ds compiled

.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 && CHECK_OOB=0 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 && IGNORE_OOB=1 PYTHONPATH=. python3 process_replay.py
         git checkout $CURRENT_HEAD  # restore to branch

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

@@ -45,10 +45,6 @@ inputs:
     description: "Install mesa"
     required: false
     default: 'false'
-  tinydreno:
-    description: "Install tinydreno"
-    required: false
-    default: 'false'
 runs:
   using: "composite"
   steps:
@@ -60,40 +56,32 @@ 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 }}-${{ runner.arch }}-python-${{ 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:
         path: ${{ github.workspace }}/.venv
-        key: venv-${{ runner.os }}-${{ runner.arch }}-python-${{ steps.setup-python.outputs.python-version }}-${{ inputs.deps }}-${{ inputs.pydeps }}-${{ env.CACHE_VERSION }}
+        key: venv-${{ runner.os }}-python-${{ steps.setup-python.outputs.python-version }}-${{ inputs.deps }}-${{ inputs.pydeps }}-${{ env.CACHE_VERSION }}
 
     # **** Caching downloads ****
 
-    - name: Cache downloads (PR)
-      if: inputs.key != '' && github.event_name == 'pull_request'
-      uses: actions/cache/restore@v4
-      with:
-        path: ${{ runner.os == 'Linux' && '~/.cache/tinygrad/downloads/' || '~/Library/Caches/tinygrad/downloads/' }}
-        key: downloads-${{ github.job }}-${{ inputs.key }}-${{ env.CACHE_VERSION }}
-    - name: Cache downloads
-      if: inputs.key != '' && github.event_name != 'pull_request'
+    - name: Cache downloads (Linux)
+      if: inputs.key != '' && runner.os == 'Linux'
       uses: actions/cache@v4
       with:
-        path: ${{ runner.os == 'Linux' && '~/.cache/tinygrad/downloads/' || '~/Library/Caches/tinygrad/downloads/' }}
+        path: ~/.cache/tinygrad/downloads/
+        key: downloads-${{ github.job }}-${{ inputs.key }}-${{ env.CACHE_VERSION }}
+    - name: Cache downloads (macOS)
+      if: inputs.key != '' && runner.os == 'macOS'
+      uses: actions/cache@v4
+      with:
+        path: ~/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-pr.outputs.cache-hit != 'true' && steps.restore-venv.outputs.cache-hit != 'true'
+      if: inputs.deps != '' && steps.restore-venv.outputs.cache-hit != 'true'
       shell: bash
       run: |
         python -m venv .venv
@@ -104,7 +92,7 @@ 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-pr.outputs.cache-hit != 'true' && steps.restore-venv.outputs.cache-hit != 'true'
+      if: inputs.deps == '' && steps.restore-venv.outputs.cache-hit != 'true'
       shell: bash
       run: |
         python -m venv .venv
@@ -149,7 +137,7 @@ 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/7.1 $(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
 
@@ -194,18 +182,12 @@ 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: ${{ runner.os }}-${{ runner.arch }}-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') && github.event_name != 'pull_request'
+      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.cuda == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true')
       uses: actions/cache@v4
       with:
         path: /var/cache/apt/archives/
-        key: ${{ runner.os }}-${{ runner.arch }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.CACHE_VERSION }}
+        key: ${{ runner.os }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.CACHE_VERSION }}
 
     - name: Run apt Update + Install
       if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.cuda == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true')
@@ -237,7 +219,7 @@ runs:
       shell: bash
       run: |
         sudo mkdir -p /usr/local/lib
-        curl -s -H "Authorization: token $GH_TOKEN" curl -s https://api.github.com/repos/tinygrad/amdcomgr_dylib/releases/latest | \
+        curl -s -H "Authorization: token $GH_TOKEN" curl -s https://api.github.com/repos/nimlgen/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
@@ -257,17 +239,8 @@ 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' && github.event_name != 'pull_request'
+      if: inputs.ocelot == 'true'
       id: cache-build
       uses: actions/cache@v4
       env:
@@ -276,7 +249,7 @@ 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-pr.outputs.cache-hit != 'true' && steps.cache-build.outputs.cache-hit != 'true'
+      if: inputs.ocelot == 'true' && steps.cache-build.outputs.cache-hit != 'true'
       shell: bash
       run: |
         git clone --recurse-submodules https://github.com/gpuocelot/gpuocelot.git ${{ github.workspace }}/gpuocelot
@@ -287,7 +260,6 @@ runs:
 
         CMAKE_ARGS="-Wno-dev -G Ninja -DOCELOT_BUILD_TOOLS=OFF -DCMAKE_BUILD_ALWAYS=0 -DBUILD_TESTS_CUDA=OFF -DCMAKE_POLICY_VERSION_MINIMUM=3.5"
         if [[ "${{ runner.os }}" == "macOS" ]]; then
-          sudo xcode-select -s /Applications/Xcode_16.2.app/Contents/Developer
           CMAKE_ARGS="$CMAKE_ARGS -DBoost_INCLUDE_DIR=$(brew --prefix boost)/include -DBoost_LIBRARY_DIR=$(brew --prefix boost)/lib"
         fi
 
@@ -331,9 +303,3 @@ runs:
       if: inputs.mesa == 'true' && runner.os == 'macOS'
       shell: bash
       run: brew install sirhcm/tinymesa/tinymesa_cpu
-
-    # *** tinydreno ***
-    - name: Install tinydreno (linux)
-      if: inputs.tinydreno == 'true' && runner.os == 'Linux'
-      shell: bash
-      run: sudo curl -fL https://github.com/sirhcm/tinydreno/raw/refs/heads/master/libllvm-qcom.so -o /usr/lib/libllvm-qcom.so

.github/workflows/autogen.yml

@@ -14,12 +14,10 @@ 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:
@@ -32,7 +30,6 @@ jobs:
     - name: Setup Environment
       uses: ./.github/actions/setup-tinygrad
       with:
-        key: 'autogen'
         opencl: 'true'
         amd: 'true'
         cuda: 'true'
@@ -41,38 +38,103 @@ 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 libdrm-dev
-    - name: Regenerate autogen files
+      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: |
-        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
+        mv tinygrad/runtime/autogen/opencl.py /tmp/opencl.py.bak
         python3 -c "from tinygrad.runtime.autogen import opencl"
-        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"
-        python3 -c "from tinygrad.runtime.autogen import llvm"
-        python3 -c "from tinygrad.runtime.autogen import webgpu"
-        python3 -c "from tinygrad.runtime.autogen import kgsl, qcom_dsp"
-        python3 -c "from tinygrad.runtime.autogen import libusb"
-        python3 -c "from tinygrad.runtime.autogen import mesa"
-        python3 -c "from tinygrad.runtime.autogen import avcodec"
-        python3 -c "from tinygrad.runtime.autogen import llvm_qcom"
-        REGEN=1 python3 -c "from tinygrad.runtime.autogen import libclang"
-    - name: Check for differences
+        diff /tmp/opencl.py.bak tinygrad/runtime/autogen/opencl.py
+    - name: Verify CUDA autogen
       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
-
+        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 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
+        REGEN=1 python3 -c "from tinygrad.runtime.autogen import libclang"
+        diff /tmp/libclang.py.bak tinygrad/runtime/autogen/libclang.py
   autogen-mac:
     name: In-tree Autogen (macos)
     runs-on: macos-14
@@ -83,29 +145,14 @@ jobs:
     - name: Setup Environment
       uses: ./.github/actions/setup-tinygrad
       with:
-        key: 'autogen-mac'
         llvm: 'true'
-    - name: Regenerate autogen files
+    - name: Verify macos autogen
       run: |
-        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)
+        mv tinygrad/runtime/autogen/metal.py /tmp/metal.py.bak
+        LIBCLANG_PATH=/opt/homebrew/opt/llvm@20/lib/libclang.dylib python3 -c "from tinygrad.runtime.autogen import metal"
+        diff /tmp/metal.py.bak tinygrad/runtime/autogen/metal.py
+  autogen-comgr-3:
+    name: In-tree Autogen (comgr 3)
     runs-on: ubuntu-24.04
     timeout-minutes: 15
     steps:
@@ -113,32 +160,17 @@ 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/6.2 $(lsb_release -cs) main
+        deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/6.4 $(lsb_release -cs) main
         EOF
         echo -e 'Package: *\nPin: release o=repo.radeon.com\nPin-Priority: 600' | sudo tee /etc/apt/preferences.d/rocm-pin-600
         sudo apt -qq update || true
         sudo apt-get install -y --no-install-recommends libclang-20-dev comgr
-    - name: Regenerate autogen files
+    - name: Verify comgr (3) autogen
       run: |
-        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
+        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

.github/workflows/benchmark.yml

@@ -16,41 +16,6 @@ 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=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:
@@ -84,19 +49,19 @@ 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
+      run: BENCHMARK_LOG=stable_diffusion JIT=1 ASSERT_MIN_STEP_TIME=720 python3.11 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
     - name: Run Stable Diffusion without fp16
-      run: BENCHMARK_LOG=stable_diffusion_fp32 JIT=1 ASSERT_MIN_STEP_TIME=720 python3.11 examples/stable_diffusion.py --seed 0 --noshow --timing
+      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
     - 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
+      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
     # 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
+      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
     - 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
+      run: BIG=2 MPS=1 python3.11 test/speed/external_test_speed_v_torch.py | tee torch_speed.txt
     - name: Test tensor cores
       run: METAL=1 python3.11 test/opt/test_tensor_cores.py
     - name: Test AMX tensor cores
@@ -106,59 +71,84 @@ 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
+      run: DEBUG=2 SHOULD_USE_TC=1 python3.11 extra/gemm/simple_matmul.py | tee matmul.txt
     - name: Run Tensor Core GEMM (half)
-      run: DEBUG=2 SHOULD_USE_TC=1 HALF=1 python3.11 extra/gemm/simple_matmul.py
+      run: DEBUG=2 SHOULD_USE_TC=1 HALF=1 python3.11 extra/gemm/simple_matmul.py | tee matmul_half.txt
     - name: Run Tensor Core GEMM (bfloat16)
-      run: DEBUG=2 SHOULD_USE_TC=1 BFLOAT16=1 python3.11 extra/gemm/simple_matmul.py
+      run: DEBUG=2 SHOULD_USE_TC=1 BFLOAT16=1 python3.11 extra/gemm/simple_matmul.py | tee matmul_bfloat16.txt
     - 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
-        BENCHMARK_LOG=llama JIT=1 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
+        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
     - 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
+      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
     - 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
-        BENCHMARK_LOG=llama_nf4 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing --quantize nf4
+        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
     - name: Run quantized LLaMA3
       run: |
-        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
+        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
     #- 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
+    #  run: python3.11 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_four_gpu.txt
     - name: Run GPT2
       run: |
-        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
+        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
     - name: Run GPT2 w HALF
-      run: BENCHMARK_LOG=gpt2_half HALF=1 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing
+      run: BENCHMARK_LOG=gpt2_half HALF=1 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
     - 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
+      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
     - 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
+      run: time PYTHONPATH=. TARGET_EVAL_ACC_PCT=96.0 python3.11 examples/beautiful_mnist.py | tee beautiful_mnist.txt
 
     # 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
+    #  run: BENCHMARK_LOG=cifar_10steps JIT=1 ASSERT_MIN_STEP_TIME=3000 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar.txt
     #- 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
+    #  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
 
     #- name: Run 10 CIFAR training steps w BF16
-    #  run: STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3.11 examples/hlb_cifar10.py
+    #  run: STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3.11 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
     # TODO: too slow
     # - name: Run 10 CIFAR training steps w winograd
-    #   run: BENCHMARK_LOG=cifar_10steps_wino JIT=1 ASSERT_MIN_STEP_TIME=150 WINO=1 STEPS=10 python3.11 examples/hlb_cifar10.py
+    #   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
     - 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
 
@@ -180,18 +170,14 @@ 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=. GMMU=0 DEBUG=2 AM_RESET=1 AMD=1 AMD_IFACE=USB time python3.11 test/test_tiny.py TestTiny.test_plus
+      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
-      run: sudo -E PYTHONPATH=. GMMU=0 AMD=1 AMD_IFACE=USB python3.11 test/test_tiny.py
+      run: sudo -E PYTHONPATH=. AMD=1 AMD_IFACE=USB python3.11 test/test_tiny.py
     - name: UsbGPU copy speeds
-      run: sudo -E PYTHONPATH=. GMMU=0 AMD=1 AMD_IFACE=USB python3.11 test/external/external_test_usb_asm24.py TestDevCopySpeeds
+      run: sudo -E PYTHONPATH=. AMD=1 AMD_IFACE=USB python3.11 test/external/external_test_usb_asm24.py TestDevCopySpeeds
     #- name: UsbGPU openpilot test
-    #  run: sudo -E PYTHONPATH=. GMMU=0 AMD=1 AMD_IFACE=USB GRAPH_ONE_KERNEL=1 python3.11 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/9118973ed03c1ae1d40cf69a29507ec2cc78efd7/selfdrive/modeld/models/supercombo.onnx
+    #  run: sudo -E PYTHONPATH=. AMD=1 AMD_IFACE=USB GRAPH_ONE_KERNEL=1 python3.11 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/9118973ed03c1ae1d40cf69a29507ec2cc78efd7/selfdrive/modeld/models/supercombo.onnx
     - name: UsbGPU (USB4/TB) boot time
       run: PYTHONPATH=. DEBUG=3 NV=1 NV_IFACE=PCI NV_NAK=1 time python3.11 test/test_tiny.py TestTiny.test_plus
     - name: UsbGPU (USB4/TB) tiny tests
@@ -229,7 +215,7 @@ 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
+      run: NV=1 CAPTURE_PROCESS_REPLAY=0 HALF=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py | tee torch_speed.txt
     - name: Test speed vs theoretical
       run: NV=1 IGNORE_BEAM_CACHE=1 CCACHE=0 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
     - name: Test benchmark allreduce
@@ -240,58 +226,79 @@ 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
-        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
+        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
     - 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
+      run: NV=1 NV_PTX=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_ptx.txt
     - name: Run Tensor Core GEMM (NV)
-      run: NV=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+      run: NV=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_nv.txt
     - 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
+      run: BENCHMARK_LOG=stable_diffusion NV=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
     # 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
+    #   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
     - 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
-        BENCHMARK_LOG=llama NV=1 JIT=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
+        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
     - 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
+      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
     # - 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
+    #   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
     # - 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
+    #   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
     - 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
+      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
     - 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
+      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
     - 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
+      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
     # - 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
+    #   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
     # - 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
+    #   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
     - name: Run Mixtral 8x7B
-      run: time BENCHMARK_LOG=mixtral NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/mixtral.py --temperature 0 --count 10 --timing
+      run: time BENCHMARK_LOG=mixtral NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/mixtral.py --temperature 0 --count 10 --timing | tee mixtral.txt
     - name: Run GPT2
       run: |
-        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
+        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
     - 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
+      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
     - 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
+      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
     - 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
 
@@ -330,30 +337,44 @@ 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 ASSERT_FPS=1400 JITBEAM=1 NV=1 PYTHONPATH=. python3 extra/hevc/decode.py
+      run: VALIDATE=1 MAX_FRAMES=100 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
+      run: time PYTHONPATH=. NV=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py | tee beautiful_mnist.txt
     - 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
+      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=120 NV=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
     - name: Run 10 CIFAR training steps w HALF
-      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=120 NV=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
+      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
     - 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
+      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
     # - name: Run 10 CIFAR training steps w winograd
-    #   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
+    #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=350 NV=1 CAPTURE_PROCESS_REPLAY=0 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_wino.txt
     - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
+      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_one_gpu.txt
     - name: Run full CIFAR training steps w 6 GPUS
-      run: time BENCHMARK_LOG=cifar_6gpu CAPTURE_PROCESS_REPLAY=0 NV=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
+      run: time BENCHMARK_LOG=cifar_6gpu CAPTURE_PROCESS_REPLAY=0 NV=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu.txt
     - name: Run MLPerf resnet eval on training data
       run: time BENCHMARK_LOG=resnet_eval NV=1 MODEL=resnet python3 examples/mlperf/model_eval.py
     - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-      run: BENCHMARK_LOG=resnet_10steps NV=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
+      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
     - 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
+      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
     - 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
+      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
     - 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
 
@@ -368,12 +389,10 @@ 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: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd rmmod
-    - name: Kill stale pids
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
+    - name: Remove amdgpu
+      run: sudo rmmod amdgpu || true
+    - name: Cleanup running AM processes
+      run: python extra/amdpci/am_smi.py --pids --kill
     #- name: Insert amdgpu
     #  run: sudo modprobe amdgpu
     - name: Symlink models and datasets
@@ -407,7 +426,7 @@ 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
+    #    LD_PRELOAD="/opt/rocm/lib/libhsa-runtime64.so" HSA=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py | tee torch_speed.txt
     - name: Test speed vs theoretical
       run: AMD=1 IGNORE_BEAM_CACHE=1 CCACHE=0 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
     - name: Test tensor cores AMD_LLVM=0
@@ -418,7 +437,7 @@ 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
+        AMD=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py | tee matmul_amd.txt
     - name: Test AMD=1
       run: DEBUG=2 AMD=1 python -m pytest -rA test/test_tiny.py
     #- name: Test HIP=1
@@ -433,39 +452,61 @@ 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
+      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
     - 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
+      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
     - 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
-        BENCHMARK_LOG=llama AMD=1 JIT=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
+        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
     - 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
+      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
     # - 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
+    #   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
     # - 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
+    #   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
     - 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
+      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
     - 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
+      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
     # - 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
+    #   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
     #- 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
+    #   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
     - name: Run Mixtral 8x7B
-      run: time BENCHMARK_LOG=mixtral AMD=1 python3 examples/mixtral.py --temperature 0 --count 10 --timing
+      run: time BENCHMARK_LOG=mixtral AMD=1 python3 examples/mixtral.py --temperature 0 --count 10 --timing | tee mixtral.txt
     - name: Run GPT2
       run: |
-        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
+        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
     - 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
+      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
     - 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
+      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
     - 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
 
@@ -480,12 +521,10 @@ 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: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd rmmod
-    - name: Kill stale pids
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
+    - name: Remove amdgpu
+      run: sudo rmmod amdgpu || true
+    - name: Cleanup running AM processes
+      run: python extra/amdpci/am_smi.py --pids --kill
     - name: Symlink models and datasets
       run: |
         mkdir -p weights
@@ -504,23 +543,31 @@ 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
+      run: time PYTHONPATH=. AMD=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py | tee beautiful_mnist.txt
     - 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
+      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=200 AMD=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
     - name: Run 10 CIFAR training steps w HALF
-      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=230 AMD=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
+      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
     # - 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
+    #   run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=288 AMD=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
     # TODO: too slow
     # - name: Run 10 CIFAR training steps w winograd
-    #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=66 AMD=1 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
+    #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=66 AMD=1 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_wino.txt
     - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
+      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_one_gpu.txt
     - name: Run full CIFAR training steps w 6 GPUS
-      run: time BENCHMARK_LOG=cifar_6gpu AMD=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
-    # 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
+      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
     - 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
 
@@ -535,12 +582,10 @@ 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: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd rmmod
-    - name: Kill stale pids
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
+    - name: Remove amdgpu
+      run: sudo rmmod amdgpu || true
+    - name: Cleanup running AM processes
+      run: python extra/amdpci/am_smi.py --pids --kill
     - name: Symlink models and datasets
       run: |
         mkdir -p weights
@@ -561,12 +606,19 @@ 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
+      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
     - 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
+      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
     - 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
+      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
     - 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
 
@@ -588,22 +640,20 @@ jobs:
         rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
     - name: reset process replay
       run: test/external/process_replay/reset.py
-    - name: openpilot compile3 0.11.0 driving_vision
-      run: BENCHMARK_LOG=openpilot_0_11_0_vision PYTHONPATH="." ASSERT_MIN_STEP_TIME=17 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/driving_vision.onnx
-    - name: IR3 openpilot compile3 0.11.0 driving_vision
-      run: BENCHMARK_LOG=ir3_openpilot_0_11_0_vision PYTHONPATH="." ASSERT_MIN_STEP_TIME=17 DEV=QCOM QCOM_IR3=1 FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/driving_vision.onnx
-    - name: openpilot compile3 0.11.0 driving_policy
-      run: BENCHMARK_LOG=openpilot_0_11_0_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=3 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/driving_policy.onnx
-    - name: openpilot compile3 0.11.0 dmonitoring
-      run: BENCHMARK_LOG=openpilot_0_11_0_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=11 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/dmonitoring_model.onnx
+    - name: openpilot compile3 0.10.0 driving_policy
+      run: BENCHMARK_LOG=openpilot_0_10_0_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=4 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.10.0/selfdrive/modeld/models/driving_policy.onnx
+    - 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: 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=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
+      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=3 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_policy.onnx
+      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
     - name: openpilot compile3 0.10.1 dmonitoring
-      run: BENCHMARK_LOG=openpilot_0_10_1_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=11 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/dmonitoring_model.onnx
+      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
     - name: benchmark MobileNetV2 on DSP
       run: |
         # generate quantized weights
@@ -615,27 +665,6 @@ 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="." GMMU=0 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="." GMMU=0 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]
@@ -647,12 +676,10 @@ 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: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd rmmod
+      run: ./extra/hcq/hcq_smi.py amd rmmod
     - name: Kill stale pids
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
+      run: ./extra/hcq/hcq_smi.py amd kill_pids
     - name: Symlink models and datasets
       run: |
         mkdir -p weights
@@ -681,7 +708,7 @@ 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
+      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
     - name: Test AMD=1
       run: DEBUG=2 AMD=1 python -m pytest -rA test/test_tiny.py
     - name: Test DISK copy time
@@ -691,20 +718,20 @@ 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
+      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee am_train_cifar_one_gpu.txt
     # - name: Run 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
+    #   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
     - 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
-    - name: Remote
-      run: |
-        pkill -f 'extra/remote/serve.py' || true
-        PYTHONPATH=. python3 extra/remote/serve.py 6482 &
-        sleep 1
-        DEBUG=2 PYTHONPATH=. REMOTE=127.0.0.1:6482 AM_RESET=1 AMD=1 AMD_IFACE=PCI python3 test/test_tiny.py
-        DEBUG=2 PYTHONPATH=. REMOTE=127.0.0.1:6482 AM_RESET=1 AMD=1 AMD_AQL=1 AMD_IFACE=PCI python3 test/test_tiny.py
-        pkill -f 'extra/remote/serve.py' || true
+      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
     - 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
 
@@ -719,12 +746,10 @@ 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: PYTHONPATH=. ./extra/hcq/hcq_smi.py nv rmmod
+      run: ./extra/hcq/hcq_smi.py nv rmmod
     - name: Kill stale pids
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py nv kill_pids
+      run: ./extra/hcq/hcq_smi.py nv kill_pids
     - name: Symlink models and datasets
       run: |
         mkdir -p weights
@@ -753,20 +778,21 @@ 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
+      run: BENCHMARK_LOG=llama3_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --benchmark --temperature 0 | tee nv_llama3_beam.txt
     - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
+      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee nv_train_cifar_one_gpu.txt
     - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-      run: BENCHMARK_LOG=resnet_10steps NV=1 MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
+      run: BENCHMARK_LOG=resnet_10steps NV=1 MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee nv_train_resnet_one_gpu.txt
     - name: Run 10 MLPerf Bert training steps (1 gpu)
       # 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
-    - name: Remote
-      run: |
-        pkill -f 'extra/remote/serve.py' || true
-        PYTHONPATH=. python3 extra/remote/serve.py 6483 &
-        sleep 1
-        DEBUG=2 PYTHONPATH=. REMOTE=127.0.0.1:6483 NV=1 python3 test/test_tiny.py
-        pkill -f 'extra/remote/serve.py' || true
+      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
     - 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

@@ -1,11 +1,11 @@
 name: Unit Tests
 env:
   # increment this when downloads substantially change to avoid the internet
-  CACHE_VERSION: '18'
+  CACHE_VERSION: '15'
   CAPTURE_PROCESS_REPLAY: 1
   GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
   PYTHONPATH: ${{ github.workspace }}
-  CHECK_OOB: 1
+  IGNORE_OOB: 0
 
 on:
   push:
@@ -26,19 +26,19 @@ jobs:
       uses: ./.github/actions/setup-tinygrad
       with:
         key: llvm-speed
-        deps: testing_unit
+        deps: testing_minimal
         llvm: 'true'
     - name: Speed Test
-      run: CPU=1 CPU_LLVM=1 THREADS=0 python3 test/speed/external_test_speed_v_torch.py
+      run: CPU=1 CPU_LLVM=1 python3 test/speed/external_test_speed_v_torch.py
     - name: Speed Test (BEAM=2)
-      run: BEAM=2 CPU=1 CPU_LLVM=1 THREADS=0 python3 test/speed/external_test_speed_v_torch.py
+      run: BEAM=2 CPU=1 CPU_LLVM=1 python3 test/speed/external_test_speed_v_torch.py
 
   docs:
     name: Docs
     runs-on: ubuntu-22.04
     timeout-minutes: 10
     env:
-      CHECK_OOB: 0
+      IGNORE_OOB: 1
     steps:
     - name: Checkout Code
       uses: actions/checkout@v4
@@ -70,7 +70,7 @@ jobs:
         source venv/bin/activate
         pip install $GITHUB_WORKSPACE
         cp $GITHUB_WORKSPACE/examples/beautiful_mnist.py .
-        BS=2 STEPS=10 MAX_BUFFER_SIZE=0 python beautiful_mnist.py
+        BS=2 STEPS=10 python beautiful_mnist.py
     - name: Test Docs Build
       run: python -m mkdocs build --strict
     - name: Test Docs
@@ -98,7 +98,7 @@ jobs:
       uses: ./.github/actions/setup-tinygrad
       with:
         key: torch-backend-pillow-torchvision-et-pt
-        deps: testing_unit
+        deps: testing_minimal
         pydeps: "pillow torchvision expecttest"
         llvm: 'true'
     - name: Install ninja
@@ -106,15 +106,15 @@ jobs:
         sudo apt update || true
         sudo apt install -y --no-install-recommends ninja-build
     - name: Test one op
-      run: FORWARD_ONLY=1 TINY_BACKEND=1 python3 test/test_tiny.py TestTiny.test_plus
+      run: FORWARD_ONLY=1 TINY_BACKEND=1 python3 test/test_ops.py TestOps.test_add
     - name: Test ResNet-18
       run: DEBUG=2 python3 extra/torch_backend/example.py
     - name: custom tests
-      run: python3 -m pytest -n auto extra/torch_backend/test.py --durations=20
+      run: python3 extra/torch_backend/test.py
     - 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/backend/test_ops.py --durations=20
+      run: CPU=1 CPU_LLVM=1 LLVMOPT=0 TINY_BACKEND=1 python3 -m pytest -n auto test/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
@@ -134,14 +134,14 @@ jobs:
       uses: ./.github/actions/setup-tinygrad
       with:
         key: torch-backend-pillow-torchvision-et-pt
-        deps: testing_unit
+        deps: testing_minimal
         llvm: 'true'
     - name: Install ninja
       run: |
         sudo apt update || true
         sudo apt install -y --no-install-recommends ninja-build
     - name: Test beautiful_mnist in torch with TINY_BACKEND
-      run: STEPS=20 CPU=1 TARGET_EVAL_ACC_PCT=90.0 MAX_BUFFER_SIZE=0 TINY_BACKEND=1 python3 examples/other_mnist/beautiful_mnist_torch.py
+      run: STEPS=20 CPU=1 TARGET_EVAL_ACC_PCT=90.0 TINY_BACKEND=1 python3 examples/other_mnist/beautiful_mnist_torch.py
     - name: Test some torch tests (expect failure)
       run: python3 -m pytest extra/torch_backend/torch_tests.py -v --tb=no || true
 
@@ -156,27 +156,27 @@ jobs:
       uses: ./.github/actions/setup-tinygrad
       with:
         key: be-minimal
-        deps: testing_unit
+        deps: testing_minimal
     - name: Test dtype with Python emulator
-      run: DEBUG=1 PYTHON=1 python3 -m pytest -n=auto test/backend/test_dtype.py test/backend/test_dtype_alu.py
+      run: DEBUG=1 PYTHON=1 python3 -m pytest -n=auto test/test_dtype.py test/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/backend/test_ops.py --durations=20
+      run: DEBUG=2 SKIP_SLOW_TEST=1 PYTHON=1 python3 -m pytest -n=auto test/test_ops.py --durations=20
     - name: Test uops with Python emulator
-      run: PYTHON=1 python3 -m pytest test/backend/test_uops.py --durations=20
+      run: PYTHON=1 python3 -m pytest test/test_uops.py --durations=20
     - name: Test symbolic with Python emulator
-      run: PYTHON=1 python3 test/backend/test_symbolic_ops.py
+      run: PYTHON=1 python3 test/test_symbolic_ops.py
     - name: test_renderer_failures with Python emulator
-      run: PYTHON=1 python3 -m pytest -rA test/backend/test_renderer_failures.py::TestRendererFailures
-    - name: Test IMAGE support
+      run: PYTHON=1 python3 -m pytest -rA test/test_renderer_failures.py::TestRendererFailures
+    - name: Test IMAGE=2 support
       run: |
-        IMAGE=1 PYTHON=1 python3 test/backend/test_ops.py TestOps.test_gemm
-        IMAGE=1 PYTHON=1 python3 test/backend/test_ops.py TestOps.test_simple_conv2d
+        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
     - name: Test emulated METAL tensor cores
       run: |
-        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/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/backend/test_ops.py TestOps.test_gemm
+      run: DEBUG=2 AMX=1 EMULATE=AMX FORWARD_ONLY=1 PYTHON=1 python3 test/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
@@ -197,9 +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/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 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_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
@@ -207,17 +207,18 @@ 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/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
+        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
 
   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
@@ -229,51 +230,19 @@ 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: Run pre-commit linting hooks
-      run: SKIP=tiny,tests,example pre-commit run --all-files
-    - name: Lint additional files with ruff
+    - name: Lint with ruff
       run: |
+        pip3 install --upgrade --force-reinstall ruff==0.14.10
+        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
         python3 -m ruff check extra/torch_backend/backend.py
-    - name: Run mypy with lineprecision report
+    - name: Run mypy
       run: |
-        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}'
+        python -m mypy --strict-equality --lineprecision-report .
         cat lineprecision.txt
     - name: Run TYPED=1
-      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
+      run: TYPED=1 python -c "import tinygrad"
 
   unittest:
     name: Unit Tests
@@ -286,19 +255,27 @@ jobs:
     - name: Setup Environment
       uses: ./.github/actions/setup-tinygrad
       with:
-        key: unittest-13
-        pydeps: "pillow ftfy regex pre-commit"
+        key: unittest-12
+        pydeps: "pillow numpy ftfy regex"
         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/null/test_device.py TestRunAsModule.test_module_runs
-        CPU=1 python -m pytest -n=auto test/unit/ --durations=20
+        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
     - name: Run GC tests
       run: python test/external/external_uop_gc.py
     - name: External Benchmark Schedule
@@ -312,8 +289,8 @@ 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 < 24000 lines
-      run: MAX_LINE_COUNT=24000 python sz.py
+    - name: Repo line count < 20000 lines
+      run: MAX_LINE_COUNT=20000 python sz.py
 
   spec:
     strategy:
@@ -333,7 +310,7 @@ jobs:
         deps: testing_unit
         python-version: '3.14'
     - name: Test SPEC=2
-      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 }}
+      run: SPEC=2 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 }}
 
   fuzzing:
     name: Fuzzing
@@ -367,13 +344,13 @@ jobs:
         uses: ./.github/actions/setup-tinygrad
         with:
           key: gpu-image
-          deps: testing_unit
+          deps: testing_minimal
           opencl: 'true'
-      - name: Test CL IMAGE=1 ops
+      - name: Test CL IMAGE=2 ops
         run: |
-          CL=1 IMAGE=1 python -m pytest -n=auto test/backend/test_ops.py --durations=20
+          CL=1 IMAGE=2 python -m pytest -n=auto test/test_ops.py --durations=20
         # TODO: training is broken
-        # CL=1 IMAGE=1 python test/models/test_end2end.py TestEnd2End.test_linear_mnist
+        # CL=1 IMAGE=2 python test/models/test_end2end.py TestEnd2End.test_linear_mnist
       - name: Run process replay tests
         uses: ./.github/actions/process-replay
 
@@ -388,14 +365,14 @@ jobs:
         uses: ./.github/actions/setup-tinygrad
         with:
           key: gen-dataset
-          deps: testing
+          deps: testing_minimal
           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/backend/test_optim.py -k "not muon"
+        run: CL=1 FUSE_OPTIM=1 python -m pytest -n=auto test/models/test_mnist.py test/test_optim.py -k "not muon"
       - name: Upload artifact
         uses: actions/upload-artifact@v4
         with:
@@ -418,13 +395,13 @@ jobs:
           llvm: 'true'
       - name: Test openpilot model kernel count and gate usage
         run: |
-          ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1486 ALLOWED_GATED_READ_IMAGE=17 FLOAT16=1 CL=1 IMAGE=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
+          ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1397 ALLOWED_GATED_READ_IMAGE=94 FLOAT16=1 CL=1 IMAGE=2 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
       - name: Test openpilot CL compile fp16
-        run: FLOAT16=1 DEBUGCL=1 CL=1 IMAGE=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
+        run: FLOAT16=1 DEBUGCL=1 CL=1 IMAGE=2 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
       - name: Test openpilot CL compile fp32 (test correctness)
-        run: CL=1 IMAGE=1 SELFTEST=1 python examples/openpilot/compile3.py https://github.com/haraschax/filedump/raw/refs/heads/master/driving_vision_fp32.onnx
+        run: DEBUGCL=1 CL=1 IMAGE=2 SELFTEST=1 python examples/openpilot/compile3.py https://github.com/haraschax/filedump/raw/refs/heads/master/driving_vision_fp32.onnx
       - name: Test openpilot LLVM compile fp16
-        run: IMAGE=1 FLOAT16=1 CPU=1 CPU_LLVM=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
+        run: FLOAT16=1 CPU=1 CPU_LLVM=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
       - name: Run process replay tests
         uses: ./.github/actions/process-replay
 
@@ -443,7 +420,7 @@ jobs:
         with:
           key: onnxoptc
           deps: testing
-          python-version: '3.12'
+          python-version: '3.11'
           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
@@ -454,7 +431,7 @@ 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/backend/test_quantize_onnx.py
+        run: CPU=1 CPU_LLVM=0 python3 test/test_quantize_onnx.py
       - name: Run process replay tests
         uses: ./.github/actions/process-replay
 
@@ -471,7 +448,7 @@ jobs:
           key: onnxoptl
           deps: testing
           pydeps: "tensorflow==2.19"
-          python-version: '3.12'
+          python-version: '3.11'
           opencl: 'true'
       - name: Test ONNX (CL)
         run: CL=1 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
@@ -484,11 +461,11 @@ 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 NULL_ALLOW_COPYOUT=1 python examples/beautiful_mnist_multigpu.py
+        run: NULL=1 python examples/beautiful_mnist_multigpu.py
       - name: Test Bert training
-        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
+        run: NULL=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 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
+        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
       - name: Run process replay tests
         uses: ./.github/actions/process-replay
 
@@ -497,7 +474,7 @@ jobs:
     runs-on: ubuntu-24.04
     timeout-minutes: 15
     env:
-      CHECK_OOB: 0
+      IGNORE_OOB: 1
     steps:
       - name: Checkout Code
         uses: actions/checkout@v4
@@ -505,13 +482,8 @@ jobs:
         uses: ./.github/actions/setup-tinygrad
         with:
           key: apps_llm
-      - name: Test 1B LLM (llama)
-        run: echo "What's a male chicken called? Answer with only one word." | MAX_BUFFER_SIZE=0 python3 -m tinygrad.apps.llm --model llama3.2:1b | tee /dev/stderr | grep -i rooster
-      - name: Test 1B LLM (llama q4)
-        run: echo "What's a male chicken called? Answer with only one word." | MAX_BUFFER_SIZE=0 python3 -m tinygrad.apps.llm --model llama3.2:1b-q4 | tee /dev/stderr | grep -i rooster
-      - name: Test 1B LLM (qwen)
-        # NOTE: qwen is dumb and only knows about female chickens
-        run: echo "What's a female chicken called? Answer with only one word." | MAX_BUFFER_SIZE=0 python3 -m tinygrad.apps.llm --model qwen3:0.6b | tee /dev/stderr | grep -i hen
+      - name: Test 1B LLM
+        run: echo "What's a male chicken called? Answer with only one word." | MAX_BUFFER_SIZE=0 python3 -m tinygrad.apps.llm | grep -i rooster
 
 # ****** Models Tests ******
 
@@ -550,7 +522,7 @@ jobs:
       with:
         key: metal
         deps: testing
-        python-version: '3.12'
+        python-version: '3.11'
     - name: Test models (Metal)
       run: METAL=1 python -m pytest -n=auto test/models --durations=20
     - name: Test LLaMA compile speed
@@ -569,15 +541,15 @@ jobs:
       uses: ./.github/actions/setup-tinygrad
       with:
         key: devectorize-minimal
-        deps: testing_unit
+        deps: testing_minimal
         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/backend/test_ops.py
+      run: CPU=1 CPU_LLVM=1 DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/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/backend/test_ops.py
+      run: CPU=1 CPU_LLVM=0 DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/test_ops.py
 
   testdsp:
     name: Linux (DSP)
@@ -590,8 +562,8 @@ jobs:
       uses: ./.github/actions/setup-tinygrad
       with:
         key: dsp-minimal
-        deps: testing_unit
-        pydeps: "onnx==1.18.0 onnxruntime ml_dtypes"
+        deps: testing_minimal
+        pydeps: "onnx==1.18.0 onnxruntime pillow"
         llvm: "true"
     - name: Set up Docker Buildx
       uses: docker/setup-buildx-action@v3
@@ -603,15 +575,15 @@ jobs:
         load: true
         tags: qemu-hexagon:latest
         cache-from: type=gha
-        cache-to: ${{ github.event_name != 'pull_request' && 'type=gha,mode=min' || '' }}
+        cache-to: 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/backend/test_transcendental.py TestTranscendentalVectorized
+      run: CC=clang-20 DEBUG=2 DSP=1 python test/test_transcendental.py TestTranscendentalVectorized
     - name: Test quantize onnx
-      run: DEBUG=2 DSP=1 python3 test/backend/test_quantize_onnx.py
+      run: DEBUG=2 DSP=1 python3 test/test_quantize_onnx.py
 
   testwebgpu:
     name: Linux (WebGPU)
@@ -624,105 +596,34 @@ jobs:
       uses: ./.github/actions/setup-tinygrad
       with:
         key: webgpu-minimal
-        deps: testing_unit
-        python-version: '3.12'
+        deps: testing_minimal
+        python-version: '3.11'
         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_tiny.py TestTiny.test_plus
+        WEBGPU=1 DEBUG=4 FORWARD_ONLY=1 python3 test/test_ops.py TestOps.test_add
     - name: Run selected webgpu tests
       run: |
-          WEBGPU=1 WEBGPU_BACKEND="WGPUBackendType_Vulkan" python3 -m pytest -n=auto test/backend --durations=20
+          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
     - 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: GMMU=0 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, cdna4]
 
-    name: Linux (${{ matrix.backend }} ${{ matrix.arch }})
+    name: Linux (${{ matrix.backend }})
     runs-on: ubuntu-22.04
-    timeout-minutes: 15
+    timeout-minutes: 20
     env:
       AMD: 1
       MOCKGPU: 1
-      MOCKGPU_ARCH: ${{ matrix.arch }}
-      SKIP_SLOW_TEST: 1
+      FORWARD_ONLY: 1
       AMD_LLVM: ${{ matrix.backend == 'amdllvm' && '1' || matrix.backend != 'amdllvm' && '0' }}
     steps:
       - name: Checkout Code
@@ -731,20 +632,77 @@ jobs:
         uses: ./.github/actions/setup-tinygrad
         with:
           key: ${{ matrix.backend }}-minimal
-          deps: testing_unit
+          deps: testing_minimal
           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_tiny.py TestTiny.test_plus
+          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
       - name: Run pytest (amd)
-        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/external/external_test_am.py test/backend/test_asm_gemm.py::TestAsmGEMM --durations=20
+        run: python -m pytest -n=auto test/test_ops.py test/test_dtype.py test/test_dtype_alu.py test/test_linearizer.py test/test_randomness.py test/test_jit.py test/test_graph.py test/test_multitensor.py test/device/test_hcq.py test/testextra/test_cfg_viz.py --durations=20
+      - name: Run pytest (amd)
+        run: python -m pytest test/external/external_test_am.py --durations=20
       - name: Run TRANSCENDENTAL math
-        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
+        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
       - name: Run process replay tests
         uses: ./.github/actions/process-replay
 
+  testamdasm:
+    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: Run RDNA3 emulator tests (AMD_LLVM=1)
+        run: AMD_LLVM=1 python -m pytest -n=auto extra/assembly/amd/ --durations 20
+      - name: Run RDNA3 dtype tests
+        run: AMD=1 PYTHON_REMU=1 MOCKGPU=1 AMD_LLVM=0 pytest -n=auto test/test_dtype_alu.py test/test_dtype.py
+      - name: Run RDNA3 dtype tests (AMD_LLVM=1)
+        run: AMD=1 PYTHON_REMU=1 MOCKGPU=1 AMD_LLVM=1 pytest -n=auto test/test_dtype_alu.py test/test_dtype.py
+
+  testamdautogen:
+    name: AMD autogen
+    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-autogen
+          pydeps: "pdfplumber"
+      - name: Verify AMD autogen is up to date
+        run: |
+          python -m extra.assembly.amd.pdf --arch all
+          git diff --exit-code extra/assembly/amd/autogen/
+
   testnvidia:
     strategy:
       fail-fast: false
@@ -764,7 +722,7 @@ jobs:
         uses: ./.github/actions/setup-tinygrad
         with:
           key: ${{ matrix.backend }}-minimal
-          deps: testing_unit
+          deps: testing_minimal
           cuda: 'true'
           ocelot: 'true'
       - name: Set env
@@ -772,12 +730,10 @@ 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_tiny.py TestTiny.test_plus
+          DEBUG=5 FORWARD_ONLY=1 python3 test/test_ops.py TestOps.test_add
       - name: Run pytest (cuda)
         # skip multitensor because it's slow
-        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
+        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
       - name: Run process replay tests
         uses: ./.github/actions/process-replay
 
@@ -797,7 +753,7 @@ jobs:
         uses: ./.github/actions/setup-tinygrad
         with:
           key: ${{ matrix.backend }}-minimal
-          deps: testing_unit
+          deps: testing_minimal
           opencl: ${{ matrix.backend == 'opencl' && 'true' }}
           llvm: ${{ matrix.backend == 'llvm' || matrix.backend == 'lvp' }}
           mesa: ${{ matrix.backend == 'lvp' && 'true' }}
@@ -806,11 +762,11 @@ jobs:
       - name: Check Device.DEFAULT and print some source
         run: |
           python3 -c "from tinygrad import Device; assert Device.DEFAULT in ['CPU','CL'], Device.DEFAULT"
-          DEBUG=5 FORWARD_ONLY=1 python3 test/test_tiny.py TestTiny.test_plus
+          DEBUG=5 FORWARD_ONLY=1 python3 test/test_ops.py TestOps.test_add
       - name: Run pytest (${{ matrix.backend }})
-        run: python -m pytest -n=auto test/backend --durations=20
+        run: python -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit --durations=20
       - name: Run TRANSCENDENTAL math
-        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
+        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 process replay tests
         uses: ./.github/actions/process-replay
 
@@ -828,29 +784,25 @@ jobs:
       with:
         key: metal
         deps: testing
-        python-version: '3.12'
+        python-version: '3.11'
         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/backend/test_ops.py TestOps.test_gemm
+      run: METAL=1 DEBUG=3 TC=2 python test/test_ops.py TestOps.test_gemm
     - name: Test tensor core ops (real)
-      run: METAL=1 DEBUG=3 python test/backend/test_ops.py TestOps.test_big_gemm
+      run: METAL=1 DEBUG=3 python test/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/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
+      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
     - name: Run pytest (amd)
       env:
         MOCKGPU: 1
@@ -873,8 +825,6 @@ 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
@@ -893,14 +843,14 @@ 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: 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
+    - 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
@@ -932,7 +882,8 @@ jobs:
         uses: ./.github/actions/setup-tinygrad
         with:
           key: macos-${{ matrix.backend }}-minimal
-          deps: testing_unit
+          deps: testing_minimal
+          pydeps: "capstone"
           llvm: ${{ matrix.backend == 'llvm' || matrix.backend == 'lvp' }}
           mesa: ${{ matrix.backend == 'lvp' && 'true' }}
       - name: Set env
@@ -942,7 +893,7 @@ jobs:
           python -c "from tinygrad import Device; assert Device.DEFAULT == {'LLVM':'CPU','LVP':'CPU'}.get(x:='${{ matrix.backend }}'.upper(), x), Device.DEFAULT"
           DEBUG=4 python3 test/test_tiny.py TestTiny.test_plus
       - name: Run pytest (${{ matrix.backend }})
-        run: python3 -m pytest -n=auto test/backend --durations=20
+        run: python3 -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit --durations=20
       - name: Run process replay tests
         uses: ./.github/actions/process-replay
       - name: Run macOS-specific unit test
@@ -975,16 +926,12 @@ jobs:
       - name: Run unit tests
         if: matrix.backend=='llvm'
         # test_newton_schulz hits RecursionError
-        run: python -m pytest -n=auto test/unit/ --ignore=test/unit/test_disk_tensor.py --ignore=test/unit/test_tar.py --ignore=test/unit/test_linalg.py --durations=20
-      - name: Run NULL backend tests
-        if: matrix.backend=='llvm'
-        shell: bash
-        run: CPU=0 CPU_LLVM=0 NULL=1 python -m pytest -n=auto test/null/ --ignore=test/null/test_elf.py --durations=20
+        run: python -m pytest -n=auto test/unit/ --ignore=test/unit/test_disk_tensor.py --ignore=test/unit/test_elf.py --ignore=test/unit/test_tar.py --ignore=test/unit/test_linalg.py --durations=20
       - name: Run pytest (${{ 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/backend/test_ops.py --durations=20
+          python -m pytest -n=auto test/test_tiny.py test/test_ops.py --durations=20
 
 # ****** Compile-only Tests ******
 
@@ -1003,38 +950,15 @@ jobs:
         uses: ./.github/actions/setup-tinygrad
         with:
           key: compile-${{ matrix.backend }}
-          deps: testing_unit
+          deps: testing_minimal
           mesa: ${{ (matrix.backend == 'ir3' || matrix.backend == 'nak') && 'true' }}
-          python-version: '3.12'
+          python-version: '3.14'
       - name: Set env
         shell: bash
-        run: printf "NULL=1\nNULL_ALLOW_COPYOUT=1\n${{ matrix.backend == 'ir3' && 'NULL_IR3=1' || matrix.backend == 'nak' && 'NULL_NAK=1' }}" >> $GITHUB_ENV
+        run: printf "NULL=1\n${{ 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/backend/test_ops.py TestOps.test_add
-          python -m pytest -n=auto test/backend/test_ops.py --durations=20
-  qcomclcompiletests:
-    name: Compile-only (QCOM CL)
-    runs-on: ubuntu-24.04-arm
-    timeout-minutes: 15
-    steps:
-      - name: Checkout Code
-        uses: actions/checkout@v4
-      - name: Setup Environment
-        uses: ./.github/actions/setup-tinygrad
-        with:
-          key: compile-qcomcl
-          deps: testing_unit
-          tinydreno: 'true'
-          python-version: '3.12'
-      - name: Set env
-        shell: bash
-        run: printf "NULL=1\nNULL_ALLOW_COPYOUT=1\nNULL_QCOMCL=1" >> $GITHUB_ENV
-      - name: Run test_ops
-        shell: bash
-        run: |
-          python -c "from tinygrad import Device; assert Device.DEFAULT == 'NULL'"
-          DEBUG=4 python3 test/backend/test_ops.py TestOps.test_add
-          python -m pytest -n=auto test/backend/test_ops.py --durations=20
+          DEBUG=4 python3 test/test_ops.py TestOps.test_add
+          python -m pytest -n=auto test/test_ops.py --durations=20

.gitignore

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

.pre-commit-config.yaml

@@ -16,7 +16,7 @@ repos:
         pass_filenames: false
       - id: mypy
         name: mypy
-        entry: python3 -m mypy
+        entry: python3 -m mypy tinygrad/ --strict-equality
         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/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
+        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
         language: system
         always_run: true
         pass_filenames: false

AGENTS.md

@@ -0,0 +1,17 @@
+# 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

@@ -0,0 +1,216 @@
+# 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.
+
+## AMD Performance Counter Profiling
+
+Set VIZ to `-2` to save performance counters traces for the AMD backend.
+
+Use the CLI in `./extra/sqtt/roc.py` to explore the trace.

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/backend/test_ops.py        # just the ops tests
+python3 test/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_kernel_graph
+::: tinygrad.schedule.rangeify.get_rangeify_map
     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 interface for asm24xx chips.
+* `USB`: USB3 interafce 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

@@ -6,7 +6,6 @@ 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
@@ -88,8 +87,4 @@ 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.bfloat16
-::: tinygrad.Tensor.double
-::: tinygrad.Tensor.long
-::: tinygrad.Tensor.short
+::: tinygrad.Tensor.bool

docs/tensor/movement.md

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

docs/tensor/ops.md

@@ -7,7 +7,6 @@
 ::: tinygrad.Tensor.any
 ::: tinygrad.Tensor.all
 ::: tinygrad.Tensor.isclose
-::: tinygrad.Tensor.allclose
 ::: tinygrad.Tensor.mean
 ::: tinygrad.Tensor.var
 ::: tinygrad.Tensor.var_mean
@@ -31,9 +30,7 @@
 ::: 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
@@ -41,9 +38,7 @@
 ::: 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
 
@@ -61,8 +56,3 @@
 ::: 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

@@ -1,196 +0,0 @@
-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

@@ -1,79 +0,0 @@
-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/beautiful_mnist.py

@@ -1,6 +1,6 @@
 # model based off https://medium.com/data-science/going-beyond-99-mnist-handwritten-digits-recognition-cfff96337392
 from typing import Callable
-from tinygrad import Tensor, TinyJit, nn, GlobalCounters, function
+from tinygrad import Tensor, TinyJit, nn, GlobalCounters
 from tinygrad.helpers import getenv, colored, trange
 from tinygrad.nn.datasets import mnist
 
@@ -15,31 +15,30 @@ class Model:
       nn.BatchNorm(64), Tensor.max_pool2d,
       lambda x: x.flatten(1), nn.Linear(576, 10)]
 
-  @function
   def __call__(self, x:Tensor) -> Tensor: return x.sequential(self.layers)
 
-  @TinyJit
-  @Tensor.train()
-  def train_step(self, X_train:Tensor, Y_train:Tensor) -> Tensor:
-    opt.zero_grad()
-    samples = Tensor.randint(getenv("BS", 512), high=X_train.shape[0])
-    loss = self(X_train[samples]).sparse_categorical_crossentropy(Y_train[samples]).backward()
-    return loss.realize(*opt.schedule_step())
-
-  @TinyJit
-  def get_test_acc(self, X_test:Tensor, Y_test:Tensor) -> Tensor: return (self(X_test).argmax(axis=1) == Y_test).mean()*100
-
 if __name__ == "__main__":
   X_train, Y_train, X_test, Y_test = mnist(fashion=getenv("FASHION"))
 
   model = Model()
   opt = (nn.optim.Muon if getenv("MUON") else nn.optim.SGD if getenv("SGD") else nn.optim.Adam)(nn.state.get_parameters(model))
 
+  @TinyJit
+  @Tensor.train()
+  def train_step() -> Tensor:
+    opt.zero_grad()
+    samples = Tensor.randint(getenv("BS", 512), high=X_train.shape[0])
+    loss = model(X_train[samples]).sparse_categorical_crossentropy(Y_train[samples]).backward()
+    return loss.realize(*opt.schedule_step())
+
+  @TinyJit
+  def get_test_acc() -> Tensor: return (model(X_test).argmax(axis=1) == Y_test).mean()*100
+
   test_acc = float('nan')
   for i in (t:=trange(getenv("STEPS", 70))):
     GlobalCounters.reset()   # NOTE: this makes it nice for DEBUG=2 timing
-    loss = model.train_step(X_train, Y_train)
-    if i%10 == 9: test_acc = model.get_test_acc(X_test, Y_test).item()
+    loss = train_step()
+    if i%10 == 9: test_acc = get_test_acc().item()
     t.set_description(f"loss: {loss.item():6.2f} test_accuracy: {test_acc:5.2f}%")
 
   # verify eval acc

examples/benchmark_onnx.py

@@ -5,7 +5,7 @@ from extra.onnx_helpers import get_example_inputs, validate
 
 def load_onnx_model(onnx_file):
   run_onnx = OnnxRunner(onnx_file)
-  run_onnx_jit = TinyJit(lambda **kwargs: next(iter(run_onnx({k:v.to(None) for k,v in kwargs.items()}).values())), prune=True)
+  run_onnx_jit = TinyJit(lambda **kwargs: next(iter(run_onnx({k:v.to(None) for k,v in kwargs.items()}).values())), prune=True, optimize=True)
   return run_onnx_jit, run_onnx.graph_inputs
 
 if __name__ == "__main__":

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)=}"
-assert EVAL_BS % len(GPUS) == 0, f"{EVAL_BS=} is not a multiple of {len(GPUS)=}"
+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"
 
 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

@@ -65,7 +65,17 @@ 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))
-      X[idx].flatten().assign(img.tobytes())
+
+      # 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!
       q_out.put(idx)
     q_out.put(None)
 
@@ -203,13 +213,12 @@ 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, seed:int|None=None):
+def batch_load_train_bert(BS:int):
   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
-    rng.shuffle(fs)
+    random.shuffle(fs)
     train_files.append(fs.pop(0))
 
   cycle_length = min(getenv("NUM_CPU_THREADS", min(os.cpu_count(), 8)), len(train_files))
@@ -254,8 +263,8 @@ 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].flatten().assign(x.tobytes())
-    Y[idx].flatten().assign(y.tobytes())
+    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()
 
     queue_out.put(idx)
   queue_out.put(None)
@@ -369,12 +378,12 @@ 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].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())
+      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()
 
-    imgs[idx].flatten().assign(img.tobytes())
+    imgs[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = img.tobytes()
 
     queue_out.put(idx)
   queue_out.put(None)
@@ -396,7 +405,6 @@ 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}")
@@ -543,7 +551,7 @@ 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: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.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.count, = struct.unpack("<Q", self.idx.read(8))
     doc_count, = struct.unpack("<Q", self.idx.read(8))
 
@@ -560,7 +568,7 @@ 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")).numpy()
+    self.bin_t = Tensor(base_path.with_name(f"{base_path.name}.bin"))
 
   def _index(self, idx) -> tuple[int, int]:
     return int(self.pointers[idx]), int(self.sizes[idx])
@@ -569,7 +577,7 @@ 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].view(self.dtype)
+    return self.bin_t[ptr:ptr+length*self.dtype.itemsize].bitcast(self.dtype).to(None)
 
 # https://docs.nvidia.com/megatron-core/developer-guide/latest/api-guide/datasets.html
 class GPTDataset:
@@ -628,7 +636,7 @@ class GPTDataset:
         sample_parts.append(self.indexed_dataset.get(int(self.doc_idx[i]), offset=int(offset), length=length))
 
     # concat all parts
-    text = np.concatenate(sample_parts, axis=0)
+    text = Tensor.cat(*sample_parts)
 
     return text
 
@@ -771,8 +779,7 @@ 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)]
-    stacked = np.stack(batch, axis=0)
-    yield Tensor(stacked, device="NPY")
+    yield Tensor.stack(batch, dim=0)
 
 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

@@ -219,18 +219,7 @@ 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
-  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
+  return BertForPretraining(**config)
 
 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, profile_marker, DEBUG
+from tinygrad.helpers import getenv, BEAM, WINO, round_up, diskcache_clear, Profiling
 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
 
@@ -13,8 +13,6 @@ from extra.bench_log import BenchEvent, WallTimeEvent
 # TODO: fix benchmark logging and use tinygrad tqdm
 from tqdm import tqdm
 
-from tinygrad.uop.ops import UOp
-
 def train_resnet():
   from extra.models import resnet
   from examples.mlperf.dataloader import batch_load_resnet
@@ -1010,7 +1008,6 @@ 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
 
@@ -1088,7 +1085,7 @@ 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, seed=seed), total=train_steps, disable=BENCHMARK))
+    train_it = iter(tqdm(batch_load_train_bert(BS), total=train_steps, disable=BENCHMARK))
     for _ in range(start_step): next(train_it) # Fast forward
   else:
     # repeat fake data
@@ -1150,7 +1147,7 @@ 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()
-        if not getenv("FP8_TRAIN"): assert not math.isnan(loss)
+        assert not math.isnan(loss)
         lr = lr.item()
 
       cl = time.perf_counter()
@@ -1163,7 +1160,7 @@ 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/mem":GlobalCounters.mem_used / 1e9, "train/GFLOPS": GlobalCounters.global_ops * 1e-9 / (cl - st), "epoch": (i+1)*GBS})
+                    "train/GFLOPS": GlobalCounters.global_ops * 1e-9 / (cl - st), "epoch": (i+1)*GBS})
 
       train_data, next_data = next_data, None
       i += 1
@@ -1284,29 +1281,21 @@ def train_bert():
         previous_step = i
 
 def train_llama3():
-  from examples.mlperf.models.flat_llama import FlatTransformer, apply_grad
+  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)
@@ -1320,12 +1309,10 @@ def train_llama3():
   opt_adamw_weight_decay = 0.1
 
   opt_gradient_clip_norm = 1.0
-  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
+  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)
 
   # ** init wandb **
   WANDB = getenv("WANDB")
@@ -1337,16 +1324,8 @@ 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 = FlatTransformer(**model_params, max_context=SEQLEN)
-
+  model = Transformer(**model_params, max_context=SEQLEN, jit=False, disable_kv_cache=True)
   params = get_parameters(model)
   # weights are all bfloat16 for now
   assert params and all(p.dtype == dtypes.bfloat16 for p in params)
@@ -1355,26 +1334,32 @@ def train_llama3():
     for v in get_parameters(model):
       v = v.assign(Tensor.empty(v.shape))
 
-  is_dp = (DP := getenv("DP", 1)) > 1
-  is_mp = (MP := getenv("MP", 1)) > 1
-  is_sharding = is_dp or is_mp
-  device_count = max(DP, MP)
-  device = tuple(f"{Device.DEFAULT}:{i}" for i in range(device_count))
+  if (DP := getenv("DP", 1)) > 1:
+    device = tuple(f"{Device.DEFAULT}:{i}" for i in range(DP))
+    for v in get_parameters(model):
+      v.shard_(device, axis=None)
 
-  model.shard(device, is_mp)
-
-  if is_dp: vocab_mask.shard_(device, axis=None).realize()
-  if is_mp: vocab_mask.shard_(device, axis=2).realize()
-
-  is_offload_optim = bool(getenv("OFFLOAD_OPTIM"))
-  is_fake_offload = Device.DEFAULT == "NULL"
-  optim_device = ("CPU" if not is_fake_offload else "NULL:99") if is_offload_optim else None
-  optim = GradAccClipAdamW(params, 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
-  grads = [Tensor.zeros_like(p).contiguous() for p in optim.params]
+  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():
+      if 'scale' in k: v.shard_(device, axis=None)  # from quantized
+      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.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)
+      elif '.feed_forward.w3.' in k: v.shard_(device, axis=0)
+      elif 'tok_embeddings.weight' in k: v.shard_(device, axis=0)
+      elif 'output.weight' in k: v.shard_(device, axis=0)
+      else:
+        # attention_norm, ffn_norm, norm
+        v.shard_(device, axis=None)
+      # 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)
   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"):
@@ -1387,131 +1372,98 @@ def train_llama3():
     load_state_dict(scheduler, safe_load(fn), realize=False)
 
   @TinyJit
-  def minibatch(tokens:Tensor):
-    if is_dp: tokens = tokens.to(None).shard(device, 0)
-    if is_mp: tokens = tokens.shard(device)
-    if not is_sharding: tokens = tokens.to(None)
-    logits:Tensor = model(tokens[:, :-1])
-    loss = vocab_mask.where(-1e9, logits).sparse_categorical_crossentropy(tokens[:, 1:])
+  @Tensor.train()
+  def train_step(model, tokens:Tensor):
+    optim.zero_grad()
+    if (DP := getenv("DP", 1)) > 1:
+      device = tuple(f"{Device.DEFAULT}:{i}" for i in range(DP))
+      tokens = tokens.shard(device, 0)
+    if (MP := getenv("MP", 1)) > 1:
+      device = tuple(f"{Device.DEFAULT}:{i}" for i in range(MP))
+      tokens = tokens.shard(device)
+    logits:Tensor = model(tokens[:, :-1], start_pos=0, temperature=math.nan)
+    loss = 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)
 
-    for i,(t,g) in enumerate(zip(optim.params, loss.gradient(*optim.params))):
-      grads[i].replace(Tensor(grads[i].uop.after(UOp.group(*apply_grad(grads[i].uop, g.uop))), device=t.device))
-
-    loss_cpu = loss.flatten().float().to("CPU")
-    return loss_cpu.realize(*grads)
-
-  @TinyJit
-  def optim_step():
-    grad_norm = optim.fstep(grads)
+    optim.step()
     scheduler.step()
 
-    for g in grads: g.assign(g.zeros_like())
-
-    lr_cpu = optim.lr.float().to("CPU")
-    grad_norm_cpu = grad_norm.float().to("CPU")
-    Tensor.realize(lr_cpu, grad_norm_cpu, *grads)
-
-    return lr_cpu, grad_norm_cpu
+    lr = optim.lr
+    loss.realize(lr)
+    return loss, lr
 
   @TinyJit
   @Tensor.train(False)
-  def eval_step(tokens:Tensor):
-    if is_dp: tokens = tokens.to(None).shard(device, 0)
-    if is_mp: tokens = tokens.shard(device)
-    if not is_sharding: tokens = tokens.to(None)
-    logits:Tensor = model(tokens[:, :-1])
-    loss = vocab_mask.where(-1e9, logits).sparse_categorical_crossentropy(tokens[:, 1:])
-    return loss.flatten().float().to("CPU")
+  def eval_step(model, tokens:Tensor):
+    if (DP := getenv("DP", 1)) > 1:
+      device = tuple(f"{Device.DEFAULT}:{i}" for i in range(DP))
+      tokens = tokens.shard(device, 0)
+    if (MP := getenv("MP", 1)) > 1:
+      device = tuple(f"{Device.DEFAULT}:{i}" for i in range(MP))
+      tokens = tokens.shard(device)
+    logits:Tensor = model(tokens[:, :-1], start_pos=0, temperature=math.nan)
+    loss = logits.sparse_categorical_crossentropy(tokens[:, 1:])
+    return loss.flatten().float()
 
   # ** data iters **
   def fake_data(bs, samples):
-    import numpy as np
     for _ in range(samples // bs):
-      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")
+      yield Tensor.randint(bs, SEQLEN + 1, low=0, high=model_params["vocab_size"], dtype=dtypes.int32, device=Device.DEFAULT)
 
   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=DATA_SEED, val=bool(TRAIN_ON_VAL), small=bool(SMALL))
+      return batch_load_llama3(BS, SAMPLES, SEQLEN, BASEDIR, seed=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(EVAL_SAMPLES, SEQLEN, BASEDIR, val=True, small=bool(SMALL))
+    eval_dataset = get_llama3_dataset(5760, SEQLEN, BASEDIR, val=True, small=bool(SMALL))
 
   def get_eval_iter():
     if eval_dataset is None:
-      return fake_data(EVAL_BS, EVAL_SAMPLES)
+      return fake_data(EVAL_BS, 5760)
     from examples.mlperf.dataloader import iterate_llama3_dataset
     return iterate_llama3_dataset(eval_dataset, EVAL_BS)
 
-  num_params = sum(p.numel() for p in params) - model_params["vocab_size"]*model_params["dim"]
-  train_iter = get_train_iter()
+  iter = get_train_iter()
   i, sequences_seen = resume_ckpt, 0
-  step_times = []
-  while i < MAX_STEPS:
+  for tokens in tqdm(iter, total=SAMPLES//GBS):
     GlobalCounters.reset()
-    actual_gbs = GBS if i >= 2 else BS
     if getenv("TRAIN", 1):
-      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 >= 2 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)
+      t = time.perf_counter()
+      loss, lr = train_step(model, tokens)
+      loss = loss.float().item()
+      lr = lr.item()
 
       i += 1
-      sequences_seen += actual_gbs
+      sequences_seen += tokens.shape[0]
 
+      sec = time.perf_counter()-t
       mem_gb = GlobalCounters.mem_used / 1e9
-      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
+      gflops = GlobalCounters.global_ops / 1e9 / sec
       tqdm.write(
-          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())))
+        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")
 
       if WANDB:
-        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
-        })
+        wandb.log({"lr": lr, "train/loss": loss, "train/step_time": sec, "train/GFLOPS": gflops, "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")
@@ -1523,29 +1475,16 @@ def train_llama3():
         fn = f"{ckpt_dir}/llama3_{i}_optim.safe"
         safe_save(get_state_dict(scheduler), fn)
 
-      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 != (sequences_seen - actual_gbs) // EVAL_FREQ and (i != 1 or EVAL_FREQ == 1)) or (BENCHMARK and i == BENCHMARK):
-      if EVAL_BS == 0: return
+    if sequences_seen % EVAL_FREQ == 0 and (i != 1 or EVAL_FREQ == 1):
       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 {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
+      tqdm.write(f"evaluating {5760//EVAL_BS} batches of {EVAL_BS} sequences")
 
+      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}")
@@ -1631,7 +1570,7 @@ 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():
@@ -1670,7 +1609,7 @@ 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/models/flat_llama.py

@@ -1,193 +0,0 @@
-import math, os
-if __name__ == "__main__":
-  os.environ["DEFAULT_FLOAT"] = "bfloat16"
-  os.environ["OPTIM_DTYPE"] = "bfloat16"
-  if "DEV" not in os.environ: os.environ["DEV"] = "NULL"
-  # CDNA
-  os.environ["EMULATE"] = "AMD_CDNA4"
-  os.environ["DEVICE_IN_FUNCTION_BUG"] = "1"
-  os.environ["ALL2ALL"] = "1"
-  os.environ["USE_ATOMICS"] = "1"
-  if "HK_FLASH_ATTENTION" not in os.environ:
-    os.environ["HK_FLASH_ATTENTION"] = "1"
-    if "ASM_GEMM" not in os.environ:
-      os.environ["ASM_GEMM"] = "1"
-from tinygrad import Tensor, nn, function, getenv, dtypes, TinyJit
-from tinygrad.helpers import Timing, colored, GlobalCounters, profile_marker
-from tinygrad.uop.ops import Ops, UOp
-from extra.models.llama import apply_rotary_emb, precompute_freqs_cis
-
-FP8 = getenv("FP8", 0)
-
-FP8_DTYPE = dtypes.fp8e4m3
-FP8_MAX = 448.0
-
-def quantize_fp8(x:Tensor):
-  scale = FP8_MAX / (x.abs().max().detach() + 1e-8)
-  x_scaled = x * scale
-  x_clamped = x_scaled + (x_scaled.detach().clamp(-FP8_MAX, FP8_MAX) - x_scaled.detach())  # STE
-  return x_clamped.cast(FP8_DTYPE), scale.float().reciprocal()
-
-def matmul(x:Tensor, w:Tensor) -> Tensor:
-  if not FP8: return x @ w.T
-  # weights are already FP8, just quantize activations
-  x_fp8, x_scale = quantize_fp8(x)
-  return x_fp8.dot(w.T, dtype=dtypes.float) * x_scale
-
-def rmsnorm(x_in:Tensor, eps:float):
-  x = x_in.float()
-  x = x * (x.square().mean(-1, keepdim=True) + eps).rsqrt()
-  return x.cast(x_in.dtype)
-
-class FlatTransformer:
-  def __init__(self, dim:int, hidden_dim:int, n_heads:int, n_layers:int, norm_eps:float, vocab_size:int, n_kv_heads:int|None=None,
-               rope_theta:int=10000, max_context:int=1024):
-    self.vocab_size = vocab_size
-    self.n_layers = n_layers
-    self.n_heads = n_heads
-    self.n_kv_heads = n_kv_heads if n_kv_heads is not None else n_heads # n_kv_heads != n_heads implies MQA [arxiv/2307.09288, A.2.1]
-    self.head_dim = dim // n_heads
-    self.n_rep = self.n_heads // self.n_kv_heads
-
-    # Attention
-    self.wqkv = self.lin_per_layer(dim, self.n_heads * self.head_dim + self.n_kv_heads * self.head_dim * 2)
-    self.wo = self.lin_per_layer(self.n_heads * self.head_dim, dim)
-
-    # FeedForward
-    self.w1 = self.lin_per_layer(dim, hidden_dim)
-    self.w2 = self.lin_per_layer(hidden_dim, dim)
-    self.w3 = self.lin_per_layer(dim, hidden_dim)
-
-    self.norm_eps = norm_eps
-    self.attention_norm = Tensor.ones(n_layers, dim).contiguous()
-    self.ffn_norm = Tensor.ones(n_layers, dim).contiguous()
-
-    # output
-    self.norm = nn.RMSNorm(dim, norm_eps)
-    self.tok_embeddings = nn.Embedding(vocab_size, dim)
-    self.output = nn.Linear(dim, vocab_size, bias=False)
-    self.freqs_cis = precompute_freqs_cis(dim // n_heads, max_context * 2, rope_theta).contiguous().requires_grad_(False)
-
-  def lin_per_layer(self, in_features:int, out_features:int):
-    bound = 1 / math.sqrt(in_features)
-    dt = FP8_DTYPE if FP8 else None
-    if getenv("ZEROS"): return Tensor.zeros(self.n_layers, out_features, in_features, dtype=dt)
-    return Tensor.uniform(self.n_layers, out_features, in_features, low=-bound, high=bound, dtype=dt)
-
-  def attention(self, x:Tensor, freqs_cis:Tensor, attention_norm:Tensor, wqkv:Tensor, wo:Tensor):
-    x = rmsnorm(x, self.norm_eps) * attention_norm
-    xqkv = matmul(x, wqkv)
-
-    bsz, seqlen, _ = xqkv.shape
-    # interleaved layout: each kv group has [n_rep q heads, 1 k head, 1 v head] for clean MP sharding
-    xqkv = xqkv.reshape(bsz, seqlen, self.n_kv_heads, self.n_rep + 2, self.head_dim)
-    xq = xqkv[:, :, :, :self.n_rep].reshape(bsz, seqlen, self.n_heads, self.head_dim)
-    xk = xqkv[:, :, :, self.n_rep].reshape(bsz, seqlen, self.n_kv_heads, self.head_dim)
-    xv = xqkv[:, :, :, self.n_rep+1].reshape(bsz, seqlen, self.n_kv_heads, self.head_dim)
-
-    xq, xk = apply_rotary_emb(xq, xk, freqs_cis)
-    xq, xk, xv = xq.transpose(1, 2), xk.transpose(1, 2), xv.transpose(1, 2)
-    attn = xq.scaled_dot_product_attention(xk, xv, is_causal=True, enable_gqa=True).transpose(1, 2)
-    attn = attn.reshape(bsz, seqlen, -1)
-    return matmul(attn, wo)
-
-  def feed_forward(self, x:Tensor, ffn_norm:Tensor, w1:Tensor, w2:Tensor, w3:Tensor):
-    x = rmsnorm(x, self.norm_eps) * ffn_norm
-    x_w1 = matmul(x, w1).silu()
-    x_w3 = matmul(x.contiguous_backward(), w3)
-    return matmul(x_w1 * x_w3, w2)
-
-  @function(precompile=True, precompile_backward=True)
-  def run_layer(self, x:Tensor, freqs_cis:Tensor,
-                attention_norm:Tensor, wqkv:Tensor, wo:Tensor,
-                ffn_norm:Tensor, w1:Tensor, w2:Tensor, w3:Tensor):
-    h = x + self.attention(x, freqs_cis, attention_norm, wqkv, wo)
-    return h + self.feed_forward(h, ffn_norm, w1, w2, w3)
-
-  def shard(self, device:tuple[str, ...], mp:bool=False):
-    from tinygrad.nn.state import get_parameters
-    if not mp:
-      for v in get_parameters(self): v.shard_(device, axis=None)
-    else:
-      # flat per-layer weights: axis 0 is n_layers, so shard axes are +1 vs per-layer Transformer
-      self.wqkv.shard_(device, axis=1).realize()          # (n_layers, out, dim) shard out
-      self.wo.shard_(device, axis=2).realize()             # (n_layers, dim, in) shard in
-      self.w1.shard_(device, axis=1).realize()             # (n_layers, hidden, dim) shard out
-      self.w2.shard_(device, axis=2).realize()             # (n_layers, dim, hidden) shard in
-      self.w3.shard_(device, axis=1).realize()             # (n_layers, hidden, dim) shard out
-      self.attention_norm.shard_(device, axis=None).realize()
-      self.ffn_norm.shard_(device, axis=None).realize()
-      self.norm.weight.shard_(device, axis=None).realize()
-      self.tok_embeddings.weight.shard_(device, axis=0).realize()
-      self.output.weight.shard_(device, axis=0).realize()
-      self.freqs_cis.shard_(device, axis=None).realize()
-
-  def __call__(self, tokens:Tensor):
-    h = self.tok_embeddings(tokens)
-    freqs_cis = self.freqs_cis.cast(h.dtype)[:, :tokens.shape[1], :, :, :]
-    for i in range(self.n_layers):
-      h = self.run_layer(h, freqs_cis,
-                         self.attention_norm[i], self.wqkv[i], self.wo[i],
-                         self.ffn_norm[i], self.w1[i], self.w2[i], self.w3[i])
-    logits = self.output(self.norm(h))
-    return logits
-
-# TODO: this shouldn't be needed, but it prevents a copy of the grads. CAT can help
-def apply_grad(old_grad:UOp, new_grad:UOp) -> list[UOp]:
-  if new_grad.op == Ops.ADD:
-    return apply_grad(old_grad, new_grad.src[0])+apply_grad(old_grad, new_grad.src[1])
-  elif new_grad.op == Ops.PAD:
-    grad_shrink = tuple([(p[0], s+p[0]) for s,p in zip(new_grad.src[0].shape, new_grad.marg)])
-    return apply_grad(old_grad.shrink(grad_shrink), new_grad.src[0])
-  else:
-    return [old_grad.store(old_grad + new_grad)]
-
-if __name__ == "__main__":
-  config = {}
-  BS                 = config["BS"]                     = getenv("BS", 16)
-  SEQLEN             = config["SEQLEN"]                 = getenv("SEQLEN", 8192)
-
-  from examples.llama3 import MODEL_PARAMS
-  model_params = MODEL_PARAMS[getenv("LLAMA3_SIZE", "8B")]["args"]
-  if (llama_layers:=getenv("LLAMA_LAYERS")) != 0: model_params['n_layers'] = llama_layers
-  model = FlatTransformer(**model_params, max_context=SEQLEN)
-  state = nn.state.get_state_dict(model)
-  print("tensor count:", len(state))
-
-  # shard the model
-  from tinygrad import Device
-  if (DP := getenv("DP", 1)) > 1:
-    model.shard(tuple(f"{Device.DEFAULT}:{i}" for i in range(DP)))
-  if (MP := getenv("MP", 1)) > 1:
-    model.shard(tuple(f"{Device.DEFAULT}:{i}" for i in range(MP)), mp=True)
-
-  # preallocate all the grad buffers and zero them out
-  grads = {x:Tensor.zeros_like(x).contiguous() for x in state.values() if x.requires_grad is None}
-
-  # print model size
-  sz = 0
-  for k,v in state.items():
-    print(f"{colored(k, 'green' if v in grads else 'white'):30s} {str(v.shape):30s} {str(v.dtype):20s} {v.device}  {v.nbytes()/1e9:.2f} GB")
-    sz += v.nbytes()
-  print(f"total sz: {sz/1e9:.2f} GB")
-
-  with Timing("fake data: "): tokens = Tensor.randint(BS, SEQLEN+1, low=0, high=model.vocab_size, dtype=dtypes.int)
-  with Timing("realize weights/grads/data: "): Tensor.realize(*state.values(), *grads.values(), tokens)
-  print("mem per device: " + ', '.join(f"{dev}: {mem/1e9:.2f} GB" for dev, mem in sorted(GlobalCounters.mem_used_per_device.items())))
-  if DP > 1: tokens = tokens.shard(tuple(f"{Device.DEFAULT}:{i}" for i in range(DP)), axis=0)
-  if MP > 1: tokens = tokens.shard(tuple(f"{Device.DEFAULT}:{i}" for i in range(MP)))
-
-  @TinyJit
-  def jit_step(tokens:Tensor):
-    with Timing("python forward: "): loss = model(tokens[:, :-1]).sparse_categorical_crossentropy(tokens[:, 1:])
-    with Timing("python backward: "):
-      for t,g in zip(grads, loss.gradient(*grads)):
-        grads[t] = Tensor(grads[t].uop.after(UOp.group(*apply_grad(grads[t].uop, g.uop))), device=t.device)
-    with Timing("run step: "): loss.realize(*grads.values())
-
-  for i in range(6):
-    GlobalCounters.reset()
-    profile_marker(f"step {i}")
-    with Timing(colored(f"*** step {i}: ", "red")):
-      jit_step(tokens)
-  print("mem per device: " + ', '.join(f"{dev}: {mem/1e9:.2f} GB" for dev, mem in sorted(GlobalCounters.mem_used_per_device.items())))

examples/mlperf/models/llama.py

@@ -1,80 +0,0 @@
-from tinygrad import Tensor, nn
-from tinygrad.helpers import getenv
-from extra.models.llama import apply_rotary_emb, precompute_freqs_cis
-
-class Attention:
-  def __init__(self, dim:int, n_heads:int, n_kv_heads:int|None=None, linear=nn.Linear):
-    self.n_heads = n_heads
-    self.n_kv_heads = n_kv_heads if n_kv_heads is not None else n_heads # n_kv_heads != n_heads implies MQA [arxiv/2307.09288, A.2.1]
-    self.head_dim = dim // n_heads
-    self.n_rep = self.n_heads // self.n_kv_heads
-
-    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)
-
-  def __call__(self, x:Tensor, freqs_cis:Tensor) -> Tensor:
-    if getenv("WQKV"):
-      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 = 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)
-
-    xq, xk = apply_rotary_emb(xq, xk, freqs_cis)
-    bsz, seqlen, _, _ = xq.shape
-
-    xq, xk, xv = xq.transpose(1, 2), xk.transpose(1, 2), xv.transpose(1, 2)
-    attn = xq.scaled_dot_product_attention(xk, xv, is_causal=True, enable_gqa=True).transpose(1, 2)
-
-    attn = attn.reshape(bsz, seqlen, -1)
-    return self.wo(attn)
-
-class FeedForward:
-  def __init__(self, dim:int, hidden_dim:int, linear=nn.Linear):
-    self.w1 = linear(dim, hidden_dim, bias=False)
-    self.w2 = linear(hidden_dim, dim, bias=False)
-    self.w3 = linear(dim, hidden_dim, bias=False) # the gate in Gated Linear Unit
-
-  def __call__(self, x:Tensor) -> Tensor:
-    w1 = self.w1(x).silu()
-    w3 = self.w3(x)
-    return self.w2(w1 * w3)
-
-class TransformerBlock:
-  def __init__(self, dim:int, hidden_dim:int, n_heads:int, n_kv_heads:int|None, norm_eps:float, linear=nn.Linear):
-    self.attention = Attention(dim, n_heads, n_kv_heads, linear)
-    self.feed_forward = FeedForward(dim, hidden_dim, linear)
-    self.attention_norm = nn.RMSNorm(dim, norm_eps)
-    self.ffn_norm = nn.RMSNorm(dim, norm_eps)
-
-  def __call__(self, x:Tensor, freqs_cis:Tensor):
-    h = x + self.attention(self.attention_norm(x), freqs_cis)
-    return h + self.feed_forward(self.ffn_norm(h))
-
-class Transformer:
-  def __init__(self, dim:int, hidden_dim:int, n_heads:int, n_layers:int, norm_eps:float, vocab_size:int, n_kv_heads:int|None=None,
-               rope_theta:int=10000, max_context:int=1024, linear=nn.Linear, embedding=nn.Embedding):
-    self.layers = [TransformerBlock(dim, hidden_dim, n_heads, n_kv_heads, norm_eps, linear) for _ in range(n_layers)]
-    self.norm = nn.RMSNorm(dim, norm_eps)
-    self.tok_embeddings = embedding(vocab_size, dim)
-    self.output = nn.Linear(dim, vocab_size, bias=False) if embedding == nn.Embedding else linear(dim, vocab_size, bias=False)
-    self.freqs_cis = precompute_freqs_cis(dim // n_heads, max_context * 2, rope_theta).contiguous().requires_grad_(False)
-
-  def __call__(self, tokens:Tensor):
-    h = self.tok_embeddings(tokens)
-    freqs_cis = self.freqs_cis.cast(h.dtype)[:, :tokens.shape[1], :, :, :]
-    for layer in self.layers: h = layer(h, freqs_cis)
-    logits = self.output(self.norm(h))
-    return logits

examples/mlperf/models/test_flat_llama.py

@@ -1,140 +0,0 @@
-import os
-os.environ["WQKV"] = "1"
-import unittest
-import numpy as np
-from tinygrad import Tensor, nn, dtypes
-from tinygrad.nn.state import get_parameters
-from tinygrad.device import is_dtype_supported
-from examples.mlperf.models.llama import Transformer
-from examples.mlperf.models.flat_llama import FlatTransformer
-
-def copy_weights(flat:FlatTransformer, ref:Transformer):
-  n_layers = flat.n_layers
-  Tensor.realize(*nn.state.get_state_dict(ref).values())
-  flat.wqkv.assign(Tensor(np.stack([ref.layers[i].attention.wqkv.weight.numpy() for i in range(n_layers)])))
-  flat.wo.assign(Tensor(np.stack([ref.layers[i].attention.wo.weight.numpy() for i in range(n_layers)])))
-  flat.w1.assign(Tensor(np.stack([ref.layers[i].feed_forward.w1.weight.numpy() for i in range(n_layers)])))
-  flat.w2.assign(Tensor(np.stack([ref.layers[i].feed_forward.w2.weight.numpy() for i in range(n_layers)])))
-  flat.w3.assign(Tensor(np.stack([ref.layers[i].feed_forward.w3.weight.numpy() for i in range(n_layers)])))
-  flat.attention_norm.assign(Tensor(np.stack([ref.layers[i].attention_norm.weight.numpy() for i in range(n_layers)])))
-  flat.ffn_norm.assign(Tensor(np.stack([ref.layers[i].ffn_norm.weight.numpy() for i in range(n_layers)])))
-  flat.norm.weight.assign(Tensor(ref.norm.weight.numpy()))
-  flat.tok_embeddings.weight.assign(Tensor(ref.tok_embeddings.weight.numpy()))
-  flat.output.weight.assign(Tensor(ref.output.weight.numpy()))
-
-class TestFlatLlama(unittest.TestCase):
-  def test_forward_match(self):
-    Tensor.manual_seed(42)
-    params = dict(dim=128, hidden_dim=256, n_heads=4, n_kv_heads=2, n_layers=2, norm_eps=1e-5, vocab_size=1024, rope_theta=10000, max_context=64)
-    ref = Transformer(**params)
-    flat = FlatTransformer(**params)
-    copy_weights(flat, ref)
-    Tensor.realize(*nn.state.get_state_dict(flat).values())
-
-    tokens = Tensor([[1, 50, 100, 999, 2]])
-    ref_logits = ref(tokens).realize()
-    flat_logits = flat(tokens).realize()
-    self.assertEqual(ref_logits.shape, flat_logits.shape)
-    diff = (ref_logits - flat_logits).abs().max().item()
-    self.assertLess(diff, 1e-5, f"forward mismatch: max abs diff {diff}")
-
-  def test_backward_match(self):
-    Tensor.manual_seed(42)
-    params = dict(dim=128, hidden_dim=256, n_heads=4, n_kv_heads=2, n_layers=2, norm_eps=1e-5, vocab_size=1024, rope_theta=10000, max_context=64)
-    ref = Transformer(**params)
-    flat = FlatTransformer(**params)
-    copy_weights(flat, ref)
-
-    for p in get_parameters(ref): p.requires_grad_(True)
-    for p in get_parameters(flat): p.requires_grad_(True)
-    Tensor.realize(*nn.state.get_state_dict(flat).values())
-
-    tokens = Tensor([[1, 50, 100, 999, 2, 10]])
-
-    ref_loss = ref(tokens[:, :-1]).sparse_categorical_crossentropy(tokens[:, 1:])
-    ref_loss.backward()
-    ref_grads = {k: v.grad.numpy() for k, v in nn.state.get_state_dict(ref).items() if v.grad is not None}
-
-    flat_loss = flat(tokens[:, :-1]).sparse_categorical_crossentropy(tokens[:, 1:])
-    flat_loss.backward()
-    flat_grads = {k: v.grad.numpy() for k, v in nn.state.get_state_dict(flat).items() if v.grad is not None}
-
-    # check loss matches
-    self.assertAlmostEqual(ref_loss.item(), flat_loss.item(), places=4)
-
-    # check output weight grad matches
-    diff = abs(ref_grads["output.weight"] - flat_grads["output.weight"]).max()
-    self.assertLess(diff, 1e-4, f"output.weight grad mismatch: max abs diff {diff}")
-
-    # check per-layer weight grads match
-    for i in range(params["n_layers"]):
-      for flat_key, ref_key in [
-        ("wqkv", f"layers.{i}.attention.wqkv.weight"),
-        ("wo", f"layers.{i}.attention.wo.weight"),
-        ("w1", f"layers.{i}.feed_forward.w1.weight"),
-        ("w2", f"layers.{i}.feed_forward.w2.weight"),
-        ("w3", f"layers.{i}.feed_forward.w3.weight"),
-      ]:
-        diff = abs(ref_grads[ref_key] - flat_grads[flat_key][i]).max()
-        self.assertLess(diff, 1e-4, f"layer {i} {flat_key} grad mismatch: max abs diff {diff}")
-
-  @unittest.skipUnless(os.getenv("CPU", "") == "1", "multi-device CPU test")
-  def test_forward_match_mp(self):
-    Tensor.manual_seed(42)
-    params = dict(dim=128, hidden_dim=256, n_heads=4, n_kv_heads=2, n_layers=2, norm_eps=1e-5, vocab_size=1024, rope_theta=10000, max_context=64)
-    from tinygrad import Device
-    devices = (f"{Device.DEFAULT}:0", f"{Device.DEFAULT}:1")
-    ref = Transformer(**params)
-    flat = FlatTransformer(**params)
-    copy_weights(flat, ref)
-    Tensor.realize(*nn.state.get_state_dict(flat).values())
-    flat.shard(devices, mp=True)
-
-    tokens = Tensor([[1, 50, 100, 999, 2]], device=devices[0])
-    ref_logits = ref(tokens.to(devices[0])).numpy()
-    flat_logits = flat(tokens.shard(devices)).numpy()
-    self.assertEqual(ref_logits.shape, flat_logits.shape)
-    np.testing.assert_allclose(flat_logits, ref_logits, atol=1e-4, rtol=1e-4)
-
-  @unittest.skipUnless(os.getenv("CPU", "") == "1", "multi-device CPU test")
-  def test_forward_match_dp(self):
-    Tensor.manual_seed(42)
-    params = dict(dim=128, hidden_dim=256, n_heads=4, n_kv_heads=2, n_layers=2, norm_eps=1e-5, vocab_size=1024, rope_theta=10000, max_context=64)
-    from tinygrad import Device
-    devices = (f"{Device.DEFAULT}:0", f"{Device.DEFAULT}:1")
-    ref = Transformer(**params)
-    flat = FlatTransformer(**params)
-    copy_weights(flat, ref)
-    Tensor.realize(*nn.state.get_state_dict(flat).values())
-    flat.shard(devices)
-
-    tokens = Tensor([[1, 50, 100, 999, 2], [2, 100, 50, 1, 999]], device=devices[0])
-    ref_logits = ref(tokens.to(devices[0])).numpy()
-    flat_logits = flat(tokens.shard(devices, axis=0)).numpy()
-    self.assertEqual(ref_logits.shape, flat_logits.shape)
-    np.testing.assert_allclose(flat_logits, ref_logits, atol=1e-4, rtol=1e-4)
-
-  @unittest.skipUnless(is_dtype_supported(dtypes.fp8e4m3), "fp8 not supported on this device")
-  def test_forward_fp8(self):
-    import examples.mlperf.models.flat_llama as flat_llama_mod
-    old_fp8 = flat_llama_mod.FP8
-    try:
-      flat_llama_mod.FP8 = 1
-      Tensor.manual_seed(42)
-      params = dict(dim=128, hidden_dim=256, n_heads=4, n_kv_heads=2, n_layers=2, norm_eps=1e-5, vocab_size=1024, rope_theta=10000, max_context=64)
-      ref = Transformer(**params)
-      flat = FlatTransformer(**params)
-      copy_weights(flat, ref)
-      Tensor.realize(*nn.state.get_state_dict(flat).values())
-
-      tokens = Tensor([[1, 50, 100, 999, 2]])
-      ref_logits = ref(tokens).numpy()
-      flat_logits = flat(tokens).numpy()
-      self.assertEqual(ref_logits.shape, flat_logits.shape)
-      # FP8 has lower precision, allow larger tolerance
-      np.testing.assert_allclose(flat_logits, ref_logits, atol=1.0, rtol=0.1)
-    finally:
-      flat_llama_mod.FP8 = old_fp8
-
-if __name__ == "__main__":
-  unittest.main()

examples/mlperf/optim.py

@@ -1,59 +0,0 @@
-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]]:
-    grads = list(grads)
-
-    for i in range(len(grads)):
-      if grads[i].device != self.m[i].device: grads[i] = 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)
-      total_norm = Tensor.stack(*[g.float().square().sum() for g in grads]).sum().sqrt().contiguous()
-      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))
-
-    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)).cast(self.m[i].dtype)
-      v_hat = (self.v[i] / (1.0 - self.b2_t)).cast(self.v[i].dtype)
-      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:
-    wd = self.wd if t.ndim >= 3 else 0.0
-    up = up.shard_like(t) + self.lr.to(t.device) * 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 CHECK_OOB=0
+export IGNORE_OOB=1
 
 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 CHECK_OOB=0
+export IGNORE_OOB=1
 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 CHECK_OOB=0
+export IGNORE_OOB=1
 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 CHECK_OOB=0
+export IGNORE_OOB=1
 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

@@ -1,20 +0,0 @@
-#!/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

@@ -1,24 +0,0 @@
-#!/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 CHECK_OOB=0
+export IGNORE_OOB=1
 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 CHECK_OOB=0
+export IGNORE_OOB=1
 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 CHECK_OOB=0
+export IGNORE_OOB=1
 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 CHECK_OOB=0
+export IGNORE_OOB=1
 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 CHECK_OOB=0
+export IGNORE_OOB=1
 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 CHECK_OOB=0
+export IGNORE_OOB=1
 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 CHECK_OOB=0
+export IGNORE_OOB=1
 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/llama405b/implementations/tinybox_8xMI350X/dev_beam.sh

@@ -1,38 +0,0 @@
-#!/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:-2}
-export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
-export ALL2ALL=${ALL2ALL:-1}
-export USE_ATOMICS=${USE_ATOMICS:-0}
-export ASM_GEMM=${ASM_GEMM:-1}
-export WQKV=${WQKV:-1}
-export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-1}
-
-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
-export DP=${DP:-1} MP=${MP:-8}
-export BS=${BS:-1} EVAL_BS=${EVAL_BS:-1} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
-
-export MODEL="llama3"
-export BASEDIR="/raid/datasets/c4/"
-export LLAMA3_SIZE=${LLAMA3_SIZE:-"405B"}
-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/llama405b/implementations/tinybox_8xMI350X/dev_run.sh

@@ -1,32 +0,0 @@
-#!/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:-0}
-export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
-export ALL2ALL=${ALL2ALL:-1}
-export USE_ATOMICS=${USE_ATOMICS:-0}
-export ASM_GEMM=${ASM_GEMM:-1}
-export WQKV=${WQKV:-1}
-export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-1}
-
-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
-export DP=${DP:-1} MP=${MP:-8}
-export BS=${BS:-1} EVAL_BS=${EVAL_BS:-1} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-1152}
-
-export MODEL="llama3"
-export BASEDIR="/raid/datasets/c4/"
-export LLAMA3_SIZE=${LLAMA3_SIZE:-"405B"}
-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/dev_beam.sh

@@ -1,42 +0,0 @@
-#!/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:-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} MP=${MP:-1} BS=${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="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 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_beam_mp.sh

@@ -1,43 +0,0 @@
-#!/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:-2}
-export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
-export ALL2ALL=${ALL2ALL:-1}
-export USE_ATOMICS=${USE_ATOMICS:-0}
-export ASM_GEMM=${ASM_GEMM:-1}
-export WQKV=${WQKV:-1}
-export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-1}
-
-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
-export DP=${DP:-1} MP=${MP:-8} BS=${BS:-1} EVAL_BS=${EVAL_BS:-1} 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=${LLAMA3_SIZE:-"8B"}
-export EVAL_TARGET=3.3 EVAL_FREQ=12288
-export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 MAX_STEPS=1200000
-export WARMUP_STEPS=$((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

@@ -1,37 +0,0 @@
-#!/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} MP=${MP:-1} BS=${BS:-8} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-4}
-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="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 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/dev_run_mp.sh

@@ -1,38 +0,0 @@
-#!/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=${USE_ATOMICS:-0}
-export ASM_GEMM=${ASM_GEMM:-1}
-export WQKV=${WQKV:-1}
-export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-1}
-
-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
-export DP=${DP:-1} MP=${MP:-8} BS=${BS:-1} EVAL_BS=${EVAL_BS:-1} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-32}
-export GBS=$((BS * GRADIENT_ACC_STEPS))
-
-export MODEL="llama3"
-export BASEDIR="/raid/datasets/c4-8b/"
-export SMALL=1
-export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
-export EVAL_TARGET=3.3 EVAL_FREQ=12288
-export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 MAX_STEPS=1200000
-export WARMUP_STEPS=$((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

@@ -1,5 +0,0 @@
-#!/bin/bash
-export BENCHMARK=5
-export EVAL_BS=0
-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" --limit 20

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

@@ -1,10 +0,0 @@
-#!/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/mnist_gan.py

@@ -1,11 +1,12 @@
 from pathlib import Path
+import numpy as np
 import torch
 from torchvision.utils import make_grid, save_image
 from tinygrad.nn.state import get_parameters
 from tinygrad.tensor import Tensor
 from tinygrad.helpers import trange
 from tinygrad.nn import optim
-from tinygrad.nn.datasets import mnist
+from extra.datasets import fetch_mnist
 
 class LinearGen:
   def __init__(self):
@@ -37,14 +38,14 @@ class LinearDisc:
     return x
 
 def make_batch(images):
-  sample = Tensor.randint(batch_size, low=0, high=images.shape[0])
-  return images[sample].reshape(batch_size, 28*28).cast('float').div(127.5).sub(1.0)
+  sample = np.random.randint(0, len(images), size=(batch_size))
+  image_b = images[sample].reshape(-1, 28*28).astype(np.float32) / 127.5 - 1.0
+  return Tensor(image_b)
 
 def make_labels(bs, col, val=-2.0):
-  y = Tensor.zeros(bs, 2)
-  if col == 0: y = y + Tensor([val, 0.0])
-  else: y = y + Tensor([0.0, val])
-  return y
+  y = np.zeros((bs, 2), np.float32)
+  y[range(bs), [col] * bs] = val  # Can we do label smoothing? i.e -2.0 changed to -1.98789.
+  return Tensor(y)
 
 def train_discriminator(optimizer, data_real, data_fake):
   real_labels = make_labels(batch_size, 1)
@@ -70,12 +71,12 @@ def train_generator(optimizer, data_fake):
 
 if __name__ == "__main__":
   # data for training and validation
-  X_train, _, _, _ = mnist()
+  images_real = np.vstack(fetch_mnist()[::2])
   ds_noise = Tensor.randn(64, 128, requires_grad=False)
   # parameters
   epochs, batch_size, k = 300, 512, 1
   sample_interval = epochs // 10
-  n_steps = X_train.shape[0] // batch_size
+  n_steps = len(images_real) // batch_size
   # models and optimizer
   generator = LinearGen()
   discriminator = LinearDisc()
@@ -83,24 +84,24 @@ if __name__ == "__main__":
   output_dir = Path(".").resolve() / "outputs"
   output_dir.mkdir(exist_ok=True)
   # optimizers
-  optim_g = optim.Adam(get_parameters(generator), lr=0.0002, b1=0.5)  # 0.0002 for equilibrium!
-  optim_d = optim.Adam(get_parameters(discriminator), lr=0.0002, b1=0.5)
+  optim_g = optim.Adam(get_parameters(generator),lr=0.0002, b1=0.5)  # 0.0002 for equilibrium!
+  optim_d = optim.Adam(get_parameters(discriminator),lr=0.0002, b1=0.5)
   # training loop
-  with Tensor.train():
-    for epoch in (t := trange(epochs)):
-      loss_g, loss_d = 0.0, 0.0
-      for _ in range(n_steps):
-        data_real = make_batch(X_train)
-        for step in range(k):  # Try with k = 5 or 7.
-          noise = Tensor.randn(batch_size, 128)
-          data_fake = generator.forward(noise).detach()
-          loss_d += train_discriminator(optim_d, data_real, data_fake)
+  Tensor.training = True
+  for epoch in (t := trange(epochs)):
+    loss_g, loss_d = 0.0, 0.0
+    for _ in range(n_steps):
+      data_real = make_batch(images_real)
+      for step in range(k):  # Try with k = 5 or 7.
         noise = Tensor.randn(batch_size, 128)
-        data_fake = generator.forward(noise)
-        loss_g += train_generator(optim_g, data_fake)
-      if (epoch + 1) % sample_interval == 0:
-        fake_images = generator.forward(ds_noise).detach().numpy()
-        fake_images = (fake_images.reshape(-1, 1, 28, 28) + 1) / 2  # 0 - 1 range.
-        save_image(make_grid(torch.tensor(fake_images)), output_dir / f"image_{epoch+1}.jpg")
-      t.set_description(f"Generator loss: {loss_g/n_steps}, Discriminator loss: {loss_d/n_steps}")
+        data_fake = generator.forward(noise).detach()
+        loss_d += train_discriminator(optim_d, data_real, data_fake)
+      noise = Tensor.randn(batch_size, 128)
+      data_fake = generator.forward(noise)
+      loss_g += train_generator(optim_g, data_fake)
+    if (epoch + 1) % sample_interval == 0:
+      fake_images = generator.forward(ds_noise).detach().numpy()
+      fake_images = (fake_images.reshape(-1, 1, 28, 28) + 1) / 2  # 0 - 1 range.
+      save_image(make_grid(torch.tensor(fake_images)), output_dir / f"image_{epoch+1}.jpg")
+    t.set_description(f"Generator loss: {loss_g/n_steps}, Discriminator loss: {loss_d/n_steps}")
   print("Training Completed!")

examples/openpilot/compile3.py

@@ -31,7 +31,7 @@ def compile(onnx_file):
   for i in range(3):
     GlobalCounters.reset()
     print(f"run {i}")
-    with Context(DEBUG=max(DEBUG.value, 2 if i == 2 else 1), OPENPILOT_HACKS=1):
+    with Context(DEBUG=max(DEBUG.value, 2 if i == 2 else 1)):
       ret = run_onnx_jit(**inputs).numpy()
     # copy i == 1 so use of JITBEAM is okay
     if i == 1: test_val = np.copy(ret)

examples/openpilot/load_pickle.py

@@ -1,16 +0,0 @@
-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

@@ -6,6 +6,7 @@ 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
@@ -335,7 +336,6 @@ 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_input_info[0][-1]
+  device = run_onnx_jit.captured.expected_st_vars_dtype_device[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_memoryview())
+        data = bytes(state_dict[name].uop.base.realized.as_buffer())
         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.randint(1, 77, low=0, high=49408, dtype=dtypes.int32)], forward = model.cond_stage_model.transformer.text_model),
+    Step(name = "textModel", input = [Tensor.randn(1, 77)], 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

@@ -7,7 +7,6 @@ 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
 
@@ -160,7 +159,7 @@ 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 = mel(sr=RATE, n_fft=N_FFT, n_mels=N_MELS).numpy() @ magnitudes
+  mel_spec = librosa.filters.mel(sr=RATE, n_fft=N_FFT, n_mels=N_MELS) @ 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

@@ -65,7 +65,7 @@ def get_bar0_size(pcibus):
 class AMSMI(AMDev):
   def __init__(self, pcibus, vram_bar:MMIOInterface, doorbell_bar:MMIOInterface, mmio_bar:MMIOInterface):
     self.pcibus = pcibus
-    self.vram, self.doorbell64, self.mmio = vram_bar, doorbell_bar, mmio_bar
+    self.vram, self.doorbell64, self.mmio, self.dma_regions = vram_bar, doorbell_bar, mmio_bar, None
     self.pci_state = self.read_pci_state()
     if self.pci_state == "D0": self._init_from_d0()
 
@@ -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:", "Processing accelerators:"]
+    remove_parts = ["Advanced Micro Devices, Inc. [AMD/ATI]", "VGA compatible controller:"]
     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.MetricsTableV0_t
-        case (13,0,12): table_t = dev.smu.smu_mod.MetricsTable_t
+        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 _: 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)|(13,0,12): return max(0, min(100, self._smuq10_round(metrics.SocketGfxBusy)))
+      case (13,0,6): 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)|(13,0,12): return max(0, min(100, self._smuq10_round(metrics.DramBandwidthUtilization)))
+      case (13,0,6): 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)|(13,0,12):
+      case (13,0,6):
         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)|(13,0,12): return {}
+      case (13,0,6): 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]
@@ -205,37 +205,38 @@ 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)|(13,0,12): return self._smuq10_round(metrics.GfxclkFrequency[0])
+      case (13,0,6): 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)|(13,0,12): return self._smuq10_round(metrics.UclkFrequency)
+      case (13,0,6): 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)|(13,0,12): return self._smuq10_round(metrics.FclkFrequency)
+      case (13,0,6): 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)|(13,0,12): return None, None
+      case (13,0,6): 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:
@@ -244,8 +245,8 @@ class SMICtx:
         entry = pt.entries[i]
 
         if (entry & am.AMDGPU_PTE_VALID) == 0: continue
-        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))
+        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))
           continue
         if (entry & am.AMDGPU_PTE_SYSTEM) != 0: continue
         usage += (1 << ((9 * (3-pt.lv)) + 12))
@@ -279,7 +280,7 @@ 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, activity_line_width, opt_text=mem_fmt)}"] \
+                    + [f"MEM Usage    {draw_bar((mem_used / mem_total) / 100, 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

@@ -1,18 +1,12 @@
 #!/usr/bin/env python3
 
-import os
 from tinygrad.helpers import Context
-from tinygrad.runtime.support.system import System, PCIDevice
-from tinygrad.runtime.support.hcq import FileIOInterface
+from tinygrad.runtime.support.system import System, PCIDevice, PCIDevImplBase
 from tinygrad.runtime.support.am.amdev import AMDev
 
 if __name__ == "__main__":
   gpus = System.pci_scan_bus(0x1002, [(0xffff, [0x74a1, 0x75a0])])
-  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) for gpu in gpus]
+  pcidevs = [PCIDevice(f"reset:{gpu}", gpu, bars=[0, 2, 5]) for gpu in gpus]
   amdevs = []
   with Context(DEBUG=2):
     for pcidev in pcidevs:

extra/amdpci/proclogs.py

@@ -7,8 +7,8 @@ class GFXFake:
   def __init__(self): self.xccs = 8
 
 class AMDFake(AMDev):
-  def __init__(self, pci_dev):
-    self.pci_dev, self.devfmt = pci_dev, pci_dev.pcibus
+  def __init__(self, pci_dev, dma_regions=None):
+    self.pci_dev, self.devfmt, self.dma_regions = pci_dev, pci_dev.pcibus, dma_regions
     self.vram, self.doorbell64, self.mmio = self.pci_dev.map_bar(0), self.pci_dev.map_bar(2, fmt='Q'), self.pci_dev.map_bar(5, fmt='I')
     self._run_discovery()
     self._build_regs()
@@ -19,9 +19,8 @@ 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, register_objects=None, *, only_xcc0: bool = False):
+def parse_amdgpu_logs(log_content, register_names=None, *, only_xcc0: bool = False):
   register_map = register_names or {}
-  register_objs = register_objects or {}
 
   def replace_register(match):
     reg = match.group(1)
@@ -38,28 +37,6 @@ def parse_amdgpu_logs(log_content, register_names=None, register_objects=None, *
   # 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 = []
@@ -73,18 +50,16 @@ 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, reg_objs, only_xcc0=only_xcc0)
+  processed_log = parse_amdgpu_logs(log_content, reg_names, only_xcc0=only_xcc0)
 
   with open(sys.argv[2], 'w') as f:
     f.write(processed_log)

extra/assembly/amd/README

@@ -0,0 +1,31 @@
+An integrated environment for AMD GPU assembly and emulation
+
+Test with `PYTHONPATH="." pytest -n12 extra/assembly/amd/`
+`AMD_LLVM=1 PYTHONPATH="." pytest -n12 extra/assembly/amd/`
+
+* pdf.py -- extract assembly format + instruction psuedocode from AMD PDF
+* dsl.py -- helpers for the autogen instruction classes in `__init__.py`. should be standalone with init
+* pcode.py -- psuedocode execution environment. psuedocode should be transformed as little as possible.
+* asm.py -- an asm/disasm function to transform to and from AMD assembly syntax
+* emu.py -- an emulator for RDNA that runs in tinygrad with `AMD=1 MOCKGPU=1 PYTHON_REMU=1`
+
+The code should be as readable and deduplicated as possible. asm and emu shouldn't be required for dsl.
+
+test_emu.py has a good set of instruction tests for the emulation, with USE_HW=1 it will compare to real hardware.
+Whenever an instruction is fixed, regression tests should be added here and confirmed with real hardware.
+
+test_llvm.py tests asm/disasm on the LLVM tests, confirming it behaves the same as LLVM.
+
+tinygrad's dtype tests should pass with and without LLVM. they run in about 12 seconds.
+
+`PYTHONPATH="." AMD=1 PYTHON_REMU=1 MOCKGPU=1 AMD_LLVM=0 pytest -n=12 test/test_dtype_alu.py test/test_dtype.py`
+`PYTHONPATH="." AMD=1 PYTHON_REMU=1 MOCKGPU=1 AMD_LLVM=1 pytest -n=12 test/test_dtype_alu.py test/test_dtype.py`
+
+The ops tests also pass, but they are very slow, so you should run them one at a time.
+
+`SKIP_SLOW_TEST=1 PYTHONPATH="." AMD=1 PYTHON_REMU=1 MOCKGPU=1 AMD_LLVM=0 pytest -n=12 test/test_ops.py`
+`SKIP_SLOW_TEST=1 PYTHONPATH="." AMD=1 PYTHON_REMU=1 MOCKGPU=1 AMD_LLVM=1 pytest -n=12 test/test_ops.py`
+
+When something is caught by main tinygrad tests, a local regression test should be added to `extra/assembly/amd/test`. While working with tinygrad, you can dump the assembly with `DEBUG=7`. These tests all pass on real hardware, so if a test is failing with `AMD=1 PYTHON_REMU=1 MOCKGPU=1` it's likely because an instruction is emulated incorrectly. You can test without `MOCKGPU=1` to test on real hardware, if it works on real hardware there's a bug in the emulator.
+
+Currently, only RDNA3 is well supported, but when finished, this will support RDNA3+RDNA4+CDNA in ~2000 lines. Count lines with `cloc --by-file extra/assembly/amd/*.py`

extra/assembly/amd/asm.py

@@ -0,0 +1,581 @@
+# RDNA3 assembler and disassembler
+from __future__ import annotations
+import re
+from extra.assembly.amd.dsl import Inst, RawImm, Reg, SrcMod, SGPR, VGPR, TTMP, s, v, ttmp, _RegFactory
+from extra.assembly.amd.dsl import VCC_LO, VCC_HI, VCC, EXEC_LO, EXEC_HI, EXEC, SCC, M0, NULL, OFF
+from extra.assembly.amd.dsl import SPECIAL_GPRS, SPECIAL_PAIRS, FLOAT_DEC, FLOAT_ENC, decode_src
+from extra.assembly.amd.autogen.rdna3 import ins
+from extra.assembly.amd.autogen.rdna3.ins import (VOP1, VOP2, VOP3, VOP3SD, VOP3P, VOPC, VOPD, VINTERP, SOP1, SOP2, SOPC, SOPK, SOPP, SMEM, DS, FLAT, MUBUF, MTBUF, MIMG, EXP,
+  VOP1Op, VOP2Op, VOP3Op, VOP3SDOp, VOPDOp, SOP1Op, SOPKOp, SOPPOp, SMEMOp, DSOp, MUBUFOp)
+
+def _matches_encoding(word: int, cls: type[Inst]) -> bool:
+  """Check if word matches the encoding pattern of an instruction class."""
+  if cls._encoding is None: return False
+  bf, val = cls._encoding
+  return ((word >> bf.lo) & bf.mask()) == val
+
+# Order matters: more specific encodings first, VOP2 last (it's a catch-all for bit31=0)
+_FORMATS_64 = [VOPD, VOP3P, VINTERP, VOP3, DS, FLAT, MUBUF, MTBUF, MIMG, SMEM, EXP]
+_FORMATS_32 = [SOP1, SOPC, SOPP, SOPK, VOPC, VOP1, SOP2, VOP2]  # SOP2/VOP2 are catch-alls
+
+def detect_format(data: bytes) -> type[Inst]:
+  """Detect instruction format from machine code bytes."""
+  assert len(data) >= 4, f"need at least 4 bytes, got {len(data)}"
+  word = int.from_bytes(data[:4], 'little')
+  # Check 64-bit formats first (bits[31:30] == 0b11)
+  if (word >> 30) == 0b11:
+    for cls in _FORMATS_64:
+      if _matches_encoding(word, cls):
+        return VOP3SD if cls is VOP3 and ((word >> 16) & 0x3ff) in Inst._VOP3SD_OPS else cls
+    raise ValueError(f"unknown 64-bit format word={word:#010x}")
+  # 32-bit formats
+  for cls in _FORMATS_32:
+    if _matches_encoding(word, cls): return cls
+  raise ValueError(f"unknown 32-bit format word={word:#010x}")
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# CONSTANTS
+# ═══════════════════════════════════════════════════════════════════════════════
+
+HWREG = {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.items()}
+MSG = {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'}
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# HELPERS
+# ═══════════════════════════════════════════════════════════════════════════════
+
+def _reg(p: str, b: int, n: int = 1) -> str: return f"{p}{b}" if n == 1 else f"{p}[{b}:{b+n-1}]"
+def _sreg(b: int, n: int = 1) -> str: return _reg("s", b, n)
+def _vreg(b: int, n: int = 1) -> str: return _reg("v", b, n)
+def _ttmp(b: int, n: int = 1) -> str: return _reg("ttmp", b - 108, n) if 108 <= b <= 123 else None
+def _sreg_or_ttmp(b: int, n: int = 1) -> str: return _ttmp(b, n) or _sreg(b, n)
+
+def _fmt_sdst(v: int, n: int = 1) -> str:
+  if v == 124: return "null"
+  if t := _ttmp(v, n): return t
+  if n > 1: return SPECIAL_PAIRS.get(v) or _sreg(v, n)
+  return SPECIAL_GPRS.get(v, f"s{v}")
+
+def _fmt_src(v: int, n: int = 1) -> str:
+  if n == 1: return decode_src(v)
+  if v >= 256: return _vreg(v - 256, n)
+  if v <= 105: return _sreg(v, n)
+  if n == 2 and v in SPECIAL_PAIRS: return SPECIAL_PAIRS[v]
+  if t := _ttmp(v, n): return t
+  return decode_src(v)
+
+def _fmt_v16(v: int, base: int = 256, hi_thresh: int = 384) -> str:
+  return f"v{(v - base) & 0x7f}.{'h' if v >= hi_thresh else 'l'}"
+
+def waitcnt(vmcnt: int = 0x3f, expcnt: int = 0x7, lgkmcnt: int = 0x3f) -> int:
+  return (expcnt & 0x7) | ((lgkmcnt & 0x3f) << 4) | ((vmcnt & 0x3f) << 10)
+
+def _has(op: str, *subs) -> bool: return any(s in op for s in subs)
+def _omod(v: int) -> str: return {1: " mul:2", 2: " mul:4", 3: " div:2"}.get(v, "")
+def _src16(inst, v: int) -> str: return _fmt_v16(v) if v >= 256 else inst.lit(v)  # format 16-bit src: vgpr.h/l or literal
+def _mods(*pairs) -> str: return " ".join(m for c, m in pairs if c)
+def _fmt_bits(label: str, val: int, count: int) -> str: return f"{label}:[{','.join(str((val >> i) & 1) for i in range(count))}]"
+
+def _vop3_src(inst, v: int, neg: int, abs_: int, hi: int, n: int, f16: bool, any_hi: bool) -> str:
+  """Format VOP3 source operand with modifiers."""
+  if n > 1: s = _fmt_src(v, n)
+  elif f16 and v >= 256: s = f"v{v - 256}.h" if hi else (f"v{v - 256}.l" if any_hi else inst.lit(v))
+  else: s = inst.lit(v)
+  if abs_: s = f"|{s}|"
+  return f"-{s}" if neg else s
+
+def _opsel_str(opsel: int, n: int, need: bool, is16_d: bool) -> str:
+  """Format op_sel modifier string."""
+  if not need: return ""
+  if is16_d and (opsel & 8): return f" op_sel:[1,1,1{',1' if n == 3 else ''}]"
+  if n == 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}]"
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# DISASSEMBLER
+# ═══════════════════════════════════════════════════════════════════════════════
+
+def _disasm_vop1(inst: VOP1) -> str:
+  name = inst.op_name.lower()
+  if inst.op in (VOP1Op.V_NOP, VOP1Op.V_PIPEFLUSH): return name
+  if inst.op == VOP1Op.V_READFIRSTLANE_B32: return f"v_readfirstlane_b32 {decode_src(inst.vdst)}, v{inst.src0 - 256 if inst.src0 >= 256 else inst.src0}"
+  # 16-bit dst: uses .h/.l suffix (determined by name pattern, not dtype - e.g. sat_pk_u8_i16 outputs 8-bit but uses 16-bit encoding)
+  parts = name.split('_')
+  is_16d = any(p in ('f16','i16','u16','b16') for p in parts[-2:-1]) or (len(parts) >= 2 and parts[-1] in ('f16','i16','u16','b16') and 'cvt' not in name)
+  dst = _vreg(inst.vdst, inst.dst_regs()) if inst.dst_regs() > 1 else _fmt_v16(inst.vdst, 0, 128) if is_16d else f"v{inst.vdst}"
+  src = _fmt_src(inst.src0, inst.src_regs(0)) if inst.src_regs(0) > 1 else _src16(inst, inst.src0) if inst.is_src_16(0) and 'sat_pk' not in name else inst.lit(inst.src0)
+  return f"{name}_e32 {dst}, {src}"
+
+def _disasm_vop2(inst: VOP2) -> str:
+  name = inst.op_name.lower()
+  suf = "" if inst.op == VOP2Op.V_DOT2ACC_F32_F16 else "_e32"
+  # fmaak: dst = src0 * vsrc1 + K, fmamk: dst = src0 * K + vsrc1
+  if inst.op in (VOP2Op.V_FMAAK_F32, VOP2Op.V_FMAAK_F16): return f"{name}{suf} v{inst.vdst}, {inst.lit(inst.src0)}, v{inst.vsrc1}, 0x{inst._literal:x}"
+  if inst.op in (VOP2Op.V_FMAMK_F32, VOP2Op.V_FMAMK_F16): return f"{name}{suf} v{inst.vdst}, {inst.lit(inst.src0)}, 0x{inst._literal:x}, v{inst.vsrc1}"
+  if inst.is_16bit(): return f"{name}{suf} {_fmt_v16(inst.vdst, 0, 128)}, {_src16(inst, inst.src0)}, {_fmt_v16(inst.vsrc1, 0, 128)}"
+  return f"{name}{suf} v{inst.vdst}, {inst.lit(inst.src0)}, v{inst.vsrc1}" + (", vcc_lo" if inst.op == VOP2Op.V_CNDMASK_B32 else "")
+
+def _disasm_vopc(inst: VOPC) -> str:
+  name = inst.op_name.lower()
+  s0 = _fmt_src(inst.src0, inst.src_regs(0)) if inst.src_regs(0) > 1 else _src16(inst, inst.src0) if inst.is_16bit() else inst.lit(inst.src0)
+  s1 = _vreg(inst.vsrc1, inst.src_regs(1)) if inst.src_regs(1) > 1 else _fmt_v16(inst.vsrc1, 0, 128) if inst.is_16bit() else f"v{inst.vsrc1}"
+  return f"{name}_e32 {s0}, {s1}" if inst.op.value >= 128 else f"{name}_e32 vcc_lo, {s0}, {s1}"
+
+NO_ARG_SOPP = {SOPPOp.S_ENDPGM, SOPPOp.S_BARRIER, SOPPOp.S_WAKEUP, SOPPOp.S_ICACHE_INV,
+               SOPPOp.S_WAIT_IDLE, SOPPOp.S_ENDPGM_SAVED, SOPPOp.S_CODE_END, SOPPOp.S_ENDPGM_ORDERED_PS_DONE}
+
+def _disasm_sopp(inst: SOPP) -> str:
+  name = inst.op_name.lower()
+  if inst.op in NO_ARG_SOPP: return name
+  if inst.op == SOPPOp.S_WAITCNT:
+    vm, exp, lgkm = (inst.simm16 >> 10) & 0x3f, inst.simm16 & 0xf, (inst.simm16 >> 4) & 0x3f
+    p = [f"vmcnt({vm})" if vm != 0x3f else "", f"expcnt({exp})" if exp != 7 else "", f"lgkmcnt({lgkm})" if lgkm != 0x3f else ""]
+    return f"s_waitcnt {' '.join(x for x in p if x) or '0'}"
+  if inst.op == SOPPOp.S_DELAY_ALU:
+    deps, skips = ['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'], ['SAME','NEXT','SKIP_1','SKIP_2','SKIP_3','SKIP_4']
+    id0, skip, id1 = inst.simm16 & 0xf, (inst.simm16 >> 4) & 0x7, (inst.simm16 >> 7) & 0xf
+    dep = lambda v: deps[v-1] if 0 < v <= len(deps) else str(v)
+    p = [f"instid0({dep(id0)})" if id0 else "", f"instskip({skips[skip]})" if skip else "", f"instid1({dep(id1)})" if id1 else ""]
+    return f"s_delay_alu {' | '.join(x for x in p if x) or '0'}"
+  return f"{name} {inst.simm16}" if name.startswith(('s_cbranch', 's_branch')) else f"{name} 0x{inst.simm16:x}"
+
+def _disasm_smem(inst: SMEM) -> str:
+  name = inst.op_name.lower()
+  if inst.op in (SMEMOp.S_GL1_INV, SMEMOp.S_DCACHE_INV): return name
+  off_s = f"{decode_src(inst.soffset)} offset:0x{inst.offset:x}" if inst.offset and inst.soffset != 124 else f"0x{inst.offset:x}" if inst.offset else decode_src(inst.soffset)
+  sbase_idx, sbase_count = inst.sbase * 2, 4 if (8 <= inst.op.value <= 12 or name == 's_atc_probe_buffer') else 2
+  sbase_str = _fmt_src(sbase_idx, sbase_count) if sbase_count == 2 else _sreg(sbase_idx, sbase_count) if sbase_idx <= 105 else _reg("ttmp", sbase_idx - 108, sbase_count)
+  if name in ('s_atc_probe', 's_atc_probe_buffer'): return f"{name} {inst.sdata}, {sbase_str}, {off_s}"
+  return f"{name} {_fmt_sdst(inst.sdata, inst.dst_regs())}, {sbase_str}, {off_s}" + _mods((inst.glc, " glc"), (inst.dlc, " dlc"))
+
+def _disasm_flat(inst: FLAT) -> str:
+  name = inst.op_name.lower()
+  seg = ['flat', 'scratch', 'global'][inst.seg] if inst.seg < 3 else 'flat'
+  instr = f"{seg}_{name.split('_', 1)[1] if '_' in name else name}"
+  off_val = inst.offset if seg == 'flat' else (inst.offset if inst.offset < 4096 else inst.offset - 8192)
+  w = inst.dst_regs() * (2 if 'cmpswap' in name else 1)
+  mods = f"{f' offset:{off_val}' if off_val else ''}{' glc' if inst.glc else ''}{' slc' if inst.slc else ''}{' dlc' if inst.dlc else ''}"
+  # saddr
+  if seg == 'flat' or inst.saddr == 0x7F: saddr_s = ""
+  elif inst.saddr == 124: saddr_s = ", off"
+  elif seg == 'scratch': saddr_s = f", {decode_src(inst.saddr)}"
+  elif inst.saddr in SPECIAL_PAIRS: saddr_s = f", {SPECIAL_PAIRS[inst.saddr]}"
+  elif t := _ttmp(inst.saddr, 2): saddr_s = f", {t}"
+  else: saddr_s = f", {_sreg(inst.saddr, 2) if inst.saddr < 106 else decode_src(inst.saddr)}"
+  # addtid: no addr
+  if 'addtid' in name: return f"{instr} v{inst.data if 'store' in name else inst.vdst}{saddr_s}{mods}"
+  # addr width
+  addr_s = "off" if not inst.sve and seg == 'scratch' else _vreg(inst.addr, 1 if seg == 'scratch' or (inst.saddr not in (0x7F, 124)) else 2)
+  data_s, vdst_s = _vreg(inst.data, w), _vreg(inst.vdst, w // 2 if 'cmpswap' in name else w)
+  if 'atomic' in name:
+    return f"{instr} {vdst_s}, {addr_s}, {data_s}{saddr_s if seg != 'flat' else ''}{mods}" if inst.glc else f"{instr} {addr_s}, {data_s}{saddr_s if seg != 'flat' else ''}{mods}"
+  if 'store' in name: return f"{instr} {addr_s}, {data_s}{saddr_s}{mods}"
+  return f"{instr} {_vreg(inst.vdst, w)}, {addr_s}{saddr_s}{mods}"
+
+def _disasm_ds(inst: DS) -> str:
+  op, name = inst.op, inst.op_name.lower()
+  gds = " gds" if inst.gds else ""
+  off = f" offset:{inst.offset0 | (inst.offset1 << 8)}" if inst.offset0 or inst.offset1 else ""
+  off2 = f" offset0:{inst.offset0} offset1:{inst.offset1}" if inst.offset0 or inst.offset1 else ""
+  w = inst.dst_regs()
+  d0, d1, dst, addr = _vreg(inst.data0, w), _vreg(inst.data1, w), _vreg(inst.vdst, w), f"v{inst.addr}"
+
+  if op == DSOp.DS_NOP: return name
+  if op == DSOp.DS_BVH_STACK_RTN_B32: return f"{name} v{inst.vdst}, {addr}, v{inst.data0}, {_vreg(inst.data1, 4)}{off}{gds}"
+  if 'gws_sema' in name and op != DSOp.DS_GWS_SEMA_BR: return f"{name}{off}{gds}"
+  if 'gws_' in name: return f"{name} {addr}{off}{gds}"
+  if op in (DSOp.DS_CONSUME, DSOp.DS_APPEND): return f"{name} v{inst.vdst}{off}{gds}"
+  if 'gs_reg' in name: return f"{name} {_vreg(inst.vdst, 2)}, v{inst.data0}{off}{gds}"
+  if '2addr' in name:
+    if 'load' in name: return f"{name} {_vreg(inst.vdst, w*2)}, {addr}{off2}{gds}"
+    if 'store' in name and 'xchg' not in name: return f"{name} {addr}, {d0}, {d1}{off2}{gds}"
+    return f"{name} {_vreg(inst.vdst, w*2)}, {addr}, {d0}, {d1}{off2}{gds}"
+  if 'load' in name: return f"{name} v{inst.vdst}{off}{gds}" if 'addtid' in name else f"{name} {dst}, {addr}{off}{gds}"
+  if 'store' in name and not _has(name, 'cmp', 'xchg'):
+    return f"{name} v{inst.data0}{off}{gds}" if 'addtid' in name else f"{name} {addr}, {d0}{off}{gds}"
+  if 'swizzle' in name or op == DSOp.DS_ORDERED_COUNT: return f"{name} v{inst.vdst}, {addr}{off}{gds}"
+  if 'permute' in name: return f"{name} v{inst.vdst}, {addr}, v{inst.data0}{off}{gds}"
+  if 'condxchg' in name: return f"{name} {_vreg(inst.vdst, 2)}, {addr}, {_vreg(inst.data0, 2)}{off}{gds}"
+  if _has(name, 'cmpstore', 'mskor', 'wrap'):
+    return f"{name} {dst}, {addr}, {d0}, {d1}{off}{gds}" if '_rtn' in name else f"{name} {addr}, {d0}, {d1}{off}{gds}"
+  return f"{name} {dst}, {addr}, {d0}{off}{gds}" if '_rtn' in name else f"{name} {addr}, {d0}{off}{gds}"
+
+def _disasm_vop3(inst: VOP3) -> str:
+  op, name = inst.op, inst.op_name.lower()
+
+  # VOP3SD (shared encoding)
+  if isinstance(op, VOP3SDOp):
+    sdst = (inst.clmp << 7) | (inst.opsel << 3) | inst.abs
+    def src(v, neg, n): s = _fmt_src(v, n) if n > 1 else inst.lit(v); return f"-{s}" if neg else s
+    s0, s1, s2 = src(inst.src0, inst.neg & 1, inst.src_regs(0)), src(inst.src1, inst.neg & 2, inst.src_regs(1)), src(inst.src2, inst.neg & 4, inst.src_regs(2))
+    dst = _vreg(inst.vdst, inst.dst_regs()) if inst.dst_regs() > 1 else f"v{inst.vdst}"
+    srcs = f"{s0}, {s1}, {s2}" if inst.num_srcs() == 3 else f"{s0}, {s1}"
+    return f"{name} {dst}, {_fmt_sdst(sdst, 1)}, {srcs}" + _omod(inst.omod)
+
+  # Detect 16-bit operand sizes (for .h/.l suffix handling)
+  is16_d = is16_s = is16_s2 = False
+  if 'cvt_pk' in name: is16_s = name.endswith('16')
+  elif m := re.match(r'v_(?:cvt|frexp_exp)_([a-z0-9_]+)_([a-z0-9]+)', name):
+    is16_d, is16_s = _has(m.group(1), 'f16','i16','u16','b16'), _has(m.group(2), 'f16','i16','u16','b16')
+    is16_s2 = is16_s
+  elif re.match(r'v_mad_[iu]32_[iu]16', name): is16_s = True
+  elif 'pack_b32' in name: is16_s = is16_s2 = True
+  else: is16_d = is16_s = is16_s2 = inst.is_16bit()
+
+  any_hi = inst.opsel != 0
+  s0 = _vop3_src(inst, inst.src0, inst.neg&1, inst.abs&1, inst.opsel&1, inst.src_regs(0), is16_s, any_hi)
+  s1 = _vop3_src(inst, inst.src1, inst.neg&2, inst.abs&2, inst.opsel&2, inst.src_regs(1), is16_s, any_hi)
+  s2 = _vop3_src(inst, inst.src2, inst.neg&4, inst.abs&4, inst.opsel&4, inst.src_regs(2), is16_s2, any_hi)
+
+  # Destination
+  dn = inst.dst_regs()
+  if op == VOP3Op.V_READLANE_B32: dst = _fmt_sdst(inst.vdst, 1)
+  elif dn > 1: dst = _vreg(inst.vdst, dn)
+  elif is16_d: dst = f"v{inst.vdst}.h" if (inst.opsel & 8) else f"v{inst.vdst}.l" if any_hi else f"v{inst.vdst}"
+  else: dst = f"v{inst.vdst}"
+
+  cl, om = " clamp" if inst.clmp else "", _omod(inst.omod)
+  nonvgpr_opsel = (inst.src0 < 256 and (inst.opsel & 1)) or (inst.src1 < 256 and (inst.opsel & 2)) or (inst.src2 < 256 and (inst.opsel & 4))
+  need_opsel = nonvgpr_opsel or (inst.opsel and not is16_s)
+
+  if inst.op < 256:  # VOPC
+    return f"{name}_e64 {s0}, {s1}" if name.startswith('v_cmpx') else f"{name}_e64 {_fmt_sdst(inst.vdst, 1)}, {s0}, {s1}"
+  if inst.op < 384:  # VOP2
+    n = inst.num_srcs()
+    os = _opsel_str(inst.opsel, n, need_opsel, is16_d)
+    return f"{name}_e64 {dst}, {s0}, {s1}, {s2}{os}{cl}{om}" if n == 3 else f"{name}_e64 {dst}, {s0}, {s1}{os}{cl}{om}"
+  if inst.op < 512:  # VOP1
+    return f"{name}_e64" if op in (VOP3Op.V_NOP, VOP3Op.V_PIPEFLUSH) else f"{name}_e64 {dst}, {s0}{_opsel_str(inst.opsel, 1, need_opsel, is16_d)}{cl}{om}"
+  # Native VOP3
+  n = inst.num_srcs()
+  os = _opsel_str(inst.opsel, n, need_opsel, is16_d)
+  return f"{name} {dst}, {s0}, {s1}, {s2}{os}{cl}{om}" if n == 3 else f"{name} {dst}, {s0}, {s1}{os}{cl}{om}"
+
+def _disasm_vop3sd(inst: VOP3SD) -> str:
+  name = inst.op_name.lower()
+  def src(v, neg, n): s = _fmt_src(v, n) if n > 1 else inst.lit(v); return f"-{s}" if neg else s
+  s0, s1, s2 = src(inst.src0, inst.neg & 1, inst.src_regs(0)), src(inst.src1, inst.neg & 2, inst.src_regs(1)), src(inst.src2, inst.neg & 4, inst.src_regs(2))
+  dst = _vreg(inst.vdst, inst.dst_regs()) if inst.dst_regs() > 1 else f"v{inst.vdst}"
+  srcs = f"{s0}, {s1}, {s2}" if inst.num_srcs() == 3 else f"{s0}, {s1}"
+  suffix = "_e64" if name.startswith('v_') and 'co_' in name else ""
+  return f"{name}{suffix} {dst}, {_fmt_sdst(inst.sdst, 1)}, {srcs}{' clamp' if inst.clmp else ''}{_omod(inst.omod)}"
+
+def _disasm_vopd(inst: VOPD) -> str:
+  lit = inst._literal or inst.literal
+  vdst_y, nx, ny = (inst.vdsty << 1) | ((inst.vdstx & 1) ^ 1), VOPDOp(inst.opx).name.lower(), VOPDOp(inst.opy).name.lower()
+  def half(n, vd, s0, vs1): return f"{n} v{vd}, {inst.lit(s0)}{f', 0x{lit:x}' if lit and _has(n, 'fmaak', 'fmamk') else ''}" if 'mov' in n else f"{n} v{vd}, {inst.lit(s0)}, v{vs1}{f', 0x{lit:x}' if lit and _has(n, 'fmaak', 'fmamk') else ''}"
+  return f"{half(nx, inst.vdstx, inst.srcx0, inst.vsrcx1)} :: {half(ny, vdst_y, inst.srcy0, inst.vsrcy1)}"
+
+def _disasm_vop3p(inst: VOP3P) -> str:
+  name = inst.op_name.lower()
+  is_wmma, n, is_fma_mix = 'wmma' in name, inst.num_srcs(), 'fma_mix' in name
+  if is_wmma:
+    sc = 2 if 'iu4' in name else 4 if 'iu8' in name else 8
+    src0, src1, src2, dst = _fmt_src(inst.src0, sc), _fmt_src(inst.src1, sc), _fmt_src(inst.src2, 8), _vreg(inst.vdst, 8)
+  else: src0, src1, src2, dst = _fmt_src(inst.src0, 1), _fmt_src(inst.src1, 1), _fmt_src(inst.src2, 1), f"v{inst.vdst}"
+  opsel_hi = inst.opsel_hi | (inst.opsel_hi2 << 2)
+  if is_fma_mix:
+    def m(s, neg, abs_): return f"-{f'|{s}|' if abs_ else s}" if neg else (f"|{s}|" if abs_ else s)
+    src0, src1, src2 = m(src0, inst.neg & 1, inst.neg_hi & 1), m(src1, inst.neg & 2, inst.neg_hi & 2), m(src2, inst.neg & 4, inst.neg_hi & 4)
+    mods = ([_fmt_bits("op_sel", inst.opsel, n)] if inst.opsel else []) + ([_fmt_bits("op_sel_hi", opsel_hi, n)] if opsel_hi else []) + (["clamp"] if inst.clmp else [])
+  else:
+    mods = ([_fmt_bits("op_sel", inst.opsel, n)] if inst.opsel else []) + ([_fmt_bits("op_sel_hi", opsel_hi, n)] if opsel_hi != (7 if n == 3 else 3) else []) + \
+           ([_fmt_bits("neg_lo", inst.neg, n)] if inst.neg else []) + ([_fmt_bits("neg_hi", inst.neg_hi, n)] if inst.neg_hi else []) + (["clamp"] if inst.clmp else [])
+  return f"{name} {dst}, {src0}, {src1}, {src2}{' ' + ' '.join(mods) if mods else ''}" if n == 3 else f"{name} {dst}, {src0}, {src1}{' ' + ' '.join(mods) if mods else ''}"
+
+def _disasm_buf(inst: MUBUF | MTBUF) -> str:
+  name = inst.op_name.lower()
+  if inst.op in (MUBUFOp.BUFFER_GL0_INV, MUBUFOp.BUFFER_GL1_INV): return name
+  w = (2 if _has(name, 'xyz', 'xyzw') else 1) if 'd16' in name else \
+      ((2 if _has(name, 'b64', 'u64', 'i64') else 1) * (2 if 'cmpswap' in name else 1)) if 'atomic' in name else \
+      {'b32':1,'b64':2,'b96':3,'b128':4,'b16':1,'x':1,'xy':2,'xyz':3,'xyzw':4}.get(name.split('_')[-1], 1)
+  if inst.tfe: w += 1
+  vaddr = _vreg(inst.vaddr, 2) if inst.offen and inst.idxen else f"v{inst.vaddr}" if inst.offen or inst.idxen else "off"
+  srsrc = _sreg_or_ttmp(inst.srsrc*4, 4)
+  mods = ([f"format:{inst.format}"] if isinstance(inst, MTBUF) else []) + [m for c, m in [(inst.idxen,"idxen"),(inst.offen,"offen"),(inst.offset,f"offset:{inst.offset}"),(inst.glc,"glc"),(inst.dlc,"dlc"),(inst.slc,"slc"),(inst.tfe,"tfe")] if c]
+  return f"{name} {_vreg(inst.vdata, w)}, {vaddr}, {srsrc}, {decode_src(inst.soffset)}{' ' + ' '.join(mods) if mods else ''}"
+
+def _mimg_vaddr_width(name: str, dim: int, a16: bool) -> int:
+  """Calculate vaddr register count for MIMG sample/gather operations."""
+  #                    1d,2d,3d,cube,1d_arr,2d_arr,2d_msaa,2d_msaa_arr
+  base =              [1, 2, 3, 3,   2,     3,     3,      4][dim]  # address coords
+  grad =              [1, 2, 3, 2,   1,     2,     2,      2][dim]  # gradient coords (for derivatives)
+  if 'get_resinfo' in name: return 1  # only mip level
+  packed, unpacked = 0, 0
+  if '_mip' in name: packed += 1
+  elif 'sample' in name or 'gather' in name:
+    if '_o' in name: unpacked += 1                                              # offset
+    if re.search(r'_c(_|$)', name): unpacked += 1                               # compare (not _cl)
+    if '_d' in name: unpacked += (grad + 1) & ~1 if '_g16' in name else grad*2  # derivatives
+    if '_b' in name: unpacked += 1                                              # bias
+    if '_l' in name and '_cl' not in name and '_lz' not in name: packed += 1    # LOD
+    if '_cl' in name: packed += 1                                               # clamp
+  return (base + packed + 1) // 2 + unpacked if a16 else base + packed + unpacked
+
+def _disasm_mimg(inst: MIMG) -> str:
+  name = inst.op_name.lower()
+  srsrc_base = inst.srsrc * 4
+  srsrc_str = _sreg_or_ttmp(srsrc_base, 8)
+  # BVH intersect ray: special case with 4 SGPR srsrc
+  if 'bvh' in name:
+    vaddr = (9 if '64' in name else 8) if inst.a16 else (12 if '64' in name else 11)
+    return f"{name} {_vreg(inst.vdata, 4)}, {_vreg(inst.vaddr, vaddr)}, {_sreg_or_ttmp(srsrc_base, 4)}{' a16' if inst.a16 else ''}"
+  # vdata width from dmask (gather4/msaa_load always 4), d16 packs, tfe adds 1
+  vdata = 4 if 'gather4' in name or 'msaa_load' in name else (bin(inst.dmask).count('1') or 1)
+  if inst.d16: vdata = (vdata + 1) // 2
+  if inst.tfe: vdata += 1
+  # vaddr width
+  dim_names = ['1d', '2d', '3d', 'cube', '1d_array', '2d_array', '2d_msaa', '2d_msaa_array']
+  dim = dim_names[inst.dim] if inst.dim < len(dim_names) else f"dim_{inst.dim}"
+  vaddr = _mimg_vaddr_width(name, inst.dim, inst.a16)
+  vaddr_str = f"v{inst.vaddr}" if vaddr == 1 else _vreg(inst.vaddr, vaddr)
+  # modifiers
+  mods = [f"dmask:0x{inst.dmask:x}"] if inst.dmask and (inst.dmask != 15 or 'atomic' in name) else []
+  mods.append(f"dim:SQ_RSRC_IMG_{dim.upper()}")
+  for flag, mod in [(inst.unrm,"unorm"),(inst.glc,"glc"),(inst.slc,"slc"),(inst.dlc,"dlc"),(inst.r128,"r128"),
+                    (inst.a16,"a16"),(inst.tfe,"tfe"),(inst.lwe,"lwe"),(inst.d16,"d16")]:
+    if flag: mods.append(mod)
+  # ssamp for sample/gather/get_lod
+  ssamp_str = ""
+  if 'sample' in name or 'gather' in name or 'get_lod' in name:
+    ssamp_str = ", " + _sreg_or_ttmp(inst.ssamp * 4, 4)
+  return f"{name} {_vreg(inst.vdata, vdata)}, {vaddr_str}, {srsrc_str}{ssamp_str} {' '.join(mods)}"
+
+def _disasm_sop1(inst: SOP1) -> str:
+  op, name = inst.op, inst.op_name.lower()
+  if op == SOP1Op.S_GETPC_B64: return f"{name} {_fmt_sdst(inst.sdst, 2)}"
+  if op in (SOP1Op.S_SETPC_B64, SOP1Op.S_RFE_B64): return f"{name} {_fmt_src(inst.ssrc0, 2)}"
+  if op == SOP1Op.S_SWAPPC_B64: return f"{name} {_fmt_sdst(inst.sdst, 2)}, {_fmt_src(inst.ssrc0, 2)}"
+  if op in (SOP1Op.S_SENDMSG_RTN_B32, SOP1Op.S_SENDMSG_RTN_B64): return f"{name} {_fmt_sdst(inst.sdst, inst.dst_regs())}, sendmsg({MSG.get(inst.ssrc0, str(inst.ssrc0))})"
+  return f"{name} {_fmt_sdst(inst.sdst, inst.dst_regs())}, {inst.lit(inst.ssrc0) if inst.src_regs(0) == 1 else _fmt_src(inst.ssrc0, inst.src_regs(0))}"
+
+def _disasm_sop2(inst: SOP2) -> str:
+  return f"{inst.op_name.lower()} {_fmt_sdst(inst.sdst, inst.dst_regs())}, {inst.lit(inst.ssrc0) if inst.ssrc0 == 255 else _fmt_src(inst.ssrc0, inst.src_regs(0))}, {inst.lit(inst.ssrc1) if inst.ssrc1 == 255 else _fmt_src(inst.ssrc1, inst.src_regs(1))}"
+
+def _disasm_sopc(inst: SOPC) -> str:
+  return f"{inst.op_name.lower()} {_fmt_src(inst.ssrc0, inst.src_regs(0))}, {_fmt_src(inst.ssrc1, inst.src_regs(1))}"
+
+def _disasm_sopk(inst: SOPK) -> str:
+  op, name = inst.op, inst.op_name.lower()
+  if op == SOPKOp.S_VERSION: return f"{name} 0x{inst.simm16:x}"
+  if op in (SOPKOp.S_SETREG_B32, SOPKOp.S_GETREG_B32):
+    hid, hoff, hsz = inst.simm16 & 0x3f, (inst.simm16 >> 6) & 0x1f, ((inst.simm16 >> 11) & 0x1f) + 1
+    hs = f"0x{inst.simm16:x}" if hid in (16, 17) else f"hwreg({HWREG.get(hid, str(hid))}, {hoff}, {hsz})"
+    return f"{name} {hs}, {_fmt_sdst(inst.sdst, 1)}" if op == SOPKOp.S_SETREG_B32 else f"{name} {_fmt_sdst(inst.sdst, 1)}, {hs}"
+  return f"{name} {_fmt_sdst(inst.sdst, inst.dst_regs())}, 0x{inst.simm16:x}"
+
+def _disasm_vinterp(inst: VINTERP) -> str:
+  mods = _mods((inst.waitexp, f"wait_exp:{inst.waitexp}"), (inst.clmp, "clamp"))
+  return f"{inst.op_name.lower()} v{inst.vdst}, {inst.lit(inst.src0, inst.neg & 1)}, {inst.lit(inst.src1, inst.neg & 2)}, {inst.lit(inst.src2, inst.neg & 4)}" + (" " + mods if mods else "")
+
+DISASM_HANDLERS = {VOP1: _disasm_vop1, VOP2: _disasm_vop2, VOPC: _disasm_vopc, VOP3: _disasm_vop3, VOP3SD: _disasm_vop3sd, VOPD: _disasm_vopd, VOP3P: _disasm_vop3p,
+                   VINTERP: _disasm_vinterp, SOPP: _disasm_sopp, SMEM: _disasm_smem, DS: _disasm_ds, FLAT: _disasm_flat, MUBUF: _disasm_buf, MTBUF: _disasm_buf,
+                   MIMG: _disasm_mimg, SOP1: _disasm_sop1, SOP2: _disasm_sop2, SOPC: _disasm_sopc, SOPK: _disasm_sopk}
+
+def disasm(inst: Inst) -> str: return DISASM_HANDLERS[type(inst)](inst)
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# ASSEMBLER
+# ═══════════════════════════════════════════════════════════════════════════════
+
+SPEC_REGS = {'vcc_lo': RawImm(106), 'vcc_hi': RawImm(107), 'vcc': RawImm(106), 'null': RawImm(124), 'off': RawImm(124), 'm0': RawImm(125),
+             'exec_lo': RawImm(126), 'exec_hi': RawImm(127), 'exec': RawImm(126), 'scc': RawImm(253), 'src_scc': RawImm(253)}
+FLOATS = {str(k): k for k in FLOAT_ENC}  # Valid float literal strings: '0.5', '-0.5', '1.0', etc.
+REG_MAP: dict[str, _RegFactory] = {'s': s, 'v': v, 't': ttmp, 'ttmp': ttmp}
+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'}
+SPEC_DSL = {'vcc_lo': 'VCC_LO', 'vcc_hi': 'VCC_HI', 'vcc': 'VCC_LO', 'null': 'NULL', 'off': 'OFF', 'm0': 'M0',
+            'exec_lo': 'EXEC_LO', 'exec_hi': 'EXEC_HI', 'exec': 'EXEC_LO', 'scc': 'SCC', 'src_scc': 'SCC'}
+
+def _op2dsl(op: str) -> str:
+  op = op.strip()
+  neg = op.startswith('-') and not (op[1:2].isdigit() or (len(op) > 2 and op[1] == '0' and op[2] in 'xX'))
+  if neg: op = op[1:]
+  abs_ = (op.startswith('|') and op.endswith('|')) or (op.startswith('abs(') and op.endswith(')'))
+  if abs_: op = op[1:-1] if op.startswith('|') else op[4:-1]
+  hi = ".h" if op.endswith('.h') else ".l" if op.endswith('.l') else ""
+  if hi: op = op[:-2]
+  lo = op.lower()
+  def wrap(b): return f"{'-' if neg else ''}abs({b}){hi}" if abs_ else f"-{b}{hi}" if neg else f"{b}{hi}"
+  if lo in SPEC_DSL: return wrap(SPEC_DSL[lo])
+  if op in FLOATS: return wrap(op)
+  rp = {'s': 's', 'v': 'v', 't': 'ttmp', 'ttmp': 'ttmp'}
+  if m := re.match(r'^([svt](?:tmp)?)\[(\d+):(\d+)\]$', lo): return wrap(f"{rp[m.group(1)]}[{m.group(2)}:{m.group(3)}]")
+  if m := re.match(r'^([svt](?:tmp)?)(\d+)$', lo): return wrap(f"{rp[m.group(1)]}[{m.group(2)}]")
+  if re.match(r'^-?\d+$|^-?0x[0-9a-fA-F]+$', op): return f"SrcMod({op}, neg={neg}, abs_={abs_})" if neg or abs_ else op
+  return wrap(op)
+
+def _parse_ops(s: str) -> list[str]:
+  ops, cur, depth, pipe = [], "", 0, False
+  for c in s:
+    if c in '[(': depth += 1
+    elif c in '])': depth -= 1
+    elif c == '|': pipe = not pipe
+    if c == ',' and depth == 0 and not pipe: ops.append(cur.strip()); cur = ""
+    else: cur += c
+  if cur.strip(): ops.append(cur.strip())
+  return ops
+
+def _extract(text: str, pat: str, flags=re.I):
+  if m := re.search(pat, text, flags): return m, text[:m.start()] + text[m.end():]
+  return None, text
+
+def get_dsl(text: str) -> str:
+  text, kw = text.strip(), []
+  # Extract modifiers
+  for pat, val in [(r'\s+mul:2(?:\s|$)', 1), (r'\s+mul:4(?:\s|$)', 2), (r'\s+div:2(?:\s|$)', 3)]:
+    if (m := _extract(text, pat))[0]: kw.append(f'omod={val}'); text = m[1]; break
+  if (m := _extract(text, r'\s+clamp(?:\s|$)'))[0]: kw.append('clmp=1'); text = m[1]
+  opsel, m, text = None, *_extract(text, r'\s+op_sel:\[([^\]]+)\]')
+  if m:
+    bits, mn = [int(x.strip()) for x in m.group(1).split(',')], text.split()[0].lower()
+    is3p = mn.startswith(('v_pk_', 'v_wmma_', 'v_dot'))
+    opsel = (bits[0] | (bits[1] << 1) | (bits[2] << 2)) if len(bits) == 3 and is3p else \
+            (bits[0] | (bits[1] << 1) | (bits[2] << 3)) if len(bits) == 3 else sum(b << i for i, b in enumerate(bits))
+  m, text = _extract(text, r'\s+wait_exp:(\d+)'); waitexp = m.group(1) if m else None
+  m, text = _extract(text, r'\s+offset:(0x[0-9a-fA-F]+|-?\d+)'); off_val = m.group(1) if m else None
+  m, text = _extract(text, r'\s+dlc(?:\s|$)'); dlc = 1 if m else None
+  m, text = _extract(text, r'\s+glc(?:\s|$)'); glc = 1 if m else None
+  m, text = _extract(text, r'\s+slc(?:\s|$)'); slc = 1 if m else None
+  m, text = _extract(text, r'\s+neg_lo:\[([^\]]+)\]'); neg_lo = sum(int(x.strip()) << i for i, x in enumerate(m.group(1).split(','))) if m else None
+  m, text = _extract(text, r'\s+neg_hi:\[([^\]]+)\]'); neg_hi = sum(int(x.strip()) << i for i, x in enumerate(m.group(1).split(','))) if m else None
+  if waitexp: kw.append(f'waitexp={waitexp}')
+
+  parts = text.replace(',', ' ').split()
+  if not parts: raise ValueError("empty instruction")
+  mn, op_str = parts[0].lower(), text[len(parts[0]):].strip()
+  ops, args = _parse_ops(op_str), [_op2dsl(o) for o in _parse_ops(op_str)]
+
+  # s_waitcnt
+  if mn == 's_waitcnt':
+    vm, exp, lgkm = 0x3f, 0x7, 0x3f
+    for p in op_str.replace(',', ' ').split():
+      if m := re.match(r'vmcnt\((\d+)\)', p): vm = int(m.group(1))
+      elif m := re.match(r'expcnt\((\d+)\)', p): exp = int(m.group(1))
+      elif m := re.match(r'lgkmcnt\((\d+)\)', p): lgkm = int(m.group(1))
+      elif re.match(r'^0x[0-9a-f]+$|^\d+$', p): return f"s_waitcnt(simm16={int(p, 0)})"
+    return f"s_waitcnt(simm16={waitcnt(vm, exp, lgkm)})"
+
+  # VOPD
+  if '::' in text:
+    xp, yp = text.split('::')
+    xps, yps = xp.strip().replace(',', ' ').split(), yp.strip().replace(',', ' ').split()
+    xo, yo = [_op2dsl(p) for p in xps[1:]], [_op2dsl(p) for p in yps[1:]]
+    vdx, sx0, vsx1 = xo[0], xo[1] if len(xo) > 1 else '0', xo[2] if len(xo) > 2 else 'v[0]'
+    vdy, sy0, vsy1 = yo[0], yo[1] if len(yo) > 1 else '0', yo[2] if len(yo) > 2 else 'v[0]'
+    lit = xo[3] if 'fmaak' in xps[0].lower() and len(xo) > 3 else yo[3] if 'fmaak' in yps[0].lower() and len(yo) > 3 else None
+    if 'fmamk' in xps[0].lower() and len(xo) > 3: lit, vsx1 = xo[2], xo[3]
+    elif 'fmamk' in yps[0].lower() and len(yo) > 3: lit, vsy1 = yo[2], yo[3]
+    return f"VOPD(VOPDOp.{xps[0].upper()}, VOPDOp.{yps[0].upper()}, vdstx={vdx}, vdsty={vdy}, srcx0={sx0}, vsrcx1={vsx1}, srcy0={sy0}, vsrcy1={vsy1}{f', literal={lit}' if lit else ''})"
+
+  # Special instructions
+  if mn == 's_setreg_imm32_b32': raise ValueError(f"unsupported: {mn}")
+  if mn in ('s_setpc_b64', 's_rfe_b64'): return f"{mn}(ssrc0={args[0]})"
+  if mn in ('s_sendmsg_rtn_b32', 's_sendmsg_rtn_b64'): return f"{mn}(sdst={args[0]}, ssrc0=RawImm({args[1].strip()}))"
+  if mn == 's_version': return f"{mn}(simm16={args[0]})"
+  if mn == 's_setreg_b32': return f"{mn}(simm16={args[0]}, sdst={args[1]})"
+
+  # SMEM
+  if mn in SMEM_OPS:
+    gs, ds = ", glc=1" if glc else "", ", dlc=1" if dlc else ""
+    if len(ops) >= 3 and re.match(r'^-?[0-9]|^-?0x', ops[2].strip().lower()):
+      return f"{mn}(sdata={args[0]}, sbase={args[1]}, offset={args[2]}, soffset=RawImm(124){gs}{ds})"
+    if off_val and len(ops) >= 3: return f"{mn}(sdata={args[0]}, sbase={args[1]}, offset={off_val}, soffset={args[2]}{gs}{ds})"
+    if len(ops) >= 3: return f"{mn}(sdata={args[0]}, sbase={args[1]}, soffset={args[2]}{gs}{ds})"
+
+  # Buffer
+  if mn.startswith('buffer_') and len(ops) >= 2 and ops[1].strip().lower() == 'off':
+    return f"{mn}(vdata={args[0]}, vaddr=0, srsrc={args[2]}, soffset={f'RawImm({args[3].strip()})' if len(args) > 3 else 'RawImm(0)'})"
+
+  # FLAT/GLOBAL/SCRATCH load/store/atomic - saddr needs RawImm(124) for off/null
+  def _saddr(a): return 'RawImm(124)' if a in ('OFF', 'NULL') else a
+  flat_mods = f"{f', offset={off_val}' if off_val else ''}{', glc=1' if glc else ''}{', slc=1' if slc else ''}{', dlc=1' if dlc else ''}"
+  for pre, flds in [('flat_load','vdst,addr,saddr'), ('global_load','vdst,addr,saddr'), ('scratch_load','vdst,addr,saddr'),
+                    ('flat_store','addr,data,saddr'), ('global_store','addr,data,saddr'), ('scratch_store','addr,data,saddr')]:
+    if mn.startswith(pre) and len(args) >= 2:
+      f0, f1, f2 = flds.split(',')
+      return f"{mn}({f0}={args[0]}, {f1}={args[1]}{f', {f2}={_saddr(args[2])}' if len(args) >= 3 else ', saddr=RawImm(124)'}{flat_mods})"
+  for pre in ('flat_atomic', 'global_atomic', 'scratch_atomic'):
+    if mn.startswith(pre):
+      if glc and len(args) >= 3: return f"{mn}(vdst={args[0]}, addr={args[1]}, data={args[2]}{f', saddr={_saddr(args[3])}' if len(args) >= 4 else ', saddr=RawImm(124)'}{flat_mods})"
+      if len(args) >= 2: return f"{mn}(addr={args[0]}, data={args[1]}{f', saddr={_saddr(args[2])}' if len(args) >= 3 else ', saddr=RawImm(124)'}{flat_mods})"
+
+  # DS instructions
+  if mn.startswith('ds_'):
+    off0, off1 = (str(int(off_val, 0) & 0xff), str((int(off_val, 0) >> 8) & 0xff)) if off_val else ("0", "0")
+    gds_s = ", gds=1" if 'gds' in text.lower().split()[-1:] else ""
+    off_kw = f", offset0={off0}, offset1={off1}{gds_s}"
+    if mn == 'ds_nop' or mn in ('ds_gws_sema_v', 'ds_gws_sema_p', 'ds_gws_sema_release_all'): return f"{mn}({off_kw.lstrip(', ')})"
+    if 'gws_' in mn: return f"{mn}(addr={args[0]}{off_kw})"
+    if 'consume' in mn or 'append' in mn: return f"{mn}(vdst={args[0]}{off_kw})"
+    if 'gs_reg' in mn: return f"{mn}(vdst={args[0]}, data0={args[1]}{off_kw})"
+    if '2addr' in mn:
+      if 'load' in mn: return f"{mn}(vdst={args[0]}, addr={args[1]}{off_kw})"
+      if 'store' in mn and 'xchg' not in mn: return f"{mn}(addr={args[0]}, data0={args[1]}, data1={args[2]}{off_kw})"
+      return f"{mn}(vdst={args[0]}, addr={args[1]}, data0={args[2]}, data1={args[3]}{off_kw})"
+    if 'load' in mn: return f"{mn}(vdst={args[0]}{off_kw})" if 'addtid' in mn else f"{mn}(vdst={args[0]}, addr={args[1]}{off_kw})"
+    if 'store' in mn and not _has(mn, 'cmp', 'xchg'):
+      return f"{mn}(data0={args[0]}{off_kw})" if 'addtid' in mn else f"{mn}(addr={args[0]}, data0={args[1]}{off_kw})"
+    if 'swizzle' in mn or 'ordered_count' in mn: return f"{mn}(vdst={args[0]}, addr={args[1]}{off_kw})"
+    if 'permute' in mn: return f"{mn}(vdst={args[0]}, addr={args[1]}, data0={args[2]}{off_kw})"
+    if 'bvh' in mn: return f"{mn}(vdst={args[0]}, addr={args[1]}, data0={args[2]}, data1={args[3]}{off_kw})"
+    if 'condxchg' in mn: return f"{mn}(vdst={args[0]}, addr={args[1]}, data0={args[2]}{off_kw})"
+    if _has(mn, 'cmpstore', 'mskor', 'wrap'):
+      return f"{mn}(vdst={args[0]}, addr={args[1]}, data0={args[2]}, data1={args[3]}{off_kw})" if '_rtn' in mn else f"{mn}(addr={args[0]}, data0={args[1]}, data1={args[2]}{off_kw})"
+    return f"{mn}(vdst={args[0]}, addr={args[1]}, data0={args[2]}{off_kw})" if '_rtn' in mn else f"{mn}(addr={args[0]}, data0={args[1]}{off_kw})"
+
+  # v_fmaak/v_fmamk literal extraction
+  lit_s = ""
+  if mn in ('v_fmaak_f32', 'v_fmaak_f16') and len(args) == 4: lit_s, args = f", literal={args[3].strip()}", args[:3]
+  elif mn in ('v_fmamk_f32', 'v_fmamk_f16') and len(args) == 4: lit_s, args = f", literal={args[2].strip()}", [args[0], args[1], args[3]]
+
+  # VCC ops cleanup
+  vcc_ops = {'v_add_co_ci_u32', 'v_sub_co_ci_u32', 'v_subrev_co_ci_u32'}
+  if mn.replace('_e32', '') in vcc_ops and len(args) >= 5: mn, args = mn.replace('_e32', '') + '_e32', [args[0], args[2], args[3]]
+  if mn.replace('_e64', '') in vcc_ops and mn.endswith('_e64'): mn = mn.replace('_e64', '')
+  if mn.startswith('v_cmp') and not mn.endswith('_e64') and len(args) >= 3 and ops[0].strip().lower() in ('vcc_lo', 'vcc_hi', 'vcc'): args = args[1:]
+  if 'cmpx' in mn and mn.endswith('_e64') and len(args) == 2: args = ['RawImm(126)'] + args
+
+  fn = mn.replace('.', '_')
+  if opsel is not None: args = [re.sub(r'\.[hl]$', '', a) for a in args]
+
+  # v_fma_mix*: extract inline neg/abs modifiers
+  if 'fma_mix' in mn and neg_lo is None and neg_hi is None:
+    inline_neg, inline_abs, clean_args = 0, 0, [args[0]]
+    for i, op in enumerate(ops[1:4]):
+      op = op.strip()
+      neg = op.startswith('-') and not (op[1:2].isdigit() or (len(op) > 2 and op[1] == '0' and op[2] in 'xX'))
+      if neg: op = op[1:]
+      abs_ = op.startswith('|') and op.endswith('|')
+      if abs_: op = op[1:-1]
+      if neg: inline_neg |= (1 << i)
+      if abs_: inline_abs |= (1 << i)
+      clean_args.append(_op2dsl(op))
+    args = clean_args + args[4:]
+    if inline_neg: neg_lo = inline_neg
+    if inline_abs: neg_hi = inline_abs
+
+  all_kw = list(kw)
+  if lit_s: all_kw.append(lit_s.lstrip(', '))
+  if opsel is not None: all_kw.append(f'opsel={opsel}')
+  if neg_lo is not None: all_kw.append(f'neg={neg_lo}')
+  if neg_hi is not None: all_kw.append(f'neg_hi={neg_hi}')
+  if 'bvh' in mn and 'intersect_ray' in mn: all_kw.extend(['dmask=15', 'unrm=1', 'r128=1'])
+
+  a_str, kw_str = ', '.join(args), ', '.join(all_kw)
+  return f"{fn}({a_str}, {kw_str})" if kw_str and a_str else f"{fn}({kw_str})" if kw_str else f"{fn}({a_str})"
+
+def asm(text: str) -> Inst:
+  dsl = get_dsl(text)
+  ns = {n: getattr(ins, n) for n in dir(ins) if not n.startswith('_')}
+  ns.update({'s': s, 'v': v, 'ttmp': ttmp, 'abs': abs, 'RawImm': RawImm, 'SrcMod': SrcMod, 'VGPR': VGPR, 'SGPR': SGPR, 'TTMP': TTMP,
+             'VCC_LO': VCC_LO, 'VCC_HI': VCC_HI, 'VCC': VCC, 'EXEC_LO': EXEC_LO, 'EXEC_HI': EXEC_HI, 'EXEC': EXEC, 'SCC': SCC, 'M0': M0, 'NULL': NULL, 'OFF': OFF})
+  try: return eval(dsl, ns)
+  except NameError:
+    if m := re.match(r'^(v_\w+)(\(.*\))$', dsl): return eval(f"{m.group(1)}_e32{m.group(2)}", ns)
+    raise

extra/assembly/amd/dsl.py

@@ -0,0 +1,533 @@
+# library for RDNA3 assembly DSL
+# mypy: ignore-errors
+from __future__ import annotations
+import struct, math, re
+from enum import IntEnum
+from functools import cache, cached_property
+from typing import overload, Annotated, TypeVar, Generic
+from extra.assembly.amd.autogen.rdna3.enum import (VOP1Op, VOP2Op, VOP3Op, VOP3SDOp, VOP3POp, VOPCOp, VOPDOp, SOP1Op, SOP2Op,
+  SOPCOp, SOPKOp, SOPPOp, SMEMOp, DSOp, FLATOp, MUBUFOp, MTBUFOp, MIMGOp, VINTERPOp)
+
+# Common masks and bit conversion functions
+MASK32, MASK64 = 0xffffffff, 0xffffffffffffffff
+_struct_f, _struct_I = struct.Struct("<f"), struct.Struct("<I")
+_struct_e, _struct_H = struct.Struct("<e"), struct.Struct("<H")
+_struct_d, _struct_Q = struct.Struct("<d"), struct.Struct("<Q")
+def _f32(i): return _struct_f.unpack(_struct_I.pack(i & MASK32))[0]
+def _i32(f):
+  if isinstance(f, int): f = float(f)
+  if math.isnan(f): return 0xffc00000 if math.copysign(1.0, f) < 0 else 0x7fc00000
+  if math.isinf(f): return 0x7f800000 if f > 0 else 0xff800000
+  try: return _struct_I.unpack(_struct_f.pack(f))[0]
+  except (OverflowError, struct.error): return 0x7f800000 if f > 0 else 0xff800000
+def _sext(v, b): return v - (1 << b) if v & (1 << (b - 1)) else v
+def _f16(i): return _struct_e.unpack(_struct_H.pack(i & 0xffff))[0]
+def _i16(f):
+  if math.isnan(f): return 0x7e00
+  if math.isinf(f): return 0x7c00 if f > 0 else 0xfc00
+  try: return _struct_H.unpack(_struct_e.pack(f))[0]
+  except (OverflowError, struct.error): return 0x7c00 if f > 0 else 0xfc00
+def _f64(i): return _struct_d.unpack(_struct_Q.pack(i & MASK64))[0]
+def _i64(f):
+  if math.isnan(f): return 0x7ff8000000000000
+  if math.isinf(f): return 0x7ff0000000000000 if f > 0 else 0xfff0000000000000
+  try: return _struct_Q.unpack(_struct_d.pack(f))[0]
+  except (OverflowError, struct.error): return 0x7ff0000000000000 if f > 0 else 0xfff0000000000000
+
+# Instruction spec - register counts and dtypes derived from instruction names
+_REGS = {'B32': 1, 'B64': 2, 'B96': 3, 'B128': 4, 'B256': 8, 'B512': 16,
+         'F32': 1, 'I32': 1, 'U32': 1, 'F64': 2, 'I64': 2, 'U64': 2,
+         'F16': 1, 'I16': 1, 'U16': 1, 'B16': 1, 'I8': 1, 'U8': 1, 'B8': 1}
+_CVT_RE = re.compile(r'CVT_([FIUB]\d+)_([FIUB]\d+)$')
+_MAD_MUL_RE = re.compile(r'(?:MAD|MUL)_([IU]\d+)_([IU]\d+)$')
+_PACK_RE = re.compile(r'PACK_([FIUB]\d+)_([FIUB]\d+)$')
+_DST_SRC_RE = re.compile(r'_([FIUB]\d+)_([FIUB]\d+)$')
+_SINGLE_RE = re.compile(r'_([FIUB](?:32|64|16|8|96|128|256|512))$')
+@cache
+def _suffix(name: str) -> tuple[str | None, str | None]:
+  name = name.upper()
+  if m := _CVT_RE.search(name): return m.group(1), m.group(2)
+  if m := _MAD_MUL_RE.search(name): return m.group(1), m.group(2)
+  if m := _PACK_RE.search(name): return m.group(1), m.group(2)
+  if m := _DST_SRC_RE.search(name): return m.group(1), m.group(2)
+  if m := _SINGLE_RE.search(name): return m.group(1), m.group(1)
+  return None, None
+_SPECIAL_REGS = {
+  'V_LSHLREV_B64': (2, 1, 2, 1), 'V_LSHRREV_B64': (2, 1, 2, 1), 'V_ASHRREV_I64': (2, 1, 2, 1),
+  'S_LSHL_B64': (2, 2, 1, 1), 'S_LSHR_B64': (2, 2, 1, 1), 'S_ASHR_I64': (2, 2, 1, 1),
+  'S_BFE_U64': (2, 2, 1, 1), 'S_BFE_I64': (2, 2, 1, 1), 'S_BFM_B64': (2, 1, 1, 1),
+  'S_BITSET0_B64': (2, 1, 1, 1), 'S_BITSET1_B64': (2, 1, 1, 1),
+  'S_BITCMP0_B64': (1, 2, 1, 1), 'S_BITCMP1_B64': (1, 2, 1, 1),
+  'V_LDEXP_F64': (2, 2, 1, 1), 'V_TRIG_PREOP_F64': (2, 2, 1, 1),
+  'V_CMP_CLASS_F64': (1, 2, 1, 1), 'V_CMPX_CLASS_F64': (1, 2, 1, 1),
+  'V_CMP_CLASS_F32': (1, 1, 1, 1), 'V_CMPX_CLASS_F32': (1, 1, 1, 1),
+  'V_CMP_CLASS_F16': (1, 1, 1, 1), 'V_CMPX_CLASS_F16': (1, 1, 1, 1),
+  'V_MAD_U64_U32': (2, 1, 1, 2), 'V_MAD_I64_I32': (2, 1, 1, 2),
+  'V_QSAD_PK_U16_U8': (2, 2, 1, 2), 'V_MQSAD_PK_U16_U8': (2, 2, 1, 2), 'V_MQSAD_U32_U8': (4, 2, 1, 4),
+}
+_SPECIAL_DTYPE = {
+  'V_LSHLREV_B64': ('B64', 'U32', 'B64', None), 'V_LSHRREV_B64': ('B64', 'U32', 'B64', None), 'V_ASHRREV_I64': ('I64', 'U32', 'I64', None),
+  'S_LSHL_B64': ('B64', 'B64', 'U32', None), 'S_LSHR_B64': ('B64', 'B64', 'U32', None), 'S_ASHR_I64': ('I64', 'I64', 'U32', None),
+  'S_BFE_U64': ('U64', 'U64', 'U32', None), 'S_BFE_I64': ('I64', 'I64', 'U32', None),
+  'S_BFM_B64': ('B64', 'U32', 'U32', None), 'S_BITSET0_B64': ('B64', 'U32', None, None), 'S_BITSET1_B64': ('B64', 'U32', None, None),
+  'S_BITCMP0_B64': ('SCC', 'B64', 'U32', None), 'S_BITCMP1_B64': ('SCC', 'B64', 'U32', None),
+  'V_LDEXP_F64': ('F64', 'F64', 'I32', None), 'V_TRIG_PREOP_F64': ('F64', 'F64', 'U32', None),
+  'V_CMP_CLASS_F64': ('VCC', 'F64', 'U32', None), 'V_CMPX_CLASS_F64': ('EXEC', 'F64', 'U32', None),
+  'V_CMP_CLASS_F32': ('VCC', 'F32', 'U32', None), 'V_CMPX_CLASS_F32': ('EXEC', 'F32', 'U32', None),
+  'V_CMP_CLASS_F16': ('VCC', 'F16', 'U32', None), 'V_CMPX_CLASS_F16': ('EXEC', 'F16', 'U32', None),
+  'V_MAD_U64_U32': ('U64', 'U32', 'U32', 'U64'), 'V_MAD_I64_I32': ('I64', 'I32', 'I32', 'I64'),
+  'V_QSAD_PK_U16_U8': ('B64', 'B64', 'B64', 'B64'), 'V_MQSAD_PK_U16_U8': ('B64', 'B64', 'B64', 'B64'),
+  'V_MQSAD_U32_U8': ('B128', 'B64', 'B64', 'B128'),
+}
+@cache
+def spec_regs(name: str) -> tuple[int, int, int, int]:
+  uname = name.upper()
+  if uname in _SPECIAL_REGS: return _SPECIAL_REGS[uname]
+  if 'SAD' in uname and 'U8' in uname and 'QSAD' not in uname and 'MQSAD' not in uname: return 1, 1, 1, 1
+  dst_suf, src_suf = _suffix(name)
+  return _REGS.get(dst_suf, 1), _REGS.get(src_suf, 1), _REGS.get(src_suf, 1), _REGS.get(src_suf, 1)
+@cache
+def spec_dtype(name: str) -> tuple[str | None, str | None, str | None, str | None]:
+  uname = name.upper()
+  if uname in _SPECIAL_DTYPE: return _SPECIAL_DTYPE[uname]
+  if 'SAD' in uname and ('U8' in uname or 'U16' in uname) and 'QSAD' not in uname and 'MQSAD' not in uname: return 'U32', 'U32', 'U32', 'U32'
+  if '_CMP_' in uname or '_CMPX_' in uname:
+    dst_suf, src_suf = _suffix(name)
+    return 'EXEC' if '_CMPX_' in uname else 'VCC', src_suf, src_suf, None
+  dst_suf, src_suf = _suffix(name)
+  return dst_suf, src_suf, src_suf, src_suf
+_F16_RE = re.compile(r'_[FIUB]16(?:_|$)')
+_F64_RE = re.compile(r'_[FIUB]64(?:_|$)')
+@cache
+def spec_is_16bit(name: str) -> bool:
+  uname = name.upper()
+  if 'SAD' in uname or 'PACK' in uname or '_PK_' in uname or 'SAT_PK' in uname or 'DOT2' in uname: return False
+  if '_F32' in uname or '_I32' in uname or '_U32' in uname or '_B32' in uname: return False
+  return bool(_F16_RE.search(uname))
+@cache
+def spec_is_64bit(name: str) -> bool: return bool(_F64_RE.search(name.upper()))
+_3SRC = {'FMA', 'MAD', 'MIN3', 'MAX3', 'MED3', 'DIV_FIX', 'DIV_FMAS', 'DIV_SCALE', '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', 'DOT4', 'DOT8', 'WMMA', 'CVT_PK_U8', 'MULLIT', 'CO_CI'}
+_2SRC = {'FMAC'}  # FMAC uses dst as implicit accumulator, so only 2 explicit sources
+def spec_num_srcs(name: str) -> int:
+  name = name.upper()
+  if any(k in name for k in _2SRC): return 2
+  return 3 if any(k in name for k in _3SRC) else 2
+def is_dtype_16(dt: str | None) -> bool: return dt is not None and '16' in dt
+def is_dtype_64(dt: str | None) -> bool: return dt is not None and '64' in dt
+
+# Bit field DSL
+class BitField:
+  def __init__(self, hi: int, lo: int, name: str | None = None): self.hi, self.lo, self.name, self._marker = hi, lo, name, None
+  def __set_name__(self, owner, name):
+    import typing
+    self.name, self._owner = name, owner
+    # Cache marker at class definition time
+    hints = typing.get_type_hints(owner, include_extras=True)
+    if name in hints:
+      hint = hints[name]
+      if typing.get_origin(hint) is Annotated:
+        args = typing.get_args(hint)
+        self._marker = args[1] if len(args) > 1 else None
+  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: return self._marker
+  @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):
+      # VOP3 with VOPC opcodes (0-255) -> VOPCOp, VOP3SD opcodes -> VOP3SDOp
+      if self.marker is VOP3Op:
+        if val < 256: return VOPCOp(val)
+        if val in Inst._VOP3SD_OPS: return VOP3SDOp(val)
+      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()
+
+# Source operand with modifiers - base class for anything that can be a src with neg/abs
+class SrcMod:
+  __slots__ = ('val', 'neg', 'abs_')
+  def __init__(self, val: int, neg: bool = False, abs_: bool = False): self.val, self.neg, self.abs_ = val, neg, abs_
+  def __repr__(self): return f"{'-' if self.neg else ''}{'|' if self.abs_ else ''}{self.val}{'|' if self.abs_ else ''}"
+  def __neg__(self): return SrcMod(self.val, not self.neg, self.abs_)
+  def __abs__(self): return SrcMod(self.val, self.neg, True)
+
+# Register types
+class Reg(SrcMod):
+  __slots__ = ('idx', 'count', 'hi')
+  def __init__(self, idx: int, count: int = 1, hi: bool = False, neg: bool = False, abs_: bool = False):
+    self.idx, self.count, self.hi = idx, count, hi
+    super().__init__(idx, neg, abs_)
+  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, not self.neg, self.abs_)
+  def __abs__(self): return self.__class__(self.idx, self.count, self.hi, self.neg, True)
+  @property
+  def l(self): return self.__class__(self.idx, self.count, False, self.neg, self.abs_)
+  @property
+  def h(self): return self.__class__(self.idx, self.count, True, self.neg, self.abs_)
+
+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")
+
+# Special registers as SrcMod objects (support -VCC_LO, abs(EXEC_LO), etc.)
+VCC_LO, VCC_HI, VCC = SrcMod(106), SrcMod(107), SrcMod(106)
+EXEC_LO, EXEC_HI, EXEC = SrcMod(126), SrcMod(127), SrcMod(126)
+SCC, M0, NULL, OFF = SrcMod(253), SrcMod(125), SrcMod(124), SrcMod(124)
+
+# 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:
+  if isinstance(val, RawImm): return val.val
+  if isinstance(val, SrcMod) and not isinstance(val, Reg): return val.val  # Special registers like VCC_LO, NULL
+  if hasattr(val, 'value'): return val.value  # IntEnum
+  if hasattr(val, 'idx'): return val.idx  # Reg
+  return val
+
+# Encoding/decoding constants
+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}
+FLOAT_DEC = {v: str(k) for k, v in FLOAT_ENC.items()}
+SPECIAL_GPRS = {106: "vcc_lo", 107: "vcc_hi", 124: "null", 125: "m0", 126: "exec_lo", 127: "exec_hi", 253: "scc"}
+SPECIAL_PAIRS = {106: "vcc", 126: "exec"}
+SRC_FIELDS = {'src0', 'src1', 'src2', 'ssrc0', 'ssrc1', 'soffset', 'srcx0', 'srcy0'}
+RAW_FIELDS = {'vdata', 'vdst', 'vaddr', 'addr', 'data', 'data0', 'data1', 'sdst', 'sdata', 'vsrc1'}
+
+def _encode_reg(val: Reg) -> int: return (108 if isinstance(val, TTMP) else 0) + val.idx
+
+def _is_inline_const(v: int) -> bool: return 0 <= v <= 127 or 128 <= v <= 208 or 240 <= v <= 255
+
+def encode_src(val) -> int:
+  if isinstance(val, VGPR): return 256 + _encode_reg(val)
+  if isinstance(val, Reg): return _encode_reg(val)
+  if isinstance(val, SrcMod) and not isinstance(val, Reg): return val.val if _is_inline_const(val.val) else 255
+  if hasattr(val, 'value'): return val.value  # IntEnum
+  if isinstance(val, float): return 128 if val == 0.0 else FLOAT_ENC.get(val, 255)
+  if isinstance(val, int): return 128 + val if 0 <= val <= 64 else 192 - val if -16 <= val <= -1 else 255
+  return 255
+
+def decode_src(val: int) -> str:
+  if val <= 105: return f"s{val}"
+  if val in SPECIAL_GPRS: return SPECIAL_GPRS[val]
+  if val in FLOAT_DEC: return FLOAT_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}"
+
+# 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 _or_field(self, name: str, bit: int):
+    cur = self._values.get(name, 0)
+    self._values[name] = (cur.val if isinstance(cur, RawImm) else cur) | bit
+
+  def _encode_src(self, name: str, val):
+    """Encode a source field, handling modifiers and literals."""
+    encoded = encode_src(val)
+    has_opsel = 'opsel' in self._fields
+    if isinstance(val, Reg) and val.hi and not has_opsel: encoded |= 0x80  # hi bit in src for VOP1/2/C
+    self._values[name] = RawImm(encoded)
+    # Handle neg/abs/opsel modifiers
+    if isinstance(val, SrcMod):
+      mod_bit = {'src0': 1, 'src1': 2, 'src2': 4}.get(name, 0)
+      if val.neg and 'neg' in self._fields: self._or_field('neg', mod_bit)
+      if val.abs_ and 'abs' in self._fields: self._or_field('abs', mod_bit)
+    if isinstance(val, Reg) and val.hi and has_opsel:
+      self._or_field('opsel', {'src0': 1, 'src1': 2, 'src2': 4}.get(name, 0))
+    # Track literal value if needed
+    if encoded == 255 and self._literal is None:
+      import struct
+      # Check if THIS source uses 64-bit encoding (not just src0)
+      src_idx = {'src0': 0, 'src1': 1, 'src2': 2, 'ssrc0': 0, 'ssrc1': 1}.get(name, 0)
+      src_regs = self.src_regs(src_idx)
+      is_64 = src_regs == 2
+      if isinstance(val, SrcMod) and not isinstance(val, Reg): lit32 = val.val & MASK32
+      elif isinstance(val, int) and not isinstance(val, IntEnum): lit32 = val & MASK32
+      elif isinstance(val, float): lit32 = (_i64(val) >> 32) if is_64 else _i32(val)  # f64: high 32 bits of f64 repr
+      else: return
+      self._literal = (lit32 << 32) if is_64 else lit32
+
+  def _encode_raw(self, name: str, val):
+    """Encode a raw register field (vdst, vdata, etc.)."""
+    if isinstance(val, Reg):
+      encoded = _encode_reg(val)
+      if val.hi and 'opsel' not in self._fields: encoded |= 0x80
+      self._values[name] = encoded
+      if name == 'vdst' and val.hi and 'opsel' in self._fields: self._or_field('opsel', 8)
+    elif hasattr(val, 'value'): self._values[name] = val.value
+
+  def _validate(self, orig_args: dict):
+    """Format-specific validation. Override in subclass or check by class name."""
+    cls_name, op = self.__class__.__name__, orig_args.get('op')
+    if hasattr(op, 'value'): op = op.value
+    # SMEM: register count must match opcode
+    if cls_name == 'SMEM' and op is not None:
+      expected = {0:1, 1:2, 2:4, 3:8, 4:16, 8:1, 9:2, 10:4, 11:8, 12:16}.get(op)
+      sdata = orig_args.get('sdata')
+      if expected and isinstance(sdata, Reg) and sdata.count != expected:
+        raise ValueError(f"SMEM op {op} expects {expected} registers, got {sdata.count}")
+    # SOP1: b32=1 reg, b64=2 regs
+    if cls_name == 'SOP1' and hasattr(orig_args.get('op'), 'name'):
+      expected = 2 if orig_args['op'].name.endswith('_B64') else 1
+      for fld in ('sdst', 'ssrc0'):
+        if isinstance(orig_args.get(fld), Reg) and orig_args[fld].count != expected:
+          raise ValueError(f"SOP1 {orig_args['op'].name} expects {expected} register(s) for {fld}, got {orig_args[fld].count}")
+
+  def __init__(self, *args, literal: int | None = None, **kwargs):
+    self._values, self._literal = dict(self._defaults), None
+    field_names = [n for n in self._fields if n != 'encoding']
+    orig_args = dict(zip(field_names, args)) | kwargs
+    self._values.update(orig_args)
+    self._validate(orig_args)
+    # Pre-shift literal for 64-bit sources (literal param is always raw 32-bit value from user)
+    if literal is not None:
+      # Find which source uses the literal (255) and check its register count
+      for n, idx in [('src0', 0), ('src1', 1), ('src2', 2), ('ssrc0', 0), ('ssrc1', 1)]:
+        v = orig_args.get(n)
+        if (isinstance(v, RawImm) and v.val == 255) or (isinstance(v, int) and v == 255):
+          self._literal = (literal << 32) if self.src_regs(idx) == 2 else literal
+          break
+      else:
+        self._literal = literal  # fallback if no literal source found
+    cls_name = self.__class__.__name__
+
+    # Format-specific setup
+    if cls_name == 'FLAT' and 'sve' in self._fields:
+      seg = self._values.get('seg', 0)
+      if (seg.val if isinstance(seg, RawImm) else seg) == 1 and isinstance(orig_args.get('addr'), VGPR): self._values['sve'] = 1
+    if cls_name == 'VOP3P':
+      op = orig_args.get('op')
+      if hasattr(op, 'value'): op = op.value
+      if op in (32, 33, 34) and 'opsel_hi' not in orig_args: self._values['opsel_hi'] = self._values['opsel_hi2'] = 0
+
+    # Encode all fields
+    for name, val in list(self._values.items()):
+      if name == 'encoding': continue
+      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 and isinstance(val, VGPR): raise TypeError(f"field '{name}' requires SGPR, got VGPR")
+      if marker is _VGPRField and 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 by field type
+      if name in SRC_FIELDS: self._encode_src(name, val)
+      elif name in RAW_FIELDS: self._encode_raw(name, val)
+      elif name == 'sbase': self._values[name] = (val.idx if isinstance(val, Reg) else val.val if isinstance(val, SrcMod) else val * 2) // 2
+      elif name in {'srsrc', 'ssamp'} and isinstance(val, Reg): self._values[name] = val.idx // 4
+      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 isinstance(val, SrcMod) and not isinstance(val, Reg): return val.val  # Special regs like VCC_LO
+    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.val // 2 if isinstance(val, SrcMod) 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)."""
+    op = self._values.get('op')
+    if op is None: return False
+    op_name = op.name if hasattr(op, 'name') else None
+    # Look up op name from int if needed (happens in from_bytes path)
+    if op_name is None and self.__class__.__name__ == 'VOP3':
+      try: op_name = VOP3Op(op).name
+      except ValueError: pass
+    if op_name is None and self.__class__.__name__ == 'VOPC':
+      try: op_name = VOPCOp(op).name
+      except ValueError: pass
+    if op_name is None: return False
+    # V_LDEXP_F64 has 32-bit integer src1, so literal is 32-bit
+    return op_name != 'V_LDEXP_F64' and 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 sources, literal is stored in high 32 bits internally, but encoded as 4 bytes
+    # Find which source uses the literal (255) and check its register count
+    lit_src_is_64 = False
+    for n, idx in [('src0', 0), ('src1', 1), ('src2', 2), ('ssrc0', 0), ('ssrc1', 1)]:
+      if n not in self._values: continue
+      v = self._values[n]
+      if (isinstance(v, RawImm) and v.val == 255) or (isinstance(v, int) and v == 255):
+        lit_src_is_64 = self.is_src_64(idx)
+        break
+    lit32 = (lit >> 32) if lit_src_is_64 else lit
+    return result + (lit32 & MASK32).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))
+    # VOPD fmaak/fmamk always have a literal (opx/opy value 1 or 2)
+    opx, opy = inst._values.get('opx', 0), inst._values.get('opy', 0)
+    has_literal = has_literal or (cls.__name__ == 'VOPD' and (opx in (1, 2) or opy in (1, 2)))
+    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.
+      # Check which source uses the literal and whether THAT source is 64-bit
+      if len(data) >= cls._size() + 4:
+        lit32 = int.from_bytes(data[cls._size():cls._size()+4], 'little')
+        # Find which source has literal (255) and check its register count
+        lit_src_is_64 = False
+        for n, idx in [('src0', 0), ('src1', 1), ('src2', 2)]:
+          if n in inst._values and isinstance(inst._values[n], RawImm) and inst._values[n].val == 255:
+            lit_src_is_64 = inst.src_regs(idx) == 2
+            break
+        inst._literal = (lit32 << 32) if lit_src_is_64 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 __getattr__(self, name: str):
+    if name.startswith('_'): raise AttributeError(name)
+    return unwrap(self._values.get(name, 0))
+
+  def lit(self, v: int, neg: bool = False) -> str:
+    s = f"0x{self._literal:x}" if v == 255 and self._literal else decode_src(v)
+    return f"-{s}" if neg else s
+
+  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)
+
+  _enum_map = {'VOP1': VOP1Op, 'VOP2': VOP2Op, 'VOP3': VOP3Op, 'VOP3SD': VOP3SDOp, 'VOP3P': VOP3POp, 'VOPC': VOPCOp,
+               'SOP1': SOP1Op, 'SOP2': SOP2Op, 'SOPC': SOPCOp, 'SOPK': SOPKOp, 'SOPP': SOPPOp,
+               'SMEM': SMEMOp, 'DS': DSOp, 'FLAT': FLATOp, 'MUBUF': MUBUFOp, 'MTBUF': MTBUFOp, 'MIMG': MIMGOp,
+               'VOPD': VOPDOp, 'VINTERP': VINTERPOp}
+  _VOP3SD_OPS = {288, 289, 290, 764, 765, 766, 767, 768, 769, 770}
+
+  @property
+  def op(self):
+    """Return the op as an enum (e.g., VOP1Op.V_MOV_B32). VOP3 returns VOPCOp/VOP3SDOp for those op ranges."""
+    val = self._values.get('op')
+    if val is None: return None
+    if hasattr(val, 'name'): return val  # already an enum
+    cls_name = self.__class__.__name__
+    assert cls_name in self._enum_map, f"no enum map for {cls_name}"
+    return self._enum_map[cls_name](val)
+
+  @cached_property
+  def op_name(self) -> str:
+    op = self.op
+    return op.name if hasattr(op, 'name') else ''
+
+  @cached_property
+  def _spec_regs(self) -> tuple[int, int, int, int]: return spec_regs(self.op_name)
+  @cached_property
+  def _spec_dtype(self) -> tuple[str | None, str | None, str | None, str | None]: return spec_dtype(self.op_name)
+  def dst_regs(self) -> int: return self._spec_regs[0]
+  def src_regs(self, n: int) -> int: return self._spec_regs[n + 1]
+  def num_srcs(self) -> int: return spec_num_srcs(self.op_name)
+  def dst_dtype(self) -> str | None: return self._spec_dtype[0]
+  def src_dtype(self, n: int) -> str | None: return self._spec_dtype[n + 1]
+  def is_src_16(self, n: int) -> bool: return self._spec_regs[n + 1] == 1 and is_dtype_16(self._spec_dtype[n + 1])
+  def is_src_64(self, n: int) -> bool: return self._spec_regs[n + 1] == 2
+  def is_16bit(self) -> bool: return spec_is_16bit(self.op_name)
+  def is_64bit(self) -> bool: return spec_is_64bit(self.op_name)
+  def is_dst_16(self) -> bool: return self._spec_regs[0] == 1 and is_dtype_16(self._spec_dtype[0])
+
+class Inst32(Inst): pass
+class Inst64(Inst): pass

extra/assembly/amd/emu.py

@@ -0,0 +1,462 @@
+# RDNA3 emulator - executes compiled pseudocode from AMD ISA PDF
+# mypy: ignore-errors
+from __future__ import annotations
+import ctypes
+from extra.assembly.amd.dsl import Inst, unwrap, FLOAT_ENC, MASK32, MASK64, _f32, _i32, _sext, _f16, _i16, _f64, _i64
+from extra.assembly.amd.pcode import Reg
+from extra.assembly.amd.asm import detect_format
+from extra.assembly.amd.autogen.rdna3.gen_pcode import get_compiled_functions
+from extra.assembly.amd.autogen.rdna3.ins import (SOP1, SOP2, SOPC, SOPK, SOPP, SMEM, VOP1, VOP2, VOP3, VOP3SD, VOP3P, VOPC, DS, FLAT, VOPD,
+  SrcEnum, SOPPOp, SMEMOp, VOP1Op, VOP2Op, VOP3Op, VOP3SDOp, VOP3POp, VOPCOp, GLOBALOp, FLATOp, DSOp, 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
+
+# Inline constants for src operands 128-254. Build tables for f32, f16, and f64 formats.
+_FLOAT_CONSTS = {v: k for k, v in FLOAT_ENC.items()} | {248: 0.15915494309189535}  # INV_2PI
+def _build_inline_consts(mask, to_bits):
+  tbl = list(range(65)) + [((-i) & mask) for i in range(1, 17)] + [0] * (127 - 81)
+  for k, v in _FLOAT_CONSTS.items(): tbl[k - 128] = to_bits(v)
+  return tbl
+_INLINE_CONSTS = _build_inline_consts(MASK32, _i32)
+_INLINE_CONSTS_F16 = _build_inline_consts(0xffff, _i16)
+_INLINE_CONSTS_F64 = _build_inline_consts(MASK64, _i64)
+
+# Helper: extract/write 16-bit half from/to 32-bit value
+def _src16(raw: int, is_hi: bool) -> int: return ((raw >> 16) & 0xffff) if is_hi else (raw & 0xffff)
+def _dst16(cur: int, val: int, is_hi: bool) -> int: return (cur & 0x0000ffff) | ((val & 0xffff) << 16) if is_hi else (cur & 0xffff0000) | (val & 0xffff)
+def _vgpr_hi(src: int) -> bool: return src >= 256 and ((src - 256) & 0x80) != 0
+def _vgpr_masked(src: int) -> int: return ((src - 256) & 0x7f) + 256 if src >= 256 else src
+
+# 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:
+  return not _valid_mem_ranges or any(s <= addr and addr + size <= s + z for s, z in _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)
+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 & MASK32, (v >> 32) & MASK32
+  @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 & MASK32, (v >> 32) & MASK32
+
+  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 & MASK32
+  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 & MASK32); self.wsgpr(i+1, (v >> 32) & MASK32)
+
+  def _rsrc_base(self, v: int, lane: int, consts):
+    if v < SGPR_COUNT: return self.sgpr[v]
+    if v == SCC: return self.scc
+    if v < 255: return consts[v - 128]
+    if v == 255: return self.literal
+    return self.vgpr[lane][v - 256] if v <= 511 else 0
+  def rsrc(self, v: int, lane: int) -> int: return self._rsrc_base(v, lane, _INLINE_CONSTS)
+  def rsrc_f16(self, v: int, lane: int) -> int: return self._rsrc_base(v, lane, _INLINE_CONSTS_F16)
+  def rsrc64(self, v: int, lane: int) -> int:
+    if 128 <= v < 255: return _INLINE_CONSTS_F64[v - 128]
+    if v == 255: return self.literal  # literal is already shifted in from_bytes for 64-bit ops
+    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()
+
+
+def decode_program(data: bytes) -> Program:
+  result: Program = {}
+  i = 0
+  while i < len(data):
+    try: inst_class = detect_format(data[i:])
+    except ValueError: break  # stop at invalid instruction (padding/metadata after code)
+    if inst_class is None: i += 4; continue
+    base_size = inst_class._size()
+    # 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():
+      if name != 'op': setattr(inst, name, unwrap(val))  # skip op to preserve property access
+    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()
+
+  # SOPP: special cases for control flow that has no pseudocode
+  if isinstance(inst, SOPP):
+    if inst.op == SOPPOp.S_ENDPGM: return -1
+    if inst.op == SOPPOp.S_BARRIER: return -2
+
+  # SMEM: memory loads (not ALU)
+  if isinstance(inst, 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) & MASK64, 4))
+    return 0
+
+  # Get op enum and lookup compiled function
+  if isinstance(inst, SOP1): ssrc0, sdst = inst.ssrc0, inst.sdst
+  elif isinstance(inst, SOP2): ssrc0, sdst = inst.ssrc0, inst.sdst
+  elif isinstance(inst, SOPC): ssrc0, sdst = inst.ssrc0, None
+  elif isinstance(inst, SOPK): ssrc0, sdst = inst.sdst, inst.sdst  # sdst is both src and dst
+  elif isinstance(inst, SOPP): ssrc0, sdst = None, None
+  else: raise NotImplementedError(f"Unknown scalar type {type(inst)}")
+
+  # SOPP has gaps in the opcode enum - treat unknown opcodes as no-ops
+  try: op = inst.op
+  except ValueError:
+    if isinstance(inst, SOPP): return 0
+    raise
+  fn = compiled.get(type(op), {}).get(op)
+  if fn is None:
+    # SOPP instructions without pseudocode (waits, hints, nops) are no-ops
+    if isinstance(inst, SOPP): return 0
+    raise NotImplementedError(f"{op.name} not in pseudocode")
+
+  # Build context - use inst methods to determine operand sizes
+  s0 = st.rsrc64(ssrc0, 0) if inst.is_src_64(0) else (st.rsrc(ssrc0, 0) if not isinstance(inst, (SOPK, SOPP)) else (st.rsgpr(inst.sdst) if isinstance(inst, SOPK) else 0))
+  s1 = st.rsrc64(inst.ssrc1, 0) if inst.is_src_64(1) else (st.rsrc(inst.ssrc1, 0) if isinstance(inst, (SOP2, SOPC)) else inst.simm16 if isinstance(inst, SOPK) else 0)
+  d0 = st.rsgpr64(sdst) if inst.dst_regs() == 2 and sdst is not None else (st.rsgpr(sdst) if sdst is not None else 0)
+  literal = inst.simm16 if isinstance(inst, (SOPK, SOPP)) else st.literal
+
+  # Create Reg objects for compiled function - mask VCC/EXEC to 32 bits for wave32
+  result = fn(Reg(s0), Reg(s1), None, Reg(d0), Reg(st.scc), Reg(st.vcc & MASK32), 0, Reg(st.exec_mask & MASK32), literal, None, PC=Reg(st.pc * 4))
+
+  # Apply results - extract values from returned Reg objects
+  if sdst is not None and 'D0' in result:
+    (st.wsgpr64 if inst.dst_regs() == 2 else st.wsgpr)(sdst, result['D0']._val)
+  if 'SCC' in result: st.scc = result['SCC']._val & 1
+  if 'EXEC' in result: st.exec_mask = result['EXEC']._val
+  if 'PC' in result:
+    # Convert absolute byte address to word delta
+    pc_val = result['PC']._val
+    new_pc = pc_val if pc_val < 0x8000000000000000 else pc_val - 0x10000000000000000
+    new_pc_words = new_pc // 4
+    return new_pc_words - st.pc - 1  # -1 because emulator adds inst_words (1 for scalar)
+  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()
+  V = st.vgpr[lane]
+
+  # Memory ops (not ALU pseudocode)
+  if isinstance(inst, 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) & MASK64 if saddr not in (NULL, 0x7f) else (addr + offset) & MASK64
+    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) & MASK32 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
+      V[vdst] = _dst16(V[vdst], val, hi)
+    elif op in FLAT_D16_STORE:
+      sz, hi = FLAT_D16_STORE[op]
+      mem_write(addr, sz, _src16(V[data_reg], hi) & ((1 << (sz * 8)) - 1))
+    else: raise NotImplementedError(f"FLAT op {op}")
+    return
+
+  if isinstance(inst, DS):
+    fn = compiled.get(DSOp, {}).get(inst.op)
+    if fn is None: raise NotImplementedError(f"DS op {inst.op.name} not in pseudocode")
+    # Prepare data registers as lists of dwords
+    data0 = [V[inst.data0 + i] for i in range(4)]  # up to 4 dwords
+    data1 = [V[inst.data1 + i] for i in range(4)] if inst.data1 else [0, 0, 0, 0]
+    result = fn(lds, V[inst.addr], data0, data1, inst.vdst, inst.offset0, inst.offset1)
+    # Write results for loads
+    if 'vdst' in result:
+      for i, val in enumerate(result['vdst']): V[inst.vdst + i] = val & MASK32
+    return
+
+  # VOPD: dual-issue, execute two ops simultaneously (read all inputs before writes)
+  if isinstance(inst, VOPD):
+    vdsty = (inst.vdsty << 1) | ((inst.vdstx & 1) ^ 1)
+    inputs = [(inst.opx, st.rsrc(inst.srcx0, lane), V[inst.vsrcx1], V[inst.vdstx], inst.vdstx),
+              (inst.opy, st.rsrc(inst.srcy0, lane), V[inst.vsrcy1], V[vdsty], vdsty)]
+    def exec_vopd(vopd_op, s0, s1, d0):
+      op = _VOPD_TO_VOP[vopd_op]
+      return compiled[type(op)][op](Reg(s0), Reg(s1), None, Reg(d0), Reg(st.scc), Reg(st.vcc), lane, Reg(st.exec_mask), st.literal, None)['D0']._val
+    for vopd_op, s0, s1, d0, dst in inputs: V[dst] = exec_vopd(vopd_op, s0, s1, d0)
+    return
+
+  # VOP3SD: has extra scalar dest for carry output
+  if isinstance(inst, VOP3SD):
+    fn = compiled[VOP3SDOp][inst.op]
+    # Read sources based on register counts from inst properties
+    def rsrc_n(src, regs): return st.rsrc64(src, lane) if regs == 2 else st.rsrc(src, lane)
+    s0, s1, s2 = rsrc_n(inst.src0, inst.src_regs(0)), rsrc_n(inst.src1, inst.src_regs(1)), rsrc_n(inst.src2, inst.src_regs(2))
+    # Carry-in ops use src2 as carry bitmask instead of VCC
+    vcc = st.rsgpr64(inst.src2) if 'CO_CI' in inst.op_name else st.vcc
+    result = fn(Reg(s0), Reg(s1), Reg(s2), Reg(V[inst.vdst]), Reg(st.scc), Reg(vcc), lane, Reg(st.exec_mask), st.literal, None)
+    d0_val = result['D0']._val
+    V[inst.vdst] = d0_val & MASK32
+    if inst.dst_regs() == 2: V[inst.vdst + 1] = (d0_val >> 32) & MASK32
+    if 'VCC' in result: st.pend_sgpr_lane(inst.sdst, lane, (result['VCC']._val >> lane) & 1)
+    return
+
+  # Get op enum and sources (None means "no source" for that operand)
+  # dst_hi: for VOP1/VOP2 16-bit dst ops, bit 7 of vdst indicates .h (high 16-bit) destination
+  dst_hi = False
+  if isinstance(inst, VOP1):
+    if inst.op == VOP1Op.V_NOP: return
+    src0, src1, src2 = inst.src0, None, None
+    dst_hi = (inst.vdst & 0x80) != 0 and inst.is_dst_16()
+    vdst = inst.vdst & 0x7f if inst.is_dst_16() else inst.vdst
+  elif isinstance(inst, VOP2):
+    src0, src1, src2 = inst.src0, inst.vsrc1 + 256, None
+    dst_hi = (inst.vdst & 0x80) != 0 and inst.is_dst_16()
+    vdst = inst.vdst & 0x7f if inst.is_dst_16() else inst.vdst
+  elif isinstance(inst, VOP3):
+    # VOP3 ops 0-255 are VOPC comparisons encoded as VOP3 - inst.op returns VOPCOp for these
+    src0, src1, src2, vdst = inst.src0, inst.src1, (None if inst.op.value < 256 else inst.src2), inst.vdst
+  elif isinstance(inst, VOPC):
+    # For 16-bit VOPC, vsrc1 uses same encoding as VOP2 16-bit: bit 7 selects hi(1) or lo(0) half
+    # vsrc1 field is 8 bits: [6:0] = VGPR index, [7] = hi flag
+    src0, src1, src2, vdst = inst.src0, inst.vsrc1 + 256, None, VCC_LO
+  elif isinstance(inst, VOP3P):
+    # VOP3P: Packed 16-bit operations using compiled functions
+    # WMMA: wave-level matrix multiply-accumulate (special handling - needs cross-lane access)
+    if 'WMMA' in inst.op_name:
+      if lane == 0:  # Only execute once per wave, write results for all lanes
+        exec_wmma(st, inst, inst.op)
+      return
+    # V_FMA_MIX: Mixed precision FMA - opsel_hi controls f32(0) vs f16(1), opsel selects which f16 half
+    # Handle inline because abs/neg must be applied AFTER type conversion
+    if inst.op in (VOP3POp.V_FMA_MIX_F32, VOP3POp.V_FMA_MIXLO_F16, VOP3POp.V_FMA_MIXHI_F16):
+      opsel, opsel_hi, opsel_hi2 = getattr(inst, 'opsel', 0), getattr(inst, 'opsel_hi', 0), getattr(inst, 'opsel_hi2', 0)
+      neg, abs_ = getattr(inst, 'neg', 0), getattr(inst, 'neg_hi', 0)  # neg_hi reused as abs for FMA_MIX
+      raws = [st.rsrc(inst.src0, lane), st.rsrc(inst.src1, lane), st.rsrc(inst.src2, lane) if inst.src2 is not None else 0]
+      is_f16 = [opsel_hi & 1, opsel_hi & 2, opsel_hi2]
+      srcs = [_f16(_src16(raws[i], bool(opsel & (1<<i)))) if is_f16[i] else _f32(raws[i]) for i in range(3)]
+      for i in range(3):
+        if abs_ & (1<<i): srcs[i] = abs(srcs[i])
+        if neg & (1<<i): srcs[i] = -srcs[i]
+      result_f = srcs[0] * srcs[1] + srcs[2]
+      V = st.vgpr[lane]
+      V[inst.vdst] = _i32(result_f) if inst.op == VOP3POp.V_FMA_MIX_F32 else _dst16(V[inst.vdst], _i16(result_f), inst.op == VOP3POp.V_FMA_MIXHI_F16)
+      return
+    # VOP3P packed ops: opsel selects halves for lo, opsel_hi for hi; neg toggles f16 sign
+    raws = [st.rsrc_f16(inst.src0, lane), st.rsrc_f16(inst.src1, lane), st.rsrc_f16(inst.src2, lane) if inst.src2 is not None else 0]
+    opsel, opsel_hi, opsel_hi2 = getattr(inst, 'opsel', 0), getattr(inst, 'opsel_hi', 3), getattr(inst, 'opsel_hi2', 1)
+    neg, neg_hi = getattr(inst, 'neg', 0), getattr(inst, 'neg_hi', 0)
+    hi_sels = [opsel_hi & 1, opsel_hi & 2, opsel_hi2]
+    srcs = [((_src16(raws[i], hi_sels[i]) ^ (0x8000 if neg_hi & (1<<i) else 0)) << 16) |
+            (_src16(raws[i], opsel & (1<<i)) ^ (0x8000 if neg & (1<<i) else 0)) for i in range(3)]
+    result = compiled[VOP3POp][inst.op](Reg(srcs[0]), Reg(srcs[1]), Reg(srcs[2]), Reg(0), Reg(st.scc), Reg(st.vcc), lane, Reg(st.exec_mask), st.literal, None)
+    st.vgpr[lane][inst.vdst] = result['D0']._val & MASK32
+    return
+  else: raise NotImplementedError(f"Unknown vector type {type(inst)}")
+
+  op_cls = type(inst.op)
+  if (fn := compiled.get(op_cls, {}).get(inst.op)) is None: raise NotImplementedError(f"{inst.op_name} not in pseudocode")
+
+  # Read sources (with VOP3 modifiers if applicable)
+  neg, abs_ = (getattr(inst, 'neg', 0), getattr(inst, 'abs', 0)) if isinstance(inst, VOP3) else (0, 0)
+  opsel = getattr(inst, 'opsel', 0) if isinstance(inst, VOP3) else 0
+  def mod_src(val: int, idx: int, is64=False) -> int:
+    to_f, to_i = (_f64, _i64) if is64 else (_f32, _i32)
+    if (abs_ >> idx) & 1: val = to_i(abs(to_f(val)))
+    if (neg >> idx) & 1: val = to_i(-to_f(val))
+    return val
+
+  # Use inst methods to determine operand sizes (inst.is_src_16, inst.is_src_64, etc.)
+  is_vop2_16bit = isinstance(inst, VOP2) and inst.is_16bit()
+
+  # Read sources based on register counts and dtypes from inst properties
+  def read_src(src, idx, regs, is_src_16):
+    if src is None: return 0
+    if regs == 2: return mod_src(st.rsrc64(src, lane), idx, is64=True)
+    if is_src_16 and isinstance(inst, VOP3):
+      raw = st.rsrc_f16(src, lane) if 128 <= src < 255 else st.rsrc(src, lane)
+      val = _src16(raw, bool(opsel & (1 << idx)))
+      if abs_ & (1 << idx): val &= 0x7fff
+      if neg & (1 << idx): val ^= 0x8000
+      return val
+    if is_src_16 and isinstance(inst, (VOP1, VOP2, VOPC)):
+      if src >= 256: return _src16(mod_src(st.rsrc(_vgpr_masked(src), lane), idx), _vgpr_hi(src))
+      return mod_src(st.rsrc_f16(src, lane), idx) & 0xffff
+    return mod_src(st.rsrc(src, lane), idx)
+
+  s0 = read_src(src0, 0, inst.src_regs(0), inst.is_src_16(0))
+  s1 = read_src(src1, 1, inst.src_regs(1), inst.is_src_16(1)) if src1 is not None else 0
+  s2 = read_src(src2, 2, inst.src_regs(2), inst.is_src_16(2)) if src2 is not None else 0
+  # Read destination (accumulator for VOP2 f16, 64-bit for 64-bit ops)
+  d0 = _src16(V[vdst], dst_hi) if is_vop2_16bit else (V[vdst] | (V[vdst + 1] << 32)) if inst.dst_regs() == 2 else V[vdst]
+
+  # V_CNDMASK_B32/B16: 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 inst.op in (VOP3Op.V_CNDMASK_B32, VOP3Op.V_CNDMASK_B16) and isinstance(inst, 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(Reg(s0), Reg(s1), Reg(s2), Reg(d0), Reg(st.scc), Reg(vcc_for_fn), lane, Reg(st.exec_mask), st.literal, st.vgpr, src0_idx, vdst)
+
+  # Apply results - extract values from returned Reg objects
+  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' in result:
+    # VOP2 carry ops write to VCC implicitly; VOPC/VOP3 write to vdst
+    st.pend_sgpr_lane(VCC_LO if isinstance(inst, VOP2) and 'CO_CI' in inst.op_name else vdst, lane, (result['VCC']._val >> lane) & 1)
+  if 'EXEC' in result:
+    # V_CMPX instructions write to EXEC per-lane (not to vdst)
+    st.pend_sgpr_lane(EXEC_LO, lane, (result['EXEC']._val >> lane) & 1)
+  elif op_cls is VOPCOp:
+    # VOPC comparison result stored in D0 bitmask, extract lane bit (non-CMPX only)
+    st.pend_sgpr_lane(vdst, lane, (result['D0']._val >> lane) & 1)
+  if op_cls is not VOPCOp and 'vgpr_write' not in result:
+    writes_to_sgpr = 'READFIRSTLANE' in inst.op_name or 'READLANE' in inst.op_name
+    d0_val = result['D0']._val
+    if writes_to_sgpr: st.wsgpr(vdst, d0_val & MASK32)
+    elif inst.dst_regs() == 2: V[vdst], V[vdst + 1] = d0_val & MASK32, (d0_val >> 32) & MASK32
+    elif inst.is_dst_16(): V[vdst] = _dst16(V[vdst], d0_val, bool(opsel & 8) if isinstance(inst, VOP3) else dst_hi)
+    else: V[vdst] = d0_val & MASK32
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# 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 16x16 f16 matrix from 16 lanes × 8 VGPRs (2 f16 per VGPR)
+  def read_f16_mat(src):
+    return [f for l in range(16) for r in range(8) for v in [st.vgpr[l][src-256+r] if src >= 256 else st.rsgpr(src+r)] for f in [_f16(v&0xffff), _f16((v>>16)&0xffff)]]
+  mat_a, mat_b = read_f16_mat(src0), read_f16_mat(src1)
+  # Read matrix C (16x16 f32) from lanes 0-31, VGPRs src2 to src2+7
+  mat_c = [_f32(st.vgpr[i % 32][src2 - 256 + i // 32] if src2 >= 256 else st.rsgpr(src2 + i // 32)) for i in range(256)]
+  # Compute D = A × B + C (16x16 matrix multiply)
+  mat_d = [sum(mat_a[row*16+k] * mat_b[col*16+k] for k in range(16)) + mat_c[row*16+col] for row in range(16) for col in range(16)]
+  # Write result - f16 packed or f32
+  if op == VOP3POp.V_WMMA_F16_16X16X16_F16:
+    for i in range(0, 256, 2):
+      st.vgpr[(i//2) % 32][vdst + (i//2)//32] = ((_i16(mat_d[i+1]) & 0xffff) << 16) | (_i16(mat_d[i]) & 0xffff)
+  else:
+    for i in range(256): st.vgpr[i % 32][vdst + i//32] = _i32(mat_d[i])
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# MAIN EXECUTION LOOP
+# ═══════════════════════════════════════════════════════════════════════════════
+
+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._words, getattr(inst, '_literal', None) or 0
+
+  if isinstance(inst, (SOP1, SOP2, SOPC, SOPK, SOPP, SMEM)):
+    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/V_READLANE write to SGPR, execute once; others execute per-lane with exec_mask
+    is_readlane = isinstance(inst, (VOP1, VOP3)) and ('READFIRSTLANE' in inst.op_name or 'READLANE' in inst.op_name)
+    exec_mask = 1 if is_readlane else st.exec_mask
+    for lane in range(1 if is_readlane else 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)
+    # Set workgroup IDs in SGPRs based on USER_SGPR_COUNT and enable flags from COMPUTE_PGM_RSRC2
+    sgpr_idx = wg_id_sgpr_base
+    for wg_id, enabled in zip(workgroup_id, wg_id_enables):
+      if enabled: st.sgpr[sgpr_idx] = wg_id; sgpr_idx += 1
+    # Set workitem IDs in VGPR0 using packed method: v0 = (Z << 20) | (Y << 10) | X
+    for i in range(n_lanes):
+      tid = wave_start + i
+      st.vgpr[i][0] = ((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:
+  program = decode_program((ctypes.c_char * lib_sz).from_address(lib).raw)
+  if not program: return -1
+  wg_id_enables = tuple(bool((rsrc2 >> (7+i)) & 1) for i in range(3))
+  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, (rsrc2 >> 1) & 0x1f, wg_id_enables)
+  return 0

extra/assembly/amd/pcode.py

@@ -0,0 +1,542 @@
+# DSL for RDNA3 pseudocode - makes pseudocode expressions work directly as Python
+import struct, math
+from extra.assembly.amd.dsl import MASK32, MASK64, _f32, _i32, _sext, _f16, _i16, _f64, _i64
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# HELPER FUNCTIONS
+# ═══════════════════════════════════════════════════════════════════════════════
+
+def _div(a, b):
+  try: return a / b
+  except ZeroDivisionError:
+    if a == 0.0 or math.isnan(a): return float("nan")
+    return math.copysign(float("inf"), a * b) if b == 0.0 else float("inf") if a > 0 else float("-inf")
+def _to_f16_bits(v): return v if isinstance(v, int) else _i16(v)
+def _isnan(x):
+  try: return math.isnan(float(x))
+  except (TypeError, ValueError): return False
+def _check_nan_type(x, quiet_bit_expected, default):
+  """Check NaN type by examining quiet bit. Returns default if can't determine."""
+  try:
+    if not math.isnan(float(x)): return False
+    if hasattr(x, '_reg') and hasattr(x, '_bits'):
+      bits = x._reg._val & ((1 << x._bits) - 1)
+      # NaN format: exponent all 1s, quiet bit, mantissa != 0
+      # f16: exp[14:10]=31, quiet=bit9, mant[8:0]  |  f32: exp[30:23]=255, quiet=bit22, mant[22:0]  |  f64: exp[62:52]=2047, quiet=bit51, mant[51:0]
+      exp_bits, quiet_pos, mant_mask = {16: (0x1f, 9, 0x3ff), 32: (0xff, 22, 0x7fffff), 64: (0x7ff, 51, 0xfffffffffffff)}.get(x._bits, (0,0,0))
+      exp_shift = {16: 10, 32: 23, 64: 52}.get(x._bits, 0)
+      if exp_bits and ((bits >> exp_shift) & exp_bits) == exp_bits and (bits & mant_mask) != 0:
+        return ((bits >> quiet_pos) & 1) == quiet_bit_expected
+    return default
+  except (TypeError, ValueError): return False
+def _isquietnan(x): return _check_nan_type(x, 1, True)  # quiet NaN has quiet bit = 1
+def _issignalnan(x): return _check_nan_type(x, 0, False)  # signaling NaN has quiet bit = 0
+def _gt_neg_zero(a, b): return (a > b) or (a == 0 and b == 0 and not math.copysign(1, a) < 0 and math.copysign(1, b) < 0)
+def _lt_neg_zero(a, b): return (a < b) or (a == 0 and b == 0 and math.copysign(1, a) < 0 and not math.copysign(1, b) < 0)
+def _fma(a, b, c): return a * b + c
+def _signext(v): return v
+def _fpop(fn): return lambda x: (x := float(x), x if math.isnan(x) or math.isinf(x) else float(fn(x)))[1]
+trunc, floor, ceil = _fpop(math.trunc), _fpop(math.floor), _fpop(math.ceil)
+class _SafeFloat(float):
+  """Float subclass that uses _div for division to handle 0/inf correctly."""
+  def __truediv__(self, o): return _div(float(self), float(o))
+  def __rtruediv__(self, o): return _div(float(o), float(self))
+def sqrt(x): return _SafeFloat(math.sqrt(x)) if x >= 0 else _SafeFloat(float("nan"))
+def log2(x): return math.log2(x) if x > 0 else (float("-inf") if x == 0 else float("nan"))
+i32_to_f32 = u32_to_f32 = i32_to_f64 = u32_to_f64 = f32_to_f64 = f64_to_f32 = float
+def _f_to_int(f, lo, hi): f = float(f); return 0 if math.isnan(f) else (hi if f >= hi else lo if f <= lo else int(f))
+def f32_to_i32(f): return _f_to_int(f, -2147483648, 2147483647)
+def f32_to_u32(f): return _f_to_int(f, 0, 4294967295)
+f64_to_i32, f64_to_u32 = f32_to_i32, f32_to_u32
+def f32_to_f16(f):
+  f = float(f)
+  if math.isnan(f): return 0x7e00  # f16 NaN
+  if math.isinf(f): return 0x7c00 if f > 0 else 0xfc00  # f16 ±infinity
+  try: return struct.unpack("<H", struct.pack("<e", f))[0]
+  except OverflowError: return 0x7c00 if f > 0 else 0xfc00  # overflow -> ±infinity
+def _f16_to_f32_bits(bits): return struct.unpack("<e", struct.pack("<H", int(bits) & 0xffff))[0]
+def f16_to_f32(v): return v if isinstance(v, float) else _f16_to_f32_bits(v)
+def i16_to_f16(v): return f32_to_f16(float(_sext(int(v) & 0xffff, 16)))
+def u16_to_f16(v): return f32_to_f16(float(int(v) & 0xffff))
+def f16_to_i16(bits): f = _f16_to_f32_bits(bits); return max(-32768, min(32767, int(f))) if not math.isnan(f) else 0
+def f16_to_u16(bits): f = _f16_to_f32_bits(bits); return max(0, min(65535, int(f))) if not math.isnan(f) else 0
+def u8_to_u32(v): return int(v) & 0xff
+def u4_to_u32(v): return int(v) & 0xf
+def _sign(f): return 1 if math.copysign(1.0, f) < 0 else 0
+def _mantissa_f32(f): return struct.unpack("<I", struct.pack("<f", f))[0] & 0x7fffff if not (math.isinf(f) or math.isnan(f)) else 0
+def _ldexp(m, e): return math.ldexp(m, e)
+def isEven(x):
+  x = float(x)
+  if math.isinf(x) or math.isnan(x): return False
+  return int(x) % 2 == 0
+def fract(x): return x - math.floor(x)
+PI = math.pi
+def _trig(fn, x):
+  # V_SIN/COS_F32: hardware does frac on input cycles before computing
+  if math.isinf(x) or math.isnan(x): return float("nan")
+  frac_cycles = fract(x / (2 * math.pi))
+  return fn(frac_cycles * 2 * math.pi)
+def sin(x): return _trig(math.sin, x)
+def cos(x): return _trig(math.cos, x)
+def pow(a, b):
+  try: return a ** b
+  except OverflowError: return float("inf") if b > 0 else 0.0
+def _brev(v, bits): return int(bin(v & ((1 << bits) - 1))[2:].zfill(bits)[::-1], 2)
+def _brev32(v): return _brev(v, 32)
+def _brev64(v): return _brev(v, 64)
+def _ctz(v, bits):
+  v, n = int(v) & ((1 << bits) - 1), 0
+  if v == 0: return bits
+  while (v & 1) == 0: v >>= 1; n += 1
+  return n
+def _ctz32(v): return _ctz(v, 32)
+def _ctz64(v): return _ctz(v, 64)
+def _exponent(f):
+  # Handle TypedView (f16/f32/f64) to get correct exponent for that type
+  if hasattr(f, '_bits') and hasattr(f, '_float') and f._float:
+    raw = f._val
+    if f._bits == 16: return (raw >> 10) & 0x1f  # f16: 5-bit exponent
+    if f._bits == 32: return (raw >> 23) & 0xff  # f32: 8-bit exponent
+    if f._bits == 64: return (raw >> 52) & 0x7ff  # f64: 11-bit exponent
+  # Fallback: convert to f32 and get exponent
+  f = float(f)
+  if math.isinf(f) or math.isnan(f): return 255
+  if f == 0.0: return 0
+  try: bits = struct.unpack("<I", struct.pack("<f", f))[0]; return (bits >> 23) & 0xff
+  except: return 0
+def _is_denorm_f32(f):
+  if not isinstance(f, float): f = _f32(int(f) & 0xffffffff)
+  if math.isinf(f) or math.isnan(f) or f == 0.0: return False
+  bits = struct.unpack("<I", struct.pack("<f", float(f)))[0]
+  return (bits >> 23) & 0xff == 0
+def _is_denorm_f64(f):
+  if not isinstance(f, float): f = _f64(int(f) & 0xffffffffffffffff)
+  if math.isinf(f) or math.isnan(f) or f == 0.0: return False
+  bits = struct.unpack("<Q", struct.pack("<d", float(f)))[0]
+  return (bits >> 52) & 0x7ff == 0
+def v_min_f32(a, b): return a if math.isnan(b) else b if math.isnan(a) else (a if _lt_neg_zero(a, b) else b)
+def v_max_f32(a, b): return a if math.isnan(b) else b if math.isnan(a) else (a if _gt_neg_zero(a, b) else b)
+v_min_f16, v_max_f16 = v_min_f32, v_max_f32
+v_min_i32, v_max_i32 = min, max
+v_min_i16, v_max_i16 = min, max
+def v_min_u32(a, b): return min(a & MASK32, b & MASK32)
+def v_max_u32(a, b): return max(a & MASK32, b & MASK32)
+def v_min_u16(a, b): return min(a & 0xffff, b & 0xffff)
+def v_max_u16(a, b): return max(a & 0xffff, b & 0xffff)
+def v_min3_f32(a, b, c): return v_min_f32(v_min_f32(a, b), c)
+def v_max3_f32(a, b, c): return v_max_f32(v_max_f32(a, b), c)
+v_min3_f16, v_max3_f16 = v_min3_f32, v_max3_f32
+v_min3_i32, v_max3_i32, v_min3_i16, v_max3_i16 = min, max, min, max
+def v_min3_u32(a, b, c): return min(a & MASK32, b & MASK32, c & MASK32)
+def v_max3_u32(a, b, c): return max(a & MASK32, b & MASK32, c & MASK32)
+def v_min3_u16(a, b, c): return min(a & 0xffff, b & 0xffff, c & 0xffff)
+def v_max3_u16(a, b, c): return max(a & 0xffff, b & 0xffff, c & 0xffff)
+def ABSDIFF(a, b): return abs(int(a) - int(b))
+
+# BF16 (bfloat16) conversion functions
+def _bf16(i):
+  """Convert bf16 bits to float. BF16 is just the top 16 bits of f32."""
+  return struct.unpack("<f", struct.pack("<I", (i & 0xffff) << 16))[0]
+def _ibf16(f):
+  """Convert float to bf16 bits (truncate to top 16 bits of f32)."""
+  if math.isnan(f): return 0x7fc0  # bf16 quiet NaN
+  if math.isinf(f): return 0x7f80 if f > 0 else 0xff80  # bf16 ±infinity
+  try: return (struct.unpack("<I", struct.pack("<f", float(f)))[0] >> 16) & 0xffff
+  except (OverflowError, struct.error): return 0x7f80 if f > 0 else 0xff80
+def bf16_to_f32(v): return _bf16(v) if isinstance(v, int) else float(v)
+def f32_to_bf16(f): return _ibf16(f)
+
+# BYTE_PERMUTE for V_PERM_B32 - select bytes from 64-bit data based on selector
+def BYTE_PERMUTE(data, sel):
+  """Select a byte from 64-bit data based on selector value.
+  sel 0-7: select byte from data (S1 is bytes 0-3, S0 is bytes 4-7 in {S0,S1})
+  sel 8-11: sign-extend from specific bytes (8->byte1, 9->byte3, 10->byte5, 11->byte7)
+  sel 12: constant 0x00
+  sel >= 13: constant 0xFF"""
+  sel = int(sel) & 0xff
+  if sel <= 7: return (int(data) >> (sel * 8)) & 0xff
+  if sel == 8: return 0xff if ((int(data) >> 15) & 1) else 0x00  # sign of byte 1
+  if sel == 9: return 0xff if ((int(data) >> 31) & 1) else 0x00  # sign of byte 3
+  if sel == 10: return 0xff if ((int(data) >> 47) & 1) else 0x00  # sign of byte 5
+  if sel == 11: return 0xff if ((int(data) >> 63) & 1) else 0x00  # sign of byte 7
+  if sel == 12: return 0x00
+  return 0xff  # sel >= 13
+
+# v_sad_u8 helper for V_SAD instructions (sum of absolute differences of 4 bytes)
+def v_sad_u8(s0, s1, s2):
+  """V_SAD_U8: Sum of absolute differences of 4 byte pairs plus accumulator."""
+  s0, s1, s2 = int(s0), int(s1), int(s2)
+  result = s2
+  for i in range(4):
+    a = (s0 >> (i * 8)) & 0xff
+    b = (s1 >> (i * 8)) & 0xff
+    result += abs(a - b)
+  return result & 0xffffffff
+
+# v_msad_u8 helper (masked SAD - skip when reference byte is 0)
+def v_msad_u8(s0, s1, s2):
+  """V_MSAD_U8: Masked sum of absolute differences (skip if reference byte is 0)."""
+  s0, s1, s2 = int(s0), int(s1), int(s2)
+  result = s2
+  for i in range(4):
+    a = (s0 >> (i * 8)) & 0xff
+    b = (s1 >> (i * 8)) & 0xff
+    if b != 0:  # Only add diff if reference (s1) byte is non-zero
+      result += abs(a - b)
+  return result & 0xffffffff
+def f16_to_snorm(f): return max(-32768, min(32767, int(round(max(-1.0, min(1.0, f)) * 32767))))
+def f16_to_unorm(f): return max(0, min(65535, int(round(max(0.0, min(1.0, f)) * 65535))))
+def f32_to_snorm(f): return max(-32768, min(32767, int(round(max(-1.0, min(1.0, f)) * 32767))))
+def f32_to_unorm(f): return max(0, min(65535, int(round(max(0.0, min(1.0, f)) * 65535))))
+def v_cvt_i16_f32(f): return max(-32768, min(32767, int(f))) if not math.isnan(f) else 0
+def v_cvt_u16_f32(f): return max(0, min(65535, int(f))) if not math.isnan(f) else 0
+def u32_to_u16(u): return int(u) & 0xffff
+def i32_to_i16(i): return ((int(i) + 32768) & 0xffff) - 32768
+def SAT8(v): return max(0, min(255, int(v)))
+def f32_to_u8(f): return max(0, min(255, int(f))) if not math.isnan(f) else 0
+def mantissa(f):
+  if f == 0.0 or math.isinf(f) or math.isnan(f): return f
+  m, _ = math.frexp(f)
+  return m  # AMD V_FREXP_MANT returns mantissa in [0.5, 1.0) range
+def signext_from_bit(val, bit):
+  bit = int(bit)
+  if bit == 0: return 0
+  mask = (1 << bit) - 1
+  val = int(val) & mask
+  if val & (1 << (bit - 1)): return val - (1 << bit)
+  return val
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# DSL EXPORTS
+# ═══════════════════════════════════════════════════════════════════════════════
+
+__all__ = [
+  # Classes
+  'Reg', 'SliceProxy', 'TypedView',
+  # Pack functions
+  '_pack', '_pack32', 'pack', 'pack32',
+  # Constants
+  'WAVE32', 'WAVE64', 'MASK32', 'MASK64', 'WAVE_MODE', 'DENORM', 'OVERFLOW_F32', 'UNDERFLOW_F32',
+  'OVERFLOW_F64', 'UNDERFLOW_F64', 'MAX_FLOAT_F32', 'ROUND_MODE', 'cvtToQuietNAN', 'DST', 'INF', 'PI',
+  'TWO_OVER_PI_1201',
+  # Aliases for pseudocode
+  's_ff1_i32_b32', 's_ff1_i32_b64', 'GT_NEG_ZERO', 'LT_NEG_ZERO',
+  'isNAN', 'isQuietNAN', 'isSignalNAN', 'fma', 'ldexp', 'sign', 'exponent', 'F', 'signext',
+  # Conversion functions
+  '_f32', '_i32', '_f16', '_i16', '_f64', '_i64', '_sext', '_to_f16_bits', '_f16_to_f32_bits',
+  'i32_to_f32', 'u32_to_f32', 'i32_to_f64', 'u32_to_f64', 'f32_to_f64', 'f64_to_f32',
+  'f32_to_i32', 'f32_to_u32', 'f64_to_i32', 'f64_to_u32', 'f32_to_f16', 'f16_to_f32',
+  'i16_to_f16', 'u16_to_f16', 'f16_to_i16', 'f16_to_u16', 'u32_to_u16', 'i32_to_i16',
+  'f16_to_snorm', 'f16_to_unorm', 'f32_to_snorm', 'f32_to_unorm', 'v_cvt_i16_f32', 'v_cvt_u16_f32',
+  'SAT8', 'f32_to_u8', 'u8_to_u32', 'u4_to_u32',
+  # BF16 conversion functions
+  '_bf16', '_ibf16', 'bf16_to_f32', 'f32_to_bf16',
+  # Math functions
+  'trunc', 'floor', 'ceil', 'sqrt', 'log2', 'sin', 'cos', 'pow', 'fract', 'isEven', 'mantissa',
+  # Min/max functions
+  'v_min_f32', 'v_max_f32', 'v_min_i32', 'v_max_i32', 'v_min_u32', 'v_max_u32',
+  'v_min_f16', 'v_max_f16', 'v_min_i16', 'v_max_i16', 'v_min_u16', 'v_max_u16',
+  'v_min3_f32', 'v_max3_f32', 'v_min3_i32', 'v_max3_i32', 'v_min3_u32', 'v_max3_u32',
+  'v_min3_f16', 'v_max3_f16', 'v_min3_i16', 'v_max3_i16', 'v_min3_u16', 'v_max3_u16',
+  'ABSDIFF',
+  # Byte/SAD helper functions
+  'BYTE_PERMUTE', 'v_sad_u8', 'v_msad_u8',
+  # Bit manipulation
+  '_brev32', '_brev64', '_ctz32', '_ctz64', '_exponent', '_is_denorm_f32', '_is_denorm_f64',
+  '_sign', '_mantissa_f32', '_div', '_isnan', '_isquietnan', '_issignalnan', '_gt_neg_zero', '_lt_neg_zero', '_fma', '_ldexp', '_signext',
+  'signext_from_bit',
+]
+
+# Aliases used in pseudocode
+s_ff1_i32_b32, s_ff1_i32_b64 = _ctz32, _ctz64
+GT_NEG_ZERO, LT_NEG_ZERO = _gt_neg_zero, _lt_neg_zero
+isNAN = _isnan
+isQuietNAN = _isquietnan
+isSignalNAN = _issignalnan
+fma, ldexp, sign, exponent = _fma, _ldexp, _sign, _exponent
+def F(x):
+  """32'F(x) or 64'F(x) - interpret x as float. If x is int, treat as bit pattern."""
+  if isinstance(x, int): return _f32(x)  # int -> interpret as f32 bits
+  if isinstance(x, TypedView): return x  # preserve TypedView for bit-pattern checks
+  return float(x)  # already a float or float-like
+signext = lambda x: int(x)  # sign-extend to full width - already handled by Python's arbitrary precision ints
+pack = lambda hi, lo: ((int(hi) & 0xffff) << 16) | (int(lo) & 0xffff)
+pack32 = lambda hi, lo: ((int(hi) & 0xffffffff) << 32) | (int(lo) & 0xffffffff)
+_pack, _pack32 = pack, pack32  # Aliases for internal use
+WAVE32, WAVE64 = True, False
+
+# Float overflow/underflow constants
+OVERFLOW_F32 = float('inf')
+UNDERFLOW_F32 = 0.0
+OVERFLOW_F64 = float('inf')
+UNDERFLOW_F64 = 0.0
+MAX_FLOAT_F32 = 3.4028235e+38  # Largest finite float32
+
+# INF object that supports .f16/.f32/.f64 access and comparison with floats
+class _Inf:
+  f16 = f32 = f64 = float('inf')
+  def __neg__(self): return _NegInf()
+  def __pos__(self): return self
+  def __float__(self): return float('inf')
+  def __eq__(self, other): return float(other) == float('inf') if not isinstance(other, _NegInf) else False
+  def __req__(self, other): return self.__eq__(other)
+class _NegInf:
+  f16 = f32 = f64 = float('-inf')
+  def __neg__(self): return _Inf()
+  def __pos__(self): return self
+  def __float__(self): return float('-inf')
+  def __eq__(self, other): return float(other) == float('-inf') if not isinstance(other, _Inf) else False
+  def __req__(self, other): return self.__eq__(other)
+INF = _Inf()
+
+# Rounding mode placeholder
+class _RoundMode:
+  NEAREST_EVEN = 0
+ROUND_MODE = _RoundMode()
+
+# Helper functions for pseudocode
+def cvtToQuietNAN(x): return float('nan')
+DST = None  # Placeholder, will be set in context
+
+# 2/PI with 1201 bits of precision for V_TRIG_PREOP_F64
+# Computed as: int((2/pi) * 2^1201) - this is the fractional part of 2/pi scaled to integer
+# The MSB (bit 1200) corresponds to 2^0 position in the fraction 0.b1200 b1199 ... b1 b0
+_TWO_OVER_PI_1201_RAW = 0x0145f306dc9c882a53f84eafa3ea69bb81b6c52b3278872083fca2c757bd778ac36e48dc74849ba5c00c925dd413a32439fc3bd63962534e7dd1046bea5d768909d338e04d68befc827323ac7306a673e93908bf177bf250763ff12fffbc0b301fde5e2316b414da3eda6cfd9e4f96136e9e8c7ecd3cbfd45aea4f758fd7cbe2f67a0e73ef14a525d4d7f6bf623f1aba10ac06608df8f6
+
+class _BigInt:
+  """Wrapper for large integers that supports bit slicing [high:low]."""
+  __slots__ = ('_val',)
+  def __init__(self, val): self._val = val
+  def __getitem__(self, key):
+    if isinstance(key, slice):
+      high, low = key.start, key.stop
+      if high < low: high, low = low, high  # Handle reversed slice
+      mask = (1 << (high - low + 1)) - 1
+      return (self._val >> low) & mask
+    return (self._val >> key) & 1
+  def __int__(self): return self._val
+  def __index__(self): return self._val
+  def __lshift__(self, n): return self._val << int(n)
+  def __rshift__(self, n): return self._val >> int(n)
+  def __and__(self, n): return self._val & int(n)
+  def __or__(self, n): return self._val | int(n)
+
+TWO_OVER_PI_1201 = _BigInt(_TWO_OVER_PI_1201_RAW)
+
+class _WaveMode:
+  IEEE = False
+WAVE_MODE = _WaveMode()
+
+class _DenormChecker:
+  """Comparator for denormalized floats. x == DENORM.f32 checks if x is denormalized."""
+  def __init__(self, bits): self._bits = bits
+  def _check(self, other):
+    return _is_denorm_f64(float(other)) if self._bits == 64 else _is_denorm_f32(float(other))
+  def __eq__(self, other): return self._check(other)
+  def __req__(self, other): return self._check(other)
+  def __ne__(self, other): return not self._check(other)
+
+class _Denorm:
+  f32 = _DenormChecker(32)
+  f64 = _DenormChecker(64)
+DENORM = _Denorm()
+
+def _brev(v, bits):
+  """Bit-reverse a value."""
+  result = 0
+  for i in range(bits): result |= ((v >> i) & 1) << (bits - 1 - i)
+  return result
+
+class SliceProxy:
+  """Proxy for D0[31:16] that supports .f16/.u16 etc getters and setters."""
+  __slots__ = ('_reg', '_high', '_low', '_reversed')
+  def __init__(self, reg, high, low):
+    self._reg = reg
+    # Handle reversed slices like [0:31] which means bit-reverse
+    if high < low: self._high, self._low, self._reversed = low, high, True
+    else: self._high, self._low, self._reversed = high, low, False
+  def _nbits(self): return self._high - self._low + 1
+  def _mask(self): return (1 << self._nbits()) - 1
+  def _get(self):
+    v = (self._reg._val >> self._low) & self._mask()
+    return _brev(v, self._nbits()) if self._reversed else v
+  def _set(self, v):
+    v = int(v)
+    if self._reversed: v = _brev(v, self._nbits())
+    self._reg._val = (self._reg._val & ~(self._mask() << self._low)) | ((v & self._mask()) << self._low)
+
+  u8 = property(lambda s: s._get() & 0xff)
+  u16 = property(lambda s: s._get() & 0xffff, lambda s, v: s._set(v))
+  u32 = property(lambda s: s._get() & MASK32, lambda s, v: s._set(v))
+  i16 = property(lambda s: _sext(s._get() & 0xffff, 16), lambda s, v: s._set(v))
+  i32 = property(lambda s: _sext(s._get() & MASK32, 32), lambda s, v: s._set(v))
+  f16 = property(lambda s: _f16(s._get()), lambda s, v: s._set(v if isinstance(v, int) else _i16(float(v))))
+  f32 = property(lambda s: _f32(s._get()), lambda s, v: s._set(_i32(float(v))))
+  bf16 = property(lambda s: _bf16(s._get()), lambda s, v: s._set(v if isinstance(v, int) else _ibf16(float(v))))
+  b16, b32 = u16, u32
+
+  def __int__(self): return self._get()
+  def __index__(self): return self._get()
+
+  # Comparison operators (compare as integers)
+  def __eq__(s, o): return s._get() == int(o)
+  def __ne__(s, o): return s._get() != int(o)
+  def __lt__(s, o): return s._get() < int(o)
+  def __le__(s, o): return s._get() <= int(o)
+  def __gt__(s, o): return s._get() > int(o)
+  def __ge__(s, o): return s._get() >= int(o)
+
+class TypedView:
+  """View for S0.u32 that supports [4:0] slicing and [bit] access."""
+  __slots__ = ('_reg', '_bits', '_signed', '_float', '_bf16')
+  def __init__(self, reg, bits, signed=False, is_float=False, is_bf16=False):
+    self._reg, self._bits, self._signed, self._float, self._bf16 = reg, bits, signed, is_float, is_bf16
+
+  @property
+  def _val(self):
+    mask = MASK64 if self._bits == 64 else MASK32 if self._bits == 32 else (1 << self._bits) - 1
+    return self._reg._val & mask
+
+  def __getitem__(self, key):
+    if isinstance(key, slice):
+      high, low = int(key.start), int(key.stop)
+      return SliceProxy(self._reg, high, low)
+    return (self._val >> int(key)) & 1
+
+  def __setitem__(self, key, value):
+    if isinstance(key, slice):
+      high, low = int(key.start), int(key.stop)
+      if high < low: high, low, value = low, high, _brev(int(value), low - high + 1)
+      mask = (1 << (high - low + 1)) - 1
+      self._reg._val = (self._reg._val & ~(mask << low)) | ((int(value) & mask) << low)
+    elif value: self._reg._val |= (1 << int(key))
+    else: self._reg._val &= ~(1 << int(key))
+
+  def __int__(self): return _sext(self._val, self._bits) if self._signed else self._val
+  def __index__(self): return int(self)
+  def __trunc__(self): return int(float(self)) if self._float else int(self)
+  def __float__(self):
+    if self._float:
+      if self._bf16: return _bf16(self._val)  # bf16 uses different conversion
+      return _f16(self._val) if self._bits == 16 else _f32(self._val) if self._bits == 32 else _f64(self._val)
+    return float(int(self))
+
+  # Arithmetic - floats use float(), ints use int()
+  def __add__(s, o): return float(s) + float(o) if s._float else int(s) + int(o)
+  def __radd__(s, o): return float(o) + float(s) if s._float else int(o) + int(s)
+  def __sub__(s, o): return float(s) - float(o) if s._float else int(s) - int(o)
+  def __rsub__(s, o): return float(o) - float(s) if s._float else int(o) - int(s)
+  def __mul__(s, o): return float(s) * float(o) if s._float else int(s) * int(o)
+  def __rmul__(s, o): return float(o) * float(s) if s._float else int(o) * int(s)
+  def __truediv__(s, o): return _div(float(s), float(o)) if s._float else _div(int(s), int(o))
+  def __rtruediv__(s, o): return _div(float(o), float(s)) if s._float else _div(int(o), int(s))
+  def __pow__(s, o): return float(s) ** float(o) if s._float else int(s) ** int(o)
+  def __rpow__(s, o): return float(o) ** float(s) if s._float else int(o) ** int(s)
+  def __neg__(s): return -float(s) if s._float else -int(s)
+  def __abs__(s): return abs(float(s)) if s._float else abs(int(s))
+
+  # Bitwise - GPU shifts mask the shift amount to valid range
+  def __and__(s, o): return int(s) & int(o)
+  def __or__(s, o): return int(s) | int(o)
+  def __xor__(s, o): return int(s) ^ int(o)
+  def __invert__(s): return ~int(s)
+  def __lshift__(s, o): n = int(o); return int(s) << n if 0 <= n < 64 else 0
+  def __rshift__(s, o): n = int(o); return int(s) >> n if 0 <= n < 64 else 0
+  def __rand__(s, o): return int(o) & int(s)
+  def __ror__(s, o): return int(o) | int(s)
+  def __rxor__(s, o): return int(o) ^ int(s)
+  def __rlshift__(s, o): n = int(s); return int(o) << n if 0 <= n < 64 else 0
+  def __rrshift__(s, o): n = int(s); return int(o) >> n if 0 <= n < 64 else 0
+
+  # Comparison - handle _DenormChecker specially
+  def __eq__(s, o):
+    if isinstance(o, _DenormChecker): return o._check(s)
+    return float(s) == float(o) if s._float else int(s) == int(o)
+  def __ne__(s, o):
+    if isinstance(o, _DenormChecker): return not o._check(s)
+    return float(s) != float(o) if s._float else int(s) != int(o)
+  def __lt__(s, o): return float(s) < float(o) if s._float else int(s) < int(o)
+  def __le__(s, o): return float(s) <= float(o) if s._float else int(s) <= int(o)
+  def __gt__(s, o): return float(s) > float(o) if s._float else int(s) > int(o)
+  def __ge__(s, o): return float(s) >= float(o) if s._float else int(s) >= int(o)
+
+  def __bool__(s): return bool(int(s))
+
+  # Allow chained type access like jump_addr.i64 when jump_addr is already a TypedView
+  # These just return self or convert appropriately
+  @property
+  def i64(s): return s if s._bits == 64 and s._signed else int(s)
+  @property
+  def u64(s): return s if s._bits == 64 and not s._signed else int(s) & MASK64
+  @property
+  def i32(s): return s if s._bits == 32 and s._signed else _sext(int(s) & MASK32, 32)
+  @property
+  def u32(s): return s if s._bits == 32 and not s._signed else int(s) & MASK32
+
+class Reg:
+  """GPU register: D0.f32 = S0.f32 + S1.f32 just works."""
+  __slots__ = ('_val',)
+  def __init__(self, val=0): self._val = int(val) & MASK64
+
+  # Typed views
+  u64 = property(lambda s: TypedView(s, 64), lambda s, v: setattr(s, '_val', int(v) & MASK64))
+  i64 = property(lambda s: TypedView(s, 64, signed=True), lambda s, v: setattr(s, '_val', int(v) & MASK64))
+  b64 = property(lambda s: TypedView(s, 64), lambda s, v: setattr(s, '_val', int(v) & MASK64))
+  f64 = property(lambda s: TypedView(s, 64, is_float=True), lambda s, v: setattr(s, '_val', v if isinstance(v, int) else _i64(float(v))))
+  u32 = property(lambda s: TypedView(s, 32), lambda s, v: setattr(s, '_val', int(v) & MASK32))
+  i32 = property(lambda s: TypedView(s, 32, signed=True), lambda s, v: setattr(s, '_val', int(v) & MASK32))
+  b32 = property(lambda s: TypedView(s, 32), lambda s, v: setattr(s, '_val', int(v) & MASK32))
+  f32 = property(lambda s: TypedView(s, 32, is_float=True), lambda s, v: setattr(s, '_val', _i32(float(v))))
+  u24 = property(lambda s: TypedView(s, 24))
+  i24 = property(lambda s: TypedView(s, 24, signed=True))
+  u16 = property(lambda s: TypedView(s, 16), lambda s, v: setattr(s, '_val', (s._val & 0xffff0000) | (int(v) & 0xffff)))
+  i16 = property(lambda s: TypedView(s, 16, signed=True), lambda s, v: setattr(s, '_val', (s._val & 0xffff0000) | (int(v) & 0xffff)))
+  b16 = property(lambda s: TypedView(s, 16), lambda s, v: setattr(s, '_val', (s._val & 0xffff0000) | (int(v) & 0xffff)))
+  f16 = property(lambda s: TypedView(s, 16, is_float=True), lambda s, v: setattr(s, '_val', (s._val & 0xffff0000) | ((v if isinstance(v, int) else _i16(float(v))) & 0xffff)))
+  bf16 = property(lambda s: TypedView(s, 16, is_float=True, is_bf16=True), lambda s, v: setattr(s, '_val', (s._val & 0xffff0000) | ((v if isinstance(v, int) else _ibf16(float(v))) & 0xffff)))
+  u8 = property(lambda s: TypedView(s, 8))
+  i8 = property(lambda s: TypedView(s, 8, signed=True))
+  u1 = property(lambda s: TypedView(s, 1))  # single bit
+
+  def __getitem__(s, key):
+    if isinstance(key, slice): return SliceProxy(s, int(key.start), int(key.stop))
+    return (s._val >> int(key)) & 1
+
+  def __setitem__(s, key, value):
+    if isinstance(key, slice):
+      high, low = int(key.start), int(key.stop)
+      mask = (1 << (high - low + 1)) - 1
+      s._val = (s._val & ~(mask << low)) | ((int(value) & mask) << low)
+    elif value: s._val |= (1 << int(key))
+    else: s._val &= ~(1 << int(key))
+
+  def __int__(s): return s._val
+  def __index__(s): return s._val
+  def __bool__(s): return bool(s._val)
+
+  # Arithmetic (for tmp = tmp + 1 patterns). Float operands trigger f32 interpretation.
+  def __add__(s, o): return (_f32(s._val) + float(o)) if isinstance(o, float) else s._val + int(o)
+  def __radd__(s, o): return (float(o) + _f32(s._val)) if isinstance(o, float) else int(o) + s._val
+  def __sub__(s, o): return (_f32(s._val) - float(o)) if isinstance(o, float) else s._val - int(o)
+  def __rsub__(s, o): return (float(o) - _f32(s._val)) if isinstance(o, float) else int(o) - s._val
+  def __mul__(s, o): return (_f32(s._val) * float(o)) if isinstance(o, float) else s._val * int(o)
+  def __rmul__(s, o): return (float(o) * _f32(s._val)) if isinstance(o, float) else int(o) * s._val
+  def __and__(s, o): return s._val & int(o)
+  def __rand__(s, o): return int(o) & s._val
+  def __or__(s, o): return s._val | int(o)
+  def __ror__(s, o): return int(o) | s._val
+  def __xor__(s, o): return s._val ^ int(o)
+  def __rxor__(s, o): return int(o) ^ s._val
+  def __lshift__(s, o): n = int(o); return s._val << n if 0 <= n < 64 else 0
+  def __rshift__(s, o): n = int(o); return s._val >> n if 0 <= n < 64 else 0
+  def __invert__(s): return ~s._val
+
+  # Comparison (for tmp >= 0x100000000 patterns)
+  def __lt__(s, o): return s._val < int(o)
+  def __le__(s, o): return s._val <= int(o)
+  def __gt__(s, o): return s._val > int(o)
+  def __ge__(s, o): return s._val >= int(o)
+  def __eq__(s, o): return s._val == int(o)
+  def __ne__(s, o): return s._val != int(o)
+
+

extra/assembly/amd/pdf.py

@@ -0,0 +1,670 @@
+# Generate AMD ISA autogen files from PDF documentation
+# Combines format/enum generation (previously in dsl.py) and pseudocode compilation (previously in pcode.py)
+# Usage: python -m extra.assembly.amd.pdf [--arch rdna3|rdna4|cdna|all]
+import re, functools
+from pathlib import Path
+from concurrent.futures import ProcessPoolExecutor
+
+PDF_URLS = {
+  "rdna3": "https://docs.amd.com/api/khub/documents/UVVZM22UN7tMUeiW_4ShTQ/content",
+  "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 type mappings and ordering
+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'}
+INST_PATTERN = re.compile(r'^([SVD]S?_[A-Z0-9_]+)\s+(\d+)\s*$', re.M)
+
+# Patterns that can't be handled by the DSL (require special handling in emu.py)
+UNSUPPORTED = ['SGPR[', 'V_SWAP', 'eval ', 'FATAL_HALT', 'HW_REGISTERS',
+               'vscnt', 'vmcnt', 'expcnt', 'lgkmcnt',
+               'CVT_OFF_TABLE', 'ThreadMask',
+               'S1[i', 'C.i32',
+               'if n.', 'DST.u32', 'addrd = DST', 'addr = DST',
+               'BARRIER_STATE', 'ReallocVgprs',
+               'GPR_IDX', 'VSKIP', 'specified in', 'TTBL',
+               'fp6', 'bf6']  # Malformed pseudocode from PDF
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# COMPILER: pseudocode -> Python (minimal transforms)
+# ═══════════════════════════════════════════════════════════════════════════════
+
+def compile_pseudocode(pseudocode: str) -> str:
+  """Compile pseudocode to Python. Transforms are minimal - most syntax just works."""
+  pseudocode = re.sub(r'\bpass\b', 'pass_', pseudocode)  # 'pass' is Python keyword
+  raw_lines = pseudocode.strip().split('\n')
+  joined_lines: list[str] = []
+  for line in raw_lines:
+    line = line.strip()
+    if joined_lines and (joined_lines[-1].rstrip().endswith(('||', '&&', '(', ',')) or
+                         (joined_lines[-1].count('(') > joined_lines[-1].count(')'))):
+      joined_lines[-1] = joined_lines[-1].rstrip() + ' ' + line
+    else:
+      joined_lines.append(line)
+
+  lines = []
+  indent, need_pass, in_first_match_loop = 0, False, False
+  declared_arrays: dict[str, int] = {}  # Track declared arrays: name -> size
+  for line in joined_lines:
+    line = line.strip()
+    if not line or line.startswith('//'): continue
+    if line.startswith('if '):
+      lines.append('  ' * indent + f"if {_expr(line[3:].rstrip(' then'), declared_arrays)}:")
+      indent += 1
+      need_pass = True
+    elif line.startswith('elsif '):
+      if need_pass: lines.append('  ' * indent + "pass")
+      indent -= 1
+      lines.append('  ' * indent + f"elif {_expr(line[6:].rstrip(' then'), declared_arrays)}:")
+      indent += 1
+      need_pass = True
+    elif line == 'else':
+      if need_pass: lines.append('  ' * indent + "pass")
+      indent -= 1
+      lines.append('  ' * indent + "else:")
+      indent += 1
+      need_pass = True
+    elif line.startswith('endif'):
+      if need_pass: lines.append('  ' * indent + "pass")
+      indent -= 1
+      need_pass = False
+    elif line.startswith('endfor'):
+      if need_pass: lines.append('  ' * indent + "pass")
+      indent -= 1
+      need_pass, in_first_match_loop = False, False
+    elif m := re.match(r'declare\s+(\w+)\s*:\s*\d+\'[FBU]\[(\d+)\]', line):
+      # Handle array declarations: declare in : 32'F[3] or declare S : 32'B[3]
+      arr_name, arr_size = m[1], int(m[2])
+      declared_arrays[arr_name] = arr_size
+      py_name = f"{arr_name}_" if arr_name == 'in' else arr_name  # 'in' is Python keyword
+      if arr_name == 'S':
+        lines.append('  ' * indent + f"{py_name} = [S0, S1, S2]")  # Map to source registers
+      else:
+        lines.append('  ' * indent + f"{py_name} = [Reg(0) for _ in range({arr_size})]")
+    elif line.startswith('declare '):
+      pass  # Ignore other declare statements
+    elif m := re.match(r'for (\w+) in (.+?)\s*:\s*(.+?) do', line):
+      start, end = _expr(m[2].strip(), declared_arrays), _expr(m[3].strip(), declared_arrays)
+      lines.append('  ' * indent + f"for {m[1]} in range({start}, int({end})+1):")
+      indent += 1
+      need_pass, in_first_match_loop = True, True
+    elif '=' in line and not line.startswith('=='):
+      need_pass = False
+      line = line.rstrip(';')
+      if m := re.match(r'\{\s*D1\.[ui]1\s*,\s*D0\.[ui]64\s*\}\s*=\s*(.+)', line):
+        rhs = _expr(m[1], declared_arrays)
+        lines.append('  ' * indent + f"_full = {rhs}")
+        lines.append('  ' * indent + f"D0.u64 = int(_full) & 0xffffffffffffffff")
+        lines.append('  ' * indent + f"D1 = Reg((int(_full) >> 64) & 1)")
+      elif any(op in line for op in ('+=', '-=', '*=', '/=', '|=', '&=', '^=')):
+        for op in ('+=', '-=', '*=', '/=', '|=', '&=', '^='):
+          if op in line:
+            lhs, rhs = line.split(op, 1)
+            lhs_s = _expr(lhs.strip(), declared_arrays)  # Transform LHS too for array access
+            lines.append('  ' * indent + f"{lhs_s} {op} {_expr(rhs.strip(), declared_arrays)}")
+            break
+      else:
+        lhs, rhs = line.split('=', 1)
+        lhs_s, rhs_s = lhs.strip(), rhs.strip()
+        lhs_t = _expr(lhs_s, declared_arrays)  # Transform LHS for array access
+        stmt = _assign(lhs_t, _expr(rhs_s, declared_arrays), declared_arrays)
+        if in_first_match_loop and rhs_s == 'i' and (lhs_s == 'tmp' or lhs_s == 'D0.i32'):
+          stmt += "; break"
+        lines.append('  ' * indent + stmt)
+  if need_pass: lines.append('  ' * indent + "pass")
+  return '\n'.join(lines)
+
+def _assign(lhs: str, rhs: str, declared_arrays: dict[str, int] | None = None) -> str:
+  # Check for array element assignment: in_[i] should not wrap in Reg()
+  if declared_arrays and re.match(r'\w+_?\[\w+\]', lhs):
+    return f"{lhs} = {rhs}"
+  if lhs in ('tmp', 'SCC', 'VCC', 'EXEC', 'D0', 'D1', 'saveexec', 'PC'):
+    return f"{lhs} = Reg({rhs})"
+  return f"{lhs} = {rhs}"
+
+def _expr(e: str, declared_arrays: dict[str, int] | None = None) -> str:
+  e = e.strip()
+  # Handle OPSEL_HI.u3[i] and OPSEL.u3[i] - bit extraction from opsel fields
+  e = re.sub(r'(OPSEL(?:_HI)?)\.u\d+\[(\w+)\]', r'((\1 >> \2) & 1)', e)
+  # Rename 'in' to 'in_' to avoid Python keyword conflict
+  e = re.sub(r'\bin\[', 'in_[', e)
+  e = e.replace('&&', ' and ').replace('||', ' or ').replace('<>', ' != ')
+  e = re.sub(r'!([^=])', r' not \1', e)
+  e = re.sub(r'\{\s*(\w+\.u32)\s*,\s*(\w+\.u32)\s*\}', r'_pack32(\1, \2)', e)
+  def pack(m):
+    hi, lo = _expr(m[1].strip(), declared_arrays), _expr(m[2].strip(), declared_arrays)
+    return f'_pack({hi}, {lo})'
+  e = re.sub(r'\{\s*([^,{}]+)\s*,\s*([^,{}]+)\s*\}', pack, e)
+  e = re.sub(r"1201'B\(2\.0\s*/\s*PI\)", "TWO_OVER_PI_1201", e)
+  e = re.sub(r"\d+'([0-9a-fA-Fx]+)[UuFf]*", r'\1', e)
+  e = re.sub(r"\d+'[FIBU]\(", "(", e)
+  e = re.sub(r'\bB\(', '(', e)
+  e = re.sub(r'([0-9a-fA-Fx])ULL\b', r'\1', e)
+  e = re.sub(r'([0-9a-fA-Fx])LL\b', r'\1', e)
+  e = re.sub(r'([0-9a-fA-Fx])U\b', r'\1', e)
+  e = re.sub(r'(\d\.?\d*)F\b', r'\1', e)
+  e = re.sub(r'(\[laneId\])\.[uib]\d+', r'\1', e)
+  e = e.replace('+INF', 'INF').replace('-INF', '(-INF)')
+  e = re.sub(r'NAN\.f\d+', 'float("nan")', e)
+  def convert_verilog_slice(m):
+    start, width = m.group(1).strip(), m.group(2).strip()
+    return f'[({start}) + ({width}) - 1 : ({start})]'
+  e = re.sub(r'\[([^:\[\]]+)\s*\+:\s*([^:\[\]]+)\]', convert_verilog_slice, e)
+  def process_brackets(s):
+    result, i = [], 0
+    while i < len(s):
+      if s[i] == '[':
+        depth, start = 1, i + 1
+        j = start
+        while j < len(s) and depth > 0:
+          if s[j] == '[': depth += 1
+          elif s[j] == ']': depth -= 1
+          j += 1
+        inner = _expr(s[start:j-1], declared_arrays)
+        result.append('[' + inner + ']')
+        i = j
+      else:
+        result.append(s[i])
+        i += 1
+    return ''.join(result)
+  e = process_brackets(e)
+  while '?' in e:
+    depth, bracket, q = 0, 0, -1
+    for i, c in enumerate(e):
+      if c == '(': depth += 1
+      elif c == ')': depth -= 1
+      elif c == '[': bracket += 1
+      elif c == ']': bracket -= 1
+      elif c == '?' and depth == 0 and bracket == 0: q = i; break
+    if q < 0: break
+    depth, bracket, col = 0, 0, -1
+    for i in range(q + 1, len(e)):
+      if e[i] == '(': depth += 1
+      elif e[i] == ')': depth -= 1
+      elif e[i] == '[': bracket += 1
+      elif e[i] == ']': bracket -= 1
+      elif e[i] == ':' and depth == 0 and bracket == 0: col = i; break
+    if col < 0: break
+    cond, t, f = e[:q].strip(), e[q+1:col].strip(), e[col+1:].strip()
+    e = f'(({t}) if ({cond}) else ({f}))'
+  return e
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# PDF PARSING WITH PAGE CACHING
+# ═══════════════════════════════════════════════════════════════════════════════
+
+class CachedPDF:
+  """PDF wrapper with page text/table caching for faster repeated access."""
+  def __init__(self, pdf):
+    self._pdf, self._text_cache, self._table_cache = pdf, {}, {}
+  def __len__(self): return len(self._pdf.pages)
+  def text(self, i):
+    if i not in self._text_cache: self._text_cache[i] = self._pdf.pages[i].extract_text() or ''
+    return self._text_cache[i]
+  def tables(self, i):
+    if i not in self._table_cache: self._table_cache[i] = [t.extract() for t in self._pdf.pages[i].find_tables()]
+    return self._table_cache[i]
+
+def _parse_bits(s: str) -> tuple[int, int] | None:
+  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]:
+  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]:
+      if m := re.search(r"(?:'b|Must be:\s*)([01_]+)", row[2]):
+        enc_bits = m.group(1).replace('_', '')
+        enc_val, declared_width, actual_width = int(enc_bits, 2), 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):
+  """Parse a single PDF and return (formats, enums, src_enum, doc_name, instructions)."""
+  import pdfplumber
+  from tinygrad.helpers import fetch
+
+  pdf = CachedPDF(pdfplumber.open(fetch(url)))
+  total_pages = len(pdf)
+
+  # Auto-detect document type
+  first_page = pdf.text(0)
+  is_cdna4, is_cdna3 = 'CDNA4' in first_page or 'CDNA 4' in first_page, 'CDNA3' in first_page or 'MI300' in first_page
+  is_cdna, is_rdna4 = is_cdna3 or is_cdna4, 'RDNA4' in first_page or 'RDNA 4' in first_page
+  is_rdna35, is_rdna3 = 'RDNA3.5' in first_page or 'RDNA 3.5' in first_page, 'RDNA3' in first_page and 'RDNA3.5' not in first_page
+  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 Microcode Formats section (for formats/enums)
+  microcode_start = next((i for i in range(int(total_pages * 0.2), total_pages)
+                          if re.search(r'\d+\.\d+\.\d+\.\s+SOP2\b|Chapter \d+\.\s+Microcode Formats', pdf.text(i))), int(total_pages * 0.9))
+  # Find Instructions section (for pseudocode)
+  instr_start = next((i for i in range(int(total_pages * 0.1), int(total_pages * 0.5))
+                      if re.search(r'Chapter \d+\.\s+Instructions\b', pdf.text(i))), total_pages // 3)
+  instr_end = next((i for start in [int(total_pages * 0.6), int(total_pages * 0.5), instr_start]
+                    for i in range(start, min(start + 100, total_pages))
+                    if re.search(r'Chapter \d+\.\s+Microcode Formats', pdf.text(i))), total_pages)
+
+  # Parse src enum from SSRC encoding table
+  src_enum = dict(SRC_EXTRAS)
+  for i in range(microcode_start, min(microcode_start + 10, total_pages)):
+    text = pdf.text(i)
+    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_]*$', name): src_enum[val] = name
+      break
+
+  # Parse opcode tables
+  full_text = '\n'.join(pdf.text(i) for i in range(microcode_start, min(microcode_start + 50, total_pages)))
+  enums: dict[str, dict[int, str]] = {}
+  for m in re.finditer(r'Table \d+\. (\w+) Opcodes(.*?)(?=Table \d+\.|\n\d+\.\d+\.\d+\.\s+\w+\s*\nDescription|$)', 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())
+
+  # Parse instruction formats
+  def is_fields_table(t): return t and len(t) > 1 and t[0] and 'Field' in str(t[0][0] or '')
+  def has_encoding(fields): return any(f[0] == 'ENCODING' for f in fields)
+  def has_header_before_fields(text): return (pos := text.find('Field Name')) != -1 and bool(re.search(r'\d+\.\d+\.\d+\.\s+\w+\s*\n', text[:pos]))
+
+  format_headers = []
+  for i in range(50):
+    if microcode_start + i >= total_pages: break
+    text = pdf.text(microcode_start + i)
+    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 < 50:
+        next_text = pdf.text(microcode_start + 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()))
+
+  formats: dict[str, list] = {}
+  for fmt_name, rel_idx, header_pos in format_headers:
+    if fmt_name in formats: continue
+    page_idx = microcode_start + rel_idx
+    text = pdf.text(page_idx)
+    field_pos = text.find('Field Name', header_pos)
+    fields = None
+    for offset in range(3):
+      if page_idx + offset >= total_pages: break
+      if offset > 0 and has_header_before_fields(pdf.text(page_idx + offset)): break
+      for t in pdf.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 pdf.tables(page_idx):
+        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 >= total_pages or has_header_before_fields(pdf.text(page_idx + pg_offset)): break
+      for t in pdf.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']]
+  if doc_name in ('RDNA3', 'RDNA3.5'):
+    if 'SOPPOp' in enums: assert 8 not in enums['SOPPOp']; enums['SOPPOp'][8] = 'S_WAITCNT_DEPCTR'
+    if 'DSOp' in enums:
+      for k, v in {24: 'DS_GWS_SEMA_RELEASE_ALL', 25: 'DS_GWS_INIT', 26: 'DS_GWS_SEMA_V', 27: 'DS_GWS_SEMA_BR', 28: 'DS_GWS_SEMA_P', 29: 'DS_GWS_BARRIER'}.items():
+        assert k not in enums['DSOp']; enums['DSOp'][k] = v
+    if 'FLATOp' in enums:
+      for k, v in {40: 'GLOBAL_LOAD_ADDTID_B32', 41: 'GLOBAL_STORE_ADDTID_B32', 55: 'FLAT_ATOMIC_CSUB_U32'}.items():
+        assert k not in enums['FLATOp']; enums['FLATOp'][k] = v
+
+  # Extract pseudocode for instructions
+  all_text = '\n'.join(pdf.text(i) for i in range(instr_start, instr_end))
+  matches = list(INST_PATTERN.finditer(all_text))
+  raw_pseudocode: dict[tuple[str, int], str] = {}
+  for i, match in enumerate(matches):
+    name, opcode = match.group(1), int(match.group(2))
+    start, end = match.end(), matches[i + 1].start() if i + 1 < len(matches) else match.end() + 2000
+    snippet = all_text[start:end].strip()
+    if pseudocode := _extract_pseudocode(snippet): raw_pseudocode[(name, opcode)] = pseudocode
+
+  return {"formats": formats, "enums": enums, "src_enum": src_enum, "doc_name": doc_name, "pseudocode": raw_pseudocode, "is_cdna": is_cdna}
+
+def _extract_pseudocode(text: str) -> str | None:
+  """Extract pseudocode from an instruction description snippet."""
+  lines, result, depth, in_lambda = text.split('\n'), [], 0, 0
+  for line in lines:
+    s = line.strip()
+    if not s or re.match(r'^\d+ of \d+$', s) or re.match(r'^\d+\.\d+\..*Instructions', s): continue
+    if s.startswith(('Notes', 'Functional examples')): break
+    if s.startswith(('"RDNA', 'AMD ', 'CDNA')): continue
+    if '= lambda(' in s: in_lambda += 1; continue
+    if in_lambda > 0:
+      if s.endswith(');'): in_lambda -= 1
+      continue
+    if s.startswith('if '): depth += 1
+    elif s.startswith('endif'): depth = max(0, depth - 1)
+    if s.endswith('.') and not any(p in s for p in ['D0', 'D1', 'S0', 'S1', 'S2', 'SCC', 'VCC', 'tmp', '=']): continue
+    if re.match(r'^[a-z].*\.$', s) and '=' not in s: continue
+    is_code = (any(p in s for p in ['D0.', 'D1.', 'S0.', 'S1.', 'S2.', 'SCC =', 'SCC ?', 'VCC', 'EXEC', 'tmp =', 'tmp[', 'lane =', 'PC =',
+                                    'D0[', 'D1[', 'S0[', 'S1[', 'S2[', 'MEM[', 'RETURN_DATA', 'DATA.', 'DATA0', 'DATA1', 'ADDR']) or
+               s.startswith(('if ', 'else', 'elsif', 'endif', 'declare ', 'for ', 'endfor', '//')) or
+               re.match(r'^[a-z_]+\s*=', s) or re.match(r'^[a-z_]+\[', s) or (depth > 0 and '=' in s))
+    if is_code: result.append(s)
+  return '\n'.join(result) if result else None
+
+def _merge_results(results: list[dict]) -> dict:
+  """Merge multiple PDF parse results into a superset."""
+  merged = {"formats": {}, "enums": {}, "src_enum": dict(SRC_EXTRAS), "doc_names": [], "pseudocode": {}, "is_cdna": False}
+  for r in results:
+    merged["doc_names"].append(r["doc_name"])
+    merged["is_cdna"] = merged["is_cdna"] or r["is_cdna"]
+    for val, name in r["src_enum"].items():
+      if val in merged["src_enum"]: assert merged["src_enum"][val] == name
+      else: merged["src_enum"][val] = name
+    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
+        else: merged["enums"][enum_name][val] = 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]}
+        for f in fields:
+          if f[0] in existing: assert existing[f[0]] == (f[1], f[2])
+          else: merged["formats"][fmt_name].append(f)
+    for key, pc in r["pseudocode"].items():
+      if key not in merged["pseudocode"]: merged["pseudocode"][key] = pc
+  return merged
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# CODE GENERATION
+# ═══════════════════════════════════════════════════════════════════════════════
+
+def _generate_enum_py(enums, src_enum, doc_name) -> str:
+  """Generate enum.py content (just enums, no dsl.py dependency)."""
+  def enum_lines(name, items): return [f"class {name}(IntEnum):"] + [f"  {n} = {v}" for v, n in sorted(items.items())] + [""]
+  lines = [f"# autogenerated from AMD {doc_name} ISA PDF by pdf.py - do not edit", "from enum import IntEnum", ""]
+  lines += enum_lines("SrcEnum", src_enum) + sum([enum_lines(n, ops) for n, ops in sorted(enums.items())], [])
+  return '\n'.join(lines)
+
+def _generate_ins_py(formats, enums, src_enum, doc_name) -> str:
+  """Generate ins.py content (instruction formats and helpers, imports dsl.py and enum.py)."""
+  def field_key(f, order): return order.index(f[0].lower()) if f[0].lower() in order else 1000
+  lines = [f"# autogenerated from AMD {doc_name} ISA PDF by pdf.py - do not edit",
+           "# ruff: noqa: F401,F403", "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",
+           "from extra.assembly.amd.autogen.{arch}.enum import *",
+           "import functools", ""]
+  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):
+      lines.append(f"  encoding = bits[{enc[1]}:{enc[2]}] == 0b{enc[3]:b}" if enc[1] != enc[2] else f"  encoding = bits[{enc[1]}] == {enc[3]}")
+    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=lambda f: field_key(f, order)):
+      ann = f":Annotated[BitField, {ftype}]" if ftype and ftype.endswith('Op') else 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=1"}.get(fmt, "")
+      tgt = {"GLOBAL": "FLAT, GLOBALOp", "SCRATCH": "FLAT, SCRATCHOp"}.get(fmt, f"{fmt}, {cls_name}")
+      if fmt in formats or fmt in ("GLOBAL", "SCRATCH"):
+        suffix = "_e32" if fmt in ("VOP1", "VOP2", "VOPC") else "_e64" if fmt == "VOP3" and op_val < 512 else ""
+        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})")
+  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 {'DPP8', 'DPP16'}]
+  if "NULL" in src_names: lines.append("OFF = NULL\n")
+  return '\n'.join(lines)
+
+def _generate_gen_pcode_py(enums, pseudocode, arch) -> str:
+  """Generate gen_pcode.py content (compiled pseudocode functions)."""
+  # Get op enums for this arch (import from .ins which re-exports from .enum)
+  import importlib
+  autogen = importlib.import_module(f"extra.assembly.amd.autogen.{arch}.ins")
+  OP_ENUMS = [getattr(autogen, name) for name in ['SOP1Op', 'SOP2Op', 'SOPCOp', 'SOPKOp', 'SOPPOp', 'VOP1Op', 'VOP2Op', 'VOP3Op', 'VOP3SDOp', 'VOP3POp', 'VOPCOp', 'VOP3AOp', 'VOP3BOp', 'DSOp'] if hasattr(autogen, name)]
+
+  # Build defined ops mapping
+  defined_ops: dict[tuple, list] = {}
+  for enum_cls in OP_ENUMS:
+    for op in enum_cls:
+      if op.name.startswith(('S_', 'V_', 'DS_')): defined_ops.setdefault((op.name, op.value), []).append((enum_cls, op))
+
+  enum_names = [e.__name__ for e in OP_ENUMS]
+  lines = [f'''# autogenerated by pdf.py - do not edit
+# to regenerate: python -m extra.assembly.amd.pdf --arch {arch}
+# ruff: noqa: E501,F405,F403
+# mypy: ignore-errors
+from extra.assembly.amd.autogen.{arch}.enum import {", ".join(enum_names)}
+from extra.assembly.amd.pcode import *
+''']
+
+  instructions: dict = {cls: {} for cls in OP_ENUMS}
+  for key, pc in pseudocode.items():
+    if key in defined_ops:
+      for enum_cls, enum_val in defined_ops[key]: instructions[enum_cls][enum_val] = pc
+
+  for enum_cls in OP_ENUMS:
+    cls_name = enum_cls.__name__
+    if not instructions.get(enum_cls): continue
+    fn_entries = []
+    for op, pc in instructions[enum_cls].items():
+      if any(p in pc for p in UNSUPPORTED): continue
+      try:
+        code = compile_pseudocode(pc)
+        code = _apply_pseudocode_fixes(op, code)
+        fn_name, fn_code = _generate_function(cls_name, op, pc, code)
+        lines.append(fn_code)
+        fn_entries.append((op, fn_name))
+      except Exception as e: print(f"  Warning: Failed to compile {op.name}: {e}")
+    if fn_entries:
+      lines.append(f'{cls_name}_FUNCTIONS = {{')
+      for op, fn_name in fn_entries: lines.append(f"  {cls_name}.{op.name}: {fn_name},")
+      lines.append('}\n')
+
+  # Add V_WRITELANE_B32 if VOP3Op exists
+  if 'VOP3Op' in enum_names:
+    lines.append('''
+# V_WRITELANE_B32: Write scalar to specific lane's VGPR (not in PDF pseudocode)
+def _VOP3Op_V_WRITELANE_B32(s0, s1, s2, d0, scc, vcc, lane, exec_mask, literal, VGPR, _vars, src0_idx=0, vdst_idx=0):
+  wr_lane = s1 & 0x1f
+  return {'d0': d0, 'scc': scc, 'vgpr_write': (wr_lane, vdst_idx, s0 & 0xffffffff)}
+VOP3Op_FUNCTIONS[VOP3Op.V_WRITELANE_B32] = _VOP3Op_V_WRITELANE_B32
+''')
+
+  lines.append('COMPILED_FUNCTIONS = {')
+  for enum_cls in OP_ENUMS:
+    if instructions.get(enum_cls): lines.append(f'  {enum_cls.__name__}: {enum_cls.__name__}_FUNCTIONS,')
+  lines.append('}\n\ndef get_compiled_functions(): return COMPILED_FUNCTIONS')
+  return '\n'.join(lines)
+
+def _apply_pseudocode_fixes(op, code: str) -> str:
+  """Apply known fixes for PDF pseudocode bugs."""
+  if op.name == 'V_DIV_FMAS_F32':
+    code = code.replace('D0.f32 = 2.0 ** 32 * fma(S0.f32, S1.f32, S2.f32)',
+                        'D0.f32 = (2.0 ** 64 if exponent(S2.f32) > 127 else 2.0 ** -64) * fma(S0.f32, S1.f32, S2.f32)')
+  if op.name == 'V_DIV_FMAS_F64':
+    code = code.replace('D0.f64 = 2.0 ** 64 * fma(S0.f64, S1.f64, S2.f64)',
+                        'D0.f64 = (2.0 ** 128 if exponent(S2.f64) > 1023 else 2.0 ** -128) * fma(S0.f64, S1.f64, S2.f64)')
+  if op.name == 'V_DIV_SCALE_F32':
+    code = code.replace('D0.f32 = float("nan")', 'VCC = Reg(0x1); D0.f32 = float("nan")')
+    code = code.replace('elif S1.f32 == DENORM.f32:\n  D0.f32 = ldexp(S0.f32, 64)', 'elif False:\n  pass')
+    code += '\nif S1.f32 == DENORM.f32:\n  D0.f32 = float("nan")'
+    code = code.replace('elif exponent(S2.f32) <= 23:\n  D0.f32 = ldexp(S0.f32, 64)', 'elif exponent(S2.f32) <= 23:\n  VCC = Reg(0x1); D0.f32 = ldexp(S0.f32, 64)')
+    code = code.replace('elif S2.f32 / S1.f32 == DENORM.f32:\n  VCC = Reg(0x1)\n  if S0.f32 == S2.f32:\n    D0.f32 = ldexp(S0.f32, 64)', 'elif S2.f32 / S1.f32 == DENORM.f32:\n  VCC = Reg(0x1)')
+  if op.name == 'V_DIV_SCALE_F64':
+    code = code.replace('D0.f64 = float("nan")', 'VCC = Reg(0x1); D0.f64 = float("nan")')
+    code = code.replace('elif S1.f64 == DENORM.f64:\n  D0.f64 = ldexp(S0.f64, 128)', 'elif False:\n  pass')
+    code += '\nif S1.f64 == DENORM.f64:\n  D0.f64 = float("nan")'
+    code = code.replace('elif exponent(S2.f64) <= 52:\n  D0.f64 = ldexp(S0.f64, 128)', 'elif exponent(S2.f64) <= 52:\n  VCC = Reg(0x1); D0.f64 = ldexp(S0.f64, 128)')
+    code = code.replace('elif S2.f64 / S1.f64 == DENORM.f64:\n  VCC = Reg(0x1)\n  if S0.f64 == S2.f64:\n    D0.f64 = ldexp(S0.f64, 128)', 'elif S2.f64 / S1.f64 == DENORM.f64:\n  VCC = Reg(0x1)')
+  if op.name == 'V_DIV_FIXUP_F32':
+    code = code.replace('D0.f32 = ((-abs(S0.f32)) if (sign_out) else (abs(S0.f32)))',
+                        'D0.f32 = ((-OVERFLOW_F32) if (sign_out) else (OVERFLOW_F32)) if isNAN(S0.f32) else ((-abs(S0.f32)) if (sign_out) else (abs(S0.f32)))')
+  if op.name == 'V_DIV_FIXUP_F64':
+    code = code.replace('D0.f64 = ((-abs(S0.f64)) if (sign_out) else (abs(S0.f64)))',
+                        'D0.f64 = ((-OVERFLOW_F64) if (sign_out) else (OVERFLOW_F64)) if isNAN(S0.f64) else ((-abs(S0.f64)) if (sign_out) else (abs(S0.f64)))')
+  if op.name == 'V_TRIG_PREOP_F64':
+    code = code.replace('result = F((TWO_OVER_PI_1201[1200 : 0] << shift.u32) & 0x1fffffffffffff)',
+                        'result = float(((TWO_OVER_PI_1201[1200 : 0] << int(shift)) >> (1201 - 53)) & 0x1fffffffffffff)')
+  return code
+
+def _generate_function(cls_name: str, op, pc: str, code: str) -> tuple[str, str]:
+  """Generate a single compiled pseudocode function."""
+  has_d1 = '{ D1' in pc
+  is_cmpx = (cls_name in ('VOPCOp', 'VOP3Op')) and 'EXEC.u64[laneId]' in pc
+  is_div_scale = 'DIV_SCALE' in op.name
+  has_sdst = cls_name == 'VOP3SDOp' and ('VCC.u64[laneId]' in pc or is_div_scale)
+  has_opsel = 'OPSEL' in pc  # FMA_MIX and similar instructions need OPSEL/OPSEL_HI
+  combined = code + pc
+
+  fn_name = f"_{cls_name}_{op.name}"
+  # Function accepts Reg objects directly (uppercase names), laneId is passed directly as int
+  params = "S0, S1, S2, D0, SCC, VCC, laneId, EXEC, literal, VGPR, src0_idx=0, vdst_idx=0, PC=None"
+  if has_opsel: params += ", OPSEL=0, OPSEL_HI=0"
+  lines = [f"def {fn_name}({params}):"]
+
+  # Registers that need special handling (not passed directly)
+  # Only init if used but not first assigned as `name = Reg(...)` in the compiled code
+  def needs_init(name): return name in combined and not re.search(rf'^\s*{name}\s*=\s*Reg\(', code, re.MULTILINE)
+  special_regs = [('D1', 'Reg(0)'), ('SIMM16', 'Reg(literal)'), ('SIMM32', 'Reg(literal)'),
+                  ('SRC0', 'Reg(src0_idx)'), ('VDST', 'Reg(vdst_idx)')]
+  if needs_init('tmp'): special_regs.insert(0, ('tmp', 'Reg(0)'))
+  if needs_init('saveexec'): special_regs.insert(0, ('saveexec', 'Reg(EXEC._val)'))
+  used = {name for name, _ in special_regs if name in combined}
+
+  # Detect which registers are modified (not just read) - look for assignments
+  modifies_d0 = is_div_scale or bool(re.search(r'\bD0\b[.\[]', combined))
+  modifies_exec = is_cmpx or bool(re.search(r'EXEC\.(u32|u64|b32|b64)\s*=', combined))
+  modifies_vcc = has_sdst or bool(re.search(r'VCC\.(u32|u64|b32|b64)\s*=|VCC\.u64\[laneId\]\s*=', combined))
+  modifies_scc = bool(re.search(r'\bSCC\s*=', combined))
+  modifies_pc = bool(re.search(r'\bPC\s*=', combined))
+
+  # Build init code for special registers
+  init_lines = []
+  if is_div_scale: init_lines.append("  D0 = Reg(S0._val)")
+  for name, init in special_regs:
+    if name in used: init_lines.append(f"  {name} = {init}")
+  if 'EXEC_LO' in code: init_lines.append("  EXEC_LO = SliceProxy(EXEC, 31, 0)")
+  if 'EXEC_HI' in code: init_lines.append("  EXEC_HI = SliceProxy(EXEC, 63, 32)")
+  if 'VCCZ' in code and not re.search(r'^\s*VCCZ\s*=', code, re.MULTILINE): init_lines.append("  VCCZ = Reg(1 if VCC._val == 0 else 0)")
+  if 'EXECZ' in code and not re.search(r'^\s*EXECZ\s*=', code, re.MULTILINE): init_lines.append("  EXECZ = Reg(1 if EXEC._val == 0 else 0)")
+  code_lines = [line for line in code.split('\n') if line.strip()]
+  if init_lines:
+    lines.extend(init_lines)
+    if code_lines: lines.append("  # --- compiled pseudocode ---")
+  for line in code_lines:
+    lines.append(f"  {line}")
+
+  # Build result dict - only include registers that are modified
+  result_items = []
+  if modifies_d0: result_items.append("'D0': D0")
+  if modifies_scc: result_items.append("'SCC': SCC")
+  if modifies_vcc: result_items.append("'VCC': VCC")
+  if modifies_exec: result_items.append("'EXEC': EXEC")
+  if has_d1: result_items.append("'D1': D1")
+  if modifies_pc: result_items.append("'PC': PC")
+  lines.append(f"  return {{{', '.join(result_items)}}}\n")
+  return fn_name, '\n'.join(lines)
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# MAIN GENERATION
+# ═══════════════════════════════════════════════════════════════════════════════
+
+def generate_arch(arch: str) -> dict:
+  """Generate enum.py, ins.py and gen_pcode.py for a single architecture."""
+  urls = PDF_URLS[arch]
+  if isinstance(urls, str): urls = [urls]
+
+  print(f"\n{'='*60}\nGenerating {arch}...")
+  print(f"Parsing {len(urls)} PDF(s)...")
+  results = [_parse_single_pdf(url) for url in urls]
+  merged = _merge_results(results) if len(results) > 1 else results[0]
+  doc_name = "+".join(merged["doc_names"]) if len(results) > 1 else merged["doc_name"]
+
+  base_path = Path(f"extra/assembly/amd/autogen/{arch}")
+  base_path.mkdir(parents=True, exist_ok=True)
+  (base_path / "__init__.py").touch()
+
+  # Write enum.py (enums only, no dsl.py dependency)
+  enum_path = base_path / "enum.py"
+  enum_content = _generate_enum_py(merged["enums"], merged["src_enum"], doc_name)
+  enum_path.write_text(enum_content)
+  print(f"Generated {enum_path}: SrcEnum ({len(merged['src_enum'])}) + {len(merged['enums'])} enums")
+
+  # Write ins.py (instruction formats and helpers, imports dsl.py and enum.py)
+  ins_path = base_path / "ins.py"
+  ins_content = _generate_ins_py(merged["formats"], merged["enums"], merged["src_enum"], doc_name).replace("{arch}", arch)
+  ins_path.write_text(ins_content)
+  print(f"Generated {ins_path}: {len(merged['formats'])} formats")
+
+  # Write gen_pcode.py (needs enum.py to exist first for imports)
+  pcode_path = base_path / "gen_pcode.py"
+  pcode_content = _generate_gen_pcode_py(merged["enums"], merged["pseudocode"], arch)
+  pcode_path.write_text(pcode_content)
+  print(f"Generated {pcode_path}: {len(merged['pseudocode'])} instructions")
+
+  return merged
+
+def _generate_arch_wrapper(arch: str):
+  """Wrapper for multiprocessing - returns arch name for ordering."""
+  generate_arch(arch)
+  return arch
+
+def generate_all():
+  """Generate all architectures in parallel."""
+  with ProcessPoolExecutor() as executor:
+    list(executor.map(_generate_arch_wrapper, PDF_URLS.keys()))
+
+if __name__ == "__main__":
+  import argparse
+  parser = argparse.ArgumentParser(description="Generate AMD ISA autogen files from PDF documentation")
+  parser.add_argument("--arch", choices=list(PDF_URLS.keys()) + ["all"], default="rdna3")
+  args = parser.parse_args()
+  if args.arch == "all": generate_all()
+  else: generate_arch(args.arch)

extra/assembly/amd/test/bench_emu.py

@@ -0,0 +1,294 @@
+#!/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

@@ -0,0 +1,196 @@
+# 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.ins 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

@@ -0,0 +1,66 @@
+"""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")
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# EXECUTION CONTEXT (for testing compiled pseudocode)
+# ═══════════════════════════════════════════════════════════════════════════════
+
+class ExecContext:
+  """Context for running compiled pseudocode in tests."""
+  def __init__(self, s0=0, s1=0, s2=0, d0=0, scc=0, vcc=0, lane=0, exec_mask=0xffffffff, literal=0, vgprs=None, src0_idx=0, vdst_idx=0):
+    from extra.assembly.amd.pcode import Reg, MASK32, MASK64, SliceProxy
+    self._Reg, self._MASK64, self._SliceProxy = Reg, MASK64, SliceProxy
+    self.S0, self.S1, self.S2 = Reg(s0), Reg(s1), Reg(s2)
+    self.D0, self.D1 = Reg(d0), Reg(0)
+    self.SCC, self.VCC, self.EXEC = Reg(scc), Reg(vcc), Reg(exec_mask)
+    self.tmp, self.saveexec = Reg(0), Reg(exec_mask)
+    self.lane, self.laneId, self.literal = lane, lane, literal
+    self.SIMM16, self.SIMM32 = Reg(literal), Reg(literal)
+    self.VGPR = vgprs if vgprs is not None else {}
+    self.SRC0, self.VDST = Reg(src0_idx), Reg(vdst_idx)
+
+  def run(self, code: str):
+    """Execute compiled code."""
+    import extra.assembly.amd.pcode as pcode
+    ns = {k: getattr(pcode, k) for k in dir(pcode) if not k.startswith('_')}
+    # Also include underscore-prefixed helpers that compiled pseudocode uses
+    for k in ['_pack', '_pack32']:
+      if hasattr(pcode, k): ns[k] = getattr(pcode, k)
+    ns.update({
+      'S0': self.S0, 'S1': self.S1, 'S2': self.S2, 'D0': self.D0, 'D1': self.D1,
+      'SCC': self.SCC, 'VCC': self.VCC, 'EXEC': self.EXEC,
+      'EXEC_LO': self._SliceProxy(self.EXEC, 31, 0), 'EXEC_HI': self._SliceProxy(self.EXEC, 63, 32),
+      'tmp': self.tmp, 'saveexec': self.saveexec,
+      'lane': self.lane, 'laneId': self.laneId, 'literal': self.literal,
+      'SIMM16': self.SIMM16, 'SIMM32': self.SIMM32, 'VGPR': self.VGPR, 'SRC0': self.SRC0, 'VDST': self.VDST,
+    })
+    exec(code, ns)
+    def _sync(ctx_reg, ns_val):
+      if isinstance(ns_val, self._Reg): ctx_reg._val = ns_val._val
+      else: ctx_reg._val = int(ns_val) & self._MASK64
+    for name in ('SCC', 'VCC', 'EXEC', 'D0', 'D1', 'tmp', 'saveexec'):
+      if ns.get(name) is not getattr(self, name): _sync(getattr(self, name), ns[name])
+
+  def result(self) -> dict: return {"d0": self.D0._val, "scc": self.SCC._val & 1}

extra/assembly/amd/test/test_compare_emulators.py

@@ -1,16 +1,18 @@
 # Test to compare Python and Rust RDNA3 emulators by running real tinygrad kernels
-import unittest, ctypes
+import unittest, ctypes, os
 from dataclasses import dataclass
 from pathlib import Path
-from tinygrad import Device
 
-from test.mockgpu.amd.emu import WaveState, _decode_at, WAVE_SIZE, VCC_LO, EXEC_LO, SCC
-from tinygrad.renderer.amd import decode_inst
-import tinygrad
-REMU_PATH = Path(tinygrad.__file__).parent.parent / "extra/remu/target/release/libremu.so"
-if not REMU_PATH.exists(): REMU_PATH = Path(tinygrad.__file__).parent.parent / "extra/remu/target/release/libremu.dylib"
+# Set environment before any tinygrad imports to use MOCKGPU
+# This allows generating AMD GPU kernels without requiring real hardware
+os.environ["AMD"] = "1"
+os.environ["MOCKGPU"] = "1"
+os.environ["PYTHON_REMU"] = "1"
 
-def set_valid_mem_ranges(ranges): pass  # emu2 doesn't need this
+from extra.assembly.amd.emu import WaveState, decode_program, step_wave, WAVE_SIZE, set_valid_mem_ranges
+from extra.assembly.amd.test.helpers import KernelInfo
+
+REMU_PATH = Path(__file__).parents[3] / "remu/target/release/libremu.so"
 
 def _is_f32_nan(bits: int) -> bool:
   """Check if 32-bit value is a NaN (exponent all 1s, mantissa non-zero)."""
@@ -21,15 +23,6 @@ def _vals_equal(a: int, b: int) -> bool:
   if a == b: return True
   return _is_f32_nan(a) and _is_f32_nan(b)
 
-@dataclass
-class KernelSnapshot:
-  code: bytes
-  src: str
-  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
-
 @dataclass
 class StateSnapshot:
   pc: int
@@ -93,71 +86,43 @@ class RustEmulator:
     return snap.to_snapshot()
 
   def free(self):
-    if self.ctx:
-      self.lib.wave_free(self.ctx)
-      self.ctx = None
+    if self.ctx: self.lib.wave_free(self.ctx); self.ctx = None
 
 class PythonEmulator:
   def __init__(self):
     self.state: WaveState | None = None
-    self.program: dict[int, tuple] = {}  # lazily populated: pc -> (name, fxn, globals)
-    self.vmem_buf = None
-    self.lds_buf = None
-    self.kernel_buf = None  # Keep kernel bytes alive
-    self.lib_addr = 0  # Base address of kernel code
+    self.program: dict | None = None
+    self.lds: bytearray | None = None
+    self.n_lanes = 0
 
   def create(self, kernel: bytes, n_lanes: int):
-    import ctypes
-    from tinygrad.device import Buffer, BufferSpec
-    from tinygrad.dtype import dtypes
-    # Store kernel in a ctypes buffer so _decode_at can read from memory at actual PC address
-    self.kernel_buf = (ctypes.c_char * len(kernel)).from_buffer_copy(kernel)
-    self.lib_addr = ctypes.addressof(self.kernel_buf)
-    self.program = {}
-    self.state = WaveState(n_lanes)
-    self.state.pc = self.lib_addr  # Set PC to code base address
-    self.vmem_buf = Buffer('CPU', 1 << 40, dtypes.uint32, options=BufferSpec(external_ptr=0)).ensure_allocated()
-    self.lds_buf = Buffer('CPU', 65536 // 4, dtypes.uint32).ensure_allocated()
-
-  def _ensure_decoded(self, pc: int):
-    if pc not in self.program:
-      runner, _ = _decode_at(pc, "rdna3")
-      self.program[pc] = (runner.p.function_name, runner._prg.fxn, runner.p.globals)
+    self.program = decode_program(kernel)
+    self.state = WaveState()
+    self.state.exec_mask = (1 << n_lanes) - 1
+    self.lds = bytearray(65536)
+    self.n_lanes = n_lanes
 
   def step(self) -> int:
-    import ctypes
-    assert self.state is not None
-    pc = self.state.pc
-    if pc == 0xFFFFFFFFFFFFFFFF: return -1
-    self._ensure_decoded(pc)
-    name, fxn, globals_list = self.program[pc]
-    buf_addrs = {0: self.state.sgpr_buf._buf.va_addr, 1: self.state.vgpr_buf._buf.va_addr,  # type: ignore[union-attr]
-                 2: self.vmem_buf._buf.va_addr, 3: self.lds_buf._buf.va_addr}  # type: ignore[union-attr]
-    fxn(*[ctypes.c_uint64(buf_addrs[g]) for g in globals_list], ctypes.c_int32(0))
-    return -1 if self.state.pc == 0xFFFFFFFFFFFFFFFF else 0
-
+    assert self.program is not None and self.state is not None and self.lds is not None
+    return step_wave(self.program, self.state, self.lds, self.n_lanes)
   def set_sgpr(self, idx: int, val: int):
     assert self.state is not None
-    self.state._write_sgpr(idx, val)
+    self.state.sgpr[idx] = val & 0xffffffff
   def set_vgpr(self, lane: int, idx: int, val: int):
     assert self.state is not None
-    self.state._write_vgpr(idx, lane, val)
+    self.state.vgpr[lane][idx] = val & 0xffffffff
 
   def get_snapshot(self) -> StateSnapshot:
     assert self.state is not None
-    sgpr = [self.state._read_sgpr(i) for i in range(128)]
-    vgpr = [[self.state._read_vgpr(reg, lane) for reg in range(256)] for lane in range(WAVE_SIZE)]
-    # Convert actual PC address to word offset for comparison with Rust emulator
-    pc_offset = (self.state.pc - self.lib_addr) // 4 if self.state.pc != 0xFFFFFFFFFFFFFFFF else 0xFFFFFFFFFFFFFFFF
-    return StateSnapshot(pc=pc_offset, scc=self.state._read_sgpr(SCC.offset), vcc=sgpr[VCC_LO.offset],
-                         exec_mask=sgpr[EXEC_LO.offset], sgpr=sgpr, vgpr=vgpr)
+    return StateSnapshot(pc=self.state.pc, scc=self.state.scc, vcc=self.state.vcc & 0xffffffff,
+                         exec_mask=self.state.exec_mask & 0xffffffff, sgpr=list(self.state.sgpr),
+                         vgpr=[list(self.state.vgpr[i]) for i in range(WAVE_SIZE)])
 
 def run_single_kernel(kernel: bytes, n_lanes: int, args_ptr: int, global_size: tuple[int, int, int],
-                      local_size: tuple[int, int, int], max_steps: int, debug: bool, trace_len: int,
-                      kernel_idx: int = 0, max_workgroups: int = 8) -> tuple[bool, str, int]:
+                      program, max_steps: int, debug: bool, trace_len: int, kernel_idx: int = 0,
+                      max_workgroups: int = 8) -> tuple[bool, str, int]:
   """Run a single kernel through both emulators. Returns (success, message, total_steps)."""
   gx, gy, gz = global_size
-  lx, ly, lz = local_size
   total_steps = 0
   wg_count = 0
 
@@ -180,53 +145,27 @@ def run_single_kernel(kernel: bytes, n_lanes: int, args_ptr: int, global_size: t
           emu.set_sgpr(13, gidx)
           emu.set_sgpr(14, gidy)
           emu.set_sgpr(15, gidz)
-          # Initialize v[0] with packed workitem IDs for each lane
-          for lane in range(n_lanes):
-            tid = lane
-            z, y, x = tid // (lx * ly), (tid // lx) % ly, tid % lx
-            emu.set_vgpr(lane, 0, (z << 20) | (y << 10) | x)
 
         step = 0
         trace: list[tuple[int, int, str, StateSnapshot, StateSnapshot]] = []
-        prev_sync_after = False  # Track if previous instruction had known Rust bugs
         try:
           while step < max_steps:
             rust_before = rust.get_snapshot()
             python_before = python.get_snapshot()
 
-            pc_addr = python.lib_addr + python_before.pc * 4  # Convert word offset to actual address
-            python._ensure_decoded(pc_addr)
-            inst_hex_name = python.program[pc_addr][0]
-            # Decode the instruction to get mnemonic for sync_after checks
-            try:
-              # Format is mnemonic_hexbytes, e.g. v_exp_f32_e32_014b027e -> hex is 014b027e
-              parts = inst_hex_name.rsplit('_', 1)
-              inst_bytes_hex = parts[1] if len(parts) == 2 else ""
-              inst_bytes = bytes.fromhex(inst_bytes_hex) if inst_bytes_hex else b''
-              decoded = decode_inst(inst_bytes) if inst_bytes else None
-              inst_mnemonic = repr(decoded).split('(')[0] if decoded else ""
-            except Exception:
-              inst_mnemonic = ""
-            # For generic instructions, use function name for sync_after check
-            if not inst_mnemonic: inst_mnemonic = inst_hex_name
-            inst_str = inst_hex_name
+            inst = program.get(python_before.pc)
+            inst_str = inst.disasm() if inst else f"unknown at PC={python_before.pc}"
             trace.append((step, python_before.pc, inst_str, rust_before, python_before))
             if len(trace) > trace_len: trace.pop(0)
 
             if debug: print(f"K{kernel_idx} WG({gidx},{gidy},{gidz}) Step {step}: PC={python_before.pc}, inst={inst_str}")
 
-            # Instructions with known Rust emulator bugs or precision differences - sync Python to Rust after execution
+            # Instructions with known Rust emulator bugs - sync Python to Rust after execution
             # v_div_scale/v_div_fixup: Rust has different VCC handling
             # v_cvt_f16_f32: Rust clears high 16 bits, but hardware (and Python) preserves them
-            # s_add_i32/s_sub_i32: Rust has incorrect SCC overflow detection
-            # v_exp_f32/v_log_f32/v_ldexp_f32: precision differences in transcendental functions
-            # s_delay_alu: Rust handles differently
-            # v_add_co_ci_u32/v_sub_co_ci_u32/v_subrev_co_ci_u32: Rust preserves inactive VCC bits, but hardware clears all bits
-            sync_after = any(x in inst_mnemonic.lower() for x in ('v_div_scale', 'v_div_fixup', 'v_cvt_f16_f32', 's_add_i32', 's_sub_i32',
-                                                                   'v_exp_f32', 'v_log_f32', 'v_ldexp_f32', 's_delay_alu',
-                                                                   'v_add_co_ci_u32', 'v_sub_co_ci_u32', 'v_subrev_co_ci_u32'))
-            # Skip comparison if previous instruction had known Rust bugs (states were synced but may still differ slightly)
-            diffs = rust_before.diff(python_before, n_lanes) if not prev_sync_after else []
+            sync_after = any(x in inst_str for x in ('v_div_scale_f32', 'v_div_scale_f64', 'v_div_fixup_f32', 'v_div_fixup_f64',
+                                                      'v_cvt_f16_f32'))
+            diffs = rust_before.diff(python_before, n_lanes)
             if diffs:
               trace_lines = []
               for idx, (s, pc, d, rb, pb) in enumerate(trace):
@@ -237,18 +176,16 @@ def run_single_kernel(kernel: bytes, n_lanes: int, args_ptr: int, global_size: t
                   python_diffs = pb.diff(next_pb, n_lanes, "->")
                   if rust_diffs: trace_lines.append(f"             rust:   {', '.join(rust_diffs[:5])}")
                   if python_diffs: trace_lines.append(f"             python: {', '.join(python_diffs[:5])}")
-                  elif rust_diffs: trace_lines.append("             python: (no changes)")
+                  elif rust_diffs: trace_lines.append(f"             python: (no changes)")
                 else:
                   # Last traced instruction - compare with current state
                   rust_diffs = rb.diff(rust_before, n_lanes, "->")
                   python_diffs = pb.diff(python_before, n_lanes, "->")
                   if rust_diffs: trace_lines.append(f"             rust:   {', '.join(rust_diffs[:5])}")
                   if python_diffs: trace_lines.append(f"             python: {', '.join(python_diffs[:5])}")
-                  elif rust_diffs: trace_lines.append("             python: (no changes)")
+                  elif rust_diffs: trace_lines.append(f"             python: (no changes)")
               trace_str = "\n".join(trace_lines)
-              msg = f"K{kernel_idx} WG({gidx},{gidy},{gidz}) Step {step} before inst '{inst_str}': states differ (rust vs python):\n  "
-              msg += "\n  ".join(diffs[:10]) + f"\n  Recent instructions:\n{trace_str}"
-              return False, msg, total_steps
+              return False, f"K{kernel_idx} WG({gidx},{gidy},{gidz}) Step {step} before inst '{inst_str}': states differ (rust vs python):\n  " + "\n  ".join(diffs[:10]) + f"\n  Recent instructions:\n{trace_str}", total_steps
 
             rust_result = rust.step()
             python_result = python.step()
@@ -258,9 +195,7 @@ def run_single_kernel(kernel: bytes, n_lanes: int, args_ptr: int, global_size: t
               if rust_result == 1 and python_result == 0:
                 raise unittest.SkipTest(f"Rust emulator doesn't support instruction: {inst_str}")
               trace_str = "\n".join(f"    step {s}: PC={pc:3d} {d}" for s, pc, d, _, _ in trace)
-              msg = (f"K{kernel_idx} WG({gidx},{gidy},{gidz}) Step {step}: different return codes: "
-                     f"rust={rust_result}, python={python_result}, inst={inst_str}\n  Recent instructions:\n{trace_str}")
-              return False, msg, total_steps
+              return False, f"K{kernel_idx} WG({gidx},{gidy},{gidz}) Step {step}: different return codes: rust={rust_result}, python={python_result}, inst={inst_str}\n  Recent instructions:\n{trace_str}", total_steps
 
             # Sync Python state to Rust after instructions with known Rust emulator differences
             if sync_after:
@@ -269,12 +204,7 @@ def run_single_kernel(kernel: bytes, n_lanes: int, args_ptr: int, global_size: t
               for lane in range(n_lanes):
                 for i in range(256): python.set_vgpr(lane, i, rust_after.vgpr[lane][i])
               assert python.state is not None
-              # Convert Rust's word-based PC to Python's actual address
-              python.state.pc = python.lib_addr + rust_after.pc * 4
-              python.state._write_sgpr(SCC.offset, rust_after.scc)
-              python.state._write_sgpr(VCC_LO.offset, rust_after.vcc)
-              python.state._write_sgpr(EXEC_LO.offset, rust_after.exec_mask)
-            prev_sync_after = sync_after
+              python.state.pc, python.state.scc, python.state.vcc, python.state.exec_mask = rust_after.pc, rust_after.scc, rust_after.vcc, rust_after.exec_mask
 
             if rust_result == -1:
               total_steps += step + 1
@@ -293,7 +223,7 @@ def run_single_kernel(kernel: bytes, n_lanes: int, args_ptr: int, global_size: t
 
   return True, f"Completed {gx*gy*gz} workgroups", total_steps
 
-def compare_emulators_multi_kernel(kernels: list[KernelSnapshot], buf_pool: dict[int, int], max_steps: int = 1000,
+def compare_emulators_multi_kernel(kernels: list[KernelInfo], buf_pool: dict[int, int], max_steps: int = 1000,
                                     debug: bool = False, trace_len: int = 10, buf_data: dict[int, bytes] | None = None) -> tuple[bool, str]:
   """Run all kernels through both emulators with shared buffer pool."""
   if buf_data is None: buf_data = {}
@@ -323,11 +253,12 @@ def compare_emulators_multi_kernel(kernels: list[KernelSnapshot], buf_pool: dict
     kernel_ranges = ranges | {(args_ptr, ctypes.sizeof(args))}
     set_valid_mem_ranges(kernel_ranges)
 
+    program = decode_program(kernel.code)
     n_lanes = kernel.local_size[0] * kernel.local_size[1] * kernel.local_size[2]
 
     ok, msg, steps = run_single_kernel(
       kernel.code, min(n_lanes, 32), args_ptr, kernel.global_size,
-      kernel.local_size, max_steps, debug, trace_len, ki
+      program, max_steps, debug, trace_len, ki
     )
     total_steps += steps
     if not ok:
@@ -353,11 +284,11 @@ def compare_emulators_with_memory(kernel: bytes, n_lanes: int, buf_sizes: list,
   ranges.add((args_ptr, ctypes.sizeof(args)))
   set_valid_mem_ranges(ranges)
 
-  # Legacy wrapper assumes local_size = (n_lanes, 1, 1)
-  ok, msg, _ = run_single_kernel(kernel, n_lanes, args_ptr, global_size, (n_lanes, 1, 1), max_steps, debug, trace_len)
+  program = decode_program(kernel)
+  ok, msg, _ = run_single_kernel(kernel, n_lanes, args_ptr, global_size, program, max_steps, debug, trace_len)
   return ok, msg
 
-def get_kernels_from_tinygrad(op_fn) -> tuple[list[KernelSnapshot], dict[int, int], dict[int, bytes]]:
+def get_kernels_from_tinygrad(op_fn) -> tuple[list[KernelInfo], dict[int, int], dict[int, bytes]]:
   """Compile a tinygrad operation and extract all kernels with their buffer mappings."""
   from tinygrad import Tensor
   from tinygrad.runtime.support.elf import elf_loader
@@ -395,9 +326,8 @@ def get_kernels_from_tinygrad(op_fn) -> tuple[list[KernelSnapshot], dict[int, in
                 buf_pool[buf_id] = b.nbytes
               buf_idxs.append(buf_id)
               buf_sizes.append(b.nbytes)
-            kernels.append(KernelSnapshot(
+            kernels.append(KernelInfo(
               code=bytes(sec.content),
-              src=lowered.prg.p.src,
               global_size=tuple(lowered.prg.p.global_size),
               local_size=tuple(lowered.prg.p.local_size),
               buf_idxs=buf_idxs,
@@ -412,7 +342,6 @@ def get_kernel_from_tinygrad(op_fn) -> tuple[bytes, tuple[int, int, int], tuple[
   k = kernels[-1]
   return k.code, k.global_size, k.local_size, k.buf_sizes
 
-@unittest.skipUnless(Device.DEFAULT == "AMD", "requires AMD device")
 class TestTinygradKernels(unittest.TestCase):
   """Compare emulators on real tinygrad-compiled kernels."""
 
@@ -449,8 +378,7 @@ class TestTinygradKernels(unittest.TestCase):
   def test_cast(self): self._test_kernel(lambda T: T.empty(32).half().float() + T.empty(32).int().float())
 
   # Pooling - regression for VCC wave32 mode
-  def test_pool2d(self):
-    self._test_kernel(lambda T: T.empty(1, 1, 8, 8).avg_pool2d(kernel_size=(4,4)) + T.empty(1, 1, 8, 8).max_pool2d(kernel_size=(4,4)))
+  def test_pool2d(self): self._test_kernel(lambda T: T.empty(1, 1, 8, 8).avg_pool2d(kernel_size=(4,4)) + T.empty(1, 1, 8, 8).max_pool2d(kernel_size=(4,4)))
 
   # Convolution
   def test_conv2d(self): self._test_kernel(lambda T: T.empty(1, 2, 8, 8).conv2d(T.empty(2, 2, 3, 3)), max_steps=50000)
@@ -462,7 +390,6 @@ class TestTinygradKernels(unittest.TestCase):
     from tinygrad import dtypes
     self._test_kernel(lambda T: T.empty(4, 4)[T.arange(4).cast(dtypes.int64), :])
   def test_gelu(self): self._test_kernel(lambda T: T.empty(32, 32).gelu())
-  def test_exp(self): self._test_kernel(lambda T: T.empty(1024).exp())
   def test_cross_entropy(self):
     import numpy as np
     np.random.seed(0)
@@ -470,59 +397,6 @@ class TestTinygradKernels(unittest.TestCase):
     x_np = np.random.randn(16, 10).astype(np.float32)
     self._test_kernel(lambda T: (T(x_np.tolist()).reshape(16,10) + 0).cross_entropy((T(classes).int().reshape(16) + 0)))
   def test_isinf(self): self._test_kernel(lambda T: T([float('-inf'), 0., float('inf'), 1.1]*8).isinf())
-  def test_sin_f64(self):
-    from tinygrad import dtypes
-    self._test_kernel(lambda T: T([2.0], dtype=dtypes.float64).sin())
-
-  def test_sin_large_f32(self):
-    """Test sin with large values that trigger Payne-Hanek range reduction."""
-    # Values around 859240 trigger the Payne-Hanek algorithm
-    # This tests the integer multiply-high instructions used in range reduction
-    self._test_kernel(lambda T: T([859240.0, 1000000.0, 100594688.0]).sin())
-
-  def test_clip_zero_one(self):
-    """Test clip(0, 1) - regression for binary_crossentropy failure."""
-    import numpy as np
-    np.random.seed(0)
-    x_np = np.random.uniform(-2, 2, (32, 10)).astype(np.float32).tolist()
-    self._test_kernel(lambda T: T(x_np).clip(0, 1))
-
-  def test_mod_int64(self):
-    """Test int64 modulo, especially edge cases like 1 % -1."""
-    from tinygrad import dtypes
-    self._test_kernel(lambda T: T([1, 10, -10, 7], dtype=dtypes.int64) % T([-1, 3, 3, -3], dtype=dtypes.int64))
-
-  def test_expand_flatten_sum(self):
-    """Test flatten of expanded tensor followed by sum.
-
-    Bug: flatten() of an expanded tensor produces wrong results for certain sizes.
-    Sizes that are multiples of 32 work (32, 48, 64), but sizes like 33, 49, 50 fail.
-    This breaks masked_select and nonzero operations.
-    """
-    import numpy as np
-    np.random.seed(0)
-    x_np = np.random.uniform(-2, 2, (33,)).astype(np.float32)
-    self._test_kernel(lambda T: (T(x_np.tolist()) > 0.5).unsqueeze(-1).expand(33, 3).flatten().sum())
-
-  @unittest.skip("slow and broken with AMD_LLVM=1")
-  def test_nonzero(self):
-    """Test nonzero operation - counts and gathers indices of non-zero elements."""
-    import numpy as np
-    np.random.seed(42)
-    x_np = np.random.rand(10, 5, 3).astype(np.float32)
-    self._test_kernel(lambda T: (T(x_np.tolist()) > 0.5).nonzero())
-
-  @unittest.skip("Precision differences in v_exp/v_log accumulate across kernels, causing memory divergence")
-  def test_softmax_argmax_fused(self):
-    """Test fused softmax+argmax - tracks exp2 precision issue.
-
-    The fused kernel recomputes softmax inline and Python emulator's exp2 polynomial
-    has up to 1 ULP error vs native exp2f, causing accumulated differences.
-    """
-    import torch
-    torch.manual_seed(0)
-    x_np = torch.rand(4, 10).numpy()
-    self._test_kernel(lambda T: T(x_np.tolist()).softmax(1).argmax())
 
 if __name__ == "__main__":
   unittest.main()

extra/assembly/amd/test/test_formats.py

@@ -0,0 +1,407 @@
+#!/usr/bin/env python3
+"""Test MUBUF, MTBUF, MIMG, EXP, DS formats against LLVM."""
+import unittest
+from extra.assembly.amd.autogen.rdna3.ins import *
+from extra.assembly.amd.dsl import encode_src, RawImm
+from extra.assembly.amd.asm import detect_format
+
+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
+
+
+class TestDetectFormat(unittest.TestCase):
+  """Test detect_format uses encoding from autogen classes."""
+
+  def test_detect_sopp(self):
+    self.assertEqual(detect_format(s_endpgm().to_bytes()), SOPP)
+    self.assertEqual(detect_format(s_nop(0).to_bytes()), SOPP)
+    self.assertEqual(detect_format(s_barrier().to_bytes()), SOPP)
+
+  def test_detect_sop1(self):
+    self.assertEqual(detect_format(s_mov_b32(s[0], 0).to_bytes()), SOP1)
+    self.assertEqual(detect_format(s_mov_b64(s[0:1], 0).to_bytes()), SOP1)
+
+  def test_detect_sop2(self):
+    self.assertEqual(detect_format(s_add_u32(s[0], s[1], s[2]).to_bytes()), SOP2)
+    self.assertEqual(detect_format(s_mul_i32(s[0], s[1], s[2]).to_bytes()), SOP2)
+
+  def test_detect_sopc(self):
+    self.assertEqual(detect_format(s_cmp_eq_i32(s[0], s[1]).to_bytes()), SOPC)
+
+  def test_detect_sopk(self):
+    self.assertEqual(detect_format(s_movk_i32(s[0], 0x1234).to_bytes()), SOPK)
+
+  def test_detect_vop1(self):
+    self.assertEqual(detect_format(v_mov_b32_e32(v[0], 0).to_bytes()), VOP1)
+    self.assertEqual(detect_format(v_rcp_f32_e32(v[0], v[1]).to_bytes()), VOP1)
+
+  def test_detect_vop2(self):
+    self.assertEqual(detect_format(v_add_f32_e32(v[0], v[1], v[2]).to_bytes()), VOP2)
+    self.assertEqual(detect_format(v_mul_f32_e32(v[0], v[1], v[2]).to_bytes()), VOP2)
+
+  def test_detect_vopc(self):
+    self.assertEqual(detect_format(v_cmp_eq_f32_e32(v[0], v[1]).to_bytes()), VOPC)
+    self.assertEqual(detect_format(v_cmp_lt_i32_e32(v[0], v[1]).to_bytes()), VOPC)
+
+  def test_detect_vop3(self):
+    self.assertEqual(detect_format(v_add_f32_e64(v[0], v[1], v[2]).to_bytes()), VOP3)
+    self.assertEqual(detect_format(v_fma_f32(v[0], v[1], v[2], v[3]).to_bytes()), VOP3)
+
+  def test_detect_vop3p(self):
+    self.assertEqual(detect_format(VOP3P(VOP3POp.V_PK_ADD_F16, v[0], v[1], v[2], v[3]).to_bytes()), VOP3P)
+
+  def test_detect_smem(self):
+    self.assertEqual(detect_format(s_load_b32(s[0], s[2:3], 0).to_bytes()), SMEM)
+    self.assertEqual(detect_format(s_load_b64(s[0:1], s[2:3], s[5]).to_bytes()), SMEM)
+
+  def test_detect_ds(self):
+    self.assertEqual(detect_format(ds_load_b32(v[0], v[1]).to_bytes()), DS)
+    self.assertEqual(detect_format(ds_store_b32(v[0], v[1]).to_bytes()), DS)
+
+  def test_detect_flat(self):
+    self.assertEqual(detect_format(global_load_b32(v[0], v[1:3], RawImm(124)).to_bytes()), FLAT)
+    self.assertEqual(detect_format(global_store_b32(v[0:2], v[2], RawImm(124)).to_bytes()), FLAT)
+
+  def test_detect_mubuf(self):
+    self.assertEqual(detect_format(buffer_load_b32(v[0], v[1], s[0:4], s[5]).to_bytes()), MUBUF)
+
+  def test_detect_mtbuf(self):
+    self.assertEqual(detect_format(tbuffer_load_format_x(v[0], v[1], s[0:4], s[5], format=22).to_bytes()), MTBUF)
+
+  def test_detect_mimg(self):
+    self.assertEqual(detect_format(image_load(v[0:4], v[4:6], s[0:8], dmask=0xf, dim=1).to_bytes()), MIMG)
+
+  def test_detect_exp(self):
+    self.assertEqual(detect_format(EXP(en=0xf, target=0, vsrc0=v[0], vsrc1=v[1], vsrc2=v[2], vsrc3=v[3]).to_bytes()), EXP)
+
+  def test_detect_vopd(self):
+    inst = VOPD(VOPDOp.V_DUAL_MOV_B32, VOPDOp.V_DUAL_MOV_B32, vdstx=v[0], vdsty=v[1], srcx0=0, srcy0=0)
+    self.assertEqual(detect_format(inst.to_bytes()), VOPD)
+
+  def test_detect_vinterp(self):
+    inst = VINTERP(VINTERPOp.V_INTERP_P10_F32, vdst=v[0], src0=v[1], src1=v[2], src2=v[3])
+    self.assertEqual(detect_format(inst.to_bytes()), VINTERP)
+
+
+if __name__ == "__main__":
+  unittest.main()

extra/assembly/amd/test/test_handwritten.py

@@ -2,35 +2,31 @@
 # the Inst constructor should be looking at the types of the fields to correctly set the value
 
 import unittest, struct
-from tinygrad.runtime.autogen.amd.rdna3.ins import *
-from tinygrad.renderer.amd.dsl import Inst
-from test.amd.test_roundtrip import compile_asm
-from test.amd.disasm import disasm
+from extra.assembly.amd.autogen.rdna3.ins import *
+from extra.assembly.amd.dsl import Inst
+from extra.assembly.amd.asm import asm
+from extra.assembly.amd.test.test_roundtrip import compile_asm
 
-class IntegrationTestBase(unittest.TestCase):
+class TestIntegration(unittest.TestCase):
   inst: Inst
-  arch: str
   def tearDown(self):
     if not hasattr(self, 'inst'): return
     b = self.inst.to_bytes()
-    st = disasm(self.inst)
-    # Test that the instruction can be compiled by LLVM and produces the same bytes
-    desc = f"{st:25s} {self.inst} {b!r}"
-    self.assertEqual(b, compile_asm(st, arch=self.arch), desc)
+    st = self.inst.disasm()
+    reasm = asm(st)
+    desc = f"{st:25s} {self.inst} {b!r} {reasm}"
+    self.assertEqual(b, compile_asm(st), desc)
+    # TODO: this compare should work for valid things
+    #self.assertEqual(self.inst, reasm)
+    self.assertEqual(repr(self.inst), repr(reasm))
     print(desc)
 
-class TestIntegration(IntegrationTestBase):
-  arch: str = "rdna3"
-
-  def test_wmma(self):
-    self.inst = v_wmma_f32_16x16x16_f16(v[0:7], v[184:191], v[136:143], v[0:7])
-
   def test_load_b128(self):
     self.inst = s_load_b128(s[4:7], s[0:1], NULL, 0)
 
   def test_load_b128_wrong_size(self):
     # this should have to be 4 regs on the loaded to
-    with self.assertRaises(TypeError):
+    with self.assertRaises(Exception):
       self.inst = s_load_b128(s[4:6], s[0:1], NULL, 0)
 
   def test_mov_b32(self):
@@ -129,17 +125,6 @@ class TestIntegration(IntegrationTestBase):
     int_inst = s_mov_b32(s[0], struct.unpack("I", struct.pack("f", 1337.0))[0])
     self.assertEqual(self.inst, int_inst)
 
-class TestIntegrationCDNA(IntegrationTestBase):
-  arch = "cdna"
-
-  def test_mfma(self):
-    from tinygrad.runtime.autogen.amd.cdna.ins import v_mfma_f32_16x16x16_f16
-    self.inst = v_mfma_f32_16x16x16_f16(v[0:3], v[0:1], v[0:1], 0)
-
-  def test_mfma_fp8(self):
-    from tinygrad.runtime.autogen.amd.cdna.ins import v_mfma_f32_16x16x128_f8f6f4
-    self.inst = v_mfma_f32_16x16x128_f8f6f4(v[0:3], v[0:5], v[0:5], 1, cbsz=2, blgp=2)
-
 class TestRegisterSliceSyntax(unittest.TestCase):
   """
   Issue: Register slice syntax should use AMD assembly convention (inclusive end).
@@ -155,18 +140,18 @@ class TestRegisterSliceSyntax(unittest.TestCase):
   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.sz, 4, "s[4:7] should give 4 registers (s4, s5, s6, s7)")
+    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)
-    d = disasm(inst)
+    disasm = inst.disasm()
     # Disasm shows s[4:7] - user should be able to copy this back
-    self.assertIn("s[4:7]", d)
+    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.sz, 4, "s[4:7] from disasm should give 4 registers")
+    self.assertEqual(reg_from_disasm.count, 4, "s[4:7] from disasm should give 4 registers")
 
 class TestInstructionEquality(unittest.TestCase):
   """

extra/assembly/amd/test/test_integration.py

@@ -0,0 +1,330 @@
+#!/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.ins 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

@@ -0,0 +1,201 @@
+#!/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.ins 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)
+      # Format 1: "// GFX11: v_foo ... ; encoding: [0x01,0x02,...]"
+      # Format 2: "// GFX11: [0x01,0x02,...]" (used by DS, older files)
+      if m := re.search(r'(?:GFX11|W32|W64)[^:]*:.*?encoding:\s*\[(.*?)\]', lines[j]):
+        hex_bytes = m.group(1).replace('0x', '').replace(',', '').replace(' ', '')
+      elif m := re.search(r'(?:GFX11|W32|W64)[^:]*:\s*\[(0x[0-9a-fA-F,x\s]+)\]', lines[j]):
+        hex_bytes = m.group(1).replace('0x', '').replace(',', '').replace(' ', '')
+      else:
+        continue
+      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_mockgpu_invalid.py

@@ -1,6 +1,6 @@
 #!/usr/bin/env python3
 """Test that invalid instructions raise exceptions through the mock GPU stack."""
-import unittest, subprocess, os, sys, time
+import unittest, subprocess, os, time
 
 class TestMockGPUInvalidInstruction(unittest.TestCase):
   def test_unsupported_instruction_raises(self):
@@ -20,12 +20,11 @@ runner = get_runner(dev.device, si.ast)
 prg = runner._prg
 lib = bytearray(prg.lib)
 
-# Find s_endpgm (0xBFB00000) and replace with V_MOVRELD_B32 (op=66) which has no pcode
-# VOP1 encoding: bits[31:25]=0x7E, op=bits[16:9], so op=66 -> 66<<9 = 0x8400
+# Find s_endpgm (0xBFB00000) and replace with invalid SOPP op=127 (0xBFFF0000)
 found = False
 for i in range(0, len(lib) - 4, 4):
   if struct.unpack("<I", lib[i:i+4])[0] == 0xBFB00000:
-    lib[i:i+4] = struct.pack("<I", 0x7E008400)
+    lib[i:i+4] = struct.pack("<I", 0xBFFF0000)
     found = True
     break
 assert found, "s_endpgm not found"
@@ -43,11 +42,12 @@ dev.synchronize()
     env["HCQDEV_WAIT_TIMEOUT_MS"] = "10000"
 
     st = time.perf_counter()
-    result = subprocess.run([sys.executable, "-c", test_code], env=env, capture_output=True, text=True, timeout=60)
+    result = subprocess.run(["python", "-c", test_code], env=env, capture_output=True, text=True, timeout=60)
     elapsed = time.perf_counter() - st
 
     self.assertNotEqual(result.returncode, 0, "should have raised")
-    self.assertTrue("Error" in result.stderr, f"expected an error in stderr, got: {result.stderr[:500]}")
+    self.assertTrue("NotImplementedError" in result.stderr or "ValueError" in result.stderr,
+                    f"expected NotImplementedError or ValueError in stderr")
     # Should exit immediately, not wait for the full timeout
     self.assertLess(elapsed, 9.0, f"should exit immediately on emulator exception, took {elapsed:.1f}s")
 

extra/assembly/amd/test/test_pcode.py

@@ -0,0 +1,404 @@
+#!/usr/bin/env python3
+"""Tests for the RDNA3 pseudocode DSL."""
+import unittest
+from extra.assembly.amd.pcode import (Reg, TypedView, SliceProxy, MASK32, MASK64,
+                                       _f32, _i32, _f16, _i16, f32_to_f16, _isnan, _bf16, _ibf16, bf16_to_f32, f32_to_bf16,
+                                       BYTE_PERMUTE, v_sad_u8, v_msad_u8)
+from extra.assembly.amd.pdf import compile_pseudocode, _expr
+from extra.assembly.amd.test.helpers import ExecContext
+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
+  endif
+endfor
+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, even when VCC=0 (no scaling needed).
+    Bug: when VCC._val == vcc (both 0), VCC 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 = Reg(0x3f800000)  # 1.0
+    S1 = Reg(0x40400000)  # 3.0
+    S2 = Reg(0x3f800000)  # 1.0 (numerator)
+    D0, SCC, VCC, EXEC = Reg(0), Reg(0), Reg(0), Reg(0xffffffff)
+    result = _VOP3SDOp_V_DIV_SCALE_F32(S0, S1, S2, D0, SCC, VCC, 0, EXEC, 0, None)
+    # Must always have VCC in result
+    self.assertIn('VCC', result, "V_DIV_SCALE_F32 must always return VCC")
+    self.assertEqual(result['VCC']._val & 1, 0, "VCC lane 0 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
+    S0, S1, S2, D0, SCC, VCC, EXEC = Reg(quiet_nan), Reg(s1_quiet), Reg(0), Reg(0), Reg(0), Reg(0), Reg(0xffffffff)
+    result = _VOPCOp_V_CMP_CLASS_F32(S0, S1, S2, D0, SCC, VCC, 0, EXEC, 0, None)
+    self.assertEqual(result['D0']._val & 1, 1, "Should detect quiet NaN with quiet NaN mask")
+    # Test signaling NaN detection (bit 0 in mask)
+    s1_signal = 0b0000000001  # bit 0 = signaling NaN
+    S0, S1 = Reg(signal_nan), Reg(s1_signal)
+    result = _VOPCOp_V_CMP_CLASS_F32(S0, S1, S2, D0, SCC, VCC, 0, EXEC, 0, None)
+    self.assertEqual(result['D0']._val & 1, 1, "Should detect signaling NaN with signaling NaN mask")
+    # Test that quiet NaN doesn't match signaling NaN mask
+    S0, S1 = Reg(quiet_nan), Reg(s1_signal)
+    result = _VOPCOp_V_CMP_CLASS_F32(S0, S1, S2, D0, SCC, VCC, 0, EXEC, 0, None)
+    self.assertEqual(result['D0']._val & 1, 0, "Quiet NaN should not match signaling NaN mask")
+    # Test that signaling NaN doesn't match quiet NaN mask
+    S0, S1 = Reg(signal_nan), Reg(s1_quiet)
+    result = _VOPCOp_V_CMP_CLASS_F32(S0, S1, S2, D0, SCC, VCC, 0, EXEC, 0, None)
+    self.assertEqual(result['D0']._val & 1, 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")
+
+class TestBF16(unittest.TestCase):
+  """Tests for BF16 (bfloat16) support."""
+
+  def test_bf16_conversion(self):
+    """Test bf16 <-> f32 conversion."""
+    # bf16 is just the top 16 bits of f32
+    # 1.0f = 0x3f800000, bf16 = 0x3f80
+    self.assertAlmostEqual(_bf16(0x3f80), 1.0, places=2)
+    self.assertEqual(_ibf16(1.0), 0x3f80)
+    # 2.0f = 0x40000000, bf16 = 0x4000
+    self.assertAlmostEqual(_bf16(0x4000), 2.0, places=2)
+    self.assertEqual(_ibf16(2.0), 0x4000)
+    # -1.0f = 0xbf800000, bf16 = 0xbf80
+    self.assertAlmostEqual(_bf16(0xbf80), -1.0, places=2)
+    self.assertEqual(_ibf16(-1.0), 0xbf80)
+
+  def test_bf16_special_values(self):
+    """Test bf16 special values (inf, nan)."""
+    import math
+    # +inf: f32 = 0x7f800000, bf16 = 0x7f80
+    self.assertTrue(math.isinf(_bf16(0x7f80)))
+    self.assertEqual(_ibf16(float('inf')), 0x7f80)
+    # -inf: f32 = 0xff800000, bf16 = 0xff80
+    self.assertTrue(math.isinf(_bf16(0xff80)))
+    self.assertEqual(_ibf16(float('-inf')), 0xff80)
+    # NaN: quiet NaN bf16 = 0x7fc0
+    self.assertTrue(math.isnan(_bf16(0x7fc0)))
+    self.assertEqual(_ibf16(float('nan')), 0x7fc0)
+
+  def test_bf16_register_property(self):
+    """Test Reg.bf16 property."""
+    r = Reg(0)
+    r.bf16 = 3.0  # 3.0f = 0x40400000, bf16 = 0x4040
+    self.assertEqual(r._val & 0xffff, 0x4040)
+    self.assertAlmostEqual(float(r.bf16), 3.0, places=1)
+
+  def test_bf16_slice_property(self):
+    """Test SliceProxy.bf16 property."""
+    r = Reg(0x40404040)  # Two bf16 3.0 values
+    self.assertAlmostEqual(r[15:0].bf16, 3.0, places=1)
+    self.assertAlmostEqual(r[31:16].bf16, 3.0, places=1)
+
+class TestBytePermute(unittest.TestCase):
+  """Tests for BYTE_PERMUTE helper function (V_PERM_B32)."""
+
+  def test_byte_select_0_to_7(self):
+    """Test selecting bytes 0-7 from 64-bit data."""
+    # data = {s0, s1} where s0 is bytes 0-3, s1 is bytes 4-7
+    # Combined: 0x0706050403020100 (byte 0 = 0x00, byte 7 = 0x07)
+    data = 0x0706050403020100
+    for i in range(8):
+      self.assertEqual(BYTE_PERMUTE(data, i), i, f"byte {i} should be {i}")
+
+  def test_sign_extend_bytes(self):
+    """Test sign extension selectors 8-11."""
+    # sel 8: sign of byte 1 (bits 15:8)
+    # sel 9: sign of byte 3 (bits 31:24)
+    # sel 10: sign of byte 5 (bits 47:40)
+    # sel 11: sign of byte 7 (bits 63:56)
+    data = 0x8000800080008000  # All relevant bytes have sign bit set
+    self.assertEqual(BYTE_PERMUTE(data, 8), 0xff)
+    self.assertEqual(BYTE_PERMUTE(data, 9), 0xff)
+    self.assertEqual(BYTE_PERMUTE(data, 10), 0xff)
+    self.assertEqual(BYTE_PERMUTE(data, 11), 0xff)
+    data = 0x7f007f007f007f00  # No sign bits set
+    self.assertEqual(BYTE_PERMUTE(data, 8), 0x00)
+    self.assertEqual(BYTE_PERMUTE(data, 9), 0x00)
+    self.assertEqual(BYTE_PERMUTE(data, 10), 0x00)
+    self.assertEqual(BYTE_PERMUTE(data, 11), 0x00)
+
+  def test_constant_zero(self):
+    """Test selector 12 returns 0x00."""
+    self.assertEqual(BYTE_PERMUTE(0xffffffffffffffff, 12), 0x00)
+
+  def test_constant_ff(self):
+    """Test selectors >= 13 return 0xFF."""
+    for sel in [13, 14, 15, 255]:
+      self.assertEqual(BYTE_PERMUTE(0, sel), 0xff, f"sel {sel} should be 0xff")
+
+class TestSADHelpers(unittest.TestCase):
+  """Tests for V_SAD_U8 and V_MSAD_U8 helper functions."""
+
+  def test_v_sad_u8_basic(self):
+    """Test v_sad_u8 with simple values."""
+    # s0 = 0x04030201, s1 = 0x04030201 -> diff = 0 for all bytes
+    result = v_sad_u8(0x04030201, 0x04030201, 0)
+    self.assertEqual(result, 0)
+    # s0 = 0x05040302, s1 = 0x04030201 -> diff = 1+1+1+1 = 4
+    result = v_sad_u8(0x05040302, 0x04030201, 0)
+    self.assertEqual(result, 4)
+
+  def test_v_sad_u8_with_accumulator(self):
+    """Test v_sad_u8 with non-zero accumulator."""
+    # s0 = 0x05040302, s1 = 0x04030201, s2 = 100 -> 4 + 100 = 104
+    result = v_sad_u8(0x05040302, 0x04030201, 100)
+    self.assertEqual(result, 104)
+
+  def test_v_sad_u8_large_diff(self):
+    """Test v_sad_u8 with maximum byte differences."""
+    # s0 = 0xffffffff, s1 = 0x00000000 -> diff = 255*4 = 1020
+    result = v_sad_u8(0xffffffff, 0x00000000, 0)
+    self.assertEqual(result, 1020)
+
+  def test_v_msad_u8_basic(self):
+    """Test v_msad_u8 masks when reference byte is 0."""
+    # s0 = 0x10101010, s1 = 0x00000000 -> all masked, result = 0
+    result = v_msad_u8(0x10101010, 0x00000000, 0)
+    self.assertEqual(result, 0)
+    # s0 = 0x10101010, s1 = 0x01010101 -> diff = |0x10-0x01|*4 = 15*4 = 60
+    result = v_msad_u8(0x10101010, 0x01010101, 0)
+    self.assertEqual(result, 60)
+
+  def test_v_msad_u8_partial_mask(self):
+    """Test v_msad_u8 with partial masking."""
+    # s0 = 0x10101010, s1 = 0x00010001 -> bytes 1 and 3 masked
+    # diff = |0x10-0x01| + |0x10-0x01| = 15 + 15 = 30
+    result = v_msad_u8(0x10101010, 0x00010001, 0)
+    self.assertEqual(result, 30)
+
+  def test_v_msad_u8_with_accumulator(self):
+    """Test v_msad_u8 with non-zero accumulator."""
+    result = v_msad_u8(0x10101010, 0x01010101, 50)
+    self.assertEqual(result, 110)  # 60 + 50
+
+if __name__ == '__main__':
+  unittest.main()

extra/assembly/amd/test/test_pdf_parser.py

@@ -0,0 +1,150 @@
+#!/usr/bin/env python3
+"""Test that PDF parser correctly extracts format fields."""
+import unittest, os
+from extra.assembly.amd.autogen.rdna3.ins 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

@@ -0,0 +1,95 @@
+#!/usr/bin/env python3
+import unittest, subprocess
+from extra.assembly.amd.autogen.rdna3.ins 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/assembly/amd/test/test_roundtrip.py

@@ -1,10 +1,11 @@
 #!/usr/bin/env python3
 """Roundtrip tests: generate tinygrad kernels, decode instructions, re-encode, verify match."""
-import unittest, io, sys, re
-from tinygrad import Device
-from tinygrad.renderer.amd import detect_format
-from test.amd.helpers import llvm_assemble, llvm_disasm, get_target, get_mattr
-from test.amd.disasm import disasm
+import unittest, io, sys, re, subprocess, os
+from extra.assembly.amd.autogen.rdna3.ins import *
+from extra.assembly.amd.dsl import Inst
+from extra.assembly.amd.asm import asm
+from extra.assembly.amd.asm import detect_format
+from extra.assembly.amd.test.helpers import get_llvm_mc, get_llvm_objdump
 
 def disassemble_lib(lib: bytes, compiler) -> list[tuple[str, bytes]]:
   """Disassemble ELF binary and return list of (instruction_text, machine_code_bytes)."""
@@ -29,51 +30,97 @@ def disassemble_lib(lib: bytes, compiler) -> list[tuple[str, bytes]]:
       continue
   return results
 
-def compile_asm(instr: str, arch: str = 'rdna3') -> bytes:
-  """Compile a single instruction using LLVM."""
-  return llvm_assemble([instr], get_target(arch), get_mattr(arch))[0]
+def compile_asm(instr: str, compiler=None) -> bytes:
+  """Compile a single instruction with llvm-mc and return the machine code bytes."""
+  llvm_mc = get_llvm_mc()
+  result = subprocess.run(
+    [llvm_mc, '-triple=amdgcn', '-mcpu=gfx1100', '-mattr=+real-true16,+wavefrontsize32', '-show-encoding'],
+    input=f".text\n{instr}\n", capture_output=True, text=True)
+  if result.returncode != 0: raise RuntimeError(f"llvm-mc failed for '{instr}': {result.stderr.strip()}")
+  # Parse encoding: [0x01,0x39,0x0a,0x7e]
+  for line in result.stdout.split('\n'):
+    if 'encoding:' in line:
+      enc = line.split('encoding:')[1].strip()
+      if enc.startswith('[') and enc.endswith(']'):
+        hex_vals = enc[1:-1].replace('0x', '').replace(',', '').replace(' ', '')
+        return bytes.fromhex(hex_vals)
+  raise RuntimeError(f"no encoding found in llvm-mc output for: {instr}")
 
-def compile_asm_batch(instrs: list[str], arch: str = 'rdna3') -> list[bytes]:
-  """Compile multiple instructions with a single LLVM emission."""
-  return llvm_assemble(instrs, get_target(arch), get_mattr(arch))
-
-def compile_and_disasm_batch(instrs: list[str], arch: str = 'rdna3') -> list[str]:
-  """Compile instructions with LLVM and get LLVM's disassembly."""
+def compile_asm_batch(instrs: list[str]) -> list[bytes]:
+  """Compile multiple instructions with a single llvm-mc call."""
   if not instrs: return []
-  mcpu, mattr = get_target(arch), get_mattr(arch)
-  code = b''.join(llvm_assemble(instrs, mcpu, mattr))
-  return llvm_disasm(code, mcpu, mattr)[:len(instrs)]
+  llvm_mc = get_llvm_mc()
+  src = ".text\n" + "\n".join(instrs) + "\n"
+  result = subprocess.run(
+    [llvm_mc, '-triple=amdgcn', '-mcpu=gfx1100', '-mattr=+real-true16,+wavefrontsize32', '-show-encoding'],
+    input=src, capture_output=True, text=True)
+  if result.returncode != 0: raise RuntimeError(f"llvm-mc batch failed: {result.stderr.strip()}")
+  # Parse all encodings in order
+  encodings = []
+  for line in result.stdout.split('\n'):
+    if 'encoding:' in line:
+      enc = line.split('encoding:')[1].strip()
+      if enc.startswith('[') and enc.endswith(']'):
+        hex_vals = enc[1:-1].replace('0x', '').replace(',', '').replace(' ', '')
+        encodings.append(bytes.fromhex(hex_vals))
+  if len(encodings) != len(instrs): raise RuntimeError(f"expected {len(instrs)} encodings, got {len(encodings)}")
+  return encodings
+
+def compile_and_disasm_batch(instrs: list[str], compiler) -> list[str]:
+  """Compile instructions with LLVM and get LLVM's disassembly."""
+  import tempfile, os
+  if not instrs: return []
+  # Build assembly source with all instructions
+  src = ".text\n.globl test\n.p2align 8\n.type test,@function\ntest:\n"
+  src += "\n".join(f"  {instr}" for instr in instrs) + "\n"
+  # Use llvm-mc to assemble to object file
+  with tempfile.NamedTemporaryFile(suffix='.o', delete=False) as f:
+    obj_path = f.name
+  try:
+    result = subprocess.run(
+      [get_llvm_mc(), '-triple=amdgcn', '-mcpu=gfx1100', '-mattr=+real-true16,+wavefrontsize32', '-filetype=obj', '-o', obj_path],
+      input=src, capture_output=True, text=True)
+    if result.returncode != 0: raise RuntimeError(f"llvm-mc failed: {result.stderr.strip()}")
+    # Disassemble with llvm-objdump
+    result = subprocess.run([get_llvm_objdump(), '-d', '--mcpu=gfx1100', obj_path], capture_output=True, text=True)
+    if result.returncode != 0: raise RuntimeError(f"llvm-objdump failed: {result.stderr.strip()}")
+    # Parse disassembly output
+    results: list[str] = []
+    for line in result.stdout.splitlines():
+      if '//' not in line: continue
+      instr = line.split('//')[0].strip()
+      if instr: results.append(instr)
+    return results[:len(instrs)]
+  finally:
+    os.unlink(obj_path)
 
-@unittest.skipUnless(Device.DEFAULT == "AMD", "requires AMD device")
 class TestTinygradKernelRoundtrip(unittest.TestCase):
   """Test roundtrip on real tinygrad-generated kernels using get_kernels_from_tinygrad pattern."""
-  arch = 'rdna3'
 
   def _test_kernel_roundtrip(self, op_fn):
     """Generate kernel from op_fn, test:
     1. decode -> reencode matches original bytes
-    2. disasm() -> LLVM asm -> bytes matches original (validates disasm correctness)
-    3. our disasm() matches LLVM's disassembly string (informational)
+    2. asm(disasm()) matches LLVM output
+    3. our disasm() matches LLVM's disassembly string exactly
     """
-    arch = self.arch
-
-    from test.amd.test_compare_emulators import get_kernels_from_tinygrad
-    from tinygrad.runtime.support.elf import elf_loader
-    from tinygrad.runtime.support.compiler_amd import HIPCompiler, AMDLLVMCompiler
-    from tinygrad.helpers import AMD_LLVM
+    from extra.assembly.amd.test.test_compare_emulators import get_kernels_from_tinygrad
+    from tinygrad.runtime.support.compiler_amd import HIPCompiler
 
     kernels, _, _ = get_kernels_from_tinygrad(op_fn)
-    # rendered source can be C or llvmir
-    compiler = (AMDLLVMCompiler if AMD_LLVM else HIPCompiler)(get_target(arch))
+    compiler = HIPCompiler('gfx1100')
 
     # First pass: decode all instructions and collect info
     decoded_instrs: list[tuple] = []  # list of (ki, offset, orig_bytes, decoded, our_disasm, decode_ok, decode_err)
     for ki, kernel in enumerate(kernels):
       offset = 0
-      code = next((s.content for s in elf_loader(compiler.compile(kernel.src))[1] if s.name == ".text"))
-      while offset < len(code):
-        remaining = code[offset:]
-        fmt = detect_format(remaining, arch)
+      while offset < len(kernel.code):
+        remaining = kernel.code[offset:]
+        fmt = detect_format(remaining)
+        if fmt is None:
+          decoded_instrs.append((ki, offset, None, None, None, False, "no format"))
+          offset += 4
+          continue
+
         base_size = fmt._size()
         if len(remaining) < base_size:
           break
@@ -83,7 +130,7 @@ class TestTinygradKernelRoundtrip(unittest.TestCase):
           size = decoded.size()  # actual size including literal
           orig_bytes = remaining[:size]
           reencoded = decoded.to_bytes()
-          our_disasm = disasm(decoded)
+          our_disasm = decoded.disasm()
           decode_ok = reencoded == orig_bytes
           decode_err: str | None = None if decode_ok else f"orig={orig_bytes.hex()} reenc={reencoded.hex()}"
           decoded_instrs.append((ki, offset, orig_bytes, decoded, our_disasm, decode_ok, decode_err))
@@ -94,22 +141,22 @@ class TestTinygradKernelRoundtrip(unittest.TestCase):
         offset += size
 
     # Collect disasm strings for batched LLVM calls - skip unknown opcodes (op_X) that LLVM can't compile
-    asm_test_instrs: list[tuple[int, str, bytes]] = []  # (idx, our_disasm, orig_bytes) for asm test
+    asm_test_instrs: list[tuple[int, str]] = []  # (idx, our_disasm) for asm test
     disasm_test_instrs: list[tuple[int, str]] = []  # (idx, our_disasm) for disasm comparison test
 
     for idx, (ki, offset, orig_bytes, decoded, our_disasm, decode_ok, decode_err) in enumerate(decoded_instrs):
       if our_disasm is None: continue
-      # Skip unknown opcodes and malformed instructions
+      # Skip unknown opcodes and malformed instructions for both tests
       if our_disasm.startswith('op_') or re.search(r', \d+, \d+, \d+,', our_disasm): continue
-      asm_test_instrs.append((idx, our_disasm, orig_bytes))
+      asm_test_instrs.append((idx, our_disasm))
       disasm_test_instrs.append((idx, our_disasm))
 
-    # Batch compile for asm test (our disasm -> LLVM asm -> bytes)
-    asm_llvm_results = compile_asm_batch([d for _, d, _ in asm_test_instrs], arch)
-    asm_llvm_map = {idx: (result, orig) for (idx, _, orig), result in zip(asm_test_instrs, asm_llvm_results)}
+    # Batch compile for asm test
+    asm_llvm_results = compile_asm_batch([d for _, d in asm_test_instrs])
+    asm_llvm_map = {idx: result for (idx, _), result in zip(asm_test_instrs, asm_llvm_results)}
 
     # Batch compile+disasm for disasm comparison test
-    disasm_llvm_results = compile_and_disasm_batch([d for _, d in disasm_test_instrs], arch)
+    disasm_llvm_results = compile_and_disasm_batch([d for _, d in disasm_test_instrs], compiler)
     disasm_llvm_map = {idx: result for (idx, _), result in zip(disasm_test_instrs, disasm_llvm_results)}
 
     # Now evaluate results
@@ -130,16 +177,20 @@ class TestTinygradKernelRoundtrip(unittest.TestCase):
         decode_failed += 1
         decode_failures.append(f"K{ki}@{offset}: {our_disasm}: {decode_err}")
 
-      # Asm test: our disasm -> LLVM asm -> compare bytes with original
+      # Asm test
       if our_disasm is None:
         asm_skipped += 1
       elif idx in asm_llvm_map:
-        llvm_bytes, orig = asm_llvm_map[idx]
-        if llvm_bytes == orig[:len(llvm_bytes)]:
-          asm_passed += 1
-        else:
-          asm_failed += 1
-          asm_failures.append(f"K{ki}@{offset}: '{our_disasm}': llvm={llvm_bytes.hex()} orig={orig[:len(llvm_bytes)].hex()}")
+        llvm_bytes = asm_llvm_map[idx]
+        try:
+          our_bytes = asm(our_disasm).to_bytes()
+          if our_bytes[:len(llvm_bytes)] == llvm_bytes:
+            asm_passed += 1
+          else:
+            asm_failed += 1
+            asm_failures.append(f"K{ki}@{offset}: '{our_disasm}': ours={our_bytes[:len(llvm_bytes)].hex()} llvm={llvm_bytes.hex()}")
+        except Exception:
+          asm_skipped += 1
       else:
         asm_skipped += 1
 
@@ -147,20 +198,20 @@ class TestTinygradKernelRoundtrip(unittest.TestCase):
       if our_disasm is None:
         disasm_skipped += 1
       elif idx in disasm_llvm_map:
-        llvm_disasm_str = disasm_llvm_map[idx]
-        if our_disasm == llvm_disasm_str:
+        llvm_disasm = disasm_llvm_map[idx]
+        if our_disasm == llvm_disasm:
           disasm_passed += 1
         else:
           disasm_failed += 1
-          disasm_failures.append(f"K{ki}@{offset}: ours='{our_disasm}' llvm='{llvm_disasm_str}'")
+          disasm_failures.append(f"K{ki}@{offset}: ours='{our_disasm}' llvm='{llvm_disasm}'")
       else:
         disasm_skipped += 1
 
-    print(f"[{arch}] decode roundtrip: {decode_passed} passed, {decode_failed} failed, {decode_skipped} skipped")
-    print(f"[{arch}] asm via llvm: {asm_passed} passed, {asm_failed} failed, {asm_skipped} skipped")
-    print(f"[{arch}] disasm vs llvm: {disasm_passed} passed, {disasm_failed} failed, {disasm_skipped} skipped")
-    self.assertEqual(decode_failed, 0, "Decode failures:\n" + "\n".join(decode_failures[:20]))
-    self.assertEqual(asm_failed, 0, "Asm failures:\n" + "\n".join(asm_failures[:20]))
+    print(f"decode roundtrip: {decode_passed} passed, {decode_failed} failed, {decode_skipped} skipped")
+    print(f"asm vs llvm: {asm_passed} passed, {asm_failed} failed, {asm_skipped} skipped")
+    print(f"disasm vs llvm: {disasm_passed} passed, {disasm_failed} failed, {disasm_skipped} skipped")
+    self.assertEqual(decode_failed, 0, f"Decode failures:\n" + "\n".join(decode_failures[:20]))
+    self.assertEqual(asm_failed, 0, f"Asm failures:\n" + "\n".join(asm_failures[:20]))
     # Note: disasm string comparison is informational only - formatting differences between LLVM versions are expected
 
   # Basic unary ops
@@ -208,10 +259,5 @@ class TestTinygradKernelRoundtrip(unittest.TestCase):
   # Fused ops
   def test_fma(self): self._test_kernel_roundtrip(lambda T: (T([1.0, 2.0]) * T([3.0, 4.0]) + T([5.0, 6.0])))
 
-class TestTinygradKernelRoundtripRDNA4(TestTinygradKernelRoundtrip): arch = 'rdna4'
-
-@unittest.skip("CDNA decode roundtrip not yet supported")
-class TestTinygradKernelRoundtripCDNA(TestTinygradKernelRoundtrip): arch = 'cdna'
-
 if __name__ == "__main__":
   unittest.main()

extra/bench_log.py

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

extra/export_model.py

@@ -1,11 +1,11 @@
 from typing import Tuple, Dict, List, Optional
-from tinygrad.dtype import DType, dtypes
+from tinygrad.dtype import DType
 from tinygrad.renderer import ProgramSpec
-from tinygrad.tensor import Tensor
-from tinygrad.device import Device
+from tinygrad.tensor import Device, Tensor
 from tinygrad.engine.jit import TinyJit
 from tinygrad.nn.state import get_state_dict
 from tinygrad.helpers import Context, to_mv
+from tinygrad.dtype import dtypes
 from tinygrad.uop.ops import Ops
 import json
 from collections import OrderedDict
@@ -13,20 +13,12 @@ from collections import OrderedDict
 EXPORT_SUPPORTED_DEVICE = ["WEBGPU", "CPU", "CUDA", "CL"]
 
 def compile_net(run:TinyJit, special_names:Dict[int,str]) -> Tuple[Dict[str,str],List[Tuple[str,List[str],List[int]]],Dict[str,Tuple[int,DType,int]],Dict[str,Tensor]]:
-  # memory-planned subbuffers can have multiple Buffer objects for the same memory region
-  canon, _seen = {}, {}
-  for ji in run.jit_cache:
-    for b in ji.bufs:
-      if b is not None: canon[id(b)] = _seen.setdefault((id(b.base._buf), b.offset, b.size, b.dtype), b)
-  special_names = {id(canon[k]): v for k, v in special_names.items() if k in canon}
-
   functions, bufs, bufs_to_save, statements, bufnum = {}, {}, {}, [], 0
   for ji in run.jit_cache:
     fxn: ProgramSpec = ji.prg.p
     functions[fxn.function_name] = fxn.src   # NOTE: this assumes all with the same name are the same
     cargs = []
     for i,arg in enumerate(ji.bufs):
-      arg = canon[id(arg)]
       key = id(arg)
       if key not in bufs:
         if key in special_names:
@@ -75,11 +67,12 @@ 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)});" for (name, args, _global_size, _local_size) in statements] + ["}"]
+    cprog += [f"void net({forward_args}) {{"] + [f"{name}({', '.join(args)}, {thread_id});" for (name, args, _global_size, _local_size) in statements] + ["}"]
     return '\n'.join(headers + cprog)
   else:
     if bufs_to_save:

extra/fp8/fp8_linear.py

@@ -1,101 +0,0 @@
-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

@@ -1,497 +0,0 @@
-# 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_ALU, NO_DS, NO_GLOBAL = getenv("NO_ALU", 0), 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)
-    if not NO_ALU:
-      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_copy_matmul.py

@@ -1,110 +0,0 @@
-from tinygrad import UOp, getenv
-from tinygrad.uop.ops import AxisType, KernelInfo, Ops
-from tinygrad.dtype import AddrSpace, dtypes
-
-N = getenv("N", 4096)
-M = getenv("M", N)
-K = getenv("K", N)
-
-WARP_SIZE = 32
-BLOCK_M, BLOCK_N = 128, 128
-BLOCK_K = getenv("BK", 16)
-assert N % BLOCK_N == 0 and M % BLOCK_M == 0 and K % BLOCK_K == 0
-
-use_wmma = getenv("WMMA")
-if use_wmma:
-  WAVES_M, WAVES_N = 2, 2
-  LANES_PER_WAVE_M, LANES_PER_WAVE_N = 2, 16
-  UNROLL_M, UNROLL_N = 1, 1
-
-  # wmma params
-  WMMA_M, WMMA_N, WMMA_K = 16, 16, 16
-  WMMA_ACC = WMMA_M // LANES_PER_WAVE_M
-else:
-  WAVES_M, WAVES_N = 4, 1
-  LANES_PER_WAVE_M, LANES_PER_WAVE_N = 4, 8
-  UNROLL_M, UNROLL_N = 4, 4
-
-# WARP_SIZE * total waves
-THREADS_PER_BLOCK = WARP_SIZE * WAVES_M * WAVES_N
-
-# accumulator size
-TM = BLOCK_M // (WAVES_M * LANES_PER_WAVE_M)
-TN = BLOCK_N // (WAVES_N * LANES_PER_WAVE_N)
-
-def block_128x128_gemm(c:UOp, a:UOp, b:UOp) -> UOp:
-  wave_m = UOp.range(WAVES_M, 2, AxisType.LOCAL)
-  wave_n = UOp.range(WAVES_N, 3, AxisType.LOCAL)
-  lane = UOp.range(WARP_SIZE, -1, AxisType.WARP)
-  tid = (wave_m * WAVES_N + wave_n) * WARP_SIZE + lane
-
-  # -- GLOBAL -> LOCAL --
-  # wmma: spatial outer, k inner (k contiguous for vectorized WMMA tile loads)
-  # gemm: k outer, spatial inner
-  A_local = UOp.placeholder((BLOCK_M, BLOCK_K) if use_wmma else (BLOCK_K, BLOCK_M), a.dtype.base, slot=0, addrspace=AddrSpace.LOCAL)
-  B_local = UOp.placeholder((BLOCK_N, BLOCK_K) if use_wmma else (BLOCK_K, BLOCK_N), b.dtype.base, slot=1, addrspace=AddrSpace.LOCAL)
-
-  a = a.reshape(K // BLOCK_K, BLOCK_K, BLOCK_M)
-  b = b.reshape(K // BLOCK_K, BLOCK_K, BLOCK_N)
-  k_tile = UOp.range(K // BLOCK_K, 100, AxisType.REDUCE)
-
-  # copy with transpose for wmma (input is k×spatial, LDS is spatial×k)
-  A_copy = A_local.permute((1,0)) if use_wmma else A_local
-  B_copy = B_local.permute((1,0)) if use_wmma else B_local
-  A_store = A_copy.reshape(-1, THREADS_PER_BLOCK)[:, tid].store(a[k_tile].reshape(-1, THREADS_PER_BLOCK)[:, tid])
-  B_store = B_copy.reshape(-1, THREADS_PER_BLOCK)[:, tid].store(b[k_tile].reshape(-1, THREADS_PER_BLOCK)[:, tid])
-  barrier = UOp.barrier(A_store, B_store)
-  A_local, B_local = A_local.after(barrier), B_local.after(barrier)
-
-  # -- COMPUTE --
-  lane_m, lane_n = lane // LANES_PER_WAVE_N, lane % LANES_PER_WAVE_N
-
-  # accumulator (unified: both paths use (TM, TN) with scalar dtypes.float)
-  acc = UOp.placeholder((TM, TN), dtypes.float, slot=2, addrspace=AddrSpace.REG)
-  acc = acc.after(acc.store(UOp.const(dtypes.float, 0).reshape((1,)*len(acc.shape)).expand(acc.shape)))
-
-  if use_wmma:
-    k = UOp.range(BLOCK_K // WMMA_K, 101, AxisType.REDUCE)
-    tile_m = UOp.range(TM // WMMA_ACC, 200, AxisType.LOOP)
-    tile_n = UOp.range(TN, 201, AxisType.LOOP)
-
-    acc_frag = acc.reshape(TM // WMMA_ACC, WMMA_ACC, TN).permute(0,2,1)[tile_m, tile_n]
-    a_frag = A_local.reshape(WAVES_M, TM // WMMA_ACC, WMMA_M, BLOCK_K // WMMA_K, WMMA_K)[wave_m, tile_m, lane_n, k]
-    b_frag = B_local.reshape(WAVES_N, TN, WMMA_N, BLOCK_K // WMMA_K, WMMA_K)[wave_n, tile_n, lane_n, k]
-
-    wmma = UOp(Ops.SHAPED_WMMA, dtypes.float, (a_frag, b_frag, acc_frag.after(k)), arg=((16, 16, 16), 'AMD', 32))
-    acc_store = acc_frag.store(wmma).end(tile_m, tile_n)
-  else:
-    # registers for LOCAL -> REG
-    a_frag = UOp.placeholder((TM//UNROLL_M, UNROLL_M), dtypes.float, slot=0, addrspace=AddrSpace.REG)
-    b_frag = UOp.placeholder((TN//UNROLL_N, UNROLL_N), dtypes.float, slot=1, addrspace=AddrSpace.REG)
-
-    k = UOp.range(BLOCK_K, 101, AxisType.REDUCE)
-    a_frag = a_frag.after(a_frag.store(A_local[k].reshape(WAVES_M, TM//UNROLL_M, LANES_PER_WAVE_M, UNROLL_M)[wave_m, :, lane_m, :]))
-    b_frag = b_frag.after(b_frag.store(B_local[k].reshape(WAVES_N, TN//UNROLL_N, LANES_PER_WAVE_N, UNROLL_N)[wave_n, :, lane_n, :]))
-
-    # FMA
-    a_frag = a_frag.reshape(TM, 1).expand(TM, TN)
-    b_frag = b_frag.reshape(1, TN).expand(TM, TN)
-    acc_store = acc.store(acc.after(k) + (a_frag * b_frag))
-
-  # store accumulator and loop
-  acc = acc.after(acc_store.end(k).barrier().end(k_tile))
-
-  # store accumulator to output (unified)
-  c = c.reshape(WAVES_M, TM//UNROLL_M, LANES_PER_WAVE_M, UNROLL_M,
-                WAVES_N, TN//UNROLL_N, LANES_PER_WAVE_N, UNROLL_N)
-  c = c.permute((0,4,2,6, 1,3,5,7)).reshape(THREADS_PER_BLOCK, TM, TN)
-  return c[tid].store(acc).end(wave_m, wave_n, lane)
-
-def amd_copy_matmul(c:UOp, a:UOp, b:UOp) -> UOp:
-  block_id_m = UOp.range(M // BLOCK_M, 0, AxisType.GLOBAL)
-  block_id_n = UOp.range(N // BLOCK_N, 1, AxisType.GLOBAL)
-  c = c.reshape(M // BLOCK_M, BLOCK_M, N // BLOCK_N, BLOCK_N)[block_id_m, :, block_id_n, :]
-  a = a.T.reshape(K, M // BLOCK_M, BLOCK_M)[:, block_id_m, :]
-  b = b.reshape(K, N // BLOCK_N, BLOCK_N)[:, block_id_n, :]
-  return block_128x128_gemm(c, a, b).end(block_id_n, block_id_m).sink(arg=KernelInfo(opts_to_apply=()))
-
-if __name__ == "__main__":
-  from amd_uop_matmul import eval_custom_matmul
-  eval_custom_matmul(amd_copy_matmul, dtypes.half if use_wmma else dtypes.float)

extra/gemm/amd_flash_attention.py

@@ -1,205 +0,0 @@
-from tinygrad import Tensor, UOp, getenv
-from tinygrad.uop.ops import AxisType, KernelInfo, Ops
-from tinygrad.dtype import AddrSpace, dtypes
-from tinygrad.helpers import DEBUG, GlobalCounters, Context
-import math
-
-BLOCK_M, BLOCK_N = 64, 64
-WARP_SIZE = 32
-WMMA_M, WMMA_N, WMMA_K = 16, 16, 16
-WAVES_M, WAVES_N = 4, 1
-LANES_PER_WAVE_M, LANES_PER_WAVE_N = 2, 16
-WMMA_ACC = WMMA_M // LANES_PER_WAVE_M
-THREADS_PER_BLOCK = WARP_SIZE * WAVES_M * WAVES_N
-LDS_PAD = 4  # pad LDS rows to reduce bank conflicts
-
-WMMA_ARG = ((WMMA_M, WMMA_N, WMMA_K), 'AMD', 32)
-LOG2E = math.log2(math.e)
-
-def warp_shfl_xor(val, offset, lane):
-  """Read val from lane ^ offset using ds_bpermute."""
-  idx = ((lane ^ offset) * 4).cast(dtypes.int)
-  return UOp(Ops.CUSTOM, dtypes.float, (idx, val),
-             arg="__builtin_bit_cast(float, __builtin_amdgcn_ds_bpermute({0}, __builtin_bit_cast(int, {1})))")
-
-def warp_reduce_max(val, lane):
-  """Tree reduce MAX across LANES_PER_WAVE_N=16 lanes."""
-  for offset in [8, 4, 2, 1]:
-    val = UOp(Ops.MAX, dtypes.float, (val, warp_shfl_xor(val, offset, lane)))
-  return val
-
-def warp_reduce_sum(val, lane):
-  """Tree reduce SUM across LANES_PER_WAVE_N=16 lanes."""
-  for offset in [8, 4, 2, 1]:
-    val = val + warp_shfl_xor(val, offset, lane)
-  return val
-
-def amd_flash_attention(o:UOp, q:UOp, k:UOp, v:UOp) -> UOp:
-  # inputs are (B*H, N, D)
-  BH, N, D = q.shape
-  assert N % BLOCK_M == 0 and N % BLOCK_N == 0, f"N={N} must be divisible by BLOCK_M={BLOCK_M} and BLOCK_N={BLOCK_N}"
-  assert D % WMMA_K == 0 and D % LANES_PER_WAVE_N == 0, f"D={D} must be divisible by WMMA_K={WMMA_K} and LANES_PER_WAVE_N={LANES_PER_WAVE_N}"
-  assert BLOCK_M % (WAVES_M * WMMA_M) == 0 and BLOCK_N % LANES_PER_WAVE_N == 0
-  TM = BLOCK_M // (WAVES_M * LANES_PER_WAVE_M)
-  TN = BLOCK_N // (WAVES_N * LANES_PER_WAVE_N)
-  TD = D // (WAVES_N * LANES_PER_WAVE_N)
-  SCALE = 1.0 / math.sqrt(D)
-
-  block_bh = UOp.range(BH, 0, AxisType.GLOBAL)
-  block_m = UOp.range(N // BLOCK_M, 1, AxisType.GLOBAL)
-
-  q = q.reshape(BH, N//BLOCK_M, BLOCK_M, D)[block_bh, block_m]
-  k = k.reshape(BH, N//BLOCK_N, BLOCK_N, D)[block_bh]
-  v = v.reshape(BH, N//BLOCK_N, BLOCK_N, D)[block_bh]
-  o = o.reshape(BH, N//BLOCK_M, BLOCK_M, D)[block_bh, block_m]
-
-  wave_m = UOp.range(WAVES_M, 2, AxisType.LOCAL)
-  wave_n = UOp.range(WAVES_N, 3, AxisType.LOCAL)
-  lane = UOp.range(WARP_SIZE, -1, AxisType.WARP)
-  tid = (wave_m * WAVES_N + wave_n) * WARP_SIZE + lane
-  lane_m = lane // LANES_PER_WAVE_N
-  lane_n = lane % LANES_PER_WAVE_N
-
-  # LDS allocation: slot 0 = Q then P (shared), slot 1 = K then V
-  # TODO: the memory planner should be able to find this reuse
-  ELEMS_PER_THREAD = BLOCK_M * D // THREADS_PER_BLOCK
-  QP_lds = UOp.placeholder((BLOCK_M, D + LDS_PAD), dtypes.half, slot=0, addrspace=AddrSpace.LOCAL)
-  KV_lds = UOp.placeholder((BLOCK_N, D + LDS_PAD), dtypes.half, slot=1, addrspace=AddrSpace.LOCAL)[:, :D]
-
-  # register state
-  acc = UOp.placeholder((TM, TD), dtypes.float, slot=2, addrspace=AddrSpace.REG)
-  m_i = UOp.placeholder((TM,), dtypes.float, slot=3, addrspace=AddrSpace.REG)
-  l_i = UOp.placeholder((TM,), dtypes.float, slot=4, addrspace=AddrSpace.REG)
-  acc = acc.after(acc.store(acc.const_like(0)))
-  m_i = m_i.after(m_i.store(m_i.const_like(-math.inf)))
-  l_i = l_i.after(l_i.store(l_i.const_like(0)))
-
-  # ====== KV tile loop ======
-  n_tile = UOp.range(N // BLOCK_N, 100, AxisType.REDUCE)
-
-  # load Q + K into LDS (Q reloaded each iteration since P overwrites slot 0)
-  Q_lds = QP_lds[:, :D]
-  Q_store = Q_lds.after(n_tile).reshape(THREADS_PER_BLOCK, ELEMS_PER_THREAD)[tid].store(
-    q.reshape(THREADS_PER_BLOCK, ELEMS_PER_THREAD)[tid])
-  K_store = KV_lds.reshape(THREADS_PER_BLOCK, ELEMS_PER_THREAD)[tid].store(
-    k[n_tile].reshape(THREADS_PER_BLOCK, ELEMS_PER_THREAD)[tid])
-  qk_load_barrier = UOp.barrier(UOp.group(Q_store, K_store))
-  Q_lds = Q_lds.after(qk_load_barrier)
-  KV_lds_k = KV_lds.after(qk_load_barrier)
-
-  # -- S = Q @ K^T via WMMA (re-init each n_tile) --
-  S_reg = UOp.placeholder((TM, TN), dtypes.float, slot=6, addrspace=AddrSpace.REG)
-  S_reg = S_reg.after(S_reg.after(n_tile).store(S_reg.const_like(0)))
-  k_qk = UOp.range(D // WMMA_K, 101, AxisType.REDUCE)
-  tm1 = UOp.range(TM // WMMA_ACC, 200, AxisType.LOOP)
-  tn1 = UOp.range(TN, 201, AxisType.LOOP)
-  S_frag = S_reg.reshape(TM // WMMA_ACC, WMMA_ACC, TN).permute(0, 2, 1)[tm1, tn1]
-  q_frag = Q_lds.reshape(WAVES_M, TM // WMMA_ACC, WMMA_M, D // WMMA_K, WMMA_K)[wave_m, tm1, lane_n, k_qk]
-  k_frag = KV_lds_k.reshape(WAVES_N, TN, WMMA_N, D // WMMA_K, WMMA_K)[wave_n, tn1, lane_n, k_qk]
-  qk = UOp(Ops.SHAPED_WMMA, dtypes.float, (q_frag, k_frag, S_frag.after(k_qk)), arg=WMMA_ARG)
-  qk_done = S_frag.store(qk).end(tm1, tn1).end(k_qk)
-  S_reg = S_reg.after(qk_done)
-
-  # -- softmax in registers with warp shuffles --
-  S_reg = S_reg.after(S_reg.store(S_reg * SCALE))
-
-  # per-thread local row max over TN=4 elements, then warp reduce across 16 lanes
-  m_ij = UOp.placeholder((TM,), dtypes.float, slot=7, addrspace=AddrSpace.REG)
-  m_ij = m_ij.after(m_ij.after(n_tile).store(m_ij.const_like(-math.inf)))
-  rm2 = UOp.range(TN, 261, AxisType.REDUCE)
-  m_ij = m_ij.after(m_ij.store(m_ij.after(rm2).maximum(S_reg[:, rm2])).end(rm2))
-  # warp reduce max (in-place)
-  ri_w = UOp.range(TM, 270, AxisType.LOOP)
-  m_ij = m_ij.after(m_ij[ri_w].store(warp_reduce_max(m_ij[ri_w], lane)).end(ri_w))
-
-  # compute P = exp(S - m_ij) in S_reg
-  S_reg = S_reg.after(S_reg.store(((S_reg - m_ij.reshape(TM, 1).expand(TM, TN)) * LOG2E).exp2()))
-
-  p_local = UOp.placeholder((TM,), dtypes.float, slot=8, addrspace=AddrSpace.REG)
-  p_local = p_local.after(p_local.after(n_tile).store(p_local.const_like(0)))
-  rp2 = UOp.range(TN, 291, AxisType.REDUCE)
-  p_local = p_local.after(p_local.store(p_local.after(rp2) + S_reg[:, rp2]).end(rp2))
-  ri_ws = UOp.range(TM, 295, AxisType.LOOP)
-  p_sum = p_local.after(p_local[ri_ws].store(warp_reduce_sum(p_local[ri_ws], lane)).end(ri_ws))
-
-  # write P = exp(S - m_ij) to P_lds (reuses slot 0, Q no longer needed)
-  P_lds = QP_lds[:, :BLOCK_N]
-  P_write = P_lds.reshape(WAVES_M, TM // WMMA_ACC, WMMA_ACC, LANES_PER_WAVE_M, WAVES_N, TN, LANES_PER_WAVE_N)
-  P_write = P_write.permute((0, 4, 3, 6, 1, 2, 5)).reshape(THREADS_PER_BLOCK, TM, TN)
-  # TODO: P_write[tid].store(S_reg.cast(dtypes.half)) — shaped store fails due to RESHAPE(DEFINE_LOCAL) surviving linearization
-  rw1 = UOp.range(TM, 296, AxisType.LOOP)
-  rw2 = UOp.range(TN, 297, AxisType.LOOP)
-  P_store = P_write[tid, rw1, rw2].store(S_reg[rw1, rw2].cast(dtypes.half)).end(rw1, rw2)
-
-  # -- online softmax correction --
-  ri4 = UOp.range(TM, 330, AxisType.LOOP)
-  m_new_val = m_i[ri4].maximum(m_ij[ri4])
-  alpha_val = ((m_i[ri4] - m_new_val) * LOG2E).exp2()
-  beta_val = ((m_ij[ri4] - m_new_val) * LOG2E).exp2()
-  rj4 = UOp.range(TD, 331, AxisType.LOOP)
-  correction = UOp.group(
-    acc[ri4, rj4].store(alpha_val * acc[ri4, rj4]).end(rj4),
-    l_i[ri4].store(alpha_val * l_i[ri4] + beta_val * p_sum[ri4]),
-    m_i[ri4].store(m_new_val),
-  ).end(ri4)
-  acc = acc.after(correction)
-  l_i = l_i.after(correction)
-  m_i = m_i.after(correction)
-
-  # load V into KV_lds (must wait for QK WMMA to finish reading K from KV_lds)
-  V_store = KV_lds.after(qk_done).reshape(THREADS_PER_BLOCK, ELEMS_PER_THREAD)[tid].store(
-    v[n_tile].reshape(THREADS_PER_BLOCK, ELEMS_PER_THREAD)[tid])
-  pv_barrier = UOp.barrier(UOp.group(P_store, V_store))
-  P_lds = P_lds.after(pv_barrier)
-  KV_lds_v = KV_lds.after(pv_barrier)
-
-  # -- acc += P @ V via WMMA --
-  k_pv = UOp.range(BLOCK_N // WMMA_K, 400, AxisType.REDUCE)
-  tm2 = UOp.range(TM // WMMA_ACC, 401, AxisType.LOOP)
-  tn2 = UOp.range(TD, 402, AxisType.LOOP)
-  acc_frag = acc.reshape(TM // WMMA_ACC, WMMA_ACC, TD).permute(0, 2, 1)[tm2, tn2]
-  p_frag = P_lds.reshape(WAVES_M, TM // WMMA_ACC, WMMA_M, BLOCK_N // WMMA_K, WMMA_K)[wave_m, tm2, lane_n, k_pv]
-  v_frag = KV_lds_v.reshape(WAVES_N, TD, WMMA_N, BLOCK_N // WMMA_K, WMMA_K)[wave_n, tn2, lane_n, k_pv]
-  pv = UOp(Ops.SHAPED_WMMA, dtypes.float, (p_frag, v_frag, acc_frag.after(k_pv)), arg=WMMA_ARG)
-
-  # end KV tile loop
-  n_tile_end = acc_frag.store(pv).end(tm2, tn2).end(k_pv).barrier().end(n_tile)
-  acc = acc.after(n_tile_end)
-  l_i = l_i.after(n_tile_end)
-  m_i = m_i.after(n_tile_end)
-
-  # normalize: acc /= l_i
-  acc = acc.after(acc.store(acc * (1 / l_i).reshape(TM, 1).expand(TM, TD)))
-
-  # store output
-  o = o.reshape(WAVES_M, TM // WMMA_ACC, WMMA_ACC, LANES_PER_WAVE_M, WAVES_N, TD, LANES_PER_WAVE_N)
-  o = o.permute((0, 4, 3, 6, 1, 2, 5)).reshape(THREADS_PER_BLOCK, TM, TD)
-  return o[tid].store(acc).end(wave_m, wave_n, lane).end(block_m, block_bh).sink(arg=KernelInfo(opts_to_apply=()))
-
-if __name__ == "__main__":
-  B, H, N, D = getenv("B", 1), getenv("H", 32), getenv("N", 1024), getenv("D", 64)
-  q = Tensor.rand(B, H, N, D).cast(dtypes.half)
-  k = Tensor.rand(B, H, N, D).cast(dtypes.half)
-  v = Tensor.rand(B, H, N, D).cast(dtypes.half)
-  o = Tensor.empty(B, H, N, D, dtype=dtypes.float)
-  with Context(DEBUG=0): Tensor.realize(q, k, v)
-
-  q_flat, k_flat, v_flat, o_flat = q.reshape(B*H, N, D), k.reshape(B*H, N, D), v.reshape(B*H, N, D), o.reshape(B*H, N, D)
-  NUM_RUNS = getenv("CNT", 5)
-  ets = []
-  with Context(DEBUG=2):
-    for _ in range(NUM_RUNS):
-      GlobalCounters.reset()
-      tst = Tensor.custom_kernel(o_flat, q_flat, k_flat, v_flat, fxn=amd_flash_attention)[0].realize()
-      ets.append(GlobalCounters.time_sum_s)
-  print(f"best time: {min(ets)*1e3:.2f}ms")
-
-  if getenv("VERIFY", 1):
-    with Context(DEBUG=0):
-      ref = q.float().scaled_dot_product_attention(k.float(), v.float()).reshape(B*H, N, D).realize()
-      err = (ref - tst).square().mean().item()
-    print(f"mean squared error {err}")
-    if err > 1e-2:
-      raise RuntimeError("flash attention is wrong!")
-    else:
-      print("flash attention is correct!")

extra/gemm/amd_uop_matmul.py

@@ -1,74 +1,98 @@
-from tinygrad import Tensor, Context, GlobalCounters, dtypes
+import numpy as np
+from tinygrad import Tensor, Device, Context, GlobalCounters, dtypes
 from tinygrad.uop.ops import UOp, KernelInfo, sint, AxisType
+from tinygrad.engine.realize import ExecItem, get_runner
 from tinygrad.dtype import AddrSpace
-from tinygrad.helpers import DEBUG, getenv
+from tinygrad.helpers import getenv
 
 N = getenv("N", 4096)
-M = getenv("M", N)
-K = getenv("K", N)
-NUM_RUNS = getenv("CNT", 5)
+M = K = N
+run_count = getenv("CNT", 5)
 
 # ---------------------------
 # launch/config constants
 # ---------------------------
 
 WARP_SIZE = 32
-BLOCK_M, BLOCK_N, BLOCK_K = 128, 128, 8
-TM, TN = 4, 4
-LANES_PER_WAVE_M, LANES_PER_WAVE_N = 4, 8
-assert N % BLOCK_N == 0 and M % BLOCK_M == 0 and K % BLOCK_K == 0
+
+# Threadblock tile sizes (block-level tile of C that a block computes)
+BLOCK_N = 128   # columns of C (N-dim) per block
+BLOCK_M = 128   # rows of C (M-dim) per block
+BLOCK_K = 8     # K-slice per block iteration
+
+# Register tile sizes (per-thread accumulator tile of C)
+TN = 4     # columns per thread
+TM = 4     # rows per thread
 
 is_kernel5 = getenv("K5", 0)
 THREADS_PER_BLOCK = 128 if is_kernel5 else 256
-WAVES_PER_BLOCK_N = 1 if is_kernel5 else 2
-WAVES_PER_BLOCK_M = THREADS_PER_BLOCK // WARP_SIZE // WAVES_PER_BLOCK_N
-REG_TILES_PER_WAVE_N = BLOCK_N // (WAVES_PER_BLOCK_N * LANES_PER_WAVE_N * TN)
-REG_TILES_PER_WAVE_M = BLOCK_M // (WAVES_PER_BLOCK_M * LANES_PER_WAVE_M * TM)
+assert THREADS_PER_BLOCK % BLOCK_N == 0, "THREADS_PER_BLOCK must be divisible by BLOCK_N"
+assert THREADS_PER_BLOCK % BLOCK_K == 0, "THREADS_PER_BLOCK must be divisible by BLOCK_K"
+assert (BLOCK_N * BLOCK_K) % THREADS_PER_BLOCK == 0
+assert (BLOCK_M * BLOCK_K) % THREADS_PER_BLOCK == 0
 
-assert WAVES_PER_BLOCK_M*REG_TILES_PER_WAVE_M*LANES_PER_WAVE_M*TM == BLOCK_M, "M reshape is wrong"
-assert WAVES_PER_BLOCK_N*REG_TILES_PER_WAVE_N*LANES_PER_WAVE_N*TN == BLOCK_N, "N reshape is wrong"
+WARPS_PER_BLOCK = THREADS_PER_BLOCK // WARP_SIZE
+WAVE_TILE_N = 128 if is_kernel5 else 64
+WAVE_TILE_M = BLOCK_N * BLOCK_M // WARPS_PER_BLOCK // WAVE_TILE_N
+assert BLOCK_N % WAVE_TILE_N == 0, "BN must be a multiple of WN"
+assert BLOCK_M % WAVE_TILE_M == 0, "BM must be a multiple of WM"
+WAVES_IN_BLOCK_X = BLOCK_N // WAVE_TILE_N
+WAVES_IN_BLOCK_Y = BLOCK_M // WAVE_TILE_M
+assert WAVES_IN_BLOCK_X * WAVES_IN_BLOCK_Y == WARPS_PER_BLOCK, "wave grid must match warps/block"
+
+LANES_PER_WAVE_X = 8
+LANES_PER_WAVE_Y = 4
+ITERS_PER_WAVE_N = WAVE_TILE_N // (LANES_PER_WAVE_X * TN)
+ITERS_PER_WAVE_M = WAVE_TILE_M // (LANES_PER_WAVE_Y * TM)
+assert WAVE_TILE_N % (LANES_PER_WAVE_X * TN) == 0, "WAVE_TILE_N must be divisible by LANES_PER_WAVE_X*TN"
+assert WAVE_TILE_M % (LANES_PER_WAVE_Y * TM) == 0, "WAVE_TILE_M must be divisible by LANES_PER_WAVE_Y*TM"
 
 def rngs_for_shape(shape:tuple[sint, ...], rng:int, axis_type=AxisType.LOOP): return [UOp.range(s, rng+i, axis_type) for i,s in enumerate(shape)]
-def copy(dest:UOp, src:UOp, rng:int, upcast=False):
+def copy(dest:UOp, src:UOp, rng:int, set=False, upcast=False):
   assert dest.shape == src.shape
   rngs = rngs_for_shape(src.shape, rng, AxisType.UPCAST if upcast else AxisType.LOOP)
-  return dest[*rngs].store(src[*rngs]).end(*rngs)
+  copy = dest[*rngs].store(src[*rngs]).end(*rngs)
+  return dest.after(copy) if set else copy
 
-def hand_spec_kernel3(c:UOp, a:UOp, b:UOp) -> UOp:
+def hand_spec_kernel3():
   # ---------------------------
-  # block indices
+  # block indices & placeholders
   # ---------------------------
-  block_id_n = UOp.special(N // BLOCK_N, "gidx0")
-  block_id_m = UOp.special(M // BLOCK_M, "gidx1")
+  blockIdx_x = UOp.special(N // BLOCK_N, "gidx0")
+  blockIdx_y = UOp.special(N // BLOCK_M, "gidx1")
+
+  a = UOp.placeholder((N, N), dtypes.float, slot=1)
+  b = UOp.placeholder((N, N), dtypes.float, slot=2)
+  c = UOp.placeholder((N, N), dtypes.float, slot=0)
 
   # index the output with the globals
-  c = c.reshape(M // BLOCK_M, BLOCK_M, N // BLOCK_N, BLOCK_N)[block_id_m, :, block_id_n, :]
+  c = c.reshape(M // BLOCK_M, BLOCK_M, N // BLOCK_N, BLOCK_N)[blockIdx_y, :, blockIdx_x, :]
 
   # open the main reduction range
-  k_tile_range = UOp.range(K // BLOCK_K, 0, AxisType.REDUCE)
-  a = a.reshape(M // BLOCK_M, BLOCK_M, K // BLOCK_K, BLOCK_K)[block_id_m, :, k_tile_range, :]
-  b = b.reshape(K // BLOCK_K, BLOCK_K, N // BLOCK_N, BLOCK_N)[k_tile_range, :, block_id_n, :]
+  k_tile_range = UOp.range(N // BLOCK_K, 0, AxisType.REDUCE)
+  a = a.reshape(M // BLOCK_M, BLOCK_M, N // BLOCK_K, BLOCK_K)[blockIdx_y, :, k_tile_range, :]
+  b = b.reshape(N // BLOCK_K, BLOCK_K, N // BLOCK_N, BLOCK_N)[k_tile_range, :, blockIdx_x, :]
 
   # globals are no longer used, they are already in the indexes
-  del block_id_m, block_id_n
+  del blockIdx_y, blockIdx_x
 
   # ---------------------------
-  # GLOBAL -> LOCAL (A_local, B_local)
+  # GLOBAL -> LOCAL (As, Bs)
   # ---------------------------
   tid = UOp.special(THREADS_PER_BLOCK, "lidx0")
 
   # A: read BM x BK tiles (permute on store into locals)
-  BM_A_local_stride = (BLOCK_M + 4) if is_kernel5 else BLOCK_M
-  A_local = UOp.placeholder((BLOCK_K, BM_A_local_stride), dtypes.float, slot=0, addrspace=AddrSpace.LOCAL).shrink_to((BLOCK_K, BLOCK_M))
-  A_local_store = copy(A_local.permute((1,0)).reshape(-1, THREADS_PER_BLOCK)[:, tid], a.reshape(-1, THREADS_PER_BLOCK)[:, tid], rng=100)
+  BM_As_stride = (BLOCK_M + 4) if is_kernel5 else BLOCK_M
+  As = UOp.placeholder((BLOCK_K, BM_As_stride), dtypes.float, slot=0, addrspace=AddrSpace.LOCAL).shrink_to((BLOCK_K, BLOCK_M))
+  As_store = copy(As.permute((1,0)).reshape(-1, THREADS_PER_BLOCK)[:, tid], a.reshape(-1, THREADS_PER_BLOCK)[:, tid], rng=100)
 
   # B: read BK x BN tiles
-  B_local = UOp.placeholder((BLOCK_K, BLOCK_N), dtypes.float, slot=1, addrspace=AddrSpace.LOCAL)
-  B_local_store = copy(B_local.reshape(-1, THREADS_PER_BLOCK)[:, tid], b.reshape(-1, THREADS_PER_BLOCK)[:, tid], rng=200)
+  Bs = UOp.placeholder((BLOCK_K, BLOCK_N), dtypes.float, slot=1, addrspace=AddrSpace.LOCAL)
+  Bs_store = copy(Bs.reshape(-1, THREADS_PER_BLOCK)[:, tid], b.reshape(-1, THREADS_PER_BLOCK)[:, tid], rng=200)
 
   # TODO: can we automate barrier?
-  barrier = UOp.barrier(A_local_store, B_local_store)
-  A_local, B_local = A_local.after(barrier), B_local.after(barrier)
+  barrier = UOp.barrier(As_store, Bs_store)
+  As, Bs = As.after(barrier), Bs.after(barrier)
 
   # open inner k range
   k = UOp.range(BLOCK_K, 3, AxisType.REDUCE)
@@ -76,30 +100,31 @@ def hand_spec_kernel3(c:UOp, a:UOp, b:UOp) -> UOp:
   # ---------------------------
   # LOCAL -> REG (per-wave tiles)
   # ---------------------------
-  warp, lane = tid // WARP_SIZE, tid % WARP_SIZE
-  waveIdx, waveIdy = warp % WAVES_PER_BLOCK_N, warp // WAVES_PER_BLOCK_N
-  laneIdx, laneIdy = lane % LANES_PER_WAVE_N, lane // LANES_PER_WAVE_N
-  assert waveIdy.vmax+1 == WAVES_PER_BLOCK_M and laneIdy.vmax+1 == LANES_PER_WAVE_M
+  waveIdx = (tid // WARP_SIZE) % WAVES_IN_BLOCK_X
+  waveIdy = (tid // WARP_SIZE) // WAVES_IN_BLOCK_X
+  assert waveIdy.vmax+1 == WAVES_IN_BLOCK_Y
 
-  A_col = UOp.placeholder((REG_TILES_PER_WAVE_M, TM), dtypes.float, slot=0, addrspace=AddrSpace.REG)
-  A_local_slice = A_local[k, :].reshape(WAVES_PER_BLOCK_M, REG_TILES_PER_WAVE_M, LANES_PER_WAVE_M, TM)[waveIdy, :, laneIdy, :]
-  A_col = A_col.after(copy(A_col, A_local_slice, 300, upcast=True))
+  laneIdx = (tid % WARP_SIZE) % LANES_PER_WAVE_X
+  laneIdy = (tid % WARP_SIZE) // LANES_PER_WAVE_X
+  assert laneIdy.vmax+1 == LANES_PER_WAVE_Y
 
-  B_row = UOp.placeholder((REG_TILES_PER_WAVE_N, TN), dtypes.float, slot=1, addrspace=AddrSpace.REG)
-  B_local_slice = B_local[k, :].reshape(WAVES_PER_BLOCK_N, REG_TILES_PER_WAVE_N, LANES_PER_WAVE_N, TN)[waveIdx, :, laneIdx, :]
-  B_row = B_row.after(copy(B_row, B_local_slice, 400, upcast=True))
+  A_col = UOp.placeholder((ITERS_PER_WAVE_M, TM), dtypes.float, slot=0, addrspace=AddrSpace.REG)
+  A_col = copy(A_col, As[k, :].reshape(WAVES_IN_BLOCK_Y, ITERS_PER_WAVE_M, LANES_PER_WAVE_Y, TM)[waveIdy, :, laneIdy, :], 300, set=True, upcast=True)
+
+  B_row = UOp.placeholder((ITERS_PER_WAVE_N, TN), dtypes.float, slot=1, addrspace=AddrSpace.REG)
+  B_row = copy(B_row, Bs[k, :].reshape(WAVES_IN_BLOCK_X, ITERS_PER_WAVE_N, LANES_PER_WAVE_X, TN)[waveIdx, :, laneIdx, :], 400, set=True, upcast=True)
 
   # ---------------------------
   # FMA: c_regs += A_col * B_row
   # ---------------------------
-  c_regs = UOp.placeholder((REG_TILES_PER_WAVE_M, TM, REG_TILES_PER_WAVE_N, TN), dtypes.float, slot=2, addrspace=AddrSpace.REG)
+  c_regs = UOp.placeholder((ITERS_PER_WAVE_M, TM, ITERS_PER_WAVE_N, TN), dtypes.float, slot=2, addrspace=AddrSpace.REG)
   i = UOp.range(c_regs.size, 16)
   c_regs = c_regs.after(c_regs.flatten()[i].store(0.0).end(i))
 
   # TODO: why don't these work as upcast?
   # why if the ranges merge is it slow?!? (if you change the order on end, they will merge. big slowdown on METAL)
-  iter_m, t_m, iter_n, t_n = rngs = rngs_for_shape(c_regs.shape, 500)
-  sink = c_regs[*rngs].store(c_regs.after(k)[*rngs] + A_col[iter_m, t_m] * B_row[iter_n, t_n]).end(iter_m, iter_n, t_m, t_n)
+  iterWaveM, yt, iterWaveN, xt = rngs = rngs_for_shape(c_regs.shape, 500)
+  sink = c_regs[*rngs].store(c_regs.after(k)[*rngs] + A_col[iterWaveM, yt] * B_row[iterWaveN, xt]).end(iterWaveM, iterWaveN, yt, xt)
 
   # Close k, sync, and close K tiles
   sink = sink.end(k).barrier().end(k_tile_range)
@@ -107,37 +132,38 @@ def hand_spec_kernel3(c:UOp, a:UOp, b:UOp) -> UOp:
   # ---------------------------
   # REG -> GLOBAL (epilogue)
   # ---------------------------
-  c = c.reshape(WAVES_PER_BLOCK_M, REG_TILES_PER_WAVE_M, LANES_PER_WAVE_M, TM,
-                WAVES_PER_BLOCK_N, REG_TILES_PER_WAVE_N, LANES_PER_WAVE_N, TN)
+  c = c.reshape(WAVES_IN_BLOCK_Y, ITERS_PER_WAVE_M, LANES_PER_WAVE_Y, TM,
+                WAVES_IN_BLOCK_X, ITERS_PER_WAVE_N, LANES_PER_WAVE_X, TN)
   c = c[waveIdy, :, laneIdy, :,
         waveIdx, :, laneIdx, :]
   sink = copy(c, c_regs.after(sink), rng=600)
 
   return sink.sink(arg=KernelInfo(opts_to_apply=())).simplify()
 
-def eval_custom_matmul(fxn, dt=dtypes.float):
-  a = Tensor.randn(M, K, dtype=dt)
-  b = Tensor.randn(K, N, dtype=dt)
-  c = Tensor.empty(M, N, dtype=dtypes.float)
-  with Context(DEBUG=0): Tensor.realize(a, b)
+def test_matmul(sink:UOp, 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)
+  Tensor.realize(a, b, hc)
+
+  ei = ExecItem(sink, [t.uop.buffer for t in [hc, a, b]], prg=get_runner(Device.DEFAULT, sink))
 
   ets = []
-  with Context(DEBUG=max(2, DEBUG.value), DEVECTORIZE=2 if dt == dtypes.half else 0):
-    for _ in range(NUM_RUNS):
-      GlobalCounters.reset()
-      tst = Tensor.custom_kernel(c, a, b, fxn=fxn)[0].realize()
-      ets.append(GlobalCounters.time_sum_s)
-  print(f"REAL TFLOPS {M * N * K * 2 / min(ets) * 1e-12:.2f}")
+  with Context(DEBUG=2):
+    for _ in range(run_count):
+      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.float() @ b.float()).realize()
+      tc = (a @ b).realize()
     with Context(DEBUG=0):
-      err = (tc - tst).square().mean().item()
+      err = (hc - tc).square().mean().item()
     print(f"mean squared error {err}")
-    if err > (1e-2 if dt == dtypes.half else 1e-6):
+    if err > 1e-06:
       raise RuntimeError("matmul is wrong!")
 
 if __name__ == "__main__":
-  eval_custom_matmul(hand_spec_kernel3)
+  test_matmul(hand_spec_kernel3(), N=N)