small changes and test fixes from kernel is call

int is less flaky

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

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

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

@@ -56,7 +56,15 @@ runs:
 
     # **** Caching packages ****
 
+    - name: Cache Python packages (PR)
+      if: github.event_name == 'pull_request'
+      id: restore-venv-pr
+      uses: actions/cache/restore@v4
+      with:
+        path: ${{ github.workspace }}/.venv
+        key: venv-${{ runner.os }}-python-${{ steps.setup-python.outputs.python-version }}-${{ inputs.deps }}-${{ inputs.pydeps }}-${{ env.CACHE_VERSION }}
     - name: Cache Python packages
+      if: github.event_name != 'pull_request'
       id: restore-venv
       uses: actions/cache@v4
       with:
@@ -65,23 +73,23 @@ runs:
 
     # **** Caching downloads ****
 
-    - name: Cache downloads (Linux)
-      if: inputs.key != '' && runner.os == 'Linux'
-      uses: actions/cache@v4
+    - name: Cache downloads (PR)
+      if: inputs.key != '' && github.event_name == 'pull_request'
+      uses: actions/cache/restore@v4
       with:
-        path: ~/.cache/tinygrad/downloads/
+        path: ${{ runner.os == 'Linux' && '~/.cache/tinygrad/downloads/' || '~/Library/Caches/tinygrad/downloads/' }}
         key: downloads-${{ github.job }}-${{ inputs.key }}-${{ env.CACHE_VERSION }}
-    - name: Cache downloads (macOS)
-      if: inputs.key != '' && runner.os == 'macOS'
+    - name: Cache downloads
+      if: inputs.key != '' && github.event_name != 'pull_request'
       uses: actions/cache@v4
       with:
-        path: ~/Library/Caches/tinygrad/downloads/
+        path: ${{ runner.os == 'Linux' && '~/.cache/tinygrad/downloads/' || '~/Library/Caches/tinygrad/downloads/' }}
         key: downloads-${{ github.job }}-${{ inputs.key }}-${{ env.CACHE_VERSION }}
 
     # **** Python deps ****
 
     - name: Install dependencies in venv (with extra)
-      if: inputs.deps != '' && steps.restore-venv.outputs.cache-hit != 'true'
+      if: inputs.deps != '' && steps.restore-venv-pr.outputs.cache-hit != 'true' && steps.restore-venv.outputs.cache-hit != 'true'
       shell: bash
       run: |
         python -m venv .venv
@@ -92,7 +100,7 @@ runs:
         fi
         python -m pip install -e ".[${{ inputs.deps }}]" ${{ inputs.pydeps }} --extra-index-url https://download.pytorch.org/whl/cpu --extra-index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/Triton-Nightly/pypi/simple/
     - name: Install dependencies in venv (without extra)
-      if: inputs.deps == '' && steps.restore-venv.outputs.cache-hit != 'true'
+      if: inputs.deps == '' && steps.restore-venv-pr.outputs.cache-hit != 'true' && steps.restore-venv.outputs.cache-hit != 'true'
       shell: bash
       run: |
         python -m venv .venv
@@ -182,8 +190,14 @@ runs:
         echo "pkgs=$pkgs" >> "$GITHUB_OUTPUT"
         echo "hash=$(echo -n "$pkgs" | sha256sum | cut -d' ' -f1)" >> "$GITHUB_OUTPUT"
 
+    - name: Cache apt (PR)
+      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.cuda == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true') && github.event_name == 'pull_request'
+      uses: actions/cache/restore@v4
+      with:
+        path: /var/cache/apt/archives/
+        key: ${{ runner.os }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.CACHE_VERSION }}
     - name: Cache apt
-      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.cuda == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true')
+      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.cuda == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true') && github.event_name != 'pull_request'
       uses: actions/cache@v4
       with:
         path: /var/cache/apt/archives/
@@ -239,8 +253,17 @@ runs:
         ln -s /opt/homebrew/opt/[email protected] /opt/homebrew/opt/boost || true
         ln -s /opt/homebrew/opt/boost/lib/libboost_atomic-mt.dylib /opt/homebrew/opt/boost/lib/libboost_atomic.dylib || true
         ln -s /opt/homebrew/opt/boost/lib/libboost_thread-mt.dylib /opt/homebrew/opt/boost/lib/libboost_thread.dylib || true
+    - name: Cache gpuocelot (PR)
+      if: inputs.ocelot == 'true' && github.event_name == 'pull_request'
+      id: cache-build-pr
+      uses: actions/cache/restore@v4
+      env:
+        cache-name: cache-gpuocelot-build-1
+      with:
+        path: ${{ github.workspace }}/gpuocelot/ocelot
+        key: ${{ runner.os }}-gpuocelot-b16039dc940dc6bc4ea0a98380495769ff35ed99-rebuild-${{ env.CACHE_VERSION }}
     - name: Cache gpuocelot
-      if: inputs.ocelot == 'true'
+      if: inputs.ocelot == 'true' && github.event_name != 'pull_request'
       id: cache-build
       uses: actions/cache@v4
       env:
@@ -249,7 +272,7 @@ runs:
         path: ${{ github.workspace }}/gpuocelot/ocelot
         key: ${{ runner.os }}-gpuocelot-b16039dc940dc6bc4ea0a98380495769ff35ed99-rebuild-${{ env.CACHE_VERSION }}
     - name: Clone/compile gpuocelot
-      if: inputs.ocelot == 'true' && steps.cache-build.outputs.cache-hit != 'true'
+      if: inputs.ocelot == 'true' && steps.cache-build-pr.outputs.cache-hit != 'true' && steps.cache-build.outputs.cache-hit != 'true'
       shell: bash
       run: |
         git clone --recurse-submodules https://github.com/gpuocelot/gpuocelot.git ${{ github.workspace }}/gpuocelot

.github/workflows/autogen.yml

@@ -14,10 +14,12 @@ on:
     paths:
     - 'tinygrad/runtime/autogen/**/*'
     - 'tinygrad/runtime/support/autogen.py'
+    - '.github/workflows/autogen.yml'
   workflow_dispatch:
     paths:
     - 'tinygrad/runtime/autogen/**/*'
     - 'tinygrad/runtime/support/autogen.py'
+    - '.github/workflows/autogen.yml'
 
 jobs:
   autogen:
@@ -38,103 +40,36 @@ jobs:
         mesa: 'true'
         pydeps: 'pyyaml mako'
     - name: Install autogen support packages
-      run: sudo apt-get install -y --no-install-recommends libclang-20-dev llvm-20-dev hip-dev libusb-1.0-0-dev
-    - name: Verify OpenCL autogen
+      run: sudo apt-get install -y --no-install-recommends libclang-20-dev llvm-20-dev hip-dev libusb-1.0-0-dev libdrm-dev
+    - name: Regenerate autogen files
       run: |
-        mv tinygrad/runtime/autogen/opencl.py /tmp/opencl.py.bak
+        find tinygrad/runtime/autogen -type f -name "*.py" -not -name "__init__.py" -not -name "comgr_3.py" -not -name "metal.py" -not -name "iokit.py" -not -name "corefoundation.py" -not -name "libclang.py" -delete
         python3 -c "from tinygrad.runtime.autogen import opencl"
-        diff /tmp/opencl.py.bak tinygrad/runtime/autogen/opencl.py
-    - name: Verify CUDA autogen
-      run: |
-        mv tinygrad/runtime/autogen/cuda.py /tmp/cuda.py.bak
-        mv tinygrad/runtime/autogen/nvrtc.py /tmp/nvrtc.py.bak
-        mv tinygrad/runtime/autogen/nvjitlink.py /tmp/nvjitlink.py.bak
-        mv tinygrad/runtime/autogen/nv_570.py /tmp/nv_570.py.bak
-        mv tinygrad/runtime/autogen/nv.py /tmp/nv.py.bak
-        python3 -c "from tinygrad.runtime.autogen import cuda, nvrtc, nvjitlink, nv_570, nv"
-        diff /tmp/cuda.py.bak tinygrad/runtime/autogen/cuda.py
-        diff /tmp/nvrtc.py.bak tinygrad/runtime/autogen/nvrtc.py
-        diff /tmp/nvjitlink.py.bak tinygrad/runtime/autogen/nvjitlink.py
-        diff /tmp/nv_570.py.bak tinygrad/runtime/autogen/nv_570.py
-        diff /tmp/nv.py.bak tinygrad/runtime/autogen/nv.py
-    - name: Verify AMD autogen
-      run: |
-        mv tinygrad/runtime/autogen/comgr.py /tmp/comgr.py.bak
-        mv tinygrad/runtime/autogen/hsa.py /tmp/hsa.py.bak
-        mv tinygrad/runtime/autogen/hip.py /tmp/hip.py.bak
-        mv tinygrad/runtime/autogen/amd_gpu.py /tmp/amd_gpu.py.bak
-        mv tinygrad/runtime/autogen/sqtt.py /tmp/sqtt.py.bak
-        mv tinygrad/runtime/autogen/rocprof.py /tmp/rocprof.py.bak
-        mv tinygrad/runtime/autogen/am/am.py /tmp/am_am.py.bak
-        mv tinygrad/runtime/autogen/am/pm4_soc15.py /tmp/am_pm4_soc15.py.bak
-        mv tinygrad/runtime/autogen/am/pm4_nv.py /tmp/am_pm4_nv.py.bak
-        mv tinygrad/runtime/autogen/am/sdma_4_0_0.py /tmp/am_sdma_4_0_0.py.bak
-        mv tinygrad/runtime/autogen/am/sdma_5_0_0.py /tmp/am_sdma_5_0_0.py.bak
-        mv tinygrad/runtime/autogen/am/sdma_6_0_0.py /tmp/am_sdma_6_0_0.py.bak
-        mv tinygrad/runtime/autogen/am/smu_v13_0_0.py /tmp/am_smu_v13_0_0.py.bak
-        mv tinygrad/runtime/autogen/am/smu_v14_0_2.py /tmp/am_smu_v14_0_2.py.bak
-        python3 -c "from tinygrad.runtime.autogen import comgr, hsa, hip, amd_gpu, sqtt, rocprof; from tinygrad.runtime.autogen.am import am, pm4_soc15, pm4_nv, sdma_4_0_0, sdma_5_0_0, sdma_6_0_0, smu_v13_0_0, smu_v14_0_2"
-        diff /tmp/comgr.py.bak tinygrad/runtime/autogen/comgr.py
-        diff /tmp/hsa.py.bak tinygrad/runtime/autogen/hsa.py
-        diff /tmp/hip.py.bak tinygrad/runtime/autogen/hip.py
-        diff /tmp/amd_gpu.py.bak tinygrad/runtime/autogen/amd_gpu.py
-        diff /tmp/sqtt.py.bak tinygrad/runtime/autogen/sqtt.py
-        diff /tmp/rocprof.py.bak tinygrad/runtime/autogen/rocprof.py
-        diff /tmp/am_am.py.bak tinygrad/runtime/autogen/am/am.py
-        diff /tmp/am_pm4_soc15.py.bak tinygrad/runtime/autogen/am/pm4_soc15.py
-        diff /tmp/am_pm4_nv.py.bak tinygrad/runtime/autogen/am/pm4_nv.py
-        diff /tmp/am_sdma_4_0_0.py.bak tinygrad/runtime/autogen/am/sdma_4_0_0.py
-        diff /tmp/am_sdma_5_0_0.py.bak tinygrad/runtime/autogen/am/sdma_5_0_0.py
-        diff /tmp/am_sdma_6_0_0.py.bak tinygrad/runtime/autogen/am/sdma_6_0_0.py
-        diff /tmp/am_smu_v13_0_0.py.bak tinygrad/runtime/autogen/am/smu_v13_0_0.py
-        diff /tmp/am_smu_v14_0_2.py.bak tinygrad/runtime/autogen/am/smu_v14_0_2.py
-    - name: Verify Linux autogen
-      run: |
-        mv tinygrad/runtime/autogen/libc.py /tmp/libc.py.bak
-        mv tinygrad/runtime/autogen/kfd.py /tmp/kfd.py.bak
-        mv tinygrad/runtime/autogen/io_uring.py /tmp/io_uring.py.bak
-        mv tinygrad/runtime/autogen/ib.py /tmp/ib.py.bak
-        mv tinygrad/runtime/autogen/pci.py /tmp/pci.py.bak
-        mv tinygrad/runtime/autogen/vfio.py /tmp/vfio.py.bak
+        python3 -c "from tinygrad.runtime.autogen import cuda, nvrtc, nvjitlink, nv_570, nv_580, nv"
+        python3 -c "from tinygrad.runtime.autogen import comgr, 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_v14_0_2"
         python3 -c "from tinygrad.runtime.autogen import libc, kfd, io_uring, ib, pci, vfio"
-        diff /tmp/libc.py.bak tinygrad/runtime/autogen/libc.py
-        diff /tmp/kfd.py.bak tinygrad/runtime/autogen/kfd.py
-        diff /tmp/io_uring.py.bak tinygrad/runtime/autogen/io_uring.py
-        diff /tmp/ib.py.bak tinygrad/runtime/autogen/ib.py
-        diff /tmp/pci.py.bak tinygrad/runtime/autogen/pci.py
-        diff /tmp/vfio.py.bak tinygrad/runtime/autogen/vfio.py
-    - name: Verify LLVM autogen
-      run: |
-        mv tinygrad/runtime/autogen/llvm.py /tmp/llvm.py.bak
         python3 -c "from tinygrad.runtime.autogen import llvm"
-        diff /tmp/llvm.py.bak tinygrad/runtime/autogen/llvm.py
-    - name: Verify WebGPU autogen
-      run: |
-        mv tinygrad/runtime/autogen/webgpu.py /tmp/webgpu.py.bak
         python3 -c "from tinygrad.runtime.autogen import webgpu"
-        diff /tmp/webgpu.py.bak tinygrad/runtime/autogen/webgpu.py
-    - name: Verify Qualcomm autogen
-      run: |
-        mv tinygrad/runtime/autogen/kgsl.py /tmp/kgsl.py.bak
-        mv tinygrad/runtime/autogen/qcom_dsp.py /tmp/qcom_dsp.py.bak
         python3 -c "from tinygrad.runtime.autogen import kgsl, qcom_dsp"
-        diff /tmp/kgsl.py.bak tinygrad/runtime/autogen/kgsl.py
-        diff /tmp/qcom_dsp.py.bak tinygrad/runtime/autogen/qcom_dsp.py
-    - name: Verify libusb autogen
-      run: |
-        mv tinygrad/runtime/autogen/libusb.py /tmp/libusb.py.bak
         python3 -c "from tinygrad.runtime.autogen import libusb"
-        diff /tmp/libusb.py.bak tinygrad/runtime/autogen/libusb.py
-    - name: Verify mesa autogen
-      run: |
-        mv tinygrad/runtime/autogen/mesa.py /tmp/mesa.py.bak
         python3 -c "from tinygrad.runtime.autogen import mesa"
-        diff /tmp/mesa.py.bak tinygrad/runtime/autogen/mesa.py
-    - name: Verify libclang autogen
-      run: |
-        cp tinygrad/runtime/autogen/libclang.py /tmp/libclang.py.bak
+        python3 -c "from tinygrad.runtime.autogen import avcodec"
         REGEN=1 python3 -c "from tinygrad.runtime.autogen import libclang"
-        diff /tmp/libclang.py.bak tinygrad/runtime/autogen/libclang.py
+    - name: Check for differences
+      run: |
+        if ! git diff --quiet; then
+          git diff > autogen-ubuntu.patch
+          echo "Autogen files out of date. Apply patch from: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}#artifacts"
+          exit 1
+        fi
+    - name: Upload patch artifact
+      if: failure()
+      uses: actions/upload-artifact@v4
+      with:
+        name: autogen-ubuntu-patch
+        path: autogen-ubuntu.patch
+
   autogen-mac:
     name: In-tree Autogen (macos)
     runs-on: macos-14
@@ -146,11 +81,24 @@ jobs:
       uses: ./.github/actions/setup-tinygrad
       with:
         llvm: 'true'
-    - name: Verify macos autogen
+    - name: Regenerate autogen files
       run: |
-        mv tinygrad/runtime/autogen/metal.py /tmp/metal.py.bak
-        LIBCLANG_PATH=/opt/homebrew/opt/llvm@20/lib/libclang.dylib python3 -c "from tinygrad.runtime.autogen import metal"
-        diff /tmp/metal.py.bak tinygrad/runtime/autogen/metal.py
+        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 > autogen-macos.patch
+          echo "Autogen files out of date. Apply patch from: ${{ github.server_url }}/${{ 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-3:
     name: In-tree Autogen (comgr 3)
     runs-on: ubuntu-24.04
@@ -169,8 +117,20 @@ jobs:
         echo -e 'Package: *\nPin: release o=repo.radeon.com\nPin-Priority: 600' | sudo tee /etc/apt/preferences.d/rocm-pin-600
         sudo apt -qq update || true
         sudo apt-get install -y --no-install-recommends libclang-20-dev comgr
-    - name: Verify comgr (3) autogen
+    - name: Regenerate autogen files
       run: |
-        mv tinygrad/runtime/autogen/comgr_3.py /tmp/comgr_3.py.bak
+        rm tinygrad/runtime/autogen/comgr_3.py
         python3 -c "from tinygrad.runtime.autogen import comgr_3"
-        diff /tmp/comgr_3.py.bak tinygrad/runtime/autogen/comgr_3.py
+    - name: Check for differences
+      run: |
+        if ! git diff --quiet; then
+          git diff > autogen-comgr3.patch
+          echo "Autogen files out of date. Apply patch from: ${{ github.server_url }}/${{ 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-comgr3-patch
+        path: autogen-comgr3.patch

.github/workflows/benchmark.yml

@@ -16,6 +16,48 @@ 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
+    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" >> $GITHUB_ENV
+        rm -f /tmp/pytest-db-ci*
+    # TODO: remove this step once all old caches are migrated
+    - name: Migrate old huggingface cache (symlinks break onnxruntime 1.24+)
+      run: |
+        cd ~/Library/Caches/tinygrad/downloads/models 2>/dev/null || exit 0
+        for old_dir in models--*; do
+          [ -d "$old_dir" ] || continue
+          repo_id=$(echo "$old_dir" | sed 's/models--//; s/--/\//g')
+          snapshot=$(ls -1 "$old_dir/snapshots" 2>/dev/null | head -1)
+          [ -n "$snapshot" ] || continue
+          mkdir -p "$repo_id"
+          cp -RLn "$old_dir/snapshots/$snapshot/"* "$repo_id/" 2>/dev/null || true
+        done
+    - name: Run pytest -nauto
+      run: |
+        source /tmp/tinygrad_pytest_ci/bin/activate
+        pytest -nauto --durations=20
+
   testmacbenchmark:
     name: Mac Benchmark
     env:
@@ -49,19 +91,19 @@ jobs:
     - name: Print macOS version
       run: sw_vers
     - name: Run Stable Diffusion
-      run: BENCHMARK_LOG=stable_diffusion JIT=1 ASSERT_MIN_STEP_TIME=720 python3.11 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
+      run: BENCHMARK_LOG=stable_diffusion JIT=1 ASSERT_MIN_STEP_TIME=720 python3.11 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing
     - name: Run Stable Diffusion without fp16
-      run: BENCHMARK_LOG=stable_diffusion_fp32 JIT=1 ASSERT_MIN_STEP_TIME=720 python3.11 examples/stable_diffusion.py --seed 0 --noshow --timing | tee sd_no_fp16.txt
+      run: BENCHMARK_LOG=stable_diffusion_fp32 JIT=1 ASSERT_MIN_STEP_TIME=720 python3.11 examples/stable_diffusion.py --seed 0 --noshow --timing
     - name: Run Stable Diffusion v2
       # TODO: very slow step time
-      run: BENCHMARK_LOG=stable_diffusion_v2 JIT=1 ASSERT_MIN_STEP_TIME=4500 python3.11 examples/sdv2.py --fp16 --seed 0 --noshow --timing | tee sdv2.txt
+      run: BENCHMARK_LOG=stable_diffusion_v2 JIT=1 ASSERT_MIN_STEP_TIME=4500 python3.11 examples/sdv2.py --fp16 --seed 0 --noshow --timing
     # process replay can't capture this, the graph is too large
     - name: Run SDXL
-      run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=5000 CAPTURE_PROCESS_REPLAY=0 JIT=1 python3.11 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
+      run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=5000 CAPTURE_PROCESS_REPLAY=0 JIT=1 python3.11 examples/sdxl.py --seed 0 --noshow --timing
     - name: Run model inference benchmark
       run: METAL=1 NOCLANG=1 python3.11 test/external/external_model_benchmark.py
     - name: Test speed vs torch
-      run: BIG=2 MPS=1 python3.11 test/speed/external_test_speed_v_torch.py | tee torch_speed.txt
+      run: BIG=2 MPS=1 python3.11 test/speed/external_test_speed_v_torch.py
     - name: Test tensor cores
       run: METAL=1 python3.11 test/opt/test_tensor_cores.py
     - name: Test AMX tensor cores
@@ -71,84 +113,59 @@ jobs:
         DEBUG=2 CPU=1 CPU_LLVM=0 AMX=1 python3.11 test/opt/test_gen_float4.py TestFloat4.test_float4_multidim_amx TestFloat4.test_float4_multidim_unaligned_load_amx
         DEBUG=2 CPU=1 CPU_LLVM=1 AMX=1 python3.11 test/opt/test_gen_float4.py TestFloat4.test_float4_multidim_amx TestFloat4.test_float4_multidim_unaligned_load_amx
     - name: Run Tensor Core GEMM (float)
-      run: DEBUG=2 SHOULD_USE_TC=1 python3.11 extra/gemm/simple_matmul.py | tee matmul.txt
+      run: DEBUG=2 SHOULD_USE_TC=1 python3.11 extra/gemm/simple_matmul.py
     - name: Run Tensor Core GEMM (half)
-      run: DEBUG=2 SHOULD_USE_TC=1 HALF=1 python3.11 extra/gemm/simple_matmul.py | tee matmul_half.txt
+      run: DEBUG=2 SHOULD_USE_TC=1 HALF=1 python3.11 extra/gemm/simple_matmul.py
     - name: Run Tensor Core GEMM (bfloat16)
-      run: DEBUG=2 SHOULD_USE_TC=1 BFLOAT16=1 python3.11 extra/gemm/simple_matmul.py | tee matmul_bfloat16.txt
+      run: DEBUG=2 SHOULD_USE_TC=1 BFLOAT16=1 python3.11 extra/gemm/simple_matmul.py
     - name: Fuzz Padded Tensor Core GEMM
       run: METAL=1 M_START=6 M_STOP=10 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=6 K_STOP=24 K_STEP=1 TC_OPT=2 DEBUG=2 python3.11 ./extra/gemm/fuzz_matmul.py
     - name: Run LLaMA
       run: |
-        BENCHMARK_LOG=llama_nojit JIT=0 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_unjitted.txt
-        BENCHMARK_LOG=llama JIT=1 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_jitted.txt
+        BENCHMARK_LOG=llama_nojit JIT=0 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
+        BENCHMARK_LOG=llama JIT=1 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
     - name: Run LLaMA with BEAM
-      run: BENCHMARK_LOG=llama_beam JITBEAM=2 IGNORE_BEAM_CACHE=1 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_beam.txt
+      run: BENCHMARK_LOG=llama_beam JITBEAM=2 IGNORE_BEAM_CACHE=1 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
     - name: Run quantized LLaMA
       run: |
-        BENCHMARK_LOG=llama_int8 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing --quantize int8 | tee llama_int8.txt
-        BENCHMARK_LOG=llama_nf4 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing --quantize nf4 | tee llama_nf4.txt
+        BENCHMARK_LOG=llama_int8 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing --quantize int8
+        BENCHMARK_LOG=llama_nf4 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing --quantize nf4
     - name: Run quantized LLaMA3
       run: |
-        BENCHMARK_LOG=llama3_int8 python3.11 examples/llama3.py --size 8B --temperature 0 --benchmark --quantize int8 | tee llama3_int8.txt
-        BENCHMARK_LOG=llama3_nf4 python3.11 examples/llama3.py --size 8B --temperature 0 --benchmark --quantize nf4 | tee llama3_nf4.txt
+        BENCHMARK_LOG=llama3_int8 python3.11 examples/llama3.py --size 8B --temperature 0 --benchmark --quantize int8
+        BENCHMARK_LOG=llama3_nf4 python3.11 examples/llama3.py --size 8B --temperature 0 --benchmark --quantize nf4
     #- name: Run LLaMA 7B on 4 (virtual) GPUs
-    #  run: python3.11 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_four_gpu.txt
+    #  run: python3.11 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing
     - name: Run GPT2
       run: |
-        BENCHMARK_LOG=gpt2_nojit JIT=0 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
-        BENCHMARK_LOG=gpt2 JIT=1 ASSERT_MIN_STEP_TIME=13 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
+        BENCHMARK_LOG=gpt2_nojit JIT=0 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
+        BENCHMARK_LOG=gpt2 JIT=1 ASSERT_MIN_STEP_TIME=13 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
     - name: Run GPT2 w HALF
-      run: BENCHMARK_LOG=gpt2_half HALF=1 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
+      run: BENCHMARK_LOG=gpt2_half HALF=1 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing
     - name: Run GPT2 w HALF/BEAM
-      run: BENCHMARK_LOG=gpt2_half_beam HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
+      run: BENCHMARK_LOG=gpt2_half_beam HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing
     - name: Run OLMoE
       run: BENCHMARK_LOG=olmoe python3.11 examples/olmoe.py
     - name: Train MNIST
-      run: time PYTHONPATH=. TARGET_EVAL_ACC_PCT=96.0 python3.11 examples/beautiful_mnist.py | tee beautiful_mnist.txt
+      run: time PYTHONPATH=. TARGET_EVAL_ACC_PCT=96.0 python3.11 examples/beautiful_mnist.py
 
     # NOTE: this is failing in CI. it is not failing on my machine and I don't really have a way to debug it
     # the error is "RuntimeError: Internal Error (0000000e:Internal Error)"
     #- name: Run 10 CIFAR training steps
-    #  run: BENCHMARK_LOG=cifar_10steps JIT=1 ASSERT_MIN_STEP_TIME=3000 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar.txt
+    #  run: BENCHMARK_LOG=cifar_10steps JIT=1 ASSERT_MIN_STEP_TIME=3000 STEPS=10 python3.11 examples/hlb_cifar10.py
     #- name: Run 10 CIFAR training steps w HALF
-    #  run: BENCHMARK_LOG=cifar_10steps_half JIT=2 ASSERT_MIN_STEP_TIME=3000 STEPS=10 DEFAULT_FLOAT=HALF python3.11 examples/hlb_cifar10.py | tee train_cifar_half.txt
+    #  run: BENCHMARK_LOG=cifar_10steps_half JIT=2 ASSERT_MIN_STEP_TIME=3000 STEPS=10 DEFAULT_FLOAT=HALF python3.11 examples/hlb_cifar10.py
 
     #- name: Run 10 CIFAR training steps w BF16
-    #  run: STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3.11 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
+    #  run: STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3.11 examples/hlb_cifar10.py
     # TODO: too slow
     # - name: Run 10 CIFAR training steps w winograd
-    #   run: BENCHMARK_LOG=cifar_10steps_wino JIT=1 ASSERT_MIN_STEP_TIME=150 WINO=1 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar_wino.txt
+    #   run: BENCHMARK_LOG=cifar_10steps_wino JIT=1 ASSERT_MIN_STEP_TIME=150 WINO=1 STEPS=10 python3.11 examples/hlb_cifar10.py
     - uses: actions/upload-artifact@v4
       with:
         name: Speed (Mac)
         path: |
           onnx_inference_speed.csv
-          torch_speed.txt
-          llama_unjitted.txt
-          llama_jitted.txt
-          llama_beam.txt
-          llama_int8.txt
-          llama_nf4.txt
-          llama3_int8.txt
-          llama3_nf4.txt
-          llama_four_gpu.txt
-          gpt2_unjitted.txt
-          gpt2_jitted.txt
-          gpt2_half.txt
-          gpt2_half_beam.txt
-          matmul.txt
-          matmul_half.txt
-          matmul_bfloat16.txt
-          sd.txt
-          sd_no_fp16.txt
-          sdv2.txt
-          sdxl.txt
-          beautiful_mnist.txt
-          train_cifar.txt
-          train_cifar_half.txt
-          train_cifar_bf16.txt
-          train_cifar_wino.txt
     - name: Run process replay tests
       run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3.11 process_replay.py
 
@@ -170,6 +187,10 @@ jobs:
       run: |
         echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
         rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
+    - name: Kill stale pids
+      run: |
+        PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
+        PYTHONPATH=. ./extra/hcq/hcq_smi.py nv kill_pids
     - name: UsbGPU boot time
       run: sudo -E PYTHONPATH=. DEBUG=2 AM_RESET=1 AMD=1 AMD_IFACE=USB time python3.11 test/test_tiny.py TestTiny.test_plus
     - name: UsbGPU tiny tests
@@ -215,7 +236,7 @@ jobs:
     - name: Run model inference benchmark
       run: NV=1 CAPTURE_PROCESS_REPLAY=0 NOCLANG=1 python3 test/external/external_model_benchmark.py
     - name: Test speed vs torch
-      run: NV=1 CAPTURE_PROCESS_REPLAY=0 HALF=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py | tee torch_speed.txt
+      run: NV=1 CAPTURE_PROCESS_REPLAY=0 HALF=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py
     - name: Test speed vs theoretical
       run: NV=1 IGNORE_BEAM_CACHE=1 CCACHE=0 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
     - name: Test benchmark allreduce
@@ -226,79 +247,58 @@ jobs:
         NV=1 NV_PTX=1 ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
     - name: Run Tensor Core GEMM (CUDA)
       run: |
-        CUDA=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul.txt
-        CUDA=1 SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_bfloat16.txt
-        CUDA=1 SHOULD_USE_TC=1 ALLOW_TF32=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py | tee matmul_tf32.txt
-        CUDA=1 SHOULD_USE_TC=1 FP8E4M3=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_fp8.txt
+        CUDA=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+        CUDA=1 SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+        CUDA=1 SHOULD_USE_TC=1 ALLOW_TF32=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
+        CUDA=1 SHOULD_USE_TC=1 FP8E4M3=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
     - name: Run Tensor Core GEMM (PTX)
-      run: NV=1 NV_PTX=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_ptx.txt
+      run: NV=1 NV_PTX=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
     - name: Run Tensor Core GEMM (NV)
-      run: NV=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_nv.txt
+      run: NV=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
     - name: Test NV=1
       run: DEBUG=2 NV=1 python -m pytest -rA test/test_tiny.py
     - name: Test CUDA=1
       run: DEBUG=2 CUDA=1 python -m pytest -rA test/test_tiny.py
     - name: Run Stable Diffusion
-      run: BENCHMARK_LOG=stable_diffusion NV=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
+      run: BENCHMARK_LOG=stable_diffusion NV=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing
     # TODO: too slow
     # - name: Run SDXL
-    #   run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=2000 CAPTURE_PROCESS_REPLAY=0 NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
+    #   run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=2000 CAPTURE_PROCESS_REPLAY=0 NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/sdxl.py --seed 0 --noshow --timing
     - name: Run LLaMA
       run: |
-        BENCHMARK_LOG=llama_nojit NV=1 JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_unjitted.txt
-        BENCHMARK_LOG=llama NV=1 JIT=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_jitted.txt
+        BENCHMARK_LOG=llama_nojit NV=1 JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
+        BENCHMARK_LOG=llama NV=1 JIT=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
     - name: Run LLaMA with BEAM
-      run: BENCHMARK_LOG=llama_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_beam.txt
+      run: BENCHMARK_LOG=llama_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
     # - name: Run LLaMA 7B on 4 GPUs
-    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_four_gpu.txt
+    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing
     # - name: Run LLaMA 7B on 6 GPUs
-    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_six_gpu.txt
+    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
     - name: Run LLaMA-3 8B BEAM
-      run: BENCHMARK_LOG=llama3_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_beam.txt
+      run: BENCHMARK_LOG=llama3_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
     - name: Run LLaMA-3 8B on 4 GPUs with BEAM
-      run: BENCHMARK_LOG=llama3_beam_4gpu NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_four_gpu.txt
+      run: BENCHMARK_LOG=llama3_beam_4gpu NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
     - name: Run quantized LLaMA3
-      run: BENCHMARK_LOG=llama3_fp8 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --temperature 0 --benchmark --quantize fp8 | tee llama3_fp8.txt
+      run: BENCHMARK_LOG=llama3_fp8 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --temperature 0 --benchmark --quantize fp8
     # - name: Run LLaMA-3 8B on 6 GPUs
-    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 6 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_six_gpu.txt
+    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 6 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
     # - name: Run LLaMA-2 70B
-    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 MAX_CONTEXT=256 python3 examples/llama.py --gen 2 --size 70B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_2_70B.txt
+    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 MAX_CONTEXT=256 python3 examples/llama.py --gen 2 --size 70B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
     - name: Run Mixtral 8x7B
-      run: time BENCHMARK_LOG=mixtral NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/mixtral.py --temperature 0 --count 10 --timing | tee mixtral.txt
+      run: time BENCHMARK_LOG=mixtral NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/mixtral.py --temperature 0 --count 10 --timing
     - name: Run GPT2
       run: |
-        BENCHMARK_LOG=gpt2_nojit NV=1 JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
-        BENCHMARK_LOG=gpt2 NV=1 JIT=1 ASSERT_MIN_STEP_TIME=4 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
+        BENCHMARK_LOG=gpt2_nojit NV=1 JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
+        BENCHMARK_LOG=gpt2 NV=1 JIT=1 ASSERT_MIN_STEP_TIME=4 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
     - name: Run GPT2 w HALF
-      run: BENCHMARK_LOG=gpt2_half NV=1 HALF=1 ASSERT_MIN_STEP_TIME=6 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
+      run: BENCHMARK_LOG=gpt2_half NV=1 HALF=1 ASSERT_MIN_STEP_TIME=6 python3 examples/gpt2.py --count 10 --temperature 0 --timing
     - name: Run GPT2 w HALF/BEAM
-      run: BENCHMARK_LOG=gpt2_half_beam NV=1 HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
+      run: BENCHMARK_LOG=gpt2_half_beam NV=1 HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing
     - uses: actions/upload-artifact@v4
       with:
         name: Speed (NVIDIA)
         path: |
           onnx_inference_speed.csv
-          torch_speed.txt
-          matmul.txt
-          matmul_bfloat16.txt
-          matmul_tf32.txt
-          matmul_ptx.txt
-          matmul_nv.txt
-          sd.txt
-          sdxl.txt
-          llama_unjitted.txt
-          llama_jitted.txt
-          llama_beam.txt
-          llama3_beam.txt
-          llama3_four_gpu.txt
-          llama3_six_gpu.txt
-          llama3_fp8.txt
-          llama_2_70B.txt
-          mixtral.txt
-          gpt2_unjitted.txt
-          gpt2_jitted.txt
-          gpt2_half.txt
-          gpt2_half_beam.txt
     - name: Run process replay tests
       run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py
 
@@ -337,44 +337,30 @@ jobs:
     # - name: Fuzz Padded Tensor Core GEMM (PTX)
     #   run: NV=1 NV_PTX=1 M_START=12 M_STOP=20 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 python3 ./extra/gemm/fuzz_matmul.py
     - name: HEVC Decode Benchmark
-      run: VALIDATE=1 MAX_FRAMES=100 NV=1 PYTHONPATH=. python3 extra/hevc/decode.py
+      run: VALIDATE=1 MAX_FRAMES=100 JITBEAM=1 NV=1 PYTHONPATH=. python3 extra/hevc/decode.py
     - name: Train MNIST
-      run: time PYTHONPATH=. NV=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py | tee beautiful_mnist.txt
+      run: time PYTHONPATH=. NV=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py
     - name: Run 10 CIFAR training steps
-      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=120 NV=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
+      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=120 NV=1 STEPS=10 python3 examples/hlb_cifar10.py
     - name: Run 10 CIFAR training steps w HALF
-      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=110 NV=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_half.txt
+      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=110 NV=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
     - name: Run 10 CIFAR training steps w BF16
-      run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=120 NV=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
+      run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=120 NV=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py
     # - name: Run 10 CIFAR training steps w winograd
-    #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=350 NV=1 CAPTURE_PROCESS_REPLAY=0 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_wino.txt
+    #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=350 NV=1 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
     - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_one_gpu.txt
+      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
     - name: Run full CIFAR training steps w 6 GPUS
-      run: time BENCHMARK_LOG=cifar_6gpu CAPTURE_PROCESS_REPLAY=0 NV=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu.txt
+      run: time BENCHMARK_LOG=cifar_6gpu CAPTURE_PROCESS_REPLAY=0 NV=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
     - name: Run MLPerf resnet eval on training data
       run: time BENCHMARK_LOG=resnet_eval NV=1 MODEL=resnet python3 examples/mlperf/model_eval.py
     - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-      run: BENCHMARK_LOG=resnet_10steps NV=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet_one_gpu.txt
+      run: BENCHMARK_LOG=resnet_10steps NV=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
     - name: Run 10 MLPerf ResNet50 training steps (6 gpu)
-      run: BENCHMARK_LOG=resnet_10steps_6gpu NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet.txt
+      run: BENCHMARK_LOG=resnet_10steps_6gpu NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py
     - name: Run 10 MLPerf Bert training steps (6 gpu)
       # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps_6gpu NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=72 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py | tee train_bert.txt
-    - uses: actions/upload-artifact@v4
-      with:
-        name: Speed (NVIDIA Training)
-        path: |
-          beautiful_mnist.txt
-          train_cifar.txt
-          train_cifar_half.txt
-          train_cifar_bf16.txt
-          train_cifar_wino.txt
-          train_cifar_one_gpu.txt
-          train_cifar_six_gpu.txt
-          train_resnet.txt
-          train_resnet_one_gpu.txt
-          train_bert.txt
+      run: BENCHMARK_LOG=bert_10steps_6gpu NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=72 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
     - name: Run process replay tests
       run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py
 
@@ -389,10 +375,12 @@ jobs:
     steps:
     - name: Checkout Code
       uses: actions/checkout@v4
-    - name: Remove amdgpu
-      run: sudo rmmod amdgpu || true
-    - name: Cleanup running AM processes
-      run: python extra/amdpci/am_smi.py --pids --kill
+    - name: Setcap to python
+      run: ./extra/amdpci/setup_python_cap.sh
+    - name: Remove amd modules
+      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd rmmod
+    - name: Kill stale pids
+      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
     #- name: Insert amdgpu
     #  run: sudo modprobe amdgpu
     - name: Symlink models and datasets
@@ -426,16 +414,18 @@ jobs:
     #- name: Test speed vs torch
     #  run: |
     #    python3 -c "import torch; print(torch.__version__)"
-    #    LD_PRELOAD="/opt/rocm/lib/libhsa-runtime64.so" HSA=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py | tee torch_speed.txt
+    #    LD_PRELOAD="/opt/rocm/lib/libhsa-runtime64.so" HSA=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py
     - name: Test speed vs theoretical
       run: AMD=1 IGNORE_BEAM_CACHE=1 CCACHE=0 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
-    - name: Test tensor cores
-      run: |
-        AMD=1 AMD_LLVM=0 python3 test/opt/test_tensor_cores.py
-        AMD=1 AMD_LLVM=1 python3 test/opt/test_tensor_cores.py
-        AMD=1 SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+    - name: Test tensor cores AMD_LLVM=0
+      run: AMD=1 AMD_LLVM=0 python3 test/opt/test_tensor_cores.py
+    # TODO: this is flaky
+    # - name: Test tensor cores AMD_LLVM=1
+    #   run: AMD=1 AMD_LLVM=1 python3 test/opt/test_tensor_cores.py
     - name: Run Tensor Core GEMM (AMD)
-      run: AMD=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py | tee matmul_amd.txt
+      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
     - name: Test AMD=1
       run: DEBUG=2 AMD=1 python -m pytest -rA test/test_tiny.py
     #- name: Test HIP=1
@@ -450,61 +440,39 @@ jobs:
     - name: Test AM warm start time
       run: time AMD=1 python3 test/test_tiny.py TestTiny.test_plus
     - name: Run Stable Diffusion
-      run: BENCHMARK_LOG=stable_diffusion ASSERT_MIN_STEP_TIME=550 AMD=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
+      run: BENCHMARK_LOG=stable_diffusion ASSERT_MIN_STEP_TIME=550 AMD=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing
     - name: Run SDXL
-      run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=3200 CAPTURE_PROCESS_REPLAY=0 AMD=1 python3 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
+      run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=3200 CAPTURE_PROCESS_REPLAY=0 AMD=1 python3 examples/sdxl.py --seed 0 --noshow --timing
     - name: Run LLaMA 7B
       run: |
-        BENCHMARK_LOG=llama_nojit AMD=1 JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_unjitted.txt
-        BENCHMARK_LOG=llama AMD=1 JIT=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_jitted.txt
+        BENCHMARK_LOG=llama_nojit AMD=1 JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
+        BENCHMARK_LOG=llama AMD=1 JIT=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
     - name: Run LLaMA 7B with BEAM
-      run: BENCHMARK_LOG=llama_beam AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_beam.txt
+      run: BENCHMARK_LOG=llama_beam AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
     # - name: Run LLaMA 7B on 4 GPUs
-    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_four_gpu.txt
+    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing
     # - name: Run LLaMA 7B on 6 GPUs
-    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_six_gpu.txt
+    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
     - name: Run LLaMA-3 8B BEAM
-      run: BENCHMARK_LOG=llama3_beam AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_beam.txt
+      run: BENCHMARK_LOG=llama3_beam AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
     - name: Run LLaMA-3 8B on 4 GPUs with BEAM
-      run: BENCHMARK_LOG=llama3_beam_4gpu AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_four_gpu.txt
+      run: BENCHMARK_LOG=llama3_beam_4gpu AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
     # - name: Run LLaMA-3 8B on 6 GPUs
-    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 6 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_six_gpu.txt
+    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 6 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
     #- name: Restore amdgpu
     #  run: sudo modprobe amdgpu
     # - name: Run LLaMA-2 70B
-    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 2 --size 70B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_2_70B.txt
+    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 2 --size 70B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
     - name: Run Mixtral 8x7B
-      run: time BENCHMARK_LOG=mixtral AMD=1 python3 examples/mixtral.py --temperature 0 --count 10 --timing | tee mixtral.txt
+      run: time BENCHMARK_LOG=mixtral AMD=1 python3 examples/mixtral.py --temperature 0 --count 10 --timing
     - name: Run GPT2
       run: |
-        BENCHMARK_LOG=gpt2_nojit AMD=1 JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
-        BENCHMARK_LOG=gpt2 AMD=1 JIT=1 ASSERT_MIN_STEP_TIME=5 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
+        BENCHMARK_LOG=gpt2_nojit AMD=1 JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
+        BENCHMARK_LOG=gpt2 AMD=1 JIT=1 ASSERT_MIN_STEP_TIME=5 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
     - name: Run GPT2 w HALF
-      run: BENCHMARK_LOG=gpt2_half AMD=1 HALF=1 ASSERT_MIN_STEP_TIME=5 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
+      run: BENCHMARK_LOG=gpt2_half AMD=1 HALF=1 ASSERT_MIN_STEP_TIME=5 python3 examples/gpt2.py --count 10 --temperature 0 --timing
     - name: Run GPT2 w HALF/BEAM
-      run: BENCHMARK_LOG=gpt2_half_beam AMD=1 HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
-    - uses: actions/upload-artifact@v4
-      with:
-        name: Speed (AMD)
-        path: |
-          onnx_inference_speed.csv
-          torch_speed.txt
-          llama_unjitted.txt
-          llama_jitted.txt
-          llama_beam.txt
-          llama3_beam.txt
-          llama3_four_gpu.txt
-          llama3_six_gpu.txt
-          llama_2_70B.txt
-          gpt2_unjitted.txt
-          gpt2_jitted.txt
-          gpt2_half.txt
-          gpt2_half_beam.txt
-          matmul.txt
-          matmul_amd.txt
-          sd.txt
-          sdxl.txt
-          mixtral.txt
+      run: BENCHMARK_LOG=gpt2_half_beam AMD=1 HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing
     - name: Run process replay tests
       run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py
 
@@ -519,10 +487,12 @@ jobs:
     steps:
     - name: Checkout Code
       uses: actions/checkout@v4
-    - name: Remove amdgpu
-      run: sudo rmmod amdgpu || true
-    - name: Cleanup running AM processes
-      run: python extra/amdpci/am_smi.py --pids --kill
+    - name: Setcap to python
+      run: ./extra/amdpci/setup_python_cap.sh
+    - name: Remove amd modules
+      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd rmmod
+    - name: Kill stale pids
+      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
     - name: Symlink models and datasets
       run: |
         mkdir -p weights
@@ -541,31 +511,20 @@ jobs:
     - name: reset process replay
       run: test/external/process_replay/reset.py
     - name: Train MNIST
-      run: time PYTHONPATH=. AMD=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py | tee beautiful_mnist.txt
+      run: time PYTHONPATH=. AMD=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py
     - name: Run 10 CIFAR training steps
-      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=200 AMD=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
+      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=200 AMD=1 STEPS=10 python3 examples/hlb_cifar10.py
     - name: Run 10 CIFAR training steps w HALF
-      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=200 AMD=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_half.txt
+      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=200 AMD=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
     # - name: Run 10 CIFAR training steps w BF16
-    #   run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=288 AMD=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
+    #   run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=288 AMD=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py
     # TODO: too slow
     # - name: Run 10 CIFAR training steps w winograd
-    #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=66 AMD=1 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_wino.txt
+    #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=66 AMD=1 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
     - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_one_gpu.txt
+      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
     - name: Run full CIFAR training steps w 6 GPUS
-      run: time BENCHMARK_LOG=cifar_6gpu AMD=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu.txt
-    - uses: actions/upload-artifact@v4
-      with:
-        name: Speed (AMD Training)
-        path: |
-          beautiful_mnist.txt
-          train_cifar.txt
-          train_cifar_half.txt
-          train_cifar_bf16.txt
-          train_cifar_wino.txt
-          train_cifar_one_gpu.txt
-          train_cifar_six_gpu.txt
+      run: time BENCHMARK_LOG=cifar_6gpu AMD=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
     - 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
 
@@ -580,10 +539,12 @@ jobs:
     steps:
     - name: Checkout Code
       uses: actions/checkout@v4
-    - name: Remove amdgpu
-      run: sudo rmmod amdgpu || true
-    - name: Cleanup running AM processes
-      run: python extra/amdpci/am_smi.py --pids --kill
+    - name: Setcap to python
+      run: ./extra/amdpci/setup_python_cap.sh
+    - name: Remove amd modules
+      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd rmmod
+    - name: Kill stale pids
+      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
     - name: Symlink models and datasets
       run: |
         mkdir -p weights
@@ -604,19 +565,12 @@ jobs:
     - name: Run MLPerf resnet eval
       run: time BENCHMARK_LOG=resnet_eval AMD=1 MODEL=resnet python3 examples/mlperf/model_eval.py
     - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-      run: BENCHMARK_LOG=resnet_10steps AMD=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet_one_gpu.txt
+      run: BENCHMARK_LOG=resnet_10steps AMD=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
     - name: Run 10 MLPerf ResNet50 training steps (6 gpu)
-      run: BENCHMARK_LOG=resnet_10steps_6gpu AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet.txt
+      run: BENCHMARK_LOG=resnet_10steps_6gpu AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py
     - name: Run 10 MLPerf Bert training steps (6 gpu)
       # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps_6gpu AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=72 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py | tee train_bert.txt
-    - uses: actions/upload-artifact@v4
-      with:
-        name: Speed (AMD MLPerf)
-        path: |
-          train_resnet.txt
-          train_resnet_one_gpu.txt
-          train_bert.txt
+      run: BENCHMARK_LOG=bert_10steps_6gpu AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=72 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
     - name: Run process replay tests
       run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py
 
@@ -639,19 +593,21 @@ jobs:
     - name: reset process replay
       run: test/external/process_replay/reset.py
     - name: openpilot compile3 0.10.0 driving_policy
-      run: BENCHMARK_LOG=openpilot_0_10_0_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=4 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.10.0/selfdrive/modeld/models/driving_policy.onnx
+      run: BENCHMARK_LOG=openpilot_0_10_0_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=3 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.10.0/selfdrive/modeld/models/driving_policy.onnx
     - name: openpilot compile3 0.10.0 dmonitoring
       run: BENCHMARK_LOG=openpilot_0_10_0_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=11 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.10.0/selfdrive/modeld/models/dmonitoring_model.onnx
     - name: DEBUG=2 openpilot compile3 0.10.1 driving_vision
       run: PYTHONPATH="." DEBUG=2 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
     - name: DEBUG=2 IMAGE=1 openpilot compile3 0.10.1 driving_vision
       run: PYTHONPATH="." DEBUG=2 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
+    - name: IMAGE=1 openpilot compile3 0.10.1 driving_vision
+      run: BENCHMARK_LOG=image_1_openpilot_0_10_1_vision PYTHONPATH="." DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
     - name: openpilot compile3 0.10.1 driving_vision
       run: BENCHMARK_LOG=openpilot_0_10_1_vision PYTHONPATH="." ASSERT_MIN_STEP_TIME=17 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
     - name: openpilot compile3 0.10.1 driving_policy
-      run: BENCHMARK_LOG=openpilot_0_10_1_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=4 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_policy.onnx
+      run: BENCHMARK_LOG=openpilot_0_10_1_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=3 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_policy.onnx
     - name: openpilot compile3 0.10.1 dmonitoring
-      run: BENCHMARK_LOG=openpilot_0_10_1_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=10 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/dmonitoring_model.onnx
+      run: BENCHMARK_LOG=openpilot_0_10_1_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=11 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/dmonitoring_model.onnx
     - name: benchmark MobileNetV2 on DSP
       run: |
         # generate quantized weights
@@ -674,10 +630,12 @@ jobs:
     steps:
     - name: Checkout Code
       uses: actions/checkout@v4
+    - name: Setcap to python
+      run: ./extra/amdpci/setup_python_cap.sh
     - name: Remove amd modules
-      run: ./extra/hcq/hcq_smi.py amd rmmod
+      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd rmmod
     - name: Kill stale pids
-      run: ./extra/hcq/hcq_smi.py amd kill_pids
+      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
     - name: Symlink models and datasets
       run: |
         mkdir -p weights
@@ -706,7 +664,7 @@ jobs:
     #     AMD=1 AMD_LLVM=1 python3 test/test_linearizer.py test/opt/test_tensor_cores.py
     #     AMD=1 SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
     - name: Run Tensor Core GEMM (AMD)
-      run: AMD=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py | tee am_matmul_amd.txt
+      run: AMD=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
     - name: Test AMD=1
       run: DEBUG=2 AMD=1 python -m pytest -rA test/test_tiny.py
     - name: Test DISK copy time
@@ -716,20 +674,12 @@ jobs:
         AMD=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyDefaulttoCPUJit
         AMD=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyCPUtoDefaultJit
     - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee am_train_cifar_one_gpu.txt
+      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
     # - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-    #   run: BENCHMARK_LOG=resnet_10steps AMD=1 MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee am_train_resnet_one_gpu.txt
+    #   run: BENCHMARK_LOG=resnet_10steps AMD=1 MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
     - name: Run 10 MLPerf Bert training steps (1 gpu)
       # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py | tee am_train_bert_one_gpu.txt
-    - uses: actions/upload-artifact@v4
-      with:
-        name: Speed (AM Driver)
-        path: |
-          am_matmul_amd.txt
-          am_train_cifar_one_gpu.txt
-          am_train_resnet_one_gpu.txt
-          am_train_bert_one_gpu.txt
+      run: BENCHMARK_LOG=bert_10steps AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
     - name: Run process replay tests
       run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py
 
@@ -744,10 +694,12 @@ jobs:
     steps:
     - name: Checkout Code
       uses: actions/checkout@v4
+    - name: Setcap to python
+      run: ./extra/amdpci/setup_python_cap.sh
     - name: Remove nv modules
-      run: ./extra/hcq/hcq_smi.py nv rmmod
+      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py nv rmmod
     - name: Kill stale pids
-      run: ./extra/hcq/hcq_smi.py nv kill_pids
+      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py nv kill_pids
     - name: Symlink models and datasets
       run: |
         mkdir -p weights
@@ -776,21 +728,13 @@ jobs:
         NV=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyDefaulttoCPUJit
         NV=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyCPUtoDefaultJit
     - name: Test LLAMA-3
-      run: BENCHMARK_LOG=llama3_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --benchmark --temperature 0 | tee nv_llama3_beam.txt
+      run: BENCHMARK_LOG=llama3_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --benchmark --temperature 0
     - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee nv_train_cifar_one_gpu.txt
+      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
     - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-      run: BENCHMARK_LOG=resnet_10steps NV=1 MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee nv_train_resnet_one_gpu.txt
+      run: BENCHMARK_LOG=resnet_10steps NV=1 MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
     - name: Run 10 MLPerf Bert training steps (1 gpu)
       # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py | tee nv_train_bert_one_gpu.txt
-    - uses: actions/upload-artifact@v4
-      with:
-        name: Speed (NV Driver)
-        path: |
-          nv_llama3_beam.txt
-          nv_train_cifar_one_gpu.txt
-          nv_train_resnet_one_gpu.txt
-          nv_train_bert_one_gpu.txt
+      run: BENCHMARK_LOG=bert_10steps NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
     - name: Run process replay tests
       run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py

.github/workflows/test.yml

@@ -5,6 +5,7 @@ env:
   CAPTURE_PROCESS_REPLAY: 1
   GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
   PYTHONPATH: ${{ github.workspace }}
+  CHECK_OOB: 1
 
 on:
   push:
@@ -25,17 +26,19 @@ jobs:
       uses: ./.github/actions/setup-tinygrad
       with:
         key: llvm-speed
-        deps: testing_minimal
+        deps: testing_unit
         llvm: 'true'
     - name: Speed Test
-      run: CPU=1 CPU_LLVM=1 python3 test/speed/external_test_speed_v_torch.py
+      run: CPU=1 CPU_LLVM=1 THREADS=0 python3 test/speed/external_test_speed_v_torch.py
     - name: Speed Test (BEAM=2)
-      run: BEAM=2 CPU=1 CPU_LLVM=1 python3 test/speed/external_test_speed_v_torch.py
+      run: BEAM=2 CPU=1 CPU_LLVM=1 THREADS=0 python3 test/speed/external_test_speed_v_torch.py
 
   docs:
     name: Docs
     runs-on: ubuntu-22.04
     timeout-minutes: 10
+    env:
+      CHECK_OOB: 0
     steps:
     - name: Checkout Code
       uses: actions/checkout@v4
@@ -95,23 +98,19 @@ jobs:
       uses: ./.github/actions/setup-tinygrad
       with:
         key: torch-backend-pillow-torchvision-et-pt
-        deps: testing_minimal
+        deps: testing_unit
         pydeps: "pillow torchvision expecttest"
         llvm: 'true'
     - name: Install ninja
       run: |
         sudo apt update || true
         sudo apt install -y --no-install-recommends ninja-build
-    - name: Lint with ruff
-      run: |
-        pip3 install --upgrade --force-reinstall ruff==0.11.0
-        python3 -m ruff check extra/torch_backend/backend.py
     - name: Test one op
       run: FORWARD_ONLY=1 TINY_BACKEND=1 python3 test/test_ops.py TestOps.test_add
     - name: Test ResNet-18
       run: DEBUG=2 python3 extra/torch_backend/example.py
-    - name: My (custom) tests
-      run: python3 extra/torch_backend/test.py
+    - name: custom tests
+      run: python3 -m pytest -n auto extra/torch_backend/test.py --durations=20
     - name: Test one op in torch tests
       run: DEBUG=2 python3 extra/torch_backend/torch_tests.py TestTinyBackendPRIVATEUSE1.test_unary_log_tiny_float32
     - name: Test Ops with TINY_BACKEND
@@ -135,7 +134,7 @@ jobs:
       uses: ./.github/actions/setup-tinygrad
       with:
         key: torch-backend-pillow-torchvision-et-pt
-        deps: testing_minimal
+        deps: testing_unit
         llvm: 'true'
     - name: Install ninja
       run: |
@@ -157,7 +156,7 @@ jobs:
       uses: ./.github/actions/setup-tinygrad
       with:
         key: be-minimal
-        deps: testing_minimal
+        deps: testing_unit
     - name: Test dtype with Python emulator
       run: DEBUG=1 PYTHON=1 python3 -m pytest -n=auto test/test_dtype.py test/test_dtype_alu.py
     - name: Test ops with Python emulator
@@ -208,18 +207,17 @@ jobs:
       run: DEBUG=2 AMX=1 EMULATE=AMX FORWARD_ONLY=1 PYTHON=1 python3 test/opt/test_tensor_cores.py
     - name: Test device flop counts
       run: |
-        DEBUG=2 EMULATE=METAL PYTHON=1 python3 ./test/test_uops_stats.py TestUOpsStatsMatmulHalf
-        DEBUG=2 EMULATE=AMD PYTHON=1 python3 ./test/test_uops_stats.py TestUOpsStatsMatmulHalf
-        DEBUG=2 EMULATE=CUDA PYTHON=1 python3 ./test/test_uops_stats.py TestUOpsStatsMatmulHalf
-        DEBUG=2 EMULATE=INTEL PYTHON=1 python3 ./test/test_uops_stats.py TestUOpsStatsMatmulHalf
-        DEBUG=2 AMX=1 EMULATE=AMX PYTHON=1 python3 ./test/test_uops_stats.py TestUOpsStats.test_simple_matmul
+        DEBUG=2 EMULATE=METAL PYTHON=1 python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
+        DEBUG=2 EMULATE=AMD PYTHON=1 python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
+        DEBUG=2 EMULATE=CUDA PYTHON=1 python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
+        DEBUG=2 EMULATE=INTEL PYTHON=1 python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
+        DEBUG=2 AMX=1 EMULATE=AMX PYTHON=1 python3 ./test/null/test_uops_stats.py TestUOpsStats.test_simple_matmul
 
   linter:
     name: Linters
     runs-on: ubuntu-latest
     timeout-minutes: 10
 
-    # TODO: run the pre-commit hook to replace a lot of this
     steps:
     - name: Checkout Code
       uses: actions/checkout@v4
@@ -231,19 +229,20 @@ jobs:
         deps: linting
     - name: Lint bad-indentation and trailing-whitespace with pylint
       run: python -m pylint --disable=all -e W0311 -e C0303 --jobs=0 --indent-string='  ' --recursive=y .
-    - name: Lint with ruff
+    - name: Run pre-commit linting hooks
+      run: SKIP=tiny,tests,example pre-commit run --all-files
+    - name: Lint additional files with ruff
       run: |
-        pip3 install --upgrade --force-reinstall ruff==0.11.0
-        python3 -m ruff check .
         python3 -m ruff check examples/mlperf/ --ignore E501
         python3 -m ruff check extra/thunder/tiny/ --ignore E501 --ignore F841 --ignore E722
-    - name: Run mypy
+        python3 -m ruff check extra/torch_backend/backend.py
+    - name: Run mypy with lineprecision report
       run: |
-        python -m mypy --strict-equality --lineprecision-report .
+        python -m mypy --lineprecision-report .
+        grep -v autogen lineprecision.txt | awk 'NR>2 {lines+=$2; precise+=$3; imprecise+=$4; any+=$5; empty+=$6} END {t=lines-empty; printf "TOTAL: %d lines, %d precise (%.1f%%), %d imprecise (%.1f%%), %d any (%.1f%%)\n", t, precise, 100*precise/t, imprecise, 100*imprecise/t, any, 100*any/t}'
         cat lineprecision.txt
-    # broken because of UPatAny
-    #- name: Run TYPED=1
-    #  run: TYPED=1 python -c "import tinygrad"
+    - name: Run TYPED=1
+      run: CHECK_OOB=0 DEV=CPU TYPED=1 python test/test_tiny.py
 
   unittest:
     name: Unit Tests
@@ -256,24 +255,30 @@ jobs:
     - name: Setup Environment
       uses: ./.github/actions/setup-tinygrad
       with:
-        key: unittest-12
-        pydeps: "pillow numpy ftfy regex"
+        key: unittest-13
+        pydeps: "pillow ftfy regex pre-commit"
         deps: testing_unit
+        llvm: 'true'
+        amd: 'true'
+    - name: Run pre-commit test hooks
+      run: SKIP=ruff,mypy pre-commit run --all-files
     - name: Check Device.DEFAULT
       run: python -c "from tinygrad import Device; assert Device.DEFAULT == 'CPU', Device.DEFAULT"
     - name: Run unit tests
       run: |
-        CPU=1 python test/unit/test_device.py TestRunAsModule.test_module_runs
-        CPU=1 python -m pytest -n=auto test/unit/ --durations=20 --deselect=test/unit/test_device.py::TestRunAsModule::test_module_runs
+        CPU=1 python test/null/test_device.py TestRunAsModule.test_module_runs
+        CPU=1 python -m pytest -n=auto test/unit/ --durations=20
+    - name: Run 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.test_multitensor.TestMultiTensor.test_data_parallel_resnet_train_step test/device/test_null.py
+      run: NULL=1 python3 -m unittest test.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 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
+      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
@@ -311,7 +316,7 @@ jobs:
         deps: testing_unit
         python-version: '3.14'
     - name: Test SPEC=2
-      run: IGNORE_OOB=0 SPEC=2 PYTHONPATH="." pytest --maxfail=10 -n auto --durations=30 --ignore=test/models --ignore test/test_custom_kernel.py --ignore test/unit/test_hashing.py --timeout 60 -k "not test_setitem_big" --splits 2 --group ${{ matrix.group }}
+      run: SPEC=2 pytest --maxfail=10 -n auto --durations=30 --ignore=test/models --ignore=test/null --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
@@ -345,7 +350,7 @@ jobs:
         uses: ./.github/actions/setup-tinygrad
         with:
           key: gpu-image
-          deps: testing_minimal
+          deps: testing_unit
           opencl: 'true'
       - name: Test CL IMAGE=2 ops
         run: |
@@ -366,7 +371,7 @@ jobs:
         uses: ./.github/actions/setup-tinygrad
         with:
           key: gen-dataset
-          deps: testing_minimal
+          deps: testing
           opencl: 'true'
       - name: Generate Dataset
         run: CL=1 extra/optimization/generate_dataset.sh
@@ -421,7 +426,7 @@ jobs:
         with:
           key: onnxoptc
           deps: testing
-          python-version: '3.11'
+          python-version: '3.12'
           llvm: 'true'
       - name: Test ONNX (CPU)
         run: CPU=1 CPU_LLVM=0 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
@@ -449,7 +454,7 @@ jobs:
           key: onnxoptl
           deps: testing
           pydeps: "tensorflow==2.19"
-          python-version: '3.11'
+          python-version: '3.12'
           opencl: 'true'
       - name: Test ONNX (CL)
         run: CL=1 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
@@ -462,11 +467,11 @@ jobs:
       - name: Test MLPerf stuff
         run: CL=1 python -m pytest -n=auto test/external/external_test_optim.py test/external/external_test_losses.py test/external/external_test_metrics.py test/external/external_test_datasets.py --durations=20
       - name: NULL=1 beautiful_mnist_multigpu
-        run: NULL=1 python examples/beautiful_mnist_multigpu.py
+        run: NULL=1 NULL_ALLOW_COPYOUT=1 python examples/beautiful_mnist_multigpu.py
       - name: Test Bert training
-        run: NULL=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=24 GPUS=4 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
+        run: NULL=1 NULL_ALLOW_COPYOUT=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=24 GPUS=4 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
       - name: Test llama 3 training
-        run: NULL=1 SAMPLES=300 BS=8 SEQLEN=512 GRADIENT_ACC_STEPS=1 FAKEDATA=1 DEFAULT_FLOAT=bfloat16 OPTIM_DTYPE=bfloat16 LLAMA3_SIZE=1B MODEL=llama3 python3 examples/mlperf/model_train.py
+        run: NULL=1 NULL_ALLOW_COPYOUT=1 SAMPLES=300 BS=8 SEQLEN=512 GRADIENT_ACC_STEPS=1 FAKEDATA=1 DEFAULT_FLOAT=bfloat16 OPTIM_DTYPE=bfloat16 LLAMA3_SIZE=1B MODEL=llama3 python3 examples/mlperf/model_train.py
       - name: Run process replay tests
         uses: ./.github/actions/process-replay
 
@@ -474,6 +479,8 @@ jobs:
     name: Test LLM
     runs-on: ubuntu-24.04
     timeout-minutes: 15
+    env:
+      CHECK_OOB: 0
     steps:
       - name: Checkout Code
         uses: actions/checkout@v4
@@ -521,7 +528,7 @@ jobs:
       with:
         key: metal
         deps: testing
-        python-version: '3.11'
+        python-version: '3.12'
     - name: Test models (Metal)
       run: METAL=1 python -m pytest -n=auto test/models --durations=20
     - name: Test LLaMA compile speed
@@ -540,7 +547,7 @@ jobs:
       uses: ./.github/actions/setup-tinygrad
       with:
         key: devectorize-minimal
-        deps: testing_minimal
+        deps: testing_unit
         pydeps: "pillow"
         llvm: "true"
     - name: Test LLVM=1 DEVECTORIZE=0
@@ -561,8 +568,8 @@ jobs:
       uses: ./.github/actions/setup-tinygrad
       with:
         key: dsp-minimal
-        deps: testing_minimal
-        pydeps: "onnx==1.18.0 onnxruntime pillow"
+        deps: testing_unit
+        pydeps: "onnx==1.18.0 onnxruntime ml_dtypes"
         llvm: "true"
     - name: Set up Docker Buildx
       uses: docker/setup-buildx-action@v3
@@ -595,8 +602,8 @@ jobs:
       uses: ./.github/actions/setup-tinygrad
       with:
         key: webgpu-minimal
-        deps: testing_minimal
-        python-version: '3.11'
+        deps: testing_unit
+        python-version: '3.12'
         webgpu: 'true'
     - name: Check Device.DEFAULT (WEBGPU) and print some source
       run: |
@@ -604,9 +611,7 @@ jobs:
         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/ --ignore=test/models --ignore=test/unit \
-          --ignore=test/test_copy_speed.py --ignore=test/test_rearrange_einops.py \
-          --ignore=test/test_fuzz_shape_ops.py --durations=20
+          WEBGPU=1 WEBGPU_BACKEND="WGPUBackendType_Vulkan" python3 -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit --ignore=test/null --durations=20
     - name: Run process replay tests
       uses: ./.github/actions/process-replay
 
@@ -631,7 +636,7 @@ jobs:
         uses: ./.github/actions/setup-tinygrad
         with:
           key: ${{ matrix.backend }}-minimal
-          deps: testing_minimal
+          deps: testing_unit
           amd: 'true'
           llvm: ${{ matrix.backend == 'amdllvm' && 'true' }}
       - name: Check Device.DEFAULT and print some source
@@ -642,19 +647,68 @@ jobs:
         if: matrix.backend=='amdllvm'
         run: python test/device/test_amd_llvm.py
       - name: Run pytest (amd)
-        run: python -m pytest -n=auto test/test_ops.py test/test_dtype.py test/test_dtype_alu.py test/test_linearizer.py test/test_randomness.py test/test_jit.py test/test_graph.py test/test_multitensor.py test/device/test_hcq.py --durations=20
+        run: python -m pytest -n=auto test/test_ops.py test/test_dtype.py test/test_dtype_alu.py test/test_linearizer.py test/test_randomness.py test/test_jit.py test/test_graph.py test/test_multitensor.py test/device/test_hcq.py test/testextra/test_cfg_viz.py --durations=20
       - name: Run pytest (amd)
         run: python -m pytest test/external/external_test_am.py --durations=20
       - name: Run TRANSCENDENTAL math
         run: TRANSCENDENTAL=2 python -m pytest -n=auto test/test_ops.py::TestOps::test_sin test/test_ops.py::TestOps::test_cos test/test_ops.py::TestOps::test_tan test/test_ops.py::TestOps::test_exp test/test_ops.py::TestOps::test_log --durations=20
       - name: Run TestOps.test_add with SQTT
         run: |
-          VIZ=1 PMC=1 DEBUG=5 python3 test/test_ops.py TestOps.test_add
-          VIZ=1 SQTT=1 DEBUG=5 python3 test/test_ops.py TestOps.test_add
+          VIZ=-2 DEBUG=5 python3 test/test_ops.py TestOps.test_add
           extra/sqtt/rgptool.py create "/tmp/profile.pkl.$USER" -o /tmp/gpu0.rgp
+      - name: Run AMD emulated mmapeak on NULL backend
+        env:
+          AMD: 0
+        run: PYTHONPATH=. NULL=1 EMULATE=AMD python extra/mmapeak/mmapeak.py
       - 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 extra.assembly.amd.generate
+          git diff --exit-code extra/assembly/amd/autogen/
+      - 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: Install rocprof-trace-decoder
+        run: sudo PYTHONPATH="." ./extra/sqtt/install_sqtt_decoder.py
+      - name: Run RDNA3 emulator tests
+        run: AMD_LLVM=0 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_LLVM=0 pytest -n=auto test/test_dtype_alu.py test/test_dtype.py --durations 20
+      - name: Run RDNA3 dtype tests (AMD_LLVM=1)
+        run: AMD_LLVM=1 pytest -n=auto test/test_dtype_alu.py test/test_dtype.py --durations 20
+      # TODO: run all once emulator is faster
+      - name: Run RDNA3 ops tests
+        run: SKIP_SLOW_TEST=1 AMD_LLVM=0 pytest -n=auto test/test_ops.py -k "test_sparse_categorical_crossentropy or test_tril or test_nonzero or test_softmax_argmax" --durations 20
+      - name: Run RDNA4 emulator tests
+        run: MOCKGPU_ARCH=rdna4 python -m pytest test/test_tiny.py -v --durations 20
+
   testnvidia:
     strategy:
       fail-fast: false
@@ -674,7 +728,7 @@ jobs:
         uses: ./.github/actions/setup-tinygrad
         with:
           key: ${{ matrix.backend }}-minimal
-          deps: testing_minimal
+          deps: testing_unit
           cuda: 'true'
           ocelot: 'true'
       - name: Set env
@@ -685,7 +739,9 @@ jobs:
           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/ --ignore=test/models --ignore=test/unit --ignore test/test_gc.py --ignore test/test_multitensor.py --durations=20
+        run: python -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit --ignore=test/null --ignore test/test_gc.py --ignore test/test_multitensor.py --durations=20
+      - name: Run TestOps.test_add with PMA
+        run: VIZ=-1 PMA=1 DEBUG=5 python3 test/test_ops.py TestOps.test_add
       - name: Run process replay tests
         uses: ./.github/actions/process-replay
 
@@ -705,7 +761,7 @@ jobs:
         uses: ./.github/actions/setup-tinygrad
         with:
           key: ${{ matrix.backend }}-minimal
-          deps: testing_minimal
+          deps: testing_unit
           opencl: ${{ matrix.backend == 'opencl' && 'true' }}
           llvm: ${{ matrix.backend == 'llvm' || matrix.backend == 'lvp' }}
           mesa: ${{ matrix.backend == 'lvp' && 'true' }}
@@ -716,7 +772,7 @@ jobs:
           python3 -c "from tinygrad import Device; assert Device.DEFAULT in ['CPU','CL'], Device.DEFAULT"
           DEBUG=5 FORWARD_ONLY=1 python3 test/test_ops.py TestOps.test_add
       - name: Run pytest (${{ matrix.backend }})
-        run: python -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit --durations=20
+        run: python -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit --ignore=test/null --durations=20
       - name: Run TRANSCENDENTAL math
         run: TRANSCENDENTAL=2 python -m pytest -n=auto test/test_ops.py::TestOps::test_sin test/test_ops.py::TestOps::test_cos test/test_ops.py::TestOps::test_tan test/test_ops.py::TestOps::test_exp test/test_ops.py::TestOps::test_log --durations=20
       - name: Run process replay tests
@@ -736,13 +792,15 @@ jobs:
       with:
         key: metal
         deps: testing
-        python-version: '3.11'
+        python-version: '3.12'
         amd: 'true'
         cuda: 'true'
         ocelot: 'true'
         llvm: 'true'
     - name: Run unit tests
       run: METAL=1 python -m pytest -n=auto test/unit/ --durations=20
+    - name: Run NULL backend tests
+      run: NULL=1 python -m pytest -n=auto test/null/ --durations=20
     - name: Run ONNX
       run: METAL=1 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
     - name: Test tensor core ops (fake)
@@ -834,8 +892,7 @@ jobs:
         uses: ./.github/actions/setup-tinygrad
         with:
           key: macos-${{ matrix.backend }}-minimal
-          deps: testing_minimal
-          pydeps: "capstone"
+          deps: testing_unit
           llvm: ${{ matrix.backend == 'llvm' || matrix.backend == 'lvp' }}
           mesa: ${{ matrix.backend == 'lvp' && 'true' }}
       - name: Set env
@@ -845,7 +902,7 @@ jobs:
           python -c "from tinygrad import Device; assert Device.DEFAULT == {'LLVM':'CPU','LVP':'CPU'}.get(x:='${{ matrix.backend }}'.upper(), x), Device.DEFAULT"
           DEBUG=4 python3 test/test_tiny.py TestTiny.test_plus
       - name: Run pytest (${{ matrix.backend }})
-        run: python3 -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit --durations=20
+        run: python3 -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit --ignore=test/null --durations=20
       - name: Run process replay tests
         uses: ./.github/actions/process-replay
       - name: Run macOS-specific unit test
@@ -878,7 +935,11 @@ jobs:
       - name: Run unit tests
         if: matrix.backend=='llvm'
         # test_newton_schulz hits RecursionError
-        run: python -m pytest -n=auto test/unit/ --ignore=test/unit/test_disk_tensor.py --ignore=test/unit/test_elf.py --ignore=test/unit/test_tar.py --ignore=test/unit/test_linalg.py --durations=20
+        run: python -m pytest -n=auto test/unit/ --ignore=test/unit/test_disk_tensor.py --ignore=test/unit/test_tar.py --ignore=test/unit/test_linalg.py --durations=20
+      - name: Run NULL backend tests
+        if: matrix.backend=='llvm'
+        shell: bash
+        run: CPU=0 CPU_LLVM=0 NULL=1 python -m pytest -n=auto test/null/ --ignore=test/null/test_elf.py --durations=20
       - name: Run pytest (${{ matrix.backend }})
         shell: bash
         run: |
@@ -902,12 +963,12 @@ jobs:
         uses: ./.github/actions/setup-tinygrad
         with:
           key: compile-${{ matrix.backend }}
-          deps: testing_minimal
+          deps: testing_unit
           mesa: ${{ (matrix.backend == 'ir3' || matrix.backend == 'nak') && 'true' }}
-          python-version: '3.14'
+          python-version: '3.12'
       - name: Set env
         shell: bash
-        run: printf "NULL=1\n${{ matrix.backend == 'ir3' && 'NULL_IR3=1' || matrix.backend == 'nak' && 'NULL_NAK=1' }}" >> $GITHUB_ENV
+        run: printf "NULL=1\nNULL_ALLOW_COPYOUT=1\n${{ matrix.backend == 'ir3' && 'NULL_IR3=1' || matrix.backend == 'nak' && 'NULL_NAK=1' }}" >> $GITHUB_ENV
       - name: Run test_ops
         shell: bash
         run: |

.gitignore

@@ -58,10 +58,11 @@ weights
 *.lprof
 comgr_*
 *.pkl
+!extra/sqtt/examples/**/*.pkl
 site/
 profile_stats
 *.log
 target
 .mypy_cache
 mutants
-.mutmut-cache
+.mutmut-cache

.pre-commit-config.yaml

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

CLAUDE.md

@@ -34,33 +34,6 @@ 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.
 
-## Directory Structure
-
-```
-tinygrad/
-├── tensor.py          # Tensor class, user API
-├── device.py          # Buffer, device management
-├── dtype.py           # Data types
-├── helpers.py         # Utilities, environment vars
-├── uop/
-│   ├── ops.py         # UOp class, Ops enum, PatternMatcher
-│   ├── spec.py        # UOp type verification
-│   └── symbolic.py    # Symbolic math simplification
-├── engine/
-│   ├── schedule.py    # Schedule creation, caching
-│   ├── realize.py     # Tensor realization
-│   ├── jit.py         # JIT compilation
-│   └── memory.py      # Memory planning
-├── schedule/
-│   ├── rangeify.py    # Convert movements to ranges
-│   └── indexing.py    # Index calculations
-├── codegen/
-│   ├── kernel.py      # Kernel optimization
-│   └── uopgraph.py    # UOp graph transformations
-├── renderer/          # Code generation (CUDA, Metal, etc.)
-└── runtime/           # Device backends
-```
-
 ## Testing
 
 ```bash
@@ -79,7 +52,7 @@ VIZ=1 python -c "from tinygrad import Tensor; Tensor.ones(10).sum().realize()"
 
 ## Common Environment Variables
 
-- `DEBUG=1-4` - Increasing verbosity
+- `DEBUG=1-7` - Increasing verbosity (7 shows assembly output)
 - `VIZ=1` - Enable graph visualization
 - `SPEC=1` - Enable UOp spec verification
 - `NOOPT=1` - Disable optimizations
@@ -100,6 +73,16 @@ VIZ=1 python -c "from tinygrad import Tensor; Tensor.ones(10).sum().realize()"
 - 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
@@ -209,9 +192,12 @@ When optimizing tinygrad internals:
 
 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
+## Pattern Matching Analysis
 
-Use `TRACK_MATCH_STATS=2` to identify expensive patterns:
+**Use the right tool:**
+
+- `TRACK_MATCH_STATS=2` - **Profiling**: identify expensive patterns
+- `VIZ=-1` - **Inspection**: see all transformations, what every match pattern does, the before/after diffs
 
 ```bash
 TRACK_MATCH_STATS=2 PYTHONPATH="." python3 test/external/external_benchmark_schedule.py
@@ -225,3 +211,17 @@ Key patterns to watch (from ResNet50 benchmark):
 - `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.
+
+```bash
+# Save the trace
+VIZ=-1 python test/test_tiny.py TestTiny.test_gemm
+
+# Explore it
+./extra/viz/cli.py --help
+```
+
+## 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.

docs/developer/developer.md

@@ -13,7 +13,7 @@ There's also a [doc describing speed](../developer/speed.md)
 
 Everything in [Tensor](../tensor/index.md) is syntactic sugar around constructing a graph of [UOps](../developer/uop.md).
 
-The `UOp` graph specifies the compute in terms of low level tinygrad ops. Not all UOps will actually become realized. There's two types of UOps, base and view. base contains compute into a contiguous buffer, and view is a view (specified by a ShapeTracker). Inputs to a base can be either base or view, inputs to a view can only be a single base.
+The `UOp` graph specifies the compute in terms of low level tinygrad ops. Not all UOps will actually become realized. There's two types of UOps, base and view. base contains compute into a contiguous buffer, and view is a view. Inputs to a base can be either base or view, inputs to a view can only be a single base.
 
 ## Scheduling
 

docs/developer/layout.md

@@ -26,9 +26,9 @@ Transforms the ast into an optimized ast. This is where BEAM search and heuristi
 
 ## tinygrad/codegen
 
-Transform the optimized ast into a linearized list of UOps.
+Transform the optimized ast into a linearized and rendered program.
 
-::: tinygrad.codegen.full_rewrite
+::: tinygrad.codegen.get_program
     options:
         members: false
         show_labels: false

docs/runtime.md

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

docs/tensor/elementwise.md

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

docs/tensor/movement.md

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

docs/tensor/ops.md

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

eslint.config.mjs

@@ -1,9 +0,0 @@
-import globals from "globals";
-import pluginJs from "@eslint/js";
-import pluginHtml from "eslint-plugin-html";
-
-export default [
-  {files: ["**/*.html"], plugins: {html: pluginHtml}, rules:{"max-len": ["error", {"code": 150}]}},
-  {languageOptions: {globals: globals.browser}},
-  pluginJs.configs.recommended,
-];

examples/anthropic_challenge.py

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

examples/audio_helpers.py

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

examples/mlperf/dataloader.py

@@ -72,7 +72,7 @@ def loader_process(q_in, q_out, X:Tensor, seed):
       #storage_tensor._copyin(img_tensor.numpy())
 
       # faster
-      X[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = img.tobytes()
+      X[idx].contiguous().realize().uop.base.realized.as_memoryview(force_zero_copy=True)[:] = img.tobytes()
 
       # ideal
       #X[idx].assign(img.tobytes())   # NOTE: this is slow!
@@ -213,12 +213,13 @@ class InterleavedDataset:
     self.queues[queue_index].queue.extend(load_file(file))
 
 # Reference: https://github.com/mlcommons/training/blob/1c8a098ae3e70962a4f7422c0b0bd35ae639e357/language_model/tensorflow/bert/run_pretraining.py, Line 394
-def batch_load_train_bert(BS:int):
+def batch_load_train_bert(BS:int, seed:int|None=None):
   from extra.datasets.wikipedia import get_wiki_train_files
+  rng = random.Random(seed)
   fs = sorted(get_wiki_train_files())
   train_files = []
   while fs: # TF shuffle
-    random.shuffle(fs)
+    rng.shuffle(fs)
     train_files.append(fs.pop(0))
 
   cycle_length = min(getenv("NUM_CPU_THREADS", min(os.cpu_count(), 8)), len(train_files))
@@ -263,8 +264,8 @@ def load_unet3d_data(preprocessed_dataset_dir, seed, queue_in, queue_out, X:Tens
       x = random_brightness_augmentation(x)
       x = gaussian_noise(x)
 
-    X[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = x.tobytes()
-    Y[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = y.tobytes()
+    X[idx].contiguous().realize().uop.base.realized.as_memoryview(force_zero_copy=True)[:] = x.tobytes()
+    Y[idx].contiguous().realize().uop.base.realized.as_memoryview(force_zero_copy=True)[:] = y.tobytes()
 
     queue_out.put(idx)
   queue_out.put(None)
@@ -378,12 +379,12 @@ def load_retinanet_data(base_dir:Path, val:bool, queue_in:Queue, queue_out:Queue
       clipped_match_idxs = np.clip(match_idxs, 0, None)
       clipped_boxes, clipped_labels = tgt["boxes"][clipped_match_idxs], tgt["labels"][clipped_match_idxs]
 
-      boxes[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = clipped_boxes.tobytes()
-      labels[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = clipped_labels.tobytes()
-      matches[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = match_idxs.tobytes()
-      anchors[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = anchor.tobytes()
+      boxes[idx].contiguous().realize().uop.base.realized.as_memoryview(force_zero_copy=True)[:] = clipped_boxes.tobytes()
+      labels[idx].contiguous().realize().uop.base.realized.as_memoryview(force_zero_copy=True)[:] = clipped_labels.tobytes()
+      matches[idx].contiguous().realize().uop.base.realized.as_memoryview(force_zero_copy=True)[:] = match_idxs.tobytes()
+      anchors[idx].contiguous().realize().uop.base.realized.as_memoryview(force_zero_copy=True)[:] = anchor.tobytes()
 
-    imgs[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = img.tobytes()
+    imgs[idx].contiguous().realize().uop.base.realized.as_memoryview(force_zero_copy=True)[:] = img.tobytes()
 
     queue_out.put(idx)
   queue_out.put(None)
@@ -763,48 +764,26 @@ class BlendedGPTDataset:
 
     return dataset_idx, dataset_sample_idx
 
-def batch_load_llama3(bs:int, samples:int, seqlen:int, base_dir:Path, seed:int=0, val:bool=True):
+def get_llama3_dataset(samples:int, seqlen:int, base_dir:Path, seed:int=0, val:bool=True, small:bool=False) -> BlendedGPTDataset:
+  if small:
+    if val:
+      return BlendedGPTDataset(
+        [base_dir / "c4-validation-91205-samples.en_text_document"], [1.0], samples, seqlen, seed, shuffle=False)
+    return BlendedGPTDataset(
+      [base_dir / "c4-train.en_6_text_document"], [1.0], samples, seqlen, seed, shuffle=True)
   if val:
-    dataset = BlendedGPTDataset([
-      base_dir / "validation" / "c4-validationn-91205-samples.en_text_document",
-    ], [
-      1.0
-    ], samples, seqlen, seed, False)
-  else:
-    dataset = BlendedGPTDataset([
-      base_dir / "c4-train.en_6_text_document",
-      base_dir / "c4-train.en_7_text_document",
-    ], [
-      1.0, 1.0
-    ], samples, seqlen, seed, True)
+    return BlendedGPTDataset(
+      [base_dir / "validation" / "c4-validationn-91205-samples.en_text_document"], [1.0], samples, seqlen, seed, shuffle=False)
+  return BlendedGPTDataset(
+    [base_dir / "c4-train.en_6_text_document", base_dir / "c4-train.en_7_text_document"], [1.0, 1.0], samples, seqlen, seed, shuffle=True)
 
-  for b in range(math.ceil(samples / bs)):
-    batch = []
-    for i in range(bs):
-      tokens = dataset.get(b * bs + i)
-      batch.append(tokens)
+def iterate_llama3_dataset(dataset:BlendedGPTDataset, bs:int):
+  for b in range(math.ceil(dataset.samples / bs)):
+    batch = [dataset.get(b * bs + i) for i in range(bs)]
     yield Tensor.stack(batch, dim=0)
 
-def batch_load_llama3_small(bs:int, samples:int, seqlen:int, base_dir:Path, seed:int=0, val:bool=True):
-  if val:
-    dataset = BlendedGPTDataset([
-      base_dir / "c4-validation-91205-samples.en_text_document",
-    ], [
-      1.0
-    ], samples, seqlen, seed, False)
-  else:
-    dataset = BlendedGPTDataset([
-      base_dir / "c4-train.en_6_text_document",
-    ], [
-      1.0
-    ], samples, seqlen, seed, True)
-
-  for b in range(math.ceil(samples / bs)):
-    batch = []
-    for i in range(bs):
-      tokens = dataset.get(b * bs + i)
-      batch.append(tokens)
-    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)
 
 if __name__ == "__main__":
   def load_unet3d(val):

examples/mlperf/helpers.py

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

examples/mlperf/model_eval.py

@@ -234,12 +234,9 @@ def eval_llama3():
     loss = logits.sparse_categorical_crossentropy(tokens[:, 1:])
     return loss.flatten().float()
 
-  if SMALL:
-    from examples.mlperf.dataloader import batch_load_llama3_small
-    iter = batch_load_llama3_small(BS, 5760, SEQLEN, BASEDIR, val=True)
-  else:
-    from examples.mlperf.dataloader import batch_load_llama3
-    iter = batch_load_llama3(BS, 5760, SEQLEN, BASEDIR, val=True)
+  from examples.mlperf.dataloader import get_llama3_dataset, iterate_llama3_dataset
+  eval_dataset = get_llama3_dataset(5760, SEQLEN, BASEDIR, val=True, small=bool(SMALL))
+  iter = iterate_llama3_dataset(eval_dataset, BS)
 
   losses = []
   for tokens in tqdm(iter, total=5760//BS):

examples/mlperf/model_train.py

@@ -3,7 +3,7 @@ from pathlib import Path
 import multiprocessing
 
 from tinygrad import Device, GlobalCounters, Tensor, TinyJit, dtypes
-from tinygrad.helpers import getenv, BEAM, WINO, round_up, diskcache_clear, Profiling
+from tinygrad.helpers import getenv, BEAM, WINO, round_up, diskcache_clear, Profiling, profile_marker
 from tinygrad.nn.state import get_parameters, get_state_dict, load_state_dict, safe_load, safe_save
 from tinygrad.nn.optim import LAMB, LARS, SGD, OptimizerGroup, Adam, AdamW
 
@@ -1008,6 +1008,7 @@ def train_bert():
   config["DISABLE_DROPOUT"] = getenv("DISABLE_DROPOUT", 0)
   config["TRAIN_BEAM"]    = TRAIN_BEAM = getenv("TRAIN_BEAM", BEAM.value)
   config["EVAL_BEAM"]     = EVAL_BEAM  = getenv("EVAL_BEAM", BEAM.value)
+  config["FP8_TRAIN"]     = getenv("FP8_TRAIN", 0)
 
   Tensor.manual_seed(seed)  # seed for weight initialization
 
@@ -1085,7 +1086,7 @@ def train_bert():
   if RUNMLPERF:
     # only load real data with RUNMLPERF
     eval_it = iter(batch_load_val_bert(EVAL_BS))
-    train_it = iter(tqdm(batch_load_train_bert(BS), total=train_steps, disable=BENCHMARK))
+    train_it = iter(tqdm(batch_load_train_bert(BS, seed=seed), total=train_steps, disable=BENCHMARK))
     for _ in range(start_step): next(train_it) # Fast forward
   else:
     # repeat fake data
@@ -1147,7 +1148,7 @@ def train_bert():
 
         device_str = parameters[0].device if isinstance(parameters[0].device, str) else f"{parameters[0].device[0]} * {len(parameters[0].device)}"
         loss = loss.item()
-        assert not math.isnan(loss)
+        if not getenv("FP8_TRAIN"): assert not math.isnan(loss)
         lr = lr.item()
 
       cl = time.perf_counter()
@@ -1160,7 +1161,7 @@ def train_bert():
       if WANDB:
         wandb.log({"lr": lr, "train/loss": loss, "train/global_norm": global_norm.item(), "train/step_time": cl - st,
                     "train/python_time": pt - st, "train/data_time": dt - pt, "train/cl_time": cl - dt,
-                    "train/GFLOPS": GlobalCounters.global_ops * 1e-9 / (cl - st), "epoch": (i+1)*GBS})
+                    "train/mem":GlobalCounters.mem_used / 1e9, "train/GFLOPS": GlobalCounters.global_ops * 1e-9 / (cl - st), "epoch": (i+1)*GBS})
 
       train_data, next_data = next_data, None
       i += 1
@@ -1285,17 +1286,23 @@ def train_llama3():
   from examples.llama3 import MODEL_PARAMS
   from examples.mlperf.lr_schedulers import CosineAnnealingLRWithWarmup
 
+  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)
   SEQLEN             = config["SEQLEN"]                 = getenv("SEQLEN", 8192)
   TRAIN_ON_VAL       = config["TRAIN_ON_VAL"]           = getenv("TRAIN_ON_VAL", 0)
   SMALL              = config["SMALL"]                  = getenv("SMALL", 0)
   SAMPLES            = config["SAMPLES"]                = getenv("SAMPLES", 5_760 if TRAIN_ON_VAL else 1_200_000 * 1152)
+  EVAL_SAMPLES       = config["EVAL_SAMPLES"]           = getenv("EVAL_SAMPLES", 5760 if not SMALL else 1024)
+  MAX_STEPS          = config["MAX_STEPS"]              = getenv("MAX_STEPS", math.ceil(1_200_000 * 1152 / GBS))
+  WARMUP_STEPS       = config["WARMUP_STEPS"]           = getenv("WARMUP_STEPS", math.ceil(8000 * 1152 / GBS))
+  LR                 = config["LR"]                     = getenv("LR", 8e-5 * GBS / 1152)
+  END_LR             = config["END_LR"]                 = getenv("END_LR", 8e-7)
   EVAL_FREQ          = config["EVAL_FREQ"]              = getenv("EVAL_FREQ", 46080)
   EVAL_BS            = config["EVAL_BS"]                = getenv("EVAL_BS", 16)
   EVAL_TARGET        = config["EVAL_TARGET"]            = getenv("EVAL_TARGET", 5.6)
@@ -1309,15 +1316,26 @@ def train_llama3():
   opt_adamw_weight_decay = 0.1
 
   opt_gradient_clip_norm = 1.0
-  opt_learning_rate_warmup_steps = getenv("WARMUP_STEPS", math.ceil(8000 * 1152 / GBS))
-  opt_learning_rate_decay_steps = getenv("MAX_STEPS", math.ceil(1_200_000 * 1152 / GBS)) - opt_learning_rate_warmup_steps
-  opt_base_learning_rate = getenv("LR", 8e-5 * GBS / 1152)  # NOTE: cannot change for benchmark
-  opt_end_learning_rate = getenv("END_LR", 8e-7)
+  opt_learning_rate_warmup_steps = WARMUP_STEPS
+  opt_learning_rate_decay_steps = MAX_STEPS - opt_learning_rate_warmup_steps
+  opt_base_learning_rate = LR
+  opt_end_learning_rate = END_LR
+
+  Tensor.manual_seed(SEED)  # seed for weight initialization
+
+  # ** init wandb **
+  WANDB = getenv("WANDB")
+  if WANDB:
+    import wandb
+    wandb_args = {"id": wandb_id, "resume": "must"} if (wandb_id := getenv("WANDB_RESUME", "")) else {}
+    wandb.init(config=config, **wandb_args, project="MLPerf-LLaMA3")
 
   model_params = MODEL_PARAMS[getenv("LLAMA3_SIZE", "8B")]["args"]
   # vocab_size from the mixtral tokenizer
   if not SMALL: model_params |= {"vocab_size": 32000}
   if (llama_layers:=getenv("LLAMA_LAYERS")) != 0: model_params['n_layers'] = llama_layers
+  print(f"model parameters: {model_params}")
+
   model = Transformer(**model_params, max_context=SEQLEN, jit=False, disable_kv_cache=True)
   params = get_parameters(model)
   # weights are all bfloat16 for now
@@ -1353,6 +1371,12 @@ def train_llama3():
 
   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)
+
+  # init grads
+  for p in optim.params:
+    p.grad = p.zeros_like().contiguous().realize()
+  grads = [p.grad for p in optim.params]
+
   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"):
@@ -1365,9 +1389,7 @@ def train_llama3():
     load_state_dict(scheduler, safe_load(fn), realize=False)
 
   @TinyJit
-  @Tensor.train()
-  def train_step(model, tokens:Tensor):
-    optim.zero_grad()
+  def minibatch(tokens:Tensor):
     if (DP := getenv("DP", 1)) > 1:
       device = tuple(f"{Device.DEFAULT}:{i}" for i in range(DP))
       tokens = tokens.shard(device, 0)
@@ -1377,27 +1399,40 @@ def train_llama3():
     logits:Tensor = model(tokens[:, :-1], start_pos=0, temperature=math.nan)
     loss = logits.sparse_categorical_crossentropy(tokens[:, 1:])
     loss.backward()
+    assert all(p.grad is g for p,g in zip(optim.params, grads))
+    Tensor.realize(loss, *grads)
+    return loss
+
+  @TinyJit
+  def optim_step():
+    for p in optim.params:
+      p.grad.assign(p.grad / grad_acc)
+
     # 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 g in grads:
+        total_norm += g.float().square().sum()
+      total_norm = total_norm.sqrt().contiguous().realize()
+      for g in grads:
+        g.assign((g * (opt_gradient_clip_norm / (total_norm + 1e-6)).clamp(max_=1.0)).cast(g.dtype)).realize()
 
     optim.step()
     scheduler.step()
 
+    for g in grads:
+      g.assign(g.zeros_like().contiguous()).realize()
+
     lr = optim.lr
-    loss.realize(lr)
-    return loss, lr
+    Tensor.realize(lr, *grads)
+
+    return lr
 
   @TinyJit
   @Tensor.train(False)
-  def eval_step(model, tokens:Tensor):
+  def eval_step(tokens:Tensor):
     if (DP := getenv("DP", 1)) > 1:
       device = tuple(f"{Device.DEFAULT}:{i}" for i in range(DP))
       tokens = tokens.shard(device, 0)
@@ -1417,64 +1452,121 @@ def train_llama3():
     if getenv("FAKEDATA", 0):
       return fake_data(BS, SAMPLES)
     else:
-      if SMALL:
-        from examples.mlperf.dataloader import batch_load_llama3_small
-        return batch_load_llama3_small(BS, SAMPLES, SEQLEN, BASEDIR, seed=SEED, val=bool(TRAIN_ON_VAL))
-      else:
-        from examples.mlperf.dataloader import batch_load_llama3
-        return batch_load_llama3(BS, SAMPLES, SEQLEN, BASEDIR, seed=SEED, val=bool(TRAIN_ON_VAL))
+      from examples.mlperf.dataloader import batch_load_llama3
+      return batch_load_llama3(BS, SAMPLES, SEQLEN, BASEDIR, seed=SEED, val=bool(TRAIN_ON_VAL), small=bool(SMALL))
+
+  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))
 
   def get_eval_iter():
-    if getenv("FAKEDATA", 0):
-      return fake_data(EVAL_BS, 5760)
-    else:
-      if SMALL:
-        from examples.mlperf.dataloader import batch_load_llama3_small
-        return batch_load_llama3_small(EVAL_BS, 5760, SEQLEN, BASEDIR, val=True)
-      else:
-        from examples.mlperf.dataloader import batch_load_llama3
-        return batch_load_llama3(EVAL_BS, 5760, SEQLEN, BASEDIR, val=True)
+    if eval_dataset is None:
+      return fake_data(EVAL_BS, EVAL_SAMPLES)
+    from examples.mlperf.dataloader import iterate_llama3_dataset
+    return iterate_llama3_dataset(eval_dataset, EVAL_BS)
 
-  iter = get_train_iter()
+  num_params = sum(p.numel() for p in params) - model_params["vocab_size"]*model_params["dim"]
+  train_iter = get_train_iter()
   i, sequences_seen = resume_ckpt, 0
-  for tokens in tqdm(iter, total=SAMPLES//GBS):
-    t = time.perf_counter()
+  step_times = []
+  while i < MAX_STEPS:
     GlobalCounters.reset()
-    loss, lr = train_step(model, tokens)
-    loss = loss.float().item()
+    if getenv("TRAIN", 1):
+      profile_marker(f"train @ {i}")
+      st = time.perf_counter()
 
-    i += 1
-    sequences_seen += tokens.shape[0]
+      stopped = False
+      for _ in range(grad_acc):
+        ist = time.perf_counter()
+        try: tokens = next(train_iter)
+        except StopIteration:
+          stopped = True
+          break
+        dt = time.perf_counter()
+        loss = minibatch(tokens)
+      if stopped: break
 
-    tqdm.write(f"{loss:.4f} loss, {lr.item():.12f} LR, {GlobalCounters.mem_used / 1e9:.2f} GB used, {time.perf_counter()-t:.2f} s")
-    if (fname:=getenv("LOSS_FILE", "")):
-      with open(fname, "a") as f:
-        f.write(f"{i} {loss:.4f} {lr.item():.12f} {GlobalCounters.mem_used / 1e9:.2f}\n")
+      gt = time.perf_counter()
+      lr = optim_step()
+      ot = time.perf_counter()
 
-    if (ckpt_freq := getenv("CKPT")) and (i % ckpt_freq == 0 and (i != 1 or ckpt_freq == 1)):
-      tqdm.write("saving checkpoint")
-      if not os.path.exists(ckpt_dir := "./ckpts"): os.mkdir(ckpt_dir)
-      fn = f"{ckpt_dir}/llama3_{i}.safe"
-      safe_save(get_state_dict(model), fn)
+      loss = loss.float().item()
+      lr = lr.item()
 
-      tqdm.write("saving optim checkpoint")
-      fn = f"{ckpt_dir}/llama3_{i}_optim.safe"
-      safe_save(get_state_dict(scheduler), fn)
+      et = time.perf_counter()
+      step_time = et - st
+      gbs_time = gt - st
+      optim_time = ot - gt
+      data_time = dt - ist
+      dev_time = step_time - data_time * grad_acc
+      if BENCHMARK: step_times.append(step_time)
 
-    if sequences_seen % EVAL_FREQ == 0 and (i != 1 or EVAL_FREQ == 1):
+      i += 1
+      sequences_seen += GBS
+
+      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
+      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, {mem_gb:.2f} GB used, {gflops:9.2f} GFLOPS, {mfu:5.2f}% MFU")
+
+      if WANDB:
+        wandb.log({
+          "lr": lr, "train/loss": loss,
+          "train/step_time": step_time,
+          "train/gbs_time": gbs_time,
+          "train/optim_time": optim_time,
+          "train/dev_time": dev_time,
+          "train/data_time": data_time,
+          "train/mem": mem_gb,
+          "train/GFLOPS": gflops,
+          "train/MFU": mfu,
+          "train/sequences_seen": sequences_seen
+        })
+
+      if (ckpt_freq := getenv("CKPT")) and (i % ckpt_freq == 0 and (i != 1 or ckpt_freq == 1)):
+        tqdm.write("saving checkpoint")
+        if not os.path.exists(ckpt_dir := "./ckpts"): os.mkdir(ckpt_dir)
+        fn = f"{ckpt_dir}/llama3_{i}.safe"
+        safe_save(get_state_dict(model), fn)
+
+        tqdm.write("saving optim checkpoint")
+        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 == 0 and (i != 1 or EVAL_FREQ == 1)) or (BENCHMARK and i == BENCHMARK):
+      if EVAL_BS == 0: return
       tqdm.write(f"evaluating after {sequences_seen} sequences")
+      profile_marker(f"eval @ {i}")
 
       # run eval
       eval_losses = []
       eval_iter = get_eval_iter()
       tqdm.write(f"evaluating {5760//EVAL_BS} batches of {EVAL_BS} sequences")
 
-      for tokens in tqdm(eval_iter, total=5760//EVAL_BS):
-        eval_losses += eval_step(model, tokens).tolist()
+      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
+
       log_perplexity = Tensor(eval_losses).mean().float().item()
 
       tqdm.write(f"eval log perplexity: {log_perplexity:.4f}")
 
+      if WANDB:
+        wandb.log({"eval/log_perplexity": log_perplexity, "eval/sequences_seen": sequences_seen})
+
       if log_perplexity < EVAL_TARGET:
         tqdm.write(f"target achieved after {sequences_seen} sequences")
         if getenv("CKPT"):
@@ -1553,7 +1645,7 @@ 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():
@@ -1592,7 +1684,7 @@ 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/training_submission_v6.0/tinycorp/benchmarks/bert/implementations/tinybox_1xMI300X/dev_beam.sh

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

@@ -0,0 +1,34 @@
+#!/usr/bin/env bash
+
+export PYTHONPATH="."
+export DEV=${DEV:-AMD}
+export CHECK_OOB=0
+export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
+
+export DEBUG=${DEBUG:-2}
+export FLASH_ATTENTION=${FLASH_ATTENTION:-1}
+export ALL2ALL=${ALL2ALL:-1}
+export USE_ATOMICS=${USE_ATOMICS:-1}
+export ASM_GEMM=${ASM_GEMM:-1}
+
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
+export DP=8 BS=16 EVAL_BS=8 GRADIENT_ACC_STEPS=1
+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="2.5e-4" END_LR="2.5e-5" WARMUP_SAMPLES=256 MAX_STEPS=1200000
+export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
+export SAMPLES=$((MAX_STEPS * GBS))
+
+export SEED=5760
+
+export JITBEAM=3
+export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
+
+export FAKEDATA=1 BENCHMARK=10 LLAMA_LAYERS=2
+
+python3 examples/mlperf/model_train.py

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

@@ -0,0 +1,33 @@
+#!/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 FLASH_ATTENTION=${FLASH_ATTENTION:-1}
+export ALL2ALL=${ALL2ALL:-1}
+export USE_ATOMICS=${USE_ATOMICS:-1}
+export ASM_GEMM=${ASM_GEMM:-1}
+
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
+export DP=${DP:-8} BS=${BS:-16} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-1}
+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="2.5e-4" END_LR="2.5e-5" WARMUP_SAMPLES=256 MAX_STEPS=1200000
+export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
+export SAMPLES=$((MAX_STEPS * GBS))
+
+export SEED=${SEED:-5760}
+
+export JITBEAM=${JITBEAM:-3}
+export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
+
+python3 examples/mlperf/model_train.py

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

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

examples/stable_diffusion.py

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

examples/test_pkl_imagenet.py

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

examples/tinychat/tinychat-browser/compile.py

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

examples/webgpu/stable_diffusion/compile.py

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

examples/whisper.py

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

extra/amdpci/am_smi.py

@@ -153,8 +153,7 @@ class SMICtx:
     tables = {}
     for dev in self.devs:
       match dev.ip_ver[am.MP1_HWIP]:
-        case (13,0,6): table_t = dev.smu.smu_mod.MetricsTableX_t
-        case (13,0,12): table_t = dev.smu.smu_mod.MetricsTableV2_t
+        case (13,0,6)|(13,0,12): table_t = dev.smu.smu_mod.MetricsTableX_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 +164,17 @@ class SMICtx:
 
   def get_gfx_activity(self, dev, metrics):
     match dev.ip_ver[am.MP1_HWIP]:
-      case (13,0,6): return max(0, min(100, self._smuq10_round(metrics.SocketGfxBusy)))
+      case (13,0,6)|(13,0,12): return max(0, min(100, self._smuq10_round(metrics.SocketGfxBusy)))
       case _: return metrics.SmuMetrics.AverageGfxActivity
 
   def get_mem_activity(self, dev, metrics):
     match dev.ip_ver[am.MP1_HWIP]:
-      case (13,0,6): return max(0, min(100, self._smuq10_round(metrics.DramBandwidthUtilization)))
+      case (13,0,6)|(13,0,12): return max(0, min(100, self._smuq10_round(metrics.DramBandwidthUtilization)))
       case _: return metrics.SmuMetrics.AverageUclkActivity
 
   def get_temps(self, dev, metrics, compact=False):
     match dev.ip_ver[am.MP1_HWIP]:
-      case (13,0,6):
+      case (13,0,6)|(13,0,12):
         temps = {
           "Hotspot": self._smuq10_round(metrics.MaxSocketTemperature),
           "HBM": self._smuq10_round(metrics.MaxHbmTemperature),
@@ -184,16 +183,16 @@ class SMICtx:
         if compact: return {k: temps[k] for k in ("Hotspot", "HBM") if temps.get(k, 0) != 0}
         return {k: v for k, v in temps.items() if v != 0}
       case _:
-        temps_keys = [(k, name) for k, name in dev.smu.smu_mod.c__EA_TEMP_e__enumvalues.items()
+        temps_keys = [(k, name) for k, name in dev.smu.smu_mod.TEMP_e.items()
                       if k < dev.smu.smu_mod.TEMP_COUNT and metrics.SmuMetrics.AvgTemperature[k] != 0]
         if compact: temps_keys = [(k, name) for k, name in temps_keys if k in (dev.smu.smu_mod.TEMP_HOTSPOT, dev.smu.smu_mod.TEMP_MEM)]
         return {name: metrics.SmuMetrics.AvgTemperature[k] for k, name in temps_keys}
 
   def get_voltage(self, dev, metrics, compact=False):
     match dev.ip_ver[am.MP1_HWIP]:
-      case (13,0,6): return {}
+      case (13,0,6)|(13,0,12): return {}
       case _:
-        voltage_keys = [(k, name) for k, name in dev.smu.smu_mod.c__EA_SVI_PLANE_e__enumvalues.items()
+        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]
         return {name: metrics.SmuMetrics.AvgVoltage[k] for k, name in voltage_keys}
 
@@ -205,33 +204,33 @@ class SMICtx:
   def get_gfx_freq(self, dev, metrics):
     if metrics is None: return 0
     match dev.ip_ver[am.MP1_HWIP]:
-      case (13,0,6): return self._smuq10_round(metrics.GfxclkFrequency[0])
+      case (13,0,6)|(13,0,12): return self._smuq10_round(metrics.GfxclkFrequency[0])
       case _:
         return metrics.SmuMetrics.AverageGfxclkFrequencyPostDs if self.get_gfx_activity(dev, metrics) <= self.get_busy_threshold(dev) else \
                metrics.SmuMetrics.AverageGfxclkFrequencyPreDs
 
   def get_mem_freq(self, dev, metrics):
     match dev.ip_ver[am.MP1_HWIP]:
-      case (13,0,6): return self._smuq10_round(metrics.UclkFrequency)
+      case (13,0,6)|(13,0,12): return self._smuq10_round(metrics.UclkFrequency)
       case _:
         return metrics.SmuMetrics.AverageMemclkFrequencyPostDs if self.get_mem_activity(dev, metrics) <= self.get_busy_threshold(dev) else \
                metrics.SmuMetrics.AverageMemclkFrequencyPreDs
 
   def get_fckl_freq(self, dev, metrics):
     match dev.ip_ver[am.MP1_HWIP]:
-      case (13,0,6): return self._smuq10_round(metrics.FclkFrequency)
+      case (13,0,6)|(13,0,12): return self._smuq10_round(metrics.FclkFrequency)
       case _:
         return metrics.SmuMetrics.AverageFclkFrequencyPostDs if self.get_mem_activity(dev, metrics) <= self.get_busy_threshold(dev) else \
                metrics.SmuMetrics.AverageFclkFrequencyPreDs
 
   def get_fan_rpm_pwm(self, dev, metrics):
     match dev.ip_ver[am.MP1_HWIP]:
-      case (13,0,6): return None, None
+      case (13,0,6)|(13,0,12): return None, None
       case _: return metrics.SmuMetrics.AvgFanRpm, metrics.SmuMetrics.AvgFanPwm
 
   def get_power(self, dev, metrics):
     match dev.ip_ver[am.MP1_HWIP]:
-      case (13,0,6): return self._smuq10_round(metrics.SocketPower), self._smuq10_round(metrics.MaxSocketPowerLimit)
+      case (13,0,6)|(13,0,12): return self._smuq10_round(metrics.SocketPower), self._smuq10_round(metrics.MaxSocketPowerLimit)
       case _: return metrics.SmuMetrics.AverageSocketPower, metrics.SmuMetrics.dGPU_W_MAX
 
   def get_mem_usage(self, dev):

extra/amdpci/proclogs.py

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

extra/assembly/amd/README

@@ -0,0 +1,39 @@
+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 pseudocode from AMD PDF
+* dsl.py -- helpers for the autogen instruction classes in `__init__.py`. should be standalone with init
+* pcode.py -- pseudocode execution environment. pseudocode 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.
+
+The autogen folder is autogenerated from the AMD PDFs with `python3 -m extra.assembly.amd.pdf --arch all`
+
+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
+If a test is failing with `AMD=1 PYTHON_REMU=1 MOCKGPU=1` it's 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.
+IMPORTANT: if a test is failing in the emulator, it's an instruction bug. Use DEBUG=7, get the instructions, and debug.
+
+Currently, only RDNA3 is well supported, but when finished, this will support RDNA3+RDNA4+CDNA in ~2000 lines.
+Get line count with `cloc --by-file extra/assembly/amd/*.py`
+

extra/assembly/amd/__init__.py

@@ -0,0 +1,67 @@
+# Instruction format detection and decoding
+from __future__ import annotations
+from extra.assembly.amd.dsl import Inst, FixedBitField, EnumBitField
+
+# SDWA/DPP variant detection: src0 field (bits 0-8) encodes the variant
+# 0xf9 (249) = SDWA, 0xfa (250) = DPP16 for CDNA (GFX9)
+_VARIANT_SRC0 = {"_SDWA_SDST": 0xf9, "_SDWA": 0xf9, "_DPP16": 0xfa}
+
+def _matches(data: bytes, cls: type[Inst]) -> bool:
+  """Check if data matches all FixedBitFields and op is in allowed."""
+  for _, field in cls._fields:
+    dword_idx = field.lo // 32
+    if len(data) < (dword_idx + 1) * 4: return False
+    word = int.from_bytes(data[dword_idx*4:(dword_idx+1)*4], 'little')
+    field_lo = field.lo % 32
+    if isinstance(field, FixedBitField):
+      if ((word >> field_lo) & field.mask) != field.default: return False
+    if isinstance(field, EnumBitField) and field.allowed is not None:
+      try: opcode = field.decode((word >> field_lo) & field.mask)
+      except ValueError: return False  # opcode not in enum
+      if opcode not in field.allowed: return False
+  # Check SDWA/DPP variant based on src0 field (bits 0-8) - only for variant classes
+  name = cls.__name__
+  word = int.from_bytes(data[:4], 'little')
+  for suffix, expected_src0 in _VARIANT_SRC0.items():
+    if name.endswith(suffix): return (word & 0x1ff) == expected_src0
+  return True
+
+# Import instruction classes for each architecture
+from extra.assembly.amd.autogen.rdna3.ins import (VOP1, VOP1_SDST, VOP1_LIT, VOP2, VOP2_LIT, VOP3, VOP3_SDST, VOP3SD, VOP3P, VOPC, VOPD, VINTERP,
+  SOP1, SOP1_LIT, SOP2, SOP2_LIT, SOPC, SOPK, SOPK_LIT, SOPP, SMEM, DS, FLAT, GLOBAL, SCRATCH)
+from extra.assembly.amd.autogen.rdna4.ins import (VOP1 as R4_VOP1, VOP1_SDST as R4_VOP1_SDST, VOP1_LIT as R4_VOP1_LIT,
+  VOP2 as R4_VOP2, VOP2_LIT as R4_VOP2_LIT, VOP3 as R4_VOP3, VOP3_SDST as R4_VOP3_SDST, VOP3SD as R4_VOP3SD, VOP3P as R4_VOP3P,
+  VOPC as R4_VOPC, VOPD as R4_VOPD, VINTERP as R4_VINTERP, SOP1 as R4_SOP1, SOP1_LIT as R4_SOP1_LIT,
+  SOP2 as R4_SOP2, SOP2_LIT as R4_SOP2_LIT, SOPC as R4_SOPC, SOPC_LIT as R4_SOPC_LIT,
+  SOPK as R4_SOPK, SOPK_LIT as R4_SOPK_LIT, SOPP as R4_SOPP,
+  SMEM as R4_SMEM, DS as R4_DS, VFLAT as R4_FLAT, VGLOBAL as R4_GLOBAL, VSCRATCH as R4_SCRATCH)
+from extra.assembly.amd.autogen.cdna.ins import (VOP1 as C_VOP1, VOP1_SDWA as C_VOP1_SDWA, VOP1_DPP16 as C_VOP1_DPP16,
+  VOP2 as C_VOP2, VOP2_LIT as C_VOP2_LIT, VOP2_SDWA as C_VOP2_SDWA, VOP2_DPP16 as C_VOP2_DPP16,
+  VOPC as C_VOPC, VOPC_SDWA_SDST as C_VOPC_SDWA_SDST,
+  VOP3 as C_VOP3, VOP3_SDST as C_VOP3_SDST, VOP3SD as C_VOP3SD, VOP3P as C_VOP3P, VOP3P_MFMA as C_VOP3P_MFMA, VOP3PX2 as C_VOP3PX2,
+  SOP1 as C_SOP1, SOP2 as C_SOP2, SOPC as C_SOPC, SOPK as C_SOPK, SOPK_LIT as C_SOPK_LIT, SOPP as C_SOPP, SMEM as C_SMEM, DS as C_DS,
+  FLAT as C_FLAT, GLOBAL as C_GLOBAL, SCRATCH as C_SCRATCH, MUBUF as C_MUBUF)
+
+# Order matters: more specific encodings first, catch-alls (SOP2, VOP2) last
+# Order: base before _LIT (base matches regular ops, _LIT catches lit-only ops excluded from base)
+_FORMATS = {
+  "rdna3": [VOPD, VOP3P, VINTERP, VOP3SD, VOP3_SDST, VOP3, DS, GLOBAL, SCRATCH, FLAT, SMEM,
+            SOP1, SOP1_LIT, SOP2, SOP2_LIT, SOPC, SOPK, SOPK_LIT, SOPP, VOPC, VOP1_SDST, VOP1, VOP1_LIT, VOP2, VOP2_LIT],
+  "rdna4": [R4_VOPD, R4_VOP3P, R4_VINTERP, R4_VOP3SD, R4_VOP3_SDST, R4_VOP3, R4_DS, R4_GLOBAL, R4_SCRATCH, R4_FLAT, R4_SMEM,
+            R4_SOP1, R4_SOP1_LIT, R4_SOPC, R4_SOPC_LIT, R4_SOPP, R4_SOPK, R4_SOPK_LIT, R4_VOPC, R4_VOP1_SDST, R4_VOP1, R4_VOP1_LIT,
+            R4_SOP2, R4_SOP2_LIT, R4_VOP2, R4_VOP2_LIT],
+  "cdna": [C_VOP3PX2, C_VOP3P_MFMA, C_VOP3P, C_VOP3SD, C_VOP3_SDST, C_VOP3, C_DS, C_GLOBAL, C_SCRATCH, C_FLAT, C_MUBUF, C_SMEM,
+            C_SOP1, C_SOPC, C_SOPP, C_SOPK, C_SOPK_LIT, C_VOPC_SDWA_SDST, C_VOPC,
+            C_VOP1_DPP16, C_VOP1_SDWA, C_VOP1, C_VOP2_DPP16, C_VOP2_SDWA, C_SOP2, C_VOP2, C_VOP2_LIT],
+}
+
+def detect_format(data: bytes, arch: str = "rdna3") -> type[Inst]:
+  """Detect instruction format from machine code bytes."""
+  assert len(data) >= 4, f"need at least 4 bytes, got {len(data)}"
+  for cls in _FORMATS[arch]:
+    if _matches(data, cls): return cls
+  raise ValueError(f"unknown {arch} format word={int.from_bytes(data[:4], 'little'):#010x}")
+
+def decode_inst(data: bytes, arch: str = "rdna3") -> Inst:
+  """Decode machine code bytes into an instruction."""
+  return detect_format(data, arch).from_bytes(data)

extra/assembly/amd/autogen/common.py

@@ -0,0 +1,324 @@
+# autogenerated from AMD ISA XML - do not edit
+from enum import Enum, auto
+
+class ReprEnum(Enum):
+  """Enum with clean repr that roundtrips with eval()."""
+  def __repr__(self): return f"{type(self).__name__}.{self.name}"
+
+class Fmt(Enum):
+  FMT_ANY = auto()
+  FMT_BUF = auto()
+  FMT_IMG = auto()
+  FMT_IMG_BVH = auto()
+  FMT_NUM_B1 = auto()
+  FMT_NUM_B1024 = auto()
+  FMT_NUM_B128 = auto()
+  FMT_NUM_B16 = auto()
+  FMT_NUM_B256 = auto()
+  FMT_NUM_B32 = auto()
+  FMT_NUM_B512 = auto()
+  FMT_NUM_B64 = auto()
+  FMT_NUM_B8 = auto()
+  FMT_NUM_B96 = auto()
+  FMT_NUM_BF16 = auto()
+  FMT_NUM_BF6 = auto()
+  FMT_NUM_BF8 = auto()
+  FMT_NUM_F16 = auto()
+  FMT_NUM_F32 = auto()
+  FMT_NUM_F64 = auto()
+  FMT_NUM_FP4 = auto()
+  FMT_NUM_FP6 = auto()
+  FMT_NUM_FP8 = auto()
+  FMT_NUM_I16 = auto()
+  FMT_NUM_I24 = auto()
+  FMT_NUM_I32 = auto()
+  FMT_NUM_I4 = auto()
+  FMT_NUM_I64 = auto()
+  FMT_NUM_I8 = auto()
+  FMT_NUM_IU4 = auto()
+  FMT_NUM_IU8 = auto()
+  FMT_NUM_M64 = auto()
+  FMT_NUM_PK16_BF16 = auto()
+  FMT_NUM_PK16_BF8 = auto()
+  FMT_NUM_PK16_F16 = auto()
+  FMT_NUM_PK16_F32 = auto()
+  FMT_NUM_PK16_FP8 = auto()
+  FMT_NUM_PK16_I32 = auto()
+  FMT_NUM_PK16_I8 = auto()
+  FMT_NUM_PK2_B16 = auto()
+  FMT_NUM_PK2_B32 = auto()
+  FMT_NUM_PK2_B64 = auto()
+  FMT_NUM_PK2_BF16 = auto()
+  FMT_NUM_PK2_BF8 = auto()
+  FMT_NUM_PK2_F16 = auto()
+  FMT_NUM_PK2_F32 = auto()
+  FMT_NUM_PK2_FP4 = auto()
+  FMT_NUM_PK2_FP8 = auto()
+  FMT_NUM_PK2_I16 = auto()
+  FMT_NUM_PK2_I8 = auto()
+  FMT_NUM_PK2_U16 = auto()
+  FMT_NUM_PK2_U8 = auto()
+  FMT_NUM_PK32_BF16 = auto()
+  FMT_NUM_PK32_BF6 = auto()
+  FMT_NUM_PK32_BF8 = auto()
+  FMT_NUM_PK32_F16 = auto()
+  FMT_NUM_PK32_F32 = auto()
+  FMT_NUM_PK32_FP6 = auto()
+  FMT_NUM_PK32_FP8 = auto()
+  FMT_NUM_PK32_I32 = auto()
+  FMT_NUM_PK32_I8 = auto()
+  FMT_NUM_PK4_B8 = auto()
+  FMT_NUM_PK4_BF16 = auto()
+  FMT_NUM_PK4_BF8 = auto()
+  FMT_NUM_PK4_F16 = auto()
+  FMT_NUM_PK4_F32 = auto()
+  FMT_NUM_PK4_F64 = auto()
+  FMT_NUM_PK4_FP8 = auto()
+  FMT_NUM_PK4_I32 = auto()
+  FMT_NUM_PK4_I8 = auto()
+  FMT_NUM_PK4_IU8 = auto()
+  FMT_NUM_PK4_U8 = auto()
+  FMT_NUM_PK8_B32 = auto()
+  FMT_NUM_PK8_BF16 = auto()
+  FMT_NUM_PK8_BF8 = auto()
+  FMT_NUM_PK8_F16 = auto()
+  FMT_NUM_PK8_FP8 = auto()
+  FMT_NUM_PK8_I4 = auto()
+  FMT_NUM_PK8_I8 = auto()
+  FMT_NUM_PK8_IU4 = auto()
+  FMT_NUM_PK8_U4 = auto()
+  FMT_NUM_PK8_U8 = auto()
+  FMT_NUM_PK_F16 = auto()
+  FMT_NUM_PK_I16 = auto()
+  FMT_NUM_PK_I8 = auto()
+  FMT_NUM_PK_U16 = auto()
+  FMT_NUM_PK_U8 = auto()
+  FMT_NUM_U16 = auto()
+  FMT_NUM_U24 = auto()
+  FMT_NUM_U32 = auto()
+  FMT_NUM_U4 = auto()
+  FMT_NUM_U64 = auto()
+  FMT_NUM_U8 = auto()
+  FMT_RSRC = auto()
+  FMT_RSRC_SCALAR = auto()
+  FMT_RSRC_SCRATCH = auto()
+  FMT_RSRC_SCRATCH_BYTE = auto()
+  FMT_RSRC_SCRATCH_STRIDE = auto()
+  FMT_RSRC_TYPED = auto()
+  FMT_RSRC_TYPED_BYTE = auto()
+  FMT_RSRC_TYPED_SCRATCH = auto()
+  FMT_RSRC_TYPED_STRIDE = auto()
+  FMT_RSRC_VECTOR = auto()
+  FMT_RSRC_VECTOR_BYTE = auto()
+  FMT_RSRC_VECTOR_STRIDE = auto()
+  FMT_SAMP = auto()
+  FMT_WMMA_AB_16X16_BF16 = auto()
+  FMT_WMMA_AB_16X16_BF8 = auto()
+  FMT_WMMA_AB_16X16_F16 = auto()
+  FMT_WMMA_AB_16X16_FP8 = auto()
+  FMT_WMMA_AB_16X16_IU4 = auto()
+  FMT_WMMA_AB_16X16_IU8 = auto()
+  FMT_WMMA_AB_16X32_BF16 = auto()
+  FMT_WMMA_AB_16X32_BF8 = auto()
+  FMT_WMMA_AB_16X32_F16 = auto()
+  FMT_WMMA_AB_16X32_FP8 = auto()
+  FMT_WMMA_AB_16X32_IU4 = auto()
+  FMT_WMMA_AB_16X32_IU8 = auto()
+  FMT_WMMA_AB_16X64_IU4 = auto()
+  FMT_WMMA_AB_BF16 = auto()
+  FMT_WMMA_AB_F16 = auto()
+  FMT_WMMA_AB_IU4 = auto()
+  FMT_WMMA_AB_IU8 = auto()
+  FMT_WMMA_DC_16X16_BF16 = auto()
+  FMT_WMMA_DC_16X16_F16 = auto()
+  FMT_WMMA_DC_16X16_F32 = auto()
+  FMT_WMMA_DC_16X16_I32 = auto()
+  FMT_WMMA_DC_BF16 = auto()
+  FMT_WMMA_DC_F16 = auto()
+  FMT_WMMA_DC_F32 = auto()
+  FMT_WMMA_DC_I32 = auto()
+  FMT_WMMA_INDEX_SET = auto()
+
+FMT_BITS = {
+  Fmt.FMT_ANY: 1,
+  Fmt.FMT_BUF: 64,
+  Fmt.FMT_IMG: 256,
+  Fmt.FMT_IMG_BVH: 128,
+  Fmt.FMT_NUM_B1: 1,
+  Fmt.FMT_NUM_B1024: 1024,
+  Fmt.FMT_NUM_B128: 128,
+  Fmt.FMT_NUM_B16: 16,
+  Fmt.FMT_NUM_B256: 256,
+  Fmt.FMT_NUM_B32: 32,
+  Fmt.FMT_NUM_B512: 512,
+  Fmt.FMT_NUM_B64: 64,
+  Fmt.FMT_NUM_B8: 8,
+  Fmt.FMT_NUM_B96: 96,
+  Fmt.FMT_NUM_BF16: 16,
+  Fmt.FMT_NUM_BF6: 6,
+  Fmt.FMT_NUM_BF8: 8,
+  Fmt.FMT_NUM_F16: 16,
+  Fmt.FMT_NUM_F32: 32,
+  Fmt.FMT_NUM_F64: 64,
+  Fmt.FMT_NUM_FP4: 4,
+  Fmt.FMT_NUM_FP6: 6,
+  Fmt.FMT_NUM_FP8: 8,
+  Fmt.FMT_NUM_I16: 16,
+  Fmt.FMT_NUM_I24: 24,
+  Fmt.FMT_NUM_I32: 32,
+  Fmt.FMT_NUM_I4: 4,
+  Fmt.FMT_NUM_I64: 64,
+  Fmt.FMT_NUM_I8: 8,
+  Fmt.FMT_NUM_IU4: 4,
+  Fmt.FMT_NUM_IU8: 8,
+  Fmt.FMT_NUM_M64: 64,
+  Fmt.FMT_NUM_PK16_BF16: 256,
+  Fmt.FMT_NUM_PK16_BF8: 128,
+  Fmt.FMT_NUM_PK16_F16: 256,
+  Fmt.FMT_NUM_PK16_F32: 512,
+  Fmt.FMT_NUM_PK16_FP8: 128,
+  Fmt.FMT_NUM_PK16_I32: 512,
+  Fmt.FMT_NUM_PK16_I8: 128,
+  Fmt.FMT_NUM_PK2_B16: 32,
+  Fmt.FMT_NUM_PK2_B32: 64,
+  Fmt.FMT_NUM_PK2_B64: 128,
+  Fmt.FMT_NUM_PK2_BF16: 32,
+  Fmt.FMT_NUM_PK2_BF8: 16,
+  Fmt.FMT_NUM_PK2_F16: 32,
+  Fmt.FMT_NUM_PK2_F32: 64,
+  Fmt.FMT_NUM_PK2_FP4: 8,
+  Fmt.FMT_NUM_PK2_FP8: 16,
+  Fmt.FMT_NUM_PK2_I16: 32,
+  Fmt.FMT_NUM_PK2_I8: 16,
+  Fmt.FMT_NUM_PK2_U16: 32,
+  Fmt.FMT_NUM_PK2_U8: 16,
+  Fmt.FMT_NUM_PK32_BF16: 512,
+  Fmt.FMT_NUM_PK32_BF6: 192,
+  Fmt.FMT_NUM_PK32_BF8: 256,
+  Fmt.FMT_NUM_PK32_F16: 512,
+  Fmt.FMT_NUM_PK32_F32: 1024,
+  Fmt.FMT_NUM_PK32_FP6: 192,
+  Fmt.FMT_NUM_PK32_FP8: 256,
+  Fmt.FMT_NUM_PK32_I32: 1024,
+  Fmt.FMT_NUM_PK32_I8: 256,
+  Fmt.FMT_NUM_PK4_B8: 32,
+  Fmt.FMT_NUM_PK4_BF16: 64,
+  Fmt.FMT_NUM_PK4_BF8: 32,
+  Fmt.FMT_NUM_PK4_F16: 64,
+  Fmt.FMT_NUM_PK4_F32: 128,
+  Fmt.FMT_NUM_PK4_F64: 256,
+  Fmt.FMT_NUM_PK4_FP8: 32,
+  Fmt.FMT_NUM_PK4_I32: 128,
+  Fmt.FMT_NUM_PK4_I8: 32,
+  Fmt.FMT_NUM_PK4_IU8: 32,
+  Fmt.FMT_NUM_PK4_U8: 32,
+  Fmt.FMT_NUM_PK8_B32: 256,
+  Fmt.FMT_NUM_PK8_BF16: 128,
+  Fmt.FMT_NUM_PK8_BF8: 64,
+  Fmt.FMT_NUM_PK8_F16: 128,
+  Fmt.FMT_NUM_PK8_FP8: 64,
+  Fmt.FMT_NUM_PK8_I4: 32,
+  Fmt.FMT_NUM_PK8_I8: 64,
+  Fmt.FMT_NUM_PK8_IU4: 32,
+  Fmt.FMT_NUM_PK8_U4: 32,
+  Fmt.FMT_NUM_PK8_U8: 64,
+  Fmt.FMT_NUM_PK_F16: 32,
+  Fmt.FMT_NUM_PK_I16: 32,
+  Fmt.FMT_NUM_PK_I8: 32,
+  Fmt.FMT_NUM_PK_U16: 32,
+  Fmt.FMT_NUM_PK_U8: 32,
+  Fmt.FMT_NUM_U16: 16,
+  Fmt.FMT_NUM_U24: 24,
+  Fmt.FMT_NUM_U32: 32,
+  Fmt.FMT_NUM_U4: 4,
+  Fmt.FMT_NUM_U64: 64,
+  Fmt.FMT_NUM_U8: 8,
+  Fmt.FMT_RSRC: 128,
+  Fmt.FMT_RSRC_SCALAR: 128,
+  Fmt.FMT_RSRC_SCRATCH: 128,
+  Fmt.FMT_RSRC_SCRATCH_BYTE: 128,
+  Fmt.FMT_RSRC_SCRATCH_STRIDE: 128,
+  Fmt.FMT_RSRC_TYPED: 128,
+  Fmt.FMT_RSRC_TYPED_BYTE: 128,
+  Fmt.FMT_RSRC_TYPED_SCRATCH: 128,
+  Fmt.FMT_RSRC_TYPED_STRIDE: 128,
+  Fmt.FMT_RSRC_VECTOR: 128,
+  Fmt.FMT_RSRC_VECTOR_BYTE: 128,
+  Fmt.FMT_RSRC_VECTOR_STRIDE: 128,
+  Fmt.FMT_SAMP: 128,
+  Fmt.FMT_WMMA_AB_16X16_BF16: 128,
+  Fmt.FMT_WMMA_AB_16X16_BF8: 64,
+  Fmt.FMT_WMMA_AB_16X16_F16: 128,
+  Fmt.FMT_WMMA_AB_16X16_FP8: 64,
+  Fmt.FMT_WMMA_AB_16X16_IU4: 32,
+  Fmt.FMT_WMMA_AB_16X16_IU8: 64,
+  Fmt.FMT_WMMA_AB_16X32_BF16: 256,
+  Fmt.FMT_WMMA_AB_16X32_BF8: 128,
+  Fmt.FMT_WMMA_AB_16X32_F16: 256,
+  Fmt.FMT_WMMA_AB_16X32_FP8: 128,
+  Fmt.FMT_WMMA_AB_16X32_IU4: 64,
+  Fmt.FMT_WMMA_AB_16X32_IU8: 128,
+  Fmt.FMT_WMMA_AB_16X64_IU4: 128,
+  Fmt.FMT_WMMA_AB_BF16: 256,
+  Fmt.FMT_WMMA_AB_F16: 256,
+  Fmt.FMT_WMMA_AB_IU4: 64,
+  Fmt.FMT_WMMA_AB_IU8: 128,
+  Fmt.FMT_WMMA_DC_16X16_BF16: 128,
+  Fmt.FMT_WMMA_DC_16X16_F16: 128,
+  Fmt.FMT_WMMA_DC_16X16_F32: 256,
+  Fmt.FMT_WMMA_DC_16X16_I32: 256,
+  Fmt.FMT_WMMA_DC_BF16: 256,
+  Fmt.FMT_WMMA_DC_F16: 256,
+  Fmt.FMT_WMMA_DC_F32: 256,
+  Fmt.FMT_WMMA_DC_I32: 256,
+  Fmt.FMT_WMMA_INDEX_SET: 32,
+}
+
+class OpType(Enum):
+  OPR_ACCVGPR = auto()
+  OPR_ATTR = auto()
+  OPR_CLAUSE = auto()
+  OPR_DELAY = auto()
+  OPR_EXEC = auto()
+  OPR_HWREG = auto()
+  OPR_LABEL = auto()
+  OPR_SDST = auto()
+  OPR_SDST_NULL = auto()
+  OPR_SENDMSG = auto()
+  OPR_SENDMSG_RTN = auto()
+  OPR_SIMM16 = auto()
+  OPR_SIMM24 = auto()
+  OPR_SIMM4 = auto()
+  OPR_SIMM5 = auto()
+  OPR_SIMM8 = auto()
+  OPR_SLEEP = auto()
+  OPR_SMEM_OFFSET = auto()
+  OPR_SMEM_OFFSET_NOK = auto()
+  OPR_SRC = auto()
+  OPR_SRC_ACCVGPR = auto()
+  OPR_SRC_NOLDS = auto()
+  OPR_SRC_NOLIT = auto()
+  OPR_SRC_SIMPLE = auto()
+  OPR_SRC_VGPR = auto()
+  OPR_SRC_VGPR_OR_ACCVGPR = auto()
+  OPR_SRC_VGPR_OR_ACCVGPR_OR_CONST = auto()
+  OPR_SRC_VGPR_OR_INLINE = auto()
+  OPR_SREG = auto()
+  OPR_SREG_LITERAL = auto()
+  OPR_SREG_M0 = auto()
+  OPR_SREG_M0_INL = auto()
+  OPR_SREG_NOVCC = auto()
+  OPR_SSRC = auto()
+  OPR_SSRC_BARRIER_ID = auto()
+  OPR_SSRC_LANESEL = auto()
+  OPR_SSRC_NOLIT = auto()
+  OPR_TGT = auto()
+  OPR_VERSION = auto()
+  OPR_VGPR = auto()
+  OPR_VGPR_OR_ACCVGPR = auto()
+  OPR_VGPR_OR_LDS = auto()
+  OPR_WAITCNT = auto()
+  OPR_WAITCNT_DEPCTR = auto()
+  OPR_WAIT_ALU = auto()
+  OPR_WAIT_EVENT = auto()
+  OPR_WAIT_MEM_DS = auto()

extra/assembly/amd/disasm.py

@@ -0,0 +1,907 @@
+# RDNA3/RDNA4/CDNA disassembler
+from __future__ import annotations
+import re, struct
+from typing import Callable
+from extra.assembly.amd.dsl import Inst, Reg
+
+# Special register mappings for disassembly
+SPECIAL_GPRS = {106: 'vcc_lo', 107: 'vcc_hi', 124: 'null', 125: 'm0', 126: 'exec_lo', 127: 'exec_hi',
+                128: '0', 240: '0.5', 241: '-0.5', 242: '1.0', 243: '-1.0', 244: '2.0', 245: '-2.0', 246: '4.0', 247: '-4.0', 248: '0x3e22f983', 253: 'scc'}
+SPECIAL_GPRS_CDNA = {106: 'vcc_lo', 107: 'vcc_hi', 124: 'm0', 126: 'exec_lo', 127: 'exec_hi',
+                     128: '0', 240: '0.5', 241: '-0.5', 242: '1.0', 243: '-1.0', 244: '2.0', 245: '-2.0', 246: '4.0', 247: '-4.0', 248: '0x3e22f983', 253: 'scc',
+                     102: 'flat_scratch_lo', 103: 'flat_scratch_hi', 104: 'xnack_mask_lo', 105: 'xnack_mask_hi',
+                     251: 'src_vccz', 252: 'src_execz'}
+SPECIAL_PAIRS = {106: 'vcc', 126: 'exec'}
+SPECIAL_PAIRS_CDNA = {106: 'vcc', 126: 'exec', 102: 'flat_scratch', 104: 'xnack_mask'}
+
+def decode_src(v, cdna: bool = False) -> str:
+  """Decode a source operand encoding to its string representation."""
+  v = _unwrap(v)
+  gprs = SPECIAL_GPRS_CDNA if cdna else SPECIAL_GPRS
+  if v in gprs: return gprs[v]
+  if v < 106: return f's{v}'
+  if 108 <= v < 124: return f'ttmp{v - 108}'
+  if 129 <= v <= 192: return str(v - 128)  # positive integers 1-64
+  if 193 <= v <= 208: return str(-(v - 192))  # negative integers -1 to -16
+  if v >= 256: return f'v{v - 256}'
+  return f's{v}'
+
+def _unwrap(v) -> int:
+  """Unwrap Reg to int offset, or return int as-is."""
+  return v.offset if isinstance(v, Reg) else v
+
+def _vi(v) -> int:
+  """Get VGPR index from Reg or int (for v[N] fields that encode as 256+N)."""
+  off = _unwrap(v)
+  return off - 256 if off >= 256 else off
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# LITERAL FORMATTING
+# ═══════════════════════════════════════════════════════════════════════════════
+
+_FLOAT_DEC = {240: 0.5, 241: -0.5, 242: 1.0, 243: -1.0, 244: 2.0, 245: -2.0, 246: 4.0, 247: -4.0}
+
+def _lit(inst, v, neg=0, cdna=None) -> str:
+  """Format literal/inline constant value."""
+  if cdna is None: cdna = _is_cdna(inst)
+  v = _unwrap(v)
+  if v == 255:
+    lit = inst._literal
+    if lit is None: return "0"
+    s = f"0x{lit:x}"
+  elif v in _FLOAT_DEC: s = str(_FLOAT_DEC[v])
+  elif 128 <= v <= 192: s = str(v - 128)
+  elif 193 <= v <= 208: s = str(-(v - 192))
+  elif v < 128: s = decode_src(v, cdna)
+  elif v >= 256: s = f"v{v - 256}"
+  else: s = decode_src(v, cdna)
+  return f"-{s}" if neg else s
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# INSTRUCTION METADATA - fallback functions when inst.num_srcs()/inst.operands unavailable
+# ═══════════════════════════════════════════════════════════════════════════════
+
+def _num_srcs(inst) -> int:
+  """Fallback: get number of source operands from instruction name."""
+  name = getattr(inst, 'op_name', '') or ''
+  n = name.upper()
+  # FMAC/MAC ops are 2-source (dst is implicit accumulator), but FMA/MAD ops are 3-source
+  if 'FMAC' in n or 'V_MAC_' in n: return 2
+  if any(x in n for x in ('FMA', 'MAD', 'CNDMASK', 'BFE', 'BFI', 'LERP', 'MED3', 'SAD', 'DIV_FMAS', 'DIV_FIXUP', 'DIV_SCALE', 'CUBE')): return 3
+  # PERMLANE_VAR ops are 2-source, but PERMLANE (non-VAR) are 3-source
+  if 'PERMLANE' in n and '_VAR' not in n: return 3
+  if any(x in n for x in ('_ADD3', '_LSHL_ADD', '_ADD_LSHL', '_LSHL_OR', '_AND_OR', 'OR3_B32', 'AND_OR_B32', 'ALIGNBIT', 'ALIGNBYTE', 'V_PERM_', 'XOR3', 'XAD', 'MULLIT', 'MINMAX', 'MAXMIN', 'MINIMUMMAXIMUM', 'MAXIMUMMINIMUM', 'MINIMUM3', 'MAXIMUM3', 'MIN3', 'MAX3', 'DOT2', 'CVT_PK_U8_F32', 'DOT4', 'DOT8', 'WMMA', 'SWMMAC')): return 3
+  return 2
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# IMPORTS
+# ═══════════════════════════════════════════════════════════════════════════════
+
+from extra.assembly.amd.autogen.rdna3.ins import (VOP1, VOP1_SDST, VOP1_SDST_LIT, VOP1_LIT, VOP2, VOP2_LIT, VOP3, VOP3_SDST, VOP3_SDST_LIT,
+  VOP3_LIT, VOP3SD, VOP3SD_LIT, VOP3P, VOP3P_LIT, VOPC, VOPC_LIT, VOPD, VOPD_LIT, VINTERP, SOP1, SOP1_LIT, SOP2, SOP2_LIT, SOPC, SOPC_LIT,
+  SOPK, SOPK_LIT, SOPP, SMEM, DS, FLAT, GLOBAL, SCRATCH, VOP2Op, VOPDOp, SOPPOp, HWREG, MSG)
+from extra.assembly.amd.autogen.rdna4.ins import (VOP1 as R4_VOP1, VOP1_SDST as R4_VOP1_SDST, VOP1_SDST_LIT as R4_VOP1_SDST_LIT, VOP1_LIT as R4_VOP1_LIT,
+  VOP2 as R4_VOP2, VOP2_LIT as R4_VOP2_LIT, VOP3 as R4_VOP3, VOP3_SDST as R4_VOP3_SDST, VOP3_SDST_LIT as R4_VOP3_SDST_LIT, VOP3_LIT as R4_VOP3_LIT,
+  VOP3SD as R4_VOP3SD, VOP3SD_LIT as R4_VOP3SD_LIT, VOP3P as R4_VOP3P, VOP3P_LIT as R4_VOP3P_LIT, VOPC as R4_VOPC, VOPC_LIT as R4_VOPC_LIT,
+  VOPD as R4_VOPD, VOPD_LIT as R4_VOPD_LIT, VINTERP as R4_VINTERP, SOP1 as R4_SOP1, SOP1_LIT as R4_SOP1_LIT, SOP2 as R4_SOP2, SOP2_LIT as R4_SOP2_LIT,
+  SOPC as R4_SOPC, SOPC_LIT as R4_SOPC_LIT, SOPK as R4_SOPK, SOPK_LIT as R4_SOPK_LIT, SOPP as R4_SOPP, SMEM as R4_SMEM, DS as R4_DS,
+  VOPDOp as R4_VOPDOp, HWREG as HWREG_RDNA4, VFLAT as R4_FLAT, VGLOBAL as R4_GLOBAL, VSCRATCH as R4_SCRATCH)
+from extra.assembly.amd.autogen.cdna.ins import FLAT as C_FLAT, HWREG as HWREG_CDNA
+
+def _is_cdna(inst: Inst) -> bool: return 'cdna' in inst.__class__.__module__
+def _is_r4(inst: Inst) -> bool: return 'rdna4' in inst.__class__.__module__
+
+# CDNA opcode name aliases for disasm (new name -> old name expected by tests)
+_CDNA_DISASM_ALIASES = {'v_fmac_f64': 'v_mul_legacy_f32', 'v_dot2c_f32_bf16': 'v_mac_f32', 'v_fmamk_f32': 'v_madmk_f32', 'v_fmaak_f32': 'v_madak_f32'}
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# HELPERS
+# ═══════════════════════════════════════════════════════════════════════════════
+
+def _reg(p: str, b: int, n: int = 1) -> str: return f"{p}{_unwrap(b)}" if n == 1 else f"{p}[{_unwrap(b)}:{_unwrap(b)+n-1}]"
+def _sreg(b: int, n: int = 1) -> str: return _reg("s", _unwrap(b), n)
+def _vreg(b: int, n: int = 1) -> str: b = _unwrap(b); return _reg("v", b - 256 if b >= 256 else b, n)
+def _areg(b: int, n: int = 1) -> str: b = _unwrap(b); return _reg("a", b - 256 if b >= 256 else b, n)  # accumulator registers for GFX90a
+def _ttmp(b, n: int = 1) -> str | None: b = _unwrap(b); return _reg("ttmp", b - 108, n) if 108 <= b <= 123 else None
+
+def _fmt_sdst(v, n: int = 1, cdna: bool = False) -> str:
+  v = _unwrap(v)
+  if t := _ttmp(v, n): return t
+  pairs = SPECIAL_PAIRS_CDNA if cdna else SPECIAL_PAIRS
+  gprs = SPECIAL_GPRS_CDNA if cdna else SPECIAL_GPRS
+  if n > 1: return pairs.get(v) or gprs.get(v) or _sreg(v, n)  # also check gprs for null/m0
+  return gprs.get(v, f"s{v}")
+
+def _fmt_src(v, n: int = 1, cdna: bool = False) -> str:
+  v = _unwrap(v)
+  if v == 253: return "src_scc"  # SCC as source operand
+  if n == 1: return decode_src(v, cdna)
+  if v >= 256: return _vreg(v, n)
+  if v <= 101: return _sreg(v, n)  # s0-s101 can be pairs, but 102+ are special on CDNA
+  pairs = SPECIAL_PAIRS_CDNA if cdna else SPECIAL_PAIRS
+  if n == 2 and v in pairs: return pairs[v]
+  if v <= 105: return _sreg(v, n)  # s102-s105 regular pairs for RDNA
+  if t := _ttmp(v, n): return t
+  return decode_src(v, cdna)
+
+def _fmt_v16(v, base: int = 256, hi_thresh: int = 384) -> str:
+  v = _unwrap(v)
+  return f"v{(v - base) & 0x7f}.{'h' if v >= hi_thresh else 'l'}"
+
+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: v = _unwrap(v); return _fmt_v16(v) if v >= 256 else _lit(inst, 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) -> str:
+  """Format VOP3 source operand with modifiers."""
+  v = _unwrap(v)
+  if v == 255: s = _lit(inst, v)  # literal constant takes priority
+  elif 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"
+  elif v == 253: s = "src_scc"  # VOP3 sources use src_scc not scc
+  else: s = _lit(inst, 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 ""
+  dst_hi = (opsel >> 3) & 1
+  if n == 1: return f" op_sel:[{opsel & 1},{dst_hi}]"
+  # Use 4-element format if bit 2 is set (src2 selection used) or if 3+ sources
+  if n == 2 and not ((opsel >> 2) & 1): return f" op_sel:[{opsel & 1},{(opsel >> 1) & 1},{dst_hi}]"
+  return f" op_sel:[{opsel & 1},{(opsel >> 1) & 1},{(opsel >> 2) & 1},{dst_hi}]"
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# DISASSEMBLER
+# ═══════════════════════════════════════════════════════════════════════════════
+
+def _disasm_vop1(inst: VOP1) -> str:
+  name, cdna = inst.op_name.lower() or f'vop1_op_{inst.op}', _is_cdna(inst)
+  name = name.replace('_e32', '')  # Strip _e32 suffix
+  if any(x in name for x in ('v_nop', 'v_pipeflush', 'v_clrexcp')): return name  # no operands
+  if 'readfirstlane' in name:
+    src = inst.src0.fmt() if inst.src0.offset >= 256 else decode_src(inst.src0.offset, cdna)
+    vdst_off = inst.vdst.offset - 256 if inst.vdst.offset >= 256 else inst.vdst.offset
+    return f"{name} {_fmt_sdst(vdst_off, 1, cdna)}, {src}"
+  bits = inst.canonical_op_bits
+  is16_dst, is16_src = not cdna and bits['d'] == 16, not cdna and bits['s0'] == 16
+  # Format dst
+  if is16_dst: dst = _fmt_v16(inst.vdst)
+  else: dst = inst.vdst.fmt()
+  # Format src
+  if inst.src0.offset == 255: src = _lit(inst, inst.src0)
+  elif is16_src and inst.src0.offset >= 256: src = _fmt_v16(inst.src0)
+  elif inst.src0.sz > 1: src = _fmt_src(inst.src0, inst.src0.sz, cdna)
+  else: src = _lit(inst, inst.src0)
+  return f"{name} {dst}, {src}"
+
+_VOP2_CARRY_OUT = {'v_add_co_u32', 'v_sub_co_u32', 'v_subrev_co_u32'}  # carry out only
+_VOP2_CARRY_INOUT = {'v_addc_co_u32', 'v_subb_co_u32', 'v_subbrev_co_u32'}  # carry in and out (CDNA)
+_VOP2_CARRY_INOUT_RDNA = {'v_add_co_ci_u32', 'v_sub_co_ci_u32', 'v_subrev_co_ci_u32'}  # carry in and out (RDNA)
+def _disasm_vop2(inst: VOP2) -> str:
+  name, cdna = inst.op_name.lower(), _is_cdna(inst)
+  if cdna: name = _CDNA_DISASM_ALIASES.get(name, name)  # apply CDNA aliases
+  suf = "" if cdna or name.endswith('_e32') or (not cdna and inst.op == VOP2Op.V_DOT2ACC_F32_F16_E32) else "_e32"
+  lit = inst._literal
+  is16 = not cdna and inst.canonical_op_bits['d'] == 16
+  # fmaak/madak: dst = src0 * vsrc1 + K, fmamk/madmk: dst = src0 * K + vsrc1
+  if 'fmaak' in name or 'madak' in name or (not cdna and inst.op in (VOP2Op.V_FMAAK_F32_E32, VOP2Op.V_FMAAK_F16_E32)):
+    if lit is None: return f"op_{inst.op.value if hasattr(inst.op, 'value') else inst.op}"
+    if is16: return f"{name}{suf} {_fmt_v16(inst.vdst)}, {_src16(inst, inst.src0)}, {_fmt_v16(inst.vsrc1)}, 0x{lit:x}"
+    return f"{name}{suf} {inst.vdst.fmt()}, {_lit(inst, inst.src0)}, {inst.vsrc1.fmt()}, 0x{lit:x}"
+  if 'fmamk' in name or 'madmk' in name or (not cdna and inst.op in (VOP2Op.V_FMAMK_F32_E32, VOP2Op.V_FMAMK_F16_E32)):
+    if lit is None: return f"op_{inst.op.value if hasattr(inst.op, 'value') else inst.op}"
+    if is16: return f"{name}{suf} {_fmt_v16(inst.vdst)}, {_src16(inst, inst.src0)}, 0x{lit:x}, {_fmt_v16(inst.vsrc1)}"
+    return f"{name}{suf} {inst.vdst.fmt()}, {_lit(inst, inst.src0)}, 0x{lit:x}, {inst.vsrc1.fmt()}"
+  if is16: return f"{name}{suf} {_fmt_v16(inst.vdst)}, {_src16(inst, inst.src0)}, {_fmt_v16(inst.vsrc1)}"
+  vcc = "vcc" if cdna else "vcc_lo"
+  basename = name.replace('_e32', '')
+  if cdna and basename in _VOP2_CARRY_OUT: return f"{name}{suf} {inst.vdst.fmt()}, {vcc}, {_lit(inst, inst.src0)}, {inst.vsrc1.fmt()}"
+  if cdna and basename in _VOP2_CARRY_INOUT: return f"{name}{suf} {inst.vdst.fmt()}, {vcc}, {_lit(inst, inst.src0)}, {inst.vsrc1.fmt()}, {vcc}"
+  if not cdna and basename in _VOP2_CARRY_INOUT_RDNA: return f"{name}{suf} {inst.vdst.fmt()}, {vcc}, {_lit(inst, inst.src0)}, {inst.vsrc1.fmt()}, {vcc}"
+  sn0 = inst.canonical_op_regs.get('s0', 1)
+  if inst.vdst.sz > 1 or sn0 > 1 or inst.vsrc1.sz > 1:
+    src0 = _lit(inst, inst.src0) if inst.src0.offset == 255 else _fmt_src(inst.src0, sn0, cdna)
+    return f"{name.replace('_e32', '')} {inst.vdst.fmt()}, {src0}, {inst.vsrc1.fmt()}"
+  return f"{name}{suf} {inst.vdst.fmt()}, {_lit(inst, inst.src0)}, {inst.vsrc1.fmt()}" + (f", {vcc}" if name == 'v_cndmask_b32' else "")
+
+def _disasm_vopc(inst: VOPC) -> str:
+  name, cdna = inst.op_name.lower(), _is_cdna(inst)
+  bits = inst.canonical_op_bits
+  is16 = bits['s0'] == 16
+  if cdna:
+    s0 = _lit(inst, inst.src0) if inst.src0.offset == 255 else _fmt_src(inst.src0, inst.src0.sz, cdna)
+    return f"{name} vcc, {s0}, {inst.vsrc1.fmt()}"  # CDNA VOPC always outputs vcc
+  # RDNA: v_cmpx_* writes to exec (no vcc), v_cmp_* writes to vcc_lo
+  has_vcc = 'cmpx' not in name
+  s0 = _lit(inst, inst.src0) if inst.src0.offset == 255 else inst.src0.fmt() if inst.src0.sz > 1 else _src16(inst, inst.src0.offset) if is16 else _lit(inst, inst.src0)
+  s1 = inst.vsrc1.fmt() if inst.vsrc1.sz > 1 else _fmt_v16(inst.vsrc1) if is16 else inst.vsrc1.fmt()
+  suf = "" if name.endswith('_e32') else "_e32"
+  return f"{name}{suf} vcc_lo, {s0}, {s1}" if has_vcc else f"{name}{suf} {s0}, {s1}"
+
+NO_ARG_SOPP = {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, SOPPOp.S_TTRACEDATA}
+
+def _disasm_sopp(inst: SOPP) -> str:
+  name, cdna = inst.op_name.lower(), _is_cdna(inst)
+  is_rdna4 = _is_r4(inst)
+  # Ops that have no argument when simm16 == 0
+  no_arg_zero = {'s_barrier', 's_wakeup', 's_icache_inv', 's_ttracedata', 's_wait_idle', 's_endpgm_saved',
+                 's_endpgm_ordered_ps_done', 's_code_end'}
+  if name in no_arg_zero: return name if inst.simm16 == 0 else f"{name} {inst.simm16}"
+  if name == 's_endpgm': return name if inst.simm16 == 0 else f"{name} {inst.simm16}"
+  if cdna:
+    if name == 's_waitcnt':
+      # GFX9 format: vmcnt[3:0]=bits[3:0], vmcnt[5:4]=bits[15:14], expcnt=bits[6:4], lgkmcnt=bits[11:8] (4 bits, max 15)
+      vm_lo, exp, lgkm, vm_hi = inst.simm16 & 0xf, (inst.simm16 >> 4) & 0x7, (inst.simm16 >> 8) & 0xf, (inst.simm16 >> 14) & 0x3
+      vm = vm_lo | (vm_hi << 4)
+      p = [f"vmcnt({vm})" if vm != 0x3f else "", f"expcnt({exp})" if exp != 7 else "", f"lgkmcnt({lgkm})" if lgkm != 0xf else ""]
+      return f"s_waitcnt {' '.join(x for x in p if x) or '0'}"
+    if name.startswith(('s_cbranch', 's_branch')): return f"{name} {inst.simm16}"
+    if name == 's_set_gpr_idx_mode':
+      flags = [n for i, n in enumerate(['SRC0', 'SRC1', 'SRC2', 'DST']) if inst.simm16 & (1 << i)]
+      return f"{name} gpr_idx({','.join(flags)})"
+    return f"{name} 0x{inst.simm16:x}" if inst.simm16 else name
+  # RDNA (use name-based checks instead of enum-based for cross-arch compatibility)
+  if name == 's_waitcnt':
+    if is_rdna4:
+      return f"{name} {inst.simm16}" if inst.simm16 else f"{name} 0"
+    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 name == 's_delay_alu':
+    deps = ['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']
+    skips = ['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'}"
+  if name.startswith(('s_cbranch', 's_branch')): return f"{name} {inst.simm16}"
+  return f"{name} 0x{inst.simm16:x}"
+
+def _disasm_smem(inst: SMEM) -> str:
+  name, cdna = inst.op_name.lower(), _is_cdna(inst)
+  if name in ('s_gl1_inv', 's_dcache_inv', 's_dcache_inv_vol', 's_dcache_wb', 's_dcache_wb_vol', 's_icache_inv'): return name
+  soe, imm = getattr(inst, 'soe', 0) or getattr(inst, 'soffset_en', 0), getattr(inst, 'imm', 1)
+  is_rdna4 = _is_r4(inst)
+  offset = inst.ioffset if is_rdna4 else getattr(inst, 'offset', 0)
+  if cdna:
+    if soe and imm: off_s = f"{decode_src(inst.soffset, cdna)} offset:0x{offset:x}"
+    elif imm: off_s = f"0x{offset:x}"
+    elif offset < 256: off_s = decode_src(offset, cdna)
+    else: off_s = decode_src(inst.soffset, cdna)
+  elif offset and inst.soffset != 124: off_s = f"{decode_src(inst.soffset, cdna)} offset:0x{offset:x}"
+  elif offset: off_s = f"0x{offset:x}"
+  else: off_s = decode_src(inst.soffset, cdna)
+  is_buffer = 'buffer' in name or 's_atc_probe_buffer' == name
+  sbase_idx, sbase_count = _unwrap(inst.sbase), 4 if is_buffer else 2
+  sbase_str = _fmt_src(sbase_idx, sbase_count, cdna) 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} {_unwrap(inst.sdata)}, {sbase_str}, {off_s}"
+  if 'prefetch' in name:
+    off = getattr(inst, 'ioffset', getattr(inst, 'offset', 0))
+    if off >= 0x800000: off = off - 0x1000000
+    off_s = f"0x{off:x}" if off > 255 else str(off)
+    soff_s = decode_src(inst.soffset, cdna) if inst.soffset != 124 else ("m0" if cdna else "null")
+    if 'pc_rel' in name: return f"{name} {off_s}, {soff_s}, {_unwrap(inst.sdata)}"
+    return f"{name} {sbase_str}, {off_s}, {soff_s}, {_unwrap(inst.sdata)}"
+  # Use get_field_bits for register count
+  dst_n = inst.canonical_op_regs.get('d', 1)
+  th, scope = getattr(inst, 'th', 0), getattr(inst, 'scope', 0)
+  if is_rdna4:  # RDNA4 uses th/scope instead of glc/dlc
+    th_names = ['TH_LOAD_RT', 'TH_LOAD_NT', 'TH_LOAD_HT', 'TH_LOAD_LU']
+    scope_names = ['SCOPE_CU', 'SCOPE_SE', 'SCOPE_DEV', 'SCOPE_SYS']
+    mods = (f" th:{th_names[th]}" if th else "") + (f" scope:{scope_names[scope]}" if scope else "")
+    return f"{name} {_fmt_sdst(inst.sdata, dst_n, cdna)}, {sbase_str}, {off_s}{mods}"
+  if th or scope:
+    th_names = ['TH_LOAD_RT', 'TH_LOAD_NT', 'TH_LOAD_HT', 'TH_LOAD_LU']
+    scope_names = ['SCOPE_CU', 'SCOPE_SE', 'SCOPE_DEV', 'SCOPE_SYS']
+    mods = (f" th:{th_names[th]}" if th else "") + (f" scope:{scope_names[scope]}" if scope else "")
+    return f"{name} {_fmt_sdst(inst.sdata, dst_n, cdna)}, {sbase_str}, {off_s}{mods}"
+  if 'discard' in name: return f"{name} {sbase_str}, {off_s}" + _mods((inst.glc, " glc"), (getattr(inst, 'dlc', 0), " dlc"))
+  if name in ('s_memrealtime', 's_memtime'): return f"{name} {_fmt_sdst(inst.sdata, dst_n, cdna)}"
+  return f"{name} {_fmt_sdst(inst.sdata, dst_n, cdna)}, {sbase_str}, {off_s}" + _mods((inst.glc, " glc"), (getattr(inst, 'dlc', 0), " dlc"))
+
+def _disasm_flat(inst: FLAT) -> str:
+  name, cdna, r4 = inst.op_name.lower(), _is_cdna(inst), _is_r4(inst)
+  acc = getattr(inst, 'acc', 0)
+  reg_fn = _areg if acc else _vreg
+  if r4: seg = 'flat' if (cls_name:=inst.__class__.__name__) == 'VFLAT' else ('global' if cls_name == 'VGLOBAL' else 'scratch')
+  else: seg = ['flat', 'scratch', 'global'][inst.seg] if inst.seg < 3 else 'flat'
+  instr = f"{seg}_{name.split('_', 1)[1] if '_' in name else name}"
+  # Global/scratch uses 13-bit signed offset
+  offset = inst.ioffset if r4 else inst.offset
+  if seg != 'flat':
+    if cdna:
+      # CDNA: bit 12 is sign bit but not in offset field
+      raw = int.from_bytes(inst.to_bytes(), 'little')
+      off_val = offset | ((raw >> 12) & 1) << 12  # get bit 12
+    else:
+      off_val = offset
+    off_val = off_val if off_val < 4096 else off_val - 8192  # sign extend 13-bit
+  else:
+    off_val = offset
+  # Use get_field_bits: data for stores/atomics, d for loads
+  regs = inst.canonical_op_regs
+  w = regs.get('data', regs.get('d', 1)) if 'store' in name or 'atomic' in name else regs.get('d', 1)
+  off_s = f" offset:{off_val}" if off_val else ""
+  if cdna: mods = f"{off_s}{' sc0' if inst.sc0 else ''}{' nt' if inst.nt else ''}{' sc1' if getattr(inst, 'sc1', 0) else ''}"
+  elif r4: mods = f"{off_s}{' scope' if inst.scope else ''}{' th' if inst.th else ''}"
+  else: mods = f"{off_s}{' glc' if inst.glc else ''}{' slc' if inst.slc else ''}{' dlc' if inst.dlc else ''}"
+  if seg == 'flat': saddr_s = ""
+  elif _unwrap(inst.saddr) in (0x7F, 124): saddr_s = ", off"
+  elif seg == 'scratch': saddr_s = f", {decode_src(inst.saddr, cdna)}"
+  elif _unwrap(inst.saddr) in (SPECIAL_PAIRS_CDNA if cdna else SPECIAL_PAIRS): saddr_s = f", {(SPECIAL_PAIRS_CDNA if cdna else SPECIAL_PAIRS)[_unwrap(inst.saddr)]}"
+  elif t := _ttmp(inst.saddr, 2): saddr_s = f", {t}"
+  else: saddr_s = f", {_sreg(inst.saddr, 2) if _unwrap(inst.saddr) < 106 else decode_src(_unwrap(inst.saddr), cdna)}"
+  if 'addtid' in name: return f"{instr} {reg_fn(inst.data if 'store' in name else inst.vdst)}{saddr_s}{mods}"
+  # RDNA4: vaddr instead of addr, vsrc instead of data
+  addr = inst.vaddr if r4 else inst.addr
+  data = inst.vsrc if r4 else inst.data
+  # load_lds_* instructions: vaddr, saddr (no vdst, data goes to LDS)
+  if 'load_lds' in name:
+    addr_w = 1 if seg == 'scratch' or (_unwrap(inst.saddr) not in (0x7F, 124)) else 2
+    addr_s = "off" if not inst.sve and seg == 'scratch' else _vreg(addr, addr_w)
+    return f"{instr} {addr_s}{saddr_s}{mods}"
+  if seg == 'flat': addr_w = 2  # flat always uses 64-bit vaddr
+  elif cdna: addr_w = 1 if seg == 'scratch' or (_unwrap(inst.saddr) not in (0x7F, 124)) else 2
+  else: addr_w = 1 if seg == 'scratch' or (_unwrap(inst.saddr) not in (0x7F, 124)) else 2
+  addr_s = "off" if not inst.sve and seg == 'scratch' else _vreg(addr, addr_w)
+  data_s, vdst_s = reg_fn(data, w), reg_fn(inst.vdst, w // 2 if 'cmpswap' in name else w)
+  if 'atomic' in name:
+    glc_or_sc0 = inst.sc0 if cdna else inst.glc
+    return f"{instr} {vdst_s}, {addr_s}, {data_s}{saddr_s if seg != 'flat' else ''}{mods}" if glc_or_sc0 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} {reg_fn(inst.vdst, w)}, {addr_s}{saddr_s}{mods}"
+
+def _disasm_ds(inst: DS) -> str:
+  op, name = inst.op, inst.op_name.lower()
+  acc = getattr(inst, 'acc', 0)
+  reg_fn = _areg if acc else _vreg
+  gds = " gds" if getattr(inst, 'gds', 0) else ""
+  off = f" offset:{inst.offset0 | (inst.offset1 << 8)}" if inst.offset0 or inst.offset1 else ""
+  off2 = (" offset0:" + str(inst.offset0) if inst.offset0 else "") + (" offset1:" + str(inst.offset1) if inst.offset1 else "")
+  # Use get_field_bits: data for stores/writes/atomics, d for loads
+  regs = inst.canonical_op_regs
+  w = regs.get('data', regs.get('d', 1)) if 'store' in name or 'write' in name or ('load' not in name and 'read' not in name) else regs.get('d', 1)
+  d0, d1, dst, addr = reg_fn(inst.data0, w), reg_fn(inst.data1, w), reg_fn(inst.vdst, w), _vreg(inst.addr)
+
+  if name == 'ds_nop': return name
+  if name == 'ds_bvh_stack_rtn_b32': return f"{name} {_vreg(inst.vdst)}, {addr}, {_vreg(inst.data0)}, {_vreg(inst.data1, 4)}{off}{gds}"
+  if 'bvh_stack_push' in name:
+    d1_regs = 8 if 'push8' in name else 4
+    vdst_regs = 2 if 'pop2' in name else 1
+    vdst_s = _vreg(inst.vdst, vdst_regs) if vdst_regs > 1 else _vreg(inst.vdst)
+    return f"{name} {vdst_s}, {addr}, {_vreg(inst.data0)}, {_vreg(inst.data1, d1_regs)}{off}{gds}"
+  if 'gws_sema' in name and 'sema_br' not in name: return f"{name}{off}{gds}"
+  if 'gws_' in name: return f"{name} {addr}{off}{gds}"
+  if name in ('ds_consume', 'ds_append'): return f"{name} {reg_fn(inst.vdst)}{off}{gds}"
+  if 'gs_reg' in name: return f"{name} {reg_fn(inst.vdst, 2)}, {reg_fn(inst.data0)}{off}{gds}"
+  if '2addr' in name:
+    if 'load' in name: return f"{name} {reg_fn(inst.vdst, regs.get('d', 1))}, {addr}{off2}{gds}"
+    if 'store' in name and 'xchg' not in name: return f"{name} {addr}, {d0}, {d1}{off2}{gds}"
+    return f"{name} {reg_fn(inst.vdst, regs.get('d', 1))}, {addr}, {d0}, {d1}{off2}{gds}"
+  if 'write2' in name: return f"{name} {addr}, {d0}, {d1}{off2}{gds}"
+  if 'read2' in name: return f"{name} {reg_fn(inst.vdst, regs.get('d', 1))}, {addr}{off2}{gds}"
+  if 'xchg2' in name: return f"{name} {reg_fn(inst.vdst, regs.get('d', 1))}, {addr}, {d0}, {d1}{off2}{gds}"
+  if 'load' in name or ('read' in name and 'read2' not in name): return f"{name} {reg_fn(inst.vdst)}{off}{gds}" if 'addtid' in name else f"{name} {dst}, {addr}{off}{gds}"
+  if ('store' in name or 'write' in name) and not _has(name, 'cmp', 'xchg', 'write2'):
+    return f"{name} {reg_fn(inst.data0)}{off}{gds}" if 'addtid' in name else f"{name} {addr}, {d0}{off}{gds}"
+  if 'swizzle' in name or name == 'ds_ordered_count': return f"{name} {reg_fn(inst.vdst)}, {addr}{off}{gds}"
+  if 'permute' in name: return f"{name} {reg_fn(inst.vdst)}, {addr}, {reg_fn(inst.data0)}{off}{gds}"
+  if 'condxchg' in name: return f"{name} {reg_fn(inst.vdst, 2)}, {addr}, {reg_fn(inst.data0, 2)}{off}{gds}"
+  if _has(name, 'cmpst', '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()
+  n_up = name.upper()
+  bits = inst.canonical_op_bits
+
+  # RDNA4 v_s_* scalar VOP3 instructions - vdst is SGPR (VGPRField adds 256)
+  if name.startswith('v_s_'):
+    src = _lit(inst, inst.src0) if _unwrap(inst.src0) == 255 else ("src_scc" if _unwrap(inst.src0) == 253 else _fmt_src(inst.src0, max(1, bits['s0'] // 32)))
+    if inst.neg & 1: src = f"-{src}"
+    if inst.abs & 1: src = f"|{src}|"
+    clamp = getattr(inst, 'cm', None) or getattr(inst, 'clmp', 0)
+    vdst_raw = _unwrap(inst.vdst)
+    return f"{name} s{vdst_raw - 256 if vdst_raw >= 256 else vdst_raw}, {src}" + (" clamp" if clamp else "") + _omod(inst.omod)
+
+  # Use get_field_bits for register sizes and 16-bit detection
+  r0, r1, r2 = max(1, bits['s0'] // 32), max(1, bits['s1'] // 32), max(1, bits['s2'] // 32)
+  dn = max(1, bits['d'] // 32)
+  is16_d, is16_s, is16_s2 = bits['d'] == 16, bits['s0'] == 16, bits['s2'] == 16
+
+  s0 = _vop3_src(inst, inst.src0, inst.neg&1, inst.abs&1, inst.opsel&1, r0, is16_s)
+  s1 = _vop3_src(inst, inst.src1, inst.neg&2, inst.abs&2, inst.opsel&2, r1, is16_s)
+  s2 = _vop3_src(inst, inst.src2, inst.neg&4, inst.abs&4, inst.opsel&4, r2, is16_s2)
+
+  # Format destination
+  if 'readlane' in name:
+    vdst_off = inst.vdst.offset - 256 if inst.vdst.offset >= 256 else inst.vdst.offset
+    dst = _fmt_sdst(vdst_off, 1)
+  elif is16_d: dst = f"{inst.vdst.fmt()}.h" if (inst.opsel & 8) else f"{inst.vdst.fmt()}.l"
+  else: dst = inst.vdst.fmt()
+
+  clamp = getattr(inst, 'cm', None) or getattr(inst, 'clmp', 0)
+  cl, om = " clamp" if clamp else "", _omod(inst.omod)
+  nonvgpr_opsel = (inst.src0.offset < 256 and (inst.opsel & 1)) or (inst.src1.offset < 256 and (inst.opsel & 2)) or (inst.src2.offset < 256 and (inst.opsel & 4))
+  need_opsel = nonvgpr_opsel or (inst.opsel and not is16_s)
+
+  op_val = inst.op.value if hasattr(inst.op, 'value') else inst.op
+  e64 = "" if name.endswith('_e64') else "_e64"
+  if op_val < 256:  # VOPC
+    vdst_off = inst.vdst.offset - 256 if inst.vdst.offset >= 256 else inst.vdst.offset
+    return f"{name}{e64} {s0}, {s1}{cl}" if name.startswith('v_cmpx') else f"{name}{e64} {_fmt_sdst(vdst_off, 1)}, {s0}, {s1}{cl}"
+  if op_val < 384:  # VOP2
+    n = inst.num_srcs() or 2
+    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 op_val < 512:  # VOP1
+    if re.match(r'v_cvt_f32_(bf|fp)8', name) and inst.opsel:
+      os = f" byte_sel:{((inst.opsel & 1) << 1) | ((inst.opsel >> 1) & 1)}"
+    else:
+      os = _opsel_str(inst.opsel, 1, need_opsel, is16_d)
+    if 'v_nop' in name or 'v_pipeflush' in name: return f"{name}{e64}"
+    return f"{name}{e64} {dst}, {s0}{os}{cl}{om}"
+  # Native VOP3
+  n = inst.num_srcs() or 2
+  os = f" byte_sel:{inst.opsel >> 2}" if 'cvt_sr' in name and inst.opsel else _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(reg, neg):
+    s = _lit(inst, reg.offset) if reg.offset == 255 else ("src_scc" if reg.offset == 253 else (reg.fmt() if reg.sz > 1 else _lit(inst, reg.offset)))
+    return f"neg({s})" if neg and reg.offset == 255 else (f"-{s}" if neg else s)
+  s0, s1, s2 = src(inst.src0, inst.neg & 1), src(inst.src1, inst.neg & 2), src(inst.src2, inst.neg & 4)
+  # VOP3SD: _co_ ops (add/sub) without _ci_ have only 2 sources, all others (mad, div_scale, _co_ci_) have 3 sources
+  has_only_two_srcs = '_co_' in name and '_ci_' not in name and 'mad' not in name
+  srcs = f"{s0}, {s1}" if has_only_two_srcs else f"{s0}, {s1}, {s2}"
+  clamp = getattr(inst, 'cm', None) or getattr(inst, 'clmp', 0)
+  return f"{name} {inst.vdst.fmt()}, {_fmt_sdst(inst.sdst, 1)}, {srcs}{' clamp' if clamp else ''}{_omod(inst.omod)}"
+
+def _disasm_vopd(inst: VOPD) -> str:
+  lit = inst._literal
+  op_enum = R4_VOPDOp if _is_r4(inst) else VOPDOp
+  nx, ny = op_enum(inst.opx).name.lower(), op_enum(inst.opy).name.lower()
+  def half(n, vd, s0, vs1):
+    vd, vs1 = _vi(vd), _vi(vs1)
+    if 'mov' in n: return f"{n} v{vd}, {_lit(inst, s0)}"
+    if 'fmamk' in n and lit: return f"{n} v{vd}, {_lit(inst, s0)}, 0x{lit:x}, v{vs1}"
+    if 'fmaak' in n and lit: return f"{n} v{vd}, {_lit(inst, s0)}, v{vs1}, 0x{lit:x}"
+    return f"{n} v{vd}, {_lit(inst, s0)}, v{vs1}"
+  return f"{half(nx, inst.vdstx, inst.srcx0, inst.vsrcx1)} :: {half(ny, inst.vdsty, inst.srcy0, inst.vsrcy1)}"
+
+def _disasm_vop3p(inst: VOP3P) -> str:
+  name = inst.op_name.lower()
+  is_wmma, is_swmmac, n, is_fma_mix = 'wmma' in name, 'swmmac' in name, inst.num_srcs() or 2, 'fma_mix' in name
+  def get_src(reg):
+    return _lit(inst, reg.offset) if reg.offset == 255 else reg.fmt()
+  src0, src1, src2, dst = get_src(inst.src0), get_src(inst.src1), get_src(inst.src2), inst.vdst.fmt()
+  opsel_hi = inst.opsel_hi | (inst.opsel_hi2 << 2)
+  clamp = getattr(inst, 'cm', None) or getattr(inst, 'clmp', 0)
+  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 clamp else [])
+  elif is_swmmac:
+    mods = ([f"index_key:{inst.opsel}"] if inst.opsel 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 clamp else [])
+  else:
+    opsel_hi_default = 7 if n == 3 else 3
+    mods = ([_fmt_bits("op_sel", inst.opsel, n)] if inst.opsel else []) + ([_fmt_bits("op_sel_hi", opsel_hi, n)] if opsel_hi != opsel_hi_default 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 clamp 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_sop1(inst: SOP1) -> str:
+  op, name, cdna = inst.op, inst.op_name.lower(), _is_cdna(inst)
+  # Use get_field_bits for register sizes
+  regs = inst.canonical_op_regs
+  dst_regs, src_regs = regs.get('d', 1), regs.get('s0', 1)
+  src = _lit(inst, inst.ssrc0) if _unwrap(inst.ssrc0) == 255 else _fmt_src(inst.ssrc0, src_regs, cdna)
+  if not cdna:
+    if 'getpc_b64' in name: return f"{name} {_fmt_sdst(inst.sdst, 2)}"
+    if 'setpc_b64' in name or 'rfe_b64' in name: return f"{name} {src}"
+    if 'swappc_b64' in name: return f"{name} {_fmt_sdst(inst.sdst, 2)}, {src}"
+    if 'sendmsg_rtn' in name:
+      v = _unwrap(inst.ssrc0)
+      try: msg_str = MSG(v).name if v != 255 else None  # MSG_RTN_ILLEGAL_MSG (255) not supported by LLVM
+      except ValueError: msg_str = None
+      return f"{name} {_fmt_sdst(inst.sdst, dst_regs)}, sendmsg({msg_str})" if msg_str else f"{name} {_fmt_sdst(inst.sdst, dst_regs)}, 0x{v:x}"
+  sop1_src_only = ('S_ALLOC_VGPR', 'S_SLEEP_VAR', 'S_BARRIER_SIGNAL', 'S_BARRIER_SIGNAL_ISFIRST', 'S_BARRIER_INIT', 'S_BARRIER_JOIN', 'S_SET_GPR_IDX_IDX',
+                   'S_CBRANCH_JOIN')
+  if inst.op_name in sop1_src_only: return f"{name} {src}"
+  if cdna:
+    if 'getpc_b64' in name: return f"{name} {_fmt_sdst(inst.sdst, 2, cdna)}"
+    if 'setpc_b64' in name or 'rfe_b64' in name: return f"{name} {src}"
+    if 'swappc_b64' in name: return f"{name} {_fmt_sdst(inst.sdst, 2, cdna)}, {src}"
+  return f"{name} {_fmt_sdst(inst.sdst, dst_regs, cdna)}, {src}"
+
+def _disasm_sop2(inst: SOP2) -> str:
+  cdna, name = _is_cdna(inst), inst.op_name.lower()
+  lit = inst._literal
+  # Use get_field_bits for register sizes
+  regs = inst.canonical_op_regs
+  dn, s0n, s1n = regs['d'], regs['s0'], regs['s1']
+  s0 = _lit(inst, inst.ssrc0) if _unwrap(inst.ssrc0) == 255 else _fmt_src(inst.ssrc0, s0n, cdna)
+  s1 = _lit(inst, inst.ssrc1) if _unwrap(inst.ssrc1) == 255 else _fmt_src(inst.ssrc1, s1n, cdna)
+  dst = _fmt_sdst(inst.sdst, dn, cdna)
+  if 'fmamk' in name and lit is not None: return f"{name} {dst}, {s0}, 0x{lit:x}, {s1}"
+  if 'fmaak' in name and lit is not None: return f"{name} {dst}, {s0}, {s1}, 0x{lit:x}"
+  if name in ('s_cbranch_g_fork', 's_rfe_restore_b64'): return f"{name} {s0}, {s1}"  # no destination
+  return f"{name} {dst}, {s0}, {s1}"
+
+def _disasm_sopc(inst: SOPC) -> str:
+  cdna, regs, name = _is_cdna(inst), inst.canonical_op_regs, inst.op_name.lower()
+  s0 = _lit(inst, inst.ssrc0) if _unwrap(inst.ssrc0) == 255 else _fmt_src(inst.ssrc0, regs['s0'], cdna)
+  if name == 's_set_gpr_idx_on':
+    imm = _unwrap(inst.ssrc1) & 0xf
+    flags = [n for i, n in enumerate(['SRC0', 'SRC1', 'SRC2', 'DST']) if imm & (1 << i)]
+    return f"{name} {s0}, gpr_idx({','.join(flags)})"
+  s1 = _lit(inst, inst.ssrc1) if _unwrap(inst.ssrc1) == 255 else _fmt_src(inst.ssrc1, regs['s1'], cdna)
+  return f"{name} {s0}, {s1}"
+
+_HWREG_BLACKLIST = {'HW_REG_PC_LO', 'HW_REG_PC_HI', 'HW_REG_IB_DBG1', 'HW_REG_FLUSH_IB', 'HW_REG_SHADER_TBA_LO', 'HW_REG_SHADER_TBA_HI',
+                    'HW_REG_SHADER_FLAT_SCRATCH_LO', 'HW_REG_SHADER_FLAT_SCRATCH_HI', 'HW_REG_SHADER_CYCLES'}
+_HWREG_BLACKLIST_CDNA = {'HW_REG_PC_LO', 'HW_REG_PC_HI', 'HW_REG_IB_DBG1', 'HW_REG_FLUSH_IB', 'HW_REG_SQ_SHADER_TBA_LO', 'HW_REG_SQ_SHADER_TBA_HI',
+                         'HW_REG_SQ_SHADER_TMA_LO', 'HW_REG_SQ_SHADER_TMA_HI', 'HW_REG_SQ_PERF_SNAPSHOT_DATA', 'HW_REG_SQ_PERF_SNAPSHOT_DATA1',
+                         'HW_REG_SQ_PERF_SNAPSHOT_PC_LO', 'HW_REG_SQ_PERF_SNAPSHOT_PC_HI', 'HW_REG_XCC_ID'}
+def _disasm_sopk(inst: SOPK) -> str:
+  op, name, cdna = inst.op, inst.op_name.lower(), _is_cdna(inst)
+  is_rdna4 = _is_r4(inst)
+  hw = HWREG_CDNA if cdna else (HWREG_RDNA4 if is_rdna4 else HWREG)
+  blacklist = _HWREG_BLACKLIST_CDNA if cdna else _HWREG_BLACKLIST
+  def fmt_hwreg(hid, hoff, hsz):
+    try: hr_name = hw(hid).name.replace("HW_REG_WAVE_", "HW_REG_")
+    except ValueError: return f"0x{inst.simm16:x}"
+    if hr_name in blacklist: return f"0x{inst.simm16:x}"
+    return f"hwreg({hr_name})" if hoff == 0 and hsz == 32 else f"hwreg({hr_name}, {hoff}, {hsz})"
+  if name == 's_setreg_imm32_b32':
+    hid, hoff, hsz = inst.simm16 & 0x3f, (inst.simm16 >> 6) & 0x1f, ((inst.simm16 >> 11) & 0x1f) + 1
+    return f"{name} {fmt_hwreg(hid, hoff, hsz)}, 0x{inst._literal:x}"
+  if name == 's_version': return f"{name} 0x{inst.simm16:x}"
+  if name in ('s_setreg_b32', 's_getreg_b32'):
+    hid, hoff, hsz = inst.simm16 & 0x3f, (inst.simm16 >> 6) & 0x1f, ((inst.simm16 >> 11) & 0x1f) + 1
+    hs = fmt_hwreg(hid, hoff, hsz)
+    return f"{name} {hs}, {_fmt_sdst(inst.sdst, 1, cdna)}" if 'setreg' in name else f"{name} {_fmt_sdst(inst.sdst, 1, cdna)}, {hs}"
+  if name in ('s_subvector_loop_begin', 's_subvector_loop_end'):
+    return f"{name} {_fmt_sdst(inst.sdst, 1)}, 0x{inst.simm16:x}"
+  return f"{name} {_fmt_sdst(inst.sdst, inst.canonical_op_regs['d'], cdna)}, 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()} {inst.vdst.fmt()}, {_lit(inst, inst.src0, inst.neg & 1)}, {_lit(inst, inst.src1, inst.neg & 2)}, {_lit(inst, inst.src2, inst.neg & 4)}" + (" " + mods if mods else "")
+
+DISASM_HANDLERS: dict[type, Callable[..., str]] = {
+  VOP1: _disasm_vop1, VOP1_SDST: _disasm_vop1, VOP1_SDST_LIT: _disasm_vop1, VOP1_LIT: _disasm_vop1,
+  VOP2: _disasm_vop2, VOP2_LIT: _disasm_vop2, VOPC: _disasm_vopc, VOPC_LIT: _disasm_vopc,
+  VOP3: _disasm_vop3, VOP3_SDST: _disasm_vop3, VOP3_SDST_LIT: _disasm_vop3, VOP3_LIT: _disasm_vop3, VOP3SD: _disasm_vop3sd, VOP3SD_LIT: _disasm_vop3sd,
+  VOPD: _disasm_vopd, VOPD_LIT: _disasm_vopd, VOP3P: _disasm_vop3p, VOP3P_LIT: _disasm_vop3p,
+  VINTERP: _disasm_vinterp, SOPP: _disasm_sopp, SMEM: _disasm_smem, DS: _disasm_ds, FLAT: _disasm_flat, GLOBAL: _disasm_flat, SCRATCH: _disasm_flat,
+  SOP1: _disasm_sop1, SOP1_LIT: _disasm_sop1, SOP2: _disasm_sop2, SOP2_LIT: _disasm_sop2,
+  SOPC: _disasm_sopc, SOPC_LIT: _disasm_sopc, SOPK: _disasm_sopk, SOPK_LIT: _disasm_sopk,
+  # RDNA4
+  R4_VOP1: _disasm_vop1, R4_VOP1_SDST: _disasm_vop1, R4_VOP1_SDST_LIT: _disasm_vop1, R4_VOP1_LIT: _disasm_vop1,
+  R4_VOP2: _disasm_vop2, R4_VOP2_LIT: _disasm_vop2, R4_VOPC: _disasm_vopc, R4_VOPC_LIT: _disasm_vopc,
+  R4_VOP3: _disasm_vop3, R4_VOP3_SDST: _disasm_vop3, R4_VOP3_SDST_LIT: _disasm_vop3, R4_VOP3_LIT: _disasm_vop3,
+  R4_VOP3SD: _disasm_vop3sd, R4_VOP3SD_LIT: _disasm_vop3sd, R4_VOP3P: _disasm_vop3p, R4_VOP3P_LIT: _disasm_vop3p,
+  R4_FLAT: _disasm_flat, R4_GLOBAL: _disasm_flat, R4_SCRATCH: _disasm_flat,
+  R4_VOPD: _disasm_vopd, R4_VOPD_LIT: _disasm_vopd, R4_VINTERP: _disasm_vinterp, R4_SOPP: _disasm_sopp, R4_SMEM: _disasm_smem, R4_DS: _disasm_ds,
+  R4_SOP1: _disasm_sop1, R4_SOP1_LIT: _disasm_sop1, R4_SOP2: _disasm_sop2, R4_SOP2_LIT: _disasm_sop2,
+  R4_SOPC: _disasm_sopc, R4_SOPC_LIT: _disasm_sopc, R4_SOPK: _disasm_sopk, R4_SOPK_LIT: _disasm_sopk}
+
+def disasm(inst: Inst) -> str: return DISASM_HANDLERS[type(inst)](inst)
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# CDNA DISASSEMBLER SUPPORT
+# ═══════════════════════════════════════════════════════════════════════════════
+
+from extra.assembly.amd.autogen.cdna.ins import (VOP1 as CDNA_VOP1, VOP1_LIT as CDNA_VOP1_LIT,
+  VOP1_SDWA as CDNA_VOP1_SDWA, VOP1_DPP16 as CDNA_VOP1_DPP16,
+  VOP2 as CDNA_VOP2, VOP2_LIT as CDNA_VOP2_LIT, VOP2_SDWA as CDNA_VOP2_SDWA, VOP2_DPP16 as CDNA_VOP2_DPP16,
+  VOPC as CDNA_VOPC, VOPC_LIT as CDNA_VOPC_LIT, VOPC_SDWA_SDST as CDNA_VOPC_SDWA_SDST,
+  VOP3 as CDNA_VOP3, VOP3_SDST as CDNA_VOP3_SDST, VOP3SD as CDNA_VOP3SD, VOP3P as CDNA_VOP3P, VOP3P_MFMA as CDNA_VOP3P_MFMA, VOP3PX2 as CDNA_VOP3PX2,
+  SOP1 as CDNA_SOP1, SOP1_LIT as CDNA_SOP1_LIT, SOP2 as CDNA_SOP2, SOP2_LIT as CDNA_SOP2_LIT,
+  SOPC as CDNA_SOPC, SOPC_LIT as CDNA_SOPC_LIT, SOPK as CDNA_SOPK, SOPK_LIT as CDNA_SOPK_LIT,
+  SOPP as CDNA_SOPP, SMEM as CDNA_SMEM, DS as CDNA_DS,
+  FLAT as CDNA_FLAT, GLOBAL as CDNA_GLOBAL, SCRATCH as CDNA_SCRATCH, MUBUF as CDNA_MUBUF)
+
+def _cdna_src(inst, v, neg, abs_=0, n=1):
+  s = _lit(inst, v) if v == 255 else _fmt_src(v, n, cdna=True)
+  if abs_: s = f"|{s}|"
+  return f"neg({s})" if neg and v == 255 else (f"-{s}" if neg else s)
+
+_CDNA_VOP3_ALIASES = {'v_fmac_f64': 'v_mul_legacy_f32', 'v_dot2c_f32_bf16': 'v_mac_f32'}
+
+def _disasm_vop3a(inst) -> str:
+  op_val = inst.op.value if hasattr(inst.op, 'value') else inst.op
+  name = inst.op_name.lower() or f'vop3a_op_{op_val}'
+  n = inst.num_srcs() or _num_srcs(inst)
+  cl, om = " clamp" if inst.clmp else "", _omod(inst.omod)
+  # _sr_ instructions use 4-element op_sel (src2 for byte selection)
+  opsel_n = 3 if '_sr_' in name and n == 2 else n
+  opsel = _opsel_str(inst.opsel, opsel_n, inst.opsel != 0, False)
+  orig_name = name
+  name = _CDNA_VOP3_ALIASES.get(name, name)
+  if name != orig_name:
+    s0, s1 = _cdna_src(inst, inst.src0, inst.neg&1, inst.abs&1, 1), _cdna_src(inst, inst.src1, inst.neg&2, inst.abs&2, 1)
+    s2 = ""
+    dst = _vreg(inst.vdst)
+  else:
+    regs = inst.canonical_op_regs
+    dregs, r0, r1, r2 = regs['d'], regs['s0'], regs['s1'], regs['s2']
+    s0, s1, s2 = _cdna_src(inst, inst.src0, inst.neg&1, inst.abs&1, r0), _cdna_src(inst, inst.src1, inst.neg&2, inst.abs&2, r1), _cdna_src(inst, inst.src2, inst.neg&4, inst.abs&4, r2)
+    dst = _vreg(inst.vdst, dregs) if dregs > 1 else _vreg(inst.vdst)
+  if op_val >= 512:
+    return f"{name} {dst}, {s0}, {s1}, {s2}{opsel}{cl}{om}" if n == 3 else f"{name} {dst}, {s0}, {s1}{opsel}{cl}{om}"
+  if op_val < 256:
+    # VOPC: vdst is actually sdst (SGPR pair), but VGPRField adds 256 to the offset
+    sdst_val = _unwrap(inst.vdst)
+    if sdst_val >= 256: sdst_val -= 256
+    sdst = _fmt_sdst(sdst_val, 2, cdna=True)
+    return f"{name} {sdst}, {s0}, {s1}{cl}"
+  if 320 <= op_val < 512:
+    if name in ('v_nop', 'v_clrexcp', 'v_nop_e64', 'v_clrexcp_e64'): return name.replace('_e64', '')
+    return f"{name} {dst}, {s0}{cl}{om}"
+  if name == 'v_cndmask_b32':
+    s2 = _fmt_src(inst.src2, 2, cdna=True)
+    return f"{name} {dst}, {s0}, {s1}, {s2}{cl}{om}"
+  return f"{name} {dst}, {s0}, {s1}, {s2}{opsel}{cl}{om}" if n == 3 else f"{name} {dst}, {s0}, {s1}{opsel}{cl}{om}"
+
+def _disasm_vop3b(inst) -> str:
+  op_val = inst.op.value if hasattr(inst.op, 'value') else inst.op
+  name, cdna = inst.op_name.lower() or f'vop3b_op_{op_val}', _is_cdna(inst)
+  n = inst.num_srcs() or _num_srcs(inst)
+  regs = inst.canonical_op_regs
+  dregs, r0, r1, r2 = regs['d'], regs['s0'], regs['s1'], regs['s2']
+  s0, s1, s2 = _cdna_src(inst, inst.src0, inst.neg&1, n=r0), _cdna_src(inst, inst.src1, inst.neg&2, n=r1), _cdna_src(inst, inst.src2, inst.neg&4, n=r2)
+  # CDNA VOP3_SDST uses vdst field for sdst (but vdst adds 256), RDNA uses separate sdst field
+  sdst_val = getattr(inst, 'sdst', None)
+  if sdst_val is None and hasattr(inst, 'vdst'):
+    sdst_val = _unwrap(inst.vdst)
+    if sdst_val >= 256: sdst_val -= 256  # VGPRField adds 256, remove it for SGPR
+  # For CDNA VOP3_SDST (VOPC->VOP3), vdst is the scalar dest (sdst), there's no vdst output
+  if cdna and 'v_cmp' in name:
+    sdst = _fmt_sdst(sdst_val, 2, cdna=True)
+    cl, om = " clamp" if inst.clmp else "", _omod(inst.omod)
+    return f"{name} {sdst}, {s0}, {s1}{cl}{om}"
+  dst = _vreg(inst.vdst, dregs) if dregs > 1 else _vreg(inst.vdst)
+  sdst = _fmt_sdst(sdst_val, 2, cdna=cdna)
+  cl, om = " clamp" if inst.clmp else "", _omod(inst.omod)
+  if name in ('v_addc_co_u32', 'v_subb_co_u32', 'v_subbrev_co_u32'):
+    s2 = _fmt_src(inst.src2, 2, cdna=cdna)
+  return f"{name} {dst}, {sdst}, {s0}, {s1}, {s2}{cl}{om}" if n == 3 else f"{name} {dst}, {sdst}, {s0}, {s1}{cl}{om}"
+
+def _disasm_cdna_vop3p(inst) -> str:
+  name, n = inst.op_name.lower(), inst.num_srcs() or 2
+  is_mfma = 'mfma' in name or 'smfmac' in name
+  is_accvgpr = 'accvgpr' in name
+  get_src = lambda v, sc: _lit(inst, v) if v == 255 else _fmt_src(v, sc, cdna=True)
+
+  # Handle accvgpr read/write (accumulator register operations)
+  if is_accvgpr:
+    src0_off = _unwrap(inst.src0)
+    vdst_off = _vi(inst.vdst)
+    if 'read' in name:
+      # v_accvgpr_read_b32 vN, aM - reads from accumulator to VGPR
+      return f"{name}_b32 v{vdst_off}, a{src0_off - 256 if src0_off >= 256 else src0_off}"
+    if 'write' in name:
+      # v_accvgpr_write_b32 aM, src - writes to accumulator from source
+      src = _lit(inst, inst.src0) if src0_off == 255 else (f"v{src0_off - 256}" if src0_off >= 256 else decode_src(src0_off, cdna=True))
+      return f"{name}_b32 a{vdst_off}, {src}"
+
+  # Handle v_mfma_ld_scale_b32 - special 2-operand format: v_mfma_ld_scale_b32 src0, src1
+  if 'ld_scale' in name:
+    src0, src1 = get_src(inst.src0, 1), get_src(inst.src1, 1)
+    mods = ([_fmt_bits("op_sel", inst.opsel, 2)] if inst.opsel else []) + \
+           ([_fmt_bits("op_sel_hi", inst.opsel_hi, 2)] if inst.opsel_hi != 3 else [])
+    return f"{name} {src0}, {src1}{' ' + ' '.join(mods) if mods else ''}"
+
+  # Handle MFMA instructions with accumulator destinations
+  if is_mfma:
+    regs = inst.canonical_op_regs
+    dregs, r0, r1, r2 = regs['d'], regs['s0'], regs['s1'], regs['s2']
+    # Infer register counts from instruction name if not in operands table (e.g., v_mfma_f32_32x32x4_xf32)
+    if dregs == 1:
+      if '32x32' in name: dregs, r0, r1, r2 = 16, 2, 2, 16
+      elif '16x16' in name: dregs, r0, r1, r2 = 4, 2, 2, 4
+    # MFMA reuses VOP3P fields differently: clmp -> acc_cd (dest is acc), opsel_hi -> acc (src1/src2 are acc)
+    # acc field (bits 60-59): bit 0 = src2 is acc (always for MFMA), bit 1 = src1 is acc
+    acc = inst.opsel_hi  # opsel_hi field maps to acc for MFMA
+    acc_cd = inst.clmp   # clmp field maps to acc_cd for MFMA (dest is accumulator)
+    is_smfmac = 'smfmac' in name  # SMFMAC has different operand semantics
+    # Format sources: src0 is always VGPR, src1/src2 depend on acc bits
+    def mfma_src(v, sc, is_acc):
+      v = _unwrap(v)
+      if v == 255: return _lit(inst, v)
+      if 128 <= v <= 208 or 240 <= v <= 248: return _lit(inst, v)
+      base = v - 256 if v >= 256 else v
+      if is_acc: return _areg(base, sc)
+      return _vreg(base, sc)
+    src0 = get_src(inst.src0, r0)  # src0 is always VGPR
+    src1 = mfma_src(inst.src1, r1, acc & 2)  # bit 1 = src1 is acc
+    # For SMFMAC, src2 is always a VGPR index (1 register), not accumulator
+    src2 = _vreg(inst.src2) if is_smfmac else mfma_src(inst.src2, r2, acc_cd)
+    dst = _areg(inst.vdst, dregs) if acc_cd else _vreg(inst.vdst, dregs)
+    # MFMA uses neg:[...] not neg_lo:[...], and doesn't support op_sel_hi or clamp
+    # Only f64 MFMA instructions support neg modifier
+    # f8f6f4 MFMA instructions support cbsz/blgp modifiers
+    mods = []
+    if 'f8f6f4' in name:
+      if inst.neg_hi: mods.append(f"cbsz:{inst.neg_hi}")
+      if inst.neg: mods.append(f"blgp:{inst.neg}")
+    elif inst.neg and 'f64' in name:
+      mods.append(_fmt_bits("neg", inst.neg, n))
+    return f"{name} {dst}, {src0}, {src1}, {src2}{' ' + ' '.join(mods) if mods else ''}"
+
+  # Standard VOP3P instructions
+  src0, src1, src2, dst = get_src(inst.src0, 1), get_src(inst.src1, 1), get_src(inst.src2, 1), _vreg(inst.vdst)
+  opsel_hi = inst.opsel_hi  # CDNA VOP3P only has 2 bits for opsel_hi (no opsel_hi2)
+  opsel_hi_default = 3  # CDNA default is 0b11 (2 bits), not 0b111 like RDNA
+  mods = ([_fmt_bits("op_sel", inst.opsel, n)] if inst.opsel else []) + ([_fmt_bits("op_sel_hi", opsel_hi, n)] if opsel_hi != opsel_hi_default 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_mubuf(inst) -> str:
+  name = inst.op_name.lower()
+  # Determine vdata register count from instruction name
+  nregs = 4 if 'xyzw' in name else 3 if 'xyz' in name else 2 if 'xy' in name or 'x2' in name or 'f64' in name or 'dwordx2' in name else 1
+  vdata = _vreg(inst.vdata, nregs)
+  vaddr = _vreg(inst.vaddr) if inst.offen or inst.idxen else None
+  srsrc = str(inst.srsrc)
+  soffset_val = _unwrap(inst.soffset)
+  soffset = f"s{soffset_val}" if soffset_val < 128 else "off"
+  offset = f" offset:{inst.offset}" if inst.offset else ""
+  offen = " offen" if inst.offen else ""
+  idxen = " idxen" if inst.idxen else ""
+  lds = " lds" if inst.lds else ""
+  sc0 = " sc0" if inst.sc0 else ""
+  sc1 = " sc1" if inst.sc1 else ""
+  nt = " nt" if inst.nt else ""
+  # Handle special cases
+  if name in ('buffer_wbl2', 'buffer_inv'):
+    return f"{name}{sc0}{sc1}"
+  if vaddr:
+    return f"{name} {vdata}, {vaddr}, {srsrc}, {soffset}{offen}{idxen}{offset}{sc0}{nt}{sc1}{lds}"
+  return f"{name} {vdata}, off, {srsrc}, {soffset}{offset}{sc0}{nt}{sc1}{lds}"
+
+_SDWA_SEL = {0: 'BYTE_0', 1: 'BYTE_1', 2: 'BYTE_2', 3: 'BYTE_3', 4: 'WORD_0', 5: 'WORD_1', 6: 'DWORD'}
+
+def _disasm_vop1_sdwa(inst) -> str:
+  name = inst.op_name.lower().replace('_e32', '')
+  regs = inst.canonical_op_regs
+  dst = _vreg(inst.vdst, regs['d'])
+  # When s0=1, vsrc0 is SGPR/constant (VGPRField adds 256, so subtract it back)
+  if inst.s0 == 0: src0 = _vreg(inst.vsrc0, regs['s0'])
+  else:
+    raw = _unwrap(inst.vsrc0) - 256  # VGPRField adds 256
+    src0 = decode_src(raw, cdna=True)  # handles SGPRs, constants, specials
+  src0_sel = _SDWA_SEL.get(inst.src0_sel, f'SEL{inst.src0_sel}')
+  mods = []
+  if inst.clmp: mods.append("clamp")
+  if inst.omod == 1: mods.append("mul:2")
+  elif inst.omod == 2: mods.append("mul:4")
+  elif inst.omod == 3: mods.append("div:2")
+  mods.append(f"src0_sel:{src0_sel}")
+  return f"{name}_sdwa {dst}, {src0} {' '.join(mods)}"
+
+def _decode_dpp(dpp: int) -> str:
+  """Decode DPP control value to string."""
+  if dpp < 0x100: return f"quad_perm:[{dpp&3},{(dpp>>2)&3},{(dpp>>4)&3},{(dpp>>6)&3}]"
+  if 0x100 <= dpp <= 0x10f: return f"row_shl:{dpp & 0xf}"
+  if 0x110 <= dpp <= 0x11f: return f"row_shr:{dpp & 0xf}"
+  if 0x120 <= dpp <= 0x12f: return f"row_ror:{dpp & 0xf}"
+  if dpp == 0x130: return "wave_shl:1"
+  if dpp == 0x134: return "wave_rol:1"
+  if dpp == 0x138: return "wave_shr:1"
+  if dpp == 0x13c: return "wave_ror:1"
+  if dpp == 0x140: return "row_mirror"
+  if dpp == 0x141: return "row_half_mirror"
+  if dpp == 0x142: return "row_bcast:15"
+  if dpp == 0x143: return "row_bcast:31"
+  if 0x150 <= dpp <= 0x15f: return f"row_newbcast:{dpp & 0xf}"
+  if 0x160 <= dpp <= 0x16f: return f"row_share:{dpp & 0xf}"
+  if 0x170 <= dpp <= 0x17f: return f"row_xmask:{dpp & 0xf}"
+  return f"dpp:{dpp:#x}"
+
+def _disasm_vop1_dpp(inst) -> str:
+  name = inst.op_name.lower().replace('_e32', '')
+  regs = inst.canonical_op_regs
+  dst, src0 = _vreg(inst.vdst, regs['d']), _vreg(inst.vsrc0, regs['s0'])
+  dpp_str = _decode_dpp(inst.dpp)
+  mods = [dpp_str]
+  if inst.row_mask != 0xf: mods.append(f"row_mask:{inst.row_mask:#x}")
+  if inst.bank_mask != 0xf: mods.append(f"bank_mask:{inst.bank_mask:#x}")
+  if inst.bc: mods.append("bound_ctrl:1")
+  return f"{name}_dpp {dst}, {src0} {' '.join(mods)}"
+
+def _disasm_vop2_sdwa(inst) -> str:
+  name, cdna = inst.op_name.lower().replace('_e32', ''), _is_cdna(inst)
+  regs = inst.canonical_op_regs
+  dst = _vreg(inst.vdst, regs['d'])
+  # When s0/s1=1, vsrc is SGPR/constant (VGPRField adds 256, so subtract it back)
+  src0 = _vreg(inst.vsrc0, regs['s0']) if inst.s0 == 0 else decode_src(_unwrap(inst.vsrc0) - 256, cdna)
+  src1 = _vreg(inst.vsrc1, regs['s1']) if inst.s1 == 0 else decode_src(_unwrap(inst.vsrc1) - 256, cdna)
+  src0_sel = _SDWA_SEL.get(inst.src0_sel, f'SEL{inst.src0_sel}')
+  src1_sel = _SDWA_SEL.get(inst.src1_sel, f'SEL{inst.src1_sel}')
+  mods = []
+  if inst.clmp: mods.append("clamp")
+  if inst.omod == 1: mods.append("mul:2")
+  elif inst.omod == 2: mods.append("mul:4")
+  elif inst.omod == 3: mods.append("div:2")
+  if inst.src0_sel != 6: mods.append(f"src0_sel:{src0_sel}")
+  if inst.src1_sel != 6: mods.append(f"src1_sel:{src1_sel}")
+  mods_str = ' '.join(mods) if mods else ""
+  # CDNA carry instructions and cndmask need vcc operands
+  if cdna and name in _VOP2_CARRY_OUT: return f"{name}_sdwa {dst}, vcc, {src0}, {src1} {mods_str}".strip()
+  if cdna and name in _VOP2_CARRY_INOUT: return f"{name}_sdwa {dst}, vcc, {src0}, {src1}, vcc {mods_str}".strip()
+  if cdna and name == 'v_cndmask_b32': return f"{name}_sdwa {dst}, {src0}, {src1}, vcc {mods_str}".strip()
+  return f"{name}_sdwa {dst}, {src0}, {src1} {mods_str}".strip()
+
+def _disasm_vop2_dpp(inst) -> str:
+  name, cdna = inst.op_name.lower().replace('_e32', ''), _is_cdna(inst)
+  regs = inst.canonical_op_regs
+  dst, src0, src1 = _vreg(inst.vdst, regs['d']), _vreg(inst.vsrc0, regs['s0']), _vreg(inst.vsrc1, regs['s1'])
+  dpp_str = _decode_dpp(inst.dpp)
+  mods = [dpp_str]
+  if inst.row_mask != 0xf: mods.append(f"row_mask:{inst.row_mask:#x}")
+  if inst.bank_mask != 0xf: mods.append(f"bank_mask:{inst.bank_mask:#x}")
+  if inst.bc: mods.append("bound_ctrl:1")
+  # CDNA carry instructions and cndmask need vcc operands
+  if cdna and name in _VOP2_CARRY_OUT: return f"{name}_dpp {dst}, vcc, {src0}, {src1} {' '.join(mods)}"
+  if cdna and name in _VOP2_CARRY_INOUT: return f"{name}_dpp {dst}, vcc, {src0}, {src1}, vcc {' '.join(mods)}"
+  if cdna and name == 'v_cndmask_b32': return f"{name}_dpp {dst}, {src0}, {src1}, vcc {' '.join(mods)}"
+  return f"{name}_dpp {dst}, {src0}, {src1} {' '.join(mods)}"
+
+def _disasm_vopc_sdwa(inst) -> str:
+  name = inst.op_name.lower().replace('_e32', '')
+  regs = inst.canonical_op_regs
+  sdst = _fmt_sdst(inst.sdst, 2, cdna=True)
+  src0 = _vreg(inst.vsrc0, regs['s0']) if getattr(inst, 's0', 0) == 0 else decode_src(_unwrap(inst.vsrc0) - 256, cdna=True)
+  src1 = _vreg(inst.vsrc1, regs['s1']) if getattr(inst, 's1', 0) == 0 else decode_src(_unwrap(inst.vsrc1) - 256, cdna=True)
+  src0_sel = _SDWA_SEL.get(inst.src0_sel, f'SEL{inst.src0_sel}')
+  src1_sel = _SDWA_SEL.get(inst.src1_sel, f'SEL{inst.src1_sel}')
+  mods = []
+  if inst.src0_sel != 6: mods.append(f"src0_sel:{src0_sel}")
+  if inst.src1_sel != 6: mods.append(f"src1_sel:{src1_sel}")
+  return f"{name}_sdwa {sdst}, {src0}, {src1} {' '.join(mods)}".strip()
+
+def _disasm_vop3px2(inst) -> str:
+  """VOP3PX2 disassembler for scaled MFMA instructions."""
+  name = inst.op_name.lower()
+  regs = inst.canonical_op_regs
+  dregs, r2 = regs['d'], regs['s2']
+  # F8F6F4 MFMA: CBSZ selects matrix A format, BLGP selects matrix B format
+  # VGPRs: FP8/BF8(0,1)=8, FP6/BF6(2,3)=6, FP4(4)=4
+  vgprs = {0: 8, 1: 8, 2: 6, 3: 6, 4: 4}
+  r0, r1 = vgprs.get(inst.cbsz, 8), vgprs.get(inst.blgp, 8)
+  def mfma_src(v, sc, is_acc):
+    v = _unwrap(v)
+    if v == 255: return _lit(inst, v)
+    base = v - 256 if v >= 256 else v
+    return _areg(base, sc) if is_acc else _vreg(base, sc)
+  src0, src1, src2 = mfma_src(inst.src0, r0, False), mfma_src(inst.src1, r1, inst.acc & 2), mfma_src(inst.src2, r2, inst.acc_cd)
+  dst = _areg(inst.vdst, dregs) if inst.acc_cd else _vreg(inst.vdst, dregs)
+  scale_src0, scale_src1 = _vreg(inst.scale_src0), _vreg(inst.scale_src1)
+  mods = []
+  if inst.opsel: mods.append(_fmt_bits("op_sel", inst.opsel, 3))
+  if inst.opsel_hi != 0: mods.append(_fmt_bits("op_sel_hi", inst.opsel_hi, 3))
+  if inst.neg: mods.append(_fmt_bits("neg", inst.neg, 3))
+  if inst.cbsz: mods.append(f"cbsz:{inst.cbsz}")
+  if inst.blgp: mods.append(f"blgp:{inst.blgp}")
+  return f"{name} {dst}, {src0}, {src1}, {src2}, {scale_src0}, {scale_src1}{' ' + ' '.join(mods) if mods else ''}"
+
+DISASM_HANDLERS.update({CDNA_VOP1: _disasm_vop1, CDNA_VOP1_LIT: _disasm_vop1,
+  CDNA_VOP1_SDWA: _disasm_vop1_sdwa, CDNA_VOP1_DPP16: _disasm_vop1_dpp,
+  CDNA_VOP2: _disasm_vop2, CDNA_VOP2_LIT: _disasm_vop2,
+  CDNA_VOP2_SDWA: _disasm_vop2_sdwa, CDNA_VOP2_DPP16: _disasm_vop2_dpp,
+  CDNA_VOPC: _disasm_vopc, CDNA_VOPC_LIT: _disasm_vopc, CDNA_VOPC_SDWA_SDST: _disasm_vopc_sdwa,
+  CDNA_SOP1: _disasm_sop1, CDNA_SOP1_LIT: _disasm_sop1, CDNA_SOP2: _disasm_sop2, CDNA_SOP2_LIT: _disasm_sop2,
+  CDNA_SOPC: _disasm_sopc, CDNA_SOPC_LIT: _disasm_sopc, CDNA_SOPK: _disasm_sopk, CDNA_SOPK_LIT: _disasm_sopk, CDNA_SOPP: _disasm_sopp,
+  CDNA_SMEM: _disasm_smem, CDNA_DS: _disasm_ds, CDNA_FLAT: _disasm_flat, CDNA_GLOBAL: _disasm_flat, CDNA_SCRATCH: _disasm_flat,
+  CDNA_VOP3: _disasm_vop3a, CDNA_VOP3_SDST: _disasm_vop3b, CDNA_VOP3SD: _disasm_vop3b, CDNA_VOP3P: _disasm_cdna_vop3p, CDNA_VOP3P_MFMA: _disasm_cdna_vop3p,
+  CDNA_MUBUF: _disasm_mubuf, CDNA_VOP3PX2: _disasm_vop3px2})

extra/assembly/amd/dsl.py

@@ -0,0 +1,450 @@
+# dsl.py - clean DSL for AMD assembly
+from typing import Any
+
+# ══════════════════════════════════════════════════════════════
+# Registers - unified src encoding space (0-511)
+# ══════════════════════════════════════════════════════════════
+
+class Reg:
+  # Register names vary by arch: RDNA has NULL@124/M0@125, CDNA has M0@124/reserved@125
+  # RDNA4 has DPP8@233, CDNA has SDWA@249/DPP@250/VCCZ@251/EXECZ@252
+  _NAMES = {102: "FLAT_SCRATCH_LO", 103: "FLAT_SCRATCH_HI", 104: "XNACK_MASK_LO", 105: "XNACK_MASK_HI",
+            106: "VCC_LO", 107: "VCC_HI", 124: "NULL", 125: "M0", 126: "EXEC_LO", 127: "EXEC_HI",
+            233: "DPP8", 234: "DPP8FI", 235: "SHARED_BASE", 236: "SHARED_LIMIT", 237: "PRIVATE_BASE", 238: "PRIVATE_LIMIT",
+            240: "0.5", 241: "-0.5", 242: "1.0", 243: "-1.0", 244: "2.0", 245: "-2.0", 246: "4.0", 247: "-4.0",
+            248: "INV_2PI", 249: "SDWA", 250: "DPP", 251: "VCCZ", 252: "EXECZ", 253: "SCC", 254: "SRC_LDS_DIRECT", 255: "LIT"}
+  _PAIRS = {106: "VCC", 126: "EXEC"}
+
+  def __init__(self, offset: int = 0, sz: int = 512, *, neg: bool = False, abs_: bool = False, hi: bool = False):
+    self.offset, self.sz = offset, sz
+    self.neg, self.abs_, self.hi = neg, abs_, hi
+
+  def __hash__(self): return hash((self.offset, self.sz, self.neg, self.abs_, self.hi))
+  def __getitem__(self, key):
+    if isinstance(key, slice):
+      start, stop = key.start or 0, key.stop or (self.sz - 1)
+      if start < 0 or stop >= self.sz: raise RuntimeError(f"slice [{start}:{stop}] out of bounds for size {self.sz}")
+      return Reg(self.offset + start, stop - start + 1)
+    if key < 0 or key >= self.sz: raise RuntimeError(f"index {key} out of bounds for size {self.sz}")
+    return Reg(self.offset + key, 1)
+  def __eq__(self, other):
+    if isinstance(other, Reg):
+      return (self.offset == other.offset and self.sz == other.sz and
+              self.neg == other.neg and self.abs_ == other.abs_ and self.hi == other.hi)
+    return NotImplemented
+  def __add__(self, other):
+    if isinstance(other, int): return Reg(self.offset + other, self.sz)
+    return NotImplemented
+  def __neg__(self) -> 'Reg': return Reg(self.offset, self.sz, neg=not self.neg, abs_=self.abs_, hi=self.hi)
+  def __abs__(self) -> 'Reg': return Reg(self.offset, self.sz, neg=self.neg, abs_=True, hi=self.hi)
+  @property
+  def h(self) -> 'Reg': return Reg(self.offset, self.sz, neg=self.neg, abs_=self.abs_, hi=True)
+  @property
+  def l(self) -> 'Reg': return Reg(self.offset, self.sz, neg=self.neg, abs_=self.abs_, hi=False)
+  def fmt(self, sz=None, parens=False, upper=False) -> str:
+    o, sz = self.offset, sz or self.sz
+    l, r = ("[", "]") if parens or sz > 1 else ("", "")  # brackets for multi-reg or when parens=True
+    if 256 <= o < 512: idx = o - 256; base = f"v{l}{idx}{r}" if sz == 1 else f"v[{idx}:{idx + sz - 1}]"
+    elif o < 106: base = f"s{l}{o}{r}" if sz == 1 else f"s[{o}:{o + sz - 1}]"
+    elif sz == 2 and o in self._PAIRS: base = self._PAIRS[o] if upper else self._PAIRS[o].lower()
+    elif o in self._NAMES: base = self._NAMES[o] if upper else self._NAMES[o].lower()  # special regs (any sz)
+    elif 108 <= o < 124: idx = o - 108; base = f"ttmp{l}{idx}{r}" if sz == 1 else f"ttmp[{idx}:{idx + sz - 1}]"
+    elif 128 <= o <= 192: base = str(o - 128)  # inline int constants (0-64)
+    elif 193 <= o <= 208: base = str(-(o - 192))  # inline negative int constants (-1 to -16)
+    else: raise RuntimeError(f"unknown register: offset={o}, sz={sz}")
+    if self.hi: base += ".h"
+    if self.abs_: base = f"abs({base})" if upper else f"|{base}|"
+    if self.neg: base = f"-{base}"
+    return base
+  def __repr__(self): return self.fmt(parens=True, upper=True)
+
+# Full src encoding space
+src = Reg(0, 512)
+
+# Slices for each region (inclusive end)
+s = src[0:105]           # SGPR0-105
+VCC_LO = src[106]
+VCC_HI = src[107]
+VCC = src[106:107]
+ttmp = src[108:123]      # TTMP0-15
+NULL = OFF = src[124]
+M0 = src[125]
+EXEC_LO = src[126]
+EXEC_HI = src[127]
+EXEC = src[126:127]
+# 128: 0, 129-192: integers 1-64, 193-208: integers -1 to -16
+# 240-248: float constants (0.5, -0.5, 1.0, -1.0, 2.0, -2.0, 4.0, -4.0, 1/(2*PI))
+INV_2PI = src[248]
+SDWA = src[249]
+DPP = DPP16 = src[250]
+VCCZ = src[251]
+EXECZ = src[252]
+SCC = src[253]
+SRC_LDS_DIRECT = src[254]
+LIT = src[255]           # literal constant marker
+v = src[256:511]         # VGPR0-255
+
+# ══════════════════════════════════════════════════════════════
+# BitField
+# ══════════════════════════════════════════════════════════════
+
+class _Bits:
+  """Helper for defining bit fields with slice syntax: bits[hi:lo] or bits[n]."""
+  def __getitem__(self, key) -> 'BitField': return BitField(key.start, key.stop) if isinstance(key, slice) else BitField(key, key)
+bits = _Bits()
+
+class BitField:
+  name: str | None
+  def __init__(self, hi: int, lo: int, default: int = 0):
+    self.hi, self.lo, self.default, self.name, self.mask = hi, lo, default, None, (1 << (hi - lo + 1)) - 1
+  def __set_name__(self, owner, name: str): self.name = name
+  def __eq__(self, other) -> 'FixedBitField':  # type: ignore[override]
+    if isinstance(other, int): return FixedBitField(self.hi, self.lo, other)
+    raise TypeError(f"BitField.__eq__ expects int, got {type(other).__name__}")
+  def enum(self, enum_cls) -> 'EnumBitField': return EnumBitField(self.hi, self.lo, enum_cls)
+  def encode(self, val) -> int:
+    assert isinstance(val, int), f"BitField.encode expects int, got {type(val).__name__}"
+    return val
+  def decode(self, val): return val
+  def set(self, raw: int, val) -> int:
+    if val is None: val = self.default
+    encoded = self.encode(val)
+    # Handle signed values: convert negative to 2's complement
+    if encoded < 0: encoded = encoded & self.mask
+    if encoded < 0 or encoded > self.mask: raise RuntimeError(f"field '{self.name}': value {encoded} doesn't fit in {self.hi - self.lo + 1} bits")
+    return (raw & ~(self.mask << self.lo)) | (encoded << self.lo)
+  def __get__(self, obj, objtype=None):
+    if obj is None: return self
+    return self.decode((obj._raw >> self.lo) & self.mask)
+  def __set__(self, obj, val): obj._raw = self.set(obj._raw, val)
+
+class FixedBitField(BitField):
+  def set(self, raw: int, val=None) -> int:
+    assert val is None, f"FixedBitField does not accept values, got {val}"
+    return super().set(raw, self.default)
+
+class EnumBitField(BitField):
+  def __init__(self, hi: int, lo: int, enum_cls, allowed: set | None = None):
+    super().__init__(hi, lo)
+    self._enum = enum_cls
+    self.allowed = allowed  # if set, only these enum values are valid for this encoding
+  def encode(self, val) -> int:
+    if not isinstance(val, self._enum): raise RuntimeError(f"expected {self._enum.__name__}, got {type(val).__name__}")
+    if self.allowed is not None and val not in self.allowed:
+      raise RuntimeError(f"opcode {val.name} not allowed in this encoding")
+    return val.value
+  def decode(self, raw): return self._enum(raw)
+
+# ══════════════════════════════════════════════════════════════
+# Typed fields
+# ══════════════════════════════════════════════════════════════
+
+import struct
+def _f32(f: float) -> int: return struct.unpack('I', struct.pack('f', f))[0]
+
+class SrcField(BitField):
+  _valid_range = (0, 511)  # inclusive
+  _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}
+
+  def __init__(self, hi: int, lo: int, default=s[0]):
+    super().__init__(hi, lo, default)
+    expected_size = self._valid_range[1] - self._valid_range[0] + 1
+    actual_size = 1 << (hi - lo + 1)
+    if actual_size != expected_size:
+      raise RuntimeError(f"{self.__class__.__name__}: field size {hi - lo + 1} bits ({actual_size}) doesn't match range {self._valid_range} ({expected_size})")
+
+  def encode(self, val) -> int:
+    """Encode value. Returns 255 (literal marker) for out-of-range values."""
+    if isinstance(val, Reg): offset = val.offset
+    elif isinstance(val, float): offset = self._FLOAT_ENC.get(val, 255)
+    elif isinstance(val, int) and 0 <= val <= 64: offset = 128 + val
+    elif isinstance(val, int) and -16 <= val < 0: offset = 192 - val
+    elif isinstance(val, int): offset = 255  # literal
+    else: raise TypeError(f"invalid src value {val}")
+    if not (self._valid_range[0] <= offset <= self._valid_range[1]):
+      raise TypeError(f"{self.__class__.__name__}: {val} (offset {offset}) out of range {self._valid_range}")
+    return offset - self._valid_range[0]
+
+  def decode(self, raw): return src[raw + self._valid_range[0]]
+
+  def __get__(self, obj, objtype=None):
+    if obj is None: return self
+    reg = self.decode((obj._raw >> self.lo) & self.mask)
+    # Resize register based on operand info (skip non-resizable special registers)
+    # VCC/EXEC pairs (106, 126), NULL (124), M0 (125), float constants (240-255)
+    if reg.offset not in (124, 125) and not 240 <= reg.offset <= 255:
+      # Map variant field names (vsrc0->src0, vsrc1->src1, etc.) for DPP/SDWA classes
+      assert self.name is not None
+      name = self.name[1:] if self.name.startswith('v') and self.name[1:] in obj.op_regs else self.name
+      if sz := obj.op_regs.get(name, 1): reg = Reg(reg.offset, sz, neg=reg.neg, abs_=reg.abs_, hi=reg.hi)
+    return reg
+
+class VGPRField(SrcField):
+  _valid_range = (256, 511)
+  def __init__(self, hi: int, lo: int, default=v[0]): super().__init__(hi, lo, default)
+  def encode(self, val) -> int:
+    if not isinstance(val, Reg): raise TypeError(f"VGPRField requires Reg, got {type(val).__name__}")
+    # For 8-bit vdst fields in VOP1/VOP2 16-bit ops, bit 7 is opsel for dest half
+    encoded = super().encode(val)
+    if val.hi and (self.hi - self.lo + 1) == 8:
+      if encoded >= 128:
+        raise ValueError(f"VGPRField: v[{encoded}].h not encodable in 8-bit field (v[0:127] only for .h)")
+      encoded |= 0x80
+    return encoded
+class SGPRField(SrcField): _valid_range = (0, 127)
+class SSrcField(SrcField): _valid_range = (0, 255)
+
+class AlignedSGPRField(BitField):
+  """SGPR field with alignment requirement. Encoded as sgpr_index // alignment."""
+  _align: int = 2
+  def encode(self, val):
+    if isinstance(val, int) and val == 0: return 0  # default: encode as s[0]
+    if not isinstance(val, Reg): raise TypeError(f"{self.__class__.__name__} requires Reg, got {type(val).__name__}")
+    if not (0 <= val.offset < 128): raise ValueError(f"{self.__class__.__name__} requires SGPR, got offset {val.offset}")
+    if val.offset & (self._align - 1): raise ValueError(f"{self.__class__.__name__} requires {self._align}-aligned SGPR, got s[{val.offset}]")
+    return val.offset >> (self._align.bit_length() - 1)
+  def decode(self, raw): return src[raw << (self._align.bit_length() - 1)]
+  def __get__(self, obj, objtype=None):
+    if obj is None: return self
+    reg = self.decode((obj._raw >> self.lo) & self.mask)
+    if sz := obj.op_regs.get(self.name, 1): reg = Reg(reg.offset, sz, neg=reg.neg, abs_=reg.abs_, hi=reg.hi)
+    return reg
+
+class SBaseField(AlignedSGPRField): _align = 2
+class SRsrcField(AlignedSGPRField): _align = 4
+
+class VDSTYField(BitField):
+  """VOPD vdsty: encoded = vgpr_idx >> 1. Actual vgpr = (encoded << 1) | ((vdstx & 1) ^ 1)."""
+  def encode(self, val):
+    if not isinstance(val, Reg): raise TypeError(f"VDSTYField requires Reg, got {type(val).__name__}")
+    if not (256 <= val.offset < 512): raise ValueError(f"VDSTYField requires VGPR, got offset {val.offset}")
+    return (val.offset - 256) >> 1
+  def __get__(self, obj, objtype=None):
+    if obj is None: return self
+    raw = (obj._raw >> self.lo) & self.mask
+    vdstx_bit0 = (obj.vdstx.offset - 256) & 1
+    vgpr_idx = (raw << 1) | (vdstx_bit0 ^ 1)
+    return Reg(256 + vgpr_idx, 1)
+
+# ══════════════════════════════════════════════════════════════
+# Operand info from XML
+# ══════════════════════════════════════════════════════════════
+
+import functools
+from extra.assembly.amd.autogen.rdna3.operands import OPERANDS as OPERANDS_RDNA3
+from extra.assembly.amd.autogen.rdna4.operands import OPERANDS as OPERANDS_RDNA4
+from extra.assembly.amd.autogen.cdna.operands import OPERANDS as OPERANDS_CDNA
+OPERANDS = {**OPERANDS_CDNA, **OPERANDS_RDNA3, **OPERANDS_RDNA4}
+
+# ══════════════════════════════════════════════════════════════
+# Inst base class
+# ══════════════════════════════════════════════════════════════
+
+def _needs_literal(val) -> bool:
+  """Check if a value needs a literal constant (can't be encoded inline)."""
+  if val is None or isinstance(val, Reg): return False
+  if isinstance(val, float): return val not in SrcField._FLOAT_ENC
+  if isinstance(val, int): return not (0 <= val <= 64 or -16 <= val < 0)
+  return False
+
+def _get_variant(cls, suffix: str):
+  """Get a variant class by suffix (e.g., '_LIT') via module lookup."""
+  import sys
+  module = sys.modules.get(cls.__module__)
+  return getattr(module, f"{cls.__name__}{suffix}", None) if module else None
+
+def _canonical_name(name: str) -> str | None:
+  """Map operand name to canonical name."""
+  if name in ('src0', 'vsrc0', 'ssrc0'): return 's0'
+  if name in ('src1', 'vsrc1', 'ssrc1'): return 's1'
+  if name == 'src2': return 's2'
+  if name in ('vdst', 'sdst', 'sdata'): return 'd'
+  if name in ('data', 'vdata', 'data0', 'vsrc'): return 'data'
+  return None
+
+class Inst:
+  _fields: list[tuple[str, BitField]]
+  _base_size: int
+
+  def __init_subclass__(cls):
+    # Collect fields from all parent classes, then override with this class's fields
+    inherited = {}
+    for base in reversed(cls.__mro__[1:]):
+      if hasattr(base, '_fields'):
+        inherited.update({name: field for name, field in base._fields})
+    inherited.update({name: val for name, val in cls.__dict__.items() if isinstance(val, BitField)})
+    cls._fields = list(inherited.items())
+    cls._base_size = (max(f.hi for _, f in cls._fields) + 8) // 8
+
+  def __new__(cls, *args, **kwargs):
+    # Auto-upgrade to variant if needed (only for base classes, not variants)
+    if not any(cls.__name__.endswith(sfx) for sfx in ('_LIT', '_DPP16', '_DPP8', '_SDWA', '_SDWA_SDST', '_MFMA')):
+      args_iter = iter(args)
+      for name, field in cls._fields:
+        if isinstance(field, FixedBitField): continue
+        val = kwargs.get(name) if name in kwargs else next(args_iter, None)
+        if not isinstance(field, SrcField): continue
+        if isinstance(val, Reg) and val.offset == 255 and (lit_cls := _get_variant(cls, '_LIT')): return lit_cls(*args, **kwargs)
+        if isinstance(val, Reg) and val.offset == 249:
+          if (sdwa_cls := _get_variant(cls, '_SDWA') or _get_variant(cls, '_SDWA_SDST')): return sdwa_cls(*args, **kwargs)
+        if isinstance(val, Reg) and val.offset == 250 and (dpp_cls := _get_variant(cls, '_DPP16')): return dpp_cls(*args, **kwargs)
+        if _needs_literal(val) and (lit_cls := _get_variant(cls, '_LIT')): return lit_cls(*args, **kwargs)
+    return object.__new__(cls)
+
+  def __init__(self, *args, **kwargs):
+    self._raw = 0
+    # Map positional args to field names (skip FixedBitFields)
+    args_iter = iter(args)
+    vals: dict[str, Any] = {}
+    for name, field in self._fields:
+      if isinstance(field, FixedBitField): vals[name] = None
+      elif name in kwargs: vals[name] = kwargs[name]
+      else: vals[name] = next(args_iter, None)
+    assert not (remaining := list(args_iter)), f"too many positional args: {remaining}"
+    # Extract modifiers from Reg objects and merge into neg/abs/opsel
+    neg_bits, abs_bits, opsel_bits = 0, 0, 0
+    for name, bit in [('src0', 0), ('src1', 1), ('src2', 2)]:
+      if name in vals and isinstance(vals[name], Reg):
+        reg = vals[name]
+        if reg.neg: neg_bits |= (1 << bit)
+        if reg.abs_: abs_bits |= (1 << bit)
+        if reg.hi: opsel_bits |= (1 << bit)
+    if 'vdst' in vals and isinstance(vals['vdst'], Reg) and vals['vdst'].hi:
+      opsel_bits |= (1 << 3)
+    if neg_bits: vals['neg'] = (vals.get('neg') or 0) | neg_bits
+    if abs_bits: vals['abs'] = (vals.get('abs') or 0) | abs_bits
+    if opsel_bits: vals['opsel'] = (vals.get('opsel') or 0) | opsel_bits
+    # For _LIT classes, capture literal value from SrcFields that encode to 255
+    literal_val = None
+    for name, field in self._fields:
+      val = vals[name]
+      if isinstance(field, SrcField) and val is not None and _needs_literal(val):
+        literal_val = _f32(val) if isinstance(val, float) else val & 0xFFFFFFFF
+    if literal_val is not None and 'literal' in vals:
+      vals['literal'] = literal_val
+    # Set all field values
+    for name, field in self._fields:
+      self._raw = field.set(self._raw, vals[name])
+    # Validate register sizes against operand info (skip special registers like NULL, VCC, EXEC, SDWA/DPP markers)
+    for name, expected in self.op_regs.items():
+      if (val := vals.get(name)) is None: continue
+      if isinstance(val, Reg) and val.sz != expected and not (106 <= val.offset <= 127 or 249 <= val.offset <= 255):
+        raise TypeError(f"{name} expects {expected} register(s), got {val.sz}")
+
+  @property
+  def op_name(self) -> str: return getattr(self, 'op').name
+  @property
+  def operands(self) -> dict: return OPERANDS.get(getattr(self, 'op'), {}) if hasattr(self, 'op') else {}
+  def _is_cdna(self) -> bool: return 'cdna' in type(self).__module__
+
+  @functools.cached_property
+  def op_bits(self) -> dict[str, int]:
+    """Get bit widths for each operand field, with WAVE32 and addr/saddr adjustments."""
+    if not hasattr(self, 'op'): return {k: v[1] for k, v in self.operands.items()}
+    bits = {k: v[1] for k, v in self.operands.items()}
+    # RDNA (WAVE32): condition masks, carry flags, and compare results are 32-bit
+    if not self._is_cdna():
+      name = self.op_name.lower()
+      if 'cndmask' in name and 'src2' in bits: bits['src2'] = 32
+      if '_co_ci_' in name and 'src2' in bits: bits['src2'] = 32  # carry-in source
+      # VOP3SD: sdst is always wavefront-size dependent (carry-out or condition mask)
+      if 'VOP3SD' in type(self).__name__ and 'sdst' in bits: bits['sdst'] = 32
+      if 'cmp' in name and 'vdst' in bits: bits['vdst'] = 32
+    # GLOBAL/FLAT: addr is 32-bit if saddr is valid SGPR, 64-bit if saddr is NULL
+    # SCRATCH: addr is always 32-bit (offset from scratch base, not absolute address)
+    if 'addr' in bits and (saddr_field := getattr(type(self), 'saddr', None)) and type(self).__name__ not in ('SCRATCH', 'VSCRATCH'):
+      saddr_val = (self._raw >> saddr_field.lo) & saddr_field.mask  # access _raw directly to avoid recursion
+      bits['addr'] = 64 if saddr_val in (124, 125) else 32  # 124=NULL, 125=M0
+    # MUBUF/MTBUF: vaddr size depends on offen/idxen (1 or 2 regs)
+    if 'vaddr' in bits and hasattr(self, 'offen') and hasattr(self, 'idxen'):
+      bits['vaddr'] = max(1, self.offen + self.idxen) * 32
+    # F8F6F4 MFMA: CBSZ selects matrix A format, BLGP selects matrix B format
+    # VGPRs: FP8/BF8(0,1)=8, FP6/BF6(2,3)=6, FP4(4)=4
+    if 'f8f6f4' in getattr(self, 'op_name', '').lower():
+      # Use explicit fields if available (VOP3PX2), else extract from VOP3P-MAI bit positions
+      cbsz = getattr(self, 'cbsz') if hasattr(type(self), 'cbsz') else (self._raw >> 8) & 0x7
+      blgp = getattr(self, 'blgp') if hasattr(type(self), 'blgp') else (self._raw >> 61) & 0x7
+      vgprs = {0: 8, 1: 8, 2: 6, 3: 6, 4: 4}
+      bits['src0'], bits['src1'] = vgprs.get(cbsz, 8) * 32, vgprs.get(blgp, 8) * 32
+    return bits
+  @property
+  def op_regs(self) -> dict[str, int]:
+    """Get register counts for each operand field."""
+    return {k: max(1, v // 32) for k, v in self.op_bits.items()}
+
+  @functools.cached_property
+  def canonical_op_bits(self) -> dict[str, int]:
+    """Get bit widths with canonical names: {'s0', 's1', 's2', 'd', 'data'}."""
+    bits = {'d': 32, 's0': 32, 's1': 32, 's2': 32, 'data': 32}
+    for name, val in self.op_bits.items():
+      if (cn := _canonical_name(name)): bits[cn] = val
+    return bits
+
+  @functools.cached_property
+  def canonical_operands(self) -> dict:
+    """Get operands with canonical names: {'s0', 's1', 's2', 'd', 'data'}."""
+    result = {}
+    for name, val in self.operands.items():
+      if (cn := _canonical_name(name)): result[cn] = val
+    return result
+
+  @property
+  def canonical_op_regs(self) -> dict[str, int]:
+    """Get register counts with canonical names: {'s0', 's1', 's2', 'd', 'data'}."""
+    return {k: max(1, v // 32) for k, v in self.canonical_op_bits.items()}
+
+  def num_srcs(self) -> int:
+    """Get number of source operands from operand info."""
+    ops = self.operands
+    if 'src2' in ops: return 3
+    if 'src1' in ops or 'vsrc1' in ops or 'ssrc1' in ops: return 2
+    if 'src0' in ops or 'vsrc0' in ops or 'ssrc0' in ops: return 1
+    return 0
+  @classmethod
+  def _size(cls) -> int: return cls._base_size
+  def size(self) -> int: return self._base_size
+  def disasm(self) -> str:
+    from extra.assembly.amd.disasm import disasm
+    return disasm(self)
+
+  def to_bytes(self) -> bytes: return self._raw.to_bytes(self._base_size, 'little')
+
+  @property
+  def _literal(self) -> int | None:
+    """Get the literal value if this instruction has one."""
+    return getattr(self, 'literal', None)
+
+  def _variant_suffix(self) -> str | None:
+    """Check if instruction needs a variant class (_LIT, _DPP8, _DPP16, _SDWA). Returns suffix or None."""
+    cls_name = type(self).__name__
+    # Don't check for variants if we're already a variant class
+    if any(s in cls_name for s in ('_LIT', '_DPP8', '_DPP16', '_SDWA')): return None
+    # VOPD: FMAMK/FMAAK opcodes always require literal (check by name since enum may differ across archs)
+    for name in ('opx', 'opy'):
+      if hasattr(self, name) and any(x in getattr(self, name).name for x in ('FMAMK', 'FMAAK')): return '_LIT'
+    for name, field in self._fields:
+      if isinstance(field, SrcField):
+        off = getattr(self, name).offset
+        if off == 255: return '_LIT'
+        if off == 249: return '_SDWA' if self._is_cdna() else '_DPP8'
+        if off == 250: return '_DPP16'
+    return None
+
+  @classmethod
+  def from_bytes(cls, data: bytes):
+    inst = object.__new__(cls)
+    inst._raw = int.from_bytes(data[:cls._base_size], 'little')
+    # Upgrade to variant class if needed (_LIT, _DPP8, _DPP16, _SDWA)
+    if (suffix := inst._variant_suffix()) and (var_cls := _get_variant(cls, suffix)) is not None:
+      return var_cls.from_bytes(data)
+    return inst
+
+  def __eq__(self, other): return type(self) is type(other) and self._raw == other._raw
+  def __hash__(self): return hash((type(self), self._raw))
+
+  def __repr__(self):
+    # collect (repr, is_default) pairs, strip trailing defaults so repr roundtrips with eval
+    name = self.op.name.lower() if hasattr(self, 'op') else type(self).__name__
+    parts = [(repr(v := getattr(self, n)), v == f.default) for n, f in self._fields if n != 'op' and not isinstance(f, FixedBitField)]
+    while parts and parts[-1][1]: parts.pop()
+    return f"{name}({', '.join(p[0] for p in parts)})"

extra/assembly/amd/generate.py

@@ -0,0 +1,480 @@
+# AMD ISA code generator - generates enum.py, ins.py, operands.py, str_pcode.py
+# Sources: XML from https://gpuopen.com/download/machine-readable-isa/latest/
+#          PDF manuals from AMD documentation
+import re, zlib, xml.etree.ElementTree as ET, zipfile
+from tinygrad.helpers import fetch
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# Configuration
+# ═══════════════════════════════════════════════════════════════════════════════
+
+ARCHS = {
+  "rdna3": {"xml": "amdgpu_isa_rdna3_5.xml", "pdf": "https://docs.amd.com/api/khub/documents/UVVZM22UN7tMUeiW_4ShTQ/content"},
+  "rdna4": {"xml": "amdgpu_isa_rdna4.xml", "pdf": "https://docs.amd.com/api/khub/documents/uQpkEvk3pv~kfAb2x~j4uw/content"},
+  "cdna": {"xml": "amdgpu_isa_cdna4.xml", "pdf": "https://www.amd.com/content/dam/amd/en/documents/instinct-tech-docs/instruction-set-architectures/amd-instinct-cdna4-instruction-set-architecture.pdf"},
+}
+XML_URL = "https://gpuopen.com/download/machine-readable-isa/latest/"
+# Map XML encoding names to codebase names
+NAME_MAP = {"VOP3_SDST_ENC": "VOP3SD", "VOP3_SDST_ENC_LIT": "VOP3SD_LIT", "VOP3_SDST_ENC_DPP16": "VOP3SD_DPP16",
+            "VOP3_SDST_ENC_DPP8": "VOP3SD_DPP8", "VOPDXY": "VOPD", "VOPDXY_LIT": "VOPD_LIT", "VDS": "DS"}
+# Instructions missing from XML but present in PDF
+FIXES = {"rdna3": {"SOPK": {22: "S_SUBVECTOR_LOOP_BEGIN", 23: "S_SUBVECTOR_LOOP_END"}, "FLAT": {55: "FLAT_ATOMIC_CSUB_U32"}},
+         "rdna4": {"SOP1": {80: "S_GET_BARRIER_STATE", 81: "S_BARRIER_INIT", 82: "S_BARRIER_JOIN"}, "SOPP": {9: "S_WAITCNT", 21: "S_BARRIER_LEAVE"}},
+         "cdna": {"DS": {152: "DS_GWS_SEMA_RELEASE_ALL", 154: "DS_GWS_SEMA_V", 156: "DS_GWS_SEMA_P"},
+                  "VOP3P": {44: "V_MFMA_LD_SCALE_B32", 62: "V_MFMA_F32_16X16X8_XF32", 63: "V_MFMA_F32_32X32X4_XF32"}}}
+# Fields missing from XML but present in hardware (format: {arch: {encoding: [(name, hi, lo), ...]}})
+FIELD_FIXES = {"cdna": {"VOP3P": [("opsel_hi2", 14, 14)]}}
+# Encoding suffixes to strip (variants we don't generate separate classes for)
+_ENC_SUFFIXES = ("_NSA1",)
+# Encoding suffix to class suffix mapping (for variants we DO generate)
+_ENC_SUFFIX_MAP = {"_INST_LITERAL": "_LIT", "_VOP_DPP16": "_DPP16", "_VOP_DPP": "_DPP16", "_VOP_DPP8": "_DPP8",
+                   "_VOP_SDWA": "_SDWA", "_VOP_SDWA_SDST_ENC": "_SDWA_SDST", "_MFMA": "_MFMA"}
+# Field name normalization
+_FIELD_RENAMES = {"opsel_hi_2": "opsel_hi2", "op_sel_hi_2": "opsel_hi2", "op_sel": "opsel", "bound_ctrl": "bc",
+                  "tgt": "target", "row_en": "row", "unorm": "unrm", "clamp": "clmp", "wait_exp": "waitexp",
+                  "simm32": "literal", "dpp_ctrl": "dpp", "acc_cd": "acc_cd", "acc": "acc",
+                  "dst_sel": "dst_sel", "dst_unused": "dst_unused", "src0_sel": "src0_sel", "src1_sel": "src1_sel"}
+# Encoding variants to skip entirely (NSA is for MIMG graphics instructions)
+_SKIP_ENCODINGS = ("NSA",)
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# XML parsing helpers
+# ═══════════════════════════════════════════════════════════════════════════════
+
+def _strip_enc(name: str) -> str:
+  """Strip ENC_ prefix and normalize encoding suffixes."""
+  name = name.removeprefix("ENC_")
+  for sfx in _ENC_SUFFIXES: name = name.replace(sfx, "")
+  # Process longer suffixes first to avoid partial matches (e.g., _VOP_DPP8 before _VOP_DPP)
+  for old, new in sorted(_ENC_SUFFIX_MAP.items(), key=lambda x: -len(x[0])): name = name.replace(old, new)
+  return name
+
+def _norm_field(name: str) -> str:
+  """Normalize field name to match expected names."""
+  for old, new in _FIELD_RENAMES.items(): name = name.replace(old, new)
+  return name
+
+def _map_flat(enc_name: str, instr_name: str) -> str:
+  """Map FLAT/GLOBAL/SCRATCH encoding to correct enum based on instruction prefix."""
+  if enc_name in ("FLAT_GLBL", "FLAT_GLOBAL"): return "GLOBAL"
+  if enc_name == "FLAT_SCRATCH": return "SCRATCH"
+  if enc_name in ("FLAT", "VFLAT", "VGLOBAL", "VSCRATCH"):
+    v = "V" if enc_name.startswith("V") else ""
+    if instr_name.startswith("GLOBAL_"): return f"{v}GLOBAL"
+    if instr_name.startswith("SCRATCH_"): return f"{v}SCRATCH"
+    return f"{v}FLAT"
+  return enc_name
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# XML parsing
+# ═══════════════════════════════════════════════════════════════════════════════
+
+def parse_xml(filename: str):
+  root = ET.fromstring(zipfile.ZipFile(fetch(XML_URL)).read(filename))
+  encodings, enums, types, fmts, op_types_set = {}, {}, {}, {}, set()
+  # Extract HWREG and MSG enums from OperandTypes
+  op_enum_map = {("OPR_HWREG", "ID"): "HWREG", ("OPR_SENDMSG_RTN", "MSG"): "MSG"}
+  for ot in root.findall(".//OperandTypes/OperandType"):
+    ot_name = ot.findtext("OperandTypeName")
+    for field in ot.findall(".//Field"):
+      if (enum_name := op_enum_map.get((ot_name, field.findtext("FieldName")))):
+        enums[enum_name] = {int(pv.findtext("Value")): pv.findtext("Name").upper() for pv in field.findall(".//PredefinedValue")}
+  # Extract DataFormats with BitCount
+  for df in root.findall("ISA/DataFormats/DataFormat"):
+    name, bits = df.findtext("DataFormatName"), df.findtext("BitCount")
+    if name and bits: fmts[name] = int(bits)
+  # Extract encoding definitions
+  for enc in root.findall("ISA/Encodings/Encoding"):
+    name = enc.findtext("EncodingName")
+    is_base = name.startswith("ENC_") or name in ("VOP3_SDST_ENC", "VOPDXY")
+    is_variant = any(sfx in name for sfx in _ENC_SUFFIX_MAP)
+    if not is_base and not is_variant: continue
+    if any(s in name for s in _SKIP_ENCODINGS): continue
+    fields = [(_norm_field(f.findtext("FieldName").lower()), int(f.find("BitLayout/Range").findtext("BitOffset") or 0) + int(f.find("BitLayout/Range").findtext("BitCount") or 0) - 1,
+               int(f.find("BitLayout/Range").findtext("BitOffset") or 0))
+              for f in enc.findall(".//MicrocodeFormat/BitMap/Field") if f.find("BitLayout/Range") is not None]
+    ident = (enc.findall("EncodingIdentifiers/EncodingIdentifier") or [None])[0]
+    enc_field = next((f for f in fields if f[0] == "encoding"), None)
+    # For multi-dword formats, encoding field may be in higher dword but identifier pattern is always in dword0; use % 32
+    enc_bits = "".join(ident.text[len(ident.text)-1-b] for b in range(enc_field[1] % 32, (enc_field[2] % 32)-1, -1)) if ident is not None and enc_field else None
+    base_name = _strip_enc(name)
+    encodings[NAME_MAP.get(base_name, base_name)] = (fields, enc_bits)
+  # Extract instruction opcodes and operand info
+  # Track which encodings each opcode appears in (for detecting LIT-only ops)
+  opcode_encs: dict[str, dict[int, set[str]]] = {}  # {base_fmt: {opcode: {enc_names}}}
+  for instr in root.findall("ISA/Instructions/Instruction"):
+    name = instr.findtext("InstructionName")
+    for enc in instr.findall("InstructionEncodings/InstructionEncoding"):
+      if enc.findtext("EncodingCondition") != "default": continue
+      base, opcode = _map_flat(_strip_enc(enc.findtext("EncodingName")), name), int(enc.findtext("Opcode") or 0)
+      enc_name = NAME_MAP.get(base, base)
+      # Encoding variants use the same Op enum as the base format
+      base_enum = enc_name
+      for sfx in ("_SDWA_SDST", "_DPP16", "_DPP8", "_SDWA", "_LIT", "_MFMA"):
+        base_enum = base_enum.replace(sfx, "")
+      # Track which encodings this opcode appears in
+      opcode_encs.setdefault(base_enum, {}).setdefault(opcode, set()).add(enc_name)
+      # ADDTID instructions go in both FLAT and GLOBAL enums (pcode uses FLATOp for these)
+      if "ADDTID" in name:
+        if base == "GLOBAL": enums.setdefault("FLAT", {})[opcode] = name
+        elif base == "VGLOBAL": enums.setdefault("VFLAT", {})[opcode] = name
+      enums.setdefault(base_enum, {})[opcode] = name
+      # Extract operand info
+      op_info = {op.findtext("FieldName").lower(): (op.findtext("DataFormatName"), int(op.findtext("OperandSize") or 0), op.findtext("OperandType"))
+                 for op in enc.findall("Operands/Operand") if op.findtext("FieldName")}
+      for fmt, _, otype in op_info.values():
+        if fmt and fmt not in fmts: fmts[fmt] = 0
+        if otype: op_types_set.add(otype)
+      if op_info: types[(name, base_enum)] = op_info
+  # Find opcodes that only exist in a specific variant encoding (no base format version)
+  suffix_only_ops: dict[str, dict[str, set[int]]] = {}  # {suffix: {base_fmt: {opcodes}}}
+  for base_fmt, opcodes in opcode_encs.items():
+    for opcode, encs in opcodes.items():
+      suffix = next((s for s in _ENC_SUFFIX_MAP.values() if all(s in e for e in encs)), None)
+      if suffix is not None: suffix_only_ops.setdefault(suffix, {}).setdefault(base_fmt, set()).add(opcode)
+  return encodings, enums, types, fmts, op_types_set, suffix_only_ops
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# PDF parsing
+# ═══════════════════════════════════════════════════════════════════════════════
+
+def extract_pdf_text(url: str) -> list[list[tuple[float, float, str, str]]]:
+  """Extract positioned text from PDF. Returns list of text elements (x, y, text, font) per page."""
+  data = fetch(url).read_bytes()
+  # Parse xref table to locate objects
+  xref: dict[int, int] = {}
+  pos = int(re.search(rb'startxref\s+(\d+)', data).group(1)) + 4
+  while data[pos:pos+7] != b'trailer':
+    while data[pos:pos+1] in b' \r\n': pos += 1
+    line_end = data.find(b'\n', pos)
+    start_obj, count = map(int, data[pos:line_end].split()[:2])
+    pos = line_end + 1
+    for i in range(count):
+      if data[pos+17:pos+18] == b'n' and (off := int(data[pos:pos+10])) > 0: xref[start_obj + i] = off
+      pos += 20
+
+  def get_stream(n: int) -> bytes:
+    obj = data[xref[n]:data.find(b'endobj', xref[n])]
+    raw = obj[obj.find(b'stream\n') + 7:obj.find(b'\nendstream')]
+    return zlib.decompress(raw) if b'/FlateDecode' in obj else raw
+
+  pages = []
+  for n in sorted(xref):
+    if b'/Type /Page' not in data[xref[n]:xref[n]+500]: continue
+    if not (m := re.search(rb'/Contents (\d+) 0 R', data[xref[n]:xref[n]+500])): continue
+    stream = get_stream(int(m.group(1))).decode('latin-1')
+    elements, font = [], ''
+    for bt in re.finditer(r'BT(.*?)ET', stream, re.S):
+      x, y = 0.0, 0.0
+      for m in re.finditer(r'(/F[\d.]+) [\d.]+ Tf|([\d.+-]+) ([\d.+-]+) Td|[\d.+-]+ [\d.+-]+ [\d.+-]+ [\d.+-]+ ([\d.+-]+) ([\d.+-]+) Tm|<([0-9A-Fa-f]+)>.*?Tj|\[([^\]]+)\] TJ', bt.group(1)):
+        if m.group(1): font = m.group(1)
+        elif m.group(2): x, y = x + float(m.group(2)), y + float(m.group(3))
+        elif m.group(4): x, y = float(m.group(4)), float(m.group(5))
+        elif m.group(6) and (t := bytes.fromhex(m.group(6)).decode('latin-1')).strip(): elements.append((x, y, t, font))
+        elif m.group(7) and (t := ''.join(bytes.fromhex(h).decode('latin-1') for h in re.findall(r'<([0-9A-Fa-f]+)>', m.group(7)))).strip(): elements.append((x, y, t, font))
+    pages.append(sorted(elements, key=lambda e: (-e[1], e[0])))
+  return pages
+
+def extract_pcode(pages: list[list[tuple[float, float, str, str]]], name_to_op: dict[str, int]) -> dict[tuple[str, int], str]:
+  """Extract pseudocode for instructions. Returns {(name, opcode): pseudocode}."""
+  # First pass: find all instruction headers across all pages
+  all_instructions: list[tuple[int, float, str, int]] = []  # (page_idx, y, name, opcode)
+  for page_idx, page in enumerate(pages):
+    by_y: dict[int, list[tuple[float, str]]] = {}
+    for x, y, t, _ in page:
+      by_y.setdefault(round(y), []).append((x, t))
+    for y, items in sorted(by_y.items(), reverse=True):
+      left = [(x, t) for x, t in items if 55 < x < 65]
+      right = [(x, t) for x, t in items if 535 < x < 550]
+      if left and right and left[0][1] in name_to_op and right[0][1].isdigit():
+        all_instructions.append((page_idx, y, left[0][1], int(right[0][1])))
+
+  # Second pass: extract pseudocode between consecutive instructions
+  pcode: dict[tuple[str, int], str] = {}
+  for i, (page_idx, y, name, opcode) in enumerate(all_instructions):
+    if i + 1 < len(all_instructions):
+      next_page, next_y = all_instructions[i + 1][0], all_instructions[i + 1][1]
+    else:
+      next_page, next_y = page_idx, 0
+    # Collect F6 text from current position to next instruction (pseudocode is at x ≈ 69)
+    lines = []
+    for p in range(page_idx, next_page + 1):
+      start_y = y if p == page_idx else 800
+      end_y = next_y if p == next_page else 0
+      lines.extend((p, y2, t) for x, y2, t, f in pages[p] if f in ('/F6.0', '/F7.0') and end_y < y2 < start_y and 60 < x < 80)
+    if lines:
+      sorted_lines = sorted(lines, key=lambda x: (x[0], -x[1]))
+      # Stop at large Y gaps (>30) - indicates section break
+      filtered = [sorted_lines[0]]
+      for j in range(1, len(sorted_lines)):
+        prev_page, prev_y, _ = sorted_lines[j-1]
+        curr_page, curr_y, _ = sorted_lines[j]
+        if curr_page == prev_page and prev_y - curr_y > 30: break
+        if curr_page != prev_page and prev_y > 60 and curr_y < 730: break
+        filtered.append(sorted_lines[j])
+      pcode_lines = [t.replace('Ê', '').strip() for _, _, t in filtered]
+      if pcode_lines: pcode[(name, opcode)] = '\n'.join(pcode_lines)
+  return pcode
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# Code generation
+# ═══════════════════════════════════════════════════════════════════════════════
+
+def write_common(all_fmts, all_op_types, path):
+  lines = ["# autogenerated from AMD ISA XML - do not edit", "from enum import Enum, auto", ""]
+  lines.append("class ReprEnum(Enum):")
+  lines.append('  """Enum with clean repr that roundtrips with eval()."""')
+  lines.append('  def __repr__(self): return f"{type(self).__name__}.{self.name}"')
+  lines.append("")
+  lines.append("class Fmt(Enum):")
+  for fmt in sorted(all_fmts.keys()): lines.append(f"  {fmt} = auto()")
+  lines.append("")
+  lines.append("FMT_BITS = {")
+  for fmt, bits in sorted(all_fmts.items()): lines.append(f"  Fmt.{fmt}: {bits},")
+  lines.append("}")
+  lines.append("")
+  lines.append("class OpType(Enum):")
+  for ot in sorted(all_op_types): lines.append(f"  {ot} = auto()")
+  with open(path, "w") as f: f.write("\n".join(lines))
+
+def write_enum(enums, path):
+  lines = ["# autogenerated from AMD ISA XML - do not edit", "from extra.assembly.amd.autogen.common import ReprEnum, Fmt, FMT_BITS, OpType  # noqa: F401", ""]
+  for name, ops in sorted(enums.items()):
+    if not ops: continue
+    suffix = "_E32" if name in ("VOP1", "VOP2", "VOPC") else "_E64" if name == "VOP3" else ""
+    lines.append(f"class {name}(ReprEnum):" if name in ("HWREG", "MSG") else f"class {name}Op(ReprEnum):")
+    aliases = []
+    for op, mem in sorted(ops.items()):
+      msuf = suffix if name != "VOP3" or op < 512 else ""
+      lines.append(f"  {mem}{msuf} = {op}")
+      if msuf: aliases.append((mem, msuf))
+    for mem, msuf in aliases: lines.append(f"  {mem} = {mem}{msuf}")
+    lines.append("")
+  with open(path, "w") as f: f.write("\n".join(lines))
+
+def write_ins(encodings, enums, suffix_only_ops, types, arch, path):
+  _VGPR_FIELDS = {"vdst", "vdstx", "vsrc0", "vsrc1", "vsrc2", "vsrc3", "vsrcx1", "vsrcy1", "vaddr", "vdata", "data", "data0", "data1", "addr", "vsrc"}
+  _VARIANT_SUFFIXES = ("_LIT", "_DPP16", "_DPP8", "_SDWA_SDST", "_SDWA", "_MFMA")
+  def get_base_fmt(fmt):
+    for sfx in _VARIANT_SUFFIXES: fmt = fmt.replace(sfx, "")
+    return fmt
+  def field_def(name, hi, lo, fmt, enc_bits=None):
+    bits = hi - lo + 1
+    base_fmt = get_base_fmt(fmt)
+    if name == "encoding" and enc_bits: return f"FixedBitField({hi}, {lo}, 0b{enc_bits})"
+    if name == "op" and fmt not in ("DPP", "SDWA"): return f"EnumBitField({hi}, {lo}, {base_fmt}Op)"
+    if name in ("opx", "opy"): return f"EnumBitField({hi}, {lo}, VOPDOp)"
+    if name == "vdsty": return f"VDSTYField({hi}, {lo})"
+    if name in _VGPR_FIELDS and bits == 8: return f"VGPRField({hi}, {lo})"
+    if name == "sbase" and bits == 6: return f"SBaseField({hi}, {lo})"
+    if name in ("srsrc", "ssamp") and bits == 5: return f"SRsrcField({hi}, {lo})"
+    if name in ("sdst", "sdata") and bits == 7: return f"SGPRField({hi}, {lo})"
+    if name in ("soffset", "saddr") and bits == 7: return f"SGPRField({hi}, {lo}, default=NULL)"
+    if name.startswith("ssrc") and bits == 8: return f"SSrcField({hi}, {lo})"
+    if name in ("saddr", "soffset") and bits == 8: return f"SSrcField({hi}, {lo}, default=NULL)"
+    if name.startswith("src") and bits == 9: return f"SrcField({hi}, {lo})"
+    # GLOBAL/SCRATCH: offset is 13-bit signed [12:0], FLAT: 12-bit unsigned (XML has 12-bit for all)
+    if name == "offset" and base_fmt in ("GLOBAL", "SCRATCH"): return f"BitField(12, {lo})"
+    if base_fmt == "VOP3P" and name == "opsel_hi": return f"BitField({hi}, {lo}, default=3)"
+    if base_fmt == "VOP3P" and name == "opsel_hi2": return f"BitField({hi}, {lo}, default=1)"
+    return f"BitField({hi}, {lo})"
+  ORDER = ['encoding', 'op', 'opx', 'opy', 'vdst', 'vdstx', 'vdsty', 'sdst', 'vdata', 'sdata', 'addr', 'vaddr', 'data', 'data0', 'data1',
+           'src0', 'srcx0', 'srcy0', 'vsrc0', 'ssrc0', 'src1', 'vsrc1', 'vsrcx1', 'vsrcy1', 'ssrc1', 'src2', 'vsrc2', 'src3', 'vsrc3',
+           'saddr', 'sbase', 'srsrc', 'ssamp', 'soffset', 'offset', 'simm16', 'literal', 'en', 'target', 'attr', 'attr_chan',
+           'omod', 'neg', 'neg_hi', 'abs', 'clmp', 'opsel', 'opsel_hi', 'waitexp', 'wait_va',
+           'dmask', 'dim', 'seg', 'format', 'offen', 'idxen', 'glc', 'dlc', 'slc', 'tfe', 'unrm', 'done', 'row',
+           'dpp', 'fi', 'bc', 'row_mask', 'bank_mask', 'src0_neg', 'src0_abs', 'src1_neg', 'src1_abs',
+           'cbsz', 'abid', 'acc_cd', 'acc', 'blgp', 'lane_sel_0', 'lane_sel_1', 'lane_sel_2', 'lane_sel_3',
+           'lane_sel_4', 'lane_sel_5', 'lane_sel_6', 'lane_sel_7', 'dst_sel', 'dst_unused', 'src0_sel', 'src1_sel']
+  sort_fields = lambda fields: sorted(fields, key=lambda f: (ORDER.index(f[0]) if f[0] in ORDER else 999, f[2]))
+
+  # Separate base encodings from variants
+  base_encodings, variant_encodings = {}, {}
+  for enc_name, data in encodings.items():
+    base = get_base_fmt(enc_name)
+    if base == enc_name: base_encodings[enc_name] = data
+    else: variant_encodings[enc_name] = data
+
+  # Build sets of ops by their vdst type from operand metadata
+  sdst_opcodes = {}  # ops where vdst is OPR_SREG (writes to SGPR)
+  for fmt, ops in enums.items():
+    for op, name in ops.items():
+      op_types = types.get((name, fmt), {})
+      vdst_type = op_types.get("vdst", (None, None, None))[2]
+      if vdst_type == "OPR_SREG": sdst_opcodes.setdefault(fmt, set()).add(op)
+
+  lines = ["# autogenerated from AMD ISA XML - do not edit", "# ruff: noqa: F401,F403",
+           "from extra.assembly.amd.dsl import *", f"from extra.assembly.amd.autogen.{arch}.enum import *", "import functools", ""]
+
+  def fmt_allowed(op_enum: str, ops: set[int]) -> str:
+    """Format allowed ops as {EnumName.MEMBER, ...}."""
+    names = [f"{op_enum}.{enums[op_enum.removesuffix('Op')][op]}" for op in sorted(ops)]
+    return "{" + ", ".join(names) + "}"
+
+  # Generate base classes first
+  for enc_name, (fields, enc_bits) in sorted(base_encodings.items()):
+    all_ops = set(enums.get(enc_name, {}).keys())
+    # Get suffix-only ops for this format (these can't be used in base class)
+    base_suffix_ops = set().union(*(d.get(enc_name, set()) for d in suffix_only_ops.values()))
+    # Exclude SDST ops from base class (they need VOP1_SDST/VOP3_SDST/VOP3B)
+    base_allowed = all_ops - base_suffix_ops - sdst_opcodes.get(enc_name, set())
+    # RDNA3 FLAT/GLOBAL/SCRATCH share encoding bits, differentiated by seg field
+    # RDNA4 VFLAT/VGLOBAL/VSCRATCH have distinct encoding bits, no seg field needed
+    has_seg_field = any(fn == "seg" for fn, _, _ in fields)
+    if enc_name in ("FLAT", "VFLAT") and has_seg_field:
+      prefix = "V" if enc_name == "VFLAT" else ""
+      for cls, seg, op_enum in [(f"{prefix}FLAT", 0, f"{prefix}FLATOp"), (f"{prefix}GLOBAL", 2, f"{prefix}GLOBALOp"), (f"{prefix}SCRATCH", 1, f"{prefix}SCRATCHOp")]:
+        cls_ops = set(enums.get(cls, {}).keys())
+        lines.append(f"class {cls}(Inst):")
+        for fn, hi, lo in sort_fields(fields):
+          if fn == "seg": lines.append(f"  seg = FixedBitField({hi}, {lo}, {seg})")
+          elif fn == "op": lines.append(f"  op = EnumBitField({hi}, {lo}, {op_enum}, {fmt_allowed(op_enum, cls_ops)})")
+          else: lines.append(f"  {fn} = {field_def(fn, hi, lo, cls, enc_bits)}")
+        lines.append("")
+    elif enc_name not in ("FLAT_GLOBAL", "FLAT_SCRATCH", "FLAT_GLBL", "DPP", "SDWA"):
+      lines.append(f"class {enc_name}(Inst):")
+      for fn, hi, lo in sort_fields(fields):
+        if fn == "op":
+          base_fmt = get_base_fmt(enc_name)
+          lines.append(f"  op = EnumBitField({hi}, {lo}, {base_fmt}Op, {fmt_allowed(f'{base_fmt}Op', base_allowed)})")
+        else:
+          lines.append(f"  {fn} = {field_def(fn, hi, lo, enc_name, enc_bits if fn == 'encoding' else None)}")
+      lines.append("")
+
+  # Generate variant classes that inherit from base (only add extra fields)
+  for enc_name, (fields, enc_bits) in sorted(variant_encodings.items()):
+    base = get_base_fmt(enc_name)
+    if base not in base_encodings: continue  # skip if no base class
+    base_fields = {f[0] for f in base_encodings[base][0]}
+    extra_fields = [(fn, hi, lo) for fn, hi, lo in fields if fn not in base_fields]
+    # Check if this is a suffix-only variant
+    variant_suffix = next((sfx for sfx in _VARIANT_SUFFIXES if enc_name.endswith(sfx)), None)
+    is_suffix_variant = variant_suffix in suffix_only_ops
+    all_ops = set(enums.get(base, {}).keys())
+    if extra_fields or is_suffix_variant:
+      lines.append(f"class {enc_name}({base}):")
+      op_field = next((f for f in base_encodings[base][0] if f[0] == "op"), None)
+      # _LIT classes: override op to allow all opcodes (base excludes lit-only ops)
+      # other classes override op to only suffix-only opcodes
+      if op_field and is_suffix_variant:
+        _, hi, lo = op_field
+        allowed_ops = all_ops if variant_suffix == "_LIT" else suffix_only_ops[variant_suffix][base]
+        lines.append(f"  op = EnumBitField({hi}, {lo}, {base}Op, {fmt_allowed(f'{base}Op', allowed_ops)})")
+      for fn, hi, lo in sort_fields(extra_fields):
+        lines.append(f"  {fn} = {field_def(fn, hi, lo, enc_name)}")
+      lines.append("")
+
+  # SDST variants (special case - redefine vdst field type, restrict to SDST ops)
+  for base, field_hi, field_lo in [("VOP1", 24, 17), ("VOP3", 7, 0)]:
+    if base not in base_encodings: continue
+    sdst_ops = sdst_opcodes.get(base, set())
+    if not sdst_ops: continue
+    # For VOP3, all ops < 256 (compare/cmpx ops) use SDST encoding
+    all_base_ops = set(enums.get(base, {}).keys())
+    if base == "VOP3": sdst_ops = sdst_ops | {op for op in all_base_ops if op < 256}
+    op_field = next((f for f in base_encodings[base][0] if f[0] == "op"), None)
+    lines.append(f"class {base}_SDST({base}):")
+    if op_field:
+      _, hi, lo = op_field
+      lines.append(f"  op = EnumBitField({hi}, {lo}, {base}Op, {fmt_allowed(f'{base}Op', sdst_ops)})")
+    lines.append(f"  vdst = SSrcField({field_hi}, {field_lo})")
+    lines.append("")
+    # SDST_LIT class (for literals with SDST destination) - same ops, just adds literal field
+    lit_enc = variant_encodings.get(f"{base}_LIT")
+    if lit_enc:
+      lit_field = next((f for f in lit_enc[0] if f[0] == "literal"), None)
+      if lit_field:
+        lines.append(f"class {base}_SDST_LIT({base}_SDST):")
+        lines.append(f"  literal = BitField({lit_field[1]}, {lit_field[2]})")
+        lines.append("")
+
+  # Instruction helpers
+  lines.append("# instruction helpers")
+  for fmt, ops in sorted(enums.items()):
+    if fmt not in base_encodings and fmt not in ("GLOBAL", "SCRATCH", "VGLOBAL", "VSCRATCH"): continue
+    suffix = "_E32" if fmt in ("VOP1", "VOP2", "VOPC") else "_E64" if fmt == "VOP3" else ""
+    op_to_suffix = {op:suffix for suffix,ops in suffix_only_ops.items() for op in ops.get(fmt, set())}
+    fmt_sdst_ops = sdst_opcodes.get(fmt, set())
+    for op, name in sorted(ops.items()):
+      msuf = suffix if fmt != "VOP3" or op < 512 else ""
+      # Determine class: SDST variants, suffix-specific variants (e.g., _MFMA, _LIT), or base
+      if fmt == "VOP1" and op in fmt_sdst_ops: cls = "VOP1_SDST"
+      elif fmt == "VOP3" and (op in fmt_sdst_ops or op < 256): cls = "VOP3_SDST"
+      elif op_to_suffix.get(op): cls = f"{fmt}{op_to_suffix[op]}"
+      else: cls = fmt
+      lines.append(f"{name.lower()}{msuf.lower()} = functools.partial({cls}, {fmt}Op.{name}{msuf})")
+  with open(path, "w") as f: f.write("\n".join(lines))
+
+def write_operands(types, enums, arch, path):
+  valid = {(name, fmt) for fmt, ops in enums.items() for name in ops.values()}
+  lines = ["# autogenerated from AMD ISA XML - do not edit",
+           "from extra.assembly.amd.autogen.common import Fmt, OpType",
+           f"from extra.assembly.amd.autogen.{arch}.enum import *", ""]
+  lines.append("# instruction operand info: {Op: {field: (Fmt, size_bits, OpType)}}")
+  lines.append("OPERANDS = {")
+  def fmt_val(v):
+    fmt, size, otype = v
+    return f"({f'Fmt.{fmt}' if fmt else 'None'}, {size}, {f'OpType.{otype}' if otype else 'None'})"
+  for (name, enc_base), fields in sorted(types.items()):
+    if (name, enc_base) not in valid: continue
+    fstr = ", ".join(f'"{k}": {fmt_val(v)}' for k, v in sorted(fields.items()))
+    lines.append(f'  {enc_base}Op.{name}: {{{fstr}}},')
+  lines.append("}")
+  with open(path, "w") as f: f.write("\n".join(lines))
+
+def write_pcode(pcode: dict[tuple[str, int], str], enums: dict[str, dict[int, str]], arch: str, path: str):
+  """Write str_pcode.py file from extracted pseudocode."""
+  entries: list[tuple[str, str, int, str]] = []
+  for fmt_name, ops in enums.items():
+    member_suffix = "_E32" if fmt_name in ("VOP1", "VOP2", "VOPC") else "_E64" if fmt_name == "VOP3" else ""
+    for opcode, name in ops.items():
+      if (name, opcode) in pcode:
+        msuf = member_suffix if fmt_name != "VOP3" or opcode < 512 else ""
+        entries.append((f"{fmt_name}Op", f"{name}{msuf}", opcode, pcode[(name, opcode)]))
+  enum_names = sorted(set(e[0] for e in entries))
+  lines = ["# autogenerated from AMD ISA PDF - do not edit", "# ruff: noqa: E501",
+           f"from extra.assembly.amd.autogen.{arch}.enum import {', '.join(enum_names)}", "", "PCODE = {"]
+  for enum_name, name, opcode, code in sorted(entries, key=lambda x: (x[0], x[2])):
+    lines.append(f"  {enum_name}.{name}: {code!r},")
+  lines.append("}")
+  with open(path, "w") as f: f.write("\n".join(lines))
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# Main
+# ═══════════════════════════════════════════════════════════════════════════════
+
+if __name__ == "__main__":
+  import pathlib
+  all_fmts, all_op_types, arch_data = {}, set(), {}
+  # First pass: parse XML for all architectures
+  for arch, cfg in ARCHS.items():
+    print(f"Parsing XML: {cfg['xml']} -> {arch}")
+    encodings, enums, types, fmts, op_types_set, suffix_only_ops = parse_xml(cfg["xml"])
+    for fmt, ops in FIXES.get(arch, {}).items(): enums.setdefault(fmt, {}).update(ops)
+    for fmt, fields in FIELD_FIXES.get(arch, {}).items():
+      if fmt in encodings: encodings[fmt] = (encodings[fmt][0] + fields, encodings[fmt][1])
+    arch_data[arch] = {"encodings": encodings, "enums": enums, "types": types, "suffix_only_ops": suffix_only_ops}
+    for fmt, bits in fmts.items():
+      assert fmt not in all_fmts or all_fmts[fmt] == bits, f"FMT_BITS mismatch for {fmt}: {all_fmts[fmt]} vs {bits}"
+      all_fmts[fmt] = bits
+    all_op_types.update(op_types_set)
+  # Write common.py
+  common_path = pathlib.Path(__file__).parent / "autogen" / "common.py"
+  write_common(all_fmts, all_op_types, common_path)
+  print(f"Wrote common.py: {len(all_fmts)} formats, {len(all_op_types)} op types")
+  # Write per-arch files from XML
+  for arch, data in arch_data.items():
+    base = pathlib.Path(__file__).parent / "autogen" / arch
+    write_enum(data["enums"], base / "enum.py")
+    write_ins(data["encodings"], data["enums"], data["suffix_only_ops"], data["types"], arch, base / "ins.py")
+    write_operands(data["types"], data["enums"], arch, base / "operands.py")
+    print(f"  {arch}: {len(data['encodings'])} encodings, {sum(len(v) for v in data['enums'].values())} instructions")
+  # Second pass: parse PDFs and write pcode
+  for arch, cfg in ARCHS.items():
+    print(f"Parsing PDF: {arch}...")
+    pages = extract_pdf_text(cfg["pdf"])
+    name_to_op = {name: op for ops in arch_data[arch]["enums"].values() for op, name in ops.items()}
+    pcode = extract_pcode(pages, name_to_op)
+    base = pathlib.Path(__file__).parent / "autogen" / arch
+    write_pcode(pcode, arch_data[arch]["enums"], arch, base / "str_pcode.py")
+    print(f"  {arch}: {len(pcode)} pcode entries")

extra/assembly/amd/sqtt.py

@@ -0,0 +1,458 @@
+"""SQTT (SQ Thread Trace) packet encoder and decoder for AMD GPUs.
+
+This module provides encoding and decoding of raw SQTT byte streams.
+The format is nibble-based with variable-width packets determined by a state machine.
+Uses BitField infrastructure from dsl.py, similar to GPU instruction encoding.
+"""
+from __future__ import annotations
+from typing import Iterator
+from enum import Enum
+from extra.assembly.amd.dsl import BitField, FixedBitField, bits
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# FIELD ENUMS
+# ═══════════════════════════════════════════════════════════════════════════════
+
+class MemSrc(Enum):
+  LDS = 0
+  LDS_ALT = 1
+  VMEM = 2
+  VMEM_ALT = 3
+
+class AluSrc(Enum):
+  NONE = 0
+  SALU = 1
+  VALU = 2
+  VALU_SALU = 3
+
+class InstOp(Enum):
+  """SQTT instruction operation types for RDNA3 (gfx1100).
+
+  Memory ops appear in two ranges depending on which SIMD executes them:
+  - 0x1x-0x2x range: ops on traced SIMD
+  - 0x5x range: ops on other SIMD (OTHER_ prefix)
+
+  GLOBAL memory ops encoding depends on addressing mode AND size:
+  - Loads: 0x21 (saddr=SGPR) or 0x22 (saddr=NULL), all sizes same
+  - Stores: base + size_offset, where VADDR is shifted +1 from SADDR
+    SADDR: 0x24(32) 0x25(64) 0x26(96) 0x27(128)
+    VADDR: 0x25(32) 0x26(64) 0x27(96) 0x28(128)
+
+  OTHER_ range follows same pattern but values overlap differently.
+  """
+  SALU = 0x0
+  SMEM = 0x1
+  JUMP = 0x3              # branch taken
+  JUMP_NO = 0x4           # branch not taken
+  MESSAGE = 0x9
+  VALU_TRANS = 0xb        # transcendental: exp, log, rcp, sqrt, sin, cos
+  VALU_64_SHIFT = 0xd     # 64-bit shifts: lshl, lshr, ashr
+  VALU_MAD64 = 0xe        # 64-bit multiply-add
+  VALU_64 = 0xf           # 64-bit: add, mul, fma, rcp, sqrt, rounding, frexp, div helpers
+  VINTERP = 0x12          # interpolation: v_interp_p10_f32, v_interp_p2_f32
+  BARRIER = 0x13
+
+  # FLAT memory ops on traced SIMD (0x1x range)
+  FLAT_LOAD = 0x1c
+  FLAT_STORE = 0x1d
+  FLAT_STORE_64 = 0x1e
+  FLAT_STORE_96 = 0x1f
+  FLAT_STORE_128 = 0x20
+
+  # GLOBAL memory ops on traced SIMD (0x2x range)
+  GLOBAL_LOAD = 0x21             # saddr=SGPR, all sizes
+  GLOBAL_LOAD_VADDR = 0x22       # saddr=NULL, all sizes
+  GLOBAL_STORE = 0x24            # saddr=SGPR, 32-bit
+  GLOBAL_STORE_64 = 0x25         # saddr=SGPR 64 or saddr=NULL 32
+  GLOBAL_STORE_96 = 0x26         # saddr=SGPR 96 or saddr=NULL 64
+  GLOBAL_STORE_128 = 0x27        # saddr=SGPR 128 or saddr=NULL 96
+  GLOBAL_STORE_VADDR_128 = 0x28  # saddr=NULL, 128-bit
+
+  # LDS ops on traced SIMD
+  LDS_LOAD = 0x29
+  LDS_STORE = 0x2b
+  LDS_STORE_64 = 0x2c
+  LDS_STORE_128 = 0x2e
+
+  # Memory ops on other SIMD (0x5x range)
+  OTHER_LDS_LOAD = 0x50
+  OTHER_LDS_STORE = 0x51
+  OTHER_LDS_STORE_64 = 0x52
+  OTHER_LDS_STORE_128 = 0x54
+  OTHER_FLAT_LOAD = 0x55
+  OTHER_FLAT_STORE = 0x56
+  OTHER_FLAT_STORE_64 = 0x57
+  OTHER_FLAT_STORE_96 = 0x58
+  OTHER_FLAT_STORE_128 = 0x59
+  OTHER_GLOBAL_LOAD = 0x5a             # saddr=SGPR, all sizes
+  OTHER_GLOBAL_LOAD_VADDR = 0x5b       # saddr=NULL or saddr=SGPR store 32
+  OTHER_GLOBAL_STORE_64 = 0x5c         # saddr=SGPR 64 or saddr=NULL 32
+  OTHER_GLOBAL_STORE_96 = 0x5d         # saddr=SGPR 96 or saddr=NULL 64
+  OTHER_GLOBAL_STORE_128 = 0x5e        # saddr=SGPR 128 or saddr=NULL 96
+  OTHER_GLOBAL_STORE_VADDR_128 = 0x5f  # saddr=NULL, 128-bit
+
+  # EXEC-modifying ops (0x7x range)
+  SALU_SAVEEXEC = 0x72    # s_*_saveexec_b32/b64
+  VALU_CMPX = 0x73        # v_cmpx_*
+
+class InstOpL4(Enum):
+  """SQTT instruction operation types for RDNA4 (gfx1200). Different encoding from RDNA3."""
+  # TODO: we need to do discovery of all of these from instructions
+  SALU = 0x0
+  SMEM = 0x1
+  UNK_02 = 0x2
+  JUMP_NO = 0x4
+  UNK_06 = 0x6
+  VMEM = 0x10
+  UNK_11 = 0x11
+  VINTERP = 0x12
+  UNK_14 = 0x14
+  OTHER_VMEM = 0x5e
+  UNK_60 = 0x60
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# PACKET TYPE BASE CLASS
+# ═══════════════════════════════════════════════════════════════════════════════
+
+class PacketType:
+  """Base class for SQTT packet types."""
+  encoding: FixedBitField
+  _raw: int
+  _time: int
+
+  def __init_subclass__(cls, **kwargs):
+    super().__init_subclass__(**kwargs)
+    cls._fields = {k: v for k, v in cls.__dict__.items() if isinstance(v, BitField)}
+    cls._size_nibbles = ((max((f.hi for f in cls._fields.values()), default=0) + 4) // 4)
+
+  @classmethod
+  def from_raw(cls, raw: int, time: int = 0):
+    inst = object.__new__(cls)
+    inst._raw, inst._time = raw, time
+    return inst
+
+  def __repr__(self) -> str:
+    fields_str = ", ".join(f"{k}={getattr(self, k)}" for k in self._fields if not k.startswith('_') and k != 'encoding')
+    return f"{self.__class__.__name__}({fields_str})"
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# TS PACKET TYPE DEFINITIONS
+# ═══════════════════════════════════════════════════════════════════════════════
+
+class TS_DELTA_S8_W3(PacketType):
+  encoding = bits[6:0] == 0b0100001
+  delta = bits[10:8]
+  _padding = bits[63:11]
+
+class TS_DELTA_S8_W3_L4(PacketType):  # Layout 4: 64->72 bits
+  encoding = bits[6:0] == 0b0100001
+  delta = bits[10:8]
+  _padding = bits[71:11]
+
+class TS_DELTA_S5_W3(PacketType):
+  encoding = bits[4:0] == 0b00110
+  delta = bits[7:5]
+  _padding = bits[51:8]
+
+class TS_DELTA_S5_W3_L4(PacketType):  # Layout 4: 52->56 bits
+  encoding = bits[4:0] == 0b00110
+  delta = bits[9:7]
+  _padding = bits[55:10]
+
+class TS_DELTA_SHORT(PacketType):
+  encoding = bits[3:0] == 0b1000
+  delta = bits[7:4]
+
+class TS_DELTA_OR_MARK(PacketType):
+  encoding = bits[6:0] == 0b0000001
+  delta = bits[47:12]
+  bit8 = bits[8:8]
+  bit9 = bits[9:9]
+  @property
+  def is_marker(self) -> bool: return bool(self.bit9 and not self.bit8)
+
+class TS_DELTA_OR_MARK_L4(PacketType):  # Layout 4: 48->64 bits
+  encoding = bits[6:0] == 0b0000001
+  delta = bits[63:12]
+  bit7 = bits[7:7]
+  bit8 = bits[8:8]
+  bit9 = bits[9:9]
+  @property
+  def is_marker(self) -> bool: return bool((self.bit9 and not self.bit8) or self.bit7)
+
+class TS_DELTA_S5_W2(PacketType):
+  encoding = bits[4:0] == 0b11100
+  delta = bits[6:5]
+  _padding = bits[47:7]
+
+class TS_DELTA_S5_W2_L4(PacketType):  # Layout 4: 48->40 bits
+  encoding = bits[4:0] == 0b11100
+  delta = bits[6:5]
+  _padding = bits[39:7]
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# PACKET TYPE DEFINITIONS
+# ═══════════════════════════════════════════════════════════════════════════════
+
+class VALUINST(PacketType):  # exclude: 1 << 2
+  encoding = bits[2:0] == 0b011
+  delta = bits[5:3]
+  flag = bits[6:6]
+  wave = bits[11:7]
+
+class VMEMEXEC(PacketType):  # exclude: 1 << 0
+  encoding = bits[3:0] == 0b1111
+  delta = bits[5:4]
+  src = bits[7:6].enum(MemSrc)
+
+class ALUEXEC(PacketType):  # exclude: 1 << 1
+  encoding = bits[3:0] == 0b1110
+  delta = bits[5:4]
+  src = bits[7:6].enum(AluSrc)
+
+class IMMEDIATE(PacketType):  # exclude: 1 << 5
+  encoding = bits[3:0] == 0b1101
+  delta = bits[6:4]
+  wave = bits[11:7]
+
+class IMMEDIATE_MASK(PacketType):  # exclude: 1 << 5
+  encoding = bits[4:0] == 0b00100
+  delta = bits[7:5]
+  mask = bits[23:8]
+
+class WAVERDY(PacketType):  # exclude: 1 << 3
+  encoding = bits[4:0] == 0b10100
+  delta = bits[7:5]
+  mask = bits[23:8]
+
+class WAVEEND(PacketType):  # exclude: 1 << 4
+  encoding = bits[4:0] == 0b10101
+  delta = bits[7:5]
+  flag7 = bits[8:8]
+  simd = bits[10:9]
+  cu_lo = bits[13:11]
+  wave = bits[19:15]
+  @property
+  def cu(self) -> int: return self.cu_lo | (self.flag7 << 3)
+
+class WAVESTART(PacketType):  # exclude: 1 << 4
+  encoding = bits[4:0] == 0b01100
+  delta = bits[6:5]
+  flag7 = bits[7:7]
+  simd = bits[9:8]
+  cu_lo = bits[12:10]
+  wave = bits[17:13]
+  id7 = bits[31:18]
+  @property
+  def cu(self) -> int: return self.cu_lo | (self.flag7 << 3)
+
+class WAVESTART_L4(PacketType):  # Layout 4 has wave field at different position
+  encoding = bits[4:0] == 0b01100
+  delta = bits[6:5]
+  flag7 = bits[7:7]
+  simd = bits[9:8]
+  cu_lo = bits[12:10]
+  wave = bits[19:15]
+  id7 = bits[31:20]
+  @property
+  def cu(self) -> int: return self.cu_lo | (self.flag7 << 3)
+
+class WAVEALLOC(PacketType):  # exclude: 1 << 10
+  encoding = bits[4:0] == 0b00101
+  delta = bits[7:5]
+  _padding = bits[19:8]
+
+class WAVEALLOC_L4(PacketType):  # Layout 4: 20->24 bits
+  encoding = bits[4:0] == 0b00101
+  delta = bits[7:5]
+  _padding = bits[23:8]
+
+class PERF(PacketType):  # exclude: 1 << 11
+  encoding = bits[4:0] == 0b10110
+  delta = bits[7:5]
+  arg = bits[27:8]
+
+class PERF_L4(PacketType):  # Layout 4: 28->32 bits
+  encoding = bits[4:0] == 0b10110
+  delta = bits[9:7]
+  arg = bits[31:10]
+
+class NOP(PacketType):
+  encoding = bits[3:0] == 0b0000
+  delta = None  # type: ignore
+  _padding = bits[3:0]
+
+class TS_WAVE_STATE(PacketType):
+  encoding = bits[6:0] == 0b1010001
+  delta = bits[15:7]
+  coarse = bits[23:16]
+  @property
+  def wave_interest(self) -> bool: return bool(self.coarse & 1)
+  @property
+  def terminate_all(self) -> bool: return bool(self.coarse & 8)
+
+class EVENT(PacketType):  # exclude: 1 << 7
+  encoding = bits[7:0] == 0b01100001
+  delta = bits[10:8]
+  event = bits[23:11]
+
+class EVENT_BIG(PacketType):
+  encoding = bits[7:0] == 0b11100001
+  delta = bits[10:8]
+  event = bits[31:11]
+
+class REG(PacketType):
+  encoding = bits[3:0] == 0b1001
+  delta = bits[6:4]
+  slot = bits[9:7]
+  hi_byte = bits[15:8]
+  subop = bits[31:16]
+  val32 = bits[63:32]
+  @property
+  def is_config(self) -> bool: return bool(self.hi_byte & 0x80)
+
+class SNAPSHOT(PacketType):
+  encoding = bits[6:0] == 0b1110001
+  delta = bits[9:7]
+  snap = bits[63:10]
+
+class LAYOUT_HEADER(PacketType):
+  encoding = bits[6:0] == 0b0010001
+  delta = None  # type: ignore
+  layout = bits[12:7]
+  simd = bits[14:13]
+  group = bits[17:15]
+  sel_a = bits[31:28]
+  sel_b = bits[36:33]
+  flag4 = bits[59:59]
+  _padding = bits[63:60]
+
+class INST(PacketType):
+  encoding = bits[2:0] == 0b010
+  delta = bits[6:4]
+  flag1 = bits[3:3]
+  flag2 = bits[7:7]
+  wave = bits[12:8]
+  op = bits[19:13].enum(InstOp)
+
+class INST_L4(PacketType):  # Layout 4: different delta position and InstOp encoding
+  encoding = bits[2:0] == 0b010
+  delta = bits[5:3]
+  flag1 = bits[6:6]
+  flag2 = bits[7:7]
+  wave = bits[12:8]
+  op = bits[19:13].enum(InstOpL4)
+
+class UTILCTR(PacketType):
+  encoding = bits[6:0] == 0b0110001
+  delta = bits[8:7]
+  ctr = bits[47:9]
+
+# Packet types with rocprof type IDs as keys
+PACKET_TYPES_L3: dict[int, type[PacketType]] = {
+  1: VALUINST, 2: VMEMEXEC, 3: ALUEXEC, 4: IMMEDIATE, 5: IMMEDIATE_MASK, 6: WAVERDY, 7: TS_DELTA_S8_W3, 8: WAVEEND,
+  9: WAVESTART, 10: TS_DELTA_S5_W2, 11: WAVEALLOC, 12: TS_DELTA_S5_W3, 13: PERF, 14: UTILCTR, 15: TS_DELTA_SHORT,
+  16: NOP, 17: TS_WAVE_STATE, 18: EVENT, 19: EVENT_BIG, 20: REG, 21: SNAPSHOT, 22: TS_DELTA_OR_MARK, 23: LAYOUT_HEADER, 24: INST,
+}
+PACKET_TYPES_L4: dict[int, type[PacketType]] = {
+  **PACKET_TYPES_L3,
+  7: TS_DELTA_S8_W3_L4, 9: WAVESTART_L4, 10: TS_DELTA_S5_W2_L4, 11: WAVEALLOC_L4,
+  12: TS_DELTA_S5_W3_L4, 13: PERF_L4, 22: TS_DELTA_OR_MARK_L4, 24: INST_L4,
+}
+def _build_decode_tables(packet_types: dict[int, type[PacketType]]) -> tuple[dict[int, tuple], bytes]:
+  # Build state table: byte -> opcode. Sort by mask specificity (more bits first), NOP last
+  sorted_types = sorted(packet_types.items(), key=lambda x: (-bin(x[1].encoding.mask).count('1'), x[0] == 16))
+  state_table = bytes(next((op for op, cls in sorted_types if (b & cls.encoding.mask) == cls.encoding.default), 16) for b in range(256))
+  # Build decode info: opcode -> (pkt_cls, nib_count, delta_lo, delta_mask, special_case)
+  # special_case: 0=none, 1=TS_DELTA_OR_MARK (check is_marker), 2=TS_DELTA_SHORT (add 8)
+  decode_info = {}
+  for opcode, pkt_cls in packet_types.items():
+    delta_field = getattr(pkt_cls, 'delta', None)
+    special = {22: 1, 15: 2}.get(opcode, 0)  # TS_DELTA_OR_MARK=22, TS_DELTA_SHORT=15
+    decode_info[opcode] = (pkt_cls, pkt_cls._size_nibbles, delta_field.lo if delta_field else 0, delta_field.mask if delta_field else 0, special)
+  return decode_info, state_table
+
+_DECODE_INFO_L3, _STATE_TABLE_L3 = _build_decode_tables(PACKET_TYPES_L3)
+_DECODE_INFO_L4, _STATE_TABLE_L4 = _build_decode_tables(PACKET_TYPES_L4)
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# DECODER
+# ═══════════════════════════════════════════════════════════════════════════════
+
+def decode(data: bytes) -> Iterator[PacketType]:
+  """Decode raw SQTT blob, yielding packet instances. Auto-detects layout from LAYOUT_HEADER."""
+  n, reg, pos, nib_off, nib_count, time = len(data), 0, 0, 0, 16, 0
+  decode_info, state_table = _DECODE_INFO_L3, _STATE_TABLE_L3  # default to layout 3, will update after seeing LAYOUT_HEADER
+
+  while pos + ((nib_count + nib_off + 1) >> 1) <= n:
+    need = nib_count - nib_off
+    # 1. if unaligned, read high nibble to align
+    if nib_off: reg, pos = (reg >> 4) | ((data[pos] >> 4) << 60), pos + 1
+    # 2. read all full bytes at once
+    if (byte_count := need >> 1):
+      chunk = int.from_bytes(data[pos:pos + byte_count], 'little')
+      reg, pos = (reg >> (byte_count * 8)) | (chunk << (64 - byte_count * 8)), pos + byte_count
+    # 3. if odd, read low nibble
+    if (nib_off := need & 1): reg = (reg >> 4) | ((data[pos] & 0xF) << 60)
+
+    opcode = state_table[reg & 0xFF]
+    pkt_cls, nib_count, delta_lo, delta_mask, special = decode_info[opcode]
+    delta = (reg >> delta_lo) & delta_mask
+    if special == 1:  # TS_DELTA_OR_MARK
+      pkt = pkt_cls.from_raw(reg, 0)  # create packet to check is_marker
+      if pkt.is_marker: delta = 0
+    elif special == 2: delta += 8  # TS_DELTA_SHORT
+    time += delta
+    pkt = pkt_cls.from_raw(reg, time)
+    # detect layout from first LAYOUT_HEADER and switch decode tables if needed
+    # NOTE: CDNA uses a completely different 16-bit header format, not nibbles - not supported here
+    if pkt_cls is LAYOUT_HEADER and pkt.layout == 4:
+      decode_info, state_table = _DECODE_INFO_L4, _STATE_TABLE_L4
+    yield pkt
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# PRINTER
+# ═══════════════════════════════════════════════════════════════════════════════
+
+PACKET_COLORS = {
+  "INST": "WHITE", "VALUINST": "BLACK", "VMEMEXEC": "yellow", "ALUEXEC": "yellow",
+  "IMMEDIATE": "YELLOW", "IMMEDIATE_MASK": "YELLOW", "WAVERDY": "cyan", "WAVEALLOC": "cyan",
+  "WAVEEND": "blue", "WAVESTART": "blue", "PERF": "magenta", "EVENT": "red", "EVENT_BIG": "red",
+  "REG": "green", "LAYOUT_HEADER": "white", "SNAPSHOT": "white", "UTILCTR": "green",
+}
+
+def format_packet(p) -> str:
+  from tinygrad.helpers import colored
+  name = type(p).__name__
+  if isinstance(p, (INST, INST_L4)):
+    op_name = p.op.name if isinstance(p.op, (InstOp, InstOpL4)) else f"0x{p.op:02x}"
+    fields = f"wave={p.wave} op={op_name}" + (" flag1" if p.flag1 else "") + (" flag2" if p.flag2 else "")
+  elif isinstance(p, VALUINST): fields = f"wave={p.wave}" + (" flag" if p.flag else "")
+  elif isinstance(p, ALUEXEC): fields = f"src={p.src.name if isinstance(p.src, AluSrc) else p.src}"
+  elif isinstance(p, VMEMEXEC): fields = f"src={p.src.name if isinstance(p.src, MemSrc) else p.src}"
+  elif isinstance(p, (WAVESTART, WAVESTART_L4, WAVEEND)): fields = f"wave={p.wave} simd={p.simd} cu={p.cu}"
+  elif hasattr(p, '_fields'):
+    filt = {'delta', 'encoding'} if not isinstance(p, (TS_DELTA_OR_MARK, TS_DELTA_OR_MARK_L4)) else {'encoding'}
+    fields = " ".join(f"{k}=0x{getattr(p, k):x}" if k in {'snap', 'val32'} else f"{k}={getattr(p, k)}"
+                      for k in p._fields if not k.startswith('_') and k not in filt)
+  else: fields = ""
+  return f"{p._time:8}: {colored(f'{name:18}', PACKET_COLORS.get(name.replace('_L4', ''), 'white'))} {fields}"
+
+def print_packets(packets) -> None:
+  from tinygrad.helpers import getenv
+  skip = {"NOP", "TS_DELTA_SHORT", "TS_WAVE_STATE", "TS_DELTA_OR_MARK",
+          "TS_DELTA_S5_W2", "TS_DELTA_S5_W3", "TS_DELTA_S8_W3", "REG", "EVENT"} if not getenv("NOSKIP") else {"NOP"}
+  for p in packets:
+    if type(p).__name__.replace("_L4", "") not in skip: print(format_packet(p))
+
+if __name__ == "__main__":
+  import sys, pickle
+  if len(sys.argv) < 2:
+    print("Usage: python sqtt.py <pkl_file>")
+    sys.exit(1)
+  with open(sys.argv[1], "rb") as f:
+    data = pickle.load(f)
+  sqtt_events = [e for e in data if type(e).__name__ == "ProfileSQTTEvent"]
+  for i, event in enumerate(sqtt_events):
+    print(f"\n=== event {i} ===")
+    print_packets(decode(event.blob))

extra/assembly/amd/sqtt_cdna.py

@@ -0,0 +1,161 @@
+"""SQTT (SQ Thread Trace) packet decoder for CDNA/MI300 GPUs.
+
+CDNA uses a completely different 16-bit header format from RDNA's nibble-based encoding.
+"""
+from __future__ import annotations
+from typing import Iterator
+from extra.assembly.amd.dsl import bits
+from extra.assembly.amd.sqtt import PacketType
+
+# CDNA pkt_fmt -> size in bytes (extracted from rocprof hash table)
+CDNA_PKT_SIZES = {0: 2, 1: 8, 2: 8, 3: 4, 4: 2, 5: 6, 6: 2, 7: 2, 8: 2, 9: 2, 10: 2, 11: 8, 12: 6, 13: 4, 14: 8, 15: 6}
+
+class CDNA_DELTA(PacketType):
+  """pkt_fmt=0: 16-bit timestamp delta packet"""
+  encoding = bits[3:0] == 0
+  delta = bits[11:4]      # (data >> 4) & 0xff
+  unk_0 = bits[12:12]     # (data >> 0xc) & 1
+  unk_1 = bits[15:13]     # (data >> 0xd)
+
+class CDNA_TIMESTAMP(PacketType):
+  """pkt_fmt=1: 64-bit timestamp packet (case 0x0)"""
+  encoding = bits[3:0] == 1
+  unk_0 = bits[15:4]
+  timestamp = bits[63:16]   # stored as (data_word >> 0x10) in low 46 bits of local_58
+
+class CDNA_PKT_2(PacketType):
+  """pkt_fmt=2: 64-bit packet (case 0x4)"""
+  encoding = bits[3:0] == 2
+  unk_0 = bits[6:5]       # (data >> 5) & 3
+  unk_1 = bits[7:7]       # (data >> 7) + 1 & 1
+  unk_padding = bits[63:8]
+
+class CDNA_WAVESTART(PacketType):
+  """pkt_fmt=3: 32-bit WAVESTART packet (case 0x8)"""
+  encoding = bits[3:0] == 3
+  unk_0 = bits[5:5]       # (data >> 5) & 1
+  unk_1 = bits[9:6]       # (data >> 6) & 0xf
+  wave = bits[13:10]      # (data >> 10) & 0xf
+  simd = bits[15:14]      # (data >> 0xe) & 3
+  cu = bits[17:16]        # (data >> 0x10) & 3
+  unk_5 = bits[19:18]     # (data >> 0x12) & 3
+  unk_6 = bits[28:22]     # (data >> 0x16) & 0x7f
+  unk_padding = bits[31:29]
+
+class CDNA_PKT_4(PacketType):
+  """pkt_fmt=4: 16-bit packet (case 0xc, same as 0x8/0x14)"""
+  encoding = bits[3:0] == 4
+  unk_0 = bits[5:5]       # (data_word >> 5) & 1
+  unk_1 = bits[9:6]       # (data_word >> 6) & 0xf
+  unk_2 = bits[13:10]     # (data_word >> 10) & 0xf
+  unk_3 = bits[15:14]     # (data_word >> 0xe)
+
+class CDNA_PKT_5(PacketType):
+  """pkt_fmt=5: 48-bit packet (case 0x10)"""
+  encoding = bits[3:0] == 5
+  unk_0 = bits[6:5]       # (data >> 5) & 3
+  unk_1 = bits[7:7]       # (data >> 7) + 1 & 1
+  unk_2 = bits[15:9]      # (data >> 9) & 0x7f
+  unk_padding = bits[47:16]
+
+class CDNA_WAVEEND(PacketType):
+  """pkt_fmt=6: 16-bit WAVEEND packet (case 0x14, same as 0x8/0xc)"""
+  encoding = bits[3:0] == 6
+  unk_0 = bits[5:5]       # (data_word >> 5) & 1
+  unk_1 = bits[9:6]       # (data_word >> 6) & 0xf
+  wave = bits[13:10]      # (data_word >> 10) & 0xf
+  simd = bits[15:14]      # (data_word >> 0xe)
+
+class CDNA_EXEC(PacketType):
+  """pkt_fmt=10: 16-bit EXEC packet (case 0x24)"""
+  encoding = bits[3:0] == 10
+  unk_0 = bits[8:5]       # (data_word >> 5) & 0xf
+  unk_1 = bits[10:9]      # (data_word >> 9) & 3
+  unk_2 = bits[15:11]     # (data_word >> 0xb)
+
+class CDNA_PKT_11(PacketType):
+  """pkt_fmt=11: 64-bit packet (case 0x28)"""
+  encoding = bits[3:0] == 11
+  unk_0 = bits[8:5]       # (data_word >> 5) & 0xf
+  unk_1 = bits[10:9]      # (data_word >> 9) & 3
+  unk_2 = bits[15:15]     # (data_word >> 0xf) & 1
+  unk_padding = bits[63:16]
+
+class CDNA_INST(PacketType):
+  """pkt_fmt=13: 32-bit INST packet (case 0x30)"""
+  encoding = bits[3:0] == 13
+  unk_0 = bits[6:5]       # (data >> 5) & 3
+  unk_1 = bits[9:8]       # (data >> 8) & 3
+  unk_2 = bits[11:10]     # (data >> 10) & 3
+  unk_3 = bits[13:12]     # (data >> 0xc) & 3
+  unk_4 = bits[15:14]     # (data >> 0xe) & 3
+  unk_5 = bits[19:18]     # (data >> 0x12) & 3
+  unk_6 = bits[21:20]     # (data >> 0x14) & 3
+  unk_7 = bits[23:22]     # (data >> 0x16) & 3
+  unk_8 = bits[25:24]     # (data >> 0x18) & 3
+  unk_9 = bits[27:26]     # (data >> 0x1a) & 3
+  unk_padding = bits[31:28]
+
+class CDNA_PKT_14(PacketType):
+  """pkt_fmt=14: 64-bit packet (case 0x34)"""
+  encoding = bits[3:0] == 14
+  unk_0 = bits[5:5]       # (data >> 5) & 1
+  unk_1 = bits[9:6]       # (data >> 6) & 0xf
+  unk_2 = bits[11:10]     # (data >> 10) & 3
+  unk_3 = bits[24:12]     # (data >> 0xc) & 0x1fff
+  unk_4 = bits[37:25]     # (data >> 0x19) & 0x1fff
+  unk_5 = bits[50:38]     # (data >> 0x26) & 0x1fff
+  unk_6 = bits[51:51]     # (data >> 0x33) & 1
+  unk_padding = bits[63:52]
+
+class CDNA_PKT_15(PacketType):
+  """pkt_fmt=15: 48-bit packet (case 0x38, same as 0x10)"""
+  encoding = bits[3:0] == 15
+  unk_0 = bits[6:5]       # (data >> 5) & 3
+  unk_1 = bits[7:7]       # (data >> 7) + 1 & 1
+  unk_2 = bits[15:9]      # (data >> 9) & 0x7f
+  unk_padding = bits[47:16]
+
+CDNA_PKT_TYPES: dict[int, type[PacketType]] = {
+  0: CDNA_DELTA, 1: CDNA_TIMESTAMP, 2: CDNA_PKT_2, 3: CDNA_WAVESTART, 4: CDNA_PKT_4,
+  5: CDNA_PKT_5, 6: CDNA_WAVEEND, 10: CDNA_EXEC, 11: CDNA_PKT_11, 13: CDNA_INST, 14: CDNA_PKT_14, 15: CDNA_PKT_15,
+}
+
+# Validate CDNA packet definitions
+for pkt_fmt, pkt_cls in CDNA_PKT_TYPES.items():
+  assert pkt_cls.encoding.default == pkt_fmt, f"{pkt_cls.__name__} encoding {pkt_cls.encoding.default} != pkt_fmt {pkt_fmt}"
+  assert CDNA_PKT_SIZES[pkt_fmt] * 2 == pkt_cls._size_nibbles, f"{pkt_cls.__name__} size {pkt_cls._size_nibbles//2} != {CDNA_PKT_SIZES[pkt_fmt]}"
+
+def decode(data: bytes) -> Iterator[PacketType]:
+  """Decode CDNA SQTT blob using 16-bit header format."""
+  pos, time, ts_offset = 0, 0, None
+  while pos + 2 <= len(data):
+    header = int.from_bytes(data[pos:pos+2], 'little')
+    pkt_fmt = header & 0xf
+    pkt_size = CDNA_PKT_SIZES[pkt_fmt]
+    if pos + pkt_size > len(data): break
+
+    raw = int.from_bytes(data[pos:pos+pkt_size], 'little')
+    # pkt_fmt=0 has delta in bits[11:4], accumulate it
+    if pkt_fmt == 0: time += ((raw >> 4) & 0xff) * 4
+    # pkt_fmt=1 with unk_0=0 is absolute timestamp - use it to anchor time
+    if pkt_fmt == 1 and ((raw >> 4) & 0xfff) == 0:
+      abs_ts = raw >> 16
+      if ts_offset is None: ts_offset = abs_ts - time  # first timestamp: save offset
+      else: time = ((abs_ts - ts_offset) & ~3) - 4     # subsequent: compute time, align to 4, subtract 4
+    pkt_cls = CDNA_PKT_TYPES[pkt_fmt]
+    yield pkt_cls.from_raw(raw, time)
+    pos += pkt_size
+
+if __name__ == "__main__":
+  import sys, pickle
+  if len(sys.argv) < 2:
+    print("Usage: python sqtt_cdna.py <pkl_file>")
+    sys.exit(1)
+  with open(sys.argv[1], "rb") as f:
+    data = pickle.load(f)
+  sqtt_events = [e for e in data if type(e).__name__ == "ProfileSQTTEvent"]
+  for i, event in enumerate(sqtt_events):
+    print(f"\n=== event {i} ===")
+    for pkt in decode(event.blob):
+      print(f"{pkt._time:8}: {pkt}")

extra/assembly/amd/sqttmap.py

@@ -0,0 +1,122 @@
+# maps SQTT trace packets to instructions.
+from dataclasses import dataclass
+from typing import Iterator
+
+from extra.assembly.amd.sqtt import decode, print_packets, INST, VALUINST, IMMEDIATE, WAVESTART, WAVEEND, InstOp, PacketType, IMMEDIATE_MASK
+from extra.assembly.amd.dsl import Inst
+from extra.assembly.amd.autogen.rdna3.ins import SOPP, s_endpgm
+from extra.assembly.amd.autogen.rdna3.enum import SOPPOp
+
+@dataclass(frozen=True)
+class InstructionInfo:
+  pc: int
+  wave: int
+  inst: Inst
+
+def map_insts(data:bytes, lib:bytes, target:int) -> Iterator[tuple[PacketType, InstructionInfo|None]]:
+  """maps SQTT packets to instructions, yields (packet, instruction_info or None)"""
+  # map pcs to insts
+  from tinygrad.viz.serve import amd_decode
+  pc_map = amd_decode(lib, target)
+
+  wave_pc:dict[int, int] = {}
+  # only processing packets on one [CU, SIMD] unit
+  def simd_select(p) -> bool: return getattr(p, "cu", 0) == 0 and getattr(p, "simd", 0) == 0
+  for p in decode(data):
+    if not simd_select(p): continue
+    if isinstance(p, WAVESTART):
+      assert p.wave not in wave_pc, "only one inflight wave per unit"
+      wave_pc[p.wave] = next(iter(pc_map))
+      continue
+    if isinstance(p, WAVEEND):
+      pc = wave_pc.pop(p.wave)
+      yield (p, InstructionInfo(pc, p.wave, s_endpgm()))
+      continue
+    # skip OTHER_ instructions, they don't belong to this unit
+    if isinstance(p, INST) and p.op.name.startswith("OTHER_"): continue
+    if isinstance(p, IMMEDIATE_MASK):
+      # immediate mask may yield multiple times per packet
+      for wave in range(16):
+        if p.mask & (1 << wave):
+          inst = pc_map[pc:=wave_pc[wave]]
+          # can this assert be more strict?
+          assert isinstance(inst, SOPP), f"IMMEDIATE_MASK packet must map to SOPP, got {inst}"
+          wave_pc[wave] += inst.size()
+          yield (p, InstructionInfo(pc, wave, inst))
+      continue
+    if isinstance(p, (VALUINST, INST, IMMEDIATE)):
+      inst = pc_map[pc:=wave_pc[p.wave]]
+      # s_delay_alu doesn't get a packet?
+      if isinstance(inst, SOPP) and inst.op in {SOPPOp.S_DELAY_ALU}:
+        wave_pc[p.wave] += inst.size()
+        inst = pc_map[pc:=wave_pc[p.wave]]
+      # identify a branch instruction, only used for asserts
+      is_branch = isinstance(inst, SOPP) and "BRANCH" in inst.op_name
+      if is_branch: assert isinstance(p, INST) and p.op in {InstOp.JUMP_NO, InstOp.JUMP}, f"branch can only be folowed by jump packets, got {p}"
+      # JUMP handling
+      if isinstance(p, INST) and p.op is InstOp.JUMP:
+        assert is_branch, f"JUMP packet must map to a branch instruction, got {inst}"
+        x = inst.simm16 & 0xffff
+        wave_pc[p.wave] += inst.size() + (x - 0x10000 if x & 0x8000 else x)*4
+      else:
+        if is_branch: assert inst.op != SOPPOp.S_BRANCH, f"S_BRANCH must have a JUMP packet, got {p}"
+        wave_pc[p.wave] += inst.size()
+      yield (p, InstructionInfo(pc, p.wave, inst))
+      continue
+    # for all other packets (VMEMEXEC, ALUEXEC, etc.), yield with None
+    yield (p, None)
+
+# test to compare every packet with the rocprof decoder
+
+def test_rocprof_inst_traces_match(sqtt, prg, target):
+  from tinygrad.viz.serve import amd_decode
+  from extra.sqtt.roc import decode as roc_decode, InstExec
+  addr_table = amd_decode(prg.lib, target)
+  disasm = {addr+prg.base:(inst.disasm(), inst.size()) for addr,inst in addr_table.items()}
+  rctx = roc_decode([sqtt], {prg.tag:disasm})
+  rwaves = rctx.inst_execs.get((sqtt.kern, sqtt.exec_tag), [])
+  rwaves_iter:dict[int, list[Iterator[InstExec]]] = {} # wave unit (0-15) -> list of inst trace iterators for all executions on that unit
+  for w in rwaves: rwaves_iter.setdefault(w.wave_id, []).append(w.unpack_insts())
+
+  passed_insts = 0
+  for pkt, info in map_insts(sqtt.blob, prg.lib, target):
+    if DEBUG >= 2: print_packets([pkt])
+    if info is None: continue
+    if DEBUG >= 2: print(f"{' '*29}{info.inst.disasm()}")
+    rocprof_inst = next(rwaves_iter[info.wave][0])
+    ref_pc = rocprof_inst.pc-prg.base
+    # always check pc matches
+    assert ref_pc == info.pc, f"pc mismatch {ref_pc}:{disasm[rocprof_inst.pc][0]} != {info.pc}:{info.inst.disasm()}"
+    # special handling for s_endpgm, it marks the wave completion.
+    if info.inst == s_endpgm():
+      completed_wave = list(rwaves_iter[info.wave].pop(0))
+      assert len(completed_wave) == 0, f"incomplete instructions in wave {info.wave}"
+    # otherwise the packet timestamp is time + "stall"
+    else:
+      assert pkt._time == rocprof_inst.time+rocprof_inst.stall
+    passed_insts += 1
+
+  for k,v in rwaves_iter.items():
+    assert len(v) == 0, f"incomplete wave {k}"
+
+  if len(rwaves):
+    print(f"passed for {passed_insts} instructions across {len(rwaves)} waves scheduled on {len(rwaves_iter)} wave units")
+
+if __name__ == "__main__":
+  import argparse, pickle, pathlib
+  from tinygrad.helpers import temp, DEBUG
+  parser = argparse.ArgumentParser()
+  parser.add_argument('--profile', type=pathlib.Path, metavar="PATH", help='Path to profile (optional file, default: latest profile)',
+                      default=pathlib.Path(temp("profile.pkl", append_user=True)))
+  parser.add_argument('--kernel', type=str, default=None, metavar="NAME", help='Kernel to focus on (optional name, default: all kernels)')
+  args = parser.parse_args()
+  with open(args.profile, "rb") as f:
+    data = pickle.load(f)
+  sqtt_events = [e for e in data if type(e).__name__ == "ProfileSQTTEvent"]
+  kern_events = {e.tag:e for e in data if type(e).__name__ == "ProfileProgramEvent"}
+  target = next((e for e in data if type(e).__name__ == "ProfileDeviceEvent" and e.device.startswith("AMD"))).props["gfx_target_version"]
+  for e in sqtt_events:
+    if args.kernel is not None and args.kernel != e.kern: continue
+    if not e.itrace: continue
+    print(f"==== {e.kern}")
+    test_rocprof_inst_traces_match(e, kern_events[e.kern], target)

extra/assembly/amd/test/bench_emu.py

@@ -0,0 +1,266 @@
+#!/usr/bin/env python3
+"""Benchmark comparing Python vs Rust RDNA3 emulators on real tinygrad kernels."""
+import ctypes, time, os
+from pathlib import Path
+
+# Set AMD=1 before importing tinygrad
+os.environ["AMD"] = "1"
+
+from extra.assembly.amd.emu import run_asm as python_run_asm, decode_program
+from extra.assembly.amd import decode_inst
+from extra.assembly.amd.autogen.rdna3.ins import SOPP, SOPPOp
+
+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, rsrc2: int, 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, rsrc2)
+
+  # 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, rsrc2)
+    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 profile_instructions(kernel: bytes):
+  """Profile individual instruction compile times."""
+  from extra.assembly.amd.emu import _get_runner, _canonical_runner_cache
+  from tinygrad.helpers import Context
+  _get_runner.cache_clear()
+  _canonical_runner_cache.clear()
+
+  results = []
+  i = 0
+  while i < len(kernel):
+    inst = decode_inst(kernel[i:])
+    if isinstance(inst, SOPP) and inst.op == SOPPOp.S_CODE_END: break
+    inst_bytes = bytes(kernel[i:i + inst.size() + 4])
+    try: inst_str = repr(inst)
+    except Exception: inst_str = f"<{type(inst).__name__}>"
+
+    # Time the full compile (sink + render + compile)
+    start = time.perf_counter()
+    with Context(CCACHE=0):
+      runner, is_new = _get_runner(inst_bytes)
+    compile_time = time.perf_counter() - start
+
+    results.append({
+      'inst_str': inst_str + ('' if is_new else ' [CACHED]'),
+      'compile_ms': compile_time * 1000 if is_new else 0,
+    })
+    i += inst.size()
+
+  return sorted(results, key=lambda x: x['compile_ms'], reverse=True)
+
+def benchmark_python_split(kernel: bytes, global_size, local_size, args_ptr, rsrc2: int, iterations: int = 5):
+  """Benchmark Python emulator with compile and execution times."""
+  from extra.assembly.amd.emu import _get_runner, _canonical_runner_cache
+  from tinygrad.helpers import Context
+  _get_runner.cache_clear()
+  _canonical_runner_cache.clear()
+  decode_program.cache_clear()
+
+  # Measure compile time (decode_program builds sinks, renders, and compiles)
+  compile_start = time.perf_counter()
+  with Context(CCACHE=0):
+    program = decode_program(kernel)
+  compile_time = time.perf_counter() - compile_start
+  n_compiled = len(_canonical_runner_cache)
+
+  # Execution time
+  exec_time = benchmark_emulator("Python", python_run_asm, kernel, global_size, local_size, args_ptr, rsrc2, iterations)
+  return compile_time, exec_time, len(program), n_compiled
+
+def get_tinygrad_kernel(op_name: str) -> tuple[bytes, tuple, tuple, list[int], dict[int, bytes], int] | None:
+  """Get a real tinygrad kernel by operation name. Returns (code, global_size, local_size, buf_sizes, buf_data, rsrc2)."""
+  try:
+    from tinygrad import Tensor
+    from tinygrad.runtime.support.elf import elf_loader
+    from tinygrad.runtime.autogen import hsa
+    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)
+        image = memoryview(bytearray(lib))
+        _, sections, _ = elf_loader(lib)
+        rodata_entry = next((sh.header.sh_addr for sh in sections if sh.name == ".rodata"), -1)
+        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
+            # Extract rsrc2 from ELF (same as ops_amd.py)
+            group_segment_size = image[rodata_entry:rodata_entry+4].cast("I")[0]
+            lds_size = ((group_segment_size + 511) // 512) & 0x1FF
+            code = hsa.amd_kernel_code_t.from_buffer_copy(bytes(image[rodata_entry:rodata_entry+256]) + b'\x00'*256)
+            rsrc2 = code.compute_pgm_rsrc2 | (lds_size << 15)
+            return (bytes(sec.content), tuple(lowered.prg.p.global_size), tuple(lowered.prg.p.local_size), buf_sizes, buf_data, rsrc2)
+    return None
+  except Exception as e:
+    print(f"  Error getting kernel: {e}")
+    return None
+
+TINYGRAD_TESTS = ["add", "mul", "reduce_sum", "softmax", "exp", "sin", "gelu", "matmul_small"]
+
+def main():
+  import argparse
+  parser = argparse.ArgumentParser(description="Benchmark RDNA3 emulators")
+  parser.add_argument("--iterations", type=int, default=3, help="Number of iterations per benchmark")
+  parser.add_argument("--profile", type=str, default=None, help="Profile instructions for a specific kernel (e.g. 'sin')")
+  parser.add_argument("--top", type=int, default=20, help="Number of top instructions to show in profile")
+  args = parser.parse_args()
+
+  # Profile mode: show individual instruction timing
+  if args.profile:
+    kernel_info = get_tinygrad_kernel(args.profile)
+    if kernel_info is None:
+      print(f"Failed to get kernel for '{args.profile}'")
+      return
+    kernel = kernel_info[0]
+    print(f"Profiling instructions for '{args.profile}' kernel...")
+    print("=" * 110)
+    results = profile_instructions(kernel)
+    print(f"{'Instruction':<90} {'Compile(ms)':>12}")
+    print("-" * 110)
+    for r in results[:args.top]:
+      inst = r['inst_str'][:87] + "..." if len(r['inst_str']) > 90 else r['inst_str']
+      print(f"{inst:<90} {r['compile_ms']:>12.3f}")
+    print("-" * 110)
+    total = sum(r['compile_ms'] for r in results)
+    print(f"{'TOTAL':<90} {total:>12.3f}")
+    return
+
+  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 = []
+
+  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, rsrc2 = kernel_info
+    buffers, args_arr, args_ptr, ranges = setup_buffers(buf_sizes, buf_data)
+
+    # Benchmark Python emulator (must be first to measure compile time before cache is populated)
+    py_compile, py_exec, n_insts, n_compiled = benchmark_python_split(kernel, global_size, local_size, args_ptr, rsrc2, args.iterations)
+
+    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_compiled} unique) × {n_workgroups} WGs × {n_threads} threads = {total_work:,} ops")
+    rust_time = benchmark_emulator("Rust", rust_remu.run_asm, kernel, global_size, local_size, args_ptr, rsrc2, args.iterations) if rust_remu else None
+
+    if py_compile is not None:
+      py_exec_rate = total_work / py_exec / 1e6
+      print(f"  Compile:        {py_compile*1000:8.3f} ms  ({n_compiled} unique)")
+      print(f"  Exec:           {py_exec*1000:8.3f} ms  ({py_exec_rate:7.2f} M ops/s)")
+    if rust_time:
+      rust_rate = total_work / rust_time / 1e6
+      speedup = py_exec / rust_time if py_exec else 0
+      print(f"  Rust:           {rust_time*1000:8.3f} ms  ({rust_rate:7.2f} M ops/s)  [{speedup:.1f}x faster]")
+
+    results.append((op_name, n_insts, n_compiled, n_workgroups, py_compile, py_exec, rust_time))
+
+  # Summary table
+  print("\n" + "=" * 110)
+  print("SUMMARY")
+  print("=" * 110)
+  print(f"{'Name':<16} {'Insts':<6} {'Unique':<6} {'WGs':<5} {'Compile (ms)':<14} {'Exec (ms)':<12} {'Rust (ms)':<12} {'Speedup':<10}")
+  print("-" * 110)
+
+  for name, n_insts, n_compiled, n_wgs, py_compile, py_exec, rust_time in results:
+    compile_ms = f"{py_compile*1000:.3f}" if py_compile else "error"
+    exec_ms = f"{py_exec*1000:.3f}" if py_exec else "error"
+    if rust_time:
+      rust_ms = f"{rust_time*1000:.3f}"
+      speedup = f"{py_exec/rust_time:.1f}x" if py_exec else "N/A"
+    else:
+      rust_ms, speedup = "N/A", "N/A"
+    print(f"{name:<16} {n_insts:<6} {n_compiled:<6} {n_wgs:<5} {compile_ms:<14} {exec_ms:<12} {rust_ms:<12} {speedup:<10}")
+
+if __name__ == "__main__":
+  main()

extra/assembly/amd/test/helpers.py

@@ -0,0 +1,38 @@
+"""Shared test helpers for RDNA3 tests."""
+import shutil
+from dataclasses import dataclass
+
+@dataclass
+class KernelInfo:
+  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
+
+# 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")
+
+ARCH_TO_TARGET:dict[str, list[str]] = {
+  "rdna3":["gfx1100"],
+  "rdna4":["gfx1200"],
+  "cdna":["gfx950", "gfx942"],
+}
+
+TARGET_TO_ARCH:dict[str, str] = {t:arch for arch,targets in ARCH_TO_TARGET.items() for t in targets}
+
+def get_target(arch:str) -> str: return ARCH_TO_TARGET[arch][0]
+
+def get_mattr(arch:str) -> str:
+  return {"rdna3":"+real-true16,+wavefrontsize32", "rdna4":"+real-true16,+wavefrontsize32", "cdna":"+wavefrontsize64"}[arch]

extra/assembly/amd/test/hw/__init__.py

@@ -0,0 +1 @@
+"""Hardware-validated emulator tests for RDNA3 instructions."""

extra/assembly/amd/test/hw/helpers.py

@@ -0,0 +1,281 @@
+"""Test infrastructure for hardware-validated RDNA3 emulator tests.
+
+Uses run_asm() with memory output, so tests can run on both emulator and real hardware.
+Set USE_HW=1 to run on both emulator and hardware, comparing results.
+"""
+import ctypes, math, os, struct
+from extra.assembly.amd.autogen.rdna3.ins import *
+
+from extra.assembly.amd.emu import run_asm
+from extra.assembly.amd.dsl import NULL, SCC, VCC_LO, VCC_HI, EXEC_LO, EXEC_HI, M0
+
+def _i32(f: float) -> int: return struct.unpack('<I', struct.pack('<f', f))[0]
+def _f32(i: int) -> float: return struct.unpack('<f', struct.pack('<I', i & 0xFFFFFFFF))[0]
+
+# f16 conversion helpers
+def f16(i: int) -> float: return struct.unpack('<e', struct.pack('<H', i & 0xFFFF))[0]
+def f32_to_f16(f: float) -> int:
+  f = float(f)
+  if math.isnan(f): return 0x7e00
+  if math.isinf(f): return 0x7c00 if f > 0 else 0xfc00
+  try: return struct.unpack('<H', struct.pack('<e', f))[0]
+  except OverflowError: return 0x7c00 if f > 0 else 0xfc00
+
+# For backwards compatibility with tests using SrcEnum.NULL etc.
+class SrcEnum:
+  NULL = NULL
+  VCC_LO = VCC_LO
+  VCC_HI = VCC_HI
+  EXEC_LO = EXEC_LO
+  EXEC_HI = EXEC_HI
+  SCC = SCC
+  M0 = M0
+  POS_HALF = 0.5
+  NEG_HALF = -0.5
+  POS_ONE = 1.0
+  NEG_ONE = -1.0
+  POS_TWO = 2.0
+  NEG_TWO = -2.0
+  POS_FOUR = 4.0
+  NEG_FOUR = -4.0
+
+VCC = VCC_LO  # For VOP3SD sdst field (VCC_LO is exported from dsl)
+USE_HW = os.environ.get("USE_HW", "0") == "1"
+FLOAT_TOLERANCE = 1e-5
+
+def get_gpu_target() -> tuple[int, int, int]:
+  """Get the GPU target as (major, minor, stepping) tuple."""
+  if not USE_HW: return (0, 0, 0)
+  from tinygrad.device import Device
+  return Device["AMD"].target
+
+def skip_unless_gfx(min_major: int, min_minor: int = 0, reason: str = ""):
+  """Skip test if GPU target is below the minimum required version."""
+  import unittest
+  def decorator(test_func):
+    if not USE_HW: return test_func
+    target = get_gpu_target()
+    if target[0] < min_major or (target[0] == min_major and target[1] < min_minor):
+      return unittest.skip(reason or f"requires gfx{min_major}{min_minor}0+")(test_func)
+    return test_func
+  return decorator
+
+# Output buffer layout: vgpr[16][32], sgpr[16], vcc, scc, exec
+N_VGPRS, N_SGPRS, WAVE_SIZE = 16, 16, 32
+VGPR_BYTES = N_VGPRS * WAVE_SIZE * 4  # 16 regs * 32 lanes * 4 bytes = 2048
+SGPR_BYTES = N_SGPRS * 4  # 16 regs * 4 bytes = 64
+OUT_BYTES = VGPR_BYTES + SGPR_BYTES + 12  # + vcc + scc + exec
+
+# Float conversion helpers
+def f2i(f: float) -> int: return _i32(f)
+def i2f(i: int) -> float: return _f32(i)
+def f2i64(f: float) -> int: return struct.unpack('<Q', struct.pack('<d', f))[0]
+def i642f(i: int) -> float: return struct.unpack('<d', struct.pack('<Q', i))[0]
+
+def assemble(instructions: list) -> bytes:
+  return b''.join(inst.to_bytes() for inst in instructions)
+
+# Simple WaveState class for test output parsing (mirrors emu.py interface for tests)
+class WaveState:
+  def __init__(self):
+    self.vgpr = [[0] * 256 for _ in range(32)]  # vgpr[lane][reg]
+    self.sgpr = [0] * 128
+    self.vcc = 0
+    self.scc = 0
+
+def get_prologue_epilogue(n_lanes: int) -> tuple[list, list]:
+  """Generate prologue and epilogue instructions for state capture."""
+  prologue = [
+    s_mov_b32(s[80], s[0]),
+    s_mov_b32(s[81], s[1]),
+    v_mov_b32_e32(v[255], v[0]),
+  ]
+  for i in range(N_VGPRS):
+    prologue.append(v_mov_b32_e32(v[i], 0))
+  for i in range(N_SGPRS):
+    prologue.append(s_mov_b32(s[i], 0))
+  prologue.append(s_mov_b32(VCC_LO, 0))
+
+  epilogue = [
+    s_mov_b32(s[90], VCC_LO),
+    s_cselect_b32(s[91], 1, 0),
+    # Save EXEC early (before we modify it for VGPR stores)
+    s_mov_b32(s[95], EXEC_LO),
+    # Restore EXEC to all active lanes for VGPR stores (test may have modified EXEC)
+    s_mov_b32(EXEC_LO, (1 << n_lanes) - 1),
+    s_load_b64(s[92:93], s[80:81], 0, soffset=NULL),
+    s_waitcnt(0),  # simm16=0 waits for all
+    v_lshlrev_b32_e32(v[240], 2, v[255]),
+  ]
+  for i in range(N_VGPRS):
+    epilogue.append(global_store_b32(addr=v[240], data=v[i], saddr=s[92:93], offset=i * WAVE_SIZE * 4))
+  epilogue.append(v_mov_b32_e32(v[241], 0))
+  epilogue.append(v_cmp_eq_u32_e32(v[255], v[241]))
+  epilogue.append(s_and_saveexec_b32(s[94], VCC_LO))
+  epilogue.append(v_mov_b32_e32(v[240], 0))
+  for i in range(N_SGPRS):
+    epilogue.append(v_mov_b32_e32(v[243], s[i]))
+    epilogue.append(global_store_b32(addr=v[240], data=v[243], saddr=s[92:93], offset=VGPR_BYTES + i * 4))
+  epilogue.append(v_mov_b32_e32(v[243], s[90]))
+  epilogue.append(global_store_b32(addr=v[240], data=v[243], saddr=s[92:93], offset=VGPR_BYTES + SGPR_BYTES))
+  epilogue.append(v_mov_b32_e32(v[243], s[91]))
+  epilogue.append(global_store_b32(addr=v[240], data=v[243], saddr=s[92:93], offset=VGPR_BYTES + SGPR_BYTES + 4))
+  # Store EXEC (saved earlier in s[95])
+  epilogue.append(v_mov_b32_e32(v[243], s[95]))
+  epilogue.append(global_store_b32(addr=v[240], data=v[243], saddr=s[92:93], offset=VGPR_BYTES + SGPR_BYTES + 8))
+  epilogue.append(s_mov_b32(EXEC_LO, s[94]))
+  epilogue.append(s_endpgm())
+  return prologue, epilogue
+
+def parse_output(out_buf: bytes, n_lanes: int) -> WaveState:
+  """Parse output buffer into WaveState."""
+  st = WaveState()
+  for i in range(N_VGPRS):
+    for lane in range(n_lanes):
+      off = i * WAVE_SIZE * 4 + lane * 4
+      st.vgpr[lane][i] = struct.unpack_from('<I', out_buf, off)[0]
+  for i in range(N_SGPRS):
+    st.sgpr[i] = struct.unpack_from('<I', out_buf, VGPR_BYTES + i * 4)[0]
+  st.vcc = struct.unpack_from('<I', out_buf, VGPR_BYTES + SGPR_BYTES)[0]
+  st.scc = struct.unpack_from('<I', out_buf, VGPR_BYTES + SGPR_BYTES + 4)[0]
+  # Store EXEC in its proper location (index 126)
+  st.sgpr[EXEC_LO.offset] = struct.unpack_from('<I', out_buf, VGPR_BYTES + SGPR_BYTES + 8)[0]
+  return st
+
+def run_program_emu(instructions: list, n_lanes: int = 1) -> WaveState:
+  """Run instructions via emulator run_asm, dump state to memory, return WaveState."""
+  out_buf = (ctypes.c_uint8 * OUT_BYTES)(*([0] * OUT_BYTES))
+  out_addr = ctypes.addressof(out_buf)
+
+  prologue, epilogue = get_prologue_epilogue(n_lanes)
+  code = assemble(prologue + instructions + epilogue)
+
+  args = (ctypes.c_uint64 * 1)(out_addr)
+  args_ptr = ctypes.addressof(args)
+  kernel_buf = (ctypes.c_char * len(code)).from_buffer_copy(code)
+  lib_ptr = ctypes.addressof(kernel_buf)
+
+  # rsrc2: USER_SGPR_COUNT=2, ENABLE_SGPR_WORKGROUP_ID_X/Y/Z=1, LDS_SIZE=128 (64KB)
+  rsrc2 = 0x19c | (128 << 15)
+  scratch_size = 0x10000  # 64KB per lane, matches .amdhsa_private_segment_fixed_size in run_program_hw
+  result = run_asm(lib_ptr, len(code), 1, 1, 1, n_lanes, 1, 1, args_ptr, rsrc2, scratch_size)
+  assert result == 0, f"run_asm failed with {result}"
+
+  return parse_output(bytes(out_buf), n_lanes)
+
+def run_program_hw(instructions: list, n_lanes: int = 1) -> WaveState:
+  """Run instructions on real AMD hardware via HIPCompiler and AMDProgram."""
+  from tinygrad.device import Device
+  from tinygrad.runtime.ops_amd import AMDProgram
+  from tinygrad.runtime.support.compiler_amd import HIPCompiler
+  from tinygrad.helpers import flat_mv
+
+  dev = Device["AMD"]
+  compiler = HIPCompiler(dev.arch)
+
+  prologue, epilogue = get_prologue_epilogue(n_lanes)
+  code = assemble(prologue + instructions + epilogue)
+
+  byte_str = ', '.join(f'0x{b:02x}' for b in code)
+  asm_src = f""".text
+.globl test
+.p2align 8
+.type test,@function
+test:
+.byte {byte_str}
+
+.rodata
+.p2align 6
+.amdhsa_kernel test
+  .amdhsa_next_free_vgpr 256
+  .amdhsa_next_free_sgpr 96
+  .amdhsa_wavefront_size32 1
+  .amdhsa_user_sgpr_kernarg_segment_ptr 1
+  .amdhsa_kernarg_size 8
+  .amdhsa_group_segment_fixed_size 65536
+  .amdhsa_private_segment_fixed_size 65536
+  .amdhsa_enable_private_segment 1
+.end_amdhsa_kernel
+
+.amdgpu_metadata
+---
+amdhsa.version:
+  - 1
+  - 0
+amdhsa.kernels:
+  - .name: test
+    .symbol: test.kd
+    .kernarg_segment_size: 8
+    .group_segment_fixed_size: 65536
+    .private_segment_fixed_size: 65536
+    .kernarg_segment_align: 8
+    .wavefront_size: 32
+    .sgpr_count: 96
+    .vgpr_count: 256
+    .max_flat_workgroup_size: 1024
+...
+.end_amdgpu_metadata
+"""
+
+  lib = compiler.compile(asm_src)
+  prg = AMDProgram(dev, "test", lib)
+
+  out_gpu = dev.allocator.alloc(OUT_BYTES)
+  assert out_gpu.va_addr % 16 == 0, f"buffer not 16-byte aligned: 0x{out_gpu.va_addr:x}"
+  prg(out_gpu, global_size=(1, 1, 1), local_size=(n_lanes, 1, 1), wait=True)
+
+  out_buf = bytearray(OUT_BYTES)
+  dev.allocator._copyout(flat_mv(memoryview(out_buf)), out_gpu)
+
+  return parse_output(bytes(out_buf), n_lanes)
+
+def compare_wave_states(emu_st: WaveState, hw_st: WaveState, n_lanes: int, n_vgprs: int = N_VGPRS, ulp_tolerance: int = 0) -> list[str]:
+  """Compare two WaveStates and return list of differences.
+
+  Args:
+    ulp_tolerance: Allow up to this many ULPs difference for float comparisons (0 = exact match required)
+  """
+  import math
+  diffs = []
+  for i in range(n_vgprs):
+    for lane in range(n_lanes):
+      emu_val = emu_st.vgpr[lane][i]
+      hw_val = hw_st.vgpr[lane][i]
+      if emu_val != hw_val:
+        emu_f, hw_f = _f32(emu_val), _f32(hw_val)
+        if math.isnan(emu_f) and math.isnan(hw_f):
+          continue
+        # Check ULP difference for floats (only for same-sign values)
+        if ulp_tolerance > 0 and (emu_val < 0x80000000) == (hw_val < 0x80000000):
+          ulp_diff = abs(int(emu_val) - int(hw_val))
+          if ulp_diff <= ulp_tolerance:
+            continue
+        diffs.append(f"v[{i}] lane {lane}: emu=0x{emu_val:08x} ({emu_f:.6g}) hw=0x{hw_val:08x} ({hw_f:.6g})")
+  for i in range(N_SGPRS):
+    emu_val = emu_st.sgpr[i]
+    hw_val = hw_st.sgpr[i]
+    if emu_val != hw_val:
+      diffs.append(f"s[{i}]: emu=0x{emu_val:08x} hw=0x{hw_val:08x}")
+  if emu_st.vcc != hw_st.vcc:
+    diffs.append(f"vcc: emu=0x{emu_st.vcc:08x} hw=0x{hw_st.vcc:08x}")
+  if emu_st.scc != hw_st.scc:
+    diffs.append(f"scc: emu={emu_st.scc} hw={hw_st.scc}")
+  return diffs
+
+def run_program(instructions: list, n_lanes: int = 1, ulp_tolerance: int = 0) -> WaveState:
+  """Run instructions and return WaveState.
+
+  If USE_HW=1, runs on both emulator and hardware, compares results, and raises if they differ.
+  Otherwise, runs only on emulator.
+
+  Args:
+    ulp_tolerance: Allow up to this many ULPs difference for float comparisons (0 = exact match required)
+  """
+  emu_st = run_program_emu(instructions, n_lanes)
+  if USE_HW:
+    hw_st = run_program_hw(instructions, n_lanes)
+    diffs = compare_wave_states(emu_st, hw_st, n_lanes, ulp_tolerance=ulp_tolerance)
+    if diffs:
+      raise AssertionError(f"Emulator vs Hardware mismatch:\n" + "\n".join(diffs))
+    return hw_st
+  return emu_st

extra/assembly/amd/test/hw/test_ds.py

@@ -0,0 +1,765 @@
+"""Tests for DS instructions - data share (LDS) operations.
+
+Includes: ds_store_b32, ds_load_b32, ds_store_2addr_*, ds_load_2addr_*,
+          ds_add_*, ds_max_*, ds_min_*, ds_and_*, ds_or_*, ds_xor_*,
+          ds_inc_*, ds_dec_*, ds_cmpstore_*, ds_storexchg_*
+"""
+import unittest
+from extra.assembly.amd.test.hw.helpers import *
+
+class TestDS2Addr(unittest.TestCase):
+  """Tests for DS_*_2ADDR instructions."""
+
+  def test_ds_store_load_2addr_b32(self):
+    """DS_STORE_2ADDR_B32 and DS_LOAD_2ADDR_B32 with offset * 4."""
+    instructions = [
+      v_mov_b32_e32(v[10], 0),
+      s_mov_b32(s[0], 0xAAAAAAAA),
+      v_mov_b32_e32(v[0], s[0]),
+      s_mov_b32(s[0], 0xBBBBBBBB),
+      v_mov_b32_e32(v[1], s[0]),
+      DS(DSOp.DS_STORE_2ADDR_B32, addr=v[10], data0=v[0], data1=v[1], vdst=v[0], offset0=0, offset1=1),
+      s_waitcnt(lgkmcnt=0),
+      DS(DSOp.DS_LOAD_2ADDR_B32, addr=v[10], vdst=v[2:3], offset0=0, offset1=1),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0xAAAAAAAA)
+    self.assertEqual(st.vgpr[0][3], 0xBBBBBBBB)
+
+  def test_ds_store_load_2addr_b64(self):
+    """DS_STORE_2ADDR_B64 and DS_LOAD_2ADDR_B64."""
+    instructions = [
+      v_mov_b32_e32(v[10], 0),
+      s_mov_b32(s[0], 0xDEADBEEF),
+      v_mov_b32_e32(v[0], s[0]),
+      s_mov_b32(s[0], 0xCAFEBABE),
+      v_mov_b32_e32(v[1], s[0]),
+      s_mov_b32(s[0], 0x12345678),
+      v_mov_b32_e32(v[2], s[0]),
+      s_mov_b32(s[0], 0x9ABCDEF0),
+      v_mov_b32_e32(v[3], s[0]),
+      DS(DSOp.DS_STORE_2ADDR_B64, addr=v[10], data0=v[0:1], data1=v[2:3], vdst=v[0], offset0=0, offset1=2),
+      s_waitcnt(lgkmcnt=0),
+      DS(DSOp.DS_LOAD_2ADDR_B64, addr=v[10], vdst=v[4:7], offset0=0, offset1=2),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][4], 0xDEADBEEF)
+    self.assertEqual(st.vgpr[0][5], 0xCAFEBABE)
+    self.assertEqual(st.vgpr[0][6], 0x12345678)
+    self.assertEqual(st.vgpr[0][7], 0x9ABCDEF0)
+
+
+class TestDS2AddrMore(unittest.TestCase):
+  """Additional DS_*_2ADDR tests."""
+
+  def test_ds_store_load_2addr_b32_nonzero_offsets(self):
+    """DS_STORE_2ADDR_B32 with non-zero offsets (offset*4 scaling)."""
+    instructions = [
+      v_mov_b32_e32(v[10], 0),
+      s_mov_b32(s[2], 0x11111111),
+      v_mov_b32_e32(v[0], s[2]),
+      s_mov_b32(s[2], 0x22222222),
+      v_mov_b32_e32(v[1], s[2]),
+      DS(DSOp.DS_STORE_2ADDR_B32, addr=v[10], data0=v[0], data1=v[1], vdst=v[0], offset0=2, offset1=5),
+      s_waitcnt(lgkmcnt=0),
+      DS(DSOp.DS_LOAD_2ADDR_B32, addr=v[10], vdst=v[2:3], offset0=2, offset1=5),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0x11111111, "v2 should have value from offset 8 (2*4)")
+    self.assertEqual(st.vgpr[0][3], 0x22222222, "v3 should have value from offset 20 (5*4)")
+
+  def test_ds_2addr_b64_no_overlap(self):
+    """DS_LOAD_2ADDR_B64 with adjacent offsets should not overlap."""
+    instructions = [
+      v_mov_b32_e32(v[10], 0),
+      s_mov_b32(s[2], 0x11111111),
+      v_mov_b32_e32(v[0], s[2]),
+      ds_store_b32(addr=v[10], data0=v[0], offset0=0),
+      s_mov_b32(s[2], 0x22222222),
+      v_mov_b32_e32(v[0], s[2]),
+      ds_store_b32(addr=v[10], data0=v[0], offset0=4),
+      s_mov_b32(s[2], 0x33333333),
+      v_mov_b32_e32(v[0], s[2]),
+      ds_store_b32(addr=v[10], data0=v[0], offset0=8),
+      s_mov_b32(s[2], 0x44444444),
+      v_mov_b32_e32(v[0], s[2]),
+      ds_store_b32(addr=v[10], data0=v[0], offset0=12),
+      s_waitcnt(lgkmcnt=0),
+      DS(DSOp.DS_LOAD_2ADDR_B64, addr=v[10], vdst=v[4:7], offset0=0, offset1=1),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][4], 0x11111111, "v4 should be 0x11111111")
+    self.assertEqual(st.vgpr[0][5], 0x22222222, "v5 should be 0x22222222")
+    self.assertEqual(st.vgpr[0][6], 0x33333333, "v6 should be 0x33333333")
+    self.assertEqual(st.vgpr[0][7], 0x44444444, "v7 should be 0x44444444")
+
+  def test_ds_load_2addr_b32_no_overwrite(self):
+    """DS_LOAD_2ADDR_B32 should only write 2 VGPRs."""
+    instructions = [
+      v_mov_b32_e32(v[10], 0),
+      s_mov_b32(s[2], 0xAAAAAAAA),
+      v_mov_b32_e32(v[0], s[2]),
+      s_mov_b32(s[2], 0xBBBBBBBB),
+      v_mov_b32_e32(v[1], s[2]),
+      DS(DSOp.DS_STORE_2ADDR_B32, addr=v[10], data0=v[0], data1=v[1], vdst=v[0], offset0=0, offset1=1),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[2], 0xDEADBEEF),
+      v_mov_b32_e32(v[4], s[2]),  # Sentinel
+      DS(DSOp.DS_LOAD_2ADDR_B32, addr=v[10], vdst=v[2:3], offset0=0, offset1=1),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0xAAAAAAAA)
+    self.assertEqual(st.vgpr[0][3], 0xBBBBBBBB)
+    self.assertEqual(st.vgpr[0][4], 0xDEADBEEF, "v4 should be untouched")
+
+  def test_ds_load_b64_no_overwrite(self):
+    """DS_LOAD_B64 should only write 2 VGPRs."""
+    instructions = [
+      v_mov_b32_e32(v[10], 0),
+      s_mov_b32(s[2], 0xDEADBEEF),
+      v_mov_b32_e32(v[0], s[2]),
+      s_mov_b32(s[2], 0xCAFEBABE),
+      v_mov_b32_e32(v[1], s[2]),
+      ds_store_b64(addr=v[10], data0=v[0:1], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[2], 0x12345678),
+      v_mov_b32_e32(v[4], s[2]),  # Sentinel
+      ds_load_b64(addr=v[10], vdst=v[2:3], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0xDEADBEEF)
+    self.assertEqual(st.vgpr[0][3], 0xCAFEBABE)
+    self.assertEqual(st.vgpr[0][4], 0x12345678, "v4 should be untouched")
+
+
+class TestDSB96(unittest.TestCase):
+  """Tests for DS_STORE_B96 and DS_LOAD_B96 (96-bit / 3 dwords)."""
+
+  def test_ds_store_load_b96(self):
+    """DS_STORE_B96 stores 3 VGPRs, DS_LOAD_B96 loads them back."""
+    instructions = [
+      v_mov_b32_e32(v[10], 0),
+      s_mov_b32(s[0], 0x11111111),
+      v_mov_b32_e32(v[0], s[0]),
+      s_mov_b32(s[0], 0x22222222),
+      v_mov_b32_e32(v[1], s[0]),
+      s_mov_b32(s[0], 0x33333333),
+      v_mov_b32_e32(v[2], s[0]),
+      ds_store_b96(addr=v[10], data0=v[0:2]),
+      s_waitcnt(lgkmcnt=0),
+      ds_load_b96(addr=v[10], vdst=v[4:6]),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][4], 0x11111111, "v4 should have first dword")
+    self.assertEqual(st.vgpr[0][5], 0x22222222, "v5 should have second dword")
+    self.assertEqual(st.vgpr[0][6], 0x33333333, "v6 should have third dword")
+
+  def test_ds_store_b96_with_offset(self):
+    """DS_STORE_B96 with non-zero offset."""
+    instructions = [
+      v_mov_b32_e32(v[10], 0),
+      s_mov_b32(s[0], 0xAAAAAAAA),
+      v_mov_b32_e32(v[0], s[0]),
+      s_mov_b32(s[0], 0xBBBBBBBB),
+      v_mov_b32_e32(v[1], s[0]),
+      s_mov_b32(s[0], 0xCCCCCCCC),
+      v_mov_b32_e32(v[2], s[0]),
+      DS(DSOp.DS_STORE_B96, addr=v[10], data0=v[0:2], offset0=12),
+      s_waitcnt(lgkmcnt=0),
+      DS(DSOp.DS_LOAD_B96, addr=v[10], vdst=v[4:6], offset0=12),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][4], 0xAAAAAAAA)
+    self.assertEqual(st.vgpr[0][5], 0xBBBBBBBB)
+    self.assertEqual(st.vgpr[0][6], 0xCCCCCCCC)
+
+
+class TestDSB128(unittest.TestCase):
+  """Tests for DS_STORE_B128 and DS_LOAD_B128 (128-bit / 4 dwords)."""
+
+  def test_ds_store_load_b128(self):
+    """DS_STORE_B128 stores 4 VGPRs, DS_LOAD_B128 loads them back."""
+    instructions = [
+      v_mov_b32_e32(v[10], 0),
+      s_mov_b32(s[0], 0x11111111),
+      v_mov_b32_e32(v[0], s[0]),
+      s_mov_b32(s[0], 0x22222222),
+      v_mov_b32_e32(v[1], s[0]),
+      s_mov_b32(s[0], 0x33333333),
+      v_mov_b32_e32(v[2], s[0]),
+      s_mov_b32(s[0], 0x44444444),
+      v_mov_b32_e32(v[3], s[0]),
+      ds_store_b128(addr=v[10], data0=v[0:3]),
+      s_waitcnt(lgkmcnt=0),
+      ds_load_b128(addr=v[10], vdst=v[4:7]),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][4], 0x11111111, "v4 should have first dword")
+    self.assertEqual(st.vgpr[0][5], 0x22222222, "v5 should have second dword")
+    self.assertEqual(st.vgpr[0][6], 0x33333333, "v6 should have third dword")
+    self.assertEqual(st.vgpr[0][7], 0x44444444, "v7 should have fourth dword")
+
+  def test_ds_store_b128_with_offset(self):
+    """DS_STORE_B128 with non-zero offset."""
+    instructions = [
+      v_mov_b32_e32(v[10], 0),
+      s_mov_b32(s[0], 0xAAAAAAAA),
+      v_mov_b32_e32(v[0], s[0]),
+      s_mov_b32(s[0], 0xBBBBBBBB),
+      v_mov_b32_e32(v[1], s[0]),
+      s_mov_b32(s[0], 0xCCCCCCCC),
+      v_mov_b32_e32(v[2], s[0]),
+      s_mov_b32(s[0], 0xDDDDDDDD),
+      v_mov_b32_e32(v[3], s[0]),
+      DS(DSOp.DS_STORE_B128, addr=v[10], data0=v[0:3], offset0=16),
+      s_waitcnt(lgkmcnt=0),
+      DS(DSOp.DS_LOAD_B128, addr=v[10], vdst=v[4:7], offset0=16),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][4], 0xAAAAAAAA)
+    self.assertEqual(st.vgpr[0][5], 0xBBBBBBBB)
+    self.assertEqual(st.vgpr[0][6], 0xCCCCCCCC)
+    self.assertEqual(st.vgpr[0][7], 0xDDDDDDDD)
+
+
+class TestDSAtomic(unittest.TestCase):
+  """Tests for DS atomic operations."""
+
+  def test_ds_max_rtn_u32(self):
+    """DS_MAX_RTN_U32: atomically store max and return old value."""
+    instructions = [
+      v_mov_b32_e32(v[10], 0),
+      s_mov_b32(s[2], 100),
+      v_mov_b32_e32(v[0], s[2]),
+      ds_store_b32(addr=v[10], data0=v[0], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[2], 200),
+      v_mov_b32_e32(v[1], s[2]),
+      ds_max_rtn_u32(addr=v[10], data0=v[1], vdst=v[2], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      ds_load_b32(addr=v[10], vdst=v[3], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 100, "v2 should have old value (100)")
+    self.assertEqual(st.vgpr[0][3], 200, "v3 should have max(100, 200) = 200")
+
+  def test_ds_min_rtn_u32(self):
+    """DS_MIN_RTN_U32: atomically store min and return old value."""
+    instructions = [
+      v_mov_b32_e32(v[10], 0),
+      s_mov_b32(s[2], 200),
+      v_mov_b32_e32(v[0], s[2]),
+      ds_store_b32(addr=v[10], data0=v[0], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[2], 100),
+      v_mov_b32_e32(v[1], s[2]),
+      ds_min_rtn_u32(addr=v[10], data0=v[1], vdst=v[2], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      ds_load_b32(addr=v[10], vdst=v[3], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 200)
+    self.assertEqual(st.vgpr[0][3], 100)
+
+  def test_ds_and_rtn_b32(self):
+    """DS_AND_RTN_B32: atomically AND and return old value."""
+    instructions = [
+      v_mov_b32_e32(v[10], 0),
+      s_mov_b32(s[2], 0xFF00FF00),
+      v_mov_b32_e32(v[0], s[2]),
+      ds_store_b32(addr=v[10], data0=v[0], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[2], 0xFFFF0000),
+      v_mov_b32_e32(v[1], s[2]),
+      ds_and_rtn_b32(addr=v[10], data0=v[1], vdst=v[2], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      ds_load_b32(addr=v[10], vdst=v[3], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0xFF00FF00)
+    self.assertEqual(st.vgpr[0][3], 0xFF000000)
+
+  def test_ds_or_rtn_b32(self):
+    """DS_OR_RTN_B32: atomically OR and return old value."""
+    instructions = [
+      v_mov_b32_e32(v[10], 0),
+      s_mov_b32(s[2], 0x00FF0000),
+      v_mov_b32_e32(v[0], s[2]),
+      ds_store_b32(addr=v[10], data0=v[0], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[2], 0x000000FF),
+      v_mov_b32_e32(v[1], s[2]),
+      ds_or_rtn_b32(addr=v[10], data0=v[1], vdst=v[2], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      ds_load_b32(addr=v[10], vdst=v[3], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0x00FF0000)
+    self.assertEqual(st.vgpr[0][3], 0x00FF00FF)
+
+  def test_ds_xor_rtn_b32(self):
+    """DS_XOR_RTN_B32: atomically XOR and return old value."""
+    instructions = [
+      v_mov_b32_e32(v[10], 0),
+      s_mov_b32(s[2], 0xAAAAAAAA),
+      v_mov_b32_e32(v[0], s[2]),
+      ds_store_b32(addr=v[10], data0=v[0], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[2], 0xFFFFFFFF),
+      v_mov_b32_e32(v[1], s[2]),
+      ds_xor_rtn_b32(addr=v[10], data0=v[1], vdst=v[2], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      ds_load_b32(addr=v[10], vdst=v[3], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0xAAAAAAAA)
+    self.assertEqual(st.vgpr[0][3], 0x55555555)
+
+  def test_ds_inc_rtn_u32(self):
+    """DS_INC_RTN_U32: increment with wrap."""
+    instructions = [
+      v_mov_b32_e32(v[10], 0),
+      s_mov_b32(s[2], 5),
+      v_mov_b32_e32(v[0], s[2]),
+      ds_store_b32(addr=v[10], data0=v[0], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[2], 10),  # limit
+      v_mov_b32_e32(v[1], s[2]),
+      ds_inc_rtn_u32(addr=v[10], data0=v[1], vdst=v[2], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      ds_load_b32(addr=v[10], vdst=v[3], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 5)
+    self.assertEqual(st.vgpr[0][3], 6)
+
+  def test_ds_dec_rtn_u32(self):
+    """DS_DEC_RTN_U32: decrement with wrap."""
+    instructions = [
+      v_mov_b32_e32(v[10], 0),
+      s_mov_b32(s[2], 5),
+      v_mov_b32_e32(v[0], s[2]),
+      ds_store_b32(addr=v[10], data0=v[0], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[2], 10),  # limit
+      v_mov_b32_e32(v[1], s[2]),
+      ds_dec_rtn_u32(addr=v[10], data0=v[1], vdst=v[2], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      ds_load_b32(addr=v[10], vdst=v[3], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 5)
+    self.assertEqual(st.vgpr[0][3], 4)
+
+  def test_ds_cmpstore_b32_match(self):
+    """DS_CMPSTORE_B32: conditional store when compare matches."""
+    instructions = [
+      v_mov_b32_e32(v[10], 0),
+      s_mov_b32(s[2], 100),
+      v_mov_b32_e32(v[0], s[2]),
+      ds_store_b32(addr=v[10], data0=v[0], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[2], 200),
+      v_mov_b32_e32(v[1], s[2]),  # new value
+      s_mov_b32(s[2], 100),
+      v_mov_b32_e32(v[2], s[2]),  # compare = 100 (matches)
+      ds_cmpstore_b32(addr=v[10], data0=v[1], data1=v[2], vdst=v[3], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      ds_load_b32(addr=v[10], vdst=v[4], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][4], 200)
+
+  def test_ds_cmpstore_b32_no_match(self):
+    """DS_CMPSTORE_B32: no store when compare doesn't match."""
+    instructions = [
+      v_mov_b32_e32(v[10], 0),
+      s_mov_b32(s[2], 100),
+      v_mov_b32_e32(v[0], s[2]),
+      ds_store_b32(addr=v[10], data0=v[0], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[2], 200),
+      v_mov_b32_e32(v[1], s[2]),  # new value
+      s_mov_b32(s[2], 50),
+      v_mov_b32_e32(v[2], s[2]),  # compare = 50 (doesn't match)
+      ds_cmpstore_b32(addr=v[10], data0=v[1], data1=v[2], vdst=v[3], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      ds_load_b32(addr=v[10], vdst=v[4], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][4], 100)
+
+  def test_ds_max_u32_no_rtn(self):
+    """DS_MAX_U32 (no RTN): atomically store max, no return value."""
+    instructions = [
+      v_mov_b32_e32(v[10], 0),
+      s_mov_b32(s[2], 100),
+      v_mov_b32_e32(v[0], s[2]),
+      ds_store_b32(addr=v[10], data0=v[0], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[2], 200),
+      v_mov_b32_e32(v[1], s[2]),
+      ds_max_u32(addr=v[10], data0=v[1], vdst=v[2], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      ds_load_b32(addr=v[10], vdst=v[3], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][3], 200, "v3 should have max(100, 200) = 200")
+
+  def test_ds_add_u32_no_rtn_preserves_vdst(self):
+    """DS_ADD_U32 (no RTN) should NOT write to vdst."""
+    instructions = [
+      v_mov_b32_e32(v[10], 0),
+      s_mov_b32(s[2], 0xDEADBEEF),
+      v_mov_b32_e32(v[2], s[2]),  # sentinel
+      s_mov_b32(s[2], 100),
+      v_mov_b32_e32(v[0], s[2]),
+      ds_store_b32(addr=v[10], data0=v[0], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[2], 50),
+      v_mov_b32_e32(v[1], s[2]),
+      ds_add_u32(addr=v[10], data0=v[1], vdst=v[2], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      ds_load_b32(addr=v[10], vdst=v[3], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0xDEADBEEF, "v2 should preserve sentinel")
+    self.assertEqual(st.vgpr[0][3], 150, "v3 should have 100 + 50 = 150")
+
+  def test_ds_add_rtn_u32_writes_vdst(self):
+    """DS_ADD_RTN_U32 should write old value to vdst."""
+    instructions = [
+      v_mov_b32_e32(v[10], 0),
+      s_mov_b32(s[2], 0xDEADBEEF),
+      v_mov_b32_e32(v[2], s[2]),  # sentinel
+      s_mov_b32(s[2], 100),
+      v_mov_b32_e32(v[0], s[2]),
+      ds_store_b32(addr=v[10], data0=v[0], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[2], 50),
+      v_mov_b32_e32(v[1], s[2]),
+      ds_add_rtn_u32(addr=v[10], data0=v[1], vdst=v[2], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      ds_load_b32(addr=v[10], vdst=v[3], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 100, "v2 should have old value (100)")
+    self.assertEqual(st.vgpr[0][3], 150, "v3 should have 100 + 50 = 150")
+
+  def test_ds_dec_rtn_u32_wrap(self):
+    """DS_DEC_RTN_U32: decrement wraps when value is 0 or > limit."""
+    instructions = [
+      v_mov_b32_e32(v[10], 0),
+      s_mov_b32(s[2], 0),  # Start at 0
+      v_mov_b32_e32(v[0], s[2]),
+      ds_store_b32(addr=v[10], data0=v[0], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[2], 10),  # limit
+      v_mov_b32_e32(v[1], s[2]),
+      ds_dec_rtn_u32(addr=v[10], data0=v[1], vdst=v[2], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      ds_load_b32(addr=v[10], vdst=v[3], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0, "v2 should have old value (0)")
+    # When mem == 0 or mem > limit, result = limit
+    self.assertEqual(st.vgpr[0][3], 10, "v3 should wrap to limit (10)")
+
+
+class TestDSStorexchg(unittest.TestCase):
+  """Tests for DS_STOREXCHG instructions."""
+
+  def test_ds_storexchg_rtn_b32(self):
+    """DS_STOREXCHG_RTN_B32: exchange value and return old."""
+    instructions = [
+      v_mov_b32_e32(v[10], 0),
+      s_mov_b32(s[0], 0xAAAAAAAA),
+      v_mov_b32_e32(v[0], s[0]),
+      ds_store_b32(addr=v[10], data0=v[0], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[0], 0xBBBBBBBB),
+      v_mov_b32_e32(v[1], s[0]),
+      DS(DSOp.DS_STOREXCHG_RTN_B32, addr=v[10], data0=v[1], vdst=v[2], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      ds_load_b32(addr=v[10], vdst=v[3], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0xAAAAAAAA)
+    self.assertEqual(st.vgpr[0][3], 0xBBBBBBBB)
+
+
+class TestDSRegisterWidth(unittest.TestCase):
+  """Regression tests: DS loads should only write correct number of VGPRs."""
+
+  def test_ds_load_b32_no_overwrite(self):
+    """DS_LOAD_B32 should only write 1 VGPR."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0),
+      s_mov_b32(s[0], 0xDEADBEEF),
+      v_mov_b32_e32(v[1], s[0]),
+      s_mov_b32(s[0], 0x11111111),
+      v_mov_b32_e32(v[2], s[0]),  # sentinel
+      ds_store_b32(addr=v[0], data0=v[1], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      ds_load_b32(addr=v[0], vdst=v[1], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][1], 0xDEADBEEF)
+    self.assertEqual(st.vgpr[0][2], 0x11111111, "v2 should be untouched")
+
+
+class TestDS2AddrStride64(unittest.TestCase):
+  """Tests for DS_*_2ADDR_STRIDE64 (offset * 256 for B32, offset * 512 for B64)."""
+
+  def test_ds_store_load_2addr_stride64_b32(self):
+    """DS_STORE_2ADDR_STRIDE64_B32: stores at ADDR + offset*256."""
+    instructions = [
+      v_mov_b32_e32(v[10], 0),
+      s_mov_b32(s[0], 0xAAAAAAAA),
+      v_mov_b32_e32(v[0], s[0]),
+      s_mov_b32(s[0], 0xBBBBBBBB),
+      v_mov_b32_e32(v[1], s[0]),
+      DS(DSOp.DS_STORE_2ADDR_STRIDE64_B32, addr=v[10], data0=v[0], data1=v[1], vdst=v[0], offset0=1, offset1=2),
+      s_waitcnt(lgkmcnt=0),
+      DS(DSOp.DS_LOAD_2ADDR_STRIDE64_B32, addr=v[10], vdst=v[2:3], offset0=1, offset1=2),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0xAAAAAAAA, "v2 from addr 256")
+    self.assertEqual(st.vgpr[0][3], 0xBBBBBBBB, "v3 from addr 512")
+
+  def test_ds_store_load_2addr_stride64_b64(self):
+    """DS_STORE_2ADDR_STRIDE64_B64: stores at ADDR + offset*512."""
+    instructions = [
+      v_mov_b32_e32(v[10], 0),
+      s_mov_b32(s[0], 0xDEADBEEF),
+      v_mov_b32_e32(v[0], s[0]),
+      s_mov_b32(s[0], 0xCAFEBABE),
+      v_mov_b32_e32(v[1], s[0]),
+      s_mov_b32(s[0], 0x12345678),
+      v_mov_b32_e32(v[2], s[0]),
+      s_mov_b32(s[0], 0x9ABCDEF0),
+      v_mov_b32_e32(v[3], s[0]),
+      DS(DSOp.DS_STORE_2ADDR_STRIDE64_B64, addr=v[10], data0=v[0:1], data1=v[2:3], vdst=v[0], offset0=1, offset1=2),
+      s_waitcnt(lgkmcnt=0),
+      DS(DSOp.DS_LOAD_2ADDR_STRIDE64_B64, addr=v[10], vdst=v[4:7], offset0=1, offset1=2),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][4], 0xDEADBEEF)
+    self.assertEqual(st.vgpr[0][5], 0xCAFEBABE)
+    self.assertEqual(st.vgpr[0][6], 0x12345678)
+    self.assertEqual(st.vgpr[0][7], 0x9ABCDEF0)
+
+  def test_ds_storexchg_2addr_rtn_b32(self):
+    """DS_STOREXCHG_2ADDR_RTN_B32: exchange at two addresses."""
+    instructions = [
+      v_mov_b32_e32(v[10], 0),
+      s_mov_b32(s[0], 0x11111111),
+      v_mov_b32_e32(v[0], s[0]),
+      s_mov_b32(s[0], 0x22222222),
+      v_mov_b32_e32(v[1], s[0]),
+      DS(DSOp.DS_STORE_2ADDR_B32, addr=v[10], data0=v[0], data1=v[1], vdst=v[0], offset0=0, offset1=1),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[0], 0xAAAAAAAA),
+      v_mov_b32_e32(v[2], s[0]),
+      s_mov_b32(s[0], 0xBBBBBBBB),
+      v_mov_b32_e32(v[3], s[0]),
+      DS(DSOp.DS_STOREXCHG_2ADDR_RTN_B32, addr=v[10], data0=v[2], data1=v[3], vdst=v[4:5], offset0=0, offset1=1),
+      s_waitcnt(lgkmcnt=0),
+      DS(DSOp.DS_LOAD_2ADDR_B32, addr=v[10], vdst=v[6:7], offset0=0, offset1=1),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][4], 0x11111111, "old val 0")
+    self.assertEqual(st.vgpr[0][5], 0x22222222, "old val 1")
+    self.assertEqual(st.vgpr[0][6], 0xAAAAAAAA, "new val 0")
+    self.assertEqual(st.vgpr[0][7], 0xBBBBBBBB, "new val 1")
+
+
+  def test_ds_storexchg_rtn_b64(self):
+    """DS_STOREXCHG_RTN_B64: exchange 64-bit value and return old."""
+    instructions = [
+      v_mov_b32_e32(v[10], 0),
+      s_mov_b32(s[0], 0xDEADBEEF),
+      v_mov_b32_e32(v[0], s[0]),   # initial low
+      s_mov_b32(s[0], 0xCAFEBABE),
+      v_mov_b32_e32(v[1], s[0]),   # initial high
+      DS(DSOp.DS_STORE_B64, addr=v[10], data0=v[0:1], vdst=v[0], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[0], 0x12345678),
+      v_mov_b32_e32(v[2], s[0]),   # new low
+      s_mov_b32(s[0], 0x9ABCDEF0),
+      v_mov_b32_e32(v[3], s[0]),   # new high
+      DS(DSOp.DS_STOREXCHG_RTN_B64, addr=v[10], data0=v[2:3], vdst=v[4:5], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      DS(DSOp.DS_LOAD_B64, addr=v[10], vdst=v[6:7], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][4], 0xDEADBEEF, "v4 should have old low dword")
+    self.assertEqual(st.vgpr[0][5], 0xCAFEBABE, "v5 should have old high dword")
+    self.assertEqual(st.vgpr[0][6], 0x12345678, "v6 should have new low dword")
+    self.assertEqual(st.vgpr[0][7], 0x9ABCDEF0, "v7 should have new high dword")
+
+  def test_ds_store_load_2addr_stride64_b64_roundtrip(self):
+    """DS_STORE_2ADDR_STRIDE64_B64 followed by DS_LOAD_2ADDR_STRIDE64_B64 works correctly."""
+    instructions = [
+      v_mov_b32_e32(v[10], 0),
+      s_mov_b32(s[0], 0x11111111),
+      v_mov_b32_e32(v[0], s[0]),
+      s_mov_b32(s[0], 0x22222222),
+      v_mov_b32_e32(v[1], s[0]),
+      DS(DSOp.DS_STORE_2ADDR_STRIDE64_B64, addr=v[10], data0=v[0:1], data1=v[0:1], vdst=v[0], offset0=1, offset1=2),
+      s_waitcnt(lgkmcnt=0),
+      DS(DSOp.DS_LOAD_2ADDR_STRIDE64_B64, addr=v[10], vdst=v[2:5], offset0=1, offset1=2),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0x11111111, "v2 should have val1 low")
+    self.assertEqual(st.vgpr[0][3], 0x22222222, "v3 should have val1 high")
+    self.assertEqual(st.vgpr[0][4], 0x11111111, "v4 should have val2 low")
+    self.assertEqual(st.vgpr[0][5], 0x22222222, "v5 should have val2 high")
+
+  def test_ds_storexchg_2addr_stride64_rtn_b32(self):
+    """DS_STOREXCHG_2ADDR_STRIDE64_RTN_B32: exchange at two addresses (offset*256)."""
+    instructions = [
+      v_mov_b32_e32(v[10], 0),
+      s_mov_b32(s[0], 0x11111111),
+      v_mov_b32_e32(v[0], s[0]),
+      s_mov_b32(s[0], 0x22222222),
+      v_mov_b32_e32(v[1], s[0]),
+      DS(DSOp.DS_STORE_2ADDR_STRIDE64_B32, addr=v[10], data0=v[0], data1=v[1], vdst=v[0], offset0=1, offset1=2),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[0], 0xAAAAAAAA),
+      v_mov_b32_e32(v[2], s[0]),
+      s_mov_b32(s[0], 0xBBBBBBBB),
+      v_mov_b32_e32(v[3], s[0]),
+      DS(DSOp.DS_STOREXCHG_2ADDR_STRIDE64_RTN_B32, addr=v[10], data0=v[2], data1=v[3], vdst=v[4:5], offset0=1, offset1=2),
+      s_waitcnt(lgkmcnt=0),
+      DS(DSOp.DS_LOAD_2ADDR_STRIDE64_B32, addr=v[10], vdst=v[6:7], offset0=1, offset1=2),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][4], 0x11111111, "v4 should have old value")
+    self.assertEqual(st.vgpr[0][5], 0x22222222, "v5 should have old value")
+    self.assertEqual(st.vgpr[0][6], 0xAAAAAAAA, "v6 should have new value")
+    self.assertEqual(st.vgpr[0][7], 0xBBBBBBBB, "v7 should have new value")
+
+  def test_ds_storexchg_2addr_stride64_rtn_b64_returns_old(self):
+    """DS_STOREXCHG_2ADDR_STRIDE64_RTN_B64: returns old values correctly."""
+    instructions = [
+      v_mov_b32_e32(v[10], 0),
+      s_mov_b32(s[0], 0x11111111),
+      v_mov_b32_e32(v[0], s[0]),
+      s_mov_b32(s[0], 0x22222222),
+      v_mov_b32_e32(v[1], s[0]),
+      DS(DSOp.DS_STORE_2ADDR_STRIDE64_B64, addr=v[10], data0=v[0:1], data1=v[0:1], vdst=v[0], offset0=1, offset1=2),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[0], 0xAAAAAAAA),
+      v_mov_b32_e32(v[6], s[0]),
+      s_mov_b32(s[0], 0xBBBBBBBB),
+      v_mov_b32_e32(v[7], s[0]),
+      DS(DSOp.DS_STOREXCHG_2ADDR_STRIDE64_RTN_B64, addr=v[10], data0=v[6:7], data1=v[6:7], vdst=v[8:11], offset0=1, offset1=2),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][8], 0x11111111, "v8 should have old val1 low")
+    self.assertEqual(st.vgpr[0][9], 0x22222222, "v9 should have old val1 high")
+    self.assertEqual(st.vgpr[0][10], 0x11111111, "v10 should have old val2 low")
+    self.assertEqual(st.vgpr[0][11], 0x22222222, "v11 should have old val2 high")
+
+
+class TestAtomicOrdering(unittest.TestCase):
+  """Tests for atomic operation return values and ordering."""
+
+  def test_ds_add_rtn_sequence(self):
+    """DS_ADD_RTN returns correct old values in sequence."""
+    instructions = [
+      v_mov_b32_e32(v[10], 0),
+      v_mov_b32_e32(v[0], 100),
+      DS(DSOp.DS_STORE_B32, addr=v[10], data0=v[0], vdst=v[0], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      v_mov_b32_e32(v[1], 25),
+      DS(DSOp.DS_ADD_RTN_U32, addr=v[10], data0=v[1], vdst=v[2], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      DS(DSOp.DS_ADD_RTN_U32, addr=v[10], data0=v[1], vdst=v[3], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+      DS(DSOp.DS_LOAD_B32, addr=v[10], vdst=v[4], offset0=0),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 100, "First add should return 100")
+    self.assertEqual(st.vgpr[0][3], 125, "Second add should return 125")
+    self.assertEqual(st.vgpr[0][4], 150, "Final value should be 150")
+
+
+class TestDsPermute(unittest.TestCase):
+  """Tests for DS_PERMUTE_B32 and DS_BPERMUTE_B32 instructions."""
+
+  def test_ds_permute_b32_identity(self):
+    """DS_PERMUTE_B32 with identity permutation (lane 0 sends to lane 0)."""
+    # For simplicity, test with single lane
+    instructions = [
+      v_mov_b32_e32(v[0], 0),  # addr = 0 (lane 0)
+      v_mov_b32_e32(v[1], 0xDEADBEEF),  # data
+      ds_permute_b32(v[2], v[0], v[1]),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    # Lane 0 sends to lane 0, so lane 0 gets 0xDEADBEEF
+    self.assertEqual(st.vgpr[0][2], 0xDEADBEEF)
+
+  def test_ds_bpermute_b32_identity(self):
+    """DS_BPERMUTE_B32 with identity permutation (each lane reads from itself)."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0),  # addr = 0 (read from lane 0)
+      v_mov_b32_e32(v[1], 0xCAFEBABE),  # data in lane 0
+      ds_bpermute_b32(v[2], v[0], v[1]),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    # Lane 0 reads from lane 0's v[1]
+    self.assertEqual(st.vgpr[0][2], 0xCAFEBABE)
+
+  def test_ds_permute_b32_broadcast(self):
+    """DS_PERMUTE_B32 broadcast - all lanes send to lane 0."""
+    # With 4 lanes, all sending to lane 0, highest lane wins
+    instructions = [
+      v_mov_b32_e32(v[0], 0),  # All lanes send to addr 0 (lane 0)
+      v_mov_b32_e32(v[1], 0x11111111),  # All lanes send same data
+      ds_permute_b32(v[2], v[0], v[1]),
+      s_waitcnt(lgkmcnt=0),
+    ]
+    st = run_program(instructions, n_lanes=4)
+    # Lane 0 receives data (highest numbered active lane wins)
+    self.assertEqual(st.vgpr[0][2], 0x11111111)
+
+
+if __name__ == '__main__':
+  unittest.main()

extra/assembly/amd/test/hw/test_flat.py

@@ -0,0 +1,363 @@
+"""Tests for FLAT instructions - flat memory operations.
+
+Includes: flat_load_*, flat_store_*, flat_atomic_*
+"""
+import unittest
+from extra.assembly.amd.test.hw.helpers import *
+
+class TestFlatAtomic(unittest.TestCase):
+  """Tests for FLAT atomic instructions."""
+
+  def _make_test(self, setup_instrs, atomic_instr, check_fn, test_offset=2000):
+    """Helper to create atomic test instructions."""
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      v_mov_b32_e32(v[0], s[2]),
+      v_mov_b32_e32(v[1], s[3]),
+    ] + setup_instrs + [atomic_instr, s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], 0),
+      v_mov_b32_e32(v[1], 0),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    check_fn(st)
+
+  def test_flat_atomic_add_u32(self):
+    """FLAT_ATOMIC_ADD_U32 adds to memory and returns old value."""
+    TEST_OFFSET = 2000
+    setup = [
+      s_mov_b32(s[0], 100),
+      v_mov_b32_e32(v[2], s[0]),
+      global_store_b32(addr=v[0:1], data=v[2], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      s_mov_b32(s[0], 50),
+      v_mov_b32_e32(v[3], s[0]),
+    ]
+    atomic = FLAT(FLATOp.FLAT_ATOMIC_ADD_U32, addr=v[0:1], data=v[3], vdst=v[4], saddr=SrcEnum.NULL, offset=TEST_OFFSET, glc=1)
+    def check(st):
+      self.assertEqual(st.vgpr[0][4], 100)
+    self._make_test(setup, atomic, check, TEST_OFFSET)
+
+  def test_flat_atomic_swap_b32(self):
+    """FLAT_ATOMIC_SWAP_B32 swaps memory value and returns old value."""
+    TEST_OFFSET = 2000
+    setup = [
+      s_mov_b32(s[0], 0xAAAAAAAA),
+      v_mov_b32_e32(v[2], s[0]),
+      global_store_b32(addr=v[0:1], data=v[2], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      s_mov_b32(s[0], 0xBBBBBBBB),
+      v_mov_b32_e32(v[3], s[0]),
+    ]
+    atomic = FLAT(FLATOp.FLAT_ATOMIC_SWAP_B32, addr=v[0:1], data=v[3], vdst=v[4], saddr=SrcEnum.NULL, offset=TEST_OFFSET, glc=1)
+    def check(st):
+      self.assertEqual(st.vgpr[0][4], 0xAAAAAAAA)
+    self._make_test(setup, atomic, check, TEST_OFFSET)
+
+  def test_flat_atomic_and_b32(self):
+    """FLAT_ATOMIC_AND_B32 ANDs with memory and returns old value."""
+    TEST_OFFSET = 2000
+    setup = [
+      s_mov_b32(s[0], 0xFF00FF00),
+      v_mov_b32_e32(v[2], s[0]),
+      global_store_b32(addr=v[0:1], data=v[2], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      s_mov_b32(s[0], 0xFFFF0000),
+      v_mov_b32_e32(v[3], s[0]),
+    ]
+    atomic = FLAT(FLATOp.FLAT_ATOMIC_AND_B32, addr=v[0:1], data=v[3], vdst=v[4], saddr=SrcEnum.NULL, offset=TEST_OFFSET, glc=1)
+    def check(st):
+      self.assertEqual(st.vgpr[0][4], 0xFF00FF00)
+    self._make_test(setup, atomic, check, TEST_OFFSET)
+
+  def test_flat_atomic_or_b32(self):
+    """FLAT_ATOMIC_OR_B32 ORs with memory and returns old value."""
+    TEST_OFFSET = 2000
+    setup = [
+      s_mov_b32(s[0], 0x00FF0000),
+      v_mov_b32_e32(v[2], s[0]),
+      global_store_b32(addr=v[0:1], data=v[2], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      s_mov_b32(s[0], 0x0000FF00),
+      v_mov_b32_e32(v[3], s[0]),
+    ]
+    atomic = FLAT(FLATOp.FLAT_ATOMIC_OR_B32, addr=v[0:1], data=v[3], vdst=v[4], saddr=SrcEnum.NULL, offset=TEST_OFFSET, glc=1)
+    def check(st):
+      self.assertEqual(st.vgpr[0][4], 0x00FF0000)
+    self._make_test(setup, atomic, check, TEST_OFFSET)
+
+  def test_flat_atomic_inc_u32(self):
+    """FLAT_ATOMIC_INC_U32 increments and returns old value."""
+    TEST_OFFSET = 2000
+    setup = [
+      s_mov_b32(s[0], 10),
+      v_mov_b32_e32(v[2], s[0]),
+      global_store_b32(addr=v[0:1], data=v[2], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      s_mov_b32(s[0], 100),  # threshold
+      v_mov_b32_e32(v[3], s[0]),
+    ]
+    atomic = FLAT(FLATOp.FLAT_ATOMIC_INC_U32, addr=v[0:1], data=v[3], vdst=v[4], saddr=SrcEnum.NULL, offset=TEST_OFFSET, glc=1)
+    def check(st):
+      self.assertEqual(st.vgpr[0][4], 10)
+    self._make_test(setup, atomic, check, TEST_OFFSET)
+
+  def test_flat_atomic_dec_u32(self):
+    """FLAT_ATOMIC_DEC_U32 decrements and returns old value."""
+    TEST_OFFSET = 2000
+    setup = [
+      s_mov_b32(s[0], 10),
+      v_mov_b32_e32(v[2], s[0]),
+      global_store_b32(addr=v[0:1], data=v[2], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      s_mov_b32(s[0], 100),
+      v_mov_b32_e32(v[3], s[0]),
+    ]
+    atomic = FLAT(FLATOp.FLAT_ATOMIC_DEC_U32, addr=v[0:1], data=v[3], vdst=v[4], saddr=SrcEnum.NULL, offset=TEST_OFFSET, glc=1)
+    def check(st):
+      self.assertEqual(st.vgpr[0][4], 10)
+    self._make_test(setup, atomic, check, TEST_OFFSET)
+
+  def test_flat_atomic_sub_u32(self):
+    """FLAT_ATOMIC_SUB_U32 subtracts from memory and returns old value."""
+    TEST_OFFSET = 2000
+    setup = [
+      s_mov_b32(s[0], 100),
+      v_mov_b32_e32(v[2], s[0]),
+      global_store_b32(addr=v[0:1], data=v[2], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      s_mov_b32(s[0], 30),
+      v_mov_b32_e32(v[3], s[0]),  # sub 30
+    ]
+    atomic = FLAT(FLATOp.FLAT_ATOMIC_SUB_U32, addr=v[0:1], data=v[3], vdst=v[4], saddr=SrcEnum.NULL, offset=TEST_OFFSET, glc=1)
+    def check(st):
+      self.assertEqual(st.vgpr[0][4], 100, "v4 should have old value (100)")
+    self._make_test(setup, atomic, check, TEST_OFFSET)
+
+  def test_flat_atomic_xor_b32(self):
+    """FLAT_ATOMIC_XOR_B32 XORs with memory and returns old value."""
+    TEST_OFFSET = 2000
+    setup = [
+      s_mov_b32(s[0], 0xAAAAAAAA),
+      v_mov_b32_e32(v[2], s[0]),
+      global_store_b32(addr=v[0:1], data=v[2], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      s_mov_b32(s[0], 0xFFFFFFFF),
+      v_mov_b32_e32(v[3], s[0]),  # XOR mask
+    ]
+    atomic = FLAT(FLATOp.FLAT_ATOMIC_XOR_B32, addr=v[0:1], data=v[3], vdst=v[4], saddr=SrcEnum.NULL, offset=TEST_OFFSET, glc=1)
+    def check(st):
+      self.assertEqual(st.vgpr[0][4], 0xAAAAAAAA, "v4 should have old value")
+    self._make_test(setup, atomic, check, TEST_OFFSET)
+
+  def test_flat_atomic_min_u32(self):
+    """FLAT_ATOMIC_MIN_U32 stores min and returns old value."""
+    TEST_OFFSET = 2000
+    setup = [
+      s_mov_b32(s[0], 100),
+      v_mov_b32_e32(v[2], s[0]),
+      global_store_b32(addr=v[0:1], data=v[2], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      s_mov_b32(s[0], 50),
+      v_mov_b32_e32(v[3], s[0]),  # compare value (smaller)
+    ]
+    atomic = FLAT(FLATOp.FLAT_ATOMIC_MIN_U32, addr=v[0:1], data=v[3], vdst=v[4], saddr=SrcEnum.NULL, offset=TEST_OFFSET, glc=1)
+    def check(st):
+      self.assertEqual(st.vgpr[0][4], 100, "v4 should have old value (100)")
+    self._make_test(setup, atomic, check, TEST_OFFSET)
+
+  def test_flat_atomic_max_u32(self):
+    """FLAT_ATOMIC_MAX_U32 stores max and returns old value."""
+    TEST_OFFSET = 2000
+    setup = [
+      s_mov_b32(s[0], 50),
+      v_mov_b32_e32(v[2], s[0]),
+      global_store_b32(addr=v[0:1], data=v[2], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      s_mov_b32(s[0], 100),
+      v_mov_b32_e32(v[3], s[0]),  # compare value (larger)
+    ]
+    atomic = FLAT(FLATOp.FLAT_ATOMIC_MAX_U32, addr=v[0:1], data=v[3], vdst=v[4], saddr=SrcEnum.NULL, offset=TEST_OFFSET, glc=1)
+    def check(st):
+      self.assertEqual(st.vgpr[0][4], 50, "v4 should have old value (50)")
+    self._make_test(setup, atomic, check, TEST_OFFSET)
+
+  def test_flat_atomic_inc_u64_returns_old_value(self):
+    """FLAT_ATOMIC_INC_U64 should return full 64-bit old value."""
+    TEST_OFFSET = 2000
+    setup = [
+      # Store initial 64-bit value: 0xCAFEBABE_DEADBEEF
+      s_mov_b32(s[0], 0xDEADBEEF),
+      v_mov_b32_e32(v[2], s[0]),
+      s_mov_b32(s[0], 0xCAFEBABE),
+      v_mov_b32_e32(v[3], s[0]),
+      global_store_b64(addr=v[0:1], data=v[2:3], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      # Threshold: 0xFFFFFFFF_FFFFFFFF
+      s_mov_b32(s[0], 0xFFFFFFFF),
+      v_mov_b32_e32(v[4], s[0]),
+      v_mov_b32_e32(v[5], s[0]),
+    ]
+    atomic = FLAT(FLATOp.FLAT_ATOMIC_INC_U64, addr=v[0:1], data=v[4:5], vdst=v[6:7], saddr=SrcEnum.NULL, offset=TEST_OFFSET, glc=1)
+    def check(st):
+      self.assertEqual(st.vgpr[0][6], 0xDEADBEEF, "v6 should have old value low dword")
+      self.assertEqual(st.vgpr[0][7], 0xCAFEBABE, "v7 should have old value high dword")
+    self._make_test(setup, atomic, check, TEST_OFFSET)
+
+  def test_flat_atomic_add_u64(self):
+    """FLAT_ATOMIC_ADD_U64 adds 64-bit value and returns old value."""
+    TEST_OFFSET = 2000
+    setup = [
+      s_mov_b32(s[0], 0x11111111),
+      v_mov_b32_e32(v[2], s[0]),
+      s_mov_b32(s[0], 0x22222222),
+      v_mov_b32_e32(v[3], s[0]),
+      global_store_b64(addr=v[0:1], data=v[2:3], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      s_mov_b32(s[0], 0x00000001),  # add 1
+      v_mov_b32_e32(v[4], s[0]),
+      s_mov_b32(s[0], 0x00000000),
+      v_mov_b32_e32(v[5], s[0]),
+    ]
+    atomic = FLAT(FLATOp.FLAT_ATOMIC_ADD_U64, addr=v[0:1], data=v[4:5], vdst=v[6:7], saddr=SrcEnum.NULL, offset=TEST_OFFSET, glc=1)
+    def check(st):
+      self.assertEqual(st.vgpr[0][6], 0x11111111, "v6 should have old value low")
+      self.assertEqual(st.vgpr[0][7], 0x22222222, "v7 should have old value high")
+    self._make_test(setup, atomic, check, TEST_OFFSET)
+
+  def test_flat_atomic_swap_b64(self):
+    """FLAT_ATOMIC_SWAP_B64 swaps 64-bit value and returns old value."""
+    TEST_OFFSET = 2000
+    setup = [
+      s_mov_b32(s[0], 0xAAAAAAAA),
+      v_mov_b32_e32(v[2], s[0]),
+      s_mov_b32(s[0], 0xBBBBBBBB),
+      v_mov_b32_e32(v[3], s[0]),
+      global_store_b64(addr=v[0:1], data=v[2:3], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      s_mov_b32(s[0], 0xCCCCCCCC),
+      v_mov_b32_e32(v[4], s[0]),
+      s_mov_b32(s[0], 0xDDDDDDDD),
+      v_mov_b32_e32(v[5], s[0]),
+    ]
+    atomic = FLAT(FLATOp.FLAT_ATOMIC_SWAP_B64, addr=v[0:1], data=v[4:5], vdst=v[6:7], saddr=SrcEnum.NULL, offset=TEST_OFFSET, glc=1)
+    def check(st):
+      self.assertEqual(st.vgpr[0][6], 0xAAAAAAAA, "v6 should have old value low")
+      self.assertEqual(st.vgpr[0][7], 0xBBBBBBBB, "v7 should have old value high")
+    self._make_test(setup, atomic, check, TEST_OFFSET)
+
+
+class TestFlatLoad(unittest.TestCase):
+  """Tests for FLAT load instructions."""
+
+  def test_flat_load_b32(self):
+    """FLAT_LOAD_B32 loads 32-bit value correctly."""
+    TEST_OFFSET = 2000
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      v_mov_b32_e32(v[0], s[2]),
+      v_mov_b32_e32(v[1], s[3]),
+      s_mov_b32(s[0], 0xDEADBEEF),
+      v_mov_b32_e32(v[2], s[0]),
+      global_store_b32(addr=v[0:1], data=v[2], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      FLAT(FLATOp.FLAT_LOAD_B32, addr=v[0:1], vdst=v[4], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], 0),
+      v_mov_b32_e32(v[1], 0),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][4], 0xDEADBEEF)
+
+  def test_flat_load_b64(self):
+    """FLAT_LOAD_B64 loads 64-bit value correctly."""
+    TEST_OFFSET = 2000
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      v_mov_b32_e32(v[0], s[2]),
+      v_mov_b32_e32(v[1], s[3]),
+      s_mov_b32(s[0], 0xDEADBEEF),
+      v_mov_b32_e32(v[2], s[0]),
+      s_mov_b32(s[0], 0xCAFEBABE),
+      v_mov_b32_e32(v[3], s[0]),
+      global_store_b64(addr=v[0:1], data=v[2:3], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      FLAT(FLATOp.FLAT_LOAD_B64, addr=v[0:1], vdst=v[4:5], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], 0),
+      v_mov_b32_e32(v[1], 0),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][4], 0xDEADBEEF)
+    self.assertEqual(st.vgpr[0][5], 0xCAFEBABE)
+
+  def test_flat_load_b96(self):
+    """FLAT_LOAD_B96 loads 96-bit (3 dword) value correctly."""
+    TEST_OFFSET = 2000
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      v_mov_b32_e32(v[0], s[2]),
+      v_mov_b32_e32(v[1], s[3]),
+      s_mov_b32(s[0], 0x11111111),
+      v_mov_b32_e32(v[2], s[0]),
+      s_mov_b32(s[0], 0x22222222),
+      v_mov_b32_e32(v[3], s[0]),
+      s_mov_b32(s[0], 0x33333333),
+      v_mov_b32_e32(v[4], s[0]),
+      global_store_b96(addr=v[0:1], data=v[2:4], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      FLAT(FLATOp.FLAT_LOAD_B96, addr=v[0:1], vdst=v[5:7], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], 0),
+      v_mov_b32_e32(v[1], 0),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][5], 0x11111111)
+    self.assertEqual(st.vgpr[0][6], 0x22222222)
+    self.assertEqual(st.vgpr[0][7], 0x33333333)
+
+  def test_flat_load_b128(self):
+    """FLAT_LOAD_B128 loads 128-bit value correctly."""
+    TEST_OFFSET = 2000
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      v_mov_b32_e32(v[0], s[2]),
+      v_mov_b32_e32(v[1], s[3]),
+      s_mov_b32(s[0], 0x11111111),
+      v_mov_b32_e32(v[2], s[0]),
+      s_mov_b32(s[0], 0x22222222),
+      v_mov_b32_e32(v[3], s[0]),
+      s_mov_b32(s[0], 0x33333333),
+      v_mov_b32_e32(v[4], s[0]),
+      s_mov_b32(s[0], 0x44444444),
+      v_mov_b32_e32(v[5], s[0]),
+      global_store_b128(addr=v[0:1], data=v[2:5], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      FLAT(FLATOp.FLAT_LOAD_B128, addr=v[0:1], vdst=v[6:9], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], 0),
+      v_mov_b32_e32(v[1], 0),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][6], 0x11111111)
+    self.assertEqual(st.vgpr[0][7], 0x22222222)
+    self.assertEqual(st.vgpr[0][8], 0x33333333)
+    self.assertEqual(st.vgpr[0][9], 0x44444444)
+
+
+if __name__ == '__main__':
+  unittest.main()

extra/assembly/amd/test/hw/test_global.py

@@ -0,0 +1,679 @@
+"""Tests for GLOBAL instructions - global memory operations.
+
+Includes: global_load_*, global_store_*, global_atomic_*, global_load_d16_*
+"""
+import unittest
+from extra.assembly.amd.test.hw.helpers import *
+
+class TestGlobalAtomic(unittest.TestCase):
+  """Tests for GLOBAL atomic instructions."""
+
+  def _make_test(self, setup_instrs, atomic_instr, check_fn, test_offset=2000):
+    """Helper to create atomic test instructions."""
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      v_mov_b32_e32(v[0], s[2]),
+      v_mov_b32_e32(v[1], s[3]),
+    ] + setup_instrs + [atomic_instr, s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], 0),
+      v_mov_b32_e32(v[1], 0),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    check_fn(st)
+
+  def test_global_atomic_add_u32(self):
+    """GLOBAL_ATOMIC_ADD_U32 adds to memory and returns old value."""
+    TEST_OFFSET = 2000
+    setup = [
+      s_mov_b32(s[0], 100),
+      v_mov_b32_e32(v[2], s[0]),
+      global_store_b32(addr=v[0:1], data=v[2], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      s_mov_b32(s[0], 50),
+      v_mov_b32_e32(v[3], s[0]),
+    ]
+    atomic = GLOBAL(GLOBALOp.GLOBAL_ATOMIC_ADD_U32, addr=v[0:1], data=v[3], vdst=v[4], saddr=SrcEnum.NULL, offset=TEST_OFFSET, glc=1)
+    def check(st):
+      self.assertEqual(st.vgpr[0][4], 100)
+    self._make_test(setup, atomic, check, TEST_OFFSET)
+
+  def test_global_atomic_add_u64(self):
+    """GLOBAL_ATOMIC_ADD_U64 adds 64-bit value and returns old value."""
+    TEST_OFFSET = 2000
+    setup = [
+      s_mov_b32(s[0], 0xFFFFFFFF),
+      v_mov_b32_e32(v[2], s[0]),
+      s_mov_b32(s[0], 0x00000000),
+      v_mov_b32_e32(v[3], s[0]),
+      global_store_b64(addr=v[0:1], data=v[2:3], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      s_mov_b32(s[0], 0x00000001),
+      v_mov_b32_e32(v[4], s[0]),
+      s_mov_b32(s[0], 0x00000000),
+      v_mov_b32_e32(v[5], s[0]),
+    ]
+    atomic = GLOBAL(GLOBALOp.GLOBAL_ATOMIC_ADD_U64, addr=v[0:1], data=v[4:5], vdst=v[6:7], saddr=SrcEnum.NULL, offset=TEST_OFFSET, glc=1)
+    def check(st):
+      self.assertEqual(st.vgpr[0][6], 0xFFFFFFFF)
+      self.assertEqual(st.vgpr[0][7], 0x00000000)
+    self._make_test(setup, atomic, check, TEST_OFFSET)
+
+
+class TestGlobalLoad(unittest.TestCase):
+  """Tests for GLOBAL load instructions."""
+
+  def test_global_load_b96(self):
+    """GLOBAL_LOAD_B96 loads 96-bit value correctly."""
+    TEST_OFFSET = 2000
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      v_mov_b32_e32(v[0], s[2]),
+      v_mov_b32_e32(v[1], s[3]),
+      s_mov_b32(s[0], 0xAAAAAAAA),
+      v_mov_b32_e32(v[2], s[0]),
+      s_mov_b32(s[0], 0xBBBBBBBB),
+      v_mov_b32_e32(v[3], s[0]),
+      s_mov_b32(s[0], 0xCCCCCCCC),
+      v_mov_b32_e32(v[4], s[0]),
+      global_store_b96(addr=v[0:1], data=v[2:4], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      GLOBAL(GLOBALOp.GLOBAL_LOAD_B96, addr=v[0:1], vdst=v[5:7], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], 0),
+      v_mov_b32_e32(v[1], 0),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][5], 0xAAAAAAAA)
+    self.assertEqual(st.vgpr[0][6], 0xBBBBBBBB)
+    self.assertEqual(st.vgpr[0][7], 0xCCCCCCCC)
+
+  def test_global_load_b128(self):
+    """GLOBAL_LOAD_B128 loads 128-bit value correctly."""
+    TEST_OFFSET = 2000
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      v_mov_b32_e32(v[0], s[2]),
+      v_mov_b32_e32(v[1], s[3]),
+      s_mov_b32(s[0], 0xDEADBEEF),
+      v_mov_b32_e32(v[2], s[0]),
+      s_mov_b32(s[0], 0xCAFEBABE),
+      v_mov_b32_e32(v[3], s[0]),
+      s_mov_b32(s[0], 0x12345678),
+      v_mov_b32_e32(v[4], s[0]),
+      s_mov_b32(s[0], 0x9ABCDEF0),
+      v_mov_b32_e32(v[5], s[0]),
+      global_store_b128(addr=v[0:1], data=v[2:5], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      GLOBAL(GLOBALOp.GLOBAL_LOAD_B128, addr=v[0:1], vdst=v[6:9], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], 0),
+      v_mov_b32_e32(v[1], 0),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][6], 0xDEADBEEF)
+    self.assertEqual(st.vgpr[0][7], 0xCAFEBABE)
+    self.assertEqual(st.vgpr[0][8], 0x12345678)
+    self.assertEqual(st.vgpr[0][9], 0x9ABCDEF0)
+
+
+class TestGlobalStore(unittest.TestCase):
+  """Tests for GLOBAL store instructions."""
+
+  def test_global_store_b8_basic(self):
+    """GLOBAL_STORE_B8 stores a single byte from VDATA[7:0]."""
+    TEST_OFFSET = 256
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      # First store 0xDEADBEEF to memory
+      s_mov_b32(s[4], 0xDEADBEEF),
+      v_mov_b32_e32(v[2], s[4]),
+      v_mov_b32_e32(v[0], 0),
+      global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      # Now store single byte 0x42 to same address (should only change byte 0)
+      v_mov_b32_e32(v[2], 0x42),
+      global_store_b8(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      # Read back and check
+      GLOBAL(GLOBALOp.GLOBAL_LOAD_B32, addr=v[0], vdst=v[3], data=v[3], saddr=s[2:3], offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], v[3]),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    # Only byte 0 should change from 0xEF to 0x42
+    self.assertEqual(st.vgpr[0][0], 0xDEADBE42, "Only byte 0 should be modified")
+
+  def test_global_store_b8_byte1(self):
+    """GLOBAL_STORE_B8 at offset+1 stores to byte 1."""
+    TEST_OFFSET = 256
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[4], 0xDEADBEEF),
+      v_mov_b32_e32(v[2], s[4]),
+      v_mov_b32_e32(v[0], 0),
+      global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[2], 0x42),
+      global_store_b8(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET+1),
+      s_waitcnt(vmcnt=0),
+      GLOBAL(GLOBALOp.GLOBAL_LOAD_B32, addr=v[0], vdst=v[3], data=v[3], saddr=s[2:3], offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], v[3]),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 0xDEAD42EF, "Only byte 1 should be modified")
+
+  def test_global_store_b16_basic(self):
+    """GLOBAL_STORE_B16 stores a 16-bit value from VDATA[15:0]."""
+    TEST_OFFSET = 256
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[4], 0xDEADBEEF),
+      v_mov_b32_e32(v[2], s[4]),
+      v_mov_b32_e32(v[0], 0),
+      global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      s_mov_b32(s[4], 0xCAFE),
+      v_mov_b32_e32(v[2], s[4]),
+      global_store_b16(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      GLOBAL(GLOBALOp.GLOBAL_LOAD_B32, addr=v[0], vdst=v[3], data=v[3], saddr=s[2:3], offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], v[3]),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 0xDEADCAFE, "Only lower 16 bits should be modified")
+
+  def test_global_store_b16_high_half(self):
+    """GLOBAL_STORE_B16 at offset+2 stores to high 16 bits."""
+    TEST_OFFSET = 256
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[4], 0xDEADBEEF),
+      v_mov_b32_e32(v[2], s[4]),
+      v_mov_b32_e32(v[0], 0),
+      global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      s_mov_b32(s[4], 0xCAFE),
+      v_mov_b32_e32(v[2], s[4]),
+      global_store_b16(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET+2),
+      s_waitcnt(vmcnt=0),
+      GLOBAL(GLOBALOp.GLOBAL_LOAD_B32, addr=v[0], vdst=v[3], data=v[3], saddr=s[2:3], offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], v[3]),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 0xCAFEBEEF, "Only upper 16 bits should be modified")
+
+  def test_global_store_b16_byte_offset_1(self):
+    """GLOBAL_STORE_B16 at byte offset 1 stores bytes 1-2 within the same word."""
+    TEST_OFFSET = 256
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[4], 0xDDCCBBAA),
+      v_mov_b32_e32(v[2], s[4]),
+      v_mov_b32_e32(v[0], 0),
+      global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      # Store 0xBEEF at byte offset 1 (bytes 1-2)
+      s_mov_b32(s[4], 0xBEEF),
+      v_mov_b32_e32(v[2], s[4]),
+      global_store_b16(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET+1),
+      s_waitcnt(vmcnt=0),
+      GLOBAL(GLOBALOp.GLOBAL_LOAD_B32, addr=v[0], vdst=v[3], data=v[3], saddr=s[2:3], offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], v[3]),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    # Bytes 1-2 should be 0xBEEF (0xEF at byte 1, 0xBE at byte 2)
+    # Original: 0xDDCCBBAA -> bytes [AA, BB, CC, DD]
+    # After:    0xDDBEEFAA -> bytes [AA, EF, BE, DD]
+    self.assertEqual(st.vgpr[0][0], 0xDDBEEFAA, "Bytes 1-2 should be 0xBEEF")
+
+  def test_global_store_b16_cross_word_boundary(self):
+    """GLOBAL_STORE_B16 at byte offset 3 crosses word boundary (byte 3 of word N, byte 0 of word N+1)."""
+    TEST_OFFSET = 256
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      # Initialize two consecutive words
+      s_mov_b32(s[4], 0xDDCCBBAA),
+      v_mov_b32_e32(v[2], s[4]),
+      v_mov_b32_e32(v[0], 0),
+      global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET),
+      s_mov_b32(s[4], 0x44332211),
+      v_mov_b32_e32(v[2], s[4]),
+      global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET+4),
+      s_waitcnt(vmcnt=0),
+      # Store 0xBEEF at byte offset 3 (crosses word boundary)
+      # Low byte (0xEF) goes to byte 3 of first word
+      # High byte (0xBE) goes to byte 0 of second word
+      s_mov_b32(s[4], 0xBEEF),
+      v_mov_b32_e32(v[2], s[4]),
+      global_store_b16(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET+3),
+      s_waitcnt(vmcnt=0),
+      # Load back both words
+      GLOBAL(GLOBALOp.GLOBAL_LOAD_B32, addr=v[0], vdst=v[3], data=v[3], saddr=s[2:3], offset=TEST_OFFSET),
+      GLOBAL(GLOBALOp.GLOBAL_LOAD_B32, addr=v[0], vdst=v[4], data=v[4], saddr=s[2:3], offset=TEST_OFFSET+4),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], v[3]),
+      v_mov_b32_e32(v[1], v[4]),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    # First word: 0xDDCCBBAA -> 0xEFCCBBAA (byte 3 becomes 0xEF)
+    # Second word: 0x44332211 -> 0x443322BE (byte 0 becomes 0xBE)
+    self.assertEqual(st.vgpr[0][0], 0xEFCCBBAA, "Byte 3 of first word should be 0xEF")
+    self.assertEqual(st.vgpr[0][1], 0x443322BE, "Byte 0 of second word should be 0xBE")
+
+  def test_global_store_b64_basic(self):
+    """GLOBAL_STORE_B64 stores 8 bytes from v[n:n+1] to memory."""
+    TEST_OFFSET = 256
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[4], 0xDEADBEEF),
+      s_mov_b32(s[5], 0xCAFEBABE),
+      v_mov_b32_e32(v[2], s[4]),
+      v_mov_b32_e32(v[3], s[5]),
+      v_mov_b32_e32(v[0], 0),
+      global_store_b64(addr=v[0], data=v[2:3], saddr=s[2:3], offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      GLOBAL(GLOBALOp.GLOBAL_LOAD_B64, addr=v[0], vdst=v[4:5], data=v[4:5], saddr=s[2:3], offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], v[4]),
+      v_mov_b32_e32(v[1], v[5]),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 0xDEADBEEF)
+    self.assertEqual(st.vgpr[0][1], 0xCAFEBABE)
+
+
+class TestD16HiLoads(unittest.TestCase):
+  """Tests for D16_HI load instructions that load into high 16 bits."""
+
+  def test_global_load_d16_hi_b16_preserves_low_bits(self):
+    """GLOBAL_LOAD_D16_HI_B16 must preserve low 16 bits of destination."""
+    TEST_OFFSET = 256
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      v_mov_b32_e32(v[0], s[2]),
+      v_mov_b32_e32(v[1], s[3]),
+      s_mov_b32(s[4], 0xCAFE),
+      v_mov_b32_e32(v[2], s[4]),
+      global_store_b16(addr=v[0:1], data=v[2], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      s_mov_b32(s[4], 0x0000BEEF),
+      v_mov_b32_e32(v[3], s[4]),
+      GLOBAL(GLOBALOp.GLOBAL_LOAD_D16_HI_B16, addr=v[0:1], vdst=v[3], data=v[3], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], v[3]),
+      v_mov_b32_e32(v[1], 0),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][0]
+    self.assertEqual(result, 0xCAFEBEEF, f"Expected 0xCAFEBEEF, got 0x{result:08x}")
+
+  def test_global_load_d16_hi_b16_data_differs_from_vdst(self):
+    """GLOBAL_LOAD_D16_HI_B16 where data field differs from vdst."""
+    TEST_OFFSET = 256
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[4], 0xCAFE),
+      v_mov_b32_e32(v[2], s[4]),
+      v_mov_b32_e32(v[3], 0),
+      global_store_b16(addr=v[3], data=v[2], saddr=s[2:3], offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      s_mov_b32(s[4], 0x0000DEAD),
+      v_mov_b32_e32(v[0], s[4]),  # data field - should NOT affect result
+      v_mov_b32_e32(v[1], 0),     # vdst - low bits should be preserved
+      GLOBAL(GLOBALOp.GLOBAL_LOAD_D16_HI_B16, addr=v[1], vdst=v[1], data=v[0], saddr=s[2:3], offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], v[1]),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][0]
+    self.assertEqual(result, 0xCAFE0000, f"Expected 0xCAFE0000, got 0x{result:08x}")
+
+  def test_global_load_d16_hi_u8_data_differs_from_vdst(self):
+    """GLOBAL_LOAD_D16_HI_U8 where data field differs from vdst."""
+    TEST_OFFSET = 256
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[4], 0xAB),
+      v_mov_b32_e32(v[2], s[4]),
+      v_mov_b32_e32(v[3], 0),
+      global_store_b8(addr=v[3], data=v[2], saddr=s[2:3], offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      s_mov_b32(s[4], 0x0000DEAD),
+      v_mov_b32_e32(v[4], s[4]),  # data field
+      s_mov_b32(s[4], 0x0000BEEF),
+      v_mov_b32_e32(v[5], s[4]),  # vdst
+      v_mov_b32_e32(v[3], 0),
+      GLOBAL(GLOBALOp.GLOBAL_LOAD_D16_HI_U8, addr=v[3], vdst=v[5], data=v[4], saddr=s[2:3], offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], v[5]),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][0]
+    self.assertEqual(result, 0x00ABBEEF, f"Expected 0x00ABBEEF, got 0x{result:08x}")
+
+  def test_global_load_d16_hi_b16_same_addr_and_dst_zero_addr(self):
+    """GLOBAL_LOAD_D16_HI_B16 with same register for addr and vdst, addr value=0."""
+    TEST_OFFSET = 256
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[4], 0xCAFE),
+      v_mov_b32_e32(v[2], s[4]),
+      v_mov_b32_e32(v[3], 0),
+      global_store_b16(addr=v[3], data=v[2], saddr=s[2:3], offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[1], 0),
+      GLOBAL(GLOBALOp.GLOBAL_LOAD_D16_HI_B16, addr=v[1], vdst=v[1], data=v[1], saddr=s[2:3], offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], v[1]),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][0]
+    self.assertEqual(result, 0xCAFE0000, f"Expected 0xCAFE0000, got 0x{result:08x}")
+
+  def test_global_load_d16_hi_b16_tril_exact_pattern(self):
+    """Exact pattern from tril() failure: data=v0 differs from vdst=v1."""
+    TEST_OFFSET = 256
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[4], 0x01010101),
+      v_mov_b32_e32(v[10], s[4]),
+      v_mov_b32_e32(v[3], 0),
+      global_store_b32(addr=v[3], data=v[10], saddr=s[2:3], offset=TEST_OFFSET),
+      global_store_b32(addr=v[3], data=v[10], saddr=s[2:3], offset=TEST_OFFSET+4),
+      s_waitcnt(vmcnt=0),
+      # Set v[0] to 0x0101 (simulating prior u16 load result)
+      s_mov_b32(s[4], 0x0101),
+      v_mov_b32_e32(v[0], s[4]),
+      # Set v[1] to 0
+      v_mov_b32_e32(v[1], 0),
+      # Load using v[1] as addr AND vdst, but v[0] as data
+      GLOBAL(GLOBALOp.GLOBAL_LOAD_D16_HI_B16, addr=v[1], vdst=v[1], data=v[0], saddr=s[2:3], offset=TEST_OFFSET+6),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], v[1]),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][0]
+    # Expected: hi=0x0101 (loaded), lo=0x0000 (from v1) -> 0x01010000
+    self.assertEqual(result, 0x01010000, f"Expected 0x01010000, got 0x{result:08x}")
+
+  def test_global_load_d16_hi_i8_data_differs_from_vdst(self):
+    """GLOBAL_LOAD_D16_HI_I8 where data field differs from vdst."""
+    TEST_OFFSET = 256
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[4], 0x80),  # negative signed byte = -128
+      v_mov_b32_e32(v[2], s[4]),
+      v_mov_b32_e32(v[3], 0),
+      global_store_b8(addr=v[3], data=v[2], saddr=s[2:3], offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      s_mov_b32(s[4], 0x0000DEAD),
+      v_mov_b32_e32(v[4], s[4]),  # data field
+      s_mov_b32(s[4], 0x0000BEEF),
+      v_mov_b32_e32(v[5], s[4]),  # vdst
+      v_mov_b32_e32(v[3], 0),
+      GLOBAL(GLOBALOp.GLOBAL_LOAD_D16_HI_I8, addr=v[3], vdst=v[5], data=v[4], saddr=s[2:3], offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], v[5]),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][0]
+    # 0x80 sign-extended = 0xFF80, lo=0xBEEF -> 0xFF80BEEF
+    self.assertEqual(result, 0xFF80BEEF, f"Expected 0xFF80BEEF, got 0x{result:08x}")
+
+  def test_global_store_b64_tril_pattern(self):
+    """Test the exact pattern from tril() kernel that was failing."""
+    TEST_OFFSET = 256
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[4], 0x01010101),
+      v_mov_b32_e32(v[10], s[4]),
+      v_mov_b32_e32(v[11], s[4]),
+      s_mov_b32(s[4], 0x01),
+      v_mov_b32_e32(v[12], s[4]),
+      v_mov_b32_e32(v[0], 0),
+      global_store_b64(addr=v[0], data=v[10:11], saddr=s[2:3], offset=TEST_OFFSET),
+      global_store_b8(addr=v[0], data=v[12], saddr=s[2:3], offset=TEST_OFFSET+8),
+      s_waitcnt(vmcnt=0),
+
+      v_mov_b32_e32(v[2], 0),
+      v_mov_b32_e32(v[1], 0),
+      GLOBAL(GLOBALOp.GLOBAL_LOAD_U16, addr=v[2], vdst=v[0], data=v[0], saddr=s[2:3], offset=TEST_OFFSET+3),
+      GLOBAL(GLOBALOp.GLOBAL_LOAD_D16_HI_B16, addr=v[1], vdst=v[1], data=v[1], saddr=s[2:3], offset=TEST_OFFSET+6),
+      GLOBAL(GLOBALOp.GLOBAL_LOAD_U8, addr=v[2], vdst=v[3], data=v[3], saddr=s[2:3], offset=TEST_OFFSET),
+      GLOBAL(GLOBALOp.GLOBAL_LOAD_U8, addr=v[2], vdst=v[4], data=v[4], saddr=s[2:3], offset=TEST_OFFSET+8),
+      s_waitcnt(vmcnt=0),
+
+      v_and_b32_e32(v[5], 0xffff, v[0]),
+      v_lshlrev_b32_e32(v[0], 24, v[0]),
+      v_lshrrev_b32_e32(v[5], 8, v[5]),
+      v_or_b32_e32(v[0], v[3], v[0]),
+      v_or_b32_e32(v[1], v[5], v[1]),
+
+      global_store_b64(addr=v[2], data=v[0:1], saddr=s[2:3], offset=TEST_OFFSET+16),
+      s_waitcnt(vmcnt=0),
+
+      GLOBAL(GLOBALOp.GLOBAL_LOAD_B64, addr=v[2], vdst=v[6:7], data=v[6:7], saddr=s[2:3], offset=TEST_OFFSET+16),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], v[6]),
+      v_mov_b32_e32(v[1], v[7]),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+
+    v0 = st.vgpr[0][0]
+    v1 = st.vgpr[0][1]
+    self.assertEqual(v0, 0x01000001, f"v0: expected 0x01000001, got 0x{v0:08x}")
+    self.assertEqual(v1, 0x01010001, f"v1: expected 0x01010001, got 0x{v1:08x}")
+
+    byte5 = (v1 >> 8) & 0xff
+    self.assertEqual(byte5, 0x00, f"byte5: expected 0x00, got 0x{byte5:02x}")
+
+
+class TestGlobalOffset(unittest.TestCase):
+  """Tests for GLOBAL instructions with different offsets.
+
+  These tests verify that instruction deduplication correctly handles different offset values.
+  If offset is made dynamic incorrectly, instructions with different offsets may load/store wrong data.
+  """
+
+  def test_global_load_different_offsets(self):
+    """Load from two different offsets and verify correct values."""
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      v_mov_b32_e32(v[0], s[2]),
+      v_mov_b32_e32(v[1], s[3]),
+      # Store 0xAAAAAAAA at offset 100
+      s_mov_b32(s[0], 0xAAAAAAAA),
+      v_mov_b32_e32(v[2], s[0]),
+      global_store_b32(addr=v[0:1], data=v[2], saddr=SrcEnum.NULL, offset=100),
+      # Store 0xBBBBBBBB at offset 200
+      s_mov_b32(s[0], 0xBBBBBBBB),
+      v_mov_b32_e32(v[2], s[0]),
+      global_store_b32(addr=v[0:1], data=v[2], saddr=SrcEnum.NULL, offset=200),
+      s_waitcnt(vmcnt=0),
+      # Load from offset 100 -> should get 0xAAAAAAAA
+      GLOBAL(GLOBALOp.GLOBAL_LOAD_B32, addr=v[0:1], vdst=v[3], saddr=SrcEnum.NULL, offset=100),
+      # Load from offset 200 -> should get 0xBBBBBBBB
+      GLOBAL(GLOBALOp.GLOBAL_LOAD_B32, addr=v[0:1], vdst=v[4], saddr=SrcEnum.NULL, offset=200),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], v[3]),
+      v_mov_b32_e32(v[1], v[4]),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 0xAAAAAAAA, f"offset 100: expected 0xAAAAAAAA, got 0x{st.vgpr[0][0]:08x}")
+    self.assertEqual(st.vgpr[0][1], 0xBBBBBBBB, f"offset 200: expected 0xBBBBBBBB, got 0x{st.vgpr[0][1]:08x}")
+
+  def test_global_store_different_offsets(self):
+    """Store to two different offsets and verify correct values."""
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      v_mov_b32_e32(v[0], s[2]),
+      v_mov_b32_e32(v[1], s[3]),
+      # Store 0x11111111 at offset 300
+      s_mov_b32(s[0], 0x11111111),
+      v_mov_b32_e32(v[2], s[0]),
+      global_store_b32(addr=v[0:1], data=v[2], saddr=SrcEnum.NULL, offset=300),
+      # Store 0x22222222 at offset 400
+      s_mov_b32(s[0], 0x22222222),
+      v_mov_b32_e32(v[3], s[0]),
+      global_store_b32(addr=v[0:1], data=v[3], saddr=SrcEnum.NULL, offset=400),
+      s_waitcnt(vmcnt=0),
+      # Load back to verify
+      GLOBAL(GLOBALOp.GLOBAL_LOAD_B32, addr=v[0:1], vdst=v[4], saddr=SrcEnum.NULL, offset=300),
+      GLOBAL(GLOBALOp.GLOBAL_LOAD_B32, addr=v[0:1], vdst=v[5], saddr=SrcEnum.NULL, offset=400),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], v[4]),
+      v_mov_b32_e32(v[1], v[5]),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 0x11111111, f"offset 300: expected 0x11111111, got 0x{st.vgpr[0][0]:08x}")
+    self.assertEqual(st.vgpr[0][1], 0x22222222, f"offset 400: expected 0x22222222, got 0x{st.vgpr[0][1]:08x}")
+
+  def test_global_negative_offset_no_saddr(self):
+    """Test negative offset without saddr (VGPR pair for address).
+    Store 0xAAAA at offset 100, 0xBBBB at offset 200.
+    Load with offset -100 from vaddr pointing to base+200 -> should get 0xAAAA (at 100).
+    Load with offset -100 from vaddr pointing to base+300 -> should get 0xBBBB (at 200)."""
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      v_mov_b32_e32(v[0], s[2]),
+      v_mov_b32_e32(v[1], s[3]),
+      # Store 0xAAAAAAAA at offset 100, 0xBBBBBBBB at offset 200
+      s_mov_b32(s[0], 0xAAAAAAAA),
+      v_mov_b32_e32(v[2], s[0]),
+      global_store_b32(addr=v[0:1], data=v[2], saddr=SrcEnum.NULL, offset=100),
+      s_mov_b32(s[0], 0xBBBBBBBB),
+      v_mov_b32_e32(v[2], s[0]),
+      global_store_b32(addr=v[0:1], data=v[2], saddr=SrcEnum.NULL, offset=200),
+      s_waitcnt(vmcnt=0),
+      # vaddr = base+200, load with offset -100 -> should get value at 100
+      s_add_u32(s[4], s[2], 200),
+      s_addc_u32(s[5], s[3], 0),
+      v_mov_b32_e32(v[4], s[4]),
+      v_mov_b32_e32(v[5], s[5]),
+      GLOBAL(GLOBALOp.GLOBAL_LOAD_B32, addr=v[4:5], vdst=v[6], saddr=SrcEnum.NULL, offset=-100),
+      # vaddr = base+300, load with offset -100 -> should get value at 200
+      s_add_u32(s[4], s[2], 300),
+      s_addc_u32(s[5], s[3], 0),
+      v_mov_b32_e32(v[4], s[4]),
+      v_mov_b32_e32(v[5], s[5]),
+      GLOBAL(GLOBALOp.GLOBAL_LOAD_B32, addr=v[4:5], vdst=v[7], saddr=SrcEnum.NULL, offset=-100),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], v[6]),
+      v_mov_b32_e32(v[1], v[7]),
+      v_mov_b32_e32(v[4], 0),
+      v_mov_b32_e32(v[5], 0),
+      v_mov_b32_e32(v[6], 0),
+      v_mov_b32_e32(v[7], 0),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+      s_mov_b32(s[4], 0),
+      s_mov_b32(s[5], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 0xAAAAAAAA, f"offset 200-100=100: expected 0xAAAAAAAA, got 0x{st.vgpr[0][0]:08x}")
+    self.assertEqual(st.vgpr[0][1], 0xBBBBBBBB, f"offset 300-100=200: expected 0xBBBBBBBB, got 0x{st.vgpr[0][1]:08x}")
+
+  def test_global_negative_offset_with_saddr(self):
+    """Test negative offset with saddr (SGPR pair for base address).
+    Store 0xAAAA at offset 100, 0xBBBB at offset 200.
+    Load with offset -100 from saddr pointing to base+200 -> should get 0xAAAA (at 100).
+    Load with offset -100 from saddr pointing to base+300 -> should get 0xBBBB (at 200)."""
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      v_mov_b32_e32(v[0], 0),
+      # Store 0xAAAAAAAA at offset 100, 0xBBBBBBBB at offset 200
+      s_mov_b32(s[0], 0xAAAAAAAA),
+      v_mov_b32_e32(v[2], s[0]),
+      global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=100),
+      s_mov_b32(s[0], 0xBBBBBBBB),
+      v_mov_b32_e32(v[2], s[0]),
+      global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=200),
+      s_waitcnt(vmcnt=0),
+      # saddr = base+200, load with offset -100 -> should get value at 100
+      s_add_u32(s[4], s[2], 200),
+      s_addc_u32(s[5], s[3], 0),
+      GLOBAL(GLOBALOp.GLOBAL_LOAD_B32, addr=v[0], vdst=v[6], saddr=s[4:5], offset=-100),
+      # saddr = base+300, load with offset -100 -> should get value at 200
+      s_add_u32(s[4], s[2], 300),
+      s_addc_u32(s[5], s[3], 0),
+      GLOBAL(GLOBALOp.GLOBAL_LOAD_B32, addr=v[0], vdst=v[7], saddr=s[4:5], offset=-100),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], v[6]),
+      v_mov_b32_e32(v[1], v[7]),
+      v_mov_b32_e32(v[6], 0),
+      v_mov_b32_e32(v[7], 0),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+      s_mov_b32(s[4], 0),
+      s_mov_b32(s[5], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 0xAAAAAAAA, f"offset 200-100=100: expected 0xAAAAAAAA, got 0x{st.vgpr[0][0]:08x}")
+    self.assertEqual(st.vgpr[0][1], 0xBBBBBBBB, f"offset 300-100=200: expected 0xBBBBBBBB, got 0x{st.vgpr[0][1]:08x}")
+
+
+if __name__ == '__main__':
+  unittest.main()

extra/assembly/amd/test/hw/test_scratch.py

@@ -0,0 +1,355 @@
+"""Tests for SCRATCH instructions - scratch (private) memory operations.
+
+Includes: scratch_load_*, scratch_store_*
+"""
+import unittest
+from extra.assembly.amd.test.hw.helpers import *
+
+class TestScratchStore(unittest.TestCase):
+  """Tests for SCRATCH store instructions."""
+
+  def test_scratch_store_b32_basic(self):
+    """SCRATCH_STORE_B32 stores 32-bit value to scratch memory."""
+    TEST_OFFSET = 256
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[4], 0xDEADBEEF),
+      v_mov_b32_e32(v[2], s[4]),
+      v_mov_b32_e32(v[0], 0),
+      # Store via scratch
+      scratch_store_b32(addr=v[0], data=v[2], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      # Load back via scratch
+      scratch_load_b32(addr=v[0], vdst=v[3], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], v[3]),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 0xDEADBEEF)
+
+  def test_scratch_store_b64_basic(self):
+    """SCRATCH_STORE_B64 stores 64-bit value to scratch memory."""
+    TEST_OFFSET = 256
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[4], 0xDEADBEEF),
+      s_mov_b32(s[5], 0xCAFEBABE),
+      v_mov_b32_e32(v[2], s[4]),
+      v_mov_b32_e32(v[3], s[5]),
+      v_mov_b32_e32(v[0], 0),
+      scratch_store_b64(addr=v[0], data=v[2:3], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      scratch_load_b64(addr=v[0], vdst=v[4:5], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], v[4]),
+      v_mov_b32_e32(v[1], v[5]),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 0xDEADBEEF)
+    self.assertEqual(st.vgpr[0][1], 0xCAFEBABE)
+
+  def test_scratch_store_b8_basic(self):
+    """SCRATCH_STORE_B8 stores single byte to scratch memory."""
+    TEST_OFFSET = 256
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      # First store full word
+      s_mov_b32(s[4], 0xDEADBEEF),
+      v_mov_b32_e32(v[2], s[4]),
+      v_mov_b32_e32(v[0], 0),
+      scratch_store_b32(addr=v[0], data=v[2], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      # Store single byte
+      v_mov_b32_e32(v[2], 0x42),
+      scratch_store_b8(addr=v[0], data=v[2], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      # Load back
+      scratch_load_b32(addr=v[0], vdst=v[3], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], v[3]),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    # Only byte 0 should change from 0xEF to 0x42
+    self.assertEqual(st.vgpr[0][0], 0xDEADBE42)
+
+  def test_scratch_store_b16_basic(self):
+    """SCRATCH_STORE_B16 stores 16-bit value to scratch memory."""
+    TEST_OFFSET = 256
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[4], 0xDEADBEEF),
+      v_mov_b32_e32(v[2], s[4]),
+      v_mov_b32_e32(v[0], 0),
+      scratch_store_b32(addr=v[0], data=v[2], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      s_mov_b32(s[4], 0xCAFE),
+      v_mov_b32_e32(v[2], s[4]),
+      scratch_store_b16(addr=v[0], data=v[2], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      scratch_load_b32(addr=v[0], vdst=v[3], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], v[3]),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 0xDEADCAFE)
+
+
+class TestScratchLoad(unittest.TestCase):
+  """Tests for SCRATCH load instructions."""
+
+  def test_scratch_load_b96(self):
+    """SCRATCH_LOAD_B96 loads 96-bit value correctly."""
+    TEST_OFFSET = 256
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      v_mov_b32_e32(v[0], 0),
+      s_mov_b32(s[4], 0xAAAAAAAA),
+      v_mov_b32_e32(v[2], s[4]),
+      s_mov_b32(s[4], 0xBBBBBBBB),
+      v_mov_b32_e32(v[3], s[4]),
+      s_mov_b32(s[4], 0xCCCCCCCC),
+      v_mov_b32_e32(v[4], s[4]),
+      scratch_store_b96(addr=v[0], data=v[2:4], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      scratch_load_b96(addr=v[0], vdst=v[5:7], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], v[5]),
+      v_mov_b32_e32(v[1], v[6]),
+      v_mov_b32_e32(v[2], v[7]),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 0xAAAAAAAA)
+    self.assertEqual(st.vgpr[0][1], 0xBBBBBBBB)
+    self.assertEqual(st.vgpr[0][2], 0xCCCCCCCC)
+
+  def test_scratch_load_b128(self):
+    """SCRATCH_LOAD_B128 loads 128-bit value correctly."""
+    TEST_OFFSET = 256
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      v_mov_b32_e32(v[0], 0),
+      s_mov_b32(s[4], 0xDEADBEEF),
+      v_mov_b32_e32(v[2], s[4]),
+      s_mov_b32(s[4], 0xCAFEBABE),
+      v_mov_b32_e32(v[3], s[4]),
+      s_mov_b32(s[4], 0x12345678),
+      v_mov_b32_e32(v[4], s[4]),
+      s_mov_b32(s[4], 0x9ABCDEF0),
+      v_mov_b32_e32(v[5], s[4]),
+      scratch_store_b128(addr=v[0], data=v[2:5], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      scratch_load_b128(addr=v[0], vdst=v[6:9], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], v[6]),
+      v_mov_b32_e32(v[1], v[7]),
+      v_mov_b32_e32(v[2], v[8]),
+      v_mov_b32_e32(v[3], v[9]),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 0xDEADBEEF)
+    self.assertEqual(st.vgpr[0][1], 0xCAFEBABE)
+    self.assertEqual(st.vgpr[0][2], 0x12345678)
+    self.assertEqual(st.vgpr[0][3], 0x9ABCDEF0)
+
+  def test_scratch_load_u8(self):
+    """SCRATCH_LOAD_U8 loads unsigned byte with zero extension."""
+    TEST_OFFSET = 256
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      v_mov_b32_e32(v[0], 0),
+      s_mov_b32(s[4], 0xDEADBEAB),
+      v_mov_b32_e32(v[2], s[4]),
+      scratch_store_b32(addr=v[0], data=v[2], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      scratch_load_u8(addr=v[0], vdst=v[3], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], v[3]),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 0xAB)
+
+  def test_scratch_load_i8(self):
+    """SCRATCH_LOAD_I8 loads signed byte with sign extension."""
+    TEST_OFFSET = 256
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      v_mov_b32_e32(v[0], 0),
+      s_mov_b32(s[4], 0x80),  # -128 as signed byte
+      v_mov_b32_e32(v[2], s[4]),
+      scratch_store_b8(addr=v[0], data=v[2], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      scratch_load_i8(addr=v[0], vdst=v[3], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], v[3]),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 0xFFFFFF80)
+
+  def test_scratch_load_u16(self):
+    """SCRATCH_LOAD_U16 loads unsigned 16-bit with zero extension."""
+    TEST_OFFSET = 256
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      v_mov_b32_e32(v[0], 0),
+      s_mov_b32(s[4], 0xDEADCAFE),
+      v_mov_b32_e32(v[2], s[4]),
+      scratch_store_b32(addr=v[0], data=v[2], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      scratch_load_u16(addr=v[0], vdst=v[3], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], v[3]),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 0xCAFE)
+
+  def test_scratch_load_i16(self):
+    """SCRATCH_LOAD_I16 loads signed 16-bit with sign extension."""
+    TEST_OFFSET = 256
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      v_mov_b32_e32(v[0], 0),
+      s_mov_b32(s[4], 0x8000),  # -32768 as signed 16-bit
+      v_mov_b32_e32(v[2], s[4]),
+      scratch_store_b16(addr=v[0], data=v[2], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      scratch_load_i16(addr=v[0], vdst=v[3], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], v[3]),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 0xFFFF8000)
+
+
+class TestScratchSVE(unittest.TestCase):
+  """Tests for SCRATCH SVE (Scratch VGPR Enable) bit behavior."""
+
+  def test_scratch_sve_zero_ignores_vaddr(self):
+    """With SVE=0, VADDR should be ignored in address calculation."""
+    TEST_OFFSET = 256
+    # Store a marker value at offset 256 (where SVE=0 should go)
+    # Then set v[0] to a non-zero value (100) and store via scratch with SVE=0
+    # If SVE=0 is handled correctly, the VADDR (100) should be IGNORED,
+    # and the store should go to offset 256, not 256+100=356
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      # First, store 0xAAAAAAAA at offset 256 with v[0]=0
+      v_mov_b32_e32(v[0], 0),
+      s_mov_b32(s[4], 0xAAAAAAAA),
+      v_mov_b32_e32(v[2], s[4]),
+      scratch_store_b32(addr=v[0], data=v[2], saddr=SrcEnum.NULL, offset=TEST_OFFSET, sve=0),
+      s_waitcnt(vmcnt=0),
+      # Now set v[0] to 100 (non-zero) and store 0xBBBBBBBB with SVE=0
+      # With SVE=0, v[0] should be IGNORED, so this should overwrite offset 256
+      v_mov_b32_e32(v[0], 100),
+      s_mov_b32(s[4], 0xBBBBBBBB),
+      v_mov_b32_e32(v[2], s[4]),
+      scratch_store_b32(addr=v[0], data=v[2], saddr=SrcEnum.NULL, offset=TEST_OFFSET, sve=0),
+      s_waitcnt(vmcnt=0),
+      # Load back from offset 256 (with v[0]=0) - should get 0xBBBBBBBB
+      v_mov_b32_e32(v[0], 0),
+      scratch_load_b32(addr=v[0], vdst=v[3], saddr=SrcEnum.NULL, offset=TEST_OFFSET, sve=0),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], v[3]),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    # If SVE=0 works correctly, v[0] should be 0xBBBBBBBB (the second store overwrote the first)
+    # If SVE=0 is wrong (VADDR used), v[0] would be 0xAAAAAAAA (stores went to different locations)
+    self.assertEqual(st.vgpr[0][0], 0xBBBBBBBB, "SVE=0 should ignore VADDR, both stores should go to same location")
+
+  def test_scratch_sve_one_uses_vaddr(self):
+    """With SVE=1, VADDR should be used as offset in address calculation."""
+    TEST_OFFSET = 256
+    # Store at offset 256 with v[0]=0, then store at offset 256 with v[0]=100 and SVE=1
+    # With SVE=1, the second store should go to 256+100=356, not 256
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      # First, store 0xAAAAAAAA at offset 256 with v[0]=0
+      v_mov_b32_e32(v[0], 0),
+      s_mov_b32(s[4], 0xAAAAAAAA),
+      v_mov_b32_e32(v[2], s[4]),
+      scratch_store_b32(addr=v[0], data=v[2], saddr=SrcEnum.NULL, offset=TEST_OFFSET, sve=1),
+      s_waitcnt(vmcnt=0),
+      # Now set v[0] to 100 and store 0xBBBBBBBB with SVE=1
+      # With SVE=1, v[0] IS used, so this should go to offset 256+100=356
+      v_mov_b32_e32(v[0], 100),
+      s_mov_b32(s[4], 0xBBBBBBBB),
+      v_mov_b32_e32(v[2], s[4]),
+      scratch_store_b32(addr=v[0], data=v[2], saddr=SrcEnum.NULL, offset=TEST_OFFSET, sve=1),
+      s_waitcnt(vmcnt=0),
+      # Load back from offset 256 (with v[0]=0) - should still be 0xAAAAAAAA
+      v_mov_b32_e32(v[0], 0),
+      scratch_load_b32(addr=v[0], vdst=v[3], saddr=SrcEnum.NULL, offset=TEST_OFFSET, sve=1),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], v[3]),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    # If SVE=1 works correctly, v[0] should be 0xAAAAAAAA (stores went to different locations)
+    self.assertEqual(st.vgpr[0][0], 0xAAAAAAAA, "SVE=1 should use VADDR, stores should go to different locations")
+
+
+class TestScratchMultiLane(unittest.TestCase):
+  """Tests for SCRATCH operations with multiple lanes."""
+
+  def test_scratch_store_load_multi_lane(self):
+    """SCRATCH store/load works correctly with multiple lanes (private per-lane memory)."""
+    TEST_OFFSET = 256
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=SrcEnum.NULL),
+      s_waitcnt(lgkmcnt=0),
+      # Each lane stores its lane ID
+      v_mov_b32_e32(v[0], 0),
+      v_mov_b32_e32(v[2], v[255]),  # v[255] has packed workitem IDs, low 10 bits = x
+      v_and_b32_e32(v[2], 0x3FF, v[2]),  # extract lane ID
+      scratch_store_b32(addr=v[0], data=v[2], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      # Load back
+      scratch_load_b32(addr=v[0], vdst=v[3], saddr=SrcEnum.NULL, offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      v_mov_b32_e32(v[0], v[3]),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=4)
+    # Each lane should have loaded its own lane ID
+    for lane in range(4):
+      self.assertEqual(st.vgpr[lane][0], lane, f"Lane {lane} should have value {lane}")
+
+
+if __name__ == '__main__':
+  unittest.main()

extra/assembly/amd/test/hw/test_smem.py

@@ -0,0 +1,448 @@
+"""Tests for SMEM instructions - scalar memory operations.
+
+Includes: s_load_b32, s_load_b64, s_load_b128, s_load_b256, s_load_b512
+Tests both immediate and register offset addressing modes.
+"""
+import unittest
+from extra.assembly.amd.test.hw.helpers import *
+
+# Use offset into output buffer for test data (output buffer is 2124 bytes)
+TEST_OFFSET = 2000
+
+# Cache invalidation sequence for scalar loads after vector stores
+# s_wait_idle waits for all outstanding memory operations including cache flushes
+CACHE_INV = [s_gl1_inv(), s_dcache_inv(), s_wait_idle()]
+
+class TestSLoadRegisterOffset(unittest.TestCase):
+  """Tests for s_load with register offset (soffset field).
+
+  Bug: s_load_b32(s[dst], s[base:base+1], s[off]) ignores the register offset
+  and only uses the immediate offset field. This causes incorrect memory loads
+  when the offset comes from a register.
+  """
+
+  def test_s_load_b32_register_offset_basic(self):
+    """s_load_b32 with register offset should load from base + reg_offset."""
+    instructions = [
+      # Load output buffer pointer from args
+      s_load_b64(s[2:3], s[80:81], 0, soffset=NULL),
+      s_waitcnt(lgkmcnt=0),
+      # Store test values to output buffer: 0xAAAAAAAA at offset, 0xBBBBBBBB at offset+4
+      s_mov_b32(s[4], 0xAAAAAAAA),
+      s_mov_b32(s[5], 0xBBBBBBBB),
+      v_mov_b32_e32(v[2], s[4]),
+      v_mov_b32_e32(v[3], s[5]),
+      v_mov_b32_e32(v[0], 0),
+      global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET),
+      global_store_b32(addr=v[0], data=v[3], saddr=s[2:3], offset=TEST_OFFSET+4),
+      s_waitcnt(vmcnt=0),
+      *CACHE_INV,
+      # Now test s_load with register offset
+      # Put offset value in s[4]: offset = 4 bytes (1 dword)
+      s_mov_b32(s[4], 4),
+      # Load from out_ptr + TEST_OFFSET + s[4] (should load 0xBBBBBBBB)
+      s_load_b32(s[5], s[2:3], s[4], offset=TEST_OFFSET),
+      s_waitcnt(0),
+      # Zero out pointer regs (different addresses in emu vs hw)
+      s_mov_b32(s[2], 0), s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[5], 0xBBBBBBBB,
+      f"s_load with reg offset 4 should load 0xBBBBBBBB: s[5]=0x{st.sgpr[5]:08x}")
+
+  def test_s_load_b32_register_offset_different_from_immediate(self):
+    """s_load_b32 with register offset loads different data than immediate offset 0."""
+    instructions = [
+      # Load output buffer pointer from args
+      s_load_b64(s[2:3], s[80:81], 0, soffset=NULL),
+      s_waitcnt(lgkmcnt=0),
+      # Store test values: 0xAAAAAAAA at offset, 0xBBBBBBBB at offset+4
+      s_mov_b32(s[4], 0xAAAAAAAA),
+      s_mov_b32(s[5], 0xBBBBBBBB),
+      v_mov_b32_e32(v[2], s[4]),
+      v_mov_b32_e32(v[3], s[5]),
+      v_mov_b32_e32(v[0], 0),
+      global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET),
+      global_store_b32(addr=v[0], data=v[3], saddr=s[2:3], offset=TEST_OFFSET+4),
+      s_waitcnt(vmcnt=0),
+      *CACHE_INV,
+      # Load with immediate offset 0
+      s_load_b32(s[5], s[2:3], NULL, offset=TEST_OFFSET),
+      s_waitcnt(0),
+      # Load with register offset 4
+      s_mov_b32(s[4], 4),
+      s_load_b32(s[6], s[2:3], s[4], offset=TEST_OFFSET),
+      s_waitcnt(0),
+      # Zero out pointer regs (different addresses in emu vs hw)
+      s_mov_b32(s[2], 0), s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    # s[5] has dword at offset 0 (0xAAAAAAAA), s[6] has dword at offset 4 (0xBBBBBBBB)
+    self.assertEqual(st.sgpr[5], 0xAAAAAAAA)
+    self.assertEqual(st.sgpr[6], 0xBBBBBBBB)
+    self.assertNotEqual(st.sgpr[5], st.sgpr[6],
+      f"s_load with reg offset 4 should load different value than offset 0: "
+      f"s[5]=0x{st.sgpr[5]:08x}, s[6]=0x{st.sgpr[6]:08x}")
+
+  def test_s_load_b32_register_offset_same_as_dst(self):
+    """s_load_b32 where soffset register is same as destination.
+
+    This is the exact pattern that exposes the bug:
+    s_load_b32(s[8], s[2:3], s[8])
+    The offset should be read BEFORE the destination is overwritten.
+    """
+    instructions = [
+      # Load output buffer pointer from args
+      s_load_b64(s[2:3], s[80:81], 0, soffset=NULL),
+      s_waitcnt(lgkmcnt=0),
+      # Store test values: 0xAAAAAAAA at offset, 0xBBBBBBBB at offset+4
+      s_mov_b32(s[6], 0xAAAAAAAA),
+      s_mov_b32(s[7], 0xBBBBBBBB),
+      v_mov_b32_e32(v[2], s[6]),
+      v_mov_b32_e32(v[3], s[7]),
+      v_mov_b32_e32(v[0], 0),
+      global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET),
+      global_store_b32(addr=v[0], data=v[3], saddr=s[2:3], offset=TEST_OFFSET+4),
+      s_waitcnt(vmcnt=0),
+      *CACHE_INV,
+      # Set up s[4] = 4 (offset in bytes)
+      s_mov_b32(s[4], 4),
+      # Load using s[4] as both offset and destination
+      # Should load from base + 4, then store result in s[4]
+      s_load_b32(s[4], s[2:3], s[4], offset=TEST_OFFSET),
+      s_waitcnt(0),
+      # Also load with immediate offset 4 for comparison
+      s_load_b32(s[5], s[2:3], NULL, offset=TEST_OFFSET+4),
+      s_waitcnt(0),
+      # Zero out pointer regs (different addresses in emu vs hw)
+      s_mov_b32(s[2], 0), s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    # s[4] and s[5] should have the same value (both loaded from offset 4 = 0xBBBBBBBB)
+    self.assertEqual(st.sgpr[4], 0xBBBBBBBB)
+    self.assertEqual(st.sgpr[4], st.sgpr[5],
+      f"s_load with reg offset s[4]=4 should match immediate offset=4: "
+      f"s[4]=0x{st.sgpr[4]:08x}, s[5]=0x{st.sgpr[5]:08x}")
+
+  def test_s_load_b32_register_offset_zero(self):
+    """s_load_b32 with register offset = 0 should be same as immediate offset 0."""
+    instructions = [
+      # Load output buffer pointer from args
+      s_load_b64(s[2:3], s[80:81], 0, soffset=NULL),
+      s_waitcnt(lgkmcnt=0),
+      # Store test value: 0xDEADBEEF at offset
+      s_mov_b32(s[7], 0xDEADBEEF),
+      v_mov_b32_e32(v[2], s[7]),
+      v_mov_b32_e32(v[0], 0),
+      global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET),
+      s_waitcnt(vmcnt=0),
+      *CACHE_INV,
+      # Load with register offset 0
+      s_mov_b32(s[4], 0),
+      s_load_b32(s[5], s[2:3], s[4], offset=TEST_OFFSET),
+      s_waitcnt(0),
+      # Load with immediate offset 0
+      s_load_b32(s[6], s[2:3], NULL, offset=TEST_OFFSET),
+      s_waitcnt(0),
+      # Zero out pointer regs (different addresses in emu vs hw)
+      s_mov_b32(s[2], 0), s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[5], 0xDEADBEEF)
+    self.assertEqual(st.sgpr[5], st.sgpr[6],
+      f"s_load with reg offset 0 should match immediate offset 0: "
+      f"s[5]=0x{st.sgpr[5]:08x}, s[6]=0x{st.sgpr[6]:08x}")
+
+  def test_s_load_b32_register_plus_immediate_offset(self):
+    """s_load_b32 with both register and immediate offset should add them."""
+    instructions = [
+      # Load output buffer pointer from args
+      s_load_b64(s[2:3], s[80:81], 0, soffset=NULL),
+      s_waitcnt(lgkmcnt=0),
+      # Store test values: 0xAAAAAAAA at offset, 0xBBBBBBBB at offset+4
+      s_mov_b32(s[8], 0xAAAAAAAA),
+      s_mov_b32(s[9], 0xBBBBBBBB),
+      v_mov_b32_e32(v[2], s[8]),
+      v_mov_b32_e32(v[3], s[9]),
+      v_mov_b32_e32(v[0], 0),
+      global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET),
+      global_store_b32(addr=v[0], data=v[3], saddr=s[2:3], offset=TEST_OFFSET+4),
+      s_waitcnt(vmcnt=0),
+      *CACHE_INV,
+      # reg offset = 4, imm offset = 0 -> total offset = 4
+      s_mov_b32(s[4], 4),
+      s_load_b32(s[5], s[2:3], s[4], offset=TEST_OFFSET),
+      s_waitcnt(0),
+      # reg offset = 0, imm offset = 4 -> total offset = 4
+      s_mov_b32(s[6], 0),
+      s_load_b32(s[7], s[2:3], s[6], offset=TEST_OFFSET+4),
+      s_waitcnt(0),
+      # Zero out pointer regs (different addresses in emu vs hw)
+      s_mov_b32(s[2], 0), s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    # Both should load from offset 4 (0xBBBBBBBB)
+    self.assertEqual(st.sgpr[5], 0xBBBBBBBB)
+    self.assertEqual(st.sgpr[7], 0xBBBBBBBB)
+    self.assertEqual(st.sgpr[5], st.sgpr[7],
+      f"reg_off=4 + imm_off=0 should equal reg_off=0 + imm_off=4: "
+      f"s[5]=0x{st.sgpr[5]:08x}, s[7]=0x{st.sgpr[7]:08x}")
+
+
+class TestSLoadMultiDword(unittest.TestCase):
+  """Tests for multi-dword s_load with register offset."""
+
+  def test_s_load_b64_register_offset(self):
+    """s_load_b64 with register offset should load 2 dwords from base + reg_offset."""
+    instructions = [
+      # Load output buffer pointer from args
+      s_load_b64(s[2:3], s[80:81], 0, soffset=NULL),
+      s_waitcnt(lgkmcnt=0),
+      # Store test values: 0xAAAAAAAA, 0xBBBBBBBB at offset
+      s_mov_b32(s[10], 0xAAAAAAAA),
+      s_mov_b32(s[11], 0xBBBBBBBB),
+      v_mov_b32_e32(v[2], s[10]),
+      v_mov_b32_e32(v[3], s[11]),
+      v_mov_b32_e32(v[0], 0),
+      global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET),
+      global_store_b32(addr=v[0], data=v[3], saddr=s[2:3], offset=TEST_OFFSET+4),
+      s_waitcnt(vmcnt=0),
+      *CACHE_INV,
+      # Load with register offset 0
+      s_mov_b32(s[4], 0),
+      s_load_b64(s[6:7], s[2:3], s[4], offset=TEST_OFFSET),
+      s_waitcnt(0),
+      # Compare with immediate offset
+      s_load_b64(s[8:9], s[2:3], NULL, offset=TEST_OFFSET),
+      s_waitcnt(0),
+      # Zero out pointer regs (different addresses in emu vs hw)
+      s_mov_b32(s[2], 0), s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[6], 0xAAAAAAAA)
+    self.assertEqual(st.sgpr[7], 0xBBBBBBBB)
+    self.assertEqual(st.sgpr[6], st.sgpr[8])
+    self.assertEqual(st.sgpr[7], st.sgpr[9])
+
+  def test_s_load_b128_register_offset(self):
+    """s_load_b128 with register offset should load 4 dwords from base + reg_offset."""
+    instructions = [
+      # Load output buffer pointer from args
+      s_load_b64(s[2:3], s[80:81], 0, soffset=NULL),
+      s_waitcnt(lgkmcnt=0),
+      # Store test values: 0xAAAAAAAA, 0xBBBBBBBB, 0xCCCCCCCC, 0xDDDDDDDD at offset
+      v_mov_b32_e32(v[0], 0),
+      s_mov_b32(s[14], 0xAAAAAAAA),
+      v_mov_b32_e32(v[2], s[14]),
+      global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET),
+      s_mov_b32(s[14], 0xBBBBBBBB),
+      v_mov_b32_e32(v[2], s[14]),
+      global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET+4),
+      s_mov_b32(s[14], 0xCCCCCCCC),
+      v_mov_b32_e32(v[2], s[14]),
+      global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET+8),
+      s_mov_b32(s[14], 0xDDDDDDDD),
+      v_mov_b32_e32(v[2], s[14]),
+      global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET+12),
+      s_waitcnt(vmcnt=0),
+      *CACHE_INV,
+      # Load with register offset 0 (s_load_b128 requires 4-aligned dest: s[4], s[8], s[12], ...)
+      s_mov_b32(s[15], 0),
+      s_load_b128(s[4:7], s[2:3], s[15], offset=TEST_OFFSET),
+      s_waitcnt(0),
+      # Compare with immediate offset
+      s_load_b128(s[8:11], s[2:3], NULL, offset=TEST_OFFSET),
+      s_waitcnt(0),
+      # Zero out pointer regs (different addresses in emu vs hw)
+      s_mov_b32(s[2], 0), s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[4], 0xAAAAAAAA)
+    self.assertEqual(st.sgpr[5], 0xBBBBBBBB)
+    self.assertEqual(st.sgpr[6], 0xCCCCCCCC)
+    self.assertEqual(st.sgpr[7], 0xDDDDDDDD)
+    self.assertEqual(st.sgpr[4], st.sgpr[8])
+    self.assertEqual(st.sgpr[5], st.sgpr[9])
+
+
+class TestSLoadLarge(unittest.TestCase):
+  """Tests for large s_load operations (s_load_b256, s_load_b512)."""
+
+  def test_s_load_b256_basic(self):
+    """s_load_b256 loads 8 consecutive dwords."""
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=NULL),
+      s_waitcnt(lgkmcnt=0),
+      v_mov_b32_e32(v[0], 0),
+      # Store 8 test values
+      s_mov_b32(s[20], 0x11111111),
+      v_mov_b32_e32(v[2], s[20]),
+      global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET),
+      s_mov_b32(s[20], 0x22222222),
+      v_mov_b32_e32(v[2], s[20]),
+      global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET+4),
+      s_mov_b32(s[20], 0x33333333),
+      v_mov_b32_e32(v[2], s[20]),
+      global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET+8),
+      s_mov_b32(s[20], 0x44444444),
+      v_mov_b32_e32(v[2], s[20]),
+      global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET+12),
+      s_mov_b32(s[20], 0x55555555),
+      v_mov_b32_e32(v[2], s[20]),
+      global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET+16),
+      s_mov_b32(s[20], 0x66666666),
+      v_mov_b32_e32(v[2], s[20]),
+      global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET+20),
+      s_mov_b32(s[20], 0x77777777),
+      v_mov_b32_e32(v[2], s[20]),
+      global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET+24),
+      s_mov_b32(s[20], 0x88888888),
+      v_mov_b32_e32(v[2], s[20]),
+      global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET+28),
+      s_waitcnt(vmcnt=0),
+      *CACHE_INV,
+      # Load all 8 dwords with s_load_b256
+      s_load_b256(s[4:11], s[2:3], NULL, offset=TEST_OFFSET),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[2], 0), s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[4], 0x11111111)
+    self.assertEqual(st.sgpr[5], 0x22222222)
+    self.assertEqual(st.sgpr[6], 0x33333333)
+    self.assertEqual(st.sgpr[7], 0x44444444)
+    self.assertEqual(st.sgpr[8], 0x55555555)
+    self.assertEqual(st.sgpr[9], 0x66666666)
+    self.assertEqual(st.sgpr[10], 0x77777777)
+    self.assertEqual(st.sgpr[11], 0x88888888)
+
+  def test_s_load_b512_basic(self):
+    """s_load_b512 loads 16 consecutive dwords."""
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=NULL),
+      s_waitcnt(lgkmcnt=0),
+      v_mov_b32_e32(v[0], 0),
+      # Store 16 test values (use a pattern: 0x10, 0x20, ..., 0x100)
+      *[instr for i in range(16) for instr in [
+        s_mov_b32(s[20], (i + 1) * 0x11111111),
+        v_mov_b32_e32(v[2], s[20]),
+        global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET + i * 4),
+      ]],
+      s_waitcnt(vmcnt=0),
+      *CACHE_INV,
+      # Load all 16 dwords with s_load_b512
+      s_load_b512(s[64:79], s[2:3], NULL, offset=TEST_OFFSET),
+      s_waitcnt(lgkmcnt=0),
+      # Copy results to lower regs for verification (since st.sgpr only has 16 regs in test)
+      s_mov_b32(s[4], s[64]),
+      s_mov_b32(s[5], s[65]),
+      s_mov_b32(s[6], s[78]),
+      s_mov_b32(s[7], s[79]),
+      s_mov_b32(s[2], 0), s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[4], 0x11111111, "first dword")
+    self.assertEqual(st.sgpr[5], 0x22222222, "second dword")
+    self.assertEqual(st.sgpr[6], 0xFFFFFFFF & (15 * 0x11111111), "15th dword")
+    self.assertEqual(st.sgpr[7], 0xFFFFFFFF & (16 * 0x11111111), "16th dword")
+
+  def test_s_load_b256_with_register_offset(self):
+    """s_load_b256 with register offset should add reg offset to address."""
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=NULL),
+      s_waitcnt(lgkmcnt=0),
+      v_mov_b32_e32(v[0], 0),
+      # Store pattern at TEST_OFFSET+8: skip first 2 dwords
+      *[instr for i in range(8) for instr in [
+        s_mov_b32(s[20], (i + 1) * 0x11111111),
+        v_mov_b32_e32(v[2], s[20]),
+        global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=TEST_OFFSET + 8 + i * 4),
+      ]],
+      s_waitcnt(vmcnt=0),
+      *CACHE_INV,
+      # Load with register offset 8
+      s_mov_b32(s[20], 8),
+      s_load_b256(s[4:11], s[2:3], s[20], offset=TEST_OFFSET),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[2], 0), s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[4], 0x11111111, "first dword at offset+8")
+    self.assertEqual(st.sgpr[5], 0x22222222, "second dword at offset+8")
+    self.assertEqual(st.sgpr[11], 0x88888888, "last dword at offset+8")
+
+
+class TestSLoadOffset(unittest.TestCase):
+  """Tests for s_load with different immediate offsets.
+
+  These tests verify that instruction deduplication correctly handles different offset values.
+  If offset is made dynamic incorrectly, instructions with different offsets may load wrong data.
+  """
+
+  def test_s_load_different_offsets(self):
+    """Load from two different offsets and verify correct values."""
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=NULL),
+      s_waitcnt(lgkmcnt=0),
+      v_mov_b32_e32(v[0], 0),
+      # Store 0xAAAAAAAA at offset 100
+      s_mov_b32(s[4], 0xAAAAAAAA),
+      v_mov_b32_e32(v[2], s[4]),
+      global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=100),
+      # Store 0xBBBBBBBB at offset 200
+      s_mov_b32(s[4], 0xBBBBBBBB),
+      v_mov_b32_e32(v[2], s[4]),
+      global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=200),
+      s_waitcnt(vmcnt=0),
+      *CACHE_INV,
+      # Load from offset 100 -> should get 0xAAAAAAAA
+      s_load_b32(s[4], s[2:3], NULL, offset=100),
+      # Load from offset 200 -> should get 0xBBBBBBBB
+      s_load_b32(s[5], s[2:3], NULL, offset=200),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[2], 0), s_mov_b32(s[3], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[4], 0xAAAAAAAA, f"offset 100: expected 0xAAAAAAAA, got 0x{st.sgpr[4]:08x}")
+    self.assertEqual(st.sgpr[5], 0xBBBBBBBB, f"offset 200: expected 0xBBBBBBBB, got 0x{st.sgpr[5]:08x}")
+
+  def test_s_load_negative_offset(self):
+    """Test negative offset (21-bit signed).
+    Store 0xAAAA at offset 100, 0xBBBB at offset 200.
+    Load with offset -100 from base+200 -> should get 0xAAAA.
+    Load with offset -100 from base+300 -> should get 0xBBBB."""
+    instructions = [
+      s_load_b64(s[2:3], s[80:81], 0, soffset=NULL),
+      s_waitcnt(lgkmcnt=0),
+      v_mov_b32_e32(v[0], 0),
+      # Store 0xAAAAAAAA at offset 100, 0xBBBBBBBB at offset 200
+      s_mov_b32(s[8], 0xAAAAAAAA),
+      v_mov_b32_e32(v[2], s[8]),
+      global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=100),
+      s_mov_b32(s[8], 0xBBBBBBBB),
+      v_mov_b32_e32(v[2], s[8]),
+      global_store_b32(addr=v[0], data=v[2], saddr=s[2:3], offset=200),
+      s_waitcnt(vmcnt=0),
+      *CACHE_INV,
+      # base+200, load with offset -100 -> should get value at 100
+      s_add_u32(s[6], s[2], 200),
+      s_addc_u32(s[7], s[3], 0),
+      s_load_b32(s[4], s[6:7], NULL, offset=-100),
+      # base+300, load with offset -100 -> should get value at 200
+      s_add_u32(s[6], s[2], 300),
+      s_addc_u32(s[7], s[3], 0),
+      s_load_b32(s[5], s[6:7], NULL, offset=-100),
+      s_waitcnt(lgkmcnt=0),
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 0),
+      s_mov_b32(s[6], 0),
+      s_mov_b32(s[7], 0),
+      s_mov_b32(s[8], 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[4], 0xAAAAAAAA, f"offset 200-100=100: expected 0xAAAAAAAA, got 0x{st.sgpr[4]:08x}")
+    self.assertEqual(st.sgpr[5], 0xBBBBBBBB, f"offset 300-100=200: expected 0xBBBBBBBB, got 0x{st.sgpr[5]:08x}")
+
+
+if __name__ == '__main__':
+  unittest.main()

extra/assembly/amd/test/hw/test_sop.py

@@ -0,0 +1,936 @@
+"""Tests for SOP instructions - scalar operations.
+
+Includes: s_add_u32, s_mov_b32, s_and_b32, s_or_b32, s_quadmask_b32, s_wqm_b32,
+          s_cbranch_vccnz, s_cbranch_vccz
+"""
+import unittest
+from extra.assembly.amd.test.hw.helpers import *
+
+class TestBasicScalar(unittest.TestCase):
+  """Tests for basic scalar operations."""
+
+  def test_s_add_u32(self):
+    """S_ADD_U32 adds two scalar values."""
+    instructions = [
+      s_mov_b32(s[0], 100),
+      s_mov_b32(s[1], 200),
+      s_add_u32(s[2], s[0], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[2], 300)
+
+  def test_s_add_u32_carry(self):
+    """S_ADD_U32 sets SCC on overflow."""
+    instructions = [
+      s_mov_b32(s[0], 64),
+      s_not_b32(s[0], s[0]),  # ~64 = 0xffffffbf
+      s_mov_b32(s[1], 64),
+      s_add_u32(s[2], s[0], s[1]),  # 0xffffffbf + 64 = 0xffffffff
+      s_mov_b32(s[3], 1),
+      s_add_u32(s[4], s[2], s[3]),  # 0xffffffff + 1 = overflow
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[4], 0)
+    self.assertEqual(st.scc, 1)
+
+  def test_s_brev_b32(self):
+    """S_BREV_B32 reverses bits of a 32-bit value."""
+    # 10 = 0b00000000000000000000000000001010
+    # reversed = 0b01010000000000000000000000000000 = 0x50000000
+    instructions = [
+      s_mov_b32(s[0], 10),
+      s_brev_b32(s[1], s[0]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[1], 0x50000000)
+
+  def test_s_brev_b32_all_ones(self):
+    """S_BREV_B32 with all ones stays all ones."""
+    instructions = [
+      s_mov_b32(s[0], 0xFFFFFFFF),
+      s_brev_b32(s[1], s[0]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[1], 0xFFFFFFFF)
+
+  def test_s_brev_b32_single_bit(self):
+    """S_BREV_B32 with bit 0 set becomes bit 31."""
+    instructions = [
+      s_mov_b32(s[0], 1),
+      s_brev_b32(s[1], s[0]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[1], 0x80000000)
+
+  @skip_unless_gfx(11, 5, "SALU FP ops require gfx1150+")
+  def test_s_fmamk_f32(self):
+    """S_FMAMK_F32: D = S0 * literal + S1."""
+    # 2.0 * 3.0 + 1.0 = 7.0
+    instructions = [
+      s_mov_b32(s[0], f2i(2.0)),
+      s_mov_b32(s[1], f2i(1.0)),
+      s_fmamk_f32(s[2], s[0], s[1], literal=f2i(3.0)),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[2], f2i(7.0))
+
+  @skip_unless_gfx(11, 5, "SALU FP ops require gfx1150+")
+  def test_s_fmamk_f32_negative(self):
+    """S_FMAMK_F32 with negative values."""
+    # -2.0 * 4.0 + 10.0 = 2.0
+    instructions = [
+      s_mov_b32(s[0], f2i(-2.0)),
+      s_mov_b32(s[1], f2i(10.0)),
+      s_fmamk_f32(s[2], s[0], s[1], literal=f2i(4.0)),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[2], f2i(2.0))
+
+
+class TestPack(unittest.TestCase):
+  """Tests for S_PACK instructions."""
+
+  def test_s_pack_ll_b32_b16(self):
+    """S_PACK_LL_B32_B16 packs low 16 bits of two sources into one 32-bit result."""
+    instructions = [
+      s_mov_b32(s[0], 0xDEADAAAA),
+      s_mov_b32(s[1], 0xDEADBBBB),
+      s_pack_ll_b32_b16(s[2], s[0], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[2], 0xBBBBAAAA)
+
+  def test_s_pack_lh_b32_b16(self):
+    """S_PACK_LH_B32_B16: D0 = { S1[31:16], S0[15:0] }."""
+    instructions = [
+      s_mov_b32(s[0], 0xDEADAAAA),
+      s_mov_b32(s[1], 0xDEADBBBB),
+      s_pack_lh_b32_b16(s[2], s[0], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[2], 0xDEADAAAA)
+
+  def test_s_pack_hh_b32_b16(self):
+    """S_PACK_HH_B32_B16: D0 = { S1[31:16], S0[31:16] }."""
+    instructions = [
+      s_mov_b32(s[0], 0xDEADAAAA),
+      s_mov_b32(s[1], 0xDEADBBBB),
+      s_pack_hh_b32_b16(s[2], s[0], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[2], 0xDEADDEAD)
+
+  def test_s_pack_hl_b32_b16(self):
+    """S_PACK_HL_B32_B16: D0 = { S1[15:0], S0[31:16] }."""
+    instructions = [
+      s_mov_b32(s[0], 0xDEADAAAA),
+      s_mov_b32(s[1], 0xDEADBBBB),
+      s_pack_hl_b32_b16(s[2], s[0], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[2], 0xBBBBDEAD)
+
+
+class TestQuadmaskWqm(unittest.TestCase):
+  """Tests for S_QUADMASK_B32 and S_WQM_B32."""
+
+  def test_s_quadmask_b32_all_quads_active(self):
+    """S_QUADMASK_B32 with all quads active."""
+    instructions = [
+      s_mov_b32(s[0], 0xFFFFFFFF),  # All lanes active
+      s_quadmask_b32(s[1], s[0]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    # Each quad (4 lanes) with any bit set -> 1 bit in result
+    # 32 lanes = 8 quads, all active -> 0xFF
+    self.assertEqual(st.sgpr[1], 0xFF)
+
+  def test_s_quadmask_b32_alternating_quads(self):
+    """S_QUADMASK_B32 with alternating quads active."""
+    instructions = [
+      s_mov_b32(s[0], 0x0F0F0F0F),  # Quads 0,2,4,6 active
+      s_quadmask_b32(s[1], s[0]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    # Quads 0,2,4,6 have at least one bit -> 0b01010101 = 0x55
+    self.assertEqual(st.sgpr[1], 0x55)
+
+  def test_s_quadmask_b32_no_quads_active(self):
+    """S_QUADMASK_B32 with no quads active."""
+    instructions = [
+      s_mov_b32(s[0], 0),
+      s_quadmask_b32(s[1], s[0]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[1], 0)
+
+  def test_s_quadmask_b32_single_lane_per_quad(self):
+    """S_QUADMASK_B32 with single lane active in each quad."""
+    instructions = [
+      s_mov_b32(s[0], 0x11111111),  # Bit 0 of each nibble
+      s_quadmask_b32(s[1], s[0]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    # All 8 quads have at least one lane -> 0xFF
+    self.assertEqual(st.sgpr[1], 0xFF)
+
+  def test_s_wqm_b32_all_active(self):
+    """S_WQM_B32 with all lanes active returns all 1s."""
+    instructions = [
+      s_mov_b32(s[0], 0xFFFFFFFF),
+      s_wqm_b32(s[1], s[0]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[1], 0xFFFFFFFF)
+
+  def test_s_wqm_b32_alternating_quads(self):
+    """S_WQM_B32 with single lane per quad expands to full quads."""
+    instructions = [
+      s_mov_b32(s[0], 0x11111111),  # One lane per quad
+      s_wqm_b32(s[1], s[0]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    # Each quad with any bit expands to all 4 bits
+    self.assertEqual(st.sgpr[1], 0xFFFFFFFF)
+
+  def test_s_wqm_b32_zero(self):
+    """S_WQM_B32 with zero input returns zero."""
+    instructions = [
+      s_mov_b32(s[0], 0),
+      s_wqm_b32(s[1], s[0]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[1], 0)
+
+
+class TestBranch(unittest.TestCase):
+  """Tests for branch instructions."""
+
+  def test_cbranch_vccnz_ignores_vcc_hi(self):
+    """S_CBRANCH_VCCNZ should only check VCC_LO in wave32."""
+    instructions = [
+      # Set VCC_LO = 0, VCC_HI = 1
+      s_mov_b32(VCC_LO, 0),
+      s_mov_b32(VCC_HI, 1),
+      v_mov_b32_e32(v[0], 0),
+      # If VCC_HI is incorrectly used, branch will be taken
+      s_cbranch_vccnz(1),  # Skip next instruction if VCC != 0
+      v_mov_b32_e32(v[0], 42),  # This should execute
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 42, "Branch should NOT be taken (VCC_LO is 0)")
+
+  def test_cbranch_vccz_ignores_vcc_hi(self):
+    """S_CBRANCH_VCCZ should only check VCC_LO in wave32."""
+    instructions = [
+      # Set VCC_LO = 1, VCC_HI = 0
+      s_mov_b32(VCC_LO, 1),
+      s_mov_b32(VCC_HI, 0),
+      v_mov_b32_e32(v[0], 0),
+      # If VCC_HI is incorrectly used, branch will be taken
+      s_cbranch_vccz(1),  # Skip next instruction if VCC == 0
+      v_mov_b32_e32(v[0], 42),  # This should execute
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 42, "Branch should NOT be taken (VCC_LO is 1)")
+
+  def test_cbranch_vccnz_branches_on_vcc_lo(self):
+    """S_CBRANCH_VCCNZ branches when VCC_LO is non-zero."""
+    instructions = [
+      s_mov_b32(VCC_LO, 1),
+      v_mov_b32_e32(v[0], 0),
+      s_cbranch_vccnz(1),  # Skip next instruction if VCC != 0
+      v_mov_b32_e32(v[0], 42),  # This should be skipped
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 0, "Branch should be taken (VCC_LO is 1)")
+
+
+class Test64BitLiterals(unittest.TestCase):
+  """Tests for 64-bit literal encoding in instructions."""
+
+  def test_64bit_literal_negative_encoding(self):
+    """64-bit literal -2^32 encodes correctly."""
+    lit = -4294967296.0  # -2^32
+    lit_bits = f2i64(lit)
+    instructions = [
+      s_mov_b32(s[0], lit_bits & 0xffffffff),
+      s_mov_b32(s[1], lit_bits >> 32),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = i642f(st.vgpr[0][0] | (st.vgpr[0][1] << 32))
+    self.assertAlmostEqual(result, -4294967296.0, places=5)
+
+class TestSCCBehavior(unittest.TestCase):
+  """Tests for SCC condition code behavior."""
+
+  def test_scc_from_s_cmp(self):
+    """SCC should be set by scalar compare."""
+    instructions = [
+      s_mov_b32(s[0], 10),
+      s_cmp_eq_u32(s[0], 10),
+      s_cselect_b32(s[1], 1, 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[1], 1, "SCC should be true")
+    self.assertEqual(st.scc, 1)
+
+  def test_scc_clear(self):
+    """SCC should be cleared by failing compare."""
+    instructions = [
+      s_mov_b32(s[0], 10),
+      s_cmp_eq_u32(s[0], 20),
+      s_cselect_b32(s[1], 1, 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[1], 0, "SCC should be false")
+    self.assertEqual(st.scc, 0)
+
+
+class TestSignedArithmetic(unittest.TestCase):
+  """Tests for S_ADD_I32, S_SUB_I32 and their SCC overflow behavior."""
+
+  def test_s_add_i32_no_overflow(self):
+    """S_ADD_I32: 1 + 1 = 2, no overflow, SCC=0."""
+    instructions = [
+      s_mov_b32(s[0], 1),
+      s_add_i32(s[1], s[0], 1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[1], 2)
+    self.assertEqual(st.scc, 0, "No overflow, SCC should be 0")
+
+  def test_s_add_i32_positive_overflow(self):
+    """S_ADD_I32: MAX_INT + 1 overflows, SCC=1."""
+    instructions = [
+      s_mov_b32(s[0], 0x7FFFFFFF),  # MAX_INT
+      s_add_i32(s[1], s[0], 1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[1], 0x80000000)  # Wraps to MIN_INT
+    self.assertEqual(st.scc, 1, "Overflow, SCC should be 1")
+
+  def test_s_add_i32_negative_no_overflow(self):
+    """S_ADD_I32: -10 + 20 = 10, no overflow."""
+    instructions = [
+      s_mov_b32(s[0], 0xFFFFFFF6),  # -10 in two's complement
+      s_mov_b32(s[1], 20),
+      s_add_i32(s[2], s[0], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[2], 10)
+    self.assertEqual(st.scc, 0)
+
+  def test_s_add_i32_negative_overflow(self):
+    """S_ADD_I32: MIN_INT + (-1) underflows, SCC=1."""
+    instructions = [
+      s_mov_b32(s[0], 0x80000000),  # MIN_INT
+      s_mov_b32(s[1], 0xFFFFFFFF),  # -1
+      s_add_i32(s[2], s[0], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[2], 0x7FFFFFFF)  # Wraps to MAX_INT
+    self.assertEqual(st.scc, 1, "Underflow, SCC should be 1")
+
+  def test_s_sub_i32_no_overflow(self):
+    """S_SUB_I32: 10 - 5 = 5, no overflow."""
+    instructions = [
+      s_mov_b32(s[0], 10),
+      s_mov_b32(s[1], 5),
+      s_sub_i32(s[2], s[0], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[2], 5)
+    self.assertEqual(st.scc, 0)
+
+  def test_s_sub_i32_overflow(self):
+    """S_SUB_I32: MAX_INT - (-1) overflows, SCC=1."""
+    instructions = [
+      s_mov_b32(s[0], 0x7FFFFFFF),  # MAX_INT
+      s_mov_b32(s[1], 0xFFFFFFFF),  # -1
+      s_sub_i32(s[2], s[0], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[2], 0x80000000)  # Wraps to MIN_INT
+    self.assertEqual(st.scc, 1, "Overflow, SCC should be 1")
+
+  def test_s_mul_hi_u32(self):
+    """S_MUL_HI_U32: high 32 bits of u32 * u32."""
+    instructions = [
+      s_mov_b32(s[0], 0x80000000),  # 2^31
+      s_mov_b32(s[1], 4),
+      s_mul_hi_u32(s[2], s[0], s[1]),  # (2^31 * 4) >> 32 = 2
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[2], 2)
+
+  def test_s_mul_hi_u32_max(self):
+    """S_MUL_HI_U32: 0xFFFFFFFF * 0xFFFFFFFF."""
+    instructions = [
+      s_mov_b32(s[0], 0xFFFFFFFF),
+      s_mov_b32(s[1], 0xFFFFFFFF),
+      s_mul_hi_u32(s[2], s[0], s[1]),  # (0xFFFFFFFF * 0xFFFFFFFF) >> 32 = 0xFFFFFFFE
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[2], 0xFFFFFFFE)
+
+  def test_s_mul_hi_i32_positive(self):
+    """S_MUL_HI_I32: positive * positive."""
+    instructions = [
+      s_mov_b32(s[0], 0x40000000),  # 2^30
+      s_mov_b32(s[1], 4),
+      s_mul_hi_i32(s[2], s[0], s[1]),  # (2^30 * 4) >> 32 = 1
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[2], 1)
+
+  def test_s_mul_hi_i32_neg_times_neg(self):
+    """S_MUL_HI_I32: (-1) * (-1) = 1, high bits = 0."""
+    instructions = [
+      s_mov_b32(s[0], 0xFFFFFFFF),  # -1
+      s_mov_b32(s[1], 0xFFFFFFFF),  # -1
+      s_mul_hi_i32(s[2], s[0], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[2], 0)
+
+  def test_s_mul_hi_i32_neg_times_pos(self):
+    """S_MUL_HI_I32: (-1) * 2 = -2, high bits = -1 (sign extension)."""
+    instructions = [
+      s_mov_b32(s[0], 0xFFFFFFFF),  # -1
+      s_mov_b32(s[1], 2),
+      s_mul_hi_i32(s[2], s[0], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[2], 0xFFFFFFFF)  # -1 sign extends
+
+  def test_s_mul_hi_i32_min_int(self):
+    """S_MUL_HI_I32: MIN_INT * 2 = -2^32, high = -1."""
+    instructions = [
+      s_mov_b32(s[0], 0x80000000),  # -2^31 (MIN_INT)
+      s_mov_b32(s[1], 2),
+      s_mul_hi_i32(s[2], s[0], s[1]),  # (-2^31 * 2) >> 32 = -1
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[2], 0xFFFFFFFF)
+
+  def test_s_mul_i32(self):
+    """S_MUL_I32: signed multiply low 32 bits."""
+    instructions = [
+      s_mov_b32(s[0], 0xFFFFFFFF),  # -1
+      s_mov_b32(s[1], 10),
+      s_mul_i32(s[2], s[0], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[2], 0xFFFFFFF6)  # -10
+
+  def test_division_sequence_from_llvm(self):
+    """Test the division sequence pattern from LLVM-generated code."""
+    # This sequence is from the sin kernel and computes integer division
+    # s10 = dividend, s18 = divisor, result in s6/s14
+    dividend = 0x28BE60DB  # Some value from the sin kernel
+    divisor = 3  # Simplified divisor
+    instructions = [
+      s_mov_b32(s[10], dividend),
+      s_mov_b32(s[18], divisor),
+      # Compute reciprocal approximation: s6 = ~0 / divisor (approx)
+      s_mov_b32(s[11], 0),
+      s_sub_i32(s[11], s[11], s[18]),  # s11 = -divisor
+      # For testing, just verify basic arithmetic works
+      s_mul_i32(s[6], s[10], 2),
+      s_add_i32(s[7], s[6], 1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[6], (dividend * 2) & 0xFFFFFFFF)
+    self.assertEqual(st.sgpr[7], ((dividend * 2) + 1) & 0xFFFFFFFF)
+
+
+class TestBitSet(unittest.TestCase):
+  """Tests for S_BITSET0_B32 and S_BITSET1_B32 instructions."""
+
+  def test_s_bitset1_b32_set_bit0(self):
+    """S_BITSET1_B32: set bit 0 in destination."""
+    instructions = [
+      s_mov_b32(s[0], 0),     # start with 0
+      s_mov_b32(s[1], 0),     # bit position = 0
+      s_bitset1_b32(s[0], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[0], 1, "Bit 0 should be set")
+
+  def test_s_bitset1_b32_set_bit31(self):
+    """S_BITSET1_B32: set bit 31 in destination."""
+    instructions = [
+      s_mov_b32(s[0], 0),     # start with 0
+      s_mov_b32(s[1], 31),    # bit position = 31
+      s_bitset1_b32(s[0], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[0], 0x80000000, "Bit 31 should be set")
+
+  def test_s_bitset1_b32_preserves_other_bits(self):
+    """S_BITSET1_B32: preserves bits not being set."""
+    instructions = [
+      s_mov_b32(s[0], 0xFF00FF00),  # existing pattern
+      s_mov_b32(s[1], 0),            # bit position = 0
+      s_bitset1_b32(s[0], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[0], 0xFF00FF01, "Should set bit 0 while preserving others")
+
+  def test_s_bitset0_b32_clear_bit0(self):
+    """S_BITSET0_B32: clear bit 0 in destination."""
+    instructions = [
+      s_mov_b32(s[0], 0xFFFFFFFF),  # start with all bits set
+      s_mov_b32(s[1], 0),            # bit position = 0
+      s_bitset0_b32(s[0], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[0], 0xFFFFFFFE, "Bit 0 should be cleared")
+
+  def test_s_bitset0_b32_clear_bit31(self):
+    """S_BITSET0_B32: clear bit 31 in destination."""
+    instructions = [
+      s_mov_b32(s[0], 0xFFFFFFFF),  # start with all bits set
+      s_mov_b32(s[1], 31),           # bit position = 31
+      s_bitset0_b32(s[0], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[0], 0x7FFFFFFF, "Bit 31 should be cleared")
+
+  def test_s_bitset1_b32_uses_low5_bits(self):
+    """S_BITSET1_B32: only uses low 5 bits of position (mod 32)."""
+    instructions = [
+      s_mov_b32(s[0], 0),
+      s_mov_b32(s[1], 32 + 5),   # position = 37, but mod 32 = 5
+      s_bitset1_b32(s[0], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[0], 0x20, "Bit 5 should be set (37 mod 32 = 5)")
+
+
+class TestBfeI64(unittest.TestCase):
+  """Tests for S_BFE_I64 - 64-bit bit field extract with sign extension.
+
+  Regression tests for sign extension bug where 32-bit masks were incorrectly
+  used for 64-bit operations, causing the high 32 bits to not be sign-extended.
+  """
+
+  def test_s_bfe_i64_positive_no_sign_extend(self):
+    """S_BFE_I64: positive value (1) in 16 bits should not sign extend."""
+    # S1 encodes: [22:16] = width, [5:0] = offset
+    # width=16, offset=0 -> S1 = (16 << 16) | 0 = 0x100000
+    instructions = [
+      s_mov_b32(s[0], 1),         # S0 lo = 1
+      s_mov_b32(s[1], 0),         # S0 hi = 0
+      s_mov_b32(s[2], 0x100000),  # width=16, offset=0
+      s_bfe_i64(s[4:5], s[0:1], s[2]),
+      v_mov_b32_e32(v[0], s[4]),
+      v_mov_b32_e32(v[1], s[5]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 1, "lo should be 1")
+    self.assertEqual(st.vgpr[0][1], 0, "hi should be 0 (no sign extend)")
+
+  def test_s_bfe_i64_negative_sign_extend(self):
+    """S_BFE_I64: 0xFFFF (-1 in 16 bits) should sign extend to 64 bits.
+
+    This is the main regression test - before the fix, hi was 0 instead of 0xFFFFFFFF.
+    """
+    instructions = [
+      s_mov_b32(s[0], 0xFFFF),    # S0 lo = -1 in 16 bits
+      s_mov_b32(s[1], 0),         # S0 hi = 0
+      s_mov_b32(s[2], 0x100000),  # width=16, offset=0
+      s_bfe_i64(s[4:5], s[0:1], s[2]),
+      v_mov_b32_e32(v[0], s[4]),
+      v_mov_b32_e32(v[1], s[5]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 0xFFFFFFFF, "lo should be 0xFFFFFFFF")
+    self.assertEqual(st.vgpr[0][1], 0xFFFFFFFF, "hi should be 0xFFFFFFFF (sign extended)")
+
+  def test_s_bfe_i64_8bit_negative_sign_extend(self):
+    """S_BFE_I64: 0xFF (-1 in 8 bits) should sign extend to 64 bits."""
+    # width=8, offset=0 -> S1 = (8 << 16) | 0 = 0x80000
+    instructions = [
+      s_mov_b32(s[0], 0xFF),      # S0 lo = -1 in 8 bits
+      s_mov_b32(s[1], 0),         # S0 hi = 0
+      s_mov_b32(s[2], 0x80000),   # width=8, offset=0
+      s_bfe_i64(s[4:5], s[0:1], s[2]),
+      v_mov_b32_e32(v[0], s[4]),
+      v_mov_b32_e32(v[1], s[5]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 0xFFFFFFFF, "lo should be 0xFFFFFFFF")
+    self.assertEqual(st.vgpr[0][1], 0xFFFFFFFF, "hi should be 0xFFFFFFFF (sign extended)")
+
+  def test_s_bfe_i64_8bit_positive(self):
+    """S_BFE_I64: 0x7F (127 in 8 bits) should not sign extend."""
+    # width=8, offset=0 -> S1 = (8 << 16) | 0 = 0x80000
+    instructions = [
+      s_mov_b32(s[0], 0x7F),      # S0 lo = 127 in 8 bits (MSB=0)
+      s_mov_b32(s[1], 0),         # S0 hi = 0
+      s_mov_b32(s[2], 0x80000),   # width=8, offset=0
+      s_bfe_i64(s[4:5], s[0:1], s[2]),
+      v_mov_b32_e32(v[0], s[4]),
+      v_mov_b32_e32(v[1], s[5]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 0x7F, "lo should be 0x7F")
+    self.assertEqual(st.vgpr[0][1], 0, "hi should be 0 (no sign extend)")
+
+  def test_s_bfe_i64_with_offset(self):
+    """S_BFE_I64: extract from non-zero bit offset with sign extension."""
+    # Extract 16 bits starting at bit 8: value 0xFF00 >> 8 = 0xFF = -1 in 8 bits? No wait...
+    # Let's put 0x8000FF00: extract 16 bits at offset 8 = 0x00FF (positive)
+    # Put 0xFF00_0000: extract 16 bits at offset 16 = 0xFF00 = -256 in signed 16-bit
+    instructions = [
+      s_mov_b32(s[0], 0xFF000000),  # bits [31:24] = 0xFF, [23:16] = 0x00
+      s_mov_b32(s[1], 0),
+      # width=16, offset=16 -> S1 = (16 << 16) | 16 = 0x100010
+      s_mov_b32(s[2], 0x100010),
+      s_bfe_i64(s[4:5], s[0:1], s[2]),
+      v_mov_b32_e32(v[0], s[4]),
+      v_mov_b32_e32(v[1], s[5]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    # Extract bits [31:16] = 0xFF00, sign bit is bit 15 of extracted = bit 31 of original = 1
+    # So result should be sign-extended 0xFF00 -> 0xFFFFFF00 in lo, 0xFFFFFFFF in hi
+    self.assertEqual(st.vgpr[0][0], 0xFFFFFF00, "lo should be sign-extended 0xFF00")
+    self.assertEqual(st.vgpr[0][1], 0xFFFFFFFF, "hi should be 0xFFFFFFFF (sign extended)")
+
+  def test_s_bfe_i64_32bit_negative(self):
+    """S_BFE_I64: extract 32 bits with sign extension."""
+    # width=32, offset=0 -> S1 = (32 << 16) | 0 = 0x200000
+    instructions = [
+      s_mov_b32(s[0], 0x80000000),  # MIN_INT32 = -2^31
+      s_mov_b32(s[1], 0),
+      s_mov_b32(s[2], 0x200000),    # width=32, offset=0
+      s_bfe_i64(s[4:5], s[0:1], s[2]),
+      v_mov_b32_e32(v[0], s[4]),
+      v_mov_b32_e32(v[1], s[5]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 0x80000000, "lo should be 0x80000000")
+    self.assertEqual(st.vgpr[0][1], 0xFFFFFFFF, "hi should be 0xFFFFFFFF (sign extended)")
+
+
+class Test64BitCompare(unittest.TestCase):
+  """Tests for 64-bit scalar compare instructions."""
+
+  def test_s_cmp_eq_u64_equal(self):
+    """S_CMP_EQ_U64: comparing equal 64-bit values sets SCC=1."""
+    val = 0x123456789ABCDEF0
+    instructions = [
+      s_mov_b32(s[0], val & 0xFFFFFFFF),
+      s_mov_b32(s[1], val >> 32),
+      s_mov_b32(s[2], val & 0xFFFFFFFF),
+      s_mov_b32(s[3], val >> 32),
+      s_cmp_eq_u64(s[0:1], s[2:3]),
+      s_cselect_b32(s[4], 1, 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.scc, 1)
+    self.assertEqual(st.sgpr[4], 1)
+
+  def test_s_cmp_eq_u64_different_upper_bits(self):
+    """S_CMP_EQ_U64: values differing only in upper 32 bits are not equal."""
+    # This is the bug case - if only lower 32 bits are compared, these would be equal
+    instructions = [
+      s_mov_b32(s[0], 0),  # lower 32 bits of value 0
+      s_mov_b32(s[1], 0),  # upper 32 bits of value 0
+      s_mov_b32(s[2], 0),  # lower 32 bits of 0x100000000
+      s_mov_b32(s[3], 1),  # upper 32 bits of 0x100000000
+      s_cmp_eq_u64(s[0:1], s[2:3]),
+      s_cselect_b32(s[4], 1, 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.scc, 0, "0 != 0x100000000, SCC should be 0")
+    self.assertEqual(st.sgpr[4], 0)
+
+  def test_s_cmp_lg_u64_different(self):
+    """S_CMP_LG_U64: different 64-bit values sets SCC=1."""
+    instructions = [
+      s_mov_b32(s[0], 0),
+      s_mov_b32(s[1], 0),  # s[0:1] = 0
+      s_mov_b32(s[2], 0),
+      s_mov_b32(s[3], 1),  # s[2:3] = 0x100000000
+      s_cmp_lg_u64(s[0:1], s[2:3]),
+      s_cselect_b32(s[4], 1, 0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.scc, 1, "0 != 0x100000000, SCC should be 1")
+    self.assertEqual(st.sgpr[4], 1)
+
+
+class TestSOPPNop(unittest.TestCase):
+  """Tests for S_NOP and other SOPP instructions with expression-based for loops.
+
+  S_NOP's pcode uses 'for i in 0U : SIMM16.u16[3 : 0].u32 do' which requires
+  the parser to handle non-constant loop bounds.
+  """
+
+  def test_s_nop_basic(self):
+    """S_NOP executes without side effects."""
+    # S_NOP with immediate 0 should just do nothing
+    instructions = [
+      s_mov_b32(s[0], 42),
+      s_nop(0),  # nop with simm16=0
+      s_mov_b32(s[1], 100),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[0], 42)
+    self.assertEqual(st.sgpr[1], 100)
+
+  def test_s_nop_with_count(self):
+    """S_NOP with count parameter executes multiple nops."""
+    # S_NOP with immediate 3 should execute 4 nops (0:3 inclusive)
+    instructions = [
+      s_mov_b32(s[0], 1),
+      s_nop(3),  # nop with simm16=3 -> 4 iterations
+      s_add_u32(s[0], s[0], 1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[0], 2)
+
+
+class TestNullRegister(unittest.TestCase):
+  """Tests for NULL register (124) behavior - writes should be discarded, reads return 0."""
+
+  def test_s_mov_b32_from_null(self):
+    """S_MOV_B32 from NULL should read as 0."""
+    instructions = [
+      s_mov_b32(s[0], 0xDEADBEEF),  # Set s[0] to sentinel
+      s_mov_b32(s[0], NULL),  # Read from NULL - should be 0
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[0], 0)
+
+  def test_s_add_u32_with_null_src(self):
+    """S_ADD_U32 with NULL as source should use 0."""
+    instructions = [
+      s_mov_b32(s[0], 100),
+      s_add_u32(s[1], s[0], NULL),  # 100 + 0 = 100
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[1], 100)
+
+  def test_s_mov_b32_to_null(self):
+    """S_MOV_B32 to NULL (sdst=124) should discard the write."""
+    instructions = [
+      s_mov_b32(s[0], 0xDEADBEEF),  # Set s[0] to sentinel
+      s_mov_b32(NULL, 42),  # Write to NULL - should be discarded
+      # s[0] should still be 0xDEADBEEF since NULL write doesn't affect it
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[0], 0xDEADBEEF)
+
+  def test_s_add_u32_to_null(self):
+    """S_ADD_U32 with sdst=NULL should discard result but still set SCC."""
+    instructions = [
+      s_mov_b32(s[0], 0xFFFFFFFF),
+      s_mov_b32(s[1], 1),
+      s_add_u32(NULL, s[0], s[1]),  # overflow, write to NULL
+      s_cselect_b32(s[2], 1, 0),  # capture SCC
+    ]
+    st = run_program(instructions, n_lanes=1)
+    # SCC should still be set from overflow even though result was discarded
+    self.assertEqual(st.sgpr[2], 1)
+    self.assertEqual(st.scc, 1)
+
+  def test_s_and_b32_to_null(self):
+    """S_AND_B32 with sdst=NULL should discard result but still set SCC."""
+    instructions = [
+      s_mov_b32(s[0], 0xFF00FF00),
+      s_mov_b32(s[1], 0x0F0F0F0F),
+      s_and_b32(NULL, s[0], s[1]),  # result=0x0F000F00, non-zero so SCC=1
+      s_cselect_b32(s[2], 1, 0),  # capture SCC
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[2], 1)  # SCC=1 because result was non-zero
+    self.assertEqual(st.scc, 1)
+
+  def test_s_or_b32_to_null_zero_result(self):
+    """S_OR_B32 with sdst=NULL and zero result should set SCC=0."""
+    instructions = [
+      s_mov_b32(s[0], 0),
+      s_mov_b32(s[1], 0),
+      s_or_b32(NULL, s[0], s[1]),  # result=0, so SCC=0
+      s_cselect_b32(s[2], 1, 0),  # capture SCC
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[2], 0)  # SCC=0 because result was zero
+    self.assertEqual(st.scc, 0)
+
+
+class Test64BitSOP1InlineConstants(unittest.TestCase):
+  """Tests for 64-bit SOP1 instructions with inline constants.
+
+  Regression tests for bug where rsrc_dyn didn't properly handle 64-bit
+  inline constants, incorrectly duplicating lo bits to hi instead of
+  zero/sign-extending.
+  """
+
+  def test_s_mov_b64_inline_0(self):
+    """S_MOV_B64 with inline constant 0."""
+    instructions = [
+      s_mov_b64(s[0:1], 0),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 0)
+    self.assertEqual(st.vgpr[0][1], 0)
+
+  def test_s_mov_b64_inline_16(self):
+    """S_MOV_B64 with inline constant 16 should set lo=16, hi=0."""
+    instructions = [
+      s_mov_b64(s[0:1], 16),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 16)
+    self.assertEqual(st.vgpr[0][1], 0)
+
+  def test_s_mov_b64_inline_64(self):
+    """S_MOV_B64 with inline constant 64 (max positive)."""
+    instructions = [
+      s_mov_b64(s[0:1], 64),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 64)
+    self.assertEqual(st.vgpr[0][1], 0)
+
+  def test_s_mov_b64_inline_neg1(self):
+    """S_MOV_B64 with inline constant -1 should sign-extend."""
+    instructions = [
+      s_mov_b64(s[0:1], -1),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 0xFFFFFFFF)
+    self.assertEqual(st.vgpr[0][1], 0xFFFFFFFF)
+
+  def test_s_mov_b64_inline_neg16(self):
+    """S_MOV_B64 with inline constant -16 should sign-extend."""
+    instructions = [
+      s_mov_b64(s[0:1], -16),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 0xFFFFFFF0)
+    self.assertEqual(st.vgpr[0][1], 0xFFFFFFFF)
+
+  def test_s_mov_b64_float_const_1_0(self):
+    """S_MOV_B64 with float inline constant 1.0 - casts F32 to F64."""
+    instructions = [
+      s_mov_b64(s[0:1], 1.0),  # inline constant 242 (1.0f)
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    # Hardware casts F32 to F64: 1.0f64 = 0x3FF0000000000000
+    self.assertEqual(st.vgpr[0][0], 0x00000000)  # lo
+    self.assertEqual(st.vgpr[0][1], 0x3FF00000)  # hi
+
+  def test_s_or_b64_inline_constant(self):
+    """S_OR_B64 with 64-bit inline constant."""
+    instructions = [
+      s_mov_b64(s[0:1], 0),
+      s_or_b64(s[2:3], s[0:1], 16),
+      v_mov_b32_e32(v[0], s[2]),
+      v_mov_b32_e32(v[1], s[3]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 16)
+    self.assertEqual(st.vgpr[0][1], 0)
+
+  def test_s_and_b64_inline_constant(self):
+    """S_AND_B64 with 64-bit inline constant."""
+    instructions = [
+      s_mov_b32(s[0], 0xFFFFFFFF),
+      s_mov_b32(s[1], 0xFFFFFFFF),
+      s_and_b64(s[2:3], s[0:1], 16),
+      v_mov_b32_e32(v[0], s[2]),
+      v_mov_b32_e32(v[1], s[3]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 16)
+    self.assertEqual(st.vgpr[0][1], 0)
+
+
+class Test64BitSOPLiterals(unittest.TestCase):
+  """Tests for 64-bit SOP instructions with 32-bit literals.
+
+  Tests the behavior when a 64-bit SOP instruction uses a 32-bit literal
+  (offset 255 in instruction encoding). The literal is zero-extended to 64 bits.
+  """
+
+  def test_s_mov_b64_literal(self):
+    """S_MOV_B64 with 32-bit literal value - zero-extended to 64 bits."""
+    instructions = [
+      s_mov_b64(s[0:1], 0x12345678),  # literal > 64, uses literal encoding
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 0x12345678)
+    self.assertEqual(st.vgpr[0][1], 0)
+
+  def test_s_or_b64_literal(self):
+    """S_OR_B64 with 32-bit literal value - zero-extended to 64 bits."""
+    instructions = [
+      s_mov_b64(s[0:1], 0),
+      s_or_b64(s[2:3], s[0:1], 0x12345678),  # literal
+      v_mov_b32_e32(v[0], s[2]),
+      v_mov_b32_e32(v[1], s[3]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 0x12345678)
+    self.assertEqual(st.vgpr[0][1], 0)
+
+  def test_s_and_b64_literal(self):
+    """S_AND_B64 with 32-bit literal value - zero-extended to 64 bits."""
+    instructions = [
+      s_mov_b32(s[0], 0xFFFFFFFF),
+      s_mov_b32(s[1], 0xFFFFFFFF),
+      s_and_b64(s[2:3], s[0:1], 0x12345678),  # literal
+      v_mov_b32_e32(v[0], s[2]),
+      v_mov_b32_e32(v[1], s[3]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 0x12345678)
+    self.assertEqual(st.vgpr[0][1], 0)
+
+  def test_s_mov_b64_literal_negative(self):
+    """S_MOV_B64 with 0xFFFFFFFF literal - zero-extended (not sign-extended)."""
+    instructions = [
+      s_mov_b64(s[0:1], 0xFFFFFFFF),  # -1 as 32-bit, but zero-extended to 64-bit
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 0xFFFFFFFF)
+    self.assertEqual(st.vgpr[0][1], 0)  # zero-extended, not sign-extended
+
+  def test_s_mov_b64_literal_high_bit(self):
+    """S_MOV_B64 with 0x80000000 literal - zero-extended (not sign-extended)."""
+    instructions = [
+      s_mov_b64(s[0:1], 0x80000000),  # high bit set, but zero-extended
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 0x80000000)
+    self.assertEqual(st.vgpr[0][1], 0)  # zero-extended, not sign-extended
+
+
+if __name__ == '__main__':
+  unittest.main()

extra/assembly/amd/test/hw/test_vop2.py

@@ -0,0 +1,983 @@
+"""Tests for VOP2 instructions - two operand vector operations.
+
+Includes: v_add_f32, v_mul_f32, v_and_b32, v_or_b32, v_xor_b32,
+          v_lshrrev_b32, v_lshlrev_b32, v_fmac_f32, v_fmaak_f32, v_fmamk_f32,
+          v_add_nc_u32, v_cndmask_b32, v_add_f16, v_mul_f16
+"""
+import unittest
+from extra.assembly.amd.test.hw.helpers import *
+
+class TestBasicArithmetic(unittest.TestCase):
+  """Tests for basic arithmetic VOP2 instructions."""
+
+  def test_v_add_f32(self):
+    """V_ADD_F32 adds two floats."""
+    instructions = [
+      v_mov_b32_e32(v[0], 1.0),
+      v_mov_b32_e32(v[1], 2.0),
+      v_add_f32_e32(v[2], v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertAlmostEqual(i2f(st.vgpr[0][2]), 3.0, places=5)
+
+  def test_v_mul_f32(self):
+    """V_MUL_F32 multiplies two floats."""
+    instructions = [
+      v_mov_b32_e32(v[0], 2.0),
+      v_mov_b32_e32(v[1], 4.0),
+      v_mul_f32_e32(v[2], v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertAlmostEqual(i2f(st.vgpr[0][2]), 8.0, places=5)
+
+  def test_v_fmac_f32(self):
+    """V_FMAC_F32: d = d + a*b using inline constants."""
+    instructions = [
+      v_mov_b32_e32(v[0], 2.0),
+      v_mov_b32_e32(v[1], 4.0),
+      v_mov_b32_e32(v[2], 1.0),
+      v_fmac_f32_e32(v[2], v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertAlmostEqual(i2f(st.vgpr[0][2]), 9.0, places=5)
+
+  def test_v_fmaak_f32(self):
+    """V_FMAAK_F32: d = a * b + K using inline constants."""
+    instructions = [
+      v_mov_b32_e32(v[0], 2.0),
+      v_mov_b32_e32(v[1], 4.0),
+      v_fmaak_f32_e32(v[2], v[0], v[1], literal=0x3f800000),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertAlmostEqual(i2f(st.vgpr[0][2]), 9.0, places=5)
+
+  def test_v_fmamk_f32_basic(self):
+    """V_FMAMK_F32: d = a * K + b."""
+    instructions = [
+      v_mov_b32_e32(v[0], 2.0),
+      v_mov_b32_e32(v[1], 1.0),
+      v_fmamk_f32_e32(v[2], v[0], v[1], literal=0x40800000),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertAlmostEqual(i2f(st.vgpr[0][2]), 9.0, places=5)
+
+  def test_v_fmamk_f32_small_constant(self):
+    """V_FMAMK_F32 with small constant."""
+    instructions = [
+      v_mov_b32_e32(v[0], 4.0),
+      v_mov_b32_e32(v[1], 1.0),
+      v_fmamk_f32_e32(v[2], v[0], v[1], literal=f2i(0.5)),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertAlmostEqual(i2f(st.vgpr[0][2]), 3.0, places=5)
+
+
+class TestBitManipulation(unittest.TestCase):
+  """Tests for bit manipulation VOP2 instructions."""
+
+  def test_v_and_b32(self):
+    """V_AND_B32 bitwise and."""
+    instructions = [
+      s_mov_b32(s[0], 0xff),
+      s_mov_b32(s[1], 0x0f),
+      v_mov_b32_e32(v[0], s[0]),
+      v_and_b32_e32(v[1], s[1], v[0]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][1], 0x0f)
+
+  def test_v_and_b32_quadrant(self):
+    """V_AND_B32 for quadrant extraction (n & 3)."""
+    instructions = [
+      s_mov_b32(s[0], 15915),
+      v_mov_b32_e32(v[0], s[0]),
+      v_and_b32_e32(v[1], 3, v[0]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][1], 15915 & 3)
+
+  def test_v_lshrrev_b32(self):
+    """V_LSHRREV_B32 logical shift right."""
+    instructions = [
+      s_mov_b32(s[0], 0xff00),
+      v_mov_b32_e32(v[0], s[0]),
+      v_lshrrev_b32_e32(v[1], 8, v[0]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][1], 0xff)
+
+  def test_v_lshlrev_b32(self):
+    """V_LSHLREV_B32 logical shift left."""
+    instructions = [
+      s_mov_b32(s[0], 0xff),
+      v_mov_b32_e32(v[0], s[0]),
+      v_lshlrev_b32_e32(v[1], 8, v[0]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][1], 0xff00)
+
+  def test_v_xor_b32(self):
+    """V_XOR_B32 bitwise xor (used in sin for sign)."""
+    instructions = [
+      s_mov_b32(s[0], 0x80000000),
+      s_mov_b32(s[1], f2i(1.0)),
+      v_mov_b32_e32(v[0], s[1]),
+      v_xor_b32_e32(v[1], s[0], v[0]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertAlmostEqual(i2f(st.vgpr[0][1]), -1.0, places=5)
+
+  def test_v_xor_b32_sign_flip(self):
+    """V_XOR_B32 for sign flip pattern."""
+    instructions = [
+      s_mov_b32(s[0], 0x80000000),
+      v_mov_b32_e32(v[0], -2.0),
+      v_xor_b32_e32(v[1], s[0], v[0]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertAlmostEqual(i2f(st.vgpr[0][1]), 2.0, places=5)
+
+
+class TestSpecialValues(unittest.TestCase):
+  """Tests for special float values - inf, nan, zero handling."""
+
+  def test_v_mul_f32_zero_times_inf(self):
+    """V_MUL_F32: 0 * inf = NaN."""
+    import math
+    instructions = [
+      v_mov_b32_e32(v[0], 0),
+      s_mov_b32(s[0], 0x7f800000),
+      v_mov_b32_e32(v[1], s[0]),
+      v_mul_f32_e32(v[2], v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertTrue(math.isnan(i2f(st.vgpr[0][2])))
+
+  def test_v_add_f32_inf_minus_inf(self):
+    """V_ADD_F32: inf + (-inf) = NaN."""
+    import math
+    instructions = [
+      s_mov_b32(s[0], 0x7f800000),
+      s_mov_b32(s[1], 0xff800000),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_add_f32_e32(v[2], v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertTrue(math.isnan(i2f(st.vgpr[0][2])))
+
+
+class TestF16Ops(unittest.TestCase):
+  """Tests for 16-bit VOP2 operations."""
+
+  def test_v_add_f16_basic(self):
+    """V_ADD_F16 adds two f16 values."""
+    instructions = [
+      s_mov_b32(s[0], 0x3c00),  # f16 1.0
+      s_mov_b32(s[1], 0x4000),  # f16 2.0
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_add_f16_e32(v[2], v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][2] & 0xffff
+    self.assertEqual(result, 0x4200, f"Expected 0x4200 (f16 3.0), got 0x{result:04x}")
+
+  def test_v_add_f16_negative(self):
+    """V_ADD_F16 with negative values."""
+    instructions = [
+      s_mov_b32(s[0], 0x3c00),  # f16 1.0
+      s_mov_b32(s[1], 0xc000),  # f16 -2.0
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_add_f16_e32(v[2], v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][2] & 0xffff
+    self.assertEqual(result, 0xbc00, f"Expected 0xbc00 (f16 -1.0), got 0x{result:04x}")
+
+  def test_v_mul_f16_basic(self):
+    """V_MUL_F16 multiplies two f16 values."""
+    instructions = [
+      s_mov_b32(s[0], 0x4000),  # f16 2.0
+      s_mov_b32(s[1], 0x4200),  # f16 3.0
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_mul_f16_e32(v[2], v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][2] & 0xffff
+    self.assertEqual(result, 0x4600, f"Expected 0x4600 (f16 6.0), got 0x{result:04x}")
+
+  def test_v_mul_f16_by_zero(self):
+    """V_MUL_F16 by zero."""
+    instructions = [
+      s_mov_b32(s[0], 0x4000),  # f16 2.0
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], 0),
+      v_mul_f16_e32(v[2], v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][2] & 0xffff
+    self.assertEqual(result, 0x0000, f"Expected 0x0000 (f16 0.0), got 0x{result:04x}")
+
+  def test_v_fmac_f16_basic(self):
+    """V_FMAC_F16: d = d + a*b."""
+    instructions = [
+      s_mov_b32(s[0], 0x4000),  # f16 2.0
+      s_mov_b32(s[1], 0x4200),  # f16 3.0
+      s_mov_b32(s[2], 0x3c00),  # f16 1.0
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_mov_b32_e32(v[2], s[2]),
+      v_fmac_f16_e32(v[2], v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][2] & 0xffff
+    # 2.0 * 3.0 + 1.0 = 7.0, f16 7.0 = 0x4700
+    self.assertEqual(result, 0x4700, f"Expected 0x4700 (f16 7.0), got 0x{result:04x}")
+
+  def test_v_max_f16_basic(self):
+    """V_MAX_F16 returns the maximum of two f16 values."""
+    instructions = [
+      s_mov_b32(s[0], 0x3c00),  # f16 1.0
+      s_mov_b32(s[1], 0x4000),  # f16 2.0
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_max_f16_e32(v[2], v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][2] & 0xffff
+    self.assertEqual(result, 0x4000, f"Expected 0x4000 (f16 2.0), got 0x{result:04x}")
+
+  def test_v_min_f16_basic(self):
+    """V_MIN_F16 returns the minimum of two f16 values."""
+    instructions = [
+      s_mov_b32(s[0], 0x3c00),  # f16 1.0
+      s_mov_b32(s[1], 0x4000),  # f16 2.0
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_min_f16_e32(v[2], v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][2] & 0xffff
+    self.assertEqual(result, 0x3c00, f"Expected 0x3c00 (f16 1.0), got 0x{result:04x}")
+
+  def test_v_fmaak_f16_basic(self):
+    """V_FMAAK_F16: d = a * b + K."""
+    instructions = [
+      s_mov_b32(s[0], 0x4000),  # f16 2.0
+      s_mov_b32(s[1], 0x4200),  # f16 3.0
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_fmaak_f16_e32(v[2], v[0], v[1], literal=0x3c00),  # + f16 1.0
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][2] & 0xffff
+    # 2.0 * 3.0 + 1.0 = 7.0, f16 7.0 = 0x4700
+    self.assertEqual(result, 0x4700, f"Expected 0x4700 (f16 7.0), got 0x{result:04x}")
+
+
+class TestHiHalfOps(unittest.TestCase):
+  """Tests for VOP2 16-bit operations with hi-half operands."""
+
+  def test_v_add_f16_src0_hi_fold(self):
+    """V_ADD_F16 with src0 hi-half fold (same register, different halves)."""
+    instructions = [
+      s_mov_b32(s[0], 0x40003c00),  # lo=f16(1.0), hi=f16(2.0)
+      v_mov_b32_e32(v[0], s[0]),
+      VOP3(VOP3Op.V_ADD_F16, vdst=v[1], src0=v[0], src1=v[0], opsel=0b0001),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][1] & 0xffff
+    self.assertEqual(result, 0x4200, f"Expected f16(3.0)=0x4200, got 0x{result:04x}")
+
+  def test_v_add_f16_src0_hi_different_reg(self):
+    """V_ADD_F16 with src0 hi-half from different register."""
+    instructions = [
+      s_mov_b32(s[0], 0x40000000),  # hi=f16(2.0), lo=0
+      s_mov_b32(s[1], 0x00003c00),  # hi=0, lo=f16(1.0)
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      VOP3(VOP3Op.V_ADD_F16, vdst=v[2], src0=v[0], src1=v[1], opsel=0b0001),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][2] & 0xffff
+    self.assertEqual(result, 0x4200, f"Expected f16(3.0)=0x4200, got 0x{result:04x}")
+
+  def test_v_mul_f16_src0_hi(self):
+    """V_MUL_F16 with src0 from high half."""
+    instructions = [
+      s_mov_b32(s[0], 0x40000000),  # hi=f16(2.0), lo=0
+      s_mov_b32(s[1], 0x00004200),  # hi=0, lo=f16(3.0)
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      VOP3(VOP3Op.V_MUL_F16, vdst=v[2], src0=v[0], src1=v[1], opsel=0b0001),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][2] & 0xffff
+    self.assertEqual(result, 0x4600, f"Expected f16(6.0)=0x4600, got 0x{result:04x}")
+
+  def test_v_mul_f16_hi_half(self):
+    """V_MUL_F16 reading from high half."""
+    instructions = [
+      s_mov_b32(s[0], 0x40003c00),  # lo=1.0, hi=2.0
+      v_mov_b32_e32(v[0], s[0]),
+      VOP3(VOP3Op.V_MUL_F16, vdst=v[1], src0=v[0], src1=v[0], opsel=0b0011),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][1] & 0xffff
+    self.assertEqual(result, 0x4400, f"Expected f16(4.0)=0x4400, got 0x{result:04x}")
+
+  def test_v_fma_f16_hi_dest(self):
+    """V_FMA_F16 writing to high half with opsel.
+
+    Uses V_FMA_F16 (not V_FMAC_F16) because it has explicit src2 operand
+    which makes opsel handling clearer.
+    """
+    instructions = [
+      s_mov_b32(s[0], 0x3c000000),  # hi=f16(1.0), lo=0
+      s_mov_b32(s[1], 0x4000),      # f16(2.0) in lo
+      s_mov_b32(s[2], 0x4200),      # f16(3.0) in lo
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_mov_b32_e32(v[2], s[2]),
+      # V_FMA_F16: dst = src0 * src1 + src2
+      # opsel=0b1100: bit2=src2 hi, bit3=dst hi
+      # So: v[0].hi = v[1].lo * v[2].lo + v[0].hi = 2.0 * 3.0 + 1.0 = 7.0
+      VOP3(VOP3Op.V_FMA_F16, vdst=v[0], src0=v[1], src1=v[2], src2=v[0], opsel=0b1100),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    hi = (st.vgpr[0][0] >> 16) & 0xffff
+    # 2.0 * 3.0 + 1.0 = 7.0, f16 7.0 = 0x4700
+    self.assertEqual(hi, 0x4700, f"Expected f16(7.0)=0x4700 in hi, got 0x{hi:04x}")
+
+  def test_v_add_f16_multilane(self):
+    """V_ADD_F16 with multiple lanes."""
+    instructions = [
+      s_mov_b32(s[0], 0x3c00),  # f16 1.0
+      s_mov_b32(s[1], 0x4000),  # f16 2.0
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_add_f16_e32(v[2], v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=4)
+    for lane in range(4):
+      result = st.vgpr[lane][2] & 0xffff
+      self.assertEqual(result, 0x4200, f"Lane {lane}: expected 0x4200, got 0x{result:04x}")
+
+
+class TestVop2F16HiHalf(unittest.TestCase):
+  """Regression tests for VOP2 f16 hi-half operand handling.
+
+  These test the bugs where:
+  1. VOP2 vsrc1 >= 384 (v[128]+) wasn't extracting hi 16 bits
+  2. VOP2 vdst >= 384 (v[128]+) wasn't preserving lo 16 bits
+  """
+
+  def test_v_add_f16_e32_vsrc1_hi_half(self):
+    """V_ADD_F16_E32 with vsrc1 from hi-half (v[128]+).
+
+    When vsrc1 >= 384 (representing v[128]+), the hardware reads from the hi 16 bits
+    of v[vsrc1-128]. The emulator must extract bits [31:16] from the actual VGPR.
+
+    Regression test for: VOP2 f16 vsrc1 hi-half extraction bug.
+    """
+    instructions = [
+      # v[0] = 0x4000_3c00: hi=f16(2.0), lo=f16(1.0)
+      s_mov_b32(s[0], 0x40003c00),
+      v_mov_b32_e32(v[0], s[0]),
+      # v_add_f16_e32 v[1], v[0], v[128]  (vsrc1=v[128] reads hi of v[0])
+      # In VOP2 encoding, vsrc1=384 means v[128], which maps to v[0].hi
+      # v[1] = v[0].lo + v[0].hi = 1.0 + 2.0 = 3.0
+      VOP2(VOP2Op.V_ADD_F16, vdst=v[1], src0=v[0], vsrc1=v[128]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][1] & 0xffff
+    # 1.0 + 2.0 = 3.0, f16 3.0 = 0x4200
+    self.assertEqual(result, 0x4200, f"Expected f16(3.0)=0x4200, got 0x{result:04x}")
+
+  def test_v_mul_f16_e32_vsrc1_hi_half(self):
+    """V_MUL_F16_E32 with vsrc1 from hi-half.
+
+    Regression test for: VOP2 f16 vsrc1 hi-half extraction bug.
+    """
+    instructions = [
+      # v[0] = 0x4200_4000: hi=f16(3.0), lo=f16(2.0)
+      s_mov_b32(s[0], 0x42004000),
+      v_mov_b32_e32(v[0], s[0]),
+      # v_mul_f16_e32 v[1], v[0], v[128]  (vsrc1=v[128] reads hi of v[0])
+      # v[1] = v[0].lo * v[0].hi = 2.0 * 3.0 = 6.0
+      VOP2(VOP2Op.V_MUL_F16, vdst=v[1], src0=v[0], vsrc1=v[128]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][1] & 0xffff
+    # 2.0 * 3.0 = 6.0, f16 6.0 = 0x4600
+    self.assertEqual(result, 0x4600, f"Expected f16(6.0)=0x4600, got 0x{result:04x}")
+
+  def test_v_add_f16_e32_vdst_hi_half(self):
+    """V_ADD_F16_E32 writing to hi-half destination (v[128]+).
+
+    When vdst >= 384 (representing v[128]+), the hardware writes to bits [31:16]
+    of v[vdst-128] while preserving bits [15:0]. The emulator must merge the result.
+
+    Regression test for: VOP2 f16 vdst hi-half write bug.
+    """
+    instructions = [
+      # v[0] = 0x0000_BEEF: lo has marker value
+      s_mov_b32(s[0], 0x0000BEEF),
+      v_mov_b32_e32(v[0], s[0]),
+      # v[1] = f16(1.0), v[2] = f16(2.0)
+      s_mov_b32(s[1], 0x3c00),
+      s_mov_b32(s[2], 0x4000),
+      v_mov_b32_e32(v[1], s[1]),
+      v_mov_b32_e32(v[2], s[2]),
+      # v_add_f16_e32 v[128], v[1], v[2]  (vdst=v[128] writes hi of v[0])
+      # v[0].hi = 1.0 + 2.0 = 3.0, v[0].lo preserved = 0xBEEF
+      VOP2(VOP2Op.V_ADD_F16, vdst=v[128], src0=v[1], vsrc1=v[2]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    hi = (st.vgpr[0][0] >> 16) & 0xffff
+    lo = st.vgpr[0][0] & 0xffff
+    # hi = 3.0 = 0x4200, lo preserved = 0xBEEF
+    self.assertEqual(hi, 0x4200, f"Expected hi=f16(3.0)=0x4200, got 0x{hi:04x}")
+    self.assertEqual(lo, 0xBEEF, f"Expected lo preserved=0xBEEF, got 0x{lo:04x}")
+
+  def test_v_mul_f16_e32_vdst_hi_half(self):
+    """V_MUL_F16_E32 writing to hi-half destination.
+
+    Regression test for: VOP2 f16 vdst hi-half write bug.
+    """
+    instructions = [
+      # v[0] = 0x0000_DEAD: lo has marker value
+      s_mov_b32(s[0], 0x0000DEAD),
+      v_mov_b32_e32(v[0], s[0]),
+      # v[1] = f16(2.0), v[2] = f16(4.0)
+      s_mov_b32(s[1], 0x4000),
+      s_mov_b32(s[2], 0x4400),
+      v_mov_b32_e32(v[1], s[1]),
+      v_mov_b32_e32(v[2], s[2]),
+      # v_mul_f16_e32 v[128], v[1], v[2]  (vdst=v[128] writes hi of v[0])
+      # v[0].hi = 2.0 * 4.0 = 8.0, v[0].lo preserved = 0xDEAD
+      VOP2(VOP2Op.V_MUL_F16, vdst=v[128], src0=v[1], vsrc1=v[2]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    hi = (st.vgpr[0][0] >> 16) & 0xffff
+    lo = st.vgpr[0][0] & 0xffff
+    # hi = 8.0 = 0x4800, lo preserved = 0xDEAD
+    self.assertEqual(hi, 0x4800, f"Expected hi=f16(8.0)=0x4800, got 0x{hi:04x}")
+    self.assertEqual(lo, 0xDEAD, f"Expected lo preserved=0xDEAD, got 0x{lo:04x}")
+
+  def test_v_add_f16_e32_both_hi_half(self):
+    """V_ADD_F16_E32 with both vsrc1 and vdst as hi-half (different underlying regs).
+
+    Tests the combination of both fixes: reading vsrc1 from hi-half AND
+    writing result to hi-half destination, using different underlying VGPRs.
+
+    Regression test for: VOP2 f16 hi-half bugs (combined).
+    """
+    instructions = [
+      # v[0] = 0x4000_xxxx: hi=f16(2.0) for vsrc1
+      s_mov_b32(s[0], 0x40000000),
+      v_mov_b32_e32(v[0], s[0]),
+      # v[1] = 0x0000_3c00: lo=f16(1.0) for src0
+      s_mov_b32(s[1], 0x00003c00),
+      v_mov_b32_e32(v[1], s[1]),
+      # v[2] = 0x0000_CAFE: lo=marker for vdst preservation
+      s_mov_b32(s[2], 0x0000CAFE),
+      v_mov_b32_e32(v[2], s[2]),
+      # v_add_f16_e32 v[130], v[1], v[128]
+      # src0 = v[1].lo = 1.0
+      # vsrc1 = v[128] reads v[0].hi = 2.0
+      # result = 1.0 + 2.0 = 3.0
+      # vdst = v[130] writes to v[2].hi, preserving v[2].lo
+      VOP2(VOP2Op.V_ADD_F16, vdst=v[130], src0=v[1], vsrc1=v[128]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    hi = (st.vgpr[0][2] >> 16) & 0xffff
+    lo = st.vgpr[0][2] & 0xffff
+    # hi = 3.0 = 0x4200, lo preserved = 0xCAFE
+    self.assertEqual(hi, 0x4200, f"Expected hi=f16(3.0)=0x4200, got 0x{hi:04x}")
+    self.assertEqual(lo, 0xCAFE, f"Expected lo preserved=0xCAFE, got 0x{lo:04x}")
+
+  def test_v_fmac_f16_e32_vsrc1_hi_half(self):
+    """V_FMAC_F16_E32 with vsrc1 from hi-half.
+
+    V_FMAC_F16: vdst = vdst + src0 * vsrc1
+
+    Regression test for: VOP2 f16 vsrc1 hi-half extraction bug.
+    """
+    instructions = [
+      # v[0] = 0x4000_3c00: hi=f16(2.0), lo=f16(1.0)
+      s_mov_b32(s[0], 0x40003c00),
+      v_mov_b32_e32(v[0], s[0]),
+      # v[1] = f16(3.0) = 0x4200
+      s_mov_b32(s[1], 0x4200),
+      v_mov_b32_e32(v[1], s[1]),
+      # v_fmac_f16_e32 v[1], v[0], v[128]
+      # vdst = v[1] = 3.0 + v[0].lo * v[0].hi = 3.0 + 1.0 * 2.0 = 5.0
+      VOP2(VOP2Op.V_FMAC_F16, vdst=v[1], src0=v[0], vsrc1=v[128]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][1] & 0xffff
+    # 3.0 + 1.0 * 2.0 = 5.0, f16 5.0 = 0x4500
+    self.assertEqual(result, 0x4500, f"Expected f16(5.0)=0x4500, got 0x{result:04x}")
+
+  def test_v_fmac_f16_e32_vdst_hi_half(self):
+    """V_FMAC_F16_E32 writing to hi-half destination.
+
+    V_FMAC_F16: vdst.h = vdst.h + src0 * vsrc1
+
+    When vdst is v[128]+, the accumulator D0 must also read from the hi-half.
+    This tests the bug where D0 was read from lo-half instead of hi-half.
+
+    Regression test for: VOP2 FMAC hi-half D0 accumulator read bug.
+    """
+    instructions = [
+      # v[0] = 0x3800_DEAD: hi=f16(0.5), lo=marker (0xDEAD)
+      s_mov_b32(s[0], 0x3800DEAD),
+      v_mov_b32_e32(v[0], s[0]),
+      # v[1] = f16(2.0) = 0x4000
+      s_mov_b32(s[1], 0x4000),
+      v_mov_b32_e32(v[1], s[1]),
+      # v[2] = f16(3.0) = 0x4200
+      s_mov_b32(s[2], 0x4200),
+      v_mov_b32_e32(v[2], s[2]),
+      # v_fmac_f16_e32 v[128], v[1], v[2]
+      # vdst = v[128] means v[0].hi
+      # D0 = v[0].hi = 0.5
+      # result = D0 + src0 * vsrc1 = 0.5 + 2.0 * 3.0 = 6.5
+      # v[0].hi = 6.5, v[0].lo preserved = 0xDEAD
+      VOP2(VOP2Op.V_FMAC_F16, vdst=v[128], src0=v[1], vsrc1=v[2]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    hi = (st.vgpr[0][0] >> 16) & 0xffff
+    lo = st.vgpr[0][0] & 0xffff
+    # hi = 6.5 = 0x4680, lo preserved = 0xDEAD
+    self.assertEqual(hi, 0x4680, f"Expected hi=f16(6.5)=0x4680, got 0x{hi:04x}")
+    self.assertEqual(lo, 0xDEAD, f"Expected lo preserved=0xDEAD, got 0x{lo:04x}")
+
+  def test_v_mul_f16_e32_src0_hi_half(self):
+    """V_MUL_F16_E32 with src0 from hi-half (src0 >= v[128]).
+
+    When src0 >= 384 (representing v[128]+), the hardware reads from the hi 16 bits
+    of v[src0-128]. The emulator must extract bits [31:16] from the actual VGPR.
+
+    Regression test for: VOP2 f16 src0 hi-half extraction bug.
+    """
+    instructions = [
+      # v[0] = 0x4000_3c00: hi=f16(2.0), lo=f16(1.0)
+      s_mov_b32(s[0], 0x40003c00),
+      v_mov_b32_e32(v[0], s[0]),
+      # v[1] = f16(3.0) = 0x4200
+      s_mov_b32(s[1], 0x4200),
+      v_mov_b32_e32(v[1], s[1]),
+      # v_mul_f16_e32 v[2], v[128], v[1]
+      # src0 = v[128] reads from v[0].hi = 2.0
+      # result = 2.0 * 3.0 = 6.0
+      VOP2(VOP2Op.V_MUL_F16, vdst=v[2], src0=v[128], vsrc1=v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][2] & 0xffff
+    # 2.0 * 3.0 = 6.0, f16 6.0 = 0x4600
+    self.assertEqual(result, 0x4600, f"Expected f16(6.0)=0x4600, got 0x{result:04x}")
+
+  def test_v_add_f16_e32_src0_hi_half(self):
+    """V_ADD_F16_E32 with src0 from hi-half (src0 >= v[128]).
+
+    Regression test for: VOP2 f16 src0 hi-half extraction bug.
+    """
+    instructions = [
+      # v[0] = 0x4000_3c00: hi=f16(2.0), lo=f16(1.0)
+      s_mov_b32(s[0], 0x40003c00),
+      v_mov_b32_e32(v[0], s[0]),
+      # v[1] = f16(5.0) = 0x4500
+      s_mov_b32(s[1], 0x4500),
+      v_mov_b32_e32(v[1], s[1]),
+      # v_add_f16_e32 v[2], v[128], v[1]
+      # src0 = v[128] reads from v[0].hi = 2.0
+      # result = 2.0 + 5.0 = 7.0
+      VOP2(VOP2Op.V_ADD_F16, vdst=v[2], src0=v[128], vsrc1=v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][2] & 0xffff
+    # 2.0 + 5.0 = 7.0, f16 7.0 = 0x4700
+    self.assertEqual(result, 0x4700, f"Expected f16(7.0)=0x4700, got 0x{result:04x}")
+
+
+class TestF16InlineConstants(unittest.TestCase):
+  """Regression tests for VOP2 F16 inline float constants.
+
+  For 16-bit VOP2 operations (v_add_f16, v_mul_f16, etc.), inline float constants
+  like 1.0, 2.0 must use F16 encoding (0x3c00, 0x4000) not F32 encoding (0x3f800000).
+
+  The emulator's rsrc() function needs bits=16 to select F16_INLINE constants.
+
+  Regression test for: VOP2 16-bit inline constant using F32 instead of F16.
+  """
+
+  def test_v_add_f16_inline_constant_1_0(self):
+    """V_ADD_F16_E32 with inline constant 1.0 should use F16 encoding."""
+    instructions = [
+      s_mov_b32(s[0], 0x3c00),  # f16 1.0
+      v_mov_b32_e32(v[0], s[0]),
+      # v_add_f16_e32 v[1], 1.0, v[0]  -- 1.0 must be F16 0x3c00, not F32 0x3f800000
+      v_add_f16_e32(v[1], 1.0, v[0]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][1] & 0xFFFF
+    # 1.0 + 1.0 = 2.0, f16 2.0 = 0x4000
+    self.assertEqual(result, 0x4000, f"Expected f16(2.0)=0x4000, got 0x{result:04x}")
+
+  def test_v_add_f16_inline_constant_2_0(self):
+    """V_ADD_F16_E32 with inline constant 2.0."""
+    instructions = [
+      s_mov_b32(s[0], 0x4200),  # f16 3.0
+      v_mov_b32_e32(v[0], s[0]),
+      v_add_f16_e32(v[1], 2.0, v[0]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][1] & 0xFFFF
+    # 2.0 + 3.0 = 5.0, f16 5.0 = 0x4500
+    self.assertEqual(result, 0x4500, f"Expected f16(5.0)=0x4500, got 0x{result:04x}")
+
+  def test_v_mul_f16_inline_constant(self):
+    """V_MUL_F16_E32 with inline constant 2.0."""
+    instructions = [
+      s_mov_b32(s[0], 0x4200),  # f16 3.0
+      v_mov_b32_e32(v[0], s[0]),
+      v_mul_f16_e32(v[1], 2.0, v[0]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][1] & 0xFFFF
+    # 2.0 * 3.0 = 6.0, f16 6.0 = 0x4600
+    self.assertEqual(result, 0x4600, f"Expected f16(6.0)=0x4600, got 0x{result:04x}")
+
+
+class TestCndmask(unittest.TestCase):
+  """Tests for V_CNDMASK_B32 and V_CNDMASK_B16."""
+
+  def test_v_cndmask_b16_select_src0(self):
+    """V_CNDMASK_B16 selects src0 when VCC bit is 0."""
+    instructions = [
+      s_mov_b32(VCC_LO, 0),  # VCC = 0
+      s_mov_b32(s[0], 0x3c00),  # f16 1.0
+      s_mov_b32(s[1], 0x4000),  # f16 2.0
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_cndmask_b16(v[2], v[0], v[1], VCC),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][2] & 0xffff
+    self.assertEqual(result, 0x3c00, f"Expected src0=0x3c00, got 0x{result:04x}")
+
+  def test_v_cndmask_b16_select_src1(self):
+    """V_CNDMASK_B16 selects src1 when VCC bit is 1."""
+    instructions = [
+      s_mov_b32(VCC_LO, 1),  # VCC = 1
+      s_mov_b32(s[0], 0x3c00),  # f16 1.0
+      s_mov_b32(s[1], 0x4000),  # f16 2.0
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_cndmask_b16(v[2], v[0], v[1], VCC),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][2] & 0xffff
+    self.assertEqual(result, 0x4000, f"Expected src1=0x4000, got 0x{result:04x}")
+
+  def test_v_cndmask_b16_write_hi(self):
+    """V_CNDMASK_B16 can write to high 16 bits with opsel."""
+    instructions = [
+      s_mov_b32(s[0], 0x3c003800),  # src0: hi=1.0, lo=0.5
+      v_mov_b32_e32(v[0], s[0]),
+      s_mov_b32(s[1], 0x4000c000),  # src1: hi=2.0, lo=-2.0
+      v_mov_b32_e32(v[1], s[1]),
+      s_mov_b32(s[2], 0xDEAD0000),  # v2 initial: hi=0xDEAD, lo=0
+      v_mov_b32_e32(v[2], s[2]),
+      s_mov_b32(VCC_LO, 0),  # vcc = 0, select src0
+      # opsel=0b1011: bit0=src0 hi, bit1=src1 hi, bit3=dst hi
+      VOP3(VOP3Op.V_CNDMASK_B16, vdst=v[2], src0=v[0], src1=v[1], src2=SrcEnum.VCC_LO, opsel=0b1011),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    hi = (st.vgpr[0][2] >> 16) & 0xffff
+    lo = st.vgpr[0][2] & 0xffff
+    # vcc=0 selects src0.h = 1.0 = 0x3c00, writes to hi
+    self.assertEqual(hi, 0x3c00, f"Expected hi=0x3c00 (1.0), got 0x{hi:04x}")
+    self.assertEqual(lo, 0x0000, f"Expected lo preserved as 0, got 0x{lo:04x}")
+
+
+class TestSpecialFloatValues(unittest.TestCase):
+  """Tests for special float value handling in VOP2 instructions."""
+
+  def test_neg_zero_add(self):
+    """-0.0 + 0.0 = +0.0 (IEEE 754)."""
+    neg_zero = 0x80000000
+    instructions = [
+      s_mov_b32(s[0], neg_zero),
+      v_mov_b32_e32(v[0], s[0]),
+      v_add_f32_e32(v[1], 0.0, v[0]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][1], 0x00000000, "Should be +0.0")
+
+  def test_neg_zero_mul(self):
+    """-0.0 * -1.0 = +0.0."""
+    neg_zero = 0x80000000
+    instructions = [
+      s_mov_b32(s[0], neg_zero),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mul_f32_e32(v[1], -1.0, v[0]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][1], 0x00000000, "Should be +0.0")
+
+  def test_inf_minus_inf(self):
+    """+inf - inf = NaN."""
+    import math
+    pos_inf = 0x7f800000
+    neg_inf = 0xff800000
+    instructions = [
+      s_mov_b32(s[0], pos_inf),
+      s_mov_b32(s[1], neg_inf),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_sub_f32_e32(v[2], v[0], v[1]),  # inf - (-inf) = inf
+      v_add_f32_e32(v[3], v[0], v[1]),  # inf + (-inf) = NaN
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], pos_inf, "inf - (-inf) = inf")
+    self.assertTrue(math.isnan(i2f(st.vgpr[0][3])), "inf + (-inf) = NaN")
+
+  def test_denormal_f32_mul_ftz(self):
+    """Denormal * normal - RDNA3 flushes denormals to zero (FTZ mode)."""
+    smallest_denorm = 0x00000001  # Smallest positive denormal
+    instructions = [
+      s_mov_b32(s[0], smallest_denorm),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mul_f32_e32(v[1], 2.0, v[0]),  # Denormal input gets flushed to 0
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][1], 0x00000000)
+
+
+class TestCarryOps(unittest.TestCase):
+  """Tests for VOP2 carry instructions (v_add_co_ci_u32, v_sub_co_ci_u32, v_subrev_co_ci_u32)."""
+
+  def test_v_subrev_co_ci_u32_no_borrow(self):
+    """V_SUBREV_CO_CI_U32: D0 = S1 - S0 - VCC_IN, when VCC_IN=0."""
+    instructions = [
+      s_mov_b32(VCC_LO, 0),  # VCC = 0 (no borrow in)
+      v_mov_b32_e32(v[0], 5),  # S0 = 5
+      v_mov_b32_e32(v[1], 10),  # S1 = 10
+      v_subrev_co_ci_u32_e32(v[2], v[0], v[1]),  # D0 = 10 - 5 - 0 = 5
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 5)
+    self.assertEqual(st.vcc, 0)  # No borrow out
+
+  def test_v_subrev_co_ci_u32_with_borrow(self):
+    """V_SUBREV_CO_CI_U32: D0 = S1 - S0 - VCC_IN, when VCC_IN=1."""
+    instructions = [
+      s_mov_b32(VCC_LO, 1),  # VCC = 1 (borrow in)
+      v_mov_b32_e32(v[0], 5),  # S0 = 5
+      v_mov_b32_e32(v[1], 10),  # S1 = 10
+      v_subrev_co_ci_u32_e32(v[2], v[0], v[1]),  # D0 = 10 - 5 - 1 = 4
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 4)
+    self.assertEqual(st.vcc, 0)  # No borrow out
+
+  def test_v_subrev_co_ci_u32_generates_borrow(self):
+    """V_SUBREV_CO_CI_U32: generates borrow when S0 + VCC_IN > S1."""
+    instructions = [
+      s_mov_b32(VCC_LO, 0),  # VCC = 0
+      v_mov_b32_e32(v[0], 10),  # S0 = 10
+      v_mov_b32_e32(v[1], 5),  # S1 = 5
+      v_subrev_co_ci_u32_e32(v[2], v[0], v[1]),  # D0 = 5 - 10 - 0 = -5 (underflow)
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0xFFFFFFFB)  # -5 as unsigned
+    self.assertEqual(st.vcc, 1)  # Borrow out
+
+  def test_v_add_co_ci_u32_no_carry(self):
+    """V_ADD_CO_CI_U32: D0 = S0 + S1 + VCC_IN, when VCC_IN=0."""
+    instructions = [
+      s_mov_b32(VCC_LO, 0),  # VCC = 0 (no carry in)
+      v_mov_b32_e32(v[0], 5),  # S0 = 5
+      v_mov_b32_e32(v[1], 10),  # S1 = 10
+      v_add_co_ci_u32_e32(v[2], v[0], v[1]),  # D0 = 5 + 10 + 0 = 15
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 15)
+    self.assertEqual(st.vcc, 0)  # No carry out
+
+  def test_v_add_co_ci_u32_with_carry(self):
+    """V_ADD_CO_CI_U32: D0 = S0 + S1 + VCC_IN, when VCC_IN=1."""
+    instructions = [
+      s_mov_b32(VCC_LO, 1),  # VCC = 1 (carry in)
+      v_mov_b32_e32(v[0], 5),  # S0 = 5
+      v_mov_b32_e32(v[1], 10),  # S1 = 10
+      v_add_co_ci_u32_e32(v[2], v[0], v[1]),  # D0 = 5 + 10 + 1 = 16
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 16)
+    self.assertEqual(st.vcc, 0)  # No carry out
+
+  def test_v_add_co_ci_u32_generates_carry(self):
+    """V_ADD_CO_CI_U32: generates carry when overflow occurs."""
+    instructions = [
+      s_mov_b32(VCC_LO, 1),  # VCC = 1 (carry in)
+      s_mov_b32(s[0], 0xFFFFFFFF),  # max u32
+      v_mov_b32_e32(v[0], s[0]),  # S0 = 0xFFFFFFFF
+      v_mov_b32_e32(v[1], 0),  # S1 = 0
+      v_add_co_ci_u32_e32(v[2], v[0], v[1]),  # D0 = 0xFFFFFFFF + 0 + 1 = 0 (overflow)
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0)  # Overflowed to 0
+    self.assertEqual(st.vcc, 1)  # Carry out
+
+  def test_v_add_co_ci_u32_clears_carry(self):
+    """V_ADD_CO_CI_U32: VCC must be updated even when no carry is generated.
+
+    This tests the case where VCC=1 going in (carry-in consumed) but the addition
+    does not overflow, so VCC must be cleared to 0.
+
+    Regression test for: VCC not being written by v_add_co_ci_u32_e32.
+    """
+    instructions = [
+      s_mov_b32(VCC_LO, 1),  # VCC = 1 (carry in)
+      v_mov_b32_e32(v[0], 1),  # S0 = 1
+      v_mov_b32_e32(v[1], 1),  # S1 = 1
+      v_add_co_ci_u32_e32(v[2], v[0], v[1]),  # D0 = 1 + 1 + 1 = 3 (no overflow)
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 3)  # 1 + 1 + 1 = 3
+    self.assertEqual(st.vcc, 0)  # No carry out - VCC must be cleared
+
+  def test_v_add_co_ci_u32_multilane_clears_vcc(self):
+    """V_ADD_CO_CI_U32 with multiple lanes: VCC bits must be updated per-lane.
+
+    When VCC has multiple bits set (one per active lane), and the addition doesn't
+    overflow for any lane, all VCC bits must be cleared.
+
+    Regression test for: VCC not being written by v_add_co_ci_u32_e32 in multi-lane case.
+    """
+    instructions = [
+      s_mov_b32(VCC_LO, 0b11),  # VCC = 0b11 (lanes 0,1 have carry-in)
+      v_mov_b32_e32(v[0], 1),  # S0 = 1 for all lanes
+      v_mov_b32_e32(v[1], 1),  # S1 = 1 for all lanes
+      v_add_co_ci_u32_e32(v[2], v[0], v[1]),  # D0 = 1 + 1 + 1 = 3 (no overflow)
+    ]
+    st = run_program(instructions, n_lanes=2)
+    self.assertEqual(st.vgpr[0][2], 3)  # lane 0: 1 + 1 + 1 = 3
+    self.assertEqual(st.vgpr[1][2], 3)  # lane 1: 1 + 1 + 1 = 3
+    self.assertEqual(st.vcc, 0)  # No carry out for any lane - all VCC bits must be cleared
+
+  def test_v_add_co_ci_u32_preserves_inactive_vcc_bits(self):
+    """V_ADD_CO_CI_U32: VCC carry-out overwrites entire VCC register.
+
+    VOP2 carry instructions write ALL VCC bits based on carry-out, clearing
+    bits for lanes that don't overflow regardless of EXEC mask.
+
+    Note: This differs from VOPC which only writes active lane bits.
+    """
+    instructions = [
+      s_mov_b32(VCC_LO, 0x00010000),  # VCC bit 16 set
+      v_mov_b32_e32(v[0], 1),  # S0 = 1
+      v_mov_b32_e32(v[1], 1),  # S1 = 1
+      v_add_co_ci_u32_e32(v[2], v[0], v[1]),  # D0 = 1 + 1 + 0 = 2 (no carry)
+    ]
+    st = run_program(instructions, n_lanes=4)
+    self.assertEqual(st.vgpr[0][2], 2)  # lane 0: 1 + 1 + 0 = 2
+    # VCC should be completely cleared (all lanes have no carry-out)
+    self.assertEqual(st.vcc, 0)
+
+  def test_v_add_co_ci_u32_all_lanes_same_result(self):
+    """V_ADD_CO_CI_U32: all active lanes should produce the same result.
+
+    When the same constant inputs are used across all lanes, each lane should
+    compute the same result and write to its own VGPR slot.
+
+    Regression test for: VGPR writes not happening for all lanes.
+    """
+    instructions = [
+      s_mov_b32(VCC_LO, 0),  # No carry-in
+      v_mov_b32_e32(v[0], 3),  # inline constant 3
+      v_mov_b32_e32(v[1], 5),  # value 5
+      v_add_co_ci_u32_e32(v[1], 3, v[1]),  # v[1] = 3 + v[1] + 0 = 3 + 5 = 8
+    ]
+    st = run_program(instructions, n_lanes=4)
+    # All 4 lanes should have v[1] = 8
+    for lane in range(4):
+      self.assertEqual(st.vgpr[lane][1], 8, f"lane {lane} should have v[1]=8")
+
+  def test_v_sub_co_ci_u32_no_borrow(self):
+    """V_SUB_CO_CI_U32: D0 = S0 - S1 - VCC_IN, when VCC_IN=0."""
+    instructions = [
+      s_mov_b32(VCC_LO, 0),  # VCC = 0 (no borrow in)
+      v_mov_b32_e32(v[0], 10),  # S0 = 10
+      v_mov_b32_e32(v[1], 5),  # S1 = 5
+      v_sub_co_ci_u32_e32(v[2], v[0], v[1]),  # D0 = 10 - 5 - 0 = 5
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 5)
+    self.assertEqual(st.vcc, 0)  # No borrow out
+
+  def test_v_sub_co_ci_u32_vop3sd_separate_carry_regs(self):
+    """VOP3SD V_SUB_CO_CI_U32: carry-in from src2, carry-out to sdst (separate registers).
+
+    This tests the VOP3SD encoding where src2 specifies the carry-in register
+    independently from sdst (carry-out). The bug was reading carry-in from sdst
+    instead of src2.
+
+    Computation: D0 = S0 - S1 - carry_in = 0 - 0 - 1 = -1 = 0xFFFFFFFF
+    """
+    instructions = [
+      s_mov_b32(s[6], 1),  # carry-in = 1 (in s[6])
+      s_mov_b32(s[10], 0),  # carry-out dest = 0 initially (in s[10])
+      # VOP3SD: v_sub_co_ci_u32(vdst, sdst, src0, src1, src2)
+      # src2 is carry-in (s[6]=1), sdst is carry-out (s[10])
+      v_sub_co_ci_u32(v[0], s[10], 0, 0, s[6]),  # D0 = 0 - 0 - 1 = -1
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 0xFFFFFFFF)  # -1 as unsigned
+    self.assertEqual(st.sgpr[10], 1)  # Borrow out to s[10]
+
+  def test_v_add_co_ci_u32_vop3sd_separate_carry_regs(self):
+    """VOP3SD V_ADD_CO_CI_U32: carry-in from src2, carry-out to sdst (separate registers).
+
+    This tests the VOP3SD encoding where src2 specifies the carry-in register
+    independently from sdst (carry-out).
+
+    Computation: D0 = S0 + S1 + carry_in = 5 + 10 + 1 = 16
+    """
+    instructions = [
+      s_mov_b32(s[6], 1),  # carry-in = 1 (in s[6])
+      s_mov_b32(s[10], 0),  # carry-out dest = 0 initially (in s[10])
+      # VOP3SD: v_add_co_ci_u32(vdst, sdst, src0, src1, src2)
+      v_add_co_ci_u32(v[0], s[10], 5, 10, s[6]),  # D0 = 5 + 10 + 1 = 16
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 16)
+    self.assertEqual(st.sgpr[10], 0)  # No carry out
+
+  def test_v_add_co_ci_u32_vop3sd_null_sdst(self):
+    """VOP3SD V_ADD_CO_CI_U32 with sdst=NULL: carry output is discarded.
+
+    When sdst=NULL (register 124), the carry-out should NOT be written anywhere.
+    We verify this by checking that VCC (which we set to a sentinel value) is unchanged.
+    """
+    instructions = [
+      s_mov_b32(VCC_LO, 0xDEADBEEF),  # Sentinel value in VCC
+      s_mov_b32(s[6], 0),  # carry-in = 0
+      # VOP3SD with NULL sdst: carry-out should be discarded
+      # Uses 0xFFFFFFFF + 1 + 0 = 0 with carry-out=1, but carry should not be written
+      v_add_co_ci_u32(v[0], NULL, 0xFFFFFFFF, 1, s[6]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][0], 0)  # 0xFFFFFFFF + 1 + 0 = 0 (overflow)
+    self.assertEqual(st.vcc, 0xDEADBEEF)  # VCC unchanged - carry was discarded
+
+
+if __name__ == '__main__':
+  unittest.main()

extra/assembly/amd/test/hw/test_vopc.py

@@ -0,0 +1,936 @@
+"""Tests for VOPC instructions - vector compare operations.
+
+Includes: v_cmp_class_f32, v_cmp_class_f16, v_cmp_eq_*, v_cmp_lt_*, v_cmp_gt_*
+"""
+import unittest
+from extra.assembly.amd.test.hw.helpers import *
+
+VCC = 106  # SGPR index for VCC_LO
+
+class TestCmpClass(unittest.TestCase):
+  """Tests for V_CMP_CLASS_F32 float classification."""
+
+  def test_cmp_class_quiet_nan(self):
+    """V_CMP_CLASS_F32 detects quiet NaN."""
+    quiet_nan = 0x7fc00000
+    instructions = [
+      s_mov_b32(s[0], quiet_nan),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], 0b0000000010),  # bit 1 = quiet NaN
+      v_cmp_class_f32_e32(v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "Should detect quiet NaN")
+
+  def test_cmp_class_signaling_nan(self):
+    """V_CMP_CLASS_F32 detects signaling NaN."""
+    signal_nan = 0x7f800001
+    instructions = [
+      s_mov_b32(s[0], signal_nan),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], 0b0000000001),  # bit 0 = signaling NaN
+      v_cmp_class_f32_e32(v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "Should detect signaling NaN")
+
+  def test_cmp_class_positive_inf(self):
+    """V_CMP_CLASS_F32 detects +inf."""
+    pos_inf = 0x7f800000
+    instructions = [
+      s_mov_b32(s[0], pos_inf),
+      s_mov_b32(s[1], 0b1000000000),  # bit 9 = +inf
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_cmp_class_f32_e32(v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "Should detect +inf")
+
+  def test_cmp_class_negative_inf(self):
+    """V_CMP_CLASS_F32 detects -inf."""
+    neg_inf = 0xff800000
+    instructions = [
+      s_mov_b32(s[0], neg_inf),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], 0b0000000100),  # bit 2 = -inf
+      v_cmp_class_f32_e32(v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "Should detect -inf")
+
+  def test_cmp_class_normal_positive(self):
+    """V_CMP_CLASS_F32 detects positive normal."""
+    instructions = [
+      v_mov_b32_e32(v[0], 1.0),
+      s_mov_b32(s[1], 0b0100000000),  # bit 8 = positive normal
+      v_mov_b32_e32(v[1], s[1]),
+      v_cmp_class_f32_e32(v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "Should detect positive normal")
+
+  def test_cmp_class_normal_negative(self):
+    """V_CMP_CLASS_F32 detects negative normal."""
+    instructions = [
+      v_mov_b32_e32(v[0], -1.0),
+      v_mov_b32_e32(v[1], 0b0000001000),  # bit 3 = negative normal
+      v_cmp_class_f32_e32(v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "Should detect negative normal")
+
+  def test_cmp_class_quiet_nan_not_signaling(self):
+    """Quiet NaN does not match signaling NaN mask."""
+    quiet_nan = 0x7fc00000
+    instructions = [
+      s_mov_b32(s[0], quiet_nan),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], 0b0000000001),  # bit 0 = signaling NaN only
+      v_cmp_class_f32_e32(v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 0, "Quiet NaN should not match signaling mask")
+
+  def test_cmp_class_signaling_nan_not_quiet(self):
+    """Signaling NaN does not match quiet NaN mask."""
+    signal_nan = 0x7f800001
+    instructions = [
+      s_mov_b32(s[0], signal_nan),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], 0b0000000010),  # bit 1 = quiet NaN only
+      v_cmp_class_f32_e32(v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 0, "Signaling NaN should not match quiet mask")
+
+  def test_v_cmp_sets_vcc_bits(self):
+    """V_CMP_EQ sets VCC bits based on per-lane comparison."""
+    instructions = [
+      s_mov_b32(s[0], 5),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[0]),
+      v_cmp_eq_u32_e32(v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=4)
+    self.assertEqual(st.vcc & 0xf, 0xf, "All lanes should match")
+
+
+class TestCmpClassF16(unittest.TestCase):
+  """Tests for V_CMP_CLASS_F16 float classification.
+
+  Class bit mapping:
+    bit 0 = signaling NaN
+    bit 1 = quiet NaN
+    bit 2 = -infinity
+    bit 3 = -normal
+    bit 4 = -denormal
+    bit 5 = -zero
+    bit 6 = +zero
+    bit 7 = +denormal
+    bit 8 = +normal
+    bit 9 = +infinity
+  """
+
+  def test_cmp_class_f16_positive_zero(self):
+    """V_CMP_CLASS_F16: +zero matches bit 6."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0x0000),  # f16 +0.0
+      v_mov_b32_e32(v[1], 0x40),     # bit 6 = +zero
+      v_cmp_class_f16_e32(v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "Should detect positive zero")
+
+  def test_cmp_class_f16_negative_zero(self):
+    """V_CMP_CLASS_F16: -zero matches bit 5."""
+    instructions = [
+      s_mov_b32(s[0], 0x8000),       # f16 -0.0
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], 0x20),     # bit 5 = -zero
+      v_cmp_class_f16_e32(v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "Should detect negative zero")
+
+  def test_cmp_class_f16_positive_normal(self):
+    """V_CMP_CLASS_F16: +1.0 (normal) matches bit 8."""
+    instructions = [
+      s_mov_b32(s[0], 0x3c00),       # f16 +1.0
+      s_mov_b32(s[1], 0x100),        # bit 8 = +normal
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_cmp_class_f16_e32(v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "Should detect positive normal")
+
+  def test_cmp_class_f16_negative_normal(self):
+    """V_CMP_CLASS_F16: -1.0 (normal) matches bit 3."""
+    instructions = [
+      s_mov_b32(s[0], 0xbc00),       # f16 -1.0
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], 0x08),     # bit 3 = -normal
+      v_cmp_class_f16_e32(v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "Should detect negative normal")
+
+  def test_cmp_class_f16_positive_infinity(self):
+    """V_CMP_CLASS_F16: +inf matches bit 9."""
+    instructions = [
+      s_mov_b32(s[0], 0x7c00),       # f16 +inf
+      s_mov_b32(s[1], 0x200),        # bit 9 = +inf
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_cmp_class_f16_e32(v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "Should detect positive infinity")
+
+  def test_cmp_class_f16_negative_infinity(self):
+    """V_CMP_CLASS_F16: -inf matches bit 2."""
+    instructions = [
+      s_mov_b32(s[0], 0xfc00),       # f16 -inf
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], 0x04),     # bit 2 = -inf
+      v_cmp_class_f16_e32(v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "Should detect negative infinity")
+
+  def test_cmp_class_f16_quiet_nan(self):
+    """V_CMP_CLASS_F16: quiet NaN matches bit 1."""
+    instructions = [
+      s_mov_b32(s[0], 0x7e00),       # f16 quiet NaN
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], 0x02),     # bit 1 = quiet NaN
+      v_cmp_class_f16_e32(v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "Should detect quiet NaN")
+
+  def test_cmp_class_f16_signaling_nan(self):
+    """V_CMP_CLASS_F16: signaling NaN matches bit 0."""
+    instructions = [
+      s_mov_b32(s[0], 0x7c01),       # f16 signaling NaN
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], 0x01),     # bit 0 = signaling NaN
+      v_cmp_class_f16_e32(v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "Should detect signaling NaN")
+
+  def test_cmp_class_f16_positive_denormal(self):
+    """V_CMP_CLASS_F16: positive denormal matches bit 7."""
+    instructions = [
+      v_mov_b32_e32(v[0], 1),        # f16 +denormal (0x0001)
+      v_mov_b32_e32(v[1], 0x80),     # bit 7 = +denormal
+      v_cmp_class_f16_e32(v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "Should detect positive denormal")
+
+  def test_cmp_class_f16_negative_denormal(self):
+    """V_CMP_CLASS_F16: negative denormal matches bit 4."""
+    instructions = [
+      s_mov_b32(s[0], 0x8001),       # f16 -denormal
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], 0x10),     # bit 4 = -denormal
+      v_cmp_class_f16_e32(v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "Should detect negative denormal")
+
+  def test_cmp_class_f16_combined_mask_zeros(self):
+    """V_CMP_CLASS_F16: mask 0x60 covers both +zero and -zero."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0),         # f16 +0.0
+      v_mov_b32_e32(v[1], 0x60),      # bits 5 and 6 (+-zero)
+      v_cmp_class_f16_e32(v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "VCC should be 1 for +zero with mask 0x60")
+
+  def test_cmp_class_f16_combined_mask_1f8(self):
+    """V_CMP_CLASS_F16: mask 0x1f8 covers -normal,-denorm,-zero,+zero,+denorm,+normal.
+
+    This is the exact mask used in the f16 sin kernel at PC=46.
+    """
+    instructions = [
+      v_mov_b32_e32(v[0], 0),         # f16 +0.0
+      s_mov_b32(s[0], 0x1f8),
+      v_mov_b32_e32(v[1], s[0]),      # mask 0x1f8
+      v_cmp_class_f16_e32(v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "VCC should be 1 for +zero with mask 0x1f8")
+
+  def test_cmp_class_f16_vop3_encoding(self):
+    """V_CMP_CLASS_F16 in VOP3 encoding (v_cmp_class_f16_e64)."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0),         # f16 +0.0
+      s_mov_b32(s[0], 0x1f8),         # class mask
+      v_cmp_class_f16_e64(VCC_LO, v[0], s[0]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "VCC should be 1 for +zero with VOP3 encoding")
+
+  def test_cmp_class_f16_vop3_normal_positive(self):
+    """V_CMP_CLASS_F16 VOP3 encoding with +1.0 (normal)."""
+    instructions = [
+      s_mov_b32(s[0], 0x3c00),        # f16 +1.0
+      v_mov_b32_e32(v[0], s[0]),
+      s_mov_b32(s[1], 0x1f8),         # class mask
+      v_cmp_class_f16_e64(VCC_LO, v[0], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "VCC should be 1 for +1.0 (normal) with mask 0x1f8")
+
+  def test_cmp_class_f16_vop3_nan_fails_mask(self):
+    """V_CMP_CLASS_F16 VOP3: NaN should NOT match mask 0x1f8 (no NaN bits set)."""
+    instructions = [
+      s_mov_b32(s[0], 0x7e00),        # f16 quiet NaN
+      v_mov_b32_e32(v[0], s[0]),
+      s_mov_b32(s[1], 0x1f8),         # class mask
+      v_cmp_class_f16_e64(VCC_LO, v[0], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 0, "VCC should be 0 for NaN with mask 0x1f8 (no NaN bits)")
+
+  def test_cmp_class_f16_vop3_inf_fails_mask(self):
+    """V_CMP_CLASS_F16 VOP3: +inf should NOT match mask 0x1f8 (no inf bits set)."""
+    instructions = [
+      s_mov_b32(s[0], 0x7c00),        # f16 +inf
+      v_mov_b32_e32(v[0], s[0]),
+      s_mov_b32(s[1], 0x1f8),         # class mask
+      v_cmp_class_f16_e64(VCC_LO, v[0], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 0, "VCC should be 0 for +inf with mask 0x1f8 (no inf bits)")
+
+
+class TestCmpInt(unittest.TestCase):
+  """Tests for integer comparison operations."""
+
+  def test_v_cmp_eq_u32(self):
+    """V_CMP_EQ_U32 sets VCC bits based on per-lane comparison."""
+    instructions = [
+      s_mov_b32(s[0], 5),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[0]),
+      v_cmp_eq_u32_e32(v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=4)
+    self.assertEqual(st.vcc & 0xf, 0xf, "All lanes should match")
+
+  def test_v_cmp_ne_u32_with_zero(self):
+    """V_CMP_NE_U32: compare with zero, used for int->bool cast."""
+    instructions = [
+      v_mov_b32_e32(v[1], 0),
+      v_cmp_eq_u32_e32(1, v[255]),  # vcc = (lane == 1)
+      v_cndmask_b32_e64(v[1], v[1], 1, VCC_LO),  # v1[lane1] = 1
+      v_cmp_ne_u32_e32(0, v[1]),  # vcc = (0 != v1)
+      v_cndmask_b32_e64(v[0], 0, 1, VCC_LO),  # v0 = vcc ? 1 : 0
+    ]
+    st = run_program(instructions, n_lanes=2)
+    self.assertEqual(st.vgpr[0][0], 0, "lane 0: 0 != 0 should be false")
+    self.assertEqual(st.vgpr[1][0], 1, "lane 1: 0 != 1 should be true")
+    self.assertEqual(st.vcc & 0x3, 0x2, "VCC should be 0b10")
+
+  def test_v_cmp_ne_u32_all_nonzero(self):
+    """V_CMP_NE_U32: all lanes have nonzero values."""
+    instructions = [
+      v_mov_b32_e32(v[1], 5),
+      v_cmp_ne_u32_e32(0, v[1]),
+    ]
+    st = run_program(instructions, n_lanes=4)
+    self.assertEqual(st.vcc & 0xf, 0xf, "All lanes should be != 0")
+
+  def test_cmp_eq_u16_opsel_lo_lo(self):
+    """V_CMP_EQ_U16 comparing lo halves."""
+    instructions = [
+      s_mov_b32(s[0], 0x12340005),  # lo=5, hi=0x1234
+      s_mov_b32(s[1], 0xABCD0005),  # lo=5, hi=0xABCD
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_cmp_eq_u16_e32(v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "Lo halves should be equal")
+
+  def test_cmp_eq_u16_opsel_hi_hi(self):
+    """V_CMP_EQ_U16 comparing hi halves with VOP3 opsel."""
+    instructions = [
+      s_mov_b32(s[2], 0x00051234),  # hi=5, lo=0x1234
+      v_mov_b32_e32(v[0], s[2]),
+      s_mov_b32(s[2], 0x0005ABCD),  # hi=5, lo=0xABCD
+      v_mov_b32_e32(v[1], s[2]),
+      v_cmp_eq_u16_e64(vdst=s[0], src0=v[0], src1=v[1], opsel=3),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[0] & 1, 1, "Hi halves should be equal: 5==5")
+
+  def test_cmp_eq_u16_opsel_hi_hi_equal(self):
+    """V_CMP_EQ_U16 VOP3 with opsel=3 compares hi halves (equal case)."""
+    instructions = [
+      s_mov_b32(s[2], 0x12340005),  # lo=5, hi=0x1234
+      v_mov_b32_e32(v[0], s[2]),
+      s_mov_b32(s[2], 0x12340009),  # lo=9, hi=0x1234
+      v_mov_b32_e32(v[1], s[2]),
+      v_cmp_eq_u16_e64(vdst=s[0], src0=v[0], src1=v[1], opsel=3),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[0] & 1, 1, "hi==hi should be true: 0x1234==0x1234")
+
+  def test_cmp_gt_u16_opsel_hi(self):
+    """V_CMP_GT_U16 VOP3 with opsel=3 compares hi halves."""
+    instructions = [
+      s_mov_b32(s[2], 0x99990005),  # lo=5, hi=0x9999
+      v_mov_b32_e32(v[0], s[2]),
+      s_mov_b32(s[2], 0x12340005),  # lo=5, hi=0x1234
+      v_mov_b32_e32(v[1], s[2]),
+      v_cmp_gt_u16_e64(vdst=s[0], src0=v[0], src1=v[1], opsel=3),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[0] & 1, 1, "hi>hi should be true: 0x9999>0x1234")
+
+
+class TestCmpFloat(unittest.TestCase):
+  """Tests for float comparison operations."""
+
+  def test_v_cmp_lt_f16_vsrc1_hi(self):
+    """V_CMP_LT_F16 with both operands from high half using VOP3 opsel."""
+    instructions = [
+      s_mov_b32(s[2], 0x3c000000),  # hi=1.0 (f16), lo=0
+      v_mov_b32_e32(v[0], s[2]),
+      s_mov_b32(s[2], 0x40000000),  # hi=2.0 (f16), lo=0
+      v_mov_b32_e32(v[1], s[2]),
+      v_cmp_lt_f16_e64(vdst=s[0], src0=v[0], src1=v[1], opsel=3),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[0] & 1, 1, "1.0 < 2.0 should be true")
+
+  def test_v_cmp_gt_f16_vsrc1_hi(self):
+    """V_CMP_GT_F16 with both operands from high half using VOP3 opsel."""
+    instructions = [
+      s_mov_b32(s[2], 0x40000000),  # hi=2.0 (f16), lo=0
+      v_mov_b32_e32(v[0], s[2]),
+      s_mov_b32(s[2], 0x3c000000),  # hi=1.0 (f16), lo=0
+      v_mov_b32_e32(v[1], s[2]),
+      v_cmp_gt_f16_e64(vdst=s[0], src0=v[0], src1=v[1], opsel=3),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.sgpr[0] & 1, 1, "2.0 > 1.0 should be true")
+
+  def test_v_cmp_eq_f16_vsrc1_hi_equal(self):
+    """v_cmp_eq_f16 with equal low and high halves."""
+    instructions = [
+      s_mov_b32(s[0], 0x42004200),  # hi=3.0 (0x4200), lo=3.0 (0x4200)
+      v_mov_b32_e32(v[0], s[0]),
+      v_cmp_eq_f16_e32(v[0], v[0].h),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "Expected vcc=1 (3.0 == 3.0)")
+
+  def test_v_cmp_neq_f16_vsrc1_hi(self):
+    """v_cmp_neq_f16 with different low and high halves."""
+    instructions = [
+      s_mov_b32(s[0], 0x40003c00),  # hi=2.0 (0x4000), lo=1.0 (0x3c00)
+      v_mov_b32_e32(v[0], s[0]),
+      v_cmp_lg_f16_e32(v[0], v[0].h),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "Expected vcc=1 (1.0 != 2.0)")
+
+  def test_v_cmp_nge_f16_inf_self(self):
+    """v_cmp_nge_f16 comparing -inf with itself (unordered less than).
+
+    Regression test: -inf < -inf should be false (IEEE 754).
+    """
+    instructions = [
+      s_mov_b32(s[0], 0xFC00FC00),  # both halves = -inf (0xFC00)
+      v_mov_b32_e32(v[0], s[0]),
+      v_cmp_nge_f16_e32(v[0], v[0].h),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 0, "Expected vcc=0 (-inf >= -inf)")
+
+  def test_v_cmp_f16_multilane(self):
+    """v_cmp_lt_f16 with vsrc1=v128 across multiple lanes."""
+    instructions = [
+      # Lane 0: v0 = 0x40003c00 (hi=2.0, lo=1.0) -> 1.0 < 2.0 = true
+      # Lane 1: v0 = 0x3c004000 (hi=1.0, lo=2.0) -> 2.0 < 1.0 = false
+      v_mov_b32_e32(v[0], 0x40003c00),  # default
+      v_cmp_eq_u32_e32(1, v[255]),  # vcc = (lane == 1)
+      v_cndmask_b32_e64(v[0], v[0], 0x3c004000, SrcEnum.VCC_LO),
+      v_cmp_lt_f16_e32(v[0], v[0].h),
+    ]
+    st = run_program(instructions, n_lanes=2)
+    self.assertEqual(st.vcc & 1, 1, "Lane 0: expected vcc=1 (1.0 < 2.0)")
+    self.assertEqual((st.vcc >> 1) & 1, 0, "Lane 1: expected vcc=0 (2.0 < 1.0)")
+
+
+class TestVOP3VOPCModifiers(unittest.TestCase):
+  """Tests for VOP3 VOPC with abs/neg modifiers."""
+
+  def test_v_cmp_ge_f32_abs_both(self):
+    """v_cmp_ge_f32 with abs on both sources: abs(0.0) >= abs(-1.0) = false.
+
+    Regression test: int16 mod operation uses v_cmp_ge_f32 with abs modifiers.
+    """
+    instructions = [
+      v_mov_b32_e32(v[0], 0.0),
+      v_mov_b32_e32(v[1], -1.0),
+      # abs=0b11 means abs(src0) and abs(src1)
+      v_cmp_ge_f32_e64(VCC_LO, v[0], v[1], abs=0b11),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 0, "abs(0.0) >= abs(-1.0) should be false")
+
+  def test_v_cmp_ge_f32_abs_negative_divisor(self):
+    """v_cmp_ge_f32 with abs: remainder check for negative divisor.
+
+    Tests the exact comparison used in int16 mod: abs(rem_f) >= abs(div_f).
+    For 1 % -1: rem_f = 0.0, div_f = -1.0, so abs(0.0) >= abs(-1.0) = false.
+    """
+    instructions = [
+      v_mov_b32_e32(v[0], 0.0),    # remainder as float
+      v_mov_b32_e32(v[1], -1.0),   # divisor as float
+      v_cmp_ge_f32_e64(VCC_LO, v[0], v[1], abs=0b11),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 0, "abs(0.0) >= abs(-1.0) should be false")
+
+  def test_v_cmp_ge_f32_abs_small_remainder(self):
+    """v_cmp_ge_f32 with abs: abs(-0.5) >= abs(-3.0) = false."""
+    instructions = [
+      v_mov_b32_e32(v[0], -0.5),
+      v_mov_b32_e32(v[1], -3.0),
+      v_cmp_ge_f32_e64(VCC_LO, v[0], v[1], abs=0b11),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 0, "abs(-0.5) >= abs(-3.0) should be false")
+
+  def test_v_cmp_ge_f32_abs_equal(self):
+    """v_cmp_ge_f32 with abs: abs(-1.0) >= abs(1.0) = true."""
+    instructions = [
+      v_mov_b32_e32(v[0], -1.0),
+      v_mov_b32_e32(v[1], 1.0),
+      v_cmp_ge_f32_e64(VCC_LO, v[0], v[1], abs=0b11),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "abs(-1.0) >= abs(1.0) should be true")
+
+
+class TestVOP3VOPC64Bit(unittest.TestCase):
+  """Tests for VOP3 VOPC with 64-bit operands."""
+
+  def test_v_cmp_lt_f64_basic(self):
+    """v_cmp_lt_f64: 0.0 < 1.0 = true."""
+    zero_f64 = f2i64(0.0)
+    one_f64 = f2i64(1.0)
+    instructions = [
+      s_mov_b32(s[0], zero_f64 & 0xffffffff),
+      s_mov_b32(s[1], zero_f64 >> 32),
+      s_mov_b32(s[2], one_f64 & 0xffffffff),
+      s_mov_b32(s[3], one_f64 >> 32),
+      v_cmp_lt_f64_e64(VCC_LO, s[0:1], s[2:3]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "0.0 < 1.0 should be true")
+
+  def test_v_cmp_lt_f64_negative(self):
+    """v_cmp_lt_f64: -1.0 < 0.0 = true."""
+    neg_one_f64 = f2i64(-1.0)
+    zero_f64 = f2i64(0.0)
+    instructions = [
+      s_mov_b32(s[0], neg_one_f64 & 0xffffffff),
+      s_mov_b32(s[1], neg_one_f64 >> 32),
+      s_mov_b32(s[2], zero_f64 & 0xffffffff),
+      s_mov_b32(s[3], zero_f64 >> 32),
+      v_cmp_lt_f64_e64(VCC_LO, s[0:1], s[2:3]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "-1.0 < 0.0 should be true")
+
+  def test_v_cmp_lt_i64_signed(self):
+    """v_cmp_lt_i64: 0 < -1 (signed) = false."""
+    instructions = [
+      s_mov_b32(s[0], 0),
+      s_mov_b32(s[1], 0),              # s[0:1] = 0
+      s_mov_b32(s[2], 0xffffffff),
+      s_mov_b32(s[3], 0xffffffff),     # s[2:3] = -1
+      v_cmp_lt_i64_e64(VCC_LO, s[0:1], s[2:3]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 0, "0 < -1 (signed) should be false")
+
+  def test_v_cmp_lt_u64_unsigned(self):
+    """v_cmp_lt_u64: 0 < 0xFFFFFFFFFFFFFFFF (unsigned) = true."""
+    instructions = [
+      s_mov_b32(s[0], 0),
+      s_mov_b32(s[1], 0),              # s[0:1] = 0
+      s_mov_b32(s[2], 0xffffffff),
+      s_mov_b32(s[3], 0xffffffff),     # s[2:3] = max uint64
+      v_cmp_lt_u64_e64(VCC_LO, s[0:1], s[2:3]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "0 < max_uint64 should be true")
+
+
+class TestVOPCF64(unittest.TestCase):
+  """Tests for VOPC (E32 encoding) with 64-bit float operands. Regression test for f64 compare bug."""
+
+  def test_v_cmp_lt_f64_e32_true(self):
+    """v_cmp_lt_f64_e32: 2.0 < 3.0 = true."""
+    lo0, hi0 = f2i64(2.0) & 0xffffffff, f2i64(2.0) >> 32
+    lo1, hi1 = f2i64(3.0) & 0xffffffff, f2i64(3.0) >> 32
+    instructions = [
+      s_mov_b32(s[0], lo0), s_mov_b32(s[1], hi0),
+      s_mov_b32(s[2], lo1), s_mov_b32(s[3], hi1),
+      v_mov_b32_e32(v[0], s[0]), v_mov_b32_e32(v[1], s[1]),
+      v_mov_b32_e32(v[2], s[2]), v_mov_b32_e32(v[3], s[3]),
+      v_cmp_lt_f64_e32(v[0:1], v[2:3]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "2.0 < 3.0 should be true")
+
+  def test_v_cmp_lt_f64_e32_false(self):
+    """v_cmp_lt_f64_e32: 3.0 < 2.0 = false."""
+    lo0, hi0 = f2i64(3.0) & 0xffffffff, f2i64(3.0) >> 32
+    lo1, hi1 = f2i64(2.0) & 0xffffffff, f2i64(2.0) >> 32
+    instructions = [
+      s_mov_b32(s[0], lo0), s_mov_b32(s[1], hi0),
+      s_mov_b32(s[2], lo1), s_mov_b32(s[3], hi1),
+      v_mov_b32_e32(v[0], s[0]), v_mov_b32_e32(v[1], s[1]),
+      v_mov_b32_e32(v[2], s[2]), v_mov_b32_e32(v[3], s[3]),
+      v_cmp_lt_f64_e32(v[0:1], v[2:3]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 0, "3.0 < 2.0 should be false")
+
+  def test_v_cmp_nlt_f64_e32_true(self):
+    """v_cmp_nlt_f64_e32: !(3.0 < 2.0) = true."""
+    lo0, hi0 = f2i64(3.0) & 0xffffffff, f2i64(3.0) >> 32
+    lo1, hi1 = f2i64(2.0) & 0xffffffff, f2i64(2.0) >> 32
+    instructions = [
+      s_mov_b32(s[0], lo0), s_mov_b32(s[1], hi0),
+      s_mov_b32(s[2], lo1), s_mov_b32(s[3], hi1),
+      v_mov_b32_e32(v[0], s[0]), v_mov_b32_e32(v[1], s[1]),
+      v_mov_b32_e32(v[2], s[2]), v_mov_b32_e32(v[3], s[3]),
+      v_cmp_nlt_f64_e32(v[0:1], v[2:3]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "!(3.0 < 2.0) should be true")
+
+  def test_v_cmp_nlt_f64_e32_false(self):
+    """v_cmp_nlt_f64_e32: !(2.0 < 3.0) = false."""
+    lo0, hi0 = f2i64(2.0) & 0xffffffff, f2i64(2.0) >> 32
+    lo1, hi1 = f2i64(3.0) & 0xffffffff, f2i64(3.0) >> 32
+    instructions = [
+      s_mov_b32(s[0], lo0), s_mov_b32(s[1], hi0),
+      s_mov_b32(s[2], lo1), s_mov_b32(s[3], hi1),
+      v_mov_b32_e32(v[0], s[0]), v_mov_b32_e32(v[1], s[1]),
+      v_mov_b32_e32(v[2], s[2]), v_mov_b32_e32(v[3], s[3]),
+      v_cmp_nlt_f64_e32(v[0:1], v[2:3]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 0, "!(2.0 < 3.0) should be false")
+
+
+class TestCmpxExec(unittest.TestCase):
+  """Tests for V_CMPX instructions that modify EXEC mask."""
+
+  def test_v_cmpx_ngt_f32_e64_all_true(self):
+    """V_CMPX_NGT_F32_E64: all lanes pass (literal <= all values)."""
+    # 131072.0 = 0x48000000
+    # All values > 131072, so !(131072 > val) = true for all
+    instructions = [
+      s_mov_b32(EXEC_LO, 0x7),  # 3 lanes active
+      v_mov_b32_e32(v[0], f2i(200000.0)),  # lane 0
+      v_cmp_eq_u32_e32(1, v[255]),
+      v_cndmask_b32_e64(v[1], v[0], f2i(300000.0), VCC_LO),  # lane 1
+      v_cmp_eq_u32_e32(2, v[255]),
+      v_cndmask_b32_e64(v[1], v[1], f2i(400000.0), VCC_LO),  # lane 2
+      # Now v[1] has: lane0=200000, lane1=300000, lane2=400000
+      # Compare: !(131072.0 > v[1]) i.e., 131072.0 <= v[1]
+      v_cmpx_ngt_f32_e64(EXEC_LO, f2i(131072.0), v[1]),
+    ]
+    st = run_program(instructions, n_lanes=3)
+    # All values > 131072, so all lanes should remain active
+    self.assertEqual(st.sgpr[EXEC_LO.offset] & 0x7, 0x7, "All 3 lanes should remain active")
+
+  def test_v_cmpx_ngt_f32_e64_some_false(self):
+    """V_CMPX_NGT_F32_E64: some lanes fail (literal > some values)."""
+    instructions = [
+      s_mov_b32(EXEC_LO, 0x7),  # 3 lanes active
+      v_mov_b32_e32(v[0], f2i(100000.0)),  # lane 0: 131072 > 100000 = true, so !(true) = false
+      v_cmp_eq_u32_e32(1, v[255]),
+      v_cndmask_b32_e64(v[1], v[0], f2i(200000.0), VCC_LO),  # lane 1: 131072 > 200000 = false, so !(false) = true
+      v_cmp_eq_u32_e32(2, v[255]),
+      v_cndmask_b32_e64(v[1], v[1], f2i(150000.0), VCC_LO),  # lane 2: 131072 > 150000 = false, so !(false) = true
+      v_cmpx_ngt_f32_e64(EXEC_LO, f2i(131072.0), v[1]),
+    ]
+    st = run_program(instructions, n_lanes=3)
+    # lane 0: fail (100000 < 131072), lanes 1,2: pass
+    self.assertEqual(st.sgpr[EXEC_LO.offset] & 0x7, 0x6, "Lanes 1,2 should be active, lane 0 inactive")
+
+  def test_v_cmpx_ngt_f32_e64_all_false(self):
+    """V_CMPX_NGT_F32_E64: all lanes fail (literal > all values)."""
+    instructions = [
+      s_mov_b32(EXEC_LO, 0x7),  # 3 lanes active
+      v_mov_b32_e32(v[0], f2i(100.0)),  # all lanes have 100.0
+      # 131072 > 100 = true, so !(true) = false for all
+      v_cmpx_ngt_f32_e64(EXEC_LO, f2i(131072.0), v[0]),
+    ]
+    st = run_program(instructions, n_lanes=3)
+    self.assertEqual(st.sgpr[EXEC_LO.offset] & 0x7, 0x0, "All lanes should be inactive")
+
+  def test_v_cmpx_ngt_f32_e64_large_values(self):
+    """V_CMPX_NGT_F32_E64: test with values that trigger Payne-Hanek in sin().
+
+    This is a regression test for the sin(859240.0) bug.
+    Values 859240, 1000000, 100594688 should all pass !(131072 > val).
+    """
+    instructions = [
+      s_mov_b32(EXEC_LO, 0x7),  # 3 lanes active
+      v_mov_b32_e32(v[0], f2i(859240.0)),   # lane 0
+      v_cmp_eq_u32_e32(1, v[255]),
+      v_cndmask_b32_e64(v[1], v[0], f2i(1000000.0), VCC_LO),   # lane 1
+      v_cmp_eq_u32_e32(2, v[255]),
+      v_cndmask_b32_e64(v[1], v[1], f2i(100594688.0), VCC_LO), # lane 2
+      v_cmpx_ngt_f32_e64(EXEC_LO, f2i(131072.0), v[1]),
+    ]
+    st = run_program(instructions, n_lanes=3)
+    # All values > 131072, so !(131072 > val) = true for all
+    self.assertEqual(st.sgpr[EXEC_LO.offset] & 0x7, 0x7, "All 3 lanes should remain active")
+
+
+class TestVCCBehavior(unittest.TestCase):
+  """Tests for VCC condition code behavior."""
+
+  def test_vcc_all_lanes_true(self):
+    """VCC should have all bits set when all lanes compare true."""
+    instructions = [
+      v_mov_b32_e32(v[0], 5),
+      v_mov_b32_e32(v[1], 5),
+      v_cmp_eq_u32_e32(v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=32)
+    self.assertEqual(st.vcc, 0xFFFFFFFF, "All 32 lanes should be true")
+
+  def test_vcc_lane_dependent(self):
+    """VCC should differ per lane based on lane_id comparison."""
+    instructions = [
+      v_mov_b32_e32(v[0], 16),
+      v_cmp_lt_u32_e32(v[255], v[0]),  # lanes 0-15 are < 16
+    ]
+    st = run_program(instructions, n_lanes=32)
+    self.assertEqual(st.vcc & 0xFFFF, 0xFFFF, "Lanes 0-15 should be true")
+    self.assertEqual(st.vcc >> 16, 0x0000, "Lanes 16-31 should be false")
+
+
+class TestCmpNge(unittest.TestCase):
+  """Tests for V_CMP_NGE (not-greater-or-equal) with NaN semantics.
+
+  NGE = !(a >= b). With NaN inputs:
+  - If either input is NaN, a >= b is false, so !(false) = true
+  - This differs from a < b which returns false for NaN inputs
+  """
+
+  def test_v_cmp_nge_f32_normal_values(self):
+    """v_cmp_nge_f32: basic comparison with normal floats."""
+    instructions = [
+      v_mov_b32_e32(v[0], f2i(1.0)),
+      v_mov_b32_e32(v[1], f2i(2.0)),
+      v_cmp_nge_f32_e32(v[0], v[1]),  # !(1.0 >= 2.0) = !(false) = true
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "!(1.0 >= 2.0) should be true")
+
+  def test_v_cmp_nge_f32_equal_values(self):
+    """v_cmp_nge_f32: equal values should return false."""
+    instructions = [
+      v_mov_b32_e32(v[0], f2i(1.0)),
+      v_mov_b32_e32(v[1], f2i(1.0)),
+      v_cmp_nge_f32_e32(v[0], v[1]),  # !(1.0 >= 1.0) = !(true) = false
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 0, "!(1.0 >= 1.0) should be false")
+
+  def test_v_cmp_nge_f32_greater_value(self):
+    """v_cmp_nge_f32: greater value should return false."""
+    instructions = [
+      v_mov_b32_e32(v[0], f2i(2.0)),
+      v_mov_b32_e32(v[1], f2i(1.0)),
+      v_cmp_nge_f32_e32(v[0], v[1]),  # !(2.0 >= 1.0) = !(true) = false
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 0, "!(2.0 >= 1.0) should be false")
+
+  def test_v_cmp_nge_f32_neg_inf(self):
+    """v_cmp_nge_f32: -inf compared to normal value."""
+    neg_inf = 0xff800000  # -inf
+    instructions = [
+      s_mov_b32(s[0], neg_inf),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], f2i(1.0)),
+      v_cmp_nge_f32_e32(v[0], v[1]),  # !(-inf >= 1.0) = !(false) = true
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "!(-inf >= 1.0) should be true")
+
+  def test_v_cmp_nge_f32_clears_inactive_vcc_bits(self):
+    """v_cmp_nge_f32 with partial EXEC clears inactive VCC bits (hardware behavior)."""
+    neg_inf = 0xff800000  # -inf
+    instructions = [
+      # Set VCC to all 1s first
+      s_mov_b32(VCC_LO, 0xFFFFFFFF),
+      # Set EXEC to only lane 0
+      s_mov_b32(EXEC_LO, 0x00000001),
+      # v0 = 1.0 for lane 0
+      v_mov_b32_e32(v[0], f2i(1.0)),
+      # Compare: !(-inf >= 1.0) = true for lane 0
+      v_cmp_nge_f32_e32(neg_inf, v[0]),
+    ]
+    st = run_program(instructions, n_lanes=16)
+    # Hardware clears inactive lane bits, only active lane results remain
+    # Lane 0 result = 1 (true), lanes 1-15 = 0 (cleared)
+    self.assertEqual(st.vcc, 0x00000001, "VCC should only have active lane results")
+
+  def test_v_cmp_nge_f32_nan_src0(self):
+    """v_cmp_nge_f32: NaN in src0 should return true (NaN >= x is false)."""
+    quiet_nan = 0x7fc00000
+    instructions = [
+      s_mov_b32(s[0], quiet_nan),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], f2i(1.0)),
+      v_cmp_nge_f32_e32(v[0], v[1]),  # !(NaN >= 1.0) = !(false) = true
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "!(NaN >= 1.0) should be true")
+
+  def test_v_cmp_nge_f32_nan_src1(self):
+    """v_cmp_nge_f32: NaN in src1 should return true (x >= NaN is false)."""
+    quiet_nan = 0x7fc00000
+    instructions = [
+      s_mov_b32(s[0], quiet_nan),
+      v_mov_b32_e32(v[0], f2i(1.0)),
+      v_mov_b32_e32(v[1], s[0]),
+      v_cmp_nge_f32_e32(v[0], v[1]),  # !(1.0 >= NaN) = !(false) = true
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "!(1.0 >= NaN) should be true")
+
+  def test_v_cmp_nge_f32_both_nan(self):
+    """v_cmp_nge_f32: both NaN should return true."""
+    quiet_nan = 0x7fc00000
+    instructions = [
+      s_mov_b32(s[0], quiet_nan),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[0]),
+      v_cmp_nge_f32_e32(v[0], v[1]),  # !(NaN >= NaN) = !(false) = true
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "!(NaN >= NaN) should be true")
+
+
+class TestCmpxPartialWavefront(unittest.TestCase):
+  """Tests for V_CMPX with partial wavefronts (fewer than 32 active lanes).
+
+  Regression tests for bug where v_cmpx incorrectly set EXEC bits for inactive
+  lanes when the wavefront had fewer than 32 lanes. This caused garbage data
+  from uninitialized lanes to corrupt memory writes.
+  """
+
+  def test_v_cmpx_eq_u32_partial_wave_3_lanes(self):
+    """V_CMPX_EQ_U32 with 3 active lanes should only affect those 3 lanes.
+
+    With n_lanes=3, initial EXEC=0x7. After v_cmpx comparing lane_id == 1,
+    only lane 1 should pass, so EXEC should become 0x2 (not have bits 3-31 set).
+    """
+    instructions = [
+      v_cmpx_eq_u32_e32(1, v[255]),  # EXEC = lanes where lane_id == 1
+    ]
+    st = run_program(instructions, n_lanes=3)
+    # Only lane 1 should be active (bit 1 set)
+    self.assertEqual(st.sgpr[EXEC_LO.offset] & 0xFFFFFFFF, 0x2,
+                     "Only lane 1 should be active after v_cmpx_eq_u32 with 3 lanes")
+
+  def test_v_cmpx_eq_u32_partial_wave_5_lanes(self):
+    """V_CMPX_EQ_U32 with 5 active lanes."""
+    instructions = [
+      v_cmpx_eq_u32_e32(3, v[255]),  # EXEC = lanes where lane_id == 3
+    ]
+    st = run_program(instructions, n_lanes=5)
+    self.assertEqual(st.sgpr[EXEC_LO.offset] & 0xFFFFFFFF, 0x8,
+                     "Only lane 3 should be active after v_cmpx_eq_u32 with 5 lanes")
+
+  def test_v_cmpx_lt_u32_partial_wave(self):
+    """V_CMPX_LT_U32 with partial wavefront."""
+    # VOPC: src0 < vsrc1, so we need v_cmpx_gt_u32 to get lane_id < 2
+    instructions = [
+      v_cmpx_gt_u32_e32(2, v[255]),  # EXEC = lanes where 2 > lane_id (i.e., lane_id < 2)
+    ]
+    st = run_program(instructions, n_lanes=4)
+    # Lanes 0,1 should be active (bits 0,1 set = 0x3)
+    self.assertEqual(st.sgpr[EXEC_LO.offset] & 0xFFFFFFFF, 0x3,
+                     "Only lanes 0,1 should be active after v_cmpx_gt_u32(2, lane_id) with 4 lanes")
+
+  def test_v_cmpx_ge_u32_partial_wave(self):
+    """V_CMPX_GE_U32 with partial wavefront."""
+    # VOPC: src0 >= vsrc1, so v_cmpx_le_u32(1, lane_id) gives lane_id >= 2? No.
+    # v_cmpx_le_u32(src0, vsrc1) = src0 <= vsrc1 = 1 <= lane_id
+    instructions = [
+      v_cmpx_le_u32_e32(2, v[255]),  # EXEC = lanes where 2 <= lane_id (i.e., lane_id >= 2)
+    ]
+    st = run_program(instructions, n_lanes=4)
+    # Lanes 2,3 should be active (bits 2,3 set = 0xC)
+    self.assertEqual(st.sgpr[EXEC_LO.offset] & 0xFFFFFFFF, 0xC,
+                     "Only lanes 2,3 should be active after v_cmpx_le_u32(2, lane_id) with 4 lanes")
+
+  def test_v_cmpx_ne_u32_partial_wave_all_pass(self):
+    """V_CMPX_NE_U32 where all active lanes pass."""
+    instructions = [
+      v_cmpx_ne_u32_e32(99, v[255]),  # EXEC = lanes where lane_id != 99
+    ]
+    st = run_program(instructions, n_lanes=3)
+    # All 3 lanes should remain active (bits 0,1,2 set = 0x7)
+    self.assertEqual(st.sgpr[EXEC_LO.offset] & 0xFFFFFFFF, 0x7,
+                     "All 3 lanes should remain active when all pass")
+
+  def test_v_cmpx_eq_u32_partial_wave_none_pass(self):
+    """V_CMPX_EQ_U32 where no active lanes pass."""
+    instructions = [
+      v_cmpx_eq_u32_e32(99, v[255]),  # EXEC = lanes where lane_id == 99
+    ]
+    st = run_program(instructions, n_lanes=3)
+    # No lanes should be active
+    self.assertEqual(st.sgpr[EXEC_LO.offset] & 0xFFFFFFFF, 0x0,
+                     "No lanes should be active when none pass")
+
+  def test_v_cmpx_f32_partial_wave(self):
+    """V_CMPX_GT_F32 with partial wavefront - float comparison."""
+    instructions = [
+      v_cvt_f32_u32_e32(v[0], v[255]),  # v[0] = float(lane_id)
+      v_mov_b32_e32(v[1], f2i(0.5)),    # v[1] = 0.5
+      v_cmpx_gt_f32_e32(v[0], v[1]),    # EXEC = lanes where v[0] > 0.5
+    ]
+    st = run_program(instructions, n_lanes=4)
+    # Lanes 1,2,3 have values > 0.5, lane 0 has 0.0
+    self.assertEqual(st.sgpr[EXEC_LO.offset] & 0xFFFFFFFF, 0xE,
+                     "Lanes 1,2,3 should be active (float > 0.5)")
+
+  def test_v_cmpx_e64_partial_wave(self):
+    """V_CMPX_EQ_U32_E64 (VOP3 encoding) with partial wavefront."""
+    instructions = [
+      v_cmpx_eq_u32_e64(EXEC_LO, v[255], 2),  # EXEC = lanes where lane_id == 2
+    ]
+    st = run_program(instructions, n_lanes=4)
+    self.assertEqual(st.sgpr[EXEC_LO.offset] & 0xFFFFFFFF, 0x4,
+                     "Only lane 2 should be active after v_cmpx_eq_u32_e64")
+
+
+if __name__ == '__main__':
+  unittest.main()

extra/assembly/amd/test/hw/test_vopd.py

@@ -0,0 +1,161 @@
+"""Tests for VOPD instructions - dual-issue vector operations.
+
+VOPD executes two operations simultaneously. Key behavior:
+- Both ops read their sources BEFORE either writes (dual-issue semantics)
+- This means if X writes to a register that Y reads, Y sees the OLD value
+- Op X can use ops 0-15 (FMAC, MUL, ADD, MOV, etc.)
+- Op Y can use ops 0-18 (includes ADD_NC_U32, LSHLREV, AND)
+"""
+import unittest
+from extra.assembly.amd.test.hw.helpers import run_program, run_program_emu, run_program_hw, compare_wave_states, \
+  v, s, v_mov_b32_e32, s_mov_b32
+from extra.assembly.amd.autogen.rdna3.ins import VOPD, VOPD_LIT, VOPDOp
+
+class TestVOPDBasic(unittest.TestCase):
+  """Basic VOPD functionality tests."""
+
+  def test_vopd_dual_mov(self):
+    """VOPD with two MOV operations to different registers."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0x12345678),
+      v_mov_b32_e32(v[1], 0xDEADBEEF),
+      # X: v[2] = v[0], Y: v[3] = v[1]
+      VOPD(VOPDOp.V_DUAL_MOV_B32, VOPDOp.V_DUAL_MOV_B32, v[2], v[3], v[0], v[1], v[0], v[0]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0x12345678)
+    self.assertEqual(st.vgpr[0][3], 0xDEADBEEF)
+
+  def test_vopd_mov_and_add(self):
+    """VOPD with MOV (X) and ADD_NC_U32 (Y) - ADD_NC_U32 can only be Y op."""
+    instructions = [
+      v_mov_b32_e32(v[0], 10),
+      v_mov_b32_e32(v[1], 5),
+      # X: v[2] = 100 (literal), Y: v[3] = v[0] + v[1] = 15
+      VOPD(VOPDOp.V_DUAL_MOV_B32, VOPDOp.V_DUAL_ADD_NC_U32, v[2], v[3], 100, v[0], v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 100)
+    self.assertEqual(st.vgpr[0][3], 15)
+
+
+class TestVOPDReadBeforeWrite(unittest.TestCase):
+  """Tests for VOPD dual-issue read-before-write semantics.
+
+  In VOPD, both X and Y operations read their sources BEFORE either writes.
+  This is critical when X's destination is Y's source.
+  """
+
+  def test_vopd_x_writes_y_reads_same_reg(self):
+    """VOPD where X writes to a register that Y reads.
+
+    X: v[2] = 0 (overwrites v[2])
+    Y: v[1] = v[2] + v[0]  (srcy0=v[2], vsrcy1=v[0])
+
+    If reads happen before writes: v[1] = OLD_v[2] + v[0] = 0xFFFFFFFF + 1 = 0
+    If writes happen before reads: v[1] = 0 + v[0] = 0 + 1 = 1
+
+    Hardware does reads-before-writes, so v[1] should be 0.
+    """
+    instructions = [
+      v_mov_b32_e32(v[0], 1),          # v[0] = 1
+      v_mov_b32_e32(v[1], 0x99999999), # v[1] = placeholder (will be overwritten)
+      v_mov_b32_e32(v[2], 0xFFFFFFFF), # v[2] = 0xFFFFFFFF
+      # X: v[2] = 0 (literal), srcx0=0, vsrcx1=v[0] (unused for MOV)
+      # Y: v[1] = srcy0 + vsrcy1 = v[2] + v[0] (should read OLD v[2] = 0xFFFFFFFF)
+      # vdsty encoding: (vdsty << 1) | ((vdstx & 1) ^ 1) where vdsty field = 0, vdstx = v[2]
+      # So vdsty_reg = (0 << 1) | ((2 & 1) ^ 1) = 0 | 1 = 1 = v[1]
+      VOPD(VOPDOp.V_DUAL_MOV_B32, VOPDOp.V_DUAL_ADD_NC_U32, v[2], v[0], 0, v[2], v[0], v[0]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    # X should have written 0 to v[2]
+    self.assertEqual(st.vgpr[0][2], 0, "X should write 0 to v[2]")
+    # Y should have read OLD v[2] (0xFFFFFFFF) and added v[0] (1)
+    # 0xFFFFFFFF + 1 = 0 (wrap around)
+    self.assertEqual(st.vgpr[0][1], 0, "Y should read OLD v[2]=0xFFFFFFFF, compute 0xFFFFFFFF+1=0")
+
+  def test_vopd_x_writes_y_reads_same_reg_v2(self):
+    """VOPD where X writes to a register that Y reads - cleaner test case.
+
+    X: v[2] = 0 (MOV)
+    Y: v[1] = v[2] + v[2] (ADD_NC_U32 with both sources from v[2])
+
+    If reads happen before writes: v[1] = OLD_v[2] + OLD_v[2] = 100 + 100 = 200
+    If writes happen before reads: v[1] = 0 + 0 = 0
+
+    Hardware does reads-before-writes, so v[1] should be 200.
+    """
+    instructions = [
+      v_mov_b32_e32(v[0], 0x88888888), # v[0] = unused placeholder
+      v_mov_b32_e32(v[1], 0x99999999), # v[1] = placeholder (will be overwritten)
+      v_mov_b32_e32(v[2], 100),        # v[2] = 100
+      # X: v[2] = 0 (literal)
+      # Y: v[1] = srcy0 + vsrcy1 = v[2] + v[2] (should read OLD v[2] = 100)
+      VOPD(VOPDOp.V_DUAL_MOV_B32, VOPDOp.V_DUAL_ADD_NC_U32, v[2], v[0], 0, v[2], v[0], v[2]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    # X should have written 0 to v[2]
+    self.assertEqual(st.vgpr[0][2], 0, "X should write 0 to v[2]")
+    # Y should have read OLD v[2] (100) twice and added them
+    self.assertEqual(st.vgpr[0][1], 200, "Y should read OLD v[2]=100 twice, compute 100+100=200")
+
+
+class TestVOPDLiterals(unittest.TestCase):
+  """Tests for VOPD instructions that use SIMM32 literals (FMAAK, FMAMK)."""
+
+  def test_vopd_fmaak_f32(self):
+    """VOPD V_DUAL_FMAAK_F32: D = S0 * S1 + SIMM32 (literal addend).
+
+    Tests that the 32-bit literal (SIMM32) is correctly passed to the instruction.
+    fma(2.0, 3.0, 10.0) = 2*3 + 10 = 16.0
+    """
+    from extra.assembly.amd.test.hw.helpers import f2i, i2f
+    instructions = [
+      v_mov_b32_e32(v[0], f2i(2.0)),  # v[0] = 2.0
+      v_mov_b32_e32(v[1], f2i(3.0)),  # v[1] = 3.0
+      # VOPD args: opx, opy, vdstx, vdsty, srcx0, srcy0, vsrcx1, vsrcy1
+      # X: v[2] = fma(srcx0, vsrcx1, SIMM32) = v[0]*v[1]+10.0 = 2*3+10 = 16
+      # Y: v[3] = srcy0 (MOV) = v[0] = 2.0
+      VOPD_LIT(VOPDOp.V_DUAL_FMAAK_F32, VOPDOp.V_DUAL_MOV_B32, v[2], v[3], v[0], v[0], v[1], v[0], literal=f2i(10.0)),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertAlmostEqual(i2f(st.vgpr[0][2]), 16.0, places=5, msg="fma(2.0, 3.0, 10.0) should be 16.0")
+
+  def test_vopd_fmamk_f32(self):
+    """VOPD V_DUAL_FMAMK_F32: D = S0 * SIMM32 + S1 (literal multiplier).
+
+    Tests that the 32-bit literal (SIMM32) is correctly used as the multiplier.
+    fma(2.0, 5.0, 3.0) = 2*5 + 3 = 13.0
+    """
+    from extra.assembly.amd.test.hw.helpers import f2i, i2f
+    instructions = [
+      v_mov_b32_e32(v[0], f2i(2.0)),  # v[0] = 2.0
+      v_mov_b32_e32(v[1], f2i(3.0)),  # v[1] = 3.0
+      # X: v[2] = fma(srcx0, SIMM32, vsrcx1) = v[0]*5.0+v[1] = 2*5+3 = 13
+      # Y: v[3] = srcy0 (MOV) = v[0] = 2.0
+      VOPD_LIT(VOPDOp.V_DUAL_FMAMK_F32, VOPDOp.V_DUAL_MOV_B32, v[2], v[3], v[0], v[0], v[1], v[0], literal=f2i(5.0)),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertAlmostEqual(i2f(st.vgpr[0][2]), 13.0, places=5, msg="fma(2.0, 5.0, 3.0) should be 13.0")
+
+
+class TestVOPDMultilane(unittest.TestCase):
+  """Tests for VOPD with multiple lanes."""
+
+  def test_vopd_multilane_mov_add(self):
+    """VOPD MOV and ADD with multiple active lanes - no register conflict."""
+    instructions = [
+      v_mov_b32_e32(v[0], 5),
+      v_mov_b32_e32(v[1], 10),
+      # X: v[2] = 100 (constant), Y: v[1] = v[0] + v[1] = 5 + 10 = 15
+      # vdsty_reg = (vdsty << 1) | ((vdstx.offset & 1) ^ 1) = (0 << 1) | ((258 & 1) ^ 1) = 0 | 1 = 1
+      VOPD(VOPDOp.V_DUAL_MOV_B32, VOPDOp.V_DUAL_ADD_NC_U32, v[2], v[0], 100, v[0], v[2], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=4)
+    for lane in range(4):
+      self.assertEqual(st.vgpr[lane][2], 100, f"Lane {lane}: v[2] should be 100")
+      self.assertEqual(st.vgpr[lane][1], 15, f"Lane {lane}: v[1] should be 15 (5+10)")
+
+
+if __name__ == '__main__':
+  unittest.main()

extra/assembly/amd/test/test_compare_emulators.py

@@ -0,0 +1,507 @@
+# Test to compare Python and Rust RDNA3 emulators by running real tinygrad kernels
+import unittest, ctypes
+from dataclasses import dataclass
+
+from extra.assembly.amd.emu import WaveState, decode_program, WAVE_SIZE, VCC_LO, EXEC_LO, SCC
+from extra.assembly.amd import decode_inst
+from extra.assembly.amd.test.helpers import KernelInfo
+from extra.assembly.amd.test.bench_emu import REMU_PATH
+
+def set_valid_mem_ranges(ranges): pass  # emu2 doesn't need this
+
+def _is_f32_nan(bits: int) -> bool:
+  """Check if 32-bit value is a NaN (exponent all 1s, mantissa non-zero)."""
+  return (bits & 0x7f800000) == 0x7f800000 and (bits & 0x007fffff) != 0
+
+def _vals_equal(a: int, b: int) -> bool:
+  """Compare two 32-bit values, treating all NaN bit patterns as equal."""
+  if a == b: return True
+  return _is_f32_nan(a) and _is_f32_nan(b)
+
+@dataclass
+class StateSnapshot:
+  pc: int
+  scc: int
+  vcc: int
+  exec_mask: int
+  sgpr: list[int]
+  vgpr: list[list[int]]
+
+  def diff(self, other: 'StateSnapshot', n_lanes: int, arrow: str = " vs ") -> list[str]:
+    """Return list of differences between two states."""
+    diffs = []
+    if self.pc != other.pc: diffs.append(f"pc: {self.pc}{arrow}{other.pc}")
+    if self.scc != other.scc: diffs.append(f"scc: {self.scc}{arrow}{other.scc}")
+    if self.vcc != other.vcc: diffs.append(f"vcc: 0x{self.vcc:08x}{arrow}0x{other.vcc:08x}")
+    if self.exec_mask != other.exec_mask: diffs.append(f"exec: 0x{self.exec_mask:08x}{arrow}0x{other.exec_mask:08x}")
+    for i, (a, b) in enumerate(zip(self.sgpr, other.sgpr)):
+      # Skip VCC_LO/HI (106/107) and EXEC_LO/HI (126/127) as they alias vcc/exec_mask which are compared separately
+      if i in (106, 107, 126, 127): continue
+      if not _vals_equal(a, b): diffs.append(f"sgpr[{i}]: 0x{a:08x}{arrow}0x{b:08x}")
+    for lane in range(n_lanes):
+      for i, (a, b) in enumerate(zip(self.vgpr[lane], other.vgpr[lane])):
+        if not _vals_equal(a, b): diffs.append(f"vgpr[{lane}][{i}]: 0x{a:08x}{arrow}0x{b:08x}")
+    return diffs
+
+class CStateSnapshot(ctypes.Structure):
+  _fields_ = [("pc", ctypes.c_uint32), ("scc", ctypes.c_uint32), ("vcc", ctypes.c_uint32), ("exec_mask", ctypes.c_uint32),
+              ("sgpr", ctypes.c_uint32 * 128), ("vgpr", (ctypes.c_uint32 * 256) * 32)]
+
+  def to_snapshot(self) -> StateSnapshot:
+    return StateSnapshot(pc=self.pc, scc=self.scc, vcc=self.vcc, exec_mask=self.exec_mask,
+                         sgpr=list(self.sgpr), vgpr=[list(self.vgpr[i]) for i in range(32)])
+
+class RustEmulator:
+  def __init__(self):
+    self.lib = ctypes.CDLL(str(REMU_PATH))
+    self.lib.wave_create.argtypes = [ctypes.c_void_p, ctypes.c_uint32, ctypes.c_uint32]
+    self.lib.wave_create.restype = ctypes.c_void_p
+    self.lib.wave_step.argtypes = [ctypes.c_void_p]
+    self.lib.wave_step.restype = ctypes.c_int32
+    self.lib.wave_get_snapshot.argtypes = [ctypes.c_void_p, ctypes.POINTER(CStateSnapshot)]
+    self.lib.wave_set_sgpr.argtypes = [ctypes.c_void_p, ctypes.c_uint32, ctypes.c_uint32]
+    self.lib.wave_set_vgpr.argtypes = [ctypes.c_void_p, ctypes.c_uint32, ctypes.c_uint32, ctypes.c_uint32]
+    self.lib.wave_init_lds.argtypes = [ctypes.c_void_p, ctypes.c_uint32]
+    self.lib.wave_free.argtypes = [ctypes.c_void_p]
+    self.ctx = None
+
+  def create(self, kernel: bytes, n_lanes: int):
+    kernel_buf = (ctypes.c_char * len(kernel)).from_buffer_copy(kernel)
+    self.ctx = self.lib.wave_create(ctypes.addressof(kernel_buf), len(kernel), n_lanes)
+    self._kernel_buf = kernel_buf
+
+  def step(self) -> int: return self.lib.wave_step(self.ctx)
+  def set_sgpr(self, idx: int, val: int): self.lib.wave_set_sgpr(self.ctx, idx, val)
+  def set_vgpr(self, lane: int, idx: int, val: int): self.lib.wave_set_vgpr(self.ctx, lane, idx, val)
+  def init_lds(self, size: int): self.lib.wave_init_lds(self.ctx, size)
+
+  def get_snapshot(self) -> StateSnapshot:
+    snap = CStateSnapshot()
+    self.lib.wave_get_snapshot(self.ctx, ctypes.byref(snap))
+    return snap.to_snapshot()
+
+  def free(self):
+    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 | None = None
+    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
+
+  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 generic instructions can read from vmem at actual PC address
+    self.kernel_buf = (ctypes.c_char * len(kernel)).from_buffer_copy(kernel)
+    self.lib_addr = ctypes.addressof(self.kernel_buf)
+    # Remap program dict to use actual addresses (like run_asm does)
+    program_raw = decode_program(kernel)
+    self.program = {self.lib_addr + offset: val for offset, val in program_raw.items()}
+    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 step(self) -> int:
+    import ctypes
+    assert self.program is not None and self.state is not None
+    pc = self.state.pc
+    if pc == 0xFFFFFFFFFFFFFFFF or pc not in self.program: return -1
+    name, fxn, globals_list, _runner = self.program[pc]
+    if fxn is None: return 1  # unsupported instruction
+    buf_addrs = {0: self.state.sgpr_buf._buf.va_addr, 1: self.state.vgpr_buf._buf.va_addr,
+                 2: self.vmem_buf._buf.va_addr, 3: self.lds_buf._buf.va_addr}
+    # Direct ctypes call - bypasses HCQ overhead
+    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
+
+  def set_sgpr(self, idx: int, val: int):
+    assert self.state is not None
+    self.state._write_sgpr(idx, val)
+  def set_vgpr(self, lane: int, idx: int, val: int):
+    assert self.state is not None
+    self.state._write_vgpr(idx, lane, val)
+
+  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)
+
+def run_single_kernel(kernel: bytes, n_lanes: int, args_ptr: int, global_size: tuple[int, int, int],
+                      local_size: tuple[int, int, 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
+
+  for gidz in range(gz):
+    for gidy in range(gy):
+      for gidx in range(gx):
+        if wg_count >= max_workgroups: return True, f"Completed {wg_count} workgroups (limit reached)", total_steps
+        wg_count += 1
+        rust = RustEmulator()
+        python = PythonEmulator()
+        rust.create(kernel, n_lanes)
+        python.create(kernel, n_lanes)
+
+        # Initialize LDS (64KB, standard size for AMD GPUs)
+        rust.init_lds(65536)
+
+        for emu in (rust, python):
+          emu.set_sgpr(0, args_ptr & 0xffffffff)
+          emu.set_sgpr(1, (args_ptr >> 32) & 0xffffffff)
+          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()
+
+            inst_info = python.program.get(python.lib_addr + python_before.pc * 4)  # Convert word offset to actual address
+            inst_hex_name = inst_info[0] if inst_info else f"unknown at PC={python_before.pc}"
+            # 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:
+              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
+            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
+            # 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 []
+            if diffs:
+              trace_lines = []
+              for idx, (s, pc, d, rb, pb) in enumerate(trace):
+                trace_lines.append(f"    step {s}: PC={pc:3d} {d}")
+                if idx < len(trace) - 1:
+                  next_rb, next_pb = trace[idx + 1][3:5]
+                  rust_diffs = rb.diff(next_rb, n_lanes, "->")
+                  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(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(f"             python: (no changes)")
+              trace_str = "\n".join(trace_lines)
+              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()
+
+            if rust_result != python_result:
+              # Rust returns 1 for unsupported instructions - skip test
+              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)
+              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:
+              rust_after = rust.get_snapshot()
+              for i in range(128): python.set_sgpr(i, rust_after.sgpr[i])
+              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
+
+            if rust_result == -1:
+              total_steps += step + 1
+              break
+            if rust_result == 1:
+              total_steps += step + 1
+              break
+            if rust_result < 0 and rust_result != -2:
+              return False, f"K{kernel_idx} WG({gidx},{gidy},{gidz}) Step {step}: error code {rust_result}", total_steps
+
+            step += 1
+          else:
+            return False, f"K{kernel_idx} WG({gidx},{gidy},{gidz}) Max steps ({max_steps}) reached", total_steps
+        finally:
+          rust.free()
+
+  return True, f"Completed {gx*gy*gz} workgroups", total_steps
+
+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 = {}
+
+  # Allocate shared buffer pool with padding for over-reads (GPU loads up to 16 bytes at once)
+  buf_id_to_ptr: dict[int, int] = {}
+  buffers = []
+  for buf_id, size in buf_pool.items():
+    padded_size = ((size + 15) // 16) * 16 + 16  # round up to 16 bytes + extra padding
+    # Initialize with data from COPY if available
+    init_data = buf_data.get(buf_id, b'\x00' * padded_size)
+    init_list = list(init_data) + [0] * (padded_size - len(init_data))
+    buf = (ctypes.c_uint8 * padded_size)(*init_list[:padded_size])
+    buffers.append((buf, padded_size))
+    buf_id_to_ptr[buf_id] = ctypes.addressof(buf)
+
+  # Set up valid memory ranges
+  ranges = {(ctypes.addressof(b), size) for b, size in buffers}
+
+  total_steps = 0
+  for ki, kernel in enumerate(kernels):
+    # Create args array for this kernel's buffers
+    args = (ctypes.c_uint64 * len(kernel.buf_idxs))(*[buf_id_to_ptr[bid] for bid in kernel.buf_idxs])
+    args_ptr = ctypes.addressof(args)
+
+    # Update valid ranges to include this args array
+    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, program, max_steps, debug, trace_len, ki
+    )
+    total_steps += steps
+    if not ok:
+      return False, msg
+
+  return True, f"Completed {len(kernels)} kernels, {total_steps} total steps"
+
+def compare_emulators_with_memory(kernel: bytes, n_lanes: int, buf_sizes: list, max_steps: int = 1000, debug: bool = False,
+                                   global_size: tuple[int, int, int] = (1, 1, 1), trace_len: int = 10) -> tuple[bool, str]:
+  """Run both emulators with memory set up for tinygrad kernels, executing all workgroups. Legacy wrapper."""
+  # Allocate buffers
+  buffers = []
+  for size in buf_sizes:
+    buf = (ctypes.c_uint8 * size)(*[0] * size)
+    buffers.append(buf)
+
+  # Create args array with buffer pointers
+  args = (ctypes.c_uint64 * len(buffers))(*[ctypes.addressof(b) for b in buffers])
+  args_ptr = ctypes.addressof(args)
+
+  # Set up valid memory ranges for Python emulator
+  ranges = {(ctypes.addressof(b), len(b)) for b in buffers}
+  ranges.add((args_ptr, ctypes.sizeof(args)))
+  set_valid_mem_ranges(ranges)
+
+  program = decode_program(kernel)
+  # 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), program, max_steps, debug, trace_len)
+  return ok, msg
+
+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
+
+  out = op_fn(Tensor)
+  sched = out.schedule()
+  kernels = []
+  buf_pool: dict[int, int] = {}  # buffer id -> size
+  buf_data: dict[int, bytes] = {}  # buffer id -> initial data from COPY
+
+  for ei in sched:
+    lowered = ei.lower()
+    if ei.ast.op.name == 'COPY':
+      # Handle COPY: extract source data to initialize destination buffer
+      if len(lowered.bufs) >= 2:
+        dst_buf, src_buf = lowered.bufs[0], lowered.bufs[1]
+        dst_id = id(dst_buf)
+        if dst_id not in buf_pool:
+          buf_pool[dst_id] = dst_buf.nbytes
+        # Get source data if it's from numpy/CPU
+        if hasattr(src_buf, 'base') and src_buf.base is not None and hasattr(src_buf.base, '_buf'):
+          src_data = bytes(src_buf.base._buf)
+          buf_data[dst_id] = src_data
+    elif ei.ast.op.name == 'SINK':
+      if 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_idxs = []
+            buf_sizes = []
+            for b in lowered.bufs:
+              buf_id = id(b)
+              if buf_id not in buf_pool:
+                buf_pool[buf_id] = b.nbytes
+              buf_idxs.append(buf_id)
+              buf_sizes.append(b.nbytes)
+            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,
+              buf_sizes=buf_sizes
+            ))
+  if not kernels: raise RuntimeError("No kernel found")
+  return kernels, buf_pool, buf_data
+
+def get_kernel_from_tinygrad(op_fn) -> tuple[bytes, tuple[int, int, int], tuple[int, int, int], list]:
+  """Compile a tinygrad operation and extract the last (main) kernel binary. Legacy wrapper."""
+  kernels, _, _ = get_kernels_from_tinygrad(op_fn)
+  k = kernels[-1]
+  return k.code, k.global_size, k.local_size, k.buf_sizes
+
+class TestTinygradKernels(unittest.TestCase):
+  """Compare emulators on real tinygrad-compiled kernels."""
+
+  def _test_kernel(self, op_fn, max_steps=10000):
+    kernels, buf_pool, buf_data = get_kernels_from_tinygrad(op_fn)
+    ok, msg = compare_emulators_multi_kernel(kernels, buf_pool, max_steps=max_steps, buf_data=buf_data)
+    self.assertTrue(ok, msg)
+
+  # Basic ops - consolidated tests covering key instruction patterns
+  def test_unary_ops(self): self._test_kernel(lambda T: T([-1.0, 0.0, 1.0, 2.0]).relu().exp().log().sqrt().reciprocal())
+  def test_binary_ops(self): self._test_kernel(lambda T: (T([1.0, 2.0]) + T([3.0, 4.0])) * T([0.5, 0.5]) - T([1.0, 1.0]))
+  def test_trig(self): self._test_kernel(lambda T: T([0.1, 1.0, 3.14, -1.0]*8).sin() + T([0.1, 1.0, 3.14, -1.0]*8).cos())
+  def test_compare(self): self._test_kernel(lambda T: (T.empty(64) < T.empty(64)).where(T.empty(64), T.empty(64)))
+  def test_bitwise(self): self._test_kernel(lambda T: (T([0xF0, 0x0F, 0xFF]*11).int() & T([0x0F, 0x0F, 0x00]*11).int()) | T([1]*33).int())
+  def test_int_ops(self): self._test_kernel(lambda T: ((T.empty(64).int() + T.empty(64).int()) * T.empty(64).int()).float())
+
+  # Reductions
+  def test_reduce(self): self._test_kernel(lambda T: T.empty(64).sum() + T.empty(64).max())
+  def test_argmax(self): self._test_kernel(lambda T: T.empty(64).argmax())
+
+  # Matmul
+  def test_gemm(self): self._test_kernel(lambda T: T.empty(8, 8) @ T.empty(8, 8), max_steps=100000)
+  @unittest.skip("Rust emulator crashes on this kernel (assertion failure in thread.rs)")
+  def test_gemm_fp16(self): self._test_kernel(lambda T: T.empty(16, 16).half() @ T.empty(16, 16).half(), max_steps=100000)
+
+  # Complex ops
+  def test_softmax(self): self._test_kernel(lambda T: T.empty(16).softmax())
+  def test_layernorm(self): self._test_kernel(lambda T: T.empty(8, 8).layernorm())
+
+  # Memory patterns
+  def test_memory(self): self._test_kernel(lambda T: T.empty(4, 4).permute(1, 0).contiguous() + T.empty(4, 1).expand(4, 4))
+
+  # Cast ops
+  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)))
+
+  # 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)
+
+  # Regression tests
+  def test_topk(self): self._test_kernel(lambda T: T.empty(64).topk(3)[0])
+  def test_interpolate(self): self._test_kernel(lambda T: T.empty(1,2,16,16).relu().cast('uint8').interpolate((8,8), mode="linear"))
+  def test_index_int64(self):
+    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)
+    classes = np.random.randint(0, 10, (16,), dtype=np.int32).tolist()
+    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_custom_kernel.py

@@ -0,0 +1,77 @@
+import unittest
+import functools
+from tinygrad import Tensor, Device, dtypes
+from tinygrad.uop.ops import UOp, Ops, KernelInfo
+from tinygrad.renderer import Estimates
+from tinygrad.runtime.support.compiler_amd import HIPCompiler
+
+from extra.assembly.amd.autogen.rdna3.ins import *
+from extra.assembly.amd.dsl import s, v, Inst
+
+def assemble_insts(insts:list[Inst], name:str, arch:str, kernarg_size:int=8) -> tuple[UOp, UOp]:
+  kd = {"kernarg_size":kernarg_size, "user_sgpr_kernarg_segment_ptr":1, "next_free_vgpr":8, "next_free_sgpr":8, "wavefront_size32":1}
+  disasm = "\n".join([inst.disasm() for inst in insts])
+  hsasrc = f".text\n.globl {name}\n.p2align 8\n.type fn_name,@function\n{name}:\n{disasm}\ns_code_end\n"
+  hsasrc += f".rodata\n.p2align 6\n.amdhsa_kernel {name}\n"+"\n".join([f".amdhsa_{k} {v}" for k,v in kd.items()])+"\n.end_amdhsa_kernel"
+  binary = HIPCompiler(arch).compile(hsasrc)
+  return UOp(Ops.SOURCE, arg=disasm), UOp(Ops.BINARY, arg=binary)
+
+def custom_add_one(A:UOp, arch:str) -> UOp:
+  A = A.flatten()
+  assert dtypes.is_float(A.dtype.base), f"buffer dtype must be float32, got {A.dtype}"
+  threads = UOp.special(A.size, "lidx0")
+  insts = [
+    s_load_b64(s[0:1], s[0:1], soffset=NULL),
+    s_waitcnt(lgkmcnt=0),
+    v_lshlrev_b32_e32(v[0], 2, v[0]), # element offset
+    global_load_b32(v[1], v[0], saddr=s[0:1]),
+    s_waitcnt(vmcnt=0),
+    v_mov_b32_e32(v[2], 1.0),
+    v_add_f32_e32(v[1], v[1], v[2]),
+    global_store_b32(addr=v[0], data=v[1], saddr=s[0:1]),
+    s_endpgm(),
+  ]
+  sink = UOp.sink(A.base, threads, arg=KernelInfo(name:=f"custom_add_one_{A.size}", estimates=Estimates(ops=A.size, mem=A.size*4*2)))
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg="AMD"), UOp(Ops.LINEAR, src=(*sink.src, sink)), *assemble_insts(insts, name, arch)))
+
+def custom_add_var(A:UOp, B:UOp, arch:str) -> UOp:
+  A,B = A.flatten(), B.flatten()
+  assert A.dtype.base == dtypes.uint32, f"buffer dtype must be uint32, got {A.dtype}"
+  threads = UOp.special(A.size, "lidx0")
+  var = UOp.variable("var", 0, 10)
+  insts = [
+    s_load_b128(s[4:7], s[0:1]),
+    s_load_b32(s[8], s[0:1], offset=0x10), # all threads load the same variable
+    s_waitcnt(lgkmcnt=0),
+    v_lshlrev_b32_e32(v[0], 2, v[0]), # element offset, different per thread
+    global_load_b32(v[1], v[0], saddr=s[6:7]),
+    s_waitcnt(vmcnt=0),
+    v_add_nc_u32_e32(v[1], s[8], v[1]),
+    global_store_b32(addr=v[0], data=v[1], saddr=s[4:5]),
+    s_endpgm(),
+  ]
+  sink = UOp.sink(A.base, B.base, var, threads, arg=KernelInfo(name:=f"custom_add_one_{A.size}"))
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg="AMD"), UOp(Ops.LINEAR, src=(*sink.src, sink)),
+                               *assemble_insts(insts, name, arch, kernarg_size=16)))
+
+class TestCustomKernel(unittest.TestCase):
+  def test_simple(self):
+    a = Tensor.full((16, 16), 1.).contiguous().realize()
+    a = Tensor.custom_kernel(a, fxn=functools.partial(custom_add_one, arch=Device[Device.DEFAULT].renderer.arch))[0]
+    ei = a.schedule()[-1].lower()
+    self.assertEqual(ei.prg.estimates.ops, a.numel())
+    self.assertEqual(ei.prg.estimates.mem, a.nbytes()*2)
+    ei.run()
+    self.assertTrue((a.numpy() == 2.).all())
+
+  def test_variable(self):
+    b = Tensor.full((16, 16), 1, dtype=dtypes.uint32).contiguous().realize()
+    a = Tensor.zeros_like(b).contiguous().realize()
+    a = Tensor.custom_kernel(a, b, fxn=functools.partial(custom_add_var, arch=Device[Device.DEFAULT].renderer.arch))[0]
+    ei = a.schedule()[-1].lower()
+    for i in range(4):
+      ei.run({"var":i})
+      self.assertTrue((a.numpy() == 1+i).all())
+
+if __name__ == "__main__":
+  unittest.main()

extra/assembly/amd/test/test_dsl2.py

@@ -0,0 +1,170 @@
+import unittest
+from extra.assembly.amd.dsl import *
+from extra.assembly.amd.dsl import VDSTYField
+from extra.assembly.amd.autogen.rdna3.enum import VOP1Op, VOP2Op
+from extra.assembly.amd.autogen.rdna3.ins import VOP1
+
+class TestRegisters(unittest.TestCase):
+  def test_vgpr_single(self):
+    self.assertEqual(repr(v[5]), "v[5]")
+    self.assertEqual(v[5].offset, 261)  # 256 + 5
+    self.assertEqual(v[5].sz, 1)
+
+  def test_sgpr_single(self):
+    self.assertEqual(repr(s[10]), "s[10]")
+    self.assertEqual(s[10].offset, 10)
+
+  def test_vgpr_range(self):
+    self.assertEqual(repr(v[0:3]), "v[0:3]")
+    self.assertEqual(v[0:3].offset, 256)
+    self.assertEqual(v[0:3].sz, 4)
+
+  def test_sgpr_range(self):
+    self.assertEqual(repr(s[4:5]), "s[4:5]")
+    self.assertEqual(s[4:5].sz, 2)
+
+  def test_ttmp_reslice(self):
+    # ttmp is src[108:123], so ttmp[0] should be src[108]
+    self.assertEqual(ttmp[0].offset, 108)
+    self.assertEqual(ttmp[1].offset, 109)
+    # ttmp[0:1] is 2 elements (inclusive slicing)
+    self.assertEqual(ttmp[0:1].offset, 108)
+    self.assertEqual(ttmp[0:1].sz, 2)
+    # ttmp[0:1][0] should be src[108]
+    self.assertEqual(ttmp[0:1][0].offset, 108)
+
+  def test_special_regs(self):
+    self.assertEqual(NULL.offset, 124)
+    self.assertEqual(M0.offset, 125)
+    self.assertEqual(EXEC_LO.offset, 126)
+    self.assertEqual(EXEC_HI.offset, 127)
+    # Check repr round-trips
+    self.assertEqual(repr(NULL), "NULL")
+    self.assertEqual(repr(M0), "M0")
+    self.assertEqual(repr(EXEC_LO), "EXEC_LO")
+    self.assertEqual(repr(EXEC), "EXEC")
+
+  def test_vcc(self):
+    self.assertEqual(VCC.offset, 106)
+    self.assertEqual(VCC.sz, 2)
+    self.assertEqual(VCC_LO.offset, 106)
+    self.assertEqual(VCC_HI.offset, 107)
+    # Check repr round-trips
+    self.assertEqual(repr(VCC_LO), "VCC_LO")
+    self.assertEqual(repr(VCC_HI), "VCC_HI")
+    self.assertEqual(repr(VCC), "VCC")
+
+  def test_float_constants(self):
+    self.assertEqual(src[240].offset, 240)
+    self.assertEqual(repr(src[240]), "0.5")
+    self.assertEqual(repr(src[242]), "1.0")
+    self.assertEqual(repr(src[243]), "-1.0")
+
+  def test_int_constants(self):
+    self.assertEqual(repr(src[128]), "0")
+    self.assertEqual(repr(src[129]), "1")
+    self.assertEqual(repr(src[192]), "64")
+    self.assertEqual(repr(src[193]), "-1")
+    self.assertEqual(repr(src[208]), "-16")
+
+class TestEnumBitField(unittest.TestCase):
+  def test_enum_name(self):
+    self.assertEqual(VOP1Op.V_MOV_B32_E32.name, "V_MOV_B32_E32")
+
+  def test_enum_value(self):
+    self.assertEqual(VOP1Op.V_MOV_B32_E32.value, 1)
+
+  def test_enum_comparison(self):
+    self.assertEqual(VOP1Op.V_MOV_B32_E32, VOP1Op.V_MOV_B32_E32)
+    self.assertNotEqual(VOP1Op.V_NOP_E32, VOP1Op.V_MOV_B32_E32)
+
+  def test_enum_different_types(self):
+    # VOP1Op and VOP2Op are different enums, even if same value
+    self.assertNotEqual(VOP1Op.V_MOV_B32_E32, VOP2Op.V_CNDMASK_B32_E32)
+
+  def test_wrong_enum_type_raises(self):
+    # Passing VOP2Op to VOP1 should raise
+    with self.assertRaises(RuntimeError):
+      VOP1(VOP2Op.V_CNDMASK_B32_E32, v[5], v[6])
+
+class TestVOP1(unittest.TestCase):
+  def test_class_setup(self):
+    self.assertEqual(VOP1._size(), 4)
+    field_names = [n for n, _ in VOP1._fields]
+    self.assertIn('encoding', field_names)
+    self.assertIn('op', field_names)
+    self.assertIn('vdst', field_names)
+    self.assertIn('src0', field_names)
+
+  def test_encoding_vgpr_vgpr(self):
+    i = VOP1(VOP1Op.V_MOV_B32_E32, v[5], v[6])
+    raw = i._raw
+    # Check each field
+    self.assertEqual((raw >> 25) & 0x7f, 0b0111111)  # encoding
+    self.assertEqual((raw >> 17) & 0xff, 5)          # vdst (just VGPR index)
+    self.assertEqual((raw >> 9) & 0xff, 1)           # op
+    self.assertEqual(raw & 0x1ff, 262)               # src0 (256 + 6)
+
+  def test_encoding_vgpr_sgpr(self):
+    i = VOP1(VOP1Op.V_MOV_B32_E32, v[5], s[10])
+    raw = i._raw
+    self.assertEqual((raw >> 17) & 0xff, 5)    # vdst (just VGPR index)
+    self.assertEqual(raw & 0x1ff, 10)          # src0 (SGPR encoded)
+
+  def test_to_bytes(self):
+    i = VOP1(VOP1Op.V_MOV_B32_E32, v[5], v[6])
+    b = i.to_bytes()
+    self.assertEqual(len(b), 4)
+    self.assertEqual(int.from_bytes(b, 'little'), i._raw)
+
+  def test_from_bytes(self):
+    i1 = VOP1(VOP1Op.V_MOV_B32_E32, v[5], v[6])
+    i2 = VOP1.from_bytes(i1.to_bytes())
+    self.assertEqual(i1._raw, i2._raw)
+
+  def test_repr(self):
+    i = VOP1(VOP1Op.V_MOV_B32_E32, v[5], v[6])
+    self.assertEqual(repr(i), "v_mov_b32_e32(v[5], v[6])")
+
+  def test_repr_sgpr_src(self):
+    i = VOP1(VOP1Op.V_MOV_B32_E32, v[5], s[10])
+    self.assertEqual(repr(i), "v_mov_b32_e32(v[5], s[10])")
+
+  def test_kwargs(self):
+    i1 = VOP1(VOP1Op.V_MOV_B32_E32, v[5], v[6])
+    i2 = VOP1(op=VOP1Op.V_MOV_B32_E32, vdst=v[5], src0=v[6])
+    self.assertEqual(i1._raw, i2._raw)
+
+  def test_kwargs_partial(self):
+    i1 = VOP1(VOP1Op.V_MOV_B32_E32, v[5], v[6])
+    i2 = VOP1(VOP1Op.V_MOV_B32_E32, src0=v[6], vdst=v[5])
+    self.assertEqual(i1._raw, i2._raw)
+
+class TestVDSTYField(unittest.TestCase):
+  def test_encode_even_vgpr(self):
+    f = VDSTYField(6, 0)  # 7-bit field
+    self.assertEqual(f.encode(v[0]), 0)
+    self.assertEqual(f.encode(v[2]), 1)
+    self.assertEqual(f.encode(v[4]), 2)
+    self.assertEqual(f.encode(v[254]), 127)
+
+  def test_encode_non_vgpr_raises(self):
+    f = VDSTYField(6, 0)
+    with self.assertRaises(ValueError) as ctx:
+      f.encode(s[0])
+    self.assertIn("VGPR", str(ctx.exception))
+
+  def test_encode_non_reg_raises(self):
+    f = VDSTYField(6, 0)
+    with self.assertRaises(TypeError) as ctx:
+      f.encode(42)
+    self.assertIn("Reg", str(ctx.exception))
+
+  def test_decode_returns_raw(self):
+    f = VDSTYField(6, 0)
+    # decode returns raw value, actual vdsty computed with vdstx context
+    self.assertEqual(f.decode(0), 0)
+    self.assertEqual(f.decode(127), 127)
+
+if __name__ == "__main__":
+  unittest.main()

extra/assembly/amd/test/test_emu2_pcode.py

@@ -0,0 +1,332 @@
+"""Tests for the pcode parser."""
+import unittest
+from collections import defaultdict
+from tinygrad.helpers import DEBUG
+from tinygrad.dtype import dtypes
+from tinygrad.uop.ops import UOp, Ops
+from extra.assembly.amd.emu import parse_pcode
+from extra.assembly.amd.pcode import parse_expr
+from extra.assembly.amd.autogen.rdna3.str_pcode import PCODE
+from extra.assembly.amd.autogen.rdna3.enum import VOP1Op, VOP2Op, VOP3Op, SOP1Op, SOP2Op, DSOp
+
+def _srcs():
+  """Create minimal source variables for pcode parsing."""
+  u32 = lambda v=0: UOp.const(dtypes.uint32, v)
+  return {'S0': u32(), 'S1': u32(), 'S2': u32(), 'SCC': u32(), 'VCC': UOp.const(dtypes.uint64, 0), 'laneId': u32()}
+
+class TestBasicParsing(unittest.TestCase):
+  """Test basic pcode parsing for common instruction patterns."""
+
+  def test_v_add_f32(self):
+    """Test parsing V_ADD_F32 pcode."""
+    _, assigns = parse_pcode(PCODE[VOP2Op.V_ADD_F32_E32], _srcs())
+    self.assertEqual(len(assigns), 1)
+    dest, _ = assigns[0]
+    self.assertTrue(dest.startswith('D0'))
+
+  def test_v_lshlrev_b32(self):
+    """Test parsing V_LSHLREV_B32 pcode."""
+    _, assigns = parse_pcode(PCODE[VOP2Op.V_LSHLREV_B32_E32], _srcs())
+    self.assertEqual(len(assigns), 1)
+
+  def test_s_cselect_b32(self):
+    """Test parsing S_CSELECT_B32 pcode with ternary."""
+    _, assigns = parse_pcode(PCODE[SOP2Op.S_CSELECT_B32], _srcs())
+    self.assertEqual(len(assigns), 1)
+
+  def test_v_add_co_ci_u32(self):
+    """Test parsing V_ADD_CO_CI_U32 with carry."""
+    _, assigns = parse_pcode(PCODE[VOP2Op.V_ADD_CO_CI_U32_E32], _srcs())
+    self.assertGreaterEqual(len(assigns), 1)
+
+class TestWithSources(unittest.TestCase):
+  """Test pcode parsing with actual source operand values."""
+
+  def test_v_add_f32_with_sources(self):
+    """Test V_ADD_F32 with actual float constants."""
+    s0 = UOp.const(dtypes.uint32, 0x3f800000)  # 1.0f
+    s1 = UOp.const(dtypes.uint32, 0x40000000)  # 2.0f
+    _, assigns = parse_pcode(PCODE[VOP2Op.V_ADD_F32_E32], {'S0': s0, 'S1': s1})
+    self.assertEqual(len(assigns), 1)
+    dest, val = assigns[0]
+    self.assertTrue(dest.startswith('D0'))
+    # Result should be an ADD operation
+    self.assertEqual(val.op, Ops.ADD)
+
+  def test_v_mul_f32_with_sources(self):
+    """Test V_MUL_F32 with actual float constants."""
+    s0 = UOp.const(dtypes.uint32, 0x40000000)  # 2.0f
+    s1 = UOp.const(dtypes.uint32, 0x40400000)  # 3.0f
+    _, assigns = parse_pcode(PCODE[VOP2Op.V_MUL_F32_E32], {'S0': s0, 'S1': s1})
+    self.assertEqual(len(assigns), 1)
+    dest, val = assigns[0]
+    self.assertEqual(val.op, Ops.MUL)
+
+class TestParseExpr(unittest.TestCase):
+  """Test the parse_expr function directly."""
+
+  def test_integer_literals(self):
+    """Test parsing integer literals."""
+    self.assertEqual(parse_expr('0', {}).arg, 0)
+    self.assertEqual(parse_expr('42', {}).arg, 42)
+    self.assertEqual(parse_expr('42U', {}).arg, 42)
+
+  def test_negative_integers(self):
+    """Test parsing negative integer literals."""
+    result = parse_expr('-1', {})
+    self.assertEqual(result.arg, -1)
+    self.assertEqual(result.dtype, dtypes.int)
+
+  def test_float_literals(self):
+    """Test parsing float literals."""
+    result = parse_expr('1.0F', {})
+    self.assertEqual(result.arg, 1.0)
+    self.assertEqual(result.dtype, dtypes.float32)
+
+  def test_hex_literals(self):
+    """Test parsing hex literals."""
+    result = parse_expr('0xFF', {})
+    self.assertEqual(result.arg, 255)
+
+  def test_variable_lookup(self):
+    """Test variable lookup in parse_expr."""
+    vars = {'x': UOp.const(dtypes.uint32, 42)}
+    result = parse_expr('x', vars)
+    self.assertEqual(result.arg, 42)
+
+  def test_binary_ops(self):
+    """Test parsing binary operations."""
+    vars = {'a': UOp.const(dtypes.uint32, 10), 'b': UOp.const(dtypes.uint32, 5)}
+
+    # Addition
+    result = parse_expr('a + b', vars)
+    self.assertEqual(result.op, Ops.ADD)
+
+    # Subtraction with constant folding
+    result = parse_expr('10 - 5', {})
+    self.assertEqual(result.op, Ops.CONST)
+    self.assertEqual(result.arg, 5)
+
+  def test_ternary(self):
+    """Test parsing ternary expressions."""
+    vars = {'cond': UOp.const(dtypes.bool, True), 'a': UOp.const(dtypes.uint32, 1), 'b': UOp.const(dtypes.uint32, 0)}
+    result = parse_expr('cond ? a : b', vars)
+    self.assertEqual(result.op, Ops.WHERE)
+
+class TestForLoopParsing(unittest.TestCase):
+  """Test for loop parsing (CLZ/CTZ patterns)."""
+
+  def test_clz_pcode_exists(self):
+    """Verify CLZ pcode is available."""
+    pcode = PCODE.get(VOP1Op.V_CLZ_I32_U32_E32)
+    self.assertIsNotNone(pcode)
+    self.assertIn('for', pcode.lower())
+
+  def test_clz_parsing(self):
+    """Test CLZ pcode parsing produces correct structure."""
+    pcode = PCODE[VOP1Op.V_CLZ_I32_U32_E32]
+    S0 = UOp.const(dtypes.uint32, 0xFFFFFFFF)  # All ones - CLZ should be 0
+    vars, assigns = parse_pcode(pcode, {'S0': S0})
+
+    self.assertEqual(len(assigns), 1)
+    dest, val = assigns[0]
+    self.assertTrue(dest.startswith('D0'))
+    # Result should be a nested WHERE structure
+    self.assertEqual(val.op, Ops.WHERE)
+
+  def test_clz_with_zero(self):
+    """Test CLZ with input 0 - should return -1."""
+    pcode = PCODE[VOP1Op.V_CLZ_I32_U32_E32]
+    S0 = UOp.const(dtypes.uint32, 0)
+    vars, assigns = parse_pcode(pcode, {'S0': S0})
+
+    # Check that the innermost value (default) is -1 (may be wrapped in CAST)
+    val = assigns[0][1]
+    # Traverse to innermost WHERE
+    while val.op == Ops.WHERE:
+      val = val.src[2]  # false branch
+    # Unwrap CAST if present
+    while val.op == Ops.CAST:
+      val = val.src[0]
+    self.assertEqual(val.arg, -1)
+
+  def test_ctz_parsing(self):
+    """Test CTZ pcode parsing."""
+    pcode = PCODE.get(VOP1Op.V_CTZ_I32_B32_E32)
+    if pcode is None:
+      self.skipTest("V_CTZ_I32_B32_E32 pcode not available")
+
+    S0 = UOp.const(dtypes.uint32, 1)  # LSB set - CTZ should be 0
+    vars, assigns = parse_pcode(pcode, {'S0': S0})
+    self.assertEqual(len(assigns), 1)
+
+class TestDSPcodePatterns(unittest.TestCase):
+  """Test DS instruction pcode patterns."""
+
+  def test_ds_load_b32_pcode(self):
+    """Test DS_LOAD_B32 pcode is parseable."""
+    pcode = PCODE.get(DSOp.DS_LOAD_B32)
+    self.assertIsNotNone(pcode)
+    self.assertIn('RETURN_DATA', pcode)
+    self.assertIn('MEM[', pcode)
+
+  def test_ds_store_b32_pcode(self):
+    """Test DS_STORE_B32 pcode is parseable."""
+    pcode = PCODE.get(DSOp.DS_STORE_B32)
+    self.assertIsNotNone(pcode)
+    self.assertIn('MEM[', pcode)
+    self.assertIn('DATA', pcode)
+
+  def test_mem_read_parsing(self):
+    """Test MEM[addr].type read expression parsing."""
+    # Create a mock LDS buffer
+    lds = UOp(Ops.PARAM, dtypes.uint32.ptr(16384), arg=3)
+    addr = UOp.const(dtypes.uint32, 0)
+    vars = {'_lds': lds, 'ADDR': addr, 'OFFSET': UOp.const(dtypes.uint32, 0)}
+
+    result = parse_expr('MEM[ADDR + OFFSET].b32', vars)
+    # Should be an INDEX operation into LDS
+    self.assertIsNotNone(result)
+
+  def test_ds_store_2addr_b32_parsing(self):
+    """Test DS_STORE_2ADDR_B32 pcode parsing produces MEM writes."""
+    pcode = PCODE.get(DSOp.DS_STORE_2ADDR_B32)
+    self.assertIsNotNone(pcode)
+    srcs = {
+      'ADDR': UOp.const(dtypes.uint32, 0),
+      'OFFSET0': UOp.const(dtypes.uint32, 0),
+      'OFFSET1': UOp.const(dtypes.uint32, 1),
+      'DATA': UOp.const(dtypes.uint32, 0xAAAAAAAA),
+      'DATA2': UOp.const(dtypes.uint32, 0xBBBBBBBB),
+    }
+    srcs['laneId'] = UOp.const(dtypes.uint32, 0)
+    _, assigns = parse_pcode(pcode, srcs)
+    # Should have 2 MEM write assignments
+    self.assertEqual(len(assigns), 2)
+    for dest, val in assigns:
+      self.assertTrue(dest.startswith('MEM['))
+      # val should be (addr, write_val) tuple
+      self.assertIsInstance(val, tuple)
+      self.assertEqual(len(val), 2)
+
+  def test_ds_load_2addr_b32_parsing(self):
+    """Test DS_LOAD_2ADDR_B32 pcode parsing produces RETURN_DATA assignments."""
+    pcode = PCODE.get(DSOp.DS_LOAD_2ADDR_B32)
+    self.assertIsNotNone(pcode)
+    lds = UOp(Ops.PARAM, dtypes.uint32.ptr(16384), arg=3)
+    srcs = {
+      'ADDR': UOp.const(dtypes.uint32, 0),
+      'OFFSET0': UOp.const(dtypes.uint32, 0),
+      'OFFSET1': UOp.const(dtypes.uint32, 1),
+      '_lds': lds,
+    }
+    srcs['laneId'] = UOp.const(dtypes.uint32, 0)
+    _, assigns = parse_pcode(pcode, srcs)
+    # Should have 2 RETURN_DATA assignments
+    self.assertEqual(len(assigns), 2)
+    self.assertEqual(assigns[0][0], 'RETURN_DATA[31:0]')
+    self.assertEqual(assigns[1][0], 'RETURN_DATA[63:32]')
+
+  def test_ds_store_address_calculation(self):
+    """Test DS_STORE_2ADDR_B32 calculates correct addresses (offset * 4)."""
+    pcode = PCODE.get(DSOp.DS_STORE_2ADDR_B32)
+    srcs = {
+      'ADDR': UOp.const(dtypes.uint32, 100),
+      'OFFSET0': UOp.const(dtypes.uint32, 2),
+      'OFFSET1': UOp.const(dtypes.uint32, 5),
+      'DATA': UOp.const(dtypes.uint32, 0xAAAAAAAA),
+      'DATA2': UOp.const(dtypes.uint32, 0xBBBBBBBB),
+    }
+    srcs['laneId'] = UOp.const(dtypes.uint32, 0)
+    _, assigns = parse_pcode(pcode, srcs)
+    # Check addresses: 100 + 2*4 = 108, 100 + 5*4 = 120
+    addr0, _ = assigns[0][1]
+    addr1, _ = assigns[1][1]
+    self.assertEqual(addr0.simplify().arg, 108)
+    self.assertEqual(addr1.simplify().arg, 120)
+
+  def test_ds_store_data_values(self):
+    """Test DS_STORE_2ADDR_B32 uses correct data values."""
+    pcode = PCODE.get(DSOp.DS_STORE_2ADDR_B32)
+    srcs = {
+      'ADDR': UOp.const(dtypes.uint32, 0),
+      'OFFSET0': UOp.const(dtypes.uint32, 0),
+      'OFFSET1': UOp.const(dtypes.uint32, 1),
+      'DATA': UOp.const(dtypes.uint32, 0xAAAAAAAA),
+      'DATA2': UOp.const(dtypes.uint32, 0xBBBBBBBB),
+    }
+    srcs['laneId'] = UOp.const(dtypes.uint32, 0)
+    _, assigns = parse_pcode(pcode, srcs)
+    _, val0 = assigns[0][1]
+    _, val1 = assigns[1][1]
+    # DATA[31:0] should preserve the value
+    self.assertEqual(val0.simplify().arg, 0xAAAAAAAA)
+    self.assertEqual(val1.simplify().arg, 0xBBBBBBBB)
+
+class TestConditionalParsing(unittest.TestCase):
+  """Test conditional (if/elsif/else) pcode parsing."""
+
+  def test_ternary_in_assignment(self):
+    """Test parsing ternary expression (which becomes WHERE)."""
+    # S_CSELECT_B32: D0.u32 = SCC ? S0.u32 : S1.u32
+    pcode = PCODE[SOP2Op.S_CSELECT_B32]
+    s0 = UOp.const(dtypes.uint32, 10)
+    s1 = UOp.const(dtypes.uint32, 20)
+    scc = UOp.const(dtypes.uint32, 1)
+    vars, assigns = parse_pcode(pcode, {'S0': s0, 'S1': s1, 'SCC': scc})
+    self.assertEqual(len(assigns), 1)
+    dest, val = assigns[0]
+    self.assertTrue(dest.startswith('D0'))
+    # Result should be a WHERE (ternary becomes WHERE)
+    self.assertEqual(val.op, Ops.WHERE)
+
+class TestAllPcode(unittest.TestCase):
+  """Test that all pcode from all architectures can be parsed."""
+
+  def _make_srcs(self):
+    """Create dummy source variables for pcode parsing."""
+    u32, u64 = lambda v=0: UOp.const(dtypes.uint32, v), lambda v=0: UOp.const(dtypes.uint64, v)
+    lds = UOp(Ops.PARAM, dtypes.uint32.ptr(16384), arg=3)
+    return {'laneId': u32(), 'laneID': u32(), 'S0': u32(), 'S1': u32(), 'S2': u32(), 'S3': u32(), 'SRC0': u32(),
+            'D0': u32(), 'D1': u32(), 'DST': u32(), 'VDST': u32(), 'SDST': u32(),
+            'VCC': u64(), 'VCCZ': u32(), 'EXEC': u64(), 'EXEC_LO': u32(), 'EXECZ': u32(), 'SCC': u32(),
+            'SIMM16': u32(), 'SIMM32': u32(), 'OFFSET': u32(), 'OFFSET0': u32(), 'OFFSET1': u32(), 'offset1': u32(),
+            'ADDR': u32(), 'ADDR_BASE': u32(), 'TADDR': u32(), 'DATA': u32(), 'DATA0': u32(), 'DATA1': u32(), 'DATA2': u32(),
+            'VDATA': u32(), 'VDATA0': u32(), 'VDATA1': u32(), 'VDATA2': u32(), 'VDATA3': u32(),
+            'OPSEL': u32(), 'OPSEL_HI': u32(), 'NEG': u32(), 'NEG_HI': u32(), 'CLAMP': u32(),
+            'M0': u32(), 'PC': u64(), 'DENORM': u32(1), 'ROUND_MODE': u32(), 'ROUND_TOWARD_ZERO': u32(), 'ROUND_NEAREST_EVEN': u32(), 'WAVE_STATUS': u32(),
+            'MAX_FLOAT_F32': u32(0x7f7fffff), 'Unsigned': u32(1), 'clampedLOD': u32(),
+            '_lds': lds, '_vmem': lds, '_active': UOp.const(dtypes.bool, True)}
+
+  def _parse_all_pcode(self, pcode_dict, arch: str, min_pct: float):
+    """Parse all pcode. RuntimeError = parser limitation (ok), other exceptions = real bugs."""
+    srcs = self._make_srcs()
+    passed, skipped, errors = 0, 0, defaultdict(list)
+    for op, pcode in pcode_dict.items():
+      try:
+        parse_pcode(pcode, srcs)
+        passed += 1
+      except RuntimeError as e: skipped += 1; errors[str(e)].append(op.name)
+      except Exception as e: self.fail(f"[{arch}] {op.name}: {e}\nPcode: {pcode[:200]}")
+    total = len(pcode_dict)
+    pct = 100 * passed / total
+    print(f"{arch}: {passed}/{total} ({pct:.1f}%) parsed, {skipped} skipped")
+    if DEBUG >= 2:
+      for err, ops in sorted(errors.items(), key=lambda x: -len(x[1])):
+        print(f"  {err}: {', '.join(ops[:5])}{'...' if len(ops) > 5 else ''} ({len(ops)})")
+    self.assertGreaterEqual(pct, min_pct, f"[{arch}] {pct:.1f}% < {min_pct}% threshold")
+
+  def test_parse_all_cdna_pcode(self):
+    from extra.assembly.amd.autogen.cdna.str_pcode import PCODE as CDNA_PCODE
+    self._parse_all_pcode(CDNA_PCODE, "CDNA", min_pct=60)
+
+  def test_parse_all_rdna3_pcode(self):
+    from extra.assembly.amd.autogen.rdna3.str_pcode import PCODE as RDNA3_PCODE
+    self._parse_all_pcode(RDNA3_PCODE, "RDNA3", min_pct=90)
+
+  def test_parse_all_rdna4_pcode(self):
+    from extra.assembly.amd.autogen.rdna4.str_pcode import PCODE as RDNA4_PCODE
+    self._parse_all_pcode(RDNA4_PCODE, "RDNA4", min_pct=65)
+
+if __name__ == "__main__":
+  unittest.main()

extra/assembly/amd/test/test_formats.py

@@ -0,0 +1,229 @@
+#!/usr/bin/env python3
+"""Test DS and other compute-relevant instruction formats.
+
+Note: Graphics-only formats (EXP, MUBUF, MTBUF, MIMG) are not supported - use GLOBAL/FLAT for memory access in compute.
+"""
+import unittest
+from extra.assembly.amd.autogen.rdna3.ins import *
+from extra.assembly.amd.dsl import VCC_HI, EXEC_LO, NULL
+OFF = NULL  # OFF is alias for NULL
+from extra.assembly.amd import detect_format
+
+
+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[0:1], 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[1:2], 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[0:1], 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[0:1])
+    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[1:2], 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[0:1], addr=v[2:3], 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[2:3], 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[2:3], 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[2:3], 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]
+    inst = VOP3(VOP3Op.V_ADD3_U32, vdst=v[5], src0=VCC_HI, 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]
+    inst = VOP3(VOP3Op.V_ADD3_U32, vdst=v[5], src0=NULL, src1=EXEC_LO, 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)
+    inst = VOP3P(VOP3POp.V_PK_ADD_F16, vdst=v[5], src0=0.5, 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(sdata=s[0], sbase=s[2:3], offset=0).to_bytes()), SMEM)
+    self.assertEqual(detect_format(s_load_b64(sdata=s[0:1], sbase=s[2:3], soffset=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(vdst=v[0], addr=v[1:2], saddr=NULL).to_bytes()), GLOBAL)
+    self.assertEqual(detect_format(global_store_b32(addr=v[0:1], data=v[2], saddr=NULL).to_bytes()), GLOBAL)
+
+  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

@@ -0,0 +1,192 @@
+# do not change these tests. we need to fix bugs to make them pass
+# the Inst constructor should be looking at the types of the fields to correctly set the value
+
+import unittest, struct
+from extra.assembly.amd.autogen.rdna3.ins import *
+from extra.assembly.amd.dsl import Inst
+from extra.assembly.amd.test.test_roundtrip import compile_asm
+
+class IntegrationTestBase(unittest.TestCase):
+  inst: Inst
+  arch: str
+  def tearDown(self):
+    if not hasattr(self, 'inst'): return
+    b = self.inst.to_bytes()
+    st = self.inst.disasm()
+    # 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)
+    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):
+      self.inst = s_load_b128(s[4:6], s[0:1], NULL, 0)
+
+  def test_mov_b32(self):
+    self.inst = s_mov_b32(s[80], s[0])
+
+  def test_mov_b64(self):
+    self.inst = s_mov_b64(s[80:81], s[0:1])
+
+  def test_mov_b32_wrong(self):
+    with self.assertRaises(Exception):
+      self.inst = s_mov_b32(s[80:81], s[0:1])
+    with self.assertRaises(Exception):
+      self.inst = s_mov_b32(s[80:81], s[0])
+    with self.assertRaises(Exception):
+      self.inst = s_mov_b32(s[80], s[0:1])
+
+  def test_mov_b64_wrong(self):
+    with self.assertRaises(Exception):
+      self.inst = s_mov_b64(s[80], s[0])
+    with self.assertRaises(Exception):
+      self.inst = s_mov_b64(s[80], s[0:1])
+    with self.assertRaises(Exception):
+      self.inst = s_mov_b64(s[80:81], s[0])
+
+  def test_load_b128_no_0(self):
+    self.inst = s_load_b128(s[4:7], s[0:1], NULL)
+
+  def test_load_b128_s(self):
+    self.inst = s_load_b128(s[4:7], s[0:1], s[8], 0)
+
+  def test_load_b128_v(self):
+    with self.assertRaises(TypeError):
+      self.inst = s_load_b128(s[4:7], s[0:1], v[8], 0)
+
+  def test_load_b128_off(self):
+    self.inst = s_load_b128(s[4:7], s[0:1], NULL, 3)
+
+  def test_simple_stos(self):
+    self.inst = s_mov_b32(s[0], s[1])
+
+  def test_simple_wrong(self):
+    with self.assertRaises(TypeError):
+      self.inst = s_mov_b32(v[0], s[1])
+
+  def test_simple_vtov(self):
+    self.inst = v_mov_b32_e32(v[0], v[1])
+
+  def test_simple_stov(self):
+    self.inst = v_mov_b32_e32(v[0], s[2])
+
+  def test_simple_float_to_v(self):
+    self.inst = v_mov_b32_e32(v[0], 1.0)
+
+  def test_simple_v_to_float(self):
+    with self.assertRaises(TypeError):
+      self.inst = v_mov_b32_e32(1, v[0])
+
+  def test_simple_int_to_v(self):
+    self.inst = v_mov_b32_e32(v[0], 1)
+
+  def test_three_add(self):
+    self.inst = v_add_co_ci_u32_e32(v[3], s[7], v[3])
+
+  def test_three_add_v(self):
+    self.inst = v_add_co_ci_u32_e32(v[3], v[7], v[3])
+
+  def test_three_add_const(self):
+    self.inst = v_add_co_ci_u32_e32(v[3], 2.0, v[3])
+
+  def test_swaitcnt_lgkm(self): self.inst = s_waitcnt(0xfc07)
+  def test_swaitcnt_vm(self): self.inst = s_waitcnt(0x03f7)
+
+  def test_vmad(self):
+    self.inst = v_mad_u64_u32(v[1:2], NULL, s[2], 3, v[1:2])
+
+  def test_large_imm(self):
+    self.inst = v_mov_b32_e32(v[0], 0x1234)
+
+  def test_dual_mov(self):
+    self.inst = VOPD(VOPDOp.V_DUAL_MOV_B32, VOPDOp.V_DUAL_MOV_B32, vdstx=v[0], vdsty=v[1], srcx0=v[2], srcy0=v[4])
+
+  def test_dual_mul(self):
+    self.inst = v_dual_mul_f32(VOPDOp.V_DUAL_MUL_F32, vdstx=v[0], vdsty=v[1], srcx0=v[2], vsrcx1=v[3], srcy0=v[4], vsrcy1=v[5])
+
+  def test_simple_int_to_s(self):
+    self.inst = s_mov_b32(s[0], 3)
+
+  def test_complex_int_to_s(self):
+    self.inst = s_mov_b32(s[0], 0x235646)
+
+  def test_simple_float_to_s(self):
+    self.inst = s_mov_b32(s[0], 1.0)
+
+  def test_complex_float_to_s(self):
+    self.inst = s_mov_b32(s[0], 1337.0)
+    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 extra.assembly.amd.autogen.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 extra.assembly.amd.autogen.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).
+
+  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.sz, 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.sz, 4, "s[4:7] from disasm should give 4 registers")
+
+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")
+
+
+if __name__ == "__main__":
+  unittest.main()

extra/assembly/amd/test/test_integration.py

@@ -0,0 +1,258 @@
+#!/usr/bin/env python3
+"""Integration test: round-trip RDNA3 assembly through AMD toolchain."""
+import unittest, io, sys
+from extra.assembly.amd.autogen.rdna3.ins import *
+
+def waitcnt(vmcnt: int = 0x3f, expcnt: int = 0x7, lgkmcnt: int = 0x3f) -> int:
+  return (expcnt & 0x7) | ((lgkmcnt & 0x3f) << 4) | ((vmcnt & 0x3f) << 10)
+
+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 DSL 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[0:1], NULL),
+      s_waitcnt(simm16=waitcnt(lgkmcnt=0)),
+      global_store_b32(addr=v[0:1], 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[0:1], 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[0:1], 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 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 +1,183 @@
+#!/usr/bin/env python3
+"""Test AMD assembler/disassembler against LLVM test vectors.
+
+Only compute-relevant instruction formats are tested. Graphics-only formats not supported:
+  - MUBUF/MTBUF: buffer instructions with resource descriptors (use GLOBAL/FLAT instead)
+  - MIMG: image/texture instructions
+  - EXP/VEXPORT: export instructions for pixel/vertex output
+  - VIMAGE/VSAMPLE: image sampling instructions (RDNA4)
+  - VBUFFER: buffer instructions (RDNA4)
+"""
+import unittest, re, subprocess, functools
+from tinygrad.helpers import fetch
+from extra.assembly.amd.disasm import disasm
+from extra.assembly.amd import decode_inst, detect_format
+from extra.assembly.amd.test.helpers import get_llvm_mc, get_target, get_mattr
+
+LLVM_BASE = "https://raw.githubusercontent.com/llvm/llvm-project/llvmorg-21.1.0/llvm/test/MC/AMDGPU"
+
+# RDNA3 (gfx11) test files for compute instructions
+# Excluded: gfx11_asm_mubuf.s, gfx11_asm_mtbuf.s, gfx11_asm_mimg.s, gfx11_asm_mubuf_alias.s, gfx11_asm_mtbuf_alias.s (graphics-only)
+RDNA_FILES = ['gfx11_asm_sop1.s', 'gfx11_asm_sop2.s', 'gfx11_asm_sopp.s', 'gfx11_asm_sopk.s', 'gfx11_asm_sopc.s',
+  'gfx11_asm_vop1.s', 'gfx11_asm_vop2.s', 'gfx11_asm_vopc.s', 'gfx11_asm_vop3.s', 'gfx11_asm_vop3p.s', 'gfx11_asm_vinterp.s',
+  'gfx11_asm_vopd.s', 'gfx11_asm_vopcx.s', 'gfx11_asm_vop3_from_vop1.s', 'gfx11_asm_vop3_from_vop2.s', 'gfx11_asm_vop3_from_vopc.s',
+  'gfx11_asm_vop3_from_vopcx.s', 'gfx11_asm_ds.s', 'gfx11_asm_smem.s', 'gfx11_asm_flat.s',
+  'gfx11_asm_wmma.s', 'gfx11_asm_vop3_features.s', 'gfx11_asm_vop3p_features.s', 'gfx11_asm_vopd_features.s',
+  'gfx11_asm_vop3_alias.s', 'gfx11_asm_vop3p_alias.s', 'gfx11_asm_vopc_alias.s', 'gfx11_asm_vopcx_alias.s', 'gfx11_asm_vinterp_alias.s',
+  'gfx11_asm_smem_alias.s']
+# CDNA (gfx9/gfx90a/gfx942/gfx950) test files for compute instructions
+# Excluded: gfx9_asm_mubuf.s, gfx9_asm_mtbuf.s, gfx90a_ldst_acc.s (has MIMG mixed in)
+# Exclude gfx90a: 'gfx90a_asm_features.s', 'mai-gfx90a.s',
+# Exclude gfx950: 'gfx950_asm_features.s' (disasm error)
+CDNA_FILES = ['gfx9_asm_sop1.s', 'gfx9_asm_sop2.s', 'gfx9_asm_sopp.s', 'gfx9_asm_sopk.s', 'gfx9_asm_sopc.s',
+  'gfx9_asm_vop1.s', 'gfx9_asm_vop2.s', 'gfx9_asm_vopc.s', 'gfx9_asm_vop3.s', 'gfx9_asm_vop3p.s',
+  'gfx9_asm_ds.s', 'gfx9_asm_flat.s', 'gfx9_asm_smem.s',
+  'flat-scratch-gfx942.s', 'gfx942_asm_features.s', 'mai-gfx942.s',
+  'gfx950_asm_vop1.s', 'gfx950_asm_read_tr.s', 'mai-gfx950.s']
+# RDNA4 (gfx12) test files for compute instructions
+# Excluded: gfx12_asm_vbuffer_mubuf.s, gfx12_asm_vbuffer_mtbuf.s, gfx12_asm_exp.s (graphics-only)
+RDNA4_FILES = ['gfx12_asm_sop1.s', 'gfx12_asm_sop2.s', 'gfx12_asm_sopp.s', 'gfx12_asm_sopk.s', 'gfx12_asm_sopc.s',
+  'gfx12_asm_vop1.s', 'gfx12_asm_vop2.s', 'gfx12_asm_vopc.s', 'gfx12_asm_vopcx.s', 'gfx12_asm_vop3.s', 'gfx12_asm_vop3c.s',
+  'gfx12_asm_vop3cx.s', 'gfx12_asm_vop3p.s', 'gfx12_asm_vop3_from_vop1.s', 'gfx12_asm_vop3_from_vop2.s',
+  'gfx12_asm_vop3p_features.s', 'gfx12_asm_vopd.s', 'gfx12_asm_vopd_features.s',
+  'gfx12_asm_ds.s', 'gfx12_asm_smem.s',
+  'gfx12_asm_wmma_w32.s']
+
+def _parse_llvm_tests(text: str, pattern: str) -> list[tuple[str, bytes]]:
+  tests = []
+  for block in text.split('\n\n'):
+    asm_text, encoding = None, None
+    for line in block.split('\n'):
+      line = line.strip()
+      if not line or line.startswith(('.', ';')): continue
+      if not line.startswith('//'):
+        asm_text = line.split('//')[0].strip() or asm_text
+      if m := re.search(pattern + r'[^:]*:.*?(?:encoding:\s*)?\[(0x[0-9a-f,x\s]+)\]', line, re.I):
+        encoding = m.group(1).replace('0x', '').replace(',', '').replace(' ', '')
+    if asm_text and encoding:
+      try: tests.append((asm_text, bytes.fromhex(encoding)))
+      except ValueError: pass
+  return tests
+
+def _get_tests_uncached(f: str, arch: str) -> list[tuple[str, bytes]]:
+  text = fetch(f"{LLVM_BASE}/{f}").read_bytes().decode('utf-8', errors='ignore')
+  if arch == "rdna3":
+    # Match GFX11 and W32 only (wavefront32 mode)
+    tests = _parse_llvm_tests(text, r'(?:GFX11|W32)')
+  elif arch == "rdna4":
+    # Match GFX12 (but not GFX1250) and W32 only (wavefront32 mode)
+    tests = _parse_llvm_tests(text, r'(?:GFX12(?!50)|W32)')
+  elif 'gfx90a' in f or 'gfx942' in f or 'gfx950' in f:
+    tests = _parse_llvm_tests(text, r'(?:GFX90A|GFX942|GFX950)')
+  else:
+    tests = _parse_llvm_tests(text, r'(?:VI9|GFX9|CHECK)')
+  # Exclude v_interp_* (graphics-only, not on CDNA)
+  if arch == "cdna": tests = [(asm, data) for asm, data in tests if not asm.startswith('v_interp_')]
+  # Filter out tests where original ASM isn't valid on target (e.g., gfx9 tests with gfx942/gfx950 constraints)
+  if arch == "cdna" and not ('gfx942' in f or 'gfx950' in f or 'gfx90a' in f): tests = _filter_valid_asm(tests, arch)
+  return tests
+
+@functools.cache
+def _get_tests(f: str, arch: str) -> list[tuple[str, bytes]]: return _get_tests_uncached(f, arch)
+
+def _compile_asm_batch(instrs: list[str], arch: str = "rdna3", mcpu: str|None = None) -> list[bytes]:
+  if not instrs: return []
+  mcpu, mattr = mcpu or get_target(arch), get_mattr(arch)
+  result = subprocess.run([get_llvm_mc(), '-triple=amdgcn', f'-mcpu={mcpu}', f'-mattr={mattr}', '-show-encoding'],
+    input=".text\n" + "\n".join(instrs) + "\n", capture_output=True, text=True, timeout=30)
+  if result.returncode != 0: raise RuntimeError(f"llvm-mc failed: {result.stderr.strip()}")
+  return [bytes.fromhex(line.split('encoding:')[1].strip()[1:-1].replace('0x', '').replace(',', '').replace(' ', ''))
+          for line in result.stdout.split('\n') if 'encoding:' in line]
+
+def _filter_valid_asm(tests: list[tuple[str, bytes]], arch: str) -> list[tuple[str, bytes]]:
+  """Filter out tests where the original ASM isn't valid on the target (e.g., gfx9 tests with gfx942/gfx950 constraints)."""
+  if not tests: return []
+  mcpu = get_target(arch)
+  # Batch assemble all instructions, parse stderr to find which lines failed
+  instrs = [asm for asm, _ in tests]
+  result = subprocess.run([get_llvm_mc(), '-triple=amdgcn', f'-mcpu={mcpu}', '-show-encoding'],
+    input=".text\n" + "\n".join(instrs) + "\n", capture_output=True, text=True, timeout=30)
+  # Parse error lines from stderr (format: "<stdin>:N:..." where N is 1-indexed, line 1 is ".text")
+  failed_lines = set()
+  for line in result.stderr.split('\n'):
+    if m := re.match(r'<stdin>:(\d+):', line): failed_lines.add(int(m.group(1)) - 1)  # -1 for .text, so line 2 -> index 1 -> tests[0]
+  # Also filter out tests where LLVM roundtrip doesn't match original (reserved bits set in original)
+  valid = [(asm, data) for i, (asm, data) in enumerate(tests) if (i + 1) not in failed_lines]
+  if not valid: return []
+  llvm_result = subprocess.run([get_llvm_mc(), '-triple=amdgcn', f'-mcpu={mcpu}', '-show-encoding'],
+    input=".text\n" + "\n".join(asm for asm, _ in valid) + "\n", capture_output=True, text=True, timeout=30)
+  llvm_bytes = [bytes.fromhex(line.split('encoding:')[1].strip()[1:-1].replace('0x', '').replace(',', '').replace(' ', ''))
+                for line in llvm_result.stdout.split('\n') if 'encoding:' in line]
+  return [(asm, data) for (asm, data), lb in zip(valid, llvm_bytes) if lb == data]
+
+def _make_test(f: str, arch: str, test_type: str):
+  def test(self):
+    tests = _get_tests(f, arch)
+    name = f"{arch}_{test_type}_{f}"
+    mcpu = "gfx942" if arch == "cdna" and "gfx942" in f else get_target(arch)
+    if test_type == "roundtrip":
+      passed, skipped = 0, 0
+      for _, data in tests:
+        try:
+          decoded = detect_format(data, arch).from_bytes(data)
+          self.assertEqual(decoded.to_bytes()[:len(data)], data)
+          passed += 1
+        except ValueError: skipped += 1  # skip invalid opcodes not in enum
+      print(f"{name}: {passed} passed, {skipped} skipped")
+      self.assertEqual(skipped, 0, f"{name}: {skipped} tests skipped, expected 0")
+    elif test_type == "repr":
+      # Test that eval(repr(inst)) reproduces the instruction
+      if arch == "rdna3": import extra.assembly.amd.autogen.rdna3.ins as ins
+      elif arch == "rdna4": import extra.assembly.amd.autogen.rdna4.ins as ins
+      elif arch == "cdna": import extra.assembly.amd.autogen.cdna.ins as ins
+      ns = {k: getattr(ins, k) for k in dir(ins) if not k.startswith('_')}
+      passed, skipped = 0, 0
+      for _, data in tests:
+        try:
+          decoded = detect_format(data, arch).from_bytes(data)
+          if decoded.to_bytes()[:len(data)] != data: skipped += 1; continue  # skip if binary roundtrip fails
+          r = repr(decoded)
+          try:
+            decoded2 = eval(r, ns)  # noqa: S307
+            if decoded == decoded2: passed += 1
+            else: skipped += 1
+          except Exception: skipped += 1
+        except ValueError: skipped += 1
+      print(f"{name}: {passed} passed, {skipped} skipped")
+      self.assertEqual(skipped, 0, f"{name}: {skipped} tests skipped, expected 0")
+    elif test_type == "disasm":
+      to_test = []
+      for _, data in tests:
+        try:
+          decoded = decode_inst(data, arch)
+          enc = decoded.to_bytes()[:len(data)]
+          # Skip if roundtrip fails, disasm fails, or op_name is missing (disasm starts with space)
+          if enc == data and (d := disasm(decoded)) and not d.startswith(' '): to_test.append((enc, d))
+        except: pass
+      skipped = len(tests) - len(to_test)
+      print(f"{name}: {len(to_test)} passed, {skipped} skipped")
+      self.assertEqual(skipped, 0, f"{name}: {skipped} tests skipped, expected 0")
+      # Compare disasm->reassemble with original encoding (filter reserved bit cases where LLVM can't reproduce)
+      llvm_bytes = _compile_asm_batch([t[1] for t in to_test], arch, mcpu)
+      valid = [(enc, d, llvm) for (enc, d), llvm in zip(to_test, llvm_bytes) if llvm == enc]
+      print(f"{name}: {len(valid)}/{len(to_test)} matched LLVM encoding")
+      for enc, _, llvm in valid: self.assertEqual(llvm, enc)
+  return test
+
+class TestLLVM(unittest.TestCase): pass
+
+for f in RDNA_FILES:
+  setattr(TestLLVM, f"test_rdna3_roundtrip_{f.replace('.s', '').replace('-', '_')}", _make_test(f, "rdna3", "roundtrip"))
+  setattr(TestLLVM, f"test_rdna3_disasm_{f.replace('.s', '').replace('-', '_')}", _make_test(f, "rdna3", "disasm"))
+  setattr(TestLLVM, f"test_rdna3_repr_{f.replace('.s', '').replace('-', '_')}", _make_test(f, "rdna3", "repr"))
+for f in CDNA_FILES:
+  setattr(TestLLVM, f"test_cdna_roundtrip_{f.replace('.s', '').replace('-', '_')}", _make_test(f, "cdna", "roundtrip"))
+  setattr(TestLLVM, f"test_cdna_disasm_{f.replace('.s', '').replace('-', '_')}", _make_test(f, "cdna", "disasm"))
+  setattr(TestLLVM, f"test_cdna_repr_{f.replace('.s', '').replace('-', '_')}", _make_test(f, "cdna", "repr"))
+for f in RDNA4_FILES:
+  setattr(TestLLVM, f"test_rdna4_roundtrip_{f.replace('.s', '').replace('-', '_')}", _make_test(f, "rdna4", "roundtrip"))
+  setattr(TestLLVM, f"test_rdna4_disasm_{f.replace('.s', '').replace('-', '_')}", _make_test(f, "rdna4", "disasm"))
+  setattr(TestLLVM, f"test_rdna4_repr_{f.replace('.s', '').replace('-', '_')}", _make_test(f, "rdna4", "repr"))
+
+if __name__ == "__main__":
+  unittest.main()

extra/assembly/amd/test/test_mockgpu_invalid.py

@@ -0,0 +1,55 @@
+#!/usr/bin/env python3
+"""Test that invalid instructions raise exceptions through the mock GPU stack."""
+import unittest, subprocess, os, time
+
+class TestMockGPUInvalidInstruction(unittest.TestCase):
+  def test_unsupported_instruction_raises(self):
+    """Test that unsupported instructions raise immediately through the full MOCKGPU stack."""
+    test_code = '''
+import struct
+from tinygrad import Device, Tensor
+from tinygrad.engine.realize import get_runner
+from tinygrad.runtime.ops_amd import AMDProgram
+
+dev = Device["AMD"]
+a = Tensor([1.0]).realize()
+b = a + 1
+si = b.schedule()[-1]
+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
+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)
+    found = True
+    break
+assert found, "s_endpgm not found"
+
+patched_prg = AMDProgram(dev, "patched", bytes(lib))
+b.uop.buffer.allocate()
+patched_prg(b.uop.buffer._buf, a.uop.buffer._buf, global_size=(1,1,1), local_size=(1,1,1))
+dev.synchronize()
+'''
+
+    env = os.environ.copy()
+    env["AMD"] = "1"
+    env["MOCKGPU"] = "1"
+    env["PYTHON_REMU"] = "1"
+    env["HCQDEV_WAIT_TIMEOUT_MS"] = "10000"
+
+    st = time.perf_counter()
+    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]}")
+    # 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")
+
+if __name__ == "__main__":
+  unittest.main()

extra/assembly/amd/test/test_pdf.py

@@ -0,0 +1,48 @@
+#!/usr/bin/env python3
+"""Test PDF pseudocode extraction from generate.py."""
+import unittest
+from extra.assembly.amd.generate import extract_pdf_text, extract_pcode, parse_xml, ARCHS, FIXES
+
+EXPECTED_PAGES = {"rdna3": 655, "rdna4": 711, "cdna": 610}
+
+class TestPcodePDF(unittest.TestCase):
+  @classmethod
+  def setUpClass(cls):
+    cls.pages = {arch: extract_pdf_text(cfg["pdf"]) for arch, cfg in ARCHS.items()}
+    cls.enums = {}
+    for arch, cfg in ARCHS.items():
+      _, enums, _, _, _, _ = parse_xml(cfg["xml"])
+      for fmt, ops in FIXES.get(arch, {}).items(): enums.setdefault(fmt, {}).update(ops)
+      cls.enums[arch] = enums
+    cls.pcode = {arch: extract_pcode(cls.pages[arch], {n: op for ops in cls.enums[arch].values() for op, n in ops.items()}) for arch in ARCHS}
+
+  def test_page_counts(self):
+    for name, exp in EXPECTED_PAGES.items():
+      self.assertEqual(len(self.pages[name]), exp, f"{name} page count")
+
+  def test_pcode_extracted(self):
+    """Check we extracted a reasonable number of pcode entries."""
+    for name in ARCHS:
+      self.assertGreater(len(self.pcode[name]), 500, f"{name} pcode count too low")
+
+  def test_pcode_rdna3_tricky(self):
+    """Test specific pseudocode patterns that are tricky to extract correctly."""
+    pcode = self.pcode['rdna3']
+    # BUFFER_ATOMIC_MAX_U64: should have 4 statements (not truncated)
+    self.assertEqual(pcode[('BUFFER_ATOMIC_MAX_U64', 72)],
+      'tmp = MEM[ADDR].u64;\nsrc = DATA.u64;\nMEM[ADDR].u64 = src >= tmp ? src : tmp;\nRETURN_DATA.u64 = tmp')
+    # GLOBAL_STORE_B128: should have 4 MEM stores (not truncated)
+    self.assertEqual(pcode[('GLOBAL_STORE_B128', 29)],
+      'MEM[ADDR].b32 = VDATA[31 : 0];\nMEM[ADDR + 4U].b32 = VDATA[63 : 32];\nMEM[ADDR + 8U].b32 = VDATA[95 : 64];\nMEM[ADDR + 12U].b32 = VDATA[127 : 96]')
+    # S_CMOVK_I32: should have full if/endif block
+    self.assertEqual(pcode[('S_CMOVK_I32', 2)],
+      "if SCC then\nD0.i32 = 32'I(signext(SIMM16.i16))\nendif")
+
+  def test_pcode_no_examples(self):
+    """Pseudocode should not contain example lines with '=>'."""
+    for name in ARCHS:
+      for (op_name, opcode), code in self.pcode[name].items():
+        self.assertNotIn('=>', code, f"{name} {op_name} contains example line with '=>'")
+
+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[2:3], saddr=OFF),
+      v_add_co_ci_u32_e32(v[1], s[5], v[1]),
+      s_waitcnt(0x03f7),  # vmcnt(0)
+      global_store_b32(addr=v[0:1], 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_rdna4_emu.py

@@ -0,0 +1,98 @@
+import unittest, ctypes
+from extra.assembly.amd.autogen.rdna4 import ins as ir4
+from extra.assembly.amd.dsl import v, s
+from extra.assembly.amd.emu import WaveState, decode_program
+from tinygrad.device import Buffer, BufferSpec
+from tinygrad.dtype import dtypes
+
+class TestRDNA4Emu(unittest.TestCase):
+  def _run(self, insts: list, sgprs: dict[int, int] = None, vgprs: dict[tuple[int, int], int] = None) -> WaveState:
+    """Run instructions and return final WaveState."""
+    # Add S_ENDPGM if not present
+    if not any(isinstance(i, ir4.SOPP) and i.op == ir4.SOPPOp.S_ENDPGM for i in insts):
+      insts = list(insts) + [ir4.SOPP(ir4.SOPPOp.S_ENDPGM, simm=0)]
+
+    # Assemble and decode
+    code = b''.join(i.to_bytes() for i in insts)
+    code_buf = (ctypes.c_uint8 * len(code)).from_buffer_copy(code)
+    code_addr = ctypes.addressof(code_buf)
+    program_raw = decode_program(code, "rdna4")
+    program = {code_addr + offset: val for offset, val in program_raw.items()}
+
+    # Setup wave state
+    st = WaveState(n_lanes=1)
+    st.pc = code_addr
+    if sgprs:
+      for idx, val in sgprs.items(): st._write_sgpr(idx, val)
+    if vgprs:
+      for (reg, lane), val in vgprs.items(): st._write_vgpr(reg, lane, val)
+
+    # Setup vmem buffer with external_ptr=0 (maps to address 0, allows any pointer access)
+    vmem_buf = Buffer('CPU', 1 << 40, dtypes.uint32, options=BufferSpec(external_ptr=0)).ensure_allocated()
+
+    # Execute
+    c_bufs = [ctypes.c_uint64(st.sgpr_buf._buf.va_addr), ctypes.c_uint64(st.vgpr_buf._buf.va_addr),
+              ctypes.c_uint64(vmem_buf._buf.va_addr), ctypes.c_uint64(0), ctypes.c_uint64(0)]
+    for _ in range(100):
+      if (pc := st.pc) == 0xFFFFFFFFFFFFFFFF or pc not in program: break
+      _, fxn, globals_list, _ = program[pc]
+      fxn(*[c_bufs[g] for g in globals_list])
+    return st
+
+  def test_vopd_dual_mov(self):
+    """Test VOPD with two V_DUAL_MOV_B32 operations: v[1]=s[1], v[2]=s[2]."""
+    insts = [ir4.VOPD(ir4.VOPDOp.V_DUAL_MOV_B32, ir4.VOPDOp.V_DUAL_MOV_B32,
+                      vdstx=v[1], vdsty=v[2], srcx0=s[1], srcy0=s[2], vsrcx1=v[0], vsrcy1=v[0])]
+    st = self._run(insts, sgprs={1: 0x40e00000, 2: 0x41100000})  # 7.0f, 9.0f
+    self.assertEqual(st._read_vgpr(1, 0), 0x40e00000)  # v[1] = 7.0
+    self.assertEqual(st._read_vgpr(2, 0), 0x41100000)  # v[2] = 9.0
+
+  def test_vopd_dual_mov_after_other_vopd(self):
+    """Test VOPD reuse: first VOPD(v[3]=0, v[0]=?), then VOPD(v[1]=s[1], v[2]=s[2])."""
+    # This matches the BEAM kernel sequence that fails
+    insts = [
+      ir4.VOPD(ir4.VOPDOp.V_DUAL_MOV_B32, ir4.VOPDOp.V_DUAL_MOV_B32,
+               vdstx=v[3], vdsty=v[0], srcx0=0, srcy0=s[0], vsrcx1=v[0], vsrcy1=v[0]),  # v[3]=0, v[0]=s[0]
+      ir4.VOPD(ir4.VOPDOp.V_DUAL_MOV_B32, ir4.VOPDOp.V_DUAL_MOV_B32,
+               vdstx=v[1], vdsty=v[2], srcx0=s[1], srcy0=s[2], vsrcx1=v[0], vsrcy1=v[0]),  # v[1]=s[1], v[2]=s[2]
+    ]
+    st = self._run(insts, sgprs={0: 0x40a00000, 1: 0x40e00000, 2: 0x41100000})  # 5.0f, 7.0f, 9.0f
+    self.assertEqual(st._read_vgpr(1, 0), 0x40e00000)  # v[1] = 7.0
+    self.assertEqual(st._read_vgpr(2, 0), 0x41100000)  # v[2] = 9.0
+
+  def test_vopd_with_s_add_f32_sequence(self):
+    """Test full BEAM kernel sequence: s_add_f32 then VOPD."""
+    # This is the exact sequence from the failing BEAM kernel
+    insts = [
+      ir4.SOP2(ir4.SOP2Op.S_ADD_F32, sdst=s[0], ssrc0=s[0], ssrc1=s[8]),   # s[0] = s[0] + s[8]
+      ir4.SOP2(ir4.SOP2Op.S_ADD_F32, sdst=s[1], ssrc0=s[1], ssrc1=s[9]),   # s[1] = s[1] + s[9]
+      ir4.SOP2(ir4.SOP2Op.S_ADD_F32, sdst=s[2], ssrc0=s[2], ssrc1=s[10]),  # s[2] = s[2] + s[10]
+      ir4.VOPD(ir4.VOPDOp.V_DUAL_MOV_B32, ir4.VOPDOp.V_DUAL_MOV_B32,
+               vdstx=v[3], vdsty=v[0], srcx0=0, srcy0=s[0], vsrcx1=v[0], vsrcy1=v[0]),
+      ir4.VOPD(ir4.VOPDOp.V_DUAL_MOV_B32, ir4.VOPDOp.V_DUAL_MOV_B32,
+               vdstx=v[1], vdsty=v[2], srcx0=s[1], srcy0=s[2], vsrcx1=v[0], vsrcy1=v[0]),
+    ]
+    # Input: s[0:2] = [1,2,3], s[8:10] = [4,5,6]
+    # After s_add_f32: s[0:2] = [5,7,9]
+    st = self._run(insts, sgprs={0: 0x3f800000, 1: 0x40000000, 2: 0x40400000,  # 1.0, 2.0, 3.0
+                                  8: 0x40800000, 9: 0x40a00000, 10: 0x40c00000})  # 4.0, 5.0, 6.0
+    self.assertEqual(st._read_vgpr(1, 0), 0x40e00000)  # v[1] = 7.0
+    self.assertEqual(st._read_vgpr(2, 0), 0x41100000)  # v[2] = 9.0
+
+  def test_s_mov_b32_then_vopd(self):
+    """Test s_mov_b32 followed by VOPD - simulates BEAM kernel sequence."""
+    # Use s_mov_b32 with SGPR source (copy from pre-initialized SGPRs)
+    # s[10:12] will have values set by test harness, copy to s[0:2], then VOPD to VGPRs
+    insts = [
+      ir4.SOP1(ir4.SOP1Op.S_MOV_B32, sdst=s[0], ssrc0=s[10]),  # s[0] = s[10]
+      ir4.SOP1(ir4.SOP1Op.S_MOV_B32, sdst=s[1], ssrc0=s[11]),  # s[1] = s[11]
+      ir4.SOP1(ir4.SOP1Op.S_MOV_B32, sdst=s[2], ssrc0=s[12]),  # s[2] = s[12]
+      ir4.VOPD(ir4.VOPDOp.V_DUAL_MOV_B32, ir4.VOPDOp.V_DUAL_MOV_B32,
+               vdstx=v[1], vdsty=v[2], srcx0=s[1], srcy0=s[2], vsrcx1=v[0], vsrcy1=v[0]),
+    ]
+    st = self._run(insts, sgprs={10: 0x40a00000, 11: 0x40e00000, 12: 0x41100000})  # 5.0, 7.0, 9.0
+    self.assertEqual(st._read_vgpr(1, 0), 0x40e00000)  # v[1] = 7.0
+    self.assertEqual(st._read_vgpr(2, 0), 0x41100000)  # v[2] = 9.0
+
+if __name__ == '__main__':
+  unittest.main()

extra/assembly/amd/test/test_roundtrip.py

@@ -0,0 +1,247 @@
+#!/usr/bin/env python3
+"""Roundtrip tests: generate tinygrad kernels, decode instructions, re-encode, verify match."""
+import unittest, io, sys, re, subprocess, os
+from extra.assembly.amd.dsl import Inst
+from extra.assembly.amd import decode_inst, detect_format
+from extra.assembly.amd.test.helpers import get_llvm_mc, get_llvm_objdump, get_target, get_mattr
+
+def disassemble_lib(lib: bytes, compiler) -> list[tuple[str, bytes]]:
+  """Disassemble ELF binary and return list of (instruction_text, machine_code_bytes)."""
+  old_stdout = sys.stdout
+  sys.stdout = io.StringIO()
+  compiler.disassemble(lib)
+  output = sys.stdout.getvalue()
+  sys.stdout = old_stdout
+
+  results = []
+  for line in output.splitlines():
+    if '//' not in line: continue
+    instr = line.split('//')[0].strip()
+    if not instr: continue
+    comment = line.split('//')[1].strip()
+    if ':' not in comment: continue
+    hex_str = comment.split(':')[1].strip().split()[0]
+    try:
+      machine_bytes = bytes.fromhex(hex_str)[::-1]  # big-endian to little-endian
+      results.append((instr, machine_bytes))
+    except ValueError:
+      continue
+  return results
+
+def compile_asm(instr: str, arch: str = 'rdna3') -> bytes:
+  """Compile a single instruction using LLVM."""
+  return compile_asm_batch([instr], arch)[0]
+
+def compile_asm_batch(instrs: list[str], arch: str = 'rdna3') -> list[bytes]:
+  """Compile multiple instructions with a single llvm-mc call."""
+  if not instrs: return []
+  result = subprocess.run([get_llvm_mc(), '-triple=amdgcn', f'-mcpu={get_target(arch)}', f'-mattr={get_mattr(arch)}', '-show-encoding'],
+                          input=".text\n" + "\n".join(instrs) + "\n", capture_output=True, text=True)
+  if result.returncode != 0: raise RuntimeError(f"llvm-mc batch failed: {result.stderr.strip()}")
+  encodings = []
+  for line in result.stdout.split('\n'):
+    if 'encoding:' in line:
+      enc = line.split('encoding:')[1].strip()
+      if enc.startswith('[') and enc.endswith(']'):
+        encodings.append(bytes.fromhex(enc[1:-1].replace('0x', '').replace(',', '').replace(' ', '')))
+  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], arch: str = 'rdna3') -> list[str]:
+  """Compile instructions with LLVM and get LLVM's disassembly."""
+  import tempfile
+  if not instrs: return []
+  mcpu, mattr = get_target(arch), get_mattr(arch)
+  src = ".text\n.globl test\n.p2align 8\n.type test,@function\ntest:\n" + "\n".join(f"  {instr}" for instr in instrs) + "\n"
+  with tempfile.NamedTemporaryFile(suffix='.o', delete=False) as f:
+    obj_path = f.name
+  try:
+    result = subprocess.run([get_llvm_mc(), '-triple=amdgcn', f'-mcpu={mcpu}', f'-mattr={mattr}', '-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()}")
+    result = subprocess.run([get_llvm_objdump(), '-d', f'--mcpu={mcpu}', obj_path], capture_output=True, text=True)
+    if result.returncode != 0: raise RuntimeError(f"llvm-objdump failed: {result.stderr.strip()}")
+    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)
+
+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)
+    """
+    arch = self.arch
+
+    from extra.assembly.amd.test.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
+
+    kernels, _, _ = get_kernels_from_tinygrad(op_fn)
+    # rendered source can be C or llvmir
+    compiler = (AMDLLVMCompiler if AMD_LLVM else HIPCompiler)(get_target(arch))
+
+    # 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)
+        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
+
+        try:
+          decoded = fmt.from_bytes(remaining)  # pass all remaining bytes so from_bytes can read literal
+          size = decoded.size()  # actual size including literal
+          orig_bytes = remaining[:size]
+          reencoded = decoded.to_bytes()
+          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))
+        except Exception as e:
+          decoded_instrs.append((ki, offset, remaining[:base_size], None, None, False, str(e)))
+          size = base_size
+
+        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
+    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
+      if our_disasm.startswith('op_') or re.search(r', \d+, \d+, \d+,', our_disasm): continue
+      asm_test_instrs.append((idx, our_disasm, orig_bytes))
+      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+disasm for disasm comparison test
+    disasm_llvm_results = compile_and_disasm_batch([d for _, d in disasm_test_instrs], arch)
+    disasm_llvm_map = {idx: result for (idx, _), result in zip(disasm_test_instrs, disasm_llvm_results)}
+
+    # Now evaluate results
+    decode_passed, decode_failed, decode_skipped = 0, 0, 0
+    asm_passed, asm_failed, asm_skipped = 0, 0, 0
+    disasm_passed, disasm_failed, disasm_skipped = 0, 0, 0
+    decode_failures: list[str] = []
+    asm_failures: list[str] = []
+    disasm_failures: list[str] = []
+
+    for idx, (ki, offset, orig_bytes, decoded, our_disasm, decode_ok, decode_err) in enumerate(decoded_instrs):
+      # Decode test
+      if decode_ok:
+        decode_passed += 1
+      elif decode_err == "no format":
+        decode_skipped += 1
+      else:
+        decode_failed += 1
+        decode_failures.append(f"K{ki}@{offset}: {our_disasm}: {decode_err}")
+
+      # Asm test: our disasm -> LLVM asm -> compare bytes with original
+      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()}")
+      else:
+        asm_skipped += 1
+
+      # Disasm comparison test
+      if our_disasm is None:
+        disasm_skipped += 1
+      elif idx in disasm_llvm_map:
+        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}'")
+      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, 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
+  def test_neg(self): self._test_kernel_roundtrip(lambda T: -T([1.0, -2.0, 3.0, -4.0]))
+  def test_relu(self): self._test_kernel_roundtrip(lambda T: T([-1.0, 0.0, 1.0, 2.0]).relu())
+  def test_exp(self): self._test_kernel_roundtrip(lambda T: T([0.0, 1.0, 2.0]).exp())
+  def test_log(self): self._test_kernel_roundtrip(lambda T: T([1.0, 2.0, 3.0]).log())
+  def test_sin(self): self._test_kernel_roundtrip(lambda T: T([0.0, 1.0, 2.0]).sin())
+  def test_sqrt(self): self._test_kernel_roundtrip(lambda T: T([1.0, 4.0, 9.0]).sqrt())
+  def test_recip(self): self._test_kernel_roundtrip(lambda T: T([1.0, 2.0, 4.0]).reciprocal())
+
+  # Binary ops
+  def test_add(self): self._test_kernel_roundtrip(lambda T: T([1.0, 2.0]) + T([3.0, 4.0]))
+  def test_sub(self): self._test_kernel_roundtrip(lambda T: T([5.0, 6.0]) - T([1.0, 2.0]))
+  def test_mul(self): self._test_kernel_roundtrip(lambda T: T([2.0, 3.0]) * T([4.0, 5.0]))
+  def test_div(self): self._test_kernel_roundtrip(lambda T: T([10.0, 20.0]) / T([2.0, 4.0]))
+  def test_max_binary(self): self._test_kernel_roundtrip(lambda T: T([1.0, 5.0]).maximum(T([3.0, 2.0])))
+
+  # Reductions
+  def test_sum_reduce(self): self._test_kernel_roundtrip(lambda T: T.empty(64).sum())
+  def test_max_reduce(self): self._test_kernel_roundtrip(lambda T: T.empty(64).max())
+  def test_mean_reduce(self): self._test_kernel_roundtrip(lambda T: T.empty(32).mean())
+
+  # Matmul
+  def test_gemm_4x4(self): self._test_kernel_roundtrip(lambda T: T.empty(4, 4) @ T.empty(4, 4))
+  def test_gemv(self): self._test_kernel_roundtrip(lambda T: T.empty(1, 16) @ T.empty(16, 16))
+
+  # Complex ops
+  def test_softmax(self): self._test_kernel_roundtrip(lambda T: T.empty(16).softmax())
+  def test_layernorm(self): self._test_kernel_roundtrip(lambda T: T.empty(8, 8).layernorm())
+
+  # Memory patterns
+  def test_contiguous(self): self._test_kernel_roundtrip(lambda T: T.empty(4, 4).permute(1, 0).contiguous())
+  def test_reshape(self): self._test_kernel_roundtrip(lambda T: (T.empty(16) + 1).reshape(4, 4).contiguous())
+  def test_expand(self): self._test_kernel_roundtrip(lambda T: T.empty(4, 1).expand(4, 4).contiguous())
+
+  # Cast ops
+  def test_cast_int(self): self._test_kernel_roundtrip(lambda T: T.empty(16).int().float())
+  def test_cast_half(self): self._test_kernel_roundtrip(lambda T: T.empty(16).half().float())
+
+  # Comparison ops
+  def test_cmp_lt(self): self._test_kernel_roundtrip(lambda T: (T.empty(64) < T.empty(64)).where(T.empty(64), T.empty(64)))
+  def test_where(self): self._test_kernel_roundtrip(lambda T: (T.empty(64) > 0).where(T.empty(64), T.empty(64)))
+
+  # 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()