\n".format(instruction, vgprIndices[0],
-                                                                vgprIndices[1], vgprIndices[2],
-                                                                vgprIndices[1], vgprIndices[2])
+    inst = instruction(v[0:vgprIndices[0]], v[vgprIndices[1]:vgprIndices[2]], v[vgprIndices[1]:vgprIndices[2]], 1)
   else:
-    instructions = "{} v[0:{}], v[{}:{}], v[{}:{}], v{}\n".format(instruction, vgprIndices[0],
-                                                                  vgprIndices[1], vgprIndices[2],
-                                                                  vgprIndices[3], vgprIndices[4],
-                                                                  vgprIndices[5])
-  src = assemblyTemplate.replace("INTERNAL_LOOP", str(INTERNAL_LOOP)).replace("INSTRUCTION", instructions*INSTRUCTIONS_PER_LOOP)
-  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)])
+    inst = instruction(v[0:vgprIndices[0]], v[vgprIndices[1]:vgprIndices[2]], v[vgprIndices[3]:vgprIndices[4]], v[vgprIndices[5]])
+  insts = repeat([inst for _ in range(INSTRUCTIONS_PER_LOOP)], n=INTERNAL_LOOP, counter_sreg=s[1])
+  def fxn(A:UOp) -> UOp:
+    threads = UOp.special(WAVE_SIZE * NUM_WAVES, "lidx0")
+    gidx = UOp.special(NUM_WORKGROUPS, "gidx0")
+    FLOPs = FLOPS_PER_MATMUL * NUM_WAVES * NUM_WORKGROUPS * INTERNAL_LOOP * INSTRUCTIONS_PER_LOOP
+    sink = UOp.sink(A.base, threads, gidx, arg=KernelInfo(inst.op.name.lower(), estimates=Estimates(ops=FLOPs, mem=0)))
+    return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg="AMD"), UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
+  dummy = Tensor.zeros(1).contiguous().realize()
+  out = Tensor.custom_kernel(dummy, fxn=fxn)[0]
+  ei = out.schedule()[-1].lower()
+  elapsed = min([ei.run(wait=True) for _ in range(2)])
   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")
+  print(f"{inst.op_name.lower():<29} : {FLOPs/elapsed/10**12:.2f} T(FL)OPS")
 
 if __name__=="__main__":
-  DEVICENUM = os.getenv("DEVICENUM", "0")
-  try:
-    DEV = Device['AMD:' + DEVICENUM]
-  except:
-    raise RuntimeError("Error while initiating AMD device")
+  DEV = Device[Device.DEFAULT]
+  arch = DEV.renderer.arch
 
-  COMPILER = HIPCompiler(DEV.arch)
-  if DEV.arch in {'gfx1100', 'gfx1103', 'gfx1151'}:
-    if DEV.arch == 'gfx1103': NUM_WORKGROUPS = 8
-    if DEV.arch == 'gfx1151': NUM_WORKGROUPS = 32
-    launchBenchmark("v_wmma_bf16_16x16x16_bf16", (7,8,15))
-    launchBenchmark("v_wmma_f16_16x16x16_f16", (7,8,15))
-    launchBenchmark("v_wmma_f32_16x16x16_bf16", (7,8,15))
-    launchBenchmark("v_wmma_f32_16x16x16_f16", (7,8,15))
-    launchBenchmark("v_wmma_i32_16x16x16_iu4", (7,8,9))
-    launchBenchmark("v_wmma_i32_16x16x16_iu8", (7,8,11))
-  elif DEV.arch == 'gfx1201':
+  if arch in {'gfx1100', 'gfx1103', 'gfx1151'}:
+    from tinygrad.runtime.autogen.amd.rdna3.ins import *
+    if arch == 'gfx1103': NUM_WORKGROUPS = 8
+    if arch == 'gfx1151': NUM_WORKGROUPS = 32
+    launchBenchmark(v_wmma_bf16_16x16x16_bf16, (7,8,15))
+    launchBenchmark(v_wmma_f16_16x16x16_f16, (7,8,15))
+    launchBenchmark(v_wmma_f32_16x16x16_bf16, (7,8,15))
+    launchBenchmark(v_wmma_f32_16x16x16_f16, (7,8,15))
+    launchBenchmark(v_wmma_i32_16x16x16_iu4, (7,8,9))
+    launchBenchmark(v_wmma_i32_16x16x16_iu8, (7,8,11))
+  elif arch in {'gfx1200', 'gfx1201'}:
+    from tinygrad.runtime.autogen.amd.rdna4.ins import *
+    # this instruction does not exist in the rdna4 isa, use the co version
+    s_sub_u32 = s_sub_co_u32
     NUM_WORKGROUPS = 64
-    launchBenchmark("v_wmma_bf16_16x16x16_bf16", (3,4,7))
-    launchBenchmark("v_wmma_f16_16x16x16_f16", (3,4,7))
-    launchBenchmark("v_wmma_f32_16x16x16_bf16", (7,8,11))
-    launchBenchmark("v_wmma_f32_16x16x16_f16", (7,8,11))
-    launchBenchmark("v_wmma_i32_16x16x16_iu4", (7,8,8))
-    launchBenchmark("v_wmma_i32_16x16x16_iu8", (7,8,9))
-    launchBenchmark("v_wmma_f32_16x16x16_fp8_fp8", (7,8,9))
-    launchBenchmark("v_wmma_f32_16x16x16_fp8_bf8", (7,8,9))
-    launchBenchmark("v_wmma_f32_16x16x16_bf8_fp8", (7,8,9))
-    launchBenchmark("v_wmma_f32_16x16x16_bf8_bf8", (7,8,9))
+    launchBenchmark(v_wmma_bf16_16x16x16_bf16, (3,4,7))
+    launchBenchmark(v_wmma_f16_16x16x16_f16, (3,4,7))
+    launchBenchmark(v_wmma_f32_16x16x16_bf16, (7,8,11))
+    launchBenchmark(v_wmma_f32_16x16x16_f16, (7,8,11))
+    launchBenchmark(v_wmma_i32_16x16x16_iu4, (7,8,8))
+    launchBenchmark(v_wmma_i32_16x16x16_iu8, (7,8,9))
+    launchBenchmark(v_wmma_f32_16x16x16_fp8_fp8, (7,8,9))
+    launchBenchmark(v_wmma_f32_16x16x16_fp8_bf8, (7,8,9))
+    launchBenchmark(v_wmma_f32_16x16x16_bf8_fp8, (7,8,9))
+    launchBenchmark(v_wmma_f32_16x16x16_bf8_bf8, (7,8,9))
     FLOPS_PER_MATMUL = 16*16*32*2
-    launchBenchmark("v_wmma_i32_16X16X32_iu4", (7,8,9))
-    launchBenchmark("v_swmmac_f32_16x16x32_f16", (7,8,11,12,19,20), False)
-    launchBenchmark("v_swmmac_f32_16x16x32_bf16", (7,8,11,12,19,20), False)
-    launchBenchmark("v_swmmac_f16_16x16x32_f16", (3,4,7,8,15,16), False)
-    launchBenchmark("v_swmmac_bf16_16x16x32_bf16", (3,4,7,8,15,16), False)
-    launchBenchmark("v_swmmac_i32_16x16x32_iu8", (7,8,9,10,13,14), False)
-    launchBenchmark("v_swmmac_i32_16x16x32_iu4", (7,8,8,9,10,11), False)
-    launchBenchmark("v_swmmac_f32_16x16x32_fp8_fp8", (7,8,9,10,13,14), False)
-    launchBenchmark("v_swmmac_f32_16x16x32_fp8_bf8", (7,8,9,10,13,14), False)
-    launchBenchmark("v_swmmac_f32_16x16x32_bf8_fp8", (7,8,9,10,13,14), False)
-    launchBenchmark("v_swmmac_f32_16x16x32_bf8_bf8", (7,8,9,10,13,14), False)
+    launchBenchmark(v_wmma_i32_16x16x32_iu4, (7,8,9))
+    launchBenchmark(v_swmmac_f32_16x16x32_f16, (7,8,11,12,19,20), False)
+    launchBenchmark(v_swmmac_f32_16x16x32_bf16, (7,8,11,12,19,20), False)
+    launchBenchmark(v_swmmac_f16_16x16x32_f16, (3,4,7,8,15,16), False)
+    launchBenchmark(v_swmmac_bf16_16x16x32_bf16, (3,4,7,8,15,16), False)
+    launchBenchmark(v_swmmac_i32_16x16x32_iu8, (7,8,9,10,13,14), False)
+    launchBenchmark(v_swmmac_i32_16x16x32_iu4, (7,8,8,9,10,11), False)
+    launchBenchmark(v_swmmac_f32_16x16x32_fp8_fp8, (7,8,9,10,13,14), False)
+    launchBenchmark(v_swmmac_f32_16x16x32_fp8_bf8, (7,8,9,10,13,14), False)
+    launchBenchmark(v_swmmac_f32_16x16x32_bf8_fp8, (7,8,9,10,13,14), False)
+    launchBenchmark(v_swmmac_f32_16x16x32_bf8_bf8, (7,8,9,10,13,14), False)
     FLOPS_PER_MATMUL = 16*16*64*2
-    launchBenchmark("v_swmmac_i32_16x16x64_iu4", (7,8,9,10,13,14), False)
-  elif DEV.arch == 'gfx950':
-    DIRECTIVE = ".amdhsa_accum_offset 4"
+    launchBenchmark(v_swmmac_i32_16x16x64_iu4, (7,8,9,10,13,14), False)
+  elif arch == 'gfx950':
+    from tinygrad.runtime.autogen.amd.cdna.ins import *
     NUM_WORKGROUPS = 256
     WAVE_SIZE = 64
     NUM_WAVES = 4
-    launchBenchmark("v_mfma_f32_16x16x16_f16", (3,0,1), accum=True)
-    launchBenchmark("v_mfma_f32_16x16x16_bf16", (3,0,1), accum=True)
+    launchBenchmark(v_mfma_f32_16x16x16_f16, (3,0,1), accum=True)
+    launchBenchmark(v_mfma_f32_16x16x16_bf16, (3,0,1), accum=True)
     FLOPS_PER_MATMUL = 16*16*32*2
-    launchBenchmark("v_mfma_f32_16x16x32_f16", (3,0,3), accum=True)
-    launchBenchmark("v_mfma_f32_16x16x32_bf16", (3,0,3), accum=True)
+    launchBenchmark(v_mfma_f32_16x16x32_f16, (3,0,3), accum=True)
+    launchBenchmark(v_mfma_f32_16x16x32_bf16, (3,0,3), accum=True)
     FLOPS_PER_MATMUL = 16*16*128*2
-    launchBenchmark("v_mfma_f32_16x16x128_f8f6f4", (3,0,7), accum=True) # fp8
-    launchBenchmark("v_mfma_f32_16x16x128_f8f6f4", (3,0,5), accum=True, extra=", cbsz:2 blgp:2") # fp6
-    launchBenchmark("v_mfma_f32_16x16x128_f8f6f4", (3,0,3), accum=True, extra=", cbsz:4 blgp:4") # fp4
+    launchBenchmark(v_mfma_f32_16x16x128_f8f6f4, (3,0,7), accum=True) # fp8
+    launchBenchmark(v_mfma_f32_16x16x128_f8f6f4, (3,0,5), accum=True, cbsz=2, blgp=2) # fp6
+    launchBenchmark(v_mfma_f32_16x16x128_f8f6f4, (3,0,3), accum=True, cbsz=4, blgp=4) # fp4
   else:
-    raise RuntimeError(f"arch {DEV.arch} not supported.")
+    raise RuntimeError(f"arch {arch} not supported.")

extra/mmapeak/template.s

@@ -1,40 +0,0 @@
-    .text
-    .globl matmul
-    .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
-
-.rodata
-.p2align 6
-.amdhsa_kernel matmul
-  .amdhsa_next_free_vgpr .amdgcn.next_free_vgpr
-  .amdhsa_next_free_sgpr .amdgcn.next_free_sgpr
-  DIRECTIVE
-.end_amdhsa_kernel
-
-.amdgpu_metadata
----
-amdhsa.version:
-  - 1
-  - 0
-amdhsa.kernels:
-  - .name: matmul
-    .symbol: matmul.kd
-    .kernarg_segment_size: 0
-    .group_segment_fixed_size: 0
-    .private_segment_fixed_size: 0
-    .kernarg_segment_align: 4
-    .wavefront_size: 32
-    .sgpr_count: 8
-    .vgpr_count: 32
-    .max_flat_workgroup_size: 1024
-...
-.end_amdgpu_metadata

extra/models/clip.py

@@ -5,7 +5,6 @@ from tinygrad.nn import Linear, LayerNorm, Embedding, Conv2d
 from typing import List, Optional, Union, Tuple, Dict
 from abc import ABC, abstractmethod
 from functools import lru_cache
-from PIL import Image
 import numpy as np
 import re, gzip
 
@@ -444,7 +443,8 @@ class OpenClipEncoder:
   # TODO:
   # Should be doable in pure tinygrad, would just require some work and verification.
   # This is very desirable since it would allow for full generation->evaluation in a single JIT call.
-  def prepare_image(self, image:Image.Image) -> Tensor:
+  def prepare_image(self, image) -> Tensor:
+    from PIL import Image
     SIZE = 224
     w, h = image.size
     scale = min(SIZE / h, SIZE / w)

extra/models/llama.py

@@ -41,9 +41,13 @@ class Attention:
     self.n_rep = self.n_heads // self.n_kv_heads
     self.max_context = max_context
 
-    self.wq = linear(dim, self.n_heads * self.head_dim, bias=False)
-    self.wk = linear(dim, self.n_kv_heads * self.head_dim, bias=False)
-    self.wv = linear(dim, self.n_kv_heads * self.head_dim, bias=False)
+    if getenv("WQKV"):
+      self.wqkv = linear(dim, self.n_heads * self.head_dim + self.n_kv_heads * self.head_dim * 2, bias=False)
+    else:
+      self.wq = linear(dim, self.n_heads * self.head_dim, bias=False)
+      self.wk = linear(dim, self.n_kv_heads * self.head_dim, bias=False)
+      self.wv = linear(dim, self.n_kv_heads * self.head_dim, bias=False)
+
     self.wo = linear(self.n_heads * self.head_dim, dim, bias=False)
 
     self.q_norm = nn.RMSNorm(dim, qk_norm) if qk_norm is not None else None
@@ -51,16 +55,21 @@ class Attention:
 
   def __call__(self, x:Tensor, start_pos:Union[Variable,int], freqs_cis:Tensor, mask:Optional[Tensor]=None) -> Tensor:
     if getenv("WQKV"):
-      if not hasattr(self, 'wqkv'): self.wqkv = Tensor.cat(self.wq.weight, self.wk.weight, self.wv.weight)
-      xqkv = x @ self.wqkv.T
-      xq, xk, xv = xqkv.split([self.wq.weight.shape[0], self.wk.weight.shape[0], self.wv.weight.shape[0]], dim=2)
+      xqkv = self.wqkv(x)
+      xqkv = xqkv.reshape(xqkv.shape[0], xqkv.shape[1], self.n_kv_heads, self.n_rep + 2, self.head_dim)
+      xq = xqkv[:, :, :, :self.n_rep].reshape(xqkv.shape[0], xqkv.shape[1], -1)
+      xk = xqkv[:, :, :, self.n_rep:self.n_rep+1].reshape(xqkv.shape[0], xqkv.shape[1], -1)
+      xv = xqkv[:, :, :, self.n_rep+1:self.n_rep+2].reshape(xqkv.shape[0], xqkv.shape[1], -1)
     else:
-      xq, xk, xv = self.wq(x), self.wk(x), self.wv(x)
+      xq, xk, xv = self.wq(x), self.wk(x.contiguous_backward()), self.wv(x)
 
     if self.q_norm is not None and self.k_norm is not None:
       xq = self.q_norm(xq)
       xk = self.k_norm(xk)
 
+    # cast_float_to_bf16 is expensive in reduction loops, break it out
+    if x.dtype == dtypes.bfloat16: xq, xk = xq.contiguous_backward(), xk.contiguous_backward()
+
     xq = xq.reshape(xq.shape[0], xq.shape[1], self.n_heads, self.head_dim)
     xk = xk.reshape(xk.shape[0], xk.shape[1], self.n_kv_heads, self.head_dim)
     xv = xv.reshape(xv.shape[0], xv.shape[1], self.n_kv_heads, self.head_dim)
@@ -86,20 +95,23 @@ class Attention:
       assert start_pos == 0
       keys, values = xk, xv
 
-    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 self.max_context:
+      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)
+    else:
+      xq, keys, values = xq.transpose(1, 2), keys.transpose(1, 2), values.transpose(1, 2)
+      attn = xq.scaled_dot_product_attention(keys, values, is_causal=True, enable_gqa=True).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]))
+        grad_q = Tensor.empty_like(q:=Tensor(kernel.src[2]))
+        grad_k = Tensor.empty_like(k:=Tensor(kernel.src[3]))
+        grad_v = Tensor.empty_like(v:=Tensor(kernel.src[4]))
         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]
@@ -194,12 +206,14 @@ class Transformer:
 
   def forward(self, tokens:Tensor, start_pos:Union[Variable,int], temperature:float, top_k:int, top_p:float, alpha_f:float, alpha_p:float):
     _bsz, seqlen = tokens.shape
-    h = self.tok_embeddings(tokens)
+    h = self.tok_embeddings(tokens).contiguous()
     freqs_cis = self.freqs_cis.cast(h.dtype)[:, start_pos:start_pos+seqlen, :, :, :]
 
-    mask = Tensor.full((1, 1, seqlen, start_pos+seqlen), float("-inf"), dtype=h.dtype, device=h.device).triu(start_pos+1) if seqlen > 1 else None
+    if self.max_context != 0 and seqlen > 1:
+      mask = Tensor.full((1, 1, seqlen, start_pos+seqlen), float("-inf"), dtype=h.dtype, device=h.device).triu(start_pos+1)
+    else: mask = None
     for layer in self.layers: h = layer(h, start_pos, freqs_cis, mask)
-    logits = self.output(self.norm(h))
+    logits = self.output(self.norm(h).contiguous().contiguous_backward()).contiguous_backward()
     if math.isnan(temperature): return logits
 
     return sample(logits[:, -1, :].flatten(), temperature, top_k, top_p, alpha_f, alpha_p)

extra/models/resnet.py

@@ -150,7 +150,7 @@ class ResNet:
         continue # Skip FC if transfer learning
 
       if 'bn' not in k and 'downsample' not in k: assert obj.shape == dat.shape, (k, obj.shape, dat.shape)
-      obj.assign(dat.to(obj.device).reshape(obj.shape))
+      obj.assign(dat.to(obj.device).cast(obj.dtype).reshape(obj.shape))
 
 ResNet18 = lambda num_classes=1000: ResNet(18, num_classes=num_classes)
 ResNet34 = lambda num_classes=1000: ResNet(34, num_classes=num_classes)

extra/models/rnnt.py

@@ -129,7 +129,7 @@ class LSTM:
       return self.do_step(x_, hc_)
 
     if hc is None:
-      hc = Tensor.zeros(self.layers, 2 * x.shape[1], self.hidden_size, requires_grad=False)
+      hc = Tensor.zeros(self.layers, 2 * x.shape[1], self.hidden_size, requires_grad=False).contiguous().realize()
 
     output = None
     for t in range(x.shape[0]):

extra/nv_gpu_driver/nv_ioctl.py

@@ -1,6 +1,7 @@
 # type: ignore
 import ctypes, ctypes.util, struct, platform, pathlib, re, time, os, signal
-from tinygrad.helpers import from_mv, to_mv, getenv, init_c_struct_t
+from tinygrad.helpers import from_mv, to_mv, getenv
+from tinygrad.runtime.support.c import init_c_struct_t
 from hexdump import hexdump
 start = time.perf_counter()
 
@@ -10,18 +11,21 @@ processor = platform.processor()
 IOCTL_SYSCALL = {"aarch64": 0x1d, "x86_64":16}[processor]
 MMAP_SYSCALL = {"aarch64": 0xde, "x86_64":0x09}[processor]
 
+IOCTL_PRINT = getenv("IOCTL_PRINT", getenv("IOCTL", 0))
+GRAB_PMA = getenv("GRAB_PMA", 0)
+
 def get_struct(argp, stype):
   return ctypes.cast(ctypes.c_void_p(argp), ctypes.POINTER(stype)).contents
 
 def dump_struct(st):
-  if getenv("IOCTL", 0) == 0: return
+  if IOCTL_PRINT == 0: return
   print("\t", st.__class__.__name__, end=" { ")
-  for v in type(st)._fields_: print(f"{v[0]}={getattr(st, v[0])}", end=" ")
+  for v in type(st)._real_fields_: print(f"{v[0]}={getattr(st, v[0])}", end=" ")
   print("}")
 
 def format_struct(s):
   sdats = []
-  for field in s._fields_:
+  for field in s._real_fields_:
     dat = getattr(s, field[0])
     if isinstance(dat, int): sdats.append(f"{field[0]}:0x{dat:X}")
     else: sdats.append(f"{field[0]}:{dat}")
@@ -58,6 +62,29 @@ def install_hook(c_function, python_function):
   return orig_func
 
 # *** ioctl lib end ***
+
+# PMA buffer tracking for raw PC sampling data (only when GRAB_PMA is enabled)
+pma_mem_handle = 0  # hMemPmaBuffer from ALLOC_PMA_STREAM
+pma_buffer_size = 0
+pma_buffer_va = 0   # actual mapped VA (found via /proc/self/maps)
+pma_get_offset = 0  # current read offset in ring buffer
+pma_pending_map = False  # flag to check for new mapping on next ioctl
+pma_maps_before = set()  # mappings before MAP_MEMORY
+pma_raw_dumps: list[bytes] = []  # raw PMA buffer dumps
+
+def get_pma_raw_dumps() -> list[bytes]: return pma_raw_dumps
+def clear_pma_raw_dumps(): pma_raw_dumps.clear()
+
+def get_proc_maps():
+  """Read current process memory mappings as set of (start, end) tuples."""
+  result = set()
+  with open("/proc/self/maps", "r") as f:
+    for line in f:
+      addr_range = line.split()[0]
+      start, end = addr_range.split("-")
+      result.add((int(start, 16), int(end, 16)))
+  return result
+
 from tinygrad.runtime.autogen import nv_570 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 \
@@ -69,6 +96,7 @@ def get_classes():
                    "NV20_SUBDEVICE_0"}
   for nm,val in nv_gpu.__dict__.items():
     if not isinstance(val, int): continue
+    if nm.endswith("PARAMETERS_MESSAGE_ID"): continue
     if 0x3000 < val < 0xffff: res[val] = nm
     if nm in known_classes: res[val] = nm
   return res
@@ -80,37 +108,92 @@ global_ioctl_id = 0
 gpus_user_modes = []
 gpus_mmio = []
 gpus_fifo = []
+offset_load = 0
 
 @ctypes.CFUNCTYPE(ctypes.c_int, ctypes.c_int, ctypes.c_ulong, ctypes.c_void_p)
 def ioctl(fd, request, argp):
   global global_ioctl_id, gpus_user_modes, gpus_mmio
+  global pma_mem_handle, pma_buffer_size, pma_buffer_va, pma_get_offset, pma_pending_map, pma_maps_before
   global_ioctl_id += 1
+
+  # Check for new PMA buffer mapping from previous MAP_MEMORY call (only when GRAB_PMA is enabled)
+  if GRAB_PMA and pma_pending_map:
+    pma_pending_map = False
+    new_maps = get_proc_maps()
+    for start, end in new_maps - pma_maps_before:
+      if end - start == pma_buffer_size:
+        pma_buffer_va = start
+        if IOCTL_PRINT >= 1: print(f"\t  PMA buffer mapped at CPU VA=0x{pma_buffer_va:x}")
+        break
+
   st = time.perf_counter()
   ret = libc.syscall(IOCTL_SYSCALL, ctypes.c_int(fd), ctypes.c_ulong(request), ctypes.c_void_p(argp))
   et = time.perf_counter()-st
   fn = os.readlink(f"/proc/self/fd/{fd}")
-  #print(f"ioctl {request:8x} {fn:20s}")
+
   idir, size, itype, nr = (request>>30), (request>>16)&0x3FFF, (request>>8)&0xFF, request&0xFF
-  if getenv("IOCTL", 0) >= 1: print(f"#{global_ioctl_id}: ", end="")
+  if IOCTL_PRINT >= 1: print(f"#{global_ioctl_id}: ", end="")
   if itype == ord(nv_gpu.NV_IOCTL_MAGIC):
     if nr == nv_gpu.NV_ESC_RM_CONTROL:
       s = get_struct(argp, nv_gpu.NVOS54_PARAMETERS)
       if s.cmd in nvcmds:
         name, struc = nvcmds[s.cmd]
-        if getenv("IOCTL", 0) >= 1:
+        if IOCTL_PRINT >= 1:
           print(f"NV_ESC_RM_CONTROL    cmd={name:30s} hClient={s.hClient}, hObject={s.hObject}, flags={s.flags}, params={s.params}, paramsSize={s.paramsSize}, status={s.status}")
 
         if struc is not None: dump_struct(get_struct(s.params, struc))
         elif hasattr(nv_gpu, name+"_PARAMS"): dump_struct(get_struct(argp, getattr(nv_gpu, name+"_PARAMS")))
         elif name == "NVA06C_CTRL_CMD_GPFIFO_SCHEDULE": dump_struct(get_struct(argp, nv_gpu.NVA06C_CTRL_GPFIFO_SCHEDULE_PARAMS))
         elif name == "NV83DE_CTRL_CMD_GET_MAPPINGS": dump_struct(get_struct(s.params, nv_gpu.NV83DE_CTRL_DEBUG_GET_MAPPINGS_PARAMETERS))
+        elif name == "NVB0CC_CTRL_CMD_SET_HS_CREDITS":
+          hs_params = get_struct(s.params, nv_gpu.NVB0CC_CTRL_SET_HS_CREDITS_PARAMS)
+          dump_struct(hs_params)
+          if IOCTL_PRINT >= 2:
+            for i in range(hs_params.numEntries):
+              print(f"\t\t", end="")
+              dump_struct(hs_params.creditInfo[i])
+
+        # PMA buffer tracking (only when GRAB_PMA is enabled)
+        if GRAB_PMA and name == "NVB0CC_CTRL_CMD_ALLOC_PMA_STREAM":
+          pma_params = get_struct(s.params, nv_gpu.struct_NVB0CC_CTRL_ALLOC_PMA_STREAM_PARAMS)
+          pma_mem_handle = pma_params.hMemPmaBuffer
+          pma_buffer_size = pma_params.pmaBufferSize
+          pma_get_offset = 0  # Reset read offset for new stream
+          if IOCTL_PRINT >= 1: print(f"\t  PMA buffer: hMem=0x{pma_mem_handle:x} size={pma_buffer_size}")
+        if GRAB_PMA and name == "NVB0CC_CTRL_CMD_PMA_STREAM_UPDATE_GET_PUT":
+          pma_update = get_struct(s.params, nv_gpu.struct_NVB0CC_CTRL_PMA_STREAM_UPDATE_GET_PUT_PARAMS)
+          if pma_update.bytesAvailable > 0 and pma_buffer_va and pma_buffer_size > 0:
+            avail = pma_update.bytesAvailable
+            read_offset = pma_get_offset
+            # Handle ring buffer wrap-around
+            if pma_get_offset + avail <= pma_buffer_size:
+              pma_data = bytes(to_mv(pma_buffer_va + pma_get_offset, avail))
+            else:
+              # Wrap around: read to end, then from start
+              first_part = pma_buffer_size - pma_get_offset
+              second_part = avail - first_part
+              pma_data = bytes(to_mv(pma_buffer_va + pma_get_offset, first_part)) + bytes(to_mv(pma_buffer_va, second_part))
+            pma_raw_dumps.append(pma_data)
+            pma_get_offset = (pma_get_offset + avail) % pma_buffer_size
+            if IOCTL_PRINT >= 2:
+              print(f"\t  PMA data: {avail} bytes from offset=0x{read_offset:x}, new offset=0x{pma_get_offset:x}")
+              hexdump(pma_data)
+
+        # Dump regOps for EXEC_REG_OPS when IOCTL >= 3
+        if name == "NVB0CC_CTRL_CMD_EXEC_REG_OPS" and struc is not None and IOCTL_PRINT >= 3:
+          reg_params = get_struct(s.params, struc)
+          for i in range(reg_params.regOpCount):
+            print(f"\t\t", end="")
+            dump_struct(reg_params.regOps[i])
+            # val = (op.regValueHi << 32) | op.regValueLo
+            # print(f"\t  regOps[{i:3d}]: op={op.regOp} type={op.regType} status={op.regStatus} offset=0x{op.regOffset:08x} value=0x{val:016x}")
       else:
-        if getenv("IOCTL", 0) >= 1: print("unhandled cmd", hex(s.cmd))
+        if IOCTL_PRINT >= 1: print("unhandled cmd", hex(s.cmd))
       # format_struct(s)
       # print(f"{(st-start)*1000:7.2f} ms +{et*1000.:7.2f} ms : {ret:2d} = {name:40s}", ' '.join(format_struct(s)))
     elif nr == nv_gpu.NV_ESC_RM_ALLOC:
       s = get_struct(argp, nv_gpu.NVOS21_PARAMETERS)
-      if getenv("IOCTL", 0) >= 1: print(f"NV_ESC_RM_ALLOC    hClass={nvclasses.get(s.hClass, f'unk=0x{s.hClass:X}'):30s}, hRoot={s.hRoot}, hObjectParent={s.hObjectParent}, pAllocParms={s.pAllocParms}, hObjectNew={s.hObjectNew} status={s.status}")
+      if IOCTL_PRINT >= 1: print(f"NV_ESC_RM_ALLOC    hClass={nvclasses.get(s.hClass, f'unk=0x{s.hClass:X}'):30s}, hRoot={s.hRoot}, hObjectParent={s.hObjectParent}, pAllocParms={s.pAllocParms}, hObjectNew={s.hObjectNew} status={s.status}")
       if s.pAllocParms is not None:
         if s.hClass == nv_gpu.NV01_DEVICE_0: dump_struct(get_struct(s.pAllocParms, nv_gpu.NV0080_ALLOC_PARAMETERS))
         if s.hClass == nv_gpu.FERMI_VASPACE_A: dump_struct(get_struct(s.pAllocParms, nv_gpu.NV_VASPACE_ALLOCATION_PARAMETERS))
@@ -118,7 +201,8 @@ def ioctl(fd, request, argp):
         if s.hClass == nv_gpu.NV1_MEMORY_USER: dump_struct(get_struct(s.pAllocParms, nv_gpu.NV_MEMORY_ALLOCATION_PARAMS))
         if s.hClass == nv_gpu.NV1_MEMORY_SYSTEM: dump_struct(get_struct(s.pAllocParms, nv_gpu.NV_MEMORY_ALLOCATION_PARAMS))
         if s.hClass == nv_gpu.GT200_DEBUGGER: dump_struct(get_struct(s.pAllocParms, nv_gpu.NV83DE_ALLOC_PARAMETERS))
-        if s.hClass == nv_gpu.AMPERE_CHANNEL_GPFIFO_A:
+        if s.hClass == nv_gpu.MAXWELL_PROFILER_DEVICE: dump_struct(get_struct(s.pAllocParms, nv_gpu.NVB2CC_ALLOC_PARAMETERS))
+        if s.hClass in {nv_gpu.AMPERE_CHANNEL_GPFIFO_A, nv_gpu.BLACKWELL_CHANNEL_GPFIFO_A}:
           sx = get_struct(s.pAllocParms, nv_gpu.NV_CHANNELGPFIFO_ALLOCATION_PARAMETERS)
           dump_struct(sx)
           gpus_fifo.append((sx.gpFifoOffset, sx.gpFifoEntries))
@@ -126,31 +210,35 @@ def ioctl(fd, request, argp):
       if s.hClass == nv_gpu.TURING_USERMODE_A: gpus_user_modes.append(s.hObjectNew)
     elif nr == nv_gpu.NV_ESC_RM_MAP_MEMORY:
       # nv_ioctl_nvos33_parameters_with_fd
-      if getenv("IOCTL", 0) >= 1:
-        s = get_struct(argp, nv_gpu.NVOS33_PARAMETERS)
+      s = get_struct(argp, nv_gpu.NVOS33_PARAMETERS)
+      if IOCTL_PRINT >= 1:
         print(f"NV_ESC_RM_MAP_MEMORY   hClient={s.hClient}, hDevice={s.hDevice}, hMemory={s.hMemory}, length={s.length} flags={s.flags} pLinearAddress={s.pLinearAddress}")
+      # Track PMA buffer mapping - save maps now, check for new mapping on next ioctl (after mmap happens)
+      if GRAB_PMA and pma_mem_handle and s.hMemory == pma_mem_handle:
+        pma_maps_before = get_proc_maps()
+        pma_pending_map = True
     elif nr == nv_gpu.NV_ESC_RM_UPDATE_DEVICE_MAPPING_INFO:
-      if getenv("IOCTL", 0) >= 1:
+      if IOCTL_PRINT >= 1:
         s = get_struct(argp, nv_gpu.NVOS56_PARAMETERS)
         print(f"NV_ESC_RM_UPDATE_DEVICE_MAPPING_INFO   hClient={s.hClient}, hDevice={s.hDevice}, hMemory={s.hMemory}, pOldCpuAddress={s.pOldCpuAddress} pNewCpuAddress={s.pNewCpuAddress} status={s.status}")
     elif nr == nv_gpu.NV_ESC_RM_ALLOC_MEMORY:
-      if getenv("IOCTL", 0) >= 1:
+      if IOCTL_PRINT >= 1:
         s = get_struct(argp, nv_gpu.nv_ioctl_nvos02_parameters_with_fd)
         print(f"NV_ESC_RM_ALLOC_MEMORY  fd={s.fd}, hRoot={s.params.hRoot}, hObjectParent={s.params.hObjectParent}, hObjectNew={s.params.hObjectNew}, hClass={s.params.hClass}, flags={s.params.flags}, pMemory={s.params.pMemory}, limit={s.params.limit}, status={s.params.status}")
     elif nr == nv_gpu.NV_ESC_ALLOC_OS_EVENT:
-      if getenv("IOCTL", 0) >= 1:
+      if IOCTL_PRINT >= 1:
         s = get_struct(argp, nv_gpu.nv_ioctl_alloc_os_event_t)
         print(f"NV_ESC_ALLOC_OS_EVENT  hClient={s.hClient} hDevice={s.hDevice} fd={s.fd} Status={s.Status}")
     elif nr == nv_gpu.NV_ESC_REGISTER_FD:
-      if getenv("IOCTL", 0) >= 1:
+      if IOCTL_PRINT >= 1:
         s = get_struct(argp, nv_gpu.nv_ioctl_register_fd_t)
         print(f"NV_ESC_REGISTER_FD  fd={s.ctl_fd}")
     elif nr in nvescs:
-      if getenv("IOCTL", 0) >= 1: print(nvescs[nr])
+      if IOCTL_PRINT >= 1: print(nvescs[nr])
     else:
-      if getenv("IOCTL", 0) >= 1: print("unhandled NR", nr)
+      if IOCTL_PRINT >= 1: print("unhandled NR", nr)
   elif fn.endswith("nvidia-uvm"):
-    if getenv("IOCTL", 0) >= 1:
+    if IOCTL_PRINT >= 1:
       print(f"{nvuvms.get(request, f'UVM UNKNOWN {request=}')}")
       if nvuvms.get(request) is not None: dump_struct(get_struct(argp, getattr(nv_gpu, nvuvms.get(request)+"_PARAMS")))
       if nvuvms.get(request) == "UVM_MAP_EXTERNAL_ALLOCATION":
@@ -159,7 +247,7 @@ def ioctl(fd, request, argp):
           print("perGpuAttributes[{i}] = ", end="")
           dump_struct(st.perGpuAttributes[i])
 
-  if getenv("IOCTL") >= 2: print("ioctl", f"{idir=} {size=} {itype=} {nr=} {fd=} {ret=}", fn)
+  if IOCTL_PRINT >= 2: print("ioctl", f"{idir=} {size=} {itype=} {nr=} {fd=} {ret=}", fn)
   return ret
 
 @ctypes.CFUNCTYPE(ctypes.c_void_p, ctypes.c_void_p, ctypes.c_size_t, ctypes.c_int, ctypes.c_int, ctypes.c_int, ctypes.c_long)
@@ -172,14 +260,14 @@ def _mmap(addr, length, prot, flags, fd, offset):
   return ret
 
 install_hook(libc.ioctl, ioctl)
-if getenv("IOCTL") >= 3: orig_mmap_mv = install_hook(libc.mmap, _mmap)
+if getenv("IOCTL") >= 4: orig_mmap_mv = install_hook(libc.mmap, _mmap)
 
 import collections
 old_gpputs = collections.defaultdict(int)
 def _dump_gpfifo(mark):
   launches = []
 
-  # print("_dump_gpfifo:", mark)
+  print("_dump_gpfifo:", mark)
   for start, size in gpus_fifo:
     gpfifo_controls = nv_gpu.AmpereAControlGPFifo.from_address(start+size*8)
     gpfifo = to_mv(start, size * 8).cast("Q")
@@ -205,7 +293,7 @@ def make_qmd_struct_type():
     fields.append((name.replace("NVC6C0_QMDV03_00_", "").lower(), ctypes.c_uint32, data[0]-data[1]+1))
     if len(fields) >= 2 and fields[-2][0].endswith('_lower') and fields[-1][0].endswith('_upper') and fields[-1][0][:-6] == fields[-2][0][:-6]:
       fields = fields[:-2] + [(fields[-1][0][:-6], ctypes.c_uint64, fields[-1][2] + fields[-2][2])]
-  return init_c_struct_t(tuple(fields))
+  return init_c_struct_t(0x40 * 4, tuple(fields))
 qmd_struct_t = make_qmd_struct_type()
 assert ctypes.sizeof(qmd_struct_t) == 0x40 * 4
 
@@ -222,7 +310,7 @@ def _dump_qmd(address, packets):
     subc = (dat>>13) & 7
     mthd = (dat<<2) & 0x7FFF
     method_name = nvqcmds.get(mthd, f"unknown method #{mthd}")
-    if getenv("IOCTL", 0) >= 1:
+    if IOCTL_PRINT >= 1:
       print(f"\t\t{method_name}, {typ=} {size=} {subc=} {mthd=}")
       for j in range(size): print(f"\t\t\t{j}: {gpfifo[i+j+1]} | 0x{gpfifo[i+j+1]:x}")
     if mthd == 792:

extra/nv_pma/.gitignore

@@ -0,0 +1 @@
+examples/

extra/nv_pma/collect.py

@@ -0,0 +1,135 @@
+import pickle, os, sys, functools, numpy as np
+from pathlib import Path
+
+os.environ["DEV"] = "CUDA"
+os.environ["PROFILE"] = os.environ.get("PROFILE", "2")
+from extra.nv_pma.cupti import cu_prof_ext
+cu_prof_ext.enable_auto()
+
+from tinygrad import Tensor, Device
+
+if not os.environ.get("IOCTL") or not os.environ.get("GRAB_PMA"):
+  print("Usage: GRAB_PMA=1 IOCTL=1 IOCTL_PRINT=0 python3 extra/nv_pma/collect.py")
+  sys.exit(1)
+
+assert Device.DEFAULT == "CUDA", "only works with CUDA"
+
+EXAMPLES_DIR = Path(__file__).parent / "examples"
+_collectors: list[tuple[str, callable]] = []
+
+def pcsampling_test(name: str):
+  def decorator(fn):
+    @functools.wraps(fn)
+    def wrapper():
+      cu_prof_ext.clear_pma_raw_dumps()
+      cu_prof_ext.clear_cupti_pc_samples()
+
+      fn()
+      Device["CUDA"].synchronize()
+
+      dumps = cu_prof_ext.get_pma_raw_dumps()
+      # from hexdump import hexdump
+      # hexdump(dumps[0][:0x40])
+
+      return {"test_name": name, "pma_raw_dumps": list(cu_prof_ext.get_pma_raw_dumps()), "cupti_pc_samples": list(cu_prof_ext.get_cupti_pc_samples())}
+    _collectors.append((name, wrapper))
+    return wrapper
+  return decorator
+
+# Refs
+
+@pcsampling_test("test_plus")
+def test_plus():
+  a = Tensor([1, 2, 3, 4])
+  b = Tensor([5, 6, 7, 8])
+  (a + b).realize()
+
+@pcsampling_test("test_matmul")
+def test_matmul():
+  a = Tensor(np.random.rand(12, 12).astype(np.float32))
+  b = Tensor(np.random.rand(12, 12).astype(np.float32))
+  (a @ b).realize()
+
+@pcsampling_test("test_reduce_sum")
+def test_reduce_sum():
+  a = Tensor(np.random.rand(1024).astype(np.float32))
+  a.sum().realize()
+
+@pcsampling_test("test_reduce_max")
+def test_reduce_max():
+  a = Tensor(np.random.rand(1024).astype(np.float32))
+  a.max().realize()
+
+@pcsampling_test("test_exp")
+def test_exp():
+  a = Tensor(np.random.rand(256).astype(np.float32))
+  a.exp().realize()
+
+@pcsampling_test("test_softmax")
+def test_softmax():
+  a = Tensor(np.random.rand(64, 64).astype(np.float32))
+  a.softmax().realize()
+
+@pcsampling_test("test_conv2d")
+def test_conv2d():
+  x = Tensor(np.random.rand(1, 3, 32, 32).astype(np.float32))
+  w = Tensor(np.random.rand(8, 3, 3, 3).astype(np.float32))
+  x.conv2d(w).realize()
+
+@pcsampling_test("test_large_matmul")
+def test_large_matmul():
+  a = Tensor(np.random.rand(128, 128).astype(np.float32))
+  b = Tensor(np.random.rand(128, 128).astype(np.float32))
+  (a @ b).realize()
+
+@pcsampling_test("test_elementwise_chain")
+def test_elementwise_chain():
+  a = Tensor(np.random.rand(512).astype(np.float32))
+  ((a + 1) * 2 - 0.5).relu().realize()
+
+@pcsampling_test("test_broadcast")
+def test_broadcast():
+  a = Tensor(np.random.rand(64, 1).astype(np.float32))
+  b = Tensor(np.random.rand(1, 64).astype(np.float32))
+  (a + b).realize()
+
+@pcsampling_test("test_plus_big")
+def test_plus_big():
+  a = Tensor(np.random.rand(64, 32).astype(np.float32))
+  b = Tensor(np.random.rand(64, 32).astype(np.float32))
+  (a + b).realize()
+
+def save_example(name: str, data: dict):
+  pma_bytes = sum(len(d) for d in data['pma_raw_dumps'])
+  cupti_samples = sum(r['samples'] for r in data['cupti_pc_samples'])
+  print(f"  PMA: {len(data['pma_raw_dumps'])} buffers, {pma_bytes} bytes")
+  print(f"  CUPTI: {len(data['cupti_pc_samples'])} records, {cupti_samples} samples")
+
+  outfile = EXAMPLES_DIR / f"{name}.pkl"
+  with open(outfile, "wb") as f:
+    pickle.dump(data, f)
+  print(f"  Saved to {outfile}")
+
+if __name__ == "__main__":
+  EXAMPLES_DIR.mkdir(exist_ok=True)
+
+  # Run specific tests if provided as arguments, otherwise run all
+  if len(sys.argv) > 1:
+    test_names = sys.argv[1:]
+    collectors = [(name, fn) for name, fn in _collectors if name in test_names]
+    if not collectors:
+      print(f"Unknown tests: {test_names}")
+      print(f"Available: {[name for name, _ in _collectors]}")
+      sys.exit(1)
+  else:
+    collectors = _collectors
+
+  for name, collect_fn in collectors:
+    print(f"\nCollecting {name}...")
+    try:
+      data = collect_fn()
+      save_example(name, data)
+    except Exception as e:
+      print(f"  ERROR: {e}")
+      import traceback
+      traceback.print_exc()

extra/nv_pma/cupti/__init__.py

@@ -0,0 +1,26 @@
+# CUPTI autogen loader for nv_pma
+# To regenerate: REGEN=1 python -c "import extra.nv_pma.cupti"
+import importlib, pathlib
+from tinygrad.helpers import getenv
+
+root = pathlib.Path(__file__).parents[3]
+here = pathlib.Path(__file__).parent
+
+def load(name, dll, files, **kwargs):
+  if not (f:=here/f"{name}.py").exists() or getenv('REGEN'):
+    kwargs['args'] = kwargs.get('args', [])
+    f.write_text(importlib.import_module("tinygrad.runtime.support.autogen").gen(name, dll, files, **kwargs))
+  return importlib.import_module(f"extra.nv_pma.cupti.{name}")
+
+def __getattr__(nm):
+  match nm:
+    case "cupti":
+      return load("cupti", "'/usr/local/cuda/targets/x86_64-linux/lib/libcupti.so'", [
+        "/usr/local/cuda/include/cupti_result.h", "/usr/local/cuda/include/cupti_activity.h",
+        "/usr/local/cuda/include/cupti_callbacks.h", "/usr/local/cuda/include/cupti_events.h",
+        "/usr/local/cuda/include/cupti_metrics.h", "/usr/local/cuda/include/cupti_driver_cbid.h",
+        "/usr/local/cuda/include/cupti_runtime_cbid.h", "/usr/local/cuda/include/cupti_profiler_target.h",
+        "/usr/local/cuda/include/cupti_profiler_host.h", "/usr/local/cuda/include/cupti_pmsampling.h",
+        "/usr/local/cuda/include/generated_cuda_meta.h", "/usr/local/cuda/include/generated_cuda_runtime_api_meta.h"
+      ], args=["-D__CUDA_API_VERSION_INTERNAL", "-I/usr/local/cuda/include"], parse_macros=False)
+    case _: raise AttributeError(f"no such autogen: {nm}")