Merge branch 'master' into kernel_is_call

fa: block skipping for fa kv bwd (#14569 )
more cleanup in create_schedule [pr] (#14566 )
2026-08-15 20:58:28 +00:00 · 2026-02-06 09:15:03 +08:00 · 2026-02-05 16:13:53 -08:00 · 2026-02-05 16:12:17 -05:00 · 2026-02-05 16:11:32 -05:00 · 2026-02-05 15:24:09 -05:00
246 changed files with 19328 additions and 7040 deletions
@@ -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
@@ -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:
@@ -145,6 +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
@@ -332,9 +378,9 @@ jobs:
    - 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: Insert amdgpu
    #  run: sudo modprobe amdgpu
    - name: Symlink models and datasets
@@ -444,9 +490,9 @@ jobs:
    - 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
@@ -496,9 +542,9 @@ jobs:
    - 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
@@ -587,9 +633,9 @@ jobs:
    - 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
@@ -651,9 +697,9 @@ jobs:
    - 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
@@ -26,12 +26,12 @@ 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
@@ -98,7 +98,7 @@ 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
@@ -134,7 +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: |
@@ -156,7 +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
@@ -207,11 +207,11 @@ 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
@@ -239,6 +239,7 @@ jobs:
    - name: Run mypy with lineprecision report
      run: |
        python -m mypy --lineprecision-report .
+        grep -v autogen lineprecision.txt | awk 'NR>2 {lines+=$2; precise+=$3; imprecise+=$4; any+=$5; empty+=$6} END {t=lines-empty; printf "TOTAL: %d lines, %d precise (%.1f%%), %d imprecise (%.1f%%), %d any (%.1f%%)\n", t, precise, 100*precise/t, imprecise, 100*imprecise/t, any, 100*any/t}'
        cat lineprecision.txt
    - name: Run TYPED=1
      run: CHECK_OOB=0 DEV=CPU TYPED=1 python test/test_tiny.py
@@ -255,26 +256,29 @@ jobs:
      uses: ./.github/actions/setup-tinygrad
      with:
        key: unittest-13
-        pydeps: "pillow numpy ftfy regex pre-commit"
+        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
@@ -312,7 +316,7 @@ jobs:
        deps: testing_unit
        python-version: '3.14'
    - name: Test SPEC=2
-      run: SPEC=2 pytest --maxfail=10 -n auto --durations=30 --ignore=test/models --ignore test/test_custom_kernel.py --ignore test/unit/test_hashing.py --ignore test/unit/test_autogen.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
@@ -346,7 +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: |
@@ -422,7 +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
@@ -450,7 +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
@@ -463,11 +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

@@ -524,7 +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
@@ -543,7 +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
@@ -564,8 +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
@@ -598,8 +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: |
@@ -607,7 +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 --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

@@ -632,7 +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
@@ -674,9 +678,9 @@ jobs:
        uses: ./.github/actions/setup-tinygrad
        with:
          key: rdna3-emu
-          deps: testing_minimal
+          deps: testing_unit
          amd: 'true'
-          python-version: '3.13'
+          python-version: '3.14'
      - name: Verify AMD autogen is up to date
        run: |
          python -m extra.assembly.amd.generate
@@ -702,6 +706,8 @@ jobs:
      # 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:
@@ -722,7 +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
@@ -733,7 +739,7 @@ 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
@@ -755,7 +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: ${{ matrix.backend == 'lvp' && 'true' }}
@@ -766,7 +772,7 @@ jobs:
          python3 -c "from tinygrad import Device; assert Device.DEFAULT in ['CPU','CL'], Device.DEFAULT"
          DEBUG=5 FORWARD_ONLY=1 python3 test/test_ops.py TestOps.test_add
      - name: Run pytest (${{ 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
@@ -786,13 +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)
@@ -884,8 +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: ${{ matrix.backend == 'lvp' && 'true' }}
      - name: Set env
@@ -895,7 +902,7 @@ jobs:
          python -c "from tinygrad import Device; assert Device.DEFAULT == {'LLVM':'CPU','LVP':'CPU'}.get(x:='${{ matrix.backend }}'.upper(), x), Device.DEFAULT"
          DEBUG=4 python3 test/test_tiny.py TestTiny.test_plus
      - name: Run pytest (${{ matrix.backend }})
-        run: python3 -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit --durations=20
+        run: python3 -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit --ignore=test/null --durations=20
      - name: Run process replay tests
        uses: ./.github/actions/process-replay
      - name: Run macOS-specific unit test
@@ -928,7 +935,11 @@ jobs:
      - name: Run unit tests
        if: matrix.backend=='llvm'
        # test_newton_schulz hits RecursionError
-        run: python -m pytest -n=auto test/unit/ --ignore=test/unit/test_disk_tensor.py --ignore=test/unit/test_elf.py --ignore=test/unit/test_tar.py --ignore=test/unit/test_linalg.py --durations=20
+        run: python -m pytest -n=auto test/unit/ --ignore=test/unit/test_disk_tensor.py --ignore=test/unit/test_tar.py --ignore=test/unit/test_linalg.py --durations=20
+      - name: Run NULL backend tests
+        if: matrix.backend=='llvm'
+        shell: bash
+        run: CPU=0 CPU_LLVM=0 NULL=1 python -m pytest -n=auto test/null/ --ignore=test/null/test_elf.py --durations=20
      - name: Run pytest (${{ matrix.backend }})
        shell: bash
        run: |
@@ -952,12 +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: |
@@ -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/unit/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
@@ -72,7 +72,7 @@ vliw_prepare = PatternMatcher([
  # cast is fake
  (UPat(Ops.CAST, name="c"), lambda c: c.src[0]),
  # rewrites to hardcode the addresses in memory
-  (UPat(Ops.DEFINE_GLOBAL, name="dg"), lambda dg: UOp.const(dtypes.uint, global_addrs[dg.arg])),
+  (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
@@ -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!
@@ -264,8 +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)
@@ -379,12 +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)
@@ -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

@@ -1292,7 +1292,6 @@ def train_llama3():
  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)
@@ -1322,6 +1321,8 @@ def train_llama3():
  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:
@@ -1370,6 +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"):
@@ -1382,9 +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)
@@ -1394,27 +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)).cast(p.grad.dtype)
+      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)
@@ -1456,35 +1474,54 @@ def train_llama3():
  while i < MAX_STEPS:
    GlobalCounters.reset()
    if getenv("TRAIN", 1):
+      profile_marker(f"train @ {i}")
      st = time.perf_counter()
-      try: tokens = next(train_iter)
-      except StopIteration: break
-      dt = time.perf_counter()
-      loss, lr = train_step(model, tokens)
+
+      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
+
+      gt = time.perf_counter()
+      lr = optim_step()
+      ot = time.perf_counter()
+
      loss = loss.float().item()
      lr = lr.item()

      et = time.perf_counter()
      step_time = et - st
-      dev_time = et - dt
-      data_time = dt - 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)

      i += 1
-      sequences_seen += tokens.shape[0]
+      sequences_seen += GBS

      mem_gb = GlobalCounters.mem_used / 1e9
      gflops = GlobalCounters.global_ops / 1e9 / dev_time
-      mfu = ((6 * num_params * SEQLEN * BS) / (dev_time * max(getenv("DP", 1), getenv("MP", 1)) * 2.3e15)) * 100
+      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 run, {dev_time:.3f} s device, {data_time:.3f} s data, {loss:.4f} loss, {lr:.12f} LR, {mem_gb:.2f} GB used, {gflops:9.2f} GFLOPS, {mfu:5.2f}% MFU")
+          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
@@ -1508,7 +1545,9 @@ def train_llama3():
              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 = []
@@ -1516,8 +1555,8 @@ def train_llama3():
      tqdm.write(f"evaluating {5760//EVAL_BS} batches of {EVAL_BS} sequences")

      for j,tokens in tqdm(enumerate(eval_iter), total=EVAL_SAMPLES//EVAL_BS):
-        eval_losses += eval_step(model, tokens).tolist()
-        
+        eval_losses += eval_step(tokens).tolist()
+
        if BENCHMARK and (j+1) == min(BENCHMARK, EVAL_SAMPLES//EVAL_BS):
          return

@@ -1606,7 +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():
@@ -1645,7 +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,
@@ -8,9 +8,11 @@ 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=8 EVAL_BS=8 GRADIENT_ACC_STEPS=1
+export DP=8 BS=16 EVAL_BS=8 GRADIENT_ACC_STEPS=1
 export GBS=$((BS * GRADIENT_ACC_STEPS))

 export MODEL="llama3"
@@ -18,7 +20,7 @@ export BASEDIR="/raid/datasets/c4-8b/"
 export SMALL=1
 export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
 export EVAL_TARGET=3.3 EVAL_FREQ=12288
-export LR="4e-4" END_LR="4e-5" WARMUP_SAMPLES=256 MAX_STEPS=1200000
+export LR="2.5e-4" END_LR="2.5e-5" WARMUP_SAMPLES=256 MAX_STEPS=1200000
 export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
 export SAMPLES=$((MAX_STEPS * GBS))

@@ -2,15 +2,18 @@

 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=8 BS=8 EVAL_BS=8 GRADIENT_ACC_STEPS=1
+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"
@@ -18,13 +21,13 @@ export BASEDIR="/raid/datasets/c4-8b/"
 export SMALL=1
 export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
 export EVAL_TARGET=3.3 EVAL_FREQ=12288
-export LR="4e-4" END_LR="4e-5" WARMUP_SAMPLES=256 MAX_STEPS=1200000
+export LR="2.5e-4" END_LR="2.5e-5" WARMUP_SAMPLES=256 MAX_STEPS=1200000
 export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
 export SAMPLES=$((MAX_STEPS * GBS))

-export SEED=5760
+export SEED=${SEED:-5760}

-export JITBEAM=3
+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
@@ -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
@@ -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)
@@ -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
@@ -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)
@@ -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)
@@ -49,10 +49,11 @@ from tinygrad.helpers import Context, DEBUG, colored
 from tinygrad.engine.realize import get_runner

 from extra.assembly.amd import decode_inst
-from extra.assembly.amd.autogen.rdna3.str_pcode import PCODE
-from extra.assembly.amd.autogen.rdna3.ins import (SOP1, SOP2, SOPC, SOPK, SOPP, SMEM, VOP1, VOP1_SDST, VOP2, VOP3, VOP3_SDST, VOP3SD, VOP3P, VOPC,
-  DS, FLAT, GLOBAL, SCRATCH, VOPD, SOPPOp, SMEMOp, VOP1Op, VOP2Op, VOP3Op, VOPDOp)
-from extra.assembly.amd.dsl import VCC_LO, EXEC_LO, SCC
+from extra.assembly.amd.autogen.rdna3.str_pcode import PCODE as PCODE_RDNA3
+from extra.assembly.amd.autogen.rdna4.str_pcode import PCODE as PCODE_RDNA4
+from extra.assembly.amd.autogen.rdna3 import ins as ir3
+from extra.assembly.amd.autogen.rdna4 import ins as ir4
+from extra.assembly.amd.dsl import VCC_LO, EXEC_LO, SCC, ttmp
 from extra.assembly.amd.autogen.common import Fmt, OpType
 from extra.assembly.amd.pcode import parse_block, _FUNCS

@@ -79,15 +80,23 @@ def _apply_src_mods(val: UOp, mod_bit: int, abs_bits: int, neg_bits: int, bits:
  if neg_bits & (1 << mod_bit): fv = fv.neg()
  return fv.bitcast(ut).cast(dtypes.uint32) if bits == 16 else fv.bitcast(ut)

-# Map VOPD ops to VOP2 ops for pcode lookup
+# Map VOPD ops to VOP2 ops for pcode lookup (both RDNA3 and RDNA4)
 VOPD_TO_VOP2 = {
-  VOPDOp.V_DUAL_FMAC_F32: VOP2Op.V_FMAC_F32_E32, VOPDOp.V_DUAL_MUL_F32: VOP2Op.V_MUL_F32_E32,
-  VOPDOp.V_DUAL_ADD_F32: VOP2Op.V_ADD_F32_E32, VOPDOp.V_DUAL_SUB_F32: VOP2Op.V_SUB_F32_E32,
-  VOPDOp.V_DUAL_SUBREV_F32: VOP2Op.V_SUBREV_F32_E32, VOPDOp.V_DUAL_MAX_F32: VOP2Op.V_MAX_F32_E32,
-  VOPDOp.V_DUAL_MIN_F32: VOP2Op.V_MIN_F32_E32, VOPDOp.V_DUAL_ADD_NC_U32: VOP2Op.V_ADD_NC_U32_E32,
-  VOPDOp.V_DUAL_LSHLREV_B32: VOP2Op.V_LSHLREV_B32_E32, VOPDOp.V_DUAL_AND_B32: VOP2Op.V_AND_B32_E32,
-  VOPDOp.V_DUAL_MOV_B32: VOP1Op.V_MOV_B32_E32, VOPDOp.V_DUAL_CNDMASK_B32: VOP2Op.V_CNDMASK_B32_E32,
-  VOPDOp.V_DUAL_FMAAK_F32: VOP2Op.V_FMAAK_F32_E32, VOPDOp.V_DUAL_FMAMK_F32: VOP2Op.V_FMAMK_F32_E32,
+  ir3.VOPDOp.V_DUAL_FMAC_F32: ir3.VOP2Op.V_FMAC_F32_E32, ir3.VOPDOp.V_DUAL_MUL_F32: ir3.VOP2Op.V_MUL_F32_E32,
+  ir3.VOPDOp.V_DUAL_ADD_F32: ir3.VOP2Op.V_ADD_F32_E32, ir3.VOPDOp.V_DUAL_SUB_F32: ir3.VOP2Op.V_SUB_F32_E32,
+  ir3.VOPDOp.V_DUAL_SUBREV_F32: ir3.VOP2Op.V_SUBREV_F32_E32, ir3.VOPDOp.V_DUAL_MAX_F32: ir3.VOP2Op.V_MAX_F32_E32,
+  ir3.VOPDOp.V_DUAL_MIN_F32: ir3.VOP2Op.V_MIN_F32_E32, ir3.VOPDOp.V_DUAL_ADD_NC_U32: ir3.VOP2Op.V_ADD_NC_U32_E32,
+  ir3.VOPDOp.V_DUAL_LSHLREV_B32: ir3.VOP2Op.V_LSHLREV_B32_E32, ir3.VOPDOp.V_DUAL_AND_B32: ir3.VOP2Op.V_AND_B32_E32,
+  ir3.VOPDOp.V_DUAL_MOV_B32: ir3.VOP1Op.V_MOV_B32_E32, ir3.VOPDOp.V_DUAL_CNDMASK_B32: ir3.VOP2Op.V_CNDMASK_B32_E32,
+  ir3.VOPDOp.V_DUAL_FMAAK_F32: ir3.VOP2Op.V_FMAAK_F32_E32, ir3.VOPDOp.V_DUAL_FMAMK_F32: ir3.VOP2Op.V_FMAMK_F32_E32,
+  # RDNA4 mappings (same VOP1/VOP2 targets, RDNA4 uses _NUM_ suffix for min/max)
+  ir4.VOPDOp.V_DUAL_FMAC_F32: ir3.VOP2Op.V_FMAC_F32_E32, ir4.VOPDOp.V_DUAL_MUL_F32: ir3.VOP2Op.V_MUL_F32_E32,
+  ir4.VOPDOp.V_DUAL_ADD_F32: ir3.VOP2Op.V_ADD_F32_E32, ir4.VOPDOp.V_DUAL_SUB_F32: ir3.VOP2Op.V_SUB_F32_E32,
+  ir4.VOPDOp.V_DUAL_SUBREV_F32: ir3.VOP2Op.V_SUBREV_F32_E32, ir4.VOPDOp.V_DUAL_MAX_NUM_F32: ir3.VOP2Op.V_MAX_F32_E32,
+  ir4.VOPDOp.V_DUAL_MIN_NUM_F32: ir3.VOP2Op.V_MIN_F32_E32, ir4.VOPDOp.V_DUAL_ADD_NC_U32: ir3.VOP2Op.V_ADD_NC_U32_E32,
+  ir4.VOPDOp.V_DUAL_LSHLREV_B32: ir3.VOP2Op.V_LSHLREV_B32_E32, ir4.VOPDOp.V_DUAL_AND_B32: ir3.VOP2Op.V_AND_B32_E32,
+  ir4.VOPDOp.V_DUAL_MOV_B32: ir3.VOP1Op.V_MOV_B32_E32, ir4.VOPDOp.V_DUAL_CNDMASK_B32: ir3.VOP2Op.V_CNDMASK_B32_E32,
+  ir4.VOPDOp.V_DUAL_FMAAK_F32: ir3.VOP2Op.V_FMAAK_F32_E32, ir4.VOPDOp.V_DUAL_FMAMK_F32: ir3.VOP2Op.V_FMAMK_F32_E32,
 }
 WAVE_SIZE = 32
 # Special registers stored after inline constants (256-259)
@@ -126,14 +135,6 @@ def _val_to_u32(val: UOp) -> UOp:
  if val.dtype in (dtypes.uint16, dtypes.int16): return val.cast(dtypes.uint32)
  return val.cast(dtypes.uint32)

-def _apply_clamp(val: UOp, clmp: int | UOp) -> UOp:
-  """Apply VOP3 clamp modifier: clamp float results to [0.0, 1.0] range."""
-  if isinstance(clmp, int) and clmp == 0: return val
-  if val.dtype not in (dtypes.float32, dtypes.half, dtypes.float64): return val
-  zero, one = UOp.const(val.dtype, 0.0), UOp.const(val.dtype, 1.0)
-  clamped = val.maximum(zero).minimum(one)
-  return clmp.ne(_c(0)).where(clamped, val) if isinstance(clmp, UOp) else clamped
-
 _pcode_fixes = {
  'V_DIV_FMAS_F32': ('D0.f32 = 2.0F ** 32 * fma(S0.f32, S1.f32, S2.f32)',
    'D0.f32 = (exponent(S2.f32) > 127) ? (2.0F ** 64 * fma(S0.f32, S1.f32, S2.f32)) : (2.0F ** -64 * fma(S0.f32, S1.f32, S2.f32))'),
@@ -146,11 +147,15 @@ _pcode_fixes = {
  'V_TRIG_PREOP_F64': ("result = 64'F((1201'B(2.0 / PI)[1200 : 0] << shift.u32) & 1201'0x1fffffffffffff)", "result = trig_preop_result(shift)"),
 }

+def _get_pcode_dict(op) -> dict:
+  """Return the PCODE dictionary for the given opcode based on its architecture."""
+  return PCODE_RDNA4 if 'rdna4' in type(op).__module__ else PCODE_RDNA3
+
 # Pcode parser
@functools.cache
 def get_pcode(op) -> str:
  op_name = op.name
-  pcode = PCODE[op]
+  pcode = _get_pcode_dict(op)[op]
  if op_name in _pcode_fixes: pcode = pcode.replace(*_pcode_fixes[op_name])
  if 'V_DIV_SCALE' in op_name:
    dt, exp_lim, ldexp_val = ('f32', '23', '64') if 'F32' in op_name else ('f64', '52', '128')
@@ -174,7 +179,12 @@ def get_pcode(op) -> str:
 def parse_pcode(pcode: str, srcs: dict[str, UOp] | None = None) -> tuple[dict, list[tuple[str, UOp]]]:
  vars: dict = srcs.copy() if srcs else {}
  assigns: list[tuple[str, UOp]] = []
-  lines = [l.strip().rstrip(';') for l in pcode.split('\n') if l.strip() and not l.strip().startswith('//')]
+  raw_lines = [l.strip().rstrip(';') for l in pcode.split('\n') if l.strip() and not l.strip().startswith('//')]
+  # TODO: pcode.py should tokenize full pcode string instead of line-by-line, then this hack can be removed
+  lines: list[str] = []
+  for l in raw_lines:
+    if lines and lines[-1].endswith('&&'): lines[-1] = lines[-1] + ' ' + l
+    else: lines.append(l)
  _, final, _ = parse_block(lines, 0, vars, assigns=assigns)
  sliced = set(d.split('[')[0] for d, _ in assigns if '[' in d)
  for var, val in final.items():
@@ -247,11 +257,11 @@ def _collect_data_slices(assigns: list[tuple[str, UOp]], data_prefix: str, pcode
 class _Ctx:
  """Context for instruction compilation - holds buffers and helpers."""
  __slots__ = ('inst_size', 'dyn_fields', '_axis_id')
-  sgpr = UOp(Ops.DEFINE_GLOBAL, dtypes.uint32.ptr(SGPR_COUNT), arg=0)
-  vgpr = UOp(Ops.DEFINE_GLOBAL, dtypes.uint32.ptr(VGPR_SIZE), arg=1)
-  vmem = UOp(Ops.DEFINE_GLOBAL, dtypes.uint32.ptr(1 << 46), arg=2)
-  lds = UOp(Ops.DEFINE_GLOBAL, dtypes.uint32.ptr(16384), arg=3)
-  scratch = UOp(Ops.DEFINE_GLOBAL, dtypes.uint8.ptr(1 << 30), arg=4)
+  sgpr = UOp(Ops.PARAM, dtypes.uint32.ptr(SGPR_COUNT), arg=0)
+  vgpr = UOp(Ops.PARAM, dtypes.uint32.ptr(VGPR_SIZE), arg=1)
+  vmem = UOp(Ops.PARAM, dtypes.uint32.ptr(1 << 46), arg=2)
+  lds = UOp(Ops.PARAM, dtypes.uint32.ptr(16384), arg=3)
+  scratch = UOp(Ops.PARAM, dtypes.uint8.ptr(1 << 30), arg=4)

  def __init__(self, inst_size: int):
    self.inst_size, self._axis_id = inst_size, 0
@@ -317,9 +327,9 @@ class _Ctx:
    return base, mask, size

  # Dynamic register access (takes UOp index instead of int)
-  def rsgpr_dyn(self, reg: UOp) -> UOp:
+  def rsgpr_dyn(self, reg: UOp, valid: UOp | None = None) -> UOp:
    """Read SGPR with dynamic register index."""
-    return self.sgpr.index(reg.cast(dtypes.int), ptr=True).load()
+    return self.sgpr.index(reg.cast(dtypes.int), valid, ptr=True).load() if valid is not None else self.sgpr.index(reg.cast(dtypes.int), ptr=True).load()

  def wsgpr_dyn(self, reg: UOp, val: UOp) -> UOp:
    """Write SGPR with dynamic register index. Writes to NULL (124) are discarded."""
@@ -336,20 +346,23 @@ class _Ctx:
    offset = reg.cast(dtypes.int) * _c(32, dtypes.int) + lane.cast(dtypes.int)
    return buf.index(offset, _lane_active(exec_mask, lane)).store(val.cast(dtypes.uint32))

-  def rsrc_dyn(self, off: UOp, lane: UOp | None, bits: int = 32, literal: UOp | None = None, is_f64: bool = False) -> UOp:
+  def rsrc_dyn(self, off: UOp, lane: UOp | None, bits: int = 32, literal: UOp | None = None, is_f64: bool = False, do_cast: bool = True) -> UOp:
    """Read source operand with dynamic offset. Handles SGPR/inline constants (<256), VGPR (>=256).
    If lane is None, only scalar access is supported (off must be < 256).
    is_f64: True for F64 operations where 64-bit literals go in high 32 bits."""
    is_float_const = (off >= _c(240)) & (off <= _c(248))
-    sgpr_lo = self.rsgpr_dyn(off)
+    is_vgpr = off >= _c(256)
+    is_sgpr = is_vgpr.ne(True)
+    sgpr_lo = self.rsgpr_dyn(off, is_sgpr)

    if lane is not None:
-      is_vgpr, vgpr_reg = off >= _c(256), off - _c(256)
+      vgpr_reg = off - _c(256)
      vgpr_lo = self.rvgpr_dyn(vgpr_reg, lane, is_vgpr)
      vgpr_val = _u64(vgpr_lo, self.rvgpr_dyn(vgpr_reg + _c(1), lane, is_vgpr)) if bits == 64 else vgpr_lo

    if bits == 64:
-      sgpr_val = _u64(sgpr_lo, self.rsgpr_dyn(off + _c(1)))
+      sgpr_hi = self.rsgpr_dyn(off + _c(1), is_sgpr)
+      sgpr_val = _u64(sgpr_lo, sgpr_hi)
      # Integer inline constants: sign-extend 32-bit value from buffer to 64-bit
      # Float constants: cast F32 to F64
      int_inline = sgpr_lo.cast(dtypes.int32).cast(dtypes.int64)
@@ -364,7 +377,7 @@ class _Ctx:
    else:
      scalar_val = sgpr_lo
      if literal is not None: scalar_val = off.eq(_c(255)).where(literal, scalar_val)
-      if bits == 16:  # Float constants: cast F32 to F16
+      if bits == 16 and do_cast:  # Float constants: cast F32 to F16
        scalar_val = is_float_const.where(scalar_val.bitcast(dtypes.float32).cast(dtypes.half).bitcast(dtypes.uint16).cast(dtypes.uint32), scalar_val)

    return is_vgpr.where(vgpr_val, scalar_val) if lane is not None else scalar_val
@@ -402,17 +415,19 @@ class _Ctx:
    return UOp.sink(*self.scalar_stores(assigns, sdst_reg, sdst_size), *self.inc_pc())

  def compile_lane_pcode(self, op, inst) -> UOp:
-    """Compile READLANE/READFIRSTLANE/WRITELANE using pcode parser."""
+    """Compile cross-lane ops (READLANE/WRITELANE/PERMLANE) using pcode parser."""
    pcode = get_pcode(op)
    op_name = op.name if hasattr(op, 'name') else str(op)
    src0_off, vdst_off = self.inst_field(type(inst).src0), self.inst_field(type(inst).vdst)
    src0_reg = (src0_off >= _c(256)).where(src0_off - _c(256), _c(0))  # VGPR index or 0
    src1_off = self.inst_field(type(inst).src1) if hasattr(type(inst), 'src1') else None
+    src2_off = self.inst_field(type(inst).src2) if hasattr(type(inst), 'src2') else None
    exec_lo = self.rsgpr_dyn(_c(EXEC_LO.offset))
    srcs = {
      'SRC0': src0_reg, 'VDST': vdst_off, 'EXEC_LO': exec_lo, 'EXEC': exec_lo.cast(dtypes.uint64), '_vgpr': self.vgpr,
      'S0': self.rsrc_dyn(src0_off, _c(0, dtypes.int)) if 'WRITELANE' in op_name else src0_reg,
      'S1': self.rsrc_dyn(src1_off, _c(0, dtypes.int)) if src1_off is not None else _c(0),
+      'S2': self.rsrc_dyn(src2_off, _c(0, dtypes.int)) if src2_off is not None else _c(0),
    }
    _, assigns = parse_pcode(pcode, srcs)
    stores = []
@@ -422,14 +437,33 @@ class _Ctx:
    return UOp.sink(*stores, *self.inc_pc())

  def compile_vop_pcode(self, op, srcs: dict[str, UOp], lane: UOp, vdst_reg: UOp, exec_mask: UOp,
-                        opsel_dst_hi: bool | UOp = False, sdst_reg: int | None = None, clmp: int | UOp = 0) -> UOp:
+                        opsel_dst_hi: bool | UOp = False, sdst_reg: int | None = None, clmp: int = 0) -> UOp:
    """Compile VOP instruction. Returns sink with stores and inc_pc."""
    pcode = get_pcode(op)
    vcc_reg = sdst_reg if sdst_reg is not None else VCC_LO.offset
    if 'VCC' not in srcs: srcs['VCC'] = self.rsgpr_dyn(_c(vcc_reg))
-    srcs.update({'EXEC': exec_mask, 'SCC': self.rsgpr_dyn(_c(SCC.offset)), 'laneId': lane})
+    srcs.update({'EXEC': exec_mask, 'SCC': self.rsgpr_dyn(_c(SCC.offset)), 'laneId': lane,
+                 'ROUND_MODE': _c(0), 'ROUND_TOWARD_ZERO': _c(0)})  # rounding mode: 0=RNE, RTZ constant
    _, assigns = parse_pcode(pcode, srcs)

+    # For integer ops with clamp, compute overflow using wide arithmetic
+    # NOTE: MUL_LO ops don't saturate - they always return the low bits
+    int_saturate = None
+    if clmp and any(p in op.name for p in ('_NC_U', '_MAD_U', '_NC_I', '_MAD_I')):
+      is_signed, is_16bit = '_I' in op.name and '_U' not in op.name, '16' in op.name
+      if not (is_16bit and is_signed):  # Skip 16-bit signed ops due to codegen issues
+        s0, s1, s2 = srcs.get('S0'), srcs.get('S1'), srcs.get('S2')
+        if s0 is not None and s1 is not None:
+          narrow_dt = dtypes.uint16 if is_16bit else (dtypes.int32 if is_signed else dtypes.uint32)
+          wide_dt = dtypes.int32 if is_16bit else dtypes.int64
+          narrow_max, narrow_min = (0xFFFF, 0) if is_16bit else ((0x7FFFFFFF, -0x80000000) if is_signed else (0xFFFFFFFF, 0))
+          def to_wide(x): return (x.bitcast(narrow_dt) if x.dtype.itemsize == narrow_dt.itemsize else x.cast(narrow_dt)).cast(wide_dt)
+          is_sub, is_mad = 'SUB' in op.name, 'MAD' in op.name
+          full = (to_wide(s0) * to_wide(s1) + to_wide(s2)) if is_mad and s2 is not None else \
+                 (to_wide(s1) - to_wide(s0)) if is_sub and 'SUBREV' in op.name else \
+                 (to_wide(s0) - to_wide(s1)) if is_sub else (to_wide(s0) + to_wide(s1))
+          int_saturate = full.clamp(narrow_min, narrow_max).cast(narrow_dt)
+
    raw_stores: list = []
    vcc_val, exec_val = None, None
    for dest, val in assigns:
@@ -444,7 +478,10 @@ class _Ctx:
                       val.cast(dtypes.uint32) if val.dtype in (dtypes.uint16, dtypes.int16) else val.cast(dtypes.uint32) & UOp.const(dtypes.uint32, slice_mask)
            raw_stores.append(('vgpr_slice', (lo_bit, width, val_bits)))
            continue
-        val = _apply_clamp(val, clmp)
+        # For integer ops with clamp, use pre-computed saturated value; for floats, clamp to [0,1]
+        if int_saturate is not None: val = int_saturate
+        elif clmp and val.dtype in (dtypes.float32, dtypes.half, dtypes.float64):
+          val = val.maximum(UOp.const(val.dtype, 0.0)).minimum(UOp.const(val.dtype, 1.0))
        if val.dtype in (dtypes.uint64, dtypes.int64, dtypes.float64):
          lo, hi = _split64(val)
          raw_stores.extend([('vgpr', self.wvgpr_dyn(vdst_reg, lane, lo, exec_mask)), ('vgpr', self.wvgpr_dyn(vdst_reg + _c(1), lane, hi, exec_mask))])
@@ -479,13 +516,14 @@ class _Ctx:
 # INSTRUCTION HANDLERS
 # ═══════════════════════════════════════════════════════════════════════════════

-def _compile_sopp(inst: SOPP, ctx: _Ctx) -> UOp:
-  simm16 = ctx.inst_field_signed(SOPP.simm16).cast(dtypes.int16)
-  if inst.op == SOPPOp.S_ENDPGM:
+def _compile_sopp(inst: ir3.SOPP | ir4.SOPP, ctx: _Ctx) -> UOp:
+  simm16 = ctx.inst_field_signed(type(inst).simm16).cast(dtypes.int16)
+  if inst.op in (ir3.SOPPOp.S_ENDPGM, ir4.SOPPOp.S_ENDPGM):
    return UOp.sink(ctx.wsgpr_dyn(_c(PC_LO_IDX), UOp.const(dtypes.uint32, 0xFFFFFFFF)),
                          ctx.wsgpr_dyn(_c(PC_HI_IDX), UOp.const(dtypes.uint32, 0xFFFFFFFF)))
+  if inst.op in (ir3.SOPPOp.S_NOP, ir4.SOPPOp.S_NOP): return UOp.sink(*ctx.inc_pc())  # S_NOP is a no-op
  # NOTE: we ignore SOPPs without PCODE
-  if inst.op in PCODE:
+  if inst.op in _get_pcode_dict(inst.op):
    pcode = get_pcode(inst.op)
    pc_bytes = ctx.rpc()  # PC is already 64-bit byte address
    vcc, exec_lo = ctx.rsgpr_dyn(_c(VCC_LO.offset)), ctx.rsgpr_dyn(_c(EXEC_LO.offset))
@@ -497,50 +535,57 @@ def _compile_sopp(inst: SOPP, ctx: _Ctx) -> UOp:
        return UOp.sink(ctx.wsgpr_dyn(_c(PC_LO_IDX), lo), ctx.wsgpr_dyn(_c(PC_HI_IDX), hi))
  return UOp.sink(*ctx.inc_pc())

-def _compile_smem(inst: SMEM, ctx: _Ctx) -> UOp:
+def _compile_smem(inst: ir3.SMEM | ir4.SMEM, ctx: _Ctx) -> UOp:
  # Cache invalidation instructions are no-ops in the emulator (we don't model caches)
-  if inst.op in (SMEMOp.S_GL1_INV, SMEMOp.S_DCACHE_INV):
+  cache_inv_ops = [ir3.SMEMOp.S_GL1_INV, ir3.SMEMOp.S_DCACHE_INV, ir4.SMEMOp.S_DCACHE_INV]
+  if hasattr(ir4.SMEMOp, 'S_GL1_INV'): cache_inv_ops.append(ir4.SMEMOp.S_GL1_INV)
+  if inst.op in cache_inv_ops:
    return UOp.sink(*ctx.inc_pc())
  # Dynamic sbase field (bits 5:0) - SGPR pair, field value * 2 = register offset
-  sbase = ctx.inst_field(SMEM.sbase) * _c(2)
+  sbase = ctx.inst_field(type(inst).sbase) * _c(2)
  # Dynamic sdata field (bits 12:6) - destination SGPR
-  sdata_reg = ctx.inst_field(SMEM.sdata)
-  offset = ctx.inst_field_signed(SMEM.offset)  # 21-bit signed immediate
-  # Dynamic soffset field (bits 63:57) - SGPR for additional offset (NULL=124 reads as 0)
-  soffset = ctx.inst_field(SMEM.soffset)
+  sdata_reg = ctx.inst_field(type(inst).sdata)
+  # RDNA4 uses 'ioffset', RDNA3 uses 'offset' - use type(inst) to get correct field
+  offset_field = type(inst).ioffset if hasattr(type(inst), 'ioffset') else type(inst).offset
+  offset = ctx.inst_field_signed(offset_field)  # signed immediate
+  # Dynamic soffset field - SGPR for additional offset (NULL=124 reads as 0)
+  soffset = ctx.inst_field(type(inst).soffset)
  addr = _u64(ctx.rsgpr_dyn(sbase), ctx.rsgpr_dyn(sbase + _c(1))) + offset.cast(dtypes.uint64) + ctx.rsgpr_dyn(soffset).cast(dtypes.uint64)
-  ndwords = {SMEMOp.S_LOAD_B32: 1, SMEMOp.S_LOAD_B64: 2, SMEMOp.S_LOAD_B128: 4, SMEMOp.S_LOAD_B256: 8, SMEMOp.S_LOAD_B512: 16}.get(inst.op, 1)
+  _SMEM_NDWORDS = {ir3.SMEMOp.S_LOAD_B32: 1, ir3.SMEMOp.S_LOAD_B64: 2, ir3.SMEMOp.S_LOAD_B128: 4,
+    ir3.SMEMOp.S_LOAD_B256: 8, ir3.SMEMOp.S_LOAD_B512: 16, ir4.SMEMOp.S_LOAD_B32: 1, ir4.SMEMOp.S_LOAD_B64: 2,
+    ir4.SMEMOp.S_LOAD_B96: 3, ir4.SMEMOp.S_LOAD_B128: 4, ir4.SMEMOp.S_LOAD_B256: 8, ir4.SMEMOp.S_LOAD_B512: 16}
+  ndwords = _SMEM_NDWORDS[inst.op]
  stores = [ctx.wsgpr_dyn(sdata_reg + _c(i), ctx.vmem.index((addr + UOp.const(dtypes.uint64, i * 4) >> UOp.const(dtypes.uint64, 2)).cast(dtypes.int)))
            for i in range(ndwords)]
  return UOp.sink(*stores, *ctx.inc_pc())

-def _compile_sop(inst: SOP1 | SOP2 | SOPC | SOPK, ctx: _Ctx) -> UOp:
+def _compile_sop(inst: ir3.SOP1 | ir3.SOP2 | ir3.SOPC | ir3.SOPK | ir4.SOP1 | ir4.SOP2 | ir4.SOPC | ir4.SOPK, ctx: _Ctx) -> UOp:
  bits = inst.canonical_op_bits
  literal = ctx.inst_field(type(inst).literal) if hasattr(type(inst), 'literal') else None

-  if isinstance(inst, SOPK):
-    sdst_off = ctx.inst_field(SOPK.sdst)
-    simm16 = ctx.inst_field(SOPK.simm16)
+  if isinstance(inst, (ir3.SOPK, ir4.SOPK)):
+    sdst_off = ctx.inst_field(type(inst).sdst)
+    simm16 = ctx.inst_field(type(inst).simm16)
    # Sign-extend simm16
    simm16_sext = simm16.cast(dtypes.int16).cast(dtypes.int32)
    srcs = {'S0': ctx.rsgpr_dyn(sdst_off), 'SIMM16': simm16_sext, 'D0': ctx.rsgpr_dyn(sdst_off)}
    dst_off, dst_size = sdst_off, 1
-  elif isinstance(inst, SOP1):
-    sdst_off = ctx.inst_field(SOP1.sdst)
-    ssrc0_off = ctx.inst_field(SOP1.ssrc0)
+  elif isinstance(inst, (ir3.SOP1, ir4.SOP1)):
+    sdst_off = ctx.inst_field(type(inst).sdst)
+    ssrc0_off = ctx.inst_field(type(inst).ssrc0)
    srcs = {'S0': ctx.rsrc_dyn(ssrc0_off, None, bits['s0'], literal)}
    dst_off, dst_size = sdst_off, bits['d'] // 32
-  elif isinstance(inst, SOP2):
-    sdst_off = ctx.inst_field(SOP2.sdst)
-    ssrc0_off = ctx.inst_field(SOP2.ssrc0)
-    ssrc1_off = ctx.inst_field(SOP2.ssrc1)
+  elif isinstance(inst, (ir3.SOP2, ir4.SOP2)):
+    sdst_off = ctx.inst_field(type(inst).sdst)
+    ssrc0_off = ctx.inst_field(type(inst).ssrc0)
+    ssrc1_off = ctx.inst_field(type(inst).ssrc1)
    srcs = {'S0': ctx.rsrc_dyn(ssrc0_off, None, bits['s0'], literal),
            'S1': ctx.rsrc_dyn(ssrc1_off, None, bits['s1'], literal)}
    if literal is not None: srcs['SIMM32'] = literal
    dst_off, dst_size = sdst_off, bits['d'] // 32
-  elif isinstance(inst, SOPC):
-    ssrc0_off = ctx.inst_field(SOPC.ssrc0)
-    ssrc1_off = ctx.inst_field(SOPC.ssrc1)
+  elif isinstance(inst, (ir3.SOPC, ir4.SOPC)):
+    ssrc0_off = ctx.inst_field(type(inst).ssrc0)
+    ssrc1_off = ctx.inst_field(type(inst).ssrc1)
    srcs = {'S0': ctx.rsrc_dyn(ssrc0_off, None, bits['s0'], literal),
            'S1': ctx.rsrc_dyn(ssrc1_off, None, bits['s1'], literal)}
    dst_off, dst_size = _c(0), 0  # SOPC writes to SCC, not sdst
@@ -549,18 +594,18 @@ def _compile_sop(inst: SOP1 | SOP2 | SOPC | SOPK, ctx: _Ctx) -> UOp:

  return ctx.compile_sop_pcode(inst.op, srcs, dst_off, dst_size)

-def _compile_vop12(inst: VOP1 | VOP1_SDST | VOP2, ctx: _Ctx) -> UOp:
+def _compile_vop12(inst: ir3.VOP1 | ir3.VOP1_SDST | ir3.VOP2 | ir4.VOP1 | ir4.VOP1_SDST | ir4.VOP2, ctx: _Ctx) -> UOp:
  op_name = _op_name(inst)
-  if op_name == 'V_READFIRSTLANE_B32_E32': return ctx.compile_lane_pcode(inst.op, inst)
+  if op_name in ('V_READFIRSTLANE_B32_E32', 'V_PERMLANE64_B32_E32'): return ctx.compile_lane_pcode(inst.op, inst)
  lane, exec_mask, bits = ctx.range(), ctx.rsgpr_dyn(_c(EXEC_LO.offset)), inst.canonical_op_bits
  literal = ctx.inst_field(type(inst).literal) if hasattr(type(inst), 'literal') else None
-  vdst_reg = ctx.inst_field(VOP1.vdst)
+  vdst_reg = ctx.inst_field(type(inst).vdst)
  write_hi_half = bits['d'] == 16 and (vdst_reg >= _c(128))
  if isinstance(write_hi_half, UOp): vdst_reg = write_hi_half.where(vdst_reg - _c(128), vdst_reg)
  elif write_hi_half: vdst_reg -= 128
-  if isinstance(inst, VOP1):
+  if isinstance(inst, (ir3.VOP1, ir4.VOP1)):
    # Handle VOP1 hi-half source operand (src0 >= v[128] for 16-bit ops)
-    src0_off = ctx.inst_field(VOP1.src0)
+    src0_off = ctx.inst_field(type(inst).src0)
    s0 = ctx.rsrc_dyn(src0_off, lane, bits['s0'], literal)
    if bits['s0'] == 16:
      src0_hi = src0_off >= _c(384)
@@ -569,13 +614,13 @@ def _compile_vop12(inst: VOP1 | VOP1_SDST | VOP2, ctx: _Ctx) -> UOp:
      s0 = src0_hi.where(_hi16(ctx.rvgpr_dyn(src0_reg, lane)), s0)
    srcs = {'S0': s0}
  else:
-    vsrc1_reg = ctx.inst_field(VOP2.vsrc1)
+    vsrc1_reg = ctx.inst_field(type(inst).vsrc1)
    vsrc1_hi = bits['s0'] == 16 and (vsrc1_reg >= _c(128))
    vsrc1_actual = _cond(vsrc1_hi, vsrc1_reg - _c(128), vsrc1_reg)
    s1 = _cond_hi16(vsrc1_hi, ctx.rvgpr_dyn(vsrc1_actual, lane))
    d0 = _cond_hi16(write_hi_half, ctx.rvgpr_dyn(vdst_reg, lane))  # FMAC/FMAMK hi-half dest needs hi-half accumulator
    # Handle VOP2 hi-half src0 operand (src0 >= v[128] for 16-bit ops)
-    src0_off = ctx.inst_field(VOP2.src0)
+    src0_off = ctx.inst_field(type(inst).src0)
    s0 = ctx.rsrc_dyn(src0_off, lane, bits['s0'], literal)
    if bits['s0'] == 16:
      src0_hi = src0_off >= _c(384)
@@ -583,19 +628,20 @@ def _compile_vop12(inst: VOP1 | VOP1_SDST | VOP2, ctx: _Ctx) -> UOp:
      src0_reg = src0_hi.where(src0_off - _c(384), _c(0))
      s0 = src0_hi.where(_hi16(ctx.rvgpr_dyn(src0_reg, lane)), s0)
    srcs = {'S0': s0, 'S1': s1, 'D0': d0}
-    if inst.op in (VOP2Op.V_FMAAK_F32_E32, VOP2Op.V_FMAMK_F32_E32, VOP2Op.V_FMAAK_F16_E32, VOP2Op.V_FMAMK_F16_E32):
+    if inst.op in (ir3.VOP2Op.V_FMAAK_F32_E32, ir3.VOP2Op.V_FMAMK_F32_E32, ir3.VOP2Op.V_FMAAK_F16_E32,
+                   ir3.VOP2Op.V_FMAMK_F16_E32):
      assert literal is not None
      srcs['SIMM32'] = literal
  return ctx.compile_vop_pcode(inst.op, srcs, lane, vdst_reg, exec_mask, opsel_dst_hi=write_hi_half)

-def _compile_vopc(inst: VOPC | VOP3, ctx: _Ctx, opsel: int = 0, abs_bits: int = 0, neg_bits: int = 0) -> UOp:
+def _compile_vopc(inst: ir3.VOPC | ir3.VOP3 | ir4.VOPC | ir4.VOP3, ctx: _Ctx, opsel: int = 0, abs_bits: int = 0, neg_bits: int = 0) -> UOp:
  exec_mask, op_name, bits = ctx.rsgpr_dyn(_c(EXEC_LO.offset)), _op_name(inst), inst.canonical_op_bits
  is_cmpx, is_vopc = 'CMPX' in op_name, hasattr(inst, 'vsrc1')  # is_vopc: e32 vs e64

  # Handle both VOPC (vsrc1) and VOP3 (src1) instruction formats - read operands dynamically
  if is_vopc:
-    src0_off = ctx.inst_field(VOPC.src0)
-    vsrc1_off = ctx.inst_field(VOPC.vsrc1)
+    src0_off = ctx.inst_field(type(inst).src0)
+    vsrc1_off = ctx.inst_field(type(inst).vsrc1)
    # For 16-bit ops, vsrc1 >= 128 means hi-half of v[vsrc1-128]
    if bits['s0'] == 16:
      vsrc1_hi = vsrc1_off >= _c(128)
@@ -604,9 +650,9 @@ def _compile_vopc(inst: VOPC | VOP3, ctx: _Ctx, opsel: int = 0, abs_bits: int =
      vsrc1_hi = False
      src1_off = _c(256) + vsrc1_off
  else:
-    src0_off = ctx.inst_field(VOP3.src0)
-    src1_off = ctx.inst_field(VOP3.src1)
-    dst_off = ctx.inst_field(VOP3.vdst)
+    src0_off = ctx.inst_field(type(inst).src0)
+    src1_off = ctx.inst_field(type(inst).src1)
+    dst_off = ctx.inst_field(type(inst).vdst)
    vsrc1_hi = False
  literal = ctx.inst_field(type(inst).literal) if hasattr(type(inst), 'literal') else None

@@ -635,7 +681,7 @@ def _compile_vopc(inst: VOPC | VOP3, ctx: _Ctx, opsel: int = 0, abs_bits: int =
    stores = [ctx.wsgpr_dyn(dst_off, new_result)] if not is_vopc else [ctx.wsgpr_dyn(_c(VCC_LO.offset), new_result)]
  return UOp.sink(*stores, *ctx.inc_pc())

-def _compile_vop3(inst: VOP3, ctx: _Ctx) -> UOp:
+def _compile_vop3(inst: ir3.VOP3 | ir4.VOP3, ctx: _Ctx) -> UOp:
  exec_mask = ctx.rsgpr_dyn(_c(EXEC_LO.offset))
  bits = inst.canonical_op_bits
  opsel, op_name = getattr(inst, 'opsel', 0) or 0, _op_name(inst)
@@ -644,18 +690,22 @@ def _compile_vop3(inst: VOP3, ctx: _Ctx) -> UOp:
  if op_name in ('V_READLANE_B32', 'V_READFIRSTLANE_B32', 'V_READFIRSTLANE_B32_E64', 'V_WRITELANE_B32'):
    return ctx.compile_lane_pcode(inst.op, inst)

+  # V_PERMLANE16_B32 / V_PERMLANEX16_B32: cross-lane swizzle via pcode
+  if 'PERMLANE16' in op_name or 'PERMLANEX16' in op_name:
+    return ctx.compile_lane_pcode(inst.op, inst)
+
  # VOP3 VOPC (v_cmp_*_e64) - delegate to unified VOPC handler
  if 'V_CMP' in op_name or 'V_CMPX' in op_name:
    return _compile_vopc(inst, ctx, opsel=opsel, abs_bits=getattr(inst, 'abs', 0) or 0, neg_bits=getattr(inst, 'neg', 0) or 0)

  # Regular VOP3 - read operands dynamically
  lane = ctx.range()
-  vdst_reg = ctx.inst_field(VOP3.vdst)
+  vdst_reg = ctx.inst_field(type(inst).vdst)
  literal = ctx.inst_field(type(inst).literal) if hasattr(type(inst), 'literal') else None
  ops = inst.canonical_operands
-  src0 = ctx.rsrc_dyn(ctx.inst_field(VOP3.src0), lane, bits['s0'], literal, 's0' in ops and ops['s0'][0] == Fmt.FMT_NUM_F64)
-  src1 = ctx.rsrc_dyn(ctx.inst_field(VOP3.src1), lane, bits['s1'], literal, 's1' in ops and ops['s1'][0] == Fmt.FMT_NUM_F64)
-  src2 = ctx.rsrc_dyn(ctx.inst_field(VOP3.src2), lane, bits['s2'], literal, 's2' in ops and ops['s2'][0] == Fmt.FMT_NUM_F64)
+  src0 = ctx.rsrc_dyn(ctx.inst_field(type(inst).src0), lane, bits['s0'], literal, 's0' in ops and ops['s0'][0] == Fmt.FMT_NUM_F64)
+  src1 = ctx.rsrc_dyn(ctx.inst_field(type(inst).src1), lane, bits['s1'], literal, 's1' in ops and ops['s1'][0] == Fmt.FMT_NUM_F64)
+  src2 = ctx.rsrc_dyn(ctx.inst_field(type(inst).src2), lane, bits['s2'], literal, 's2' in ops and ops['s2'][0] == Fmt.FMT_NUM_F64)
  if bits['s0'] == 16:
    src0 = _apply_opsel(src0, 0, opsel)
    src1 = _apply_opsel(src1, 1, opsel)
@@ -665,19 +715,19 @@ def _compile_vop3(inst: VOP3, ctx: _Ctx) -> UOp:
  src1 = _apply_src_mods(src1, 1, abs_bits, neg_bits, bits['s1'])
  src2 = _apply_src_mods(src2, 2, abs_bits, neg_bits, bits['s2'])
  srcs = {'S0': src0, 'S1': src1, 'S2': src2}
-  if inst.op in (VOP3Op.V_CNDMASK_B32_E64, VOP3Op.V_CNDMASK_B16) and src2 is not None: srcs['VCC'] = src2
+  if inst.op in (ir3.VOP3Op.V_CNDMASK_B32_E64, ir3.VOP3Op.V_CNDMASK_B16) and src2 is not None: srcs['VCC'] = src2
  # FMAC instructions need D0 (accumulator) from destination register
  if 'FMAC' in op_name: srcs['D0'] = ctx.rvgpr_dyn(vdst_reg, lane)
  opsel_dst_hi = bool(opsel & 0b1000) and bits['d'] == 16
  return ctx.compile_vop_pcode(inst.op, srcs, lane, vdst_reg, exec_mask, opsel_dst_hi=opsel_dst_hi, clmp=getattr(inst, 'clmp', 0))

-def _compile_vop3sd(inst: VOP3SD, ctx: _Ctx) -> UOp:
+def _compile_vop3sd(inst: ir3.VOP3SD | ir4.VOP3SD, ctx: _Ctx) -> UOp:
  exec_mask = ctx.rsgpr_dyn(_c(EXEC_LO.offset))
  bits, pcode, ops = inst.canonical_op_bits, get_pcode(inst.op), inst.canonical_operands

  # Read operands dynamically from instruction encoding
-  vdst_reg, sdst_off = ctx.inst_field(VOP3SD.vdst), ctx.inst_field(VOP3SD.sdst)
-  src0_off, src1_off, src2_off = ctx.inst_field(VOP3SD.src0), ctx.inst_field(VOP3SD.src1), ctx.inst_field(VOP3SD.src2)
+  vdst_reg, sdst_off = ctx.inst_field(type(inst).vdst), ctx.inst_field(type(inst).sdst)
+  src0_off, src1_off, src2_off = ctx.inst_field(type(inst).src0), ctx.inst_field(type(inst).src1), ctx.inst_field(type(inst).src2)
  literal = ctx.inst_field(type(inst).literal) if hasattr(type(inst), 'literal') else None

  has_carry_in = 's2' in ops and ops['s2'][2] == OpType.OPR_SREG
@@ -695,6 +745,7 @@ def _compile_vop3sd(inst: VOP3SD, ctx: _Ctx) -> UOp:
  _, assigns = parse_pcode(pcode, srcs)

  has_per_lane_vcc = any('[laneId]' in dest for dest, _ in assigns if dest.startswith('VCC') or dest.startswith('D0.u64'))
+  clmp = getattr(inst, 'clmp', 0)
  if has_per_lane_vcc:
    # VCC computation: RANGE+REDUCE gets axis ID first (lower ID = runs first)
    # This ensures VCC reads source values BEFORE VGPR stores modify them
@@ -706,11 +757,17 @@ def _compile_vop3sd(inst: VOP3SD, ctx: _Ctx) -> UOp:
    final_vcc = ctx.unroll_lanes(get_vcc_bit, exec_mask)
    # VGPR stores: RANGE gets axis ID second (higher ID = runs after VCC loop)
    lane3 = ctx.range()
-    d0_val = None
+    d0_val, vcc_per_lane = None, None
    for dest, val in parse_pcode(pcode, load_srcs(lane3))[1]:
      if dest.startswith('D0') and '[laneId]' not in dest: d0_val = val
+      if dest.startswith('VCC') or (dest.startswith('D0.u64') and '[laneId]' in dest): vcc_per_lane = val
    vgpr_stores = []
    if d0_val is not None:
+      # Apply clamp using carry/borrow bit: ADD overflow->0xFFFFFFFF, SUB underflow->0
+      if clmp and vcc_per_lane is not None:
+        is_sub = 'SUB' in inst.op.name
+        sat_val = _c(0) if is_sub else _c(0xFFFFFFFF)
+        d0_val = vcc_per_lane.cast(dtypes.bool).where(sat_val, d0_val.cast(dtypes.uint32))
      if d0_val.dtype in (dtypes.uint64, dtypes.int64, dtypes.float64):
        lo, hi = _split64(d0_val)
        vgpr_stores.extend([ctx.wvgpr_dyn(vdst_reg, lane3, lo, exec_mask), ctx.wvgpr_dyn(vdst_reg + _c(1), lane3, hi, exec_mask)])
@@ -723,13 +780,13 @@ def _compile_vop3sd(inst: VOP3SD, ctx: _Ctx) -> UOp:
  else:
    return ctx.compile_vop_pcode(inst.op, srcs, lane, vdst_reg, exec_mask, sdst_reg=inst.sdst.offset)

-def _compile_wmma(inst: VOP3P, ctx: _Ctx) -> UOp:
+def _compile_wmma(inst: ir3.VOP3P | ir4.VOP3P, ctx: _Ctx) -> UOp:
  op_name = _op_name(inst)
  exec_mask = ctx.rsgpr_dyn(_c(EXEC_LO.offset))
-  vdst_reg = ctx.inst_field(VOP3P.vdst)
-  src0_r = ctx.inst_field(VOP3P.src0) - _c(256)
-  src1_r = ctx.inst_field(VOP3P.src1) - _c(256)
-  src2_r = ctx.inst_field(VOP3P.src2) - _c(256)
+  vdst_reg = ctx.inst_field(type(inst).vdst)
+  src0_r = ctx.inst_field(type(inst).src0) - _c(256)
+  src1_r = ctx.inst_field(type(inst).src1) - _c(256)
+  src2_r = ctx.inst_field(type(inst).src2) - _c(256)
  is_f16_output = 'F16_16X16X16_F16' in op_name or 'BF16_16X16X16_BF16' in op_name  # F16/BF16 output vs F32 output
  is_bf16 = 'BF16' in op_name
  cvt = _FUNCS['bf16_to_f32'] if is_bf16 else _FUNCS['f16_to_f32']
@@ -756,16 +813,17 @@ def _compile_wmma(inst: VOP3P, ctx: _Ctx) -> UOp:
    stores = [ctx.wvgpr_dyn(vdst_reg + _c(i // 32), UOp.const(dtypes.int, i % 32), mat_d[i].bitcast(dtypes.uint32), exec_mask) for i in range(256)]
  return UOp.sink(*stores, *ctx.inc_pc())

-def _compile_vop3p(inst: VOP3P, ctx: _Ctx) -> UOp:
+def _compile_vop3p(inst: ir3.VOP3P | ir4.VOP3P, ctx: _Ctx) -> UOp:
  op_name = _op_name(inst)
  if 'WMMA' in op_name and ('16X16X16_F16' in op_name or '16X16X16_BF16' in op_name): return _compile_wmma(inst, ctx)

  lane = ctx.range()
  exec_mask = ctx.rsgpr_dyn(_c(EXEC_LO.offset))
-  vdst_reg = ctx.inst_field(VOP3P.vdst)
-  src0 = ctx.rsrc_dyn(ctx.inst_field(VOP3P.src0), lane, 16)
-  src1 = ctx.rsrc_dyn(ctx.inst_field(VOP3P.src1), lane, 16)
-  src2 = ctx.rsrc_dyn(ctx.inst_field(VOP3P.src2), lane, 16)
+  vdst_reg = ctx.inst_field(type(inst).vdst)
+  do_cast = any(x in op_name for x in ('F16', 'F32', 'BF16')) and 'IU' not in op_name
+  src0 = ctx.rsrc_dyn(ctx.inst_field(type(inst).src0), lane, 16, do_cast=do_cast)
+  src1 = ctx.rsrc_dyn(ctx.inst_field(type(inst).src1), lane, 16, do_cast=do_cast)
+  src2 = ctx.rsrc_dyn(ctx.inst_field(type(inst).src2), lane, 16, do_cast=do_cast)
  opsel, opsel_hi = getattr(inst, 'opsel', 0) or 0, getattr(inst, 'opsel_hi', 3) if getattr(inst, 'opsel_hi', 3) is not None else 3
  opsel_hi2 = getattr(inst, 'opsel_hi2', 1) if getattr(inst, 'opsel_hi2', 1) is not None else 1
  neg, neg_hi = getattr(inst, 'neg', 0) or 0, getattr(inst, 'neg_hi', 0) or 0
@@ -787,7 +845,7 @@ def _compile_vop3p(inst: VOP3P, ctx: _Ctx) -> UOp:
    s0_mod = apply_neg_mix(apply_abs(src0, 1, 1, 1), 1, 1, 1)
    s1_mod = apply_neg_mix(apply_abs(src1, 2, 2, 2), 2, 2, 2)
    s2_mod = apply_neg_mix(apply_abs(src2, 4, 4, 4), 4, 4, 4)
-    srcs = {'S0': s0_mod, 'S1': s1_mod, 'S2': s2_mod,
+    srcs = {'S@0': s0_mod, 'S@1': s1_mod, 'S@2': s2_mod,
            'OPSEL_HI': UOp.const(dtypes.uint32, combined_opsel_hi), 'OPSEL': UOp.const(dtypes.uint32, opsel)}
  else:
    def get_half_bits(val: UOp, use_hi: bool, apply_neg: bool = False) -> UOp:
@@ -796,24 +854,28 @@ def _compile_vop3p(inst: VOP3P, ctx: _Ctx) -> UOp:
      return bits
    def build_remapped_src(src: UOp, opsel_lo_bit: int, opsel_hi_bit: int, neg_lo_bit: int, neg_hi_bit: int) -> UOp:
      return get_half_bits(src, bool(opsel_lo_bit), bool(neg_lo_bit)) | (get_half_bits(src, bool(opsel_hi_bit), bool(neg_hi_bit)) << UOp.const(dtypes.uint32, 16))
-    s0_new = build_remapped_src(src0, opsel & 1, opsel_hi & 1, neg & 1, neg_hi & 1)
-    s1_new = build_remapped_src(src1, opsel & 2, opsel_hi & 2, neg & 2, neg_hi & 2)
-    s2_new = build_remapped_src(src2, opsel & 4, 1 if opsel_hi2 else 0, neg & 4, neg_hi & 4)
-    srcs = {'S0': s0_new, 'S1': s1_new, 'S2': s2_new}
+    # DOT IU instructions use NEG bits for signed/unsigned selection, not fp16 negation
+    is_dot_iu = 'DOT' in op_name and 'IU' in op_name
+    n0, n1, n2, nh0, nh1, nh2 = (0, 0, 0, 0, 0, 0) if is_dot_iu else (neg & 1, neg & 2, neg & 4, neg_hi & 1, neg_hi & 2, neg_hi & 4)
+    srcs = {'S0': build_remapped_src(src0, opsel & 1, opsel_hi & 1, n0, nh0),
+            'S1': build_remapped_src(src1, opsel & 2, opsel_hi & 2, n1, nh1),
+            'S2': build_remapped_src(src2, opsel & 4, 1 if opsel_hi2 else 0, n2, nh2)}
+    if is_dot_iu: srcs['NEG'] = UOp.const(dtypes.uint32, neg)
  return ctx.compile_vop_pcode(inst.op, srcs, lane, vdst_reg, exec_mask)

-def _compile_vopd(inst: VOPD, ctx: _Ctx) -> UOp:
+def _compile_vopd(inst: ir3.VOPD | ir4.VOPD, ctx: _Ctx) -> UOp:
  exec_mask = ctx.rsgpr_dyn(_c(EXEC_LO.offset))
-  # Read operands dynamically
-  vdstx_reg = ctx.inst_field(VOPD.vdstx)
+  # Read operands dynamically - use type(inst) to get correct field descriptors
+  inst_type = type(inst)
+  vdstx_reg = ctx.inst_field(inst_type.vdstx)
  # vdsty has complex encoding: actual = (raw << 1) | ((vdstx & 1) ^ 1)
-  vdsty_raw = ctx.inst_field(VOPD.vdsty)
+  vdsty_raw = ctx.inst_field(inst_type.vdsty)
  vdsty_reg = (vdsty_raw << _c(1)) | ((vdstx_reg & _c(1)) ^ _c(1))
-  srcx0_off = ctx.inst_field(VOPD.srcx0)
-  srcy0_off = ctx.inst_field(VOPD.srcy0)
-  vsrcx1_reg = ctx.inst_field(VOPD.vsrcx1)
-  vsrcy1_reg = ctx.inst_field(VOPD.vsrcy1)
-  literal = ctx.inst_field(type(inst).literal) if hasattr(type(inst), 'literal') else None
+  srcx0_off = ctx.inst_field(inst_type.srcx0)
+  srcy0_off = ctx.inst_field(inst_type.srcy0)
+  vsrcx1_reg = ctx.inst_field(inst_type.vsrcx1)
+  vsrcy1_reg = ctx.inst_field(inst_type.vsrcy1)
+  literal = ctx.inst_field(inst_type.literal) if hasattr(inst_type, 'literal') else None

  lane = ctx.range()
  srcy0, srcy1 = ctx.rsrc_dyn(srcy0_off, lane, literal=literal), ctx.rvgpr_dyn(vsrcy1_reg, lane)
@@ -824,49 +886,64 @@ def _compile_vopd(inst: VOPD, ctx: _Ctx) -> UOp:
    assert vop is not None, f"no VOP mapping for VOPD {label}: {op}"
    if label == 'Y': srcs = {'S0': srcy0, 'S1': srcy1, 'D0': ctx.rvgpr_dyn(vdst_reg, lane)}
    else: srcs = {'S0': ctx.rsrc_dyn(src0_off, lane, literal=literal), 'S1': ctx.rvgpr_dyn(vsrc1_reg, lane), 'D0': ctx.rvgpr_dyn(vdst_reg, lane)}
-    if op in (VOPDOp.V_DUAL_FMAAK_F32, VOPDOp.V_DUAL_FMAMK_F32):
+    if op in (ir3.VOPDOp.V_DUAL_FMAAK_F32, ir3.VOPDOp.V_DUAL_FMAMK_F32, ir4.VOPDOp.V_DUAL_FMAAK_F32, ir4.VOPDOp.V_DUAL_FMAMK_F32):
      assert literal is not None
      srcs['SIMM32'] = literal
-    if op == VOPDOp.V_DUAL_CNDMASK_B32: srcs['VCC'] = ctx.rsgpr_dyn(_c(VCC_LO.offset))
+    if op in (ir3.VOPDOp.V_DUAL_CNDMASK_B32, ir4.VOPDOp.V_DUAL_CNDMASK_B32): srcs['VCC'] = ctx.rsgpr_dyn(_c(VCC_LO.offset))
    pcode = get_pcode(vop)
    srcs.update({'VCC': ctx.rsgpr_dyn(_c(VCC_LO.offset)), 'EXEC': exec_mask, 'SCC': ctx.rsgpr_dyn(_c(SCC.offset)), 'laneId': lane})
    for dest, val in parse_pcode(pcode, srcs)[1]:
      if dest.startswith('D0'): all_stores.append(ctx.wvgpr_dyn(vdst_reg, lane, _val_to_u32(val), exec_mask, after=srcy1))
  return UOp.sink(UOp.group(*all_stores).end(lane), *ctx.inc_pc())

-def _compile_mem_op(inst: DS | FLAT | GLOBAL | SCRATCH, ctx: _Ctx) -> UOp:
+def _compile_mem_op(inst: ir3.DS | ir3.FLAT | ir3.GLOBAL | ir3.SCRATCH | ir4.DS | ir4.VFLAT | ir4.VGLOBAL | ir4.VSCRATCH, ctx: _Ctx) -> UOp:
  """Unified memory operation compiler for DS, FLAT, GLOBAL, SCRATCH."""
  exec_mask, op_name = ctx.rsgpr_dyn(_c(EXEC_LO.offset)), _op_name(inst)
  pcode = get_pcode(inst.op)

-  is_lds = isinstance(inst, DS)
-  is_scratch = isinstance(inst, SCRATCH)
+  is_lds = isinstance(inst, (ir3.DS, ir4.DS))
+  is_scratch = isinstance(inst, (ir3.SCRATCH, ir4.VSCRATCH))
  mem = ctx.lds if is_lds else ctx.scratch if is_scratch else ctx.vmem
  addr_shift = UOp.const(dtypes.uint32 if is_lds else dtypes.uint64, 2)

  # Extract register info - all dynamic for deduplication
  if is_lds:
-    addr_reg = ctx.inst_field(DS.addr)
-    vdata_reg = ctx.inst_field(DS.data0)
-    vdst_reg = ctx.inst_field(DS.vdst)
-    offset0 = ctx.inst_field(DS.offset0)
-    offset1 = ctx.inst_field(DS.offset1)
+    addr_reg = ctx.inst_field(type(inst).addr)
+    vdata_reg = ctx.inst_field(type(inst).data0)
+    vdst_reg = ctx.inst_field(type(inst).vdst)
+    offset0 = ctx.inst_field(type(inst).offset0)
+    offset1 = ctx.inst_field(type(inst).offset1)
    offset = offset0  # DS uses offset0 as primary offset
    saddr_reg = None
-  else:
+  elif isinstance(inst, (ir4.VGLOBAL, ir4.VSCRATCH, ir4.VFLAT)):  # RDNA4: vaddr, vsrc, ioffset
+    addr_reg = ctx.inst_field(type(inst).vaddr)
+    vdata_reg = ctx.inst_field(type(inst).vsrc)
+    vdst_reg = ctx.inst_field(type(inst).vdst)
+    offset = ctx.inst_field_signed(type(inst).ioffset)
+    offset0, offset1 = _c(0), _c(0)
+    saddr_reg = ctx.inst_field(type(inst).saddr) if hasattr(type(inst), 'saddr') else None
+  else:  # RDNA3: addr, data, offset
    addr_reg = ctx.inst_field(type(inst).addr)
    vdata_reg = ctx.inst_field(type(inst).data)
    vdst_reg = ctx.inst_field(type(inst).vdst)
    offset = ctx.inst_field_signed(type(inst).offset)
    offset0, offset1 = _c(0), _c(0)
-    # Dynamic saddr - read field, NULL (124) or >= 128 means no saddr
-    saddr_reg = ctx.inst_field(type(inst).saddr) if hasattr(inst, 'saddr') else None
+    saddr_reg = ctx.inst_field(type(inst).saddr) if hasattr(type(inst), 'saddr') else None

  # Data width from canonical_op_bits (32/64/96/128), default to 32 for untyped ops
  data_bits_mem = inst.canonical_op_bits.get('data', 32)
  is_atomic, glc = 'ATOMIC' in op_name, getattr(inst, 'glc', 0)
  has_data1 = is_lds and hasattr(inst, 'data1') and inst.data1 is not None
-  data1_reg = ctx.inst_field(DS.data1) if is_lds else _c(0)
+  data1_reg = ctx.inst_field(type(inst).data1) if is_lds else _c(0)
+
+  # DS_PERMUTE/DS_BPERMUTE: cross-lane VGPR access via pcode
+  if is_lds and 'PERMUTE' in op_name:
+    pcode = get_pcode(inst.op)
+    srcs = {'ADDR': addr_reg, 'DATA0': vdata_reg, 'VDST': vdst_reg, 'OFFSET': offset,
+            'EXEC': exec_mask.cast(dtypes.uint64), '_vgpr': ctx.vgpr}
+    _, assigns = parse_pcode(pcode, srcs)
+    stores = [ctx.vgpr.index(val[0].cast(dtypes.int)).store(val[1].cast(dtypes.uint32)) for dest, val in assigns if dest.startswith('VGPR[')]
+    return UOp.sink(*stores, *ctx.inc_pc())

  def make_addr(lane: UOp) -> UOp:
    if is_lds: return ctx.rvgpr_dyn(addr_reg, lane)
@@ -908,14 +985,26 @@ def _compile_mem_op(inst: DS | FLAT | GLOBAL | SCRATCH, ctx: _Ctx) -> UOp:
      else:
        data = {'DATA': _u64(ctx.rvgpr_dyn(vdata_reg, lane), ctx.rvgpr_dyn(vdata_reg + _c(1), lane)),
                'DATA2': _u64(ctx.rvgpr_dyn(data1_reg, lane), ctx.rvgpr_dyn(data1_reg + _c(1), lane)) if has_data1 else UOp.const(dtypes.uint64, 0)}
-      return {'ADDR': addr, 'ADDR_BASE': addr, 'OFFSET': offset, 'OFFSET0': offset0, 'OFFSET1': offset1, '_lds': mem, 'laneId': lane, **data}
+      # RDNA3 uses ADDR/OFFSET, RDNA4 uses vgpr_a/offset (lowercase) + CalcDsAddr function
+      return {'ADDR': addr, 'ADDR_BASE': addr, 'OFFSET': offset, 'OFFSET0': offset0, 'OFFSET1': offset1, '_lds': mem, 'laneId': lane,
+              'vgpr_a': ctx.rvgpr_dyn(addr_reg, lane), 'offset': offset, **data}
    active = _lane_active(exec_mask, lane)
+    # saddr < 124 means valid SGPR pair, otherwise use 0 (NULL means no saddr contribution)
+    use_saddr = (saddr_reg < _c(124)) if saddr_reg is not None else UOp.const(dtypes.bool, False)
+    saddr_raw = _u64(ctx.rsgpr_dyn(saddr_reg), ctx.rsgpr_dyn(saddr_reg + _c(1))) if saddr_reg is not None else UOp.const(dtypes.uint64, 0)
+    saddr_base = use_saddr.where(saddr_raw, UOp.const(dtypes.uint64, 0))
+    # Sign-extend offset to 64-bit for the final address calculation
+    ioffset64 = offset.cast(dtypes.int64).cast(dtypes.uint64)
+    # v_addr for CalcGlobalAddr: when saddr valid, use low 32 bits as offset; otherwise full 64-bit address. Include ioffset.
+    vaddr_full = _u64(ctx.rvgpr_dyn(addr_reg, lane), ctx.rvgpr_dyn(addr_reg + _c(1), lane))
+    vaddr_lo = ctx.rvgpr_dyn(addr_reg, lane).cast(dtypes.uint64)
+    vaddr_base = use_saddr.where(vaddr_lo + ioffset64, vaddr_full + ioffset64)
    if is_atomic:
      return {'ADDR': addr, 'DATA': _u64(ctx.rvgpr_dyn(vdata_reg, lane), ctx.rvgpr_dyn(vdata_reg + _c(1), lane)) if data_bits_mem == 64 else ctx.rvgpr_dyn(vdata_reg, lane),
-              '_vmem': mem, '_active': active, 'laneId': lane}
+              '_vmem': mem, '_active': active, 'laneId': lane, 'v_addr': vaddr_base, 's_saddr': saddr_base}
    vdata = ctx.rvgpr_dyn(vdata_reg, lane).cast(dtypes.uint64) if 'STORE' in op_name else ctx.rvgpr_dyn(vdst_reg, lane) if 'D16' in op_name else UOp.const(dtypes.uint32, 0)
    if 'STORE' in op_name and data_bits_mem >= 64: vdata = vdata | (ctx.rvgpr_dyn(vdata_reg + _c(1), lane).cast(dtypes.uint64) << UOp.const(dtypes.uint64, 32))
-    srcs = {'ADDR': addr, 'VDATA': vdata, '_vmem': mem, '_active': active, 'laneId': lane}
+    srcs = {'ADDR': addr, 'VDATA': vdata, '_vmem': mem, '_active': active, 'laneId': lane, 'v_addr': vaddr_base, 's_saddr': saddr_base}
    for i in range(data_bits_mem // 32): srcs[f'VDATA{i}'] = ctx.rvgpr_dyn(vdata_reg + _c(i), lane) if 'STORE' in op_name else UOp.const(dtypes.uint32, 0)
    return srcs

@@ -966,10 +1055,15 @@ def _compile_mem_op(inst: DS | FLAT | GLOBAL | SCRATCH, ctx: _Ctx) -> UOp:

 # Dispatch table: instruction type -> handler function
 _INST_HANDLERS: dict[type, Callable[..., UOp]] = {
-  SOPP: _compile_sopp, SMEM: _compile_smem, SOP1: _compile_sop, SOP2: _compile_sop, SOPC: _compile_sop, SOPK: _compile_sop,
-  VOP1: _compile_vop12, VOP1_SDST: _compile_vop12, VOP2: _compile_vop12, VOPC: _compile_vopc, VOP3: _compile_vop3, VOP3_SDST: _compile_vop3,
-  VOP3SD: _compile_vop3sd, VOP3P: _compile_vop3p, VOPD: _compile_vopd,
-  DS: _compile_mem_op, FLAT: _compile_mem_op, GLOBAL: _compile_mem_op, SCRATCH: _compile_mem_op,
+  ir3.SOPP: _compile_sopp, ir3.SMEM: _compile_smem, ir3.SOP1: _compile_sop, ir3.SOP2: _compile_sop, ir3.SOPC: _compile_sop, ir3.SOPK: _compile_sop,
+  ir3.VOP1: _compile_vop12, ir3.VOP1_SDST: _compile_vop12, ir3.VOP2: _compile_vop12, ir3.VOPC: _compile_vopc, ir3.VOP3: _compile_vop3,
+  ir3.VOP3_SDST: _compile_vop3, ir3.VOP3SD: _compile_vop3sd, ir3.VOP3P: _compile_vop3p, ir3.VOPD: _compile_vopd,
+  ir3.DS: _compile_mem_op, ir3.FLAT: _compile_mem_op, ir3.GLOBAL: _compile_mem_op, ir3.SCRATCH: _compile_mem_op,
+  # RDNA4 instruction classes
+  ir4.SOPP: _compile_sopp, ir4.SMEM: _compile_smem, ir4.SOP1: _compile_sop, ir4.SOP2: _compile_sop, ir4.SOPC: _compile_sop, ir4.SOPK: _compile_sop,
+  ir4.VOP1: _compile_vop12, ir4.VOP1_SDST: _compile_vop12, ir4.VOP2: _compile_vop12, ir4.VOPC: _compile_vopc, ir4.VOP3: _compile_vop3,
+  ir4.VOP3_SDST: _compile_vop3, ir4.VOP3SD: _compile_vop3sd, ir4.VOP3P: _compile_vop3p, ir4.VOPD: _compile_vopd,
+  ir4.DS: _compile_mem_op, ir4.VFLAT: _compile_mem_op, ir4.VGLOBAL: _compile_mem_op, ir4.VSCRATCH: _compile_mem_op,
 }

 # ═══════════════════════════════════════════════════════════════════════════════
@@ -979,9 +1073,9 @@ _INST_HANDLERS: dict[type, Callable[..., UOp]] = {
 _canonical_runner_cache: list[tuple[int, int, int, object]] = []  # [(base, mask, size, runner), ...]

@functools.cache
-def _get_runner(inst_bytes: bytes):
+def _get_runner(inst_bytes: bytes, arch: str = "rdna3"):
  """Build and compile instruction to CompiledRunner. Cached by instruction bytes, with canonical dedup."""
-  inst = decode_inst(inst_bytes)
+  inst = decode_inst(inst_bytes, arch)
  inst_size = inst.size()
  inst_int = int.from_bytes(inst_bytes[:inst_size], 'little')

@@ -1004,21 +1098,21 @@ def _get_runner(inst_bytes: bytes):
  canonical_name = f"{_op_name(inst).lower()}_{base.to_bytes(size, 'little').hex()}"
  sink = sink.replace(arg=KernelInfo(name=canonical_name)).rtag(1)

-  with Context(NOOPT=1, CHECK_OOB=0, TUPLE_ORDER=0):
+  with Context(NOOPT=1, CHECK_OOB=0, TUPLE_ORDER=0, EMULATED_DTYPES=""):
    runner = get_runner('CPU', sink)
  _canonical_runner_cache.append((base, mask, size, runner))
  return runner, True

@functools.cache
-def decode_program(data: bytes) -> dict[int, tuple[str, Callable, list[int], Any]]:
+def decode_program(data: bytes, arch: str = "rdna3") -> dict[int, tuple[str, Callable, list[int], Any]]:
  """Decode program to {pc: (name, fxn, globals, runner)}."""
  result: dict[int, tuple[str, Callable, list[int], Any]] = {}
  i = 0
  while i < len(data):
-    inst = decode_inst(data[i:])
-    if isinstance(inst, SOPP) and inst.op == SOPPOp.S_CODE_END: break
+    inst = decode_inst(data[i:], arch)
+    if hasattr(inst, 'op') and inst.op in (ir3.SOPPOp.S_CODE_END, ir4.SOPPOp.S_CODE_END): break
    try:
-      runner, is_new = _get_runner(bytes(data[i:i + inst.size() + 4]))
+      runner, is_new = _get_runner(bytes(data[i:i + inst.size() + 4]), arch)
      if DEBUG >= 3:
        try: inst_str = repr(inst)
        except Exception: inst_str = f"<{type(inst).__name__} at PC={i}>"
@@ -1047,8 +1141,8 @@ class WaveState:
    self.n_lanes = n_lanes
    self.vgpr_buf = Buffer('CPU', VGPR_SIZE, dtypes.uint32).ensure_allocated()
    self.sgpr_buf = Buffer('CPU', SGPR_COUNT, dtypes.uint32).ensure_allocated()
-    self._vgpr_mv = self.vgpr_buf.as_buffer(force_zero_copy=True).cast('I')
-    self._sgpr_mv = self.sgpr_buf.as_buffer(force_zero_copy=True).cast('I')
+    self._vgpr_mv = self.vgpr_buf.as_memoryview(force_zero_copy=True).cast('I')
+    self._sgpr_mv = self.sgpr_buf.as_memoryview(force_zero_copy=True).cast('I')
    # Zero memory using ctypes memset (much faster than Python loops)
    ctypes.memset(self.vgpr_buf._buf.va_addr, 0, VGPR_SIZE * 4)
    ctypes.memset(self.sgpr_buf._buf.va_addr, 0, SGPR_COUNT * 4)
@@ -1077,9 +1171,9 @@ class WaveState:
 # ═══════════════════════════════════════════════════════════════════════════════

 def run_asm(lib: int, lib_sz: int, gx: int, gy: int, gz: int, lx: int, ly: int, lz: int, args_ptr: int, rsrc2: int = 0x19c,
-            scratch_size: int = 0) -> int:
+            scratch_size: int = 0, arch: str = "rdna3") -> int:
  """Execute AMD assembly program. scratch_size is private_segment_fixed_size from kernel descriptor (per-lane)."""
-  program_raw = decode_program(bytes((ctypes.c_char * lib_sz).from_address(lib).raw))
+  program_raw = decode_program(bytes((ctypes.c_char * lib_sz).from_address(lib).raw), arch)
  program = {lib + offset: val for offset, val in program_raw.items()}  # Remap to actual addresses
  lds_size = ((rsrc2 & hsa.AMD_COMPUTE_PGM_RSRC_TWO_GRANULATED_LDS_SIZE) >> hsa.AMD_COMPUTE_PGM_RSRC_TWO_GRANULATED_LDS_SIZE_SHIFT) * 512
  total_threads = lx * ly * lz
@@ -1107,6 +1201,12 @@ def run_asm(lib: int, lib_sz: int, gx: int, gy: int, gz: int, lx: int, ly: int,
                                 (hsa.AMD_COMPUTE_PGM_RSRC_TWO_ENABLE_SGPR_WORKGROUP_ID_Z, gidz)]:
              if rsrc2 & enabled: st._write_sgpr(sgpr_idx, gid); sgpr_idx += 1

+            # RDNA4 uses TTMP registers for workgroup IDs: ttmp[9]=gidx, ttmp[10]=gidy, ttmp[11]=gidz
+            if arch == "rdna4":
+              st._write_sgpr(ttmp[9].offset, gidx)
+              st._write_sgpr(ttmp[10].offset, gidy)
+              st._write_sgpr(ttmp[11].offset, gidz)
+
            # v0 = packed workitem IDs, scratch stride in secret SGPR
            for lane in range(n_lanes):
              tid = wave_start + lane
@@ -1123,7 +1223,7 @@ def run_asm(lib: int, lib_sz: int, gx: int, gy: int, gz: int, lx: int, ly: int,
              assert fxn is not None, f"[emu] No fxn for {name} at PC={pc}"
              assert 4 not in globals_list or scratch_buf, f"SCRATCH instruction {name} but scratch_size=0"
              if DEBUG >= 6:
-                inst = decode_inst(bytes((ctypes.c_char * 12).from_address(pc).raw))
+                inst = decode_inst(bytes((ctypes.c_char * 12).from_address(pc).raw), arch)
                print(f"[emu] exec PC={pc:X}: {inst!r}")
              fxn(*[c_bufs[g] for g in globals_list])
            else: raise RuntimeError("exceeded 1M instructions, likely infinite loop")
@@ -94,13 +94,19 @@ def _trig_reduce(x, phase=0.0):
  return UOp(Ops.SIN, x.dtype, (x - n * _const(x.dtype, 6.283185307179586),))

 def _signext(val: UOp) -> UOp:
-  for bits, mask, ext in [(8, 0xFF, 0xFFFFFF00), (16, 0xFFFF, 0xFFFF0000)]:
+  for bits, mask, ext in [(4, 0xF, 0xFFFFFFF0), (8, 0xFF, 0xFFFFFF00), (16, 0xFFFF, 0xFFFF0000)]:
    if (val.op == Ops.AND and len(val.src) == 2 and val.src[1].op == Ops.CONST and val.src[1].arg == mask) or val.dtype.itemsize == bits // 8:
      v32 = val.cast(dtypes.uint32) if val.dtype != dtypes.uint32 else val
      sb = (v32 >> _u32(bits - 1)) & _u32(1)
      return sb.ne(_u32(0)).where(v32 | _u32(ext), v32).cast(dtypes.int)
  return val.cast(dtypes.int64) if val.dtype in (dtypes.int, dtypes.int32) else val

+def _signext_4bit(val: UOp) -> UOp:
+  """Sign extend a 4-bit value to 32-bit signed integer."""
+  v32 = val.cast(dtypes.uint32) if val.dtype != dtypes.uint32 else val
+  sb = (v32 >> _u32(3)) & _u32(1)  # sign bit at position 3
+  return sb.ne(_u32(0)).where(v32 | _u32(0xFFFFFFF0), v32).bitcast(dtypes.int)
+
 def _abs(val: UOp) -> UOp:
  if val.dtype not in (dtypes.float32, dtypes.float64, dtypes.half): return val
  _, _, _, _, shift = _float_info(val)
@@ -194,6 +200,17 @@ def _ff1(val: UOp, bits: int) -> UOp:
    result = cond.where(_const(dtypes.int, i), result)
  return result

+def _sad_u8(a: UOp, b: UOp, acc: UOp, masked: bool = False) -> UOp:
+  """Sum of absolute differences of 4 unsigned bytes + accumulator. If masked, skips bytes where a == 0."""
+  a, b, acc = a.cast(dtypes.uint32), b.cast(dtypes.uint32), acc.cast(dtypes.uint32)
+  result = acc
+  for i in range(4):
+    a_byte = (a >> _u32(i * 8)) & _u32(0xFF)
+    b_byte = (b >> _u32(i * 8)) & _u32(0xFF)
+    diff = (a_byte > b_byte).where(a_byte - b_byte, b_byte - a_byte)
+    result = result + (a_byte.ne(_u32(0)).where(diff, _u32(0)) if masked else diff)
+  return result
+
 _FUNCS: dict[str, Callable[..., UOp]] = {
  'sqrt': lambda a: UOp(Ops.SQRT, a.dtype, (a,)), 'trunc': lambda a: UOp(Ops.TRUNC, a.dtype, (a,)),
  'log2': lambda a: UOp(Ops.LOG2, a.dtype, (a,)), 'sin': lambda a: _trig_reduce(a),
@@ -227,11 +244,53 @@ _FUNCS: dict[str, Callable[..., UOp]] = {
  'signext_from_bit': _signext_from_bit, 'ldexp': _ldexp, 'frexp_mant': _frexp_mant, 'mantissa': _frexp_mant,
  'frexp_exp': _frexp_exp, 'trig_preop_result': _trig_preop,
  's_ff1_i32_b32': lambda a: _ff1(a, 32), 's_ff1_i32_b64': lambda a: _ff1(a, 64),
+  # Normalization conversions: map [-1,1] or [0,1] to integer range
+  # Use floor(x + 0.5) for round-to-nearest
+  # SNORM: round(value * 32767), range is [-32767, 32767] (hardware behavior)
+  'f16_to_snorm': lambda a: _floor(_f16_extract(a).cast(dtypes.float32) * _const(dtypes.float32, 32767) + _const(dtypes.float32, 0.5)).cast(dtypes.int).cast(dtypes.int16),
+  'f16_to_unorm': lambda a: _floor(_f16_extract(a).cast(dtypes.float32) * _const(dtypes.float32, 65535) + _const(dtypes.float32, 0.5)).cast(dtypes.uint16),
+  'f32_to_snorm': lambda a: _floor(a.bitcast(dtypes.float32) * _const(dtypes.float32, 32767) + _const(dtypes.float32, 0.5)).cast(dtypes.int).cast(dtypes.int16),
+  'f32_to_unorm': lambda a: _floor(a.bitcast(dtypes.float32) * _const(dtypes.float32, 65535) + _const(dtypes.float32, 0.5)).cast(dtypes.uint16),
+  'f32_to_u8': lambda a: _f_to_u(a.bitcast(dtypes.float32), dtypes.uint8),
+  # Integer truncation conversions
+  'i32_to_i16': lambda a: a.cast(dtypes.int).cast(dtypes.int16),
+  'u32_to_u16': lambda a: a.cast(dtypes.uint32).cast(dtypes.uint16),
+  'u16_to_u32': lambda a: (a.cast(dtypes.uint32) & _u32(0xFFFF)),
+  'u8_to_u32': lambda a: (a.cast(dtypes.uint32) & _u32(0xFF)),
+  'u4_to_u32': lambda a: (a.cast(dtypes.uint32) & _u32(0xF)),
+  # Signed extraction with sign extension for dot products
+  'i16_to_i32': lambda a: _signext(a.cast(dtypes.uint32) & _u32(0xFFFF)),
+  'i8_to_i32': lambda a: _signext(a.cast(dtypes.uint32) & _u32(0xFF)),
+  'i4_to_i32': lambda a: _signext_4bit(a.cast(dtypes.uint32) & _u32(0xF)),
+  # Float to int16 conversions
+  'v_cvt_i16_f32': lambda a: UOp(Ops.TRUNC, dtypes.float32, (a.bitcast(dtypes.float32),)).cast(dtypes.int16),
+  'v_cvt_u16_f32': lambda a: _f_to_u(a.bitcast(dtypes.float32), dtypes.uint16),
+  # SAD (Sum of Absolute Differences) - sum |a_i - b_i| for 4 bytes + accumulator
+  'v_sad_u8': lambda a, b, c: _sad_u8(a, b, c),
+  'v_msad_u8': lambda a, b, c: _sad_u8(a, b, c, masked=True),
+  # System NOPs - these are scheduling hints, no effect on emulation
+  'MIN': lambda a, b: (a < b).where(a, b),
+  's_nop': lambda a: _u32(0),
+  # Address calculation for memory operations
+  'CalcDsAddr': lambda a, o, *r: a.cast(dtypes.uint32) + o.cast(dtypes.uint32),
+  'CalcGlobalAddr': lambda v, s, *r: v.cast(dtypes.uint64) + s.cast(dtypes.uint64),
 }
 for is_max, name in [(False, 'min'), (True, 'max')]:
  for dt, sfx in [(dtypes.float32, 'f32'), (dtypes.int, 'i32'), (dtypes.uint32, 'u32'), (dtypes.int16, 'i16'), (dtypes.uint16, 'u16')]:
    _FUNCS[f'v_{name}_{sfx}'] = lambda *a, im=is_max, d=dt: _minmax_reduce(im, d, *a)
    _FUNCS[f'v_{name}3_{sfx}'] = lambda *a, im=is_max, d=dt: _minmax_reduce(im, d, *a)
+# f16 min/max/min3/max3/med3
+for is_max, name in [(False, 'min'), (True, 'max')]:
+  _FUNCS[f'v_{name}_f16'] = lambda *a, im=is_max: _minmax_reduce(im, dtypes.half, *[_f16_extract(x) for x in a])
+  _FUNCS[f'v_{name}3_f16'] = lambda *a, im=is_max: _minmax_reduce(im, dtypes.half, *[_f16_extract(x) for x in a])
+  _FUNCS[f'v_{name}_num_f16'] = lambda *a, im=is_max: _minmax_reduce(im, dtypes.half, *[_f16_extract(x) for x in a])
+  _FUNCS[f'v_{name}_num_f32'] = lambda *a, im=is_max: _minmax_reduce(im, dtypes.float32, *a)
+  _FUNCS[f'v_{name}3_num_f16'] = lambda *a, im=is_max: _minmax_reduce(im, dtypes.half, *[_f16_extract(x) for x in a])
+  _FUNCS[f'v_{name}3_num_f32'] = lambda *a, im=is_max: _minmax_reduce(im, dtypes.float32, *a)
+  _FUNCS[f'v_{name}imum_f16'] = lambda *a, im=is_max: _minmax_reduce(im, dtypes.half, *[_f16_extract(x) for x in a])
+  _FUNCS[f'v_{name}imum_f32'] = lambda *a, im=is_max: _minmax_reduce(im, dtypes.float32, *a)
+  _FUNCS[f'v_{name}imum3_f16'] = lambda *a, im=is_max: _minmax_reduce(im, dtypes.half, *[_f16_extract(x) for x in a])
+  _FUNCS[f'v_{name}imum3_f32'] = lambda *a, im=is_max: _minmax_reduce(im, dtypes.float32, *a)

 # ═══════════════════════════════════════════════════════════════════════════════
 # TOKENIZER/PARSER
@@ -239,7 +298,7 @@ for is_max, name in [(False, 'min'), (True, 'max')]:

 DTYPES = {'u32': dtypes.uint32, 'i32': dtypes.int, 'f32': dtypes.float32, 'b32': dtypes.uint32, 'u64': dtypes.uint64, 'i64': dtypes.int64,
          'f64': dtypes.float64, 'b64': dtypes.uint64, 'u16': dtypes.uint16, 'i16': dtypes.short, 'f16': dtypes.half, 'b16': dtypes.uint16,
-          'u8': dtypes.uint8, 'i8': dtypes.int8, 'b8': dtypes.uint8, 'u1': dtypes.uint32}
+          'u8': dtypes.uint8, 'i8': dtypes.int8, 'b8': dtypes.uint8, 'u4': dtypes.uint8, 'i4': dtypes.int8, 'u1': dtypes.uint32}
 _BITS_DT = {8: dtypes.uint8, 16: dtypes.uint16, 32: dtypes.uint32, 64: dtypes.uint64}
 _NUM_SUFFIXES = ('ULL', 'LL', 'UL', 'U', 'L', 'F', 'f')
 def _strip_suffix(num: str) -> tuple[str, str]:
@@ -381,7 +440,7 @@ class Parser:
      self.eat('COMMA')
      lo = self.parse()
      self.eat('RBRACE')
-      return (hi.cast(dtypes.uint64) << _u64(32)) | lo.cast(dtypes.uint64)
+      return (hi.cast(dt:=_BITS_DT.get((s:=lo.dtype.bitsize) * 2, dtypes.uint64)) << _const(dt, s)) | lo.cast(dt)
    if self.at('NUM'):
      num = self.eat('NUM').val
      if self.try_eat('QUOTE'):
@@ -396,7 +455,7 @@ class Parser:
        self.eat('DOT')
        dt_name = self.eat('IDENT').val
        return self._handle_mem_load(addr, DTYPES.get(dt_name, dtypes.uint32))
-      if name == 'VGPR':
+      if name == 'VGPR' and self.at('LBRACKET'):
        self.eat('LBRACKET')
        lane = self.parse()
        self.eat('RBRACKET')
@@ -423,7 +482,21 @@ class Parser:
      if self.try_eat('LBRACE'):
        idx = self.eat('NUM').val
        self.eat('RBRACE')
-        elem = self.vars.get(f'{name}{idx}', _u32(0))
+        # Handle VGPR{lane}[reg] - 2D array access after loop unrolling
+        if name == 'VGPR' and self.at('LBRACKET'):
+          self.eat('LBRACKET')
+          reg = self.parse()
+          self.eat('RBRACKET')
+          vgpr = self.vars.get('_vgpr')
+          if vgpr is None: return _u32(0)
+          return vgpr.index(_to_u32(reg) * _u32(32) + _u32(int(idx)), ptr=True).load()
+        elem = self.vars.get(f'{name}@{idx}', self.vars.get(f'{name}{idx}'))
+        if elem is None:
+          # Extract bit idx from base variable (like var[idx])
+          base = self.vars.get(name)
+          assert isinstance(base, UOp), f"unknown variable: {name}{idx}"
+          dt = dtypes.uint64 if base.dtype in (dtypes.uint64, dtypes.int64) else dtypes.uint32
+          elem = (base.cast(dt) >> _const(dt, int(idx))) & _const(dt, 1)
        if self.try_eat('DOT'):
          dt_name = self.eat('IDENT').val
          return _cast_to(elem, DTYPES.get(dt_name, dtypes.uint32))
@@ -432,27 +505,17 @@ class Parser:
        return elem
      if self.at('LBRACKET') and name not in self.vars:
        self.eat('LBRACKET')
-        if self.at('NUM'):
-          idx_num = int(self.peek().val)
-          if f'{name}{idx_num}' in self.vars:
-            self.eat('NUM')
-            self.eat('RBRACKET')
-            elem = self.vars[f'{name}{idx_num}']
-            if self.try_eat('DOT'): return _cast_to(elem, DTYPES.get(self.eat('IDENT').val, dtypes.uint32))
-            return elem
        first = self.parse()
        return self._handle_bracket_rest(first, _u32(0), name)
      if name in self.vars:
        v = self.vars[name]
-        return v if isinstance(v, UOp) else _u32(0) if isinstance(v, dict) else _u32(0)
+        assert isinstance(v, UOp), f"expected UOp for {name}, got {type(v)}"
+        return v
      raise RuntimeError(f"unknown variable: {name}")
    raise RuntimeError(f"unexpected token in primary: {self.peek()}")

  def _handle_dot(self, base, field: str) -> UOp:
-    if isinstance(base, str): return _u32(0)
-    if not isinstance(base, UOp):
-      if isinstance(base, dict): return base.get(field, _u32(0))
-      return _u32(0)
+    assert isinstance(base, UOp), f"expected UOp for dot access, got {type(base)}"
    if field == 'u64' and self.at('LBRACKET') and self.peek(1).type == 'IDENT' and self.peek(1).val == 'laneId':
      self.eat('LBRACKET')
      self.eat_val('laneId', 'IDENT')
@@ -467,6 +530,7 @@ class Parser:
    if dt == base.dtype: return base
    if dt.itemsize == 2 and base.dtype.itemsize == 4:
      return (base & _const(base.dtype, 0xFFFF)).cast(dtypes.uint16) if dt == dtypes.uint16 else (base & _const(base.dtype, 0xFFFF)).cast(dtypes.uint16).bitcast(dt)
+    if field == 'i4': return _signext_4bit(base)
    return _cast_to(base, dt)

  def _handle_bracket(self, base, var_name: str | None = None) -> UOp:
@@ -509,16 +573,18 @@ class Parser:
      var_name = self._find_var_name(base)
    if first.op == Ops.CONST:
      idx = int(first.arg)
-      if var_name and f'{var_name}{idx}' in self.vars:
-        v = self.vars[f'{var_name}{idx}']
+      # Check for array element (var@idx)
+      if var_name and f'{var_name}@{idx}' in self.vars:
+        v = self.vars[f'{var_name}@{idx}']
        return _cast_to(v, dt_suffix) if dt_suffix else v
+      # Bit extraction
      dt = dtypes.uint64 if base.dtype in (dtypes.uint64, dtypes.int64) else dtypes.uint32
      base_cast = base.cast(dt) if base.dtype != dt else base
      result = ((base_cast >> _const(dt, idx)) & _const(dt, 1))
      return _cast_to(result, dt_suffix) if dt_suffix else result
    if var_name:
      idx_u32 = _to_u32(first)
-      elems = [(i, self.vars[f'{var_name}{i}']) for i in range(256) if f'{var_name}{i}' in self.vars]
+      elems = [(i, self.vars[f'{var_name}@{i}']) for i in range(256) if f'{var_name}@{i}' in self.vars]
      if elems:
        result = elems[-1][1]
        for ei, ev in reversed(elems[:-1]):
@@ -537,7 +603,7 @@ class Parser:
    self.eat('RBRACE')
    var_name = self._find_var_name(base)
    if var_name:
-      elem = self.vars.get(f'{var_name}{idx}', _u32(0))
+      elem = self.vars.get(f'{var_name}@{idx}', _u32(0))  # use @ to avoid collision with temps like A4
      if self.try_eat('DOT'):
        dt_name = self.eat('IDENT').val
        return _cast_to(elem, DTYPES.get(dt_name, dtypes.uint32))
@@ -599,13 +665,14 @@ class Parser:
    raise RuntimeError(f"unexpected token after {bits}': {self.peek()}")

  def _parse_number(self, num: str) -> UOp:
-    if num.startswith('0x') or num.startswith('0X'): return _const(dtypes.uint64, int(num.rstrip('ULul'), 16))
-    suffix, num = _strip_suffix(num)
-    if '.' in num or suffix in ('F', 'f'):
-      return _const(dtypes.float32 if suffix in ('F', 'f') else dtypes.float64, float(num))
-    val = int(num)
-    if 'ULL' in suffix: return _const(dtypes.uint64, val)
-    if 'LL' in suffix or 'L' in suffix: return _const(dtypes.uint64, val)
+    if num.startswith('0x') or num.startswith('0X'):
+      is_u64 = num.upper().endswith('ULL') or num.upper().endswith('LL') or num.upper().endswith('UL')
+      return _const(dtypes.uint64 if is_u64 else dtypes.uint32, int(num.rstrip('ULul'), 16))
+    suffix, num_str = _strip_suffix(num)
+    if '.' in num_str or suffix in ('F', 'f'):
+      return _const(dtypes.float32 if suffix in ('F', 'f') else dtypes.float64, float(num_str))
+    val = int(num_str)
+    if 'ULL' in suffix or 'LL' in suffix or 'L' in suffix: return _const(dtypes.uint64, val)
    if 'U' in suffix: return _const(dtypes.uint32, val)
    return _const(dtypes.int if val < 0 else dtypes.uint32, val)

@@ -623,7 +690,8 @@ class Parser:
      if ';' in body or '\n' in body or 'return' in body.lower():
        lines = [l.strip() for l in body.replace(';', '\n').split('\n') if l.strip() and not l.strip().startswith('//')]
        _, _, result = parse_block(lines, 0, lv, self.funcs)
-        return result if result is not None else _u32(0)
+        assert result is not None, f"lambda {name} must return a value"
+        return result
      return parse_expr(body, lv, self.funcs)
    if name in self.funcs:
      return self.funcs[name](*args)
@@ -631,7 +699,7 @@ class Parser:

  def _handle_mem_load(self, addr: UOp, dt) -> UOp:
    mem = self.vars.get('_vmem') if '_vmem' in self.vars else self.vars.get('_lds')
-    if mem is None: return _const(dt, 0)
+    assert mem is not None, "memory load requires _vmem or _lds"
    adt = dtypes.uint64 if addr.dtype == dtypes.uint64 else dtypes.uint32
    active = self.vars.get('_active')
    gate = (active,) if active is not None else ()
@@ -693,29 +761,9 @@ def parse_tokens(toks: list[Token], vars: dict[str, VarVal], funcs: dict | None

 # Unified block parser for pcode
 def _subst_loop_var(line: str, loop_var: str, val: int) -> str:
-  """Substitute loop variable and evaluate bracket expressions.
-  Converts var[loop_var] to var{val} for array element access (like the old regex parser)."""
+  """Substitute loop variable with its value."""
  toks = tokenize(line)
-  # First pass: convert var[loop_var] to var{loop_var} to mark for array element assignment
-  result_toks: list[Token] = []
-  j = 0
-  while j < len(toks):
-    t = toks[j]
-    # Check for pattern: IDENT[loop_var] where it's not preceded by a dot (not .type[...])
-    if t.type == 'IDENT' and j+3 < len(toks) and toks[j+1].type == 'LBRACKET' and toks[j+2].type == 'IDENT' and toks[j+2].val == loop_var and toks[j+3].type == 'RBRACKET':
-      # Check that it's not .type[loop_var]
-      if not result_toks or result_toks[-1].type != 'DOT':
-        result_toks.append(t)
-        result_toks.append(Token('LBRACE', '{'))
-        result_toks.append(Token('NUM', str(val)))
-        result_toks.append(Token('RBRACE', '}'))
-        j += 4
-        continue
-    result_toks.append(t)
-    j += 1
-  # Second pass: substitute loop variable in remaining positions
-  subst_parts = [str(val) if t.type == 'IDENT' and t.val == loop_var else t.val for t in result_toks if t.type != 'EOF']
-  return ' '.join(subst_parts)
+  return ' '.join(str(val) if t.type == 'IDENT' and t.val == loop_var else t.val for t in toks if t.type != 'EOF')

 def _set_bits(old: UOp, val: UOp, width: int, offset: int) -> UOp:
  """Set bits [offset:offset+width) in old to val, masking and shifting appropriately."""
@@ -765,8 +813,9 @@ def parse_block(lines: list[str], start: int, vars: dict[str, VarVal], funcs: di
      def parse_bound():
        if p.at('NUM') and p.peek(1).type == 'QUOTE': p.eat('NUM'); p.eat('QUOTE')
        if p.at('NUM'): return int(p.eat('NUM').val.rstrip('UuLl'))
-        expr = p.parse()
-        return int(expr.arg) if expr.op == Ops.CONST else 0
+        expr = p.parse().simplify()
+        assert expr.op == Ops.CONST, f"loop bound must be constant, got {expr}"
+        return int(expr.arg)
      start_val = parse_bound()
      p.eat('COLON')
      end_val = parse_bound()
@@ -787,7 +836,7 @@ def parse_block(lines: list[str], start: int, vars: dict[str, VarVal], funcs: di
      if found_var: vars[found_var] = block_assigns[found_var] = _const(dtypes.bool, False)
      for loop_i in range(start_val, end_val + 1):
        subst_lines = [_subst_loop_var(bl, loop_var, loop_i) for bl in body_lines if not (has_break and bl.strip().lower() == 'break')]
-        _, iter_assigns, _ = parse_block(subst_lines, 0, vars, funcs, assigns)
+        _, iter_assigns, _ = parse_block(subst_lines, 0, {**vars, **block_assigns}, funcs, assigns)
        if has_break:
          assert found_var is not None
          found = block_assigns.get(found_var, vars.get(found_var))
@@ -812,7 +861,9 @@ def parse_block(lines: list[str], start: int, vars: dict[str, VarVal], funcs: di

    # declare
    if first == 'declare':
-      if '[' not in line and len(toks) >= 2 and toks[1].type == 'IDENT': vars[toks[1].val] = _u32(0)
+      # Initialize scalar declarations (skip arrays and vars already passed as srcs)
+      if '[' not in line and len(toks) >= 2 and toks[1].type == 'IDENT':
+        vars.setdefault(toks[1].val, _u32(0))
      i += 1; continue

    # lambda definition
@@ -870,6 +921,7 @@ def parse_block(lines: list[str], start: int, vars: dict[str, VarVal], funcs: di
      j, lane_toks = _match_bracket(toks, 1)
      if j < len(toks) and toks[j].type == 'LBRACKET':
        j, reg_toks = _match_bracket(toks, j)
+        if j < len(toks) and toks[j].type == 'DOT': j += 2  # skip .type suffix
        if j < len(toks) and toks[j].type == 'EQUALS': j += 1
        ln, rg, val = parse_tokens(lane_toks, vars, funcs), parse_tokens(reg_toks, vars, funcs), parse_tokens(toks[j:], vars, funcs)
        if assigns is not None: assigns.append((f'VGPR[{_tok_str(lane_toks)}][{_tok_str(reg_toks)}]', (_to_u32(rg) * _u32(32) + _to_u32(ln), val)))
@@ -933,19 +985,32 @@ def parse_block(lines: list[str], start: int, vars: dict[str, VarVal], funcs: di
            block_assigns[var] = vars[var] = _set_bit(existing, _to_u32(parse_tokens(bit_toks, vars, funcs)), parse_tokens(toks[j+1:], vars, funcs))
            i += 1; continue

-    # Array element: var{idx} = value
-    if len(toks) >= 5 and toks[0].type == 'IDENT' and toks[1].type == 'LBRACE' and toks[2].type == 'NUM':
-      var, idx = toks[0].val, int(toks[2].val)
-      j = 4
-      while j < len(toks) and toks[j].type != 'EQUALS': j += 1
-      if j < len(toks):
-        val = parse_tokens(toks[j+1:], vars, funcs)
-        existing = block_assigns.get(var, vars.get(var))
-        if existing is not None and isinstance(existing, UOp):
-          block_assigns[var] = vars[var] = _set_bit(existing, _u32(idx), val)
-        else:
-          block_assigns[f'{var}{idx}'] = vars[f'{var}{idx}'] = val
-        i += 1; continue
+    # Array element: var[idx] = value (static index) or var[expr] = value (dynamic)
+    if len(toks) >= 4 and toks[0].type == 'IDENT' and toks[1].type == 'LBRACKET':
+      var = toks[0].val
+      j, idx_toks = _match_bracket(toks, 1)
+      if j < len(toks) and toks[j].type == 'EQUALS':
+        # Static index: var[NUM] = value
+        if len(idx_toks) == 1 and idx_toks[0].type == 'NUM':
+          idx = int(idx_toks[0].val.rstrip('UuLl'))
+          val = parse_tokens(toks[j+1:], vars, funcs)
+          existing = block_assigns.get(var, vars.get(var))
+          if existing is not None and isinstance(existing, UOp):
+            block_assigns[var] = vars[var] = _set_bit(existing, _u32(idx), val)
+          else:
+            block_assigns[f'{var}@{idx}'] = vars[f'{var}@{idx}'] = val
+          i += 1; continue
+        # Dynamic index: var[expr] = value where var has @-elements
+        elems = [(k.split('@')[1], v) for k, v in {**vars, **block_assigns}.items() if k.startswith(f'{var}@') and isinstance(v, UOp)]
+        if elems:
+          idx_expr = parse_tokens(idx_toks, vars, funcs)
+          val = parse_tokens(toks[j+1:], vars, funcs)
+          for elem_idx_str, old_elem in elems:
+            elem_idx = int(elem_idx_str)
+            cond = _to_u32(idx_expr).eq(_u32(elem_idx))
+            new_val = cond.where(val.cast(old_elem.dtype) if val.dtype != old_elem.dtype else val, old_elem)
+            block_assigns[f'{var}@{elem_idx}'] = vars[f'{var}@{elem_idx}'] = new_val
+          i += 1; continue

    # Compound assignment: var += or var -=
    assign_op = next((j for j, t in enumerate(toks) if t.type == 'ASSIGN_OP'), None)
@@ -992,13 +1057,14 @@ def parse_block(lines: list[str], start: int, vars: dict[str, VarVal], funcs: di
      def parse_cond(s, kw):
        ll = s.lower()
        return _to_bool(parse_expr(s[ll.find(kw) + len(kw):ll.rfind('then')].strip(), vars, funcs))
-      def not_static_false(c): return c.op != Ops.CONST or c.arg is not False
+      def is_const(c, v): return c.op == Ops.CONST and c.arg is v
      cond = parse_cond(line, 'if')
-      conditions: list[tuple[UOp, UOp | dict[str, VarVal] | None]] = [(cond, None)] if not_static_false(cond) else []
+      conditions: list[tuple[UOp, UOp | dict[str, VarVal] | None]] = [(cond, None)] if not is_const(cond, False) else []
      else_branch: tuple[UOp | None, dict[str, VarVal]] = (None, {})
      vars_snap = dict(vars)
+      static_true = is_const(cond, True)  # track if any condition is statically true
      i += 1
-      i, branch, ret = parse_block(lines, i, vars, funcs, assigns)
+      i, branch, ret = parse_block(lines, i, vars, funcs, assigns if not is_const(cond, False) else None)
      if conditions: conditions[0] = (cond, ret if ret is not None else branch)
      vars.clear(); vars.update(vars_snap)
      while i < len(lines):
@@ -1007,12 +1073,16 @@ def parse_block(lines: list[str], start: int, vars: dict[str, VarVal], funcs: di
        lf = ltoks[0].val.lower()
        if lf == 'elsif':
          c = parse_cond(lines[i], 'elsif')
-          i += 1; i, branch, ret = parse_block(lines, i, vars, funcs, assigns)
-          if not_static_false(c): conditions.append((c, ret if ret is not None else branch))
+          take = not static_true and not is_const(c, False)
+          i += 1; i, branch, ret = parse_block(lines, i, vars, funcs, assigns if take else None)
+          if take:
+            conditions.append((c, ret if ret is not None else branch))
+            if is_const(c, True): static_true = True
          vars.clear(); vars.update(vars_snap)
        elif lf == 'else':
-          i += 1; i, branch, ret = parse_block(lines, i, vars, funcs, assigns)
-          else_branch = (ret, branch)
+          i += 1
+          i, branch, ret = parse_block(lines, i, vars, funcs, assigns if not static_true else None)
+          if not static_true: else_branch = (ret, branch)
          vars.clear(); vars.update(vars_snap)
        elif lf == 'endif': i += 1; break
        else: break
@@ -1024,17 +1094,21 @@ def parse_block(lines: list[str], start: int, vars: dict[str, VarVal], funcs: di
            if rv.dtype != result.dtype and rv.dtype.itemsize == result.dtype.itemsize: result = result.cast(rv.dtype)
            result = c.where(rv, result)
        return i, block_assigns, result
-      # Main style: merge variable assignments with WHERE
-      else_assigns = else_branch[1]
-      all_vars = set().union(*[ba.keys() for _, ba in conditions if isinstance(ba, dict)], else_assigns.keys())
-      for var in all_vars:
-        res: Any = else_assigns.get(var, block_assigns.get(var, vars.get(var, _u32(0))))
-        for cond, ba in reversed(conditions):
-          if isinstance(ba, dict) and var in ba:
-            tv = ba[var]
-            if isinstance(tv, UOp) and isinstance(res, UOp):
-              res = cond.where(tv, res.cast(tv.dtype) if tv.dtype != res.dtype and tv.dtype.itemsize == res.dtype.itemsize else res)
-        block_assigns[var] = vars[var] = res
+      # If statically true, use that branch directly; otherwise merge with WHERE
+      if static_true:
+        ba = next((b for c, b in conditions if is_const(c, True) and isinstance(b, dict)), {})
+        block_assigns.update(ba); vars.update(ba)
+      else:
+        else_assigns = else_branch[1]
+        all_vars = set().union(*[ba.keys() for _, ba in conditions if isinstance(ba, dict)], else_assigns.keys())
+        for var in all_vars:
+          res: Any = else_assigns.get(var, block_assigns.get(var, vars.get(var, _u32(0))))
+          for cond, ba in reversed(conditions):
+            if isinstance(ba, dict) and var in ba:
+              tv = ba[var]
+              if isinstance(tv, UOp) and isinstance(res, UOp):
+                res = cond.where(tv, res.cast(tv.dtype) if tv.dtype != res.dtype and tv.dtype.itemsize == res.dtype.itemsize else res)
+          block_assigns[var] = vars[var] = res
      continue

    # Regular assignment: var = value
@@ -2,11 +2,8 @@
 from dataclasses import dataclass
 from typing import Iterator

-from tinygrad.runtime.support.elf import elf_loader
-
 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 import decode_inst
 from extra.assembly.amd.autogen.rdna3.ins import SOPP, s_endpgm
 from extra.assembly.amd.autogen.rdna3.enum import SOPPOp

@@ -16,19 +13,11 @@ class InstructionInfo:
  wave: int
  inst: Inst

-def map_insts(data:bytes, lib:bytes) -> Iterator[tuple[PacketType, InstructionInfo|None]]:
+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
-  pc_map:dict[int, Inst] = {}
-  image, sections, _ = elf_loader(lib)
-  text = next((sh for sh in sections if sh.name == ".text"), None)
-  assert text is not None, "no .text section found"
-  text_off, text_size = text.header.sh_addr, text.header.sh_size
-  offset = text_off
-  while offset < text_off + text_size:
-    inst = decode_inst(image[offset:])
-    pc_map[offset-text_off] = inst
-    offset += inst.size()
+  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
@@ -37,7 +26,7 @@ def map_insts(data:bytes, lib:bytes) -> Iterator[tuple[PacketType, InstructionIn
    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] = 0
+      wave_pc[p.wave] = next(iter(pc_map))
      continue
    if isinstance(p, WAVEEND):
      pc = wave_pc.pop(p.wave)
@@ -80,22 +69,22 @@ def map_insts(data:bytes, lib:bytes) -> Iterator[tuple[PacketType, InstructionIn
 # test to compare every packet with the rocprof decoder

 def test_rocprof_inst_traces_match(sqtt, prg, target):
-  from tinygrad.viz.serve import llvm_disasm
+  from tinygrad.viz.serve import amd_decode
  from extra.sqtt.roc import decode as roc_decode, InstExec
-  disasm = {addr+prg.base:inst_disasm for addr, inst_disasm in llvm_disasm(target, prg.lib).items()}
-  rctx = roc_decode([sqtt], {prg.name:disasm})
-  rwaves = rctx.inst_execs[(sqtt.kern, sqtt.exec_tag)]
+  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())
-  rwaves_base = next(iter(disasm)) # base program counter

  passed_insts = 0
-  for pkt, info in map_insts(sqtt.blob, prg.lib):
+  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-rwaves_base
+    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.
@@ -110,7 +99,8 @@ def test_rocprof_inst_traces_match(sqtt, prg, target):
  for k,v in rwaves_iter.items():
    assert len(v) == 0, f"incomplete wave {k}"

-  print(f"passed for {passed_insts} instructions across {len(rwaves)} waves scheduled on {len(rwaves_iter)} wave units")
+  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
@@ -123,7 +113,7 @@ if __name__ == "__main__":
  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.name:e for e in data if type(e).__name__ == "ProfileProgramEvent"}
+  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
@@ -13,7 +13,7 @@ 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 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
@@ -43,6 +43,23 @@ 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
@@ -212,8 +229,12 @@ amdhsa.kernels:

  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) -> list[str]:
-  """Compare two WaveStates and return list of differences."""
+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):
@@ -224,6 +245,11 @@ def compare_wave_states(emu_st: WaveState, hw_st: WaveState, n_lanes: int, n_vgp
        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]
@@ -236,16 +262,19 @@ def compare_wave_states(emu_st: WaveState, hw_st: WaveState, n_lanes: int, n_vgp
    diffs.append(f"scc: emu={emu_st.scc} hw={hw_st.scc}")
  return diffs

-def run_program(instructions: list, n_lanes: int = 1) -> WaveState:
+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)
+    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
@@ -719,5 +719,47 @@ class TestAtomicOrdering(unittest.TestCase):
    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()
@@ -62,6 +62,7 @@ class TestBasicScalar(unittest.TestCase):
    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
@@ -73,6 +74,7 @@ class TestBasicScalar(unittest.TestCase):
    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
@@ -85,6 +87,50 @@ class TestBasicScalar(unittest.TestCase):
    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."""

@@ -255,7 +255,6 @@ class TestF16Conversions(unittest.TestCase):

  def test_v_cvt_f16_f32_small(self):
    """V_CVT_F16_F32 converts small f32 value."""
-    from extra.assembly.amd.test.hw.helpers import f32_to_f16
    instructions = [
      v_mov_b32_e32(v[0], 0.5),
      v_cvt_f16_f32_e32(v[1], v[0]),
@@ -293,7 +292,6 @@ class TestF16Conversions(unittest.TestCase):

  def test_v_cvt_f16_f32_reads_full_32bit_source(self):
    """V_CVT_F16_F32 must read full 32-bit f32 source."""
-    from extra.assembly.amd.test.hw.helpers import _f16
    instructions = [
      s_mov_b32(s[0], 0x3fc00000),  # f32 1.5
      v_mov_b32_e32(v[0], s[0]),
@@ -302,7 +300,7 @@ class TestF16Conversions(unittest.TestCase):
    st = run_program(instructions, n_lanes=1)
    result = st.vgpr[0][1]
    lo_bits = result & 0xffff
-    self.assertEqual(lo_bits, 0x3e00, f"Expected f16(1.5)=0x3e00, got 0x{lo_bits:04x} ({_f16(lo_bits)})")
+    self.assertEqual(lo_bits, 0x3e00, f"Expected f16(1.5)=0x3e00, got 0x{lo_bits:04x} ({f16(lo_bits)})")

  def test_v_cvt_i16_f16_zero(self):
    """V_CVT_I16_F16 converts f16 zero to i16 zero."""
@@ -696,7 +694,6 @@ class TestCvtF16Modifiers(unittest.TestCase):

  def test_v_cvt_f32_f16_abs_negative(self):
    """V_CVT_F32_F16 with |abs| on negative value."""
-    from extra.assembly.amd.test.hw.helpers import f32_to_f16
    f16_neg1 = f32_to_f16(-1.0)  # 0xbc00
    instructions = [
      s_mov_b32(s[0], f16_neg1),
@@ -709,7 +706,6 @@ class TestCvtF16Modifiers(unittest.TestCase):

  def test_v_cvt_f32_f16_abs_positive(self):
    """V_CVT_F32_F16 with |abs| on positive value (should stay positive)."""
-    from extra.assembly.amd.test.hw.helpers import f32_to_f16
    f16_2 = f32_to_f16(2.0)  # 0x4000
    instructions = [
      s_mov_b32(s[0], f16_2),
@@ -722,7 +718,6 @@ class TestCvtF16Modifiers(unittest.TestCase):

  def test_v_cvt_f32_f16_neg_positive(self):
    """V_CVT_F32_F16 with neg on positive value."""
-    from extra.assembly.amd.test.hw.helpers import f32_to_f16
    f16_2 = f32_to_f16(2.0)  # 0x4000
    instructions = [
      s_mov_b32(s[0], f16_2),
@@ -735,7 +730,6 @@ class TestCvtF16Modifiers(unittest.TestCase):

  def test_v_cvt_f32_f16_neg_negative(self):
    """V_CVT_F32_F16 with neg on negative value (double negative)."""
-    from extra.assembly.amd.test.hw.helpers import f32_to_f16
    f16_neg2 = f32_to_f16(-2.0)  # 0xc000
    instructions = [
      s_mov_b32(s[0], f16_neg2),
@@ -748,7 +742,6 @@ class TestCvtF16Modifiers(unittest.TestCase):

  def test_v_cvt_f16_f32_then_pack_for_wmma(self):
    """CVT F32->F16 followed by pack (common WMMA pattern)."""
-    from extra.assembly.amd.test.hw.helpers import _f16
    f32_val = 3.5
    instructions = [
      s_mov_b32(s[0], f2i(f32_val)),
@@ -757,8 +750,8 @@ class TestCvtF16Modifiers(unittest.TestCase):
      v_pack_b32_f16(v[2], v[1], v[1]),  # Pack same value
    ]
    st = run_program(instructions, n_lanes=1)
-    lo = _f16(st.vgpr[0][2] & 0xffff)
-    hi = _f16((st.vgpr[0][2] >> 16) & 0xffff)
+    lo = f16(st.vgpr[0][2] & 0xffff)
+    hi = f16((st.vgpr[0][2] >> 16) & 0xffff)
    self.assertAlmostEqual(lo, f32_val, places=1)
    self.assertAlmostEqual(hi, f32_val, places=1)

@@ -804,7 +797,6 @@ class TestConversionRounding(unittest.TestCase):

  def test_f16_to_f32_precision(self):
    """F16 to F32 conversion precision."""
-    from extra.assembly.amd.test.hw.helpers import f32_to_f16
    f16_val = f32_to_f16(1.5)
    instructions = [
      s_mov_b32(s[0], f16_val),
@@ -816,7 +808,6 @@ class TestConversionRounding(unittest.TestCase):

  def test_f16_denormal_to_f32(self):
    """F16 denormal converts to small positive f32."""
-    from extra.assembly.amd.test.hw.helpers import _f16
    f16_denorm = 0x0001  # Smallest positive f16 denormal
    instructions = [
      v_mov_b32_e32(v[0], f16_denorm),
@@ -1512,5 +1503,82 @@ class TestReciprocalF16(unittest.TestCase):
    self.assertAlmostEqual(result, 0.25, places=2, msg="1/4.0 should be 0.25")


+class TestCvtNormF16(unittest.TestCase):
+  """Tests for V_CVT_NORM_I16_F16 and V_CVT_NORM_U16_F16."""
+
+  def test_cvt_norm_i16_f16_positive(self):
+    """V_CVT_NORM_I16_F16: f16 1.0 -> i16 max (32767)."""
+    instructions = [
+      s_mov_b32(s[0], f32_to_f16(1.0)),
+      v_mov_b32_e32(v[0], s[0]),
+      v_cvt_norm_i16_f16_e32(v[1], v[0]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][1] & 0xffff
+    self.assertEqual(result, 32767)
+
+  def test_cvt_norm_i16_f16_negative(self):
+    """V_CVT_NORM_I16_F16: f16 -1.0 -> i16 -32767 (0x8001)."""
+    instructions = [
+      s_mov_b32(s[0], f32_to_f16(-1.0)),
+      v_mov_b32_e32(v[0], s[0]),
+      v_cvt_norm_i16_f16_e32(v[1], v[0]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][1] & 0xffff
+    self.assertEqual(result, 0x8001)  # -32767, hardware uses symmetric range
+
+  def test_cvt_norm_i16_f16_zero(self):
+    """V_CVT_NORM_I16_F16: f16 0.0 -> i16 0."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0),
+      v_cvt_norm_i16_f16_e32(v[1], v[0]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][1] & 0xffff
+    self.assertEqual(result, 0)
+
+  def test_cvt_norm_u16_f16_one(self):
+    """V_CVT_NORM_U16_F16: f16 1.0 -> u16 max (65535)."""
+    instructions = [
+      s_mov_b32(s[0], f32_to_f16(1.0)),
+      v_mov_b32_e32(v[0], s[0]),
+      v_cvt_norm_u16_f16_e32(v[1], v[0]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][1] & 0xffff
+    self.assertEqual(result, 65535)
+
+  def test_cvt_norm_u16_f16_half(self):
+    """V_CVT_NORM_U16_F16: f16 0.5 -> u16 ~32768."""
+    instructions = [
+      s_mov_b32(s[0], f32_to_f16(0.5)),
+      v_mov_b32_e32(v[0], s[0]),
+      v_cvt_norm_u16_f16_e32(v[1], v[0]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][1] & 0xffff
+    self.assertAlmostEqual(result, 32768, delta=1)
+
+
+class TestPermlane64(unittest.TestCase):
+  """Tests for V_PERMLANE64_B32 instruction (wave64 cross-half swap)."""
+
+  def test_v_permlane64_b32_is_nop_in_wave32(self):
+    """V_PERMLANE64_B32 is a NOP in wave32 mode.
+
+    Per AMD pcode: "if WAVE32 then s_nop(...) else ... endif"
+    The emulator runs in wave32 mode, so this instruction should not modify registers.
+    """
+    instructions = [
+      v_mov_b32_e32(v[0], 0xCAFEBABE),  # source
+      v_mov_b32_e32(v[1], 0x12345678),  # dest (should be preserved)
+      v_permlane64_b32_e32(v[1], v[0]),  # NOP in wave32
+    ]
+    st = run_program(instructions, n_lanes=1)
+    # Dest register should be unchanged (NOP behavior in wave32)
+    self.assertEqual(st.vgpr[0][1], 0x12345678)
+
+
 if __name__ == '__main__':
  unittest.main()
@@ -857,7 +857,6 @@ class TestF16Modifiers(unittest.TestCase):

  def test_v_fma_f16_inline_const_1_0(self):
    """V_FMA_F16: a*b + 1.0 should use f16 inline constant."""
-    from extra.assembly.amd.test.hw.helpers import f32_to_f16, _f16
    f16_a = f32_to_f16(0.325928)  # ~0x3537
    f16_b = f32_to_f16(-0.486572)  # ~0xb7c9
    instructions = [
@@ -868,13 +867,12 @@ class TestF16Modifiers(unittest.TestCase):
      v_fma_f16(v[4], v[4], v[6], 1.0),  # 1.0 is inline constant
    ]
    st = run_program(instructions, n_lanes=1)
-    result = _f16(st.vgpr[0][4] & 0xffff)
+    result = f16(st.vgpr[0][4] & 0xffff)
    expected = 0.325928 * (-0.486572) + 1.0
    self.assertAlmostEqual(result, expected, delta=0.01)

  def test_v_fma_f16_inline_const_0_5(self):
    """V_FMA_F16: a*b + 0.5 should use f16 inline constant."""
-    from extra.assembly.amd.test.hw.helpers import f32_to_f16, _f16
    f16_a = f32_to_f16(2.0)
    f16_b = f32_to_f16(3.0)
    instructions = [
@@ -885,13 +883,12 @@ class TestF16Modifiers(unittest.TestCase):
      v_fma_f16(v[2], v[0], v[1], 0.5),  # 0.5 is inline constant
    ]
    st = run_program(instructions, n_lanes=1)
-    result = _f16(st.vgpr[0][2] & 0xffff)
+    result = f16(st.vgpr[0][2] & 0xffff)
    expected = 2.0 * 3.0 + 0.5
    self.assertAlmostEqual(result, expected, delta=0.01)

  def test_v_fma_f16_inline_const_neg_1_0(self):
    """V_FMA_F16: a*b + (-1.0) should use f16 inline constant."""
-    from extra.assembly.amd.test.hw.helpers import f32_to_f16, _f16
    f16_a = f32_to_f16(2.0)
    f16_b = f32_to_f16(3.0)
    instructions = [
@@ -902,13 +899,12 @@ class TestF16Modifiers(unittest.TestCase):
      v_fma_f16(v[2], v[0], v[1], -1.0),  # -1.0 is inline constant
    ]
    st = run_program(instructions, n_lanes=1)
-    result = _f16(st.vgpr[0][2] & 0xffff)
+    result = f16(st.vgpr[0][2] & 0xffff)
    expected = 2.0 * 3.0 + (-1.0)
    self.assertAlmostEqual(result, expected, delta=0.01)

  def test_v_add_f16_abs_both(self):
    """V_ADD_F16 with abs on both operands."""
-    from extra.assembly.amd.test.hw.helpers import f32_to_f16, _f16
    f16_neg2 = f32_to_f16(-2.0)
    f16_neg3 = f32_to_f16(-3.0)
    instructions = [
@@ -919,12 +915,11 @@ class TestF16Modifiers(unittest.TestCase):
      v_add_f16_e64(v[2], abs(v[0]), abs(v[1])),  # |-2| + |-3| = 5
    ]
    st = run_program(instructions, n_lanes=1)
-    result = _f16(st.vgpr[0][2] & 0xffff)
+    result = f16(st.vgpr[0][2] & 0xffff)
    self.assertAlmostEqual(result, 5.0, delta=0.01)

  def test_v_mul_f16_neg_abs(self):
    """V_MUL_F16 with neg on one operand and abs on another."""
-    from extra.assembly.amd.test.hw.helpers import f32_to_f16, _f16
    f16_2 = f32_to_f16(2.0)
    f16_neg3 = f32_to_f16(-3.0)
    instructions = [
@@ -935,7 +930,7 @@ class TestF16Modifiers(unittest.TestCase):
      v_mul_f16_e64(v[2], -v[0], abs(v[1])),  # -(2) * |-3| = -6
    ]
    st = run_program(instructions, n_lanes=1)
-    result = _f16(st.vgpr[0][2] & 0xffff)
+    result = f16(st.vgpr[0][2] & 0xffff)
    self.assertAlmostEqual(result, -6.0, delta=0.01)

  def test_v_fmac_f16_hi_dest(self):
@@ -943,7 +938,6 @@ class TestF16Modifiers(unittest.TestCase):

    This tests the case from AMD_LLVM sin(0) where V_FMAC_F16 writes to v0.h.
    """
-    from extra.assembly.amd.test.hw.helpers import _f16
    instructions = [
      s_mov_b32(s[0], 0x38003c00),  # v0 = {hi=0.5, lo=1.0}
      v_mov_b32_e32(v[0], s[0]),
@@ -954,8 +948,8 @@ class TestF16Modifiers(unittest.TestCase):
    ]
    st = run_program(instructions, n_lanes=1)
    v0 = st.vgpr[0][0]
-    result_hi = _f16((v0 >> 16) & 0xffff)
-    result_lo = _f16(v0 & 0xffff)
+    result_hi = f16((v0 >> 16) & 0xffff)
+    result_lo = f16(v0 & 0xffff)
    self.assertAlmostEqual(result_hi, 0.5, delta=0.01, msg=f"Expected hi=0.5, got {result_hi}")
    self.assertAlmostEqual(result_lo, 1.0, delta=0.01, msg=f"Expected lo=1.0, got {result_lo}")

@@ -2955,5 +2949,694 @@ class TestVOP3Clamp(unittest.TestCase):
    self.assertAlmostEqual(i2f(st.vgpr[3][1]), 1.0, places=5, msg="lane 3: 2.5 should clamp to 1.0")


+class TestVOP3ClampUint32(unittest.TestCase):
+  """Tests for VOP3 clamp modifier on unsigned 32-bit integer operations."""
+
+  def test_v_sub_nc_u32_e64_clamp_underflow(self):
+    """V_SUB_NC_U32_E64 with clamp: 0 - 1 should saturate to 0."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0),
+      v_mov_b32_e32(v[1], 1),
+      v_sub_nc_u32_e64(v[2], v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0, f"expected 0, got 0x{st.vgpr[0][2]:08x}")
+
+  def test_v_sub_nc_u32_e64_clamp_no_underflow(self):
+    """V_SUB_NC_U32_E64 with clamp: 100 - 50 = 50 (no saturation needed)."""
+    instructions = [
+      v_mov_b32_e32(v[0], 100),
+      v_mov_b32_e32(v[1], 50),
+      v_sub_nc_u32_e64(v[2], v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 50, f"expected 50, got {st.vgpr[0][2]}")
+
+  def test_v_add_nc_u32_e64_clamp_overflow(self):
+    """V_ADD_NC_U32_E64 with clamp: 0xFFFFFFFF + 1 should saturate to 0xFFFFFFFF."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0xFFFFFFFF),
+      v_mov_b32_e32(v[1], 1),
+      v_add_nc_u32_e64(v[2], v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0xFFFFFFFF, f"expected 0xFFFFFFFF, got 0x{st.vgpr[0][2]:08x}")
+
+  def test_v_add_nc_u32_e64_clamp_no_overflow(self):
+    """V_ADD_NC_U32_E64 with clamp: 100 + 50 = 150 (no saturation needed)."""
+    instructions = [
+      v_mov_b32_e32(v[0], 100),
+      v_mov_b32_e32(v[1], 50),
+      v_add_nc_u32_e64(v[2], v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 150, f"expected 150, got {st.vgpr[0][2]}")
+
+
+class TestVOP3ClampUint16(unittest.TestCase):
+  """Tests for VOP3 clamp modifier on unsigned 16-bit integer operations."""
+
+  def test_v_sub_nc_u16_clamp_underflow(self):
+    """V_SUB_NC_U16 with clamp: 0 - 1 should saturate to 0."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0),
+      v_mov_b32_e32(v[1], 1),
+      v_sub_nc_u16(v[2], v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2] & 0xFFFF, 0, f"expected 0, got 0x{st.vgpr[0][2] & 0xFFFF:04x}")
+
+  def test_v_sub_nc_u16_clamp_no_underflow(self):
+    """V_SUB_NC_U16 with clamp: 100 - 50 = 50 (no saturation needed)."""
+    instructions = [
+      v_mov_b32_e32(v[0], 100),
+      v_mov_b32_e32(v[1], 50),
+      v_sub_nc_u16(v[2], v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2] & 0xFFFF, 50, f"expected 50, got {st.vgpr[0][2] & 0xFFFF}")
+
+  def test_v_add_nc_u16_clamp_overflow(self):
+    """V_ADD_NC_U16 with clamp: 0xFFFF + 1 should saturate to 0xFFFF."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0xFFFF),
+      v_mov_b32_e32(v[1], 1),
+      v_add_nc_u16(v[2], v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2] & 0xFFFF, 0xFFFF, f"expected 0xFFFF, got 0x{st.vgpr[0][2] & 0xFFFF:04x}")
+
+  def test_v_add_nc_u16_clamp_no_overflow(self):
+    """V_ADD_NC_U16 with clamp: 100 + 50 = 150 (no saturation needed)."""
+    instructions = [
+      v_mov_b32_e32(v[0], 100),
+      v_mov_b32_e32(v[1], 50),
+      v_add_nc_u16(v[2], v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2] & 0xFFFF, 150, f"expected 150, got {st.vgpr[0][2] & 0xFFFF}")
+
+
+class TestVOP3ClampInt32(unittest.TestCase):
+  """Tests for VOP3 clamp modifier on signed 32-bit integer operations."""
+
+  def test_v_add_nc_i32_clamp_overflow(self):
+    """V_ADD_NC_I32 with clamp: INT_MAX + 1 should saturate to INT_MAX."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0x7FFFFFFF),  # S0 = INT_MAX
+      v_mov_b32_e32(v[1], 1),            # S1 = 1
+      v_add_nc_i32(v[2], v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0x7FFFFFFF, f"expected 0x7FFFFFFF, got 0x{st.vgpr[0][2]:08x}")
+
+  def test_v_add_nc_i32_clamp_underflow(self):
+    """V_ADD_NC_I32 with clamp: INT_MIN + (-1) should saturate to INT_MIN."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0x80000000),  # S0 = INT_MIN
+      v_mov_b32_e32(v[1], 0xFFFFFFFF),  # S1 = -1
+      v_add_nc_i32(v[2], v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0x80000000, f"expected 0x80000000, got 0x{st.vgpr[0][2]:08x}")
+
+  def test_v_sub_nc_i32_clamp_underflow(self):
+    """V_SUB_NC_I32 with clamp: INT_MIN - 1 should saturate to INT_MIN."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0x80000000),  # S0 = INT_MIN
+      v_mov_b32_e32(v[1], 1),            # S1 = 1
+      v_sub_nc_i32(v[2], v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0x80000000, f"expected 0x80000000, got 0x{st.vgpr[0][2]:08x}")
+
+  def test_v_sub_nc_i32_clamp_overflow(self):
+    """V_SUB_NC_I32 with clamp: INT_MAX - (-1) should saturate to INT_MAX."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0x7FFFFFFF),  # S0 = INT_MAX
+      v_mov_b32_e32(v[1], 0xFFFFFFFF),  # S1 = -1
+      v_sub_nc_i32(v[2], v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0x7FFFFFFF, f"expected 0x7FFFFFFF, got 0x{st.vgpr[0][2]:08x}")
+
+  def test_v_add_nc_i32_no_saturation_positive(self):
+    """V_ADD_NC_I32 with clamp: 100 + 200 = 300 (no saturation needed)."""
+    instructions = [
+      v_mov_b32_e32(v[0], 100),
+      v_mov_b32_e32(v[1], 200),
+      v_add_nc_i32(v[2], v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 300, f"expected 300, got {st.vgpr[0][2]}")
+
+  def test_v_add_nc_i32_no_saturation_negative(self):
+    """V_ADD_NC_I32 with clamp: -100 + -200 = -300 (no saturation needed)."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0xFFFFFF9C),  # -100
+      v_mov_b32_e32(v[1], 0xFFFFFF38),  # -200
+      v_add_nc_i32(v[2], v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    expected = 0xFFFFFED4  # -300
+    self.assertEqual(st.vgpr[0][2], expected, f"expected 0x{expected:08x}, got 0x{st.vgpr[0][2]:08x}")
+
+
+class TestVOP3ClampCarry(unittest.TestCase):
+  """Tests for VOP3 clamp modifier on carry operations (VOP3SD)."""
+
+  def test_v_add_co_u32_clamp_overflow(self):
+    """V_ADD_CO_U32 with clamp: 0xFFFFFFFF + 1 should saturate to 0xFFFFFFFF."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0xFFFFFFFF),
+      v_mov_b32_e32(v[1], 1),
+      v_add_co_u32(v[2], VCC, v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0xFFFFFFFF, f"expected 0xFFFFFFFF, got 0x{st.vgpr[0][2]:08x}")
+
+  def test_v_add_co_u32_clamp_no_overflow(self):
+    """V_ADD_CO_U32 with clamp: 100 + 200 = 300 (no saturation)."""
+    instructions = [
+      v_mov_b32_e32(v[0], 100),
+      v_mov_b32_e32(v[1], 200),
+      v_add_co_u32(v[2], VCC, v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 300, f"expected 300, got {st.vgpr[0][2]}")
+
+  def test_v_sub_co_u32_clamp_underflow(self):
+    """V_SUB_CO_U32 with clamp: 0 - 1 should saturate to 0."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0),
+      v_mov_b32_e32(v[1], 1),
+      v_sub_co_u32(v[2], VCC, v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0, f"expected 0, got 0x{st.vgpr[0][2]:08x}")
+
+  def test_v_sub_co_u32_clamp_no_underflow(self):
+    """V_SUB_CO_U32 with clamp: 300 - 100 = 200 (no saturation)."""
+    instructions = [
+      v_mov_b32_e32(v[0], 300),
+      v_mov_b32_e32(v[1], 100),
+      v_sub_co_u32(v[2], VCC, v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 200, f"expected 200, got {st.vgpr[0][2]}")
+
+  def test_v_subrev_co_u32_clamp_underflow(self):
+    """V_SUBREV_CO_U32 with clamp: 1 - 0 reversed = 0 - 1 should saturate to 0."""
+    instructions = [
+      v_mov_b32_e32(v[0], 1),  # This becomes the subtrahend
+      v_mov_b32_e32(v[1], 0),  # This becomes the minuend (0 - 1)
+      v_subrev_co_u32(v[2], VCC, v[0], v[1], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0, f"expected 0, got 0x{st.vgpr[0][2]:08x}")
+
+  def test_v_add_co_ci_u32_clamp_overflow(self):
+    """V_ADD_CO_CI_U32 with clamp: 0xFFFFFFFF + 1 + 0 should saturate to 0xFFFFFFFF."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0xFFFFFFFF),
+      v_mov_b32_e32(v[1], 1),
+      s_mov_b64(VCC, 0),  # No carry in
+      v_add_co_ci_u32(v[2], VCC, v[0], v[1], VCC, clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0xFFFFFFFF, f"expected 0xFFFFFFFF, got 0x{st.vgpr[0][2]:08x}")
+
+  def test_v_add_co_ci_u32_clamp_overflow_with_carry(self):
+    """V_ADD_CO_CI_U32 with clamp: 0xFFFFFFFE + 1 + 1 should saturate to 0xFFFFFFFF."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0xFFFFFFFE),
+      v_mov_b32_e32(v[1], 1),
+      s_mov_b64(VCC, 1),  # Carry in = 1
+      v_add_co_ci_u32(v[2], VCC, v[0], v[1], VCC, clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0xFFFFFFFF, f"expected 0xFFFFFFFF, got 0x{st.vgpr[0][2]:08x}")
+
+  def test_v_sub_co_ci_u32_clamp_underflow(self):
+    """V_SUB_CO_CI_U32 with clamp: 0 - 1 - 0 should saturate to 0."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0),
+      v_mov_b32_e32(v[1], 1),
+      s_mov_b64(VCC, 0),  # No borrow in
+      v_sub_co_ci_u32(v[2], VCC, v[0], v[1], VCC, clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0, f"expected 0, got 0x{st.vgpr[0][2]:08x}")
+
+  def test_v_subrev_co_ci_u32_clamp_underflow(self):
+    """V_SUBREV_CO_CI_U32 with clamp: reversed 1 - 0 - 0 = 0 - 1 should saturate to 0."""
+    instructions = [
+      v_mov_b32_e32(v[0], 1),
+      v_mov_b32_e32(v[1], 0),
+      s_mov_b64(VCC, 0),
+      v_subrev_co_ci_u32(v[2], VCC, v[0], v[1], VCC, clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][2], 0, f"expected 0, got 0x{st.vgpr[0][2]:08x}")
+
+
+class TestVOP3ClampMAD(unittest.TestCase):
+  """Tests for VOP3 clamp modifier on MAD (multiply-add) operations."""
+
+  def test_v_mad_u16_clamp_overflow(self):
+    """V_MAD_U16 with clamp: 0xFFFF * 2 + 0 should saturate to 0xFFFF."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0xFFFF),
+      v_mov_b32_e32(v[1], 2),
+      v_mov_b32_e32(v[2], 0),
+      v_mad_u16(v[3], v[0], v[1], v[2], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][3] & 0xFFFF, 0xFFFF, f"expected 0xFFFF, got 0x{st.vgpr[0][3] & 0xFFFF:04x}")
+
+  def test_v_mad_u16_clamp_overflow_with_add(self):
+    """V_MAD_U16 with clamp: 0x8000 * 2 + 0x1000 should saturate to 0xFFFF."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0x8000),  # 32768
+      v_mov_b32_e32(v[1], 2),        # * 2 = 65536
+      v_mov_b32_e32(v[2], 0x1000),   # + 4096 = 69632 > 0xFFFF
+      v_mad_u16(v[3], v[0], v[1], v[2], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][3] & 0xFFFF, 0xFFFF, f"expected 0xFFFF, got 0x{st.vgpr[0][3] & 0xFFFF:04x}")
+
+  def test_v_mad_u16_no_overflow(self):
+    """V_MAD_U16 with clamp: 100 * 100 + 50 = 10050 (no saturation)."""
+    instructions = [
+      v_mov_b32_e32(v[0], 100),
+      v_mov_b32_e32(v[1], 100),
+      v_mov_b32_e32(v[2], 50),
+      v_mad_u16(v[3], v[0], v[1], v[2], clmp=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][3] & 0xFFFF, 10050, f"expected 10050, got {st.vgpr[0][3] & 0xFFFF}")
+
+  def test_v_mad_u16_no_clamp(self):
+    """V_MAD_U16 without clamp: 0xFFFF * 2 + 0 should wrap to 0xFFFE."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0xFFFF),
+      v_mov_b32_e32(v[1], 2),
+      v_mov_b32_e32(v[2], 0),
+      v_mad_u16(v[3], v[0], v[1], v[2], clmp=0),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    # 0xFFFF * 2 = 0x1FFFE, low 16 bits = 0xFFFE
+    self.assertEqual(st.vgpr[0][3] & 0xFFFF, 0xFFFE, f"expected 0xFFFE, got 0x{st.vgpr[0][3] & 0xFFFF:04x}")
+
+
+class TestCvtPkF16(unittest.TestCase):
+  """Tests for V_CVT_PK_RTZ_F16_F32 - pack two f32 to f16 with round toward zero."""
+
+  def test_cvt_pk_rtz_f16_f32_basic(self):
+    """V_CVT_PK_RTZ_F16_F32: basic pack of two f32 values."""
+    instructions = [
+      v_mov_b32_e32(v[0], 1.0),
+      v_mov_b32_e32(v[1], 2.0),
+      v_cvt_pk_rtz_f16_f32_e64(v[2], v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][2]
+    lo_f16 = f16(result & 0xffff)
+    hi_f16 = f16((result >> 16) & 0xffff)
+    self.assertAlmostEqual(lo_f16, 1.0, delta=0.01)
+    self.assertAlmostEqual(hi_f16, 2.0, delta=0.01)
+
+
+class TestCvtPkNorm(unittest.TestCase):
+  """Tests for V_CVT_PK_NORM_I16_F32 and V_CVT_PK_NORM_U16_F32."""
+
+  def test_cvt_pk_norm_i16_f32_basic(self):
+    """V_CVT_PK_NORM_I16_F32: pack two f32 to normalized i16."""
+    instructions = [
+      v_mov_b32_e32(v[0], 1.0),
+      v_mov_b32_e32(v[1], -1.0),
+      v_cvt_pk_norm_i16_f32(v[2], v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][2]
+    lo = result & 0xffff
+    hi = (result >> 16) & 0xffff
+    self.assertEqual(lo, 32767)
+    self.assertEqual(hi, 0x8001)  # -32767, hardware uses symmetric range
+
+  def test_cvt_pk_norm_u16_f32_basic(self):
+    """V_CVT_PK_NORM_U16_F32: pack two f32 to normalized u16."""
+    instructions = [
+      v_mov_b32_e32(v[0], 1.0),
+      v_mov_b32_e32(v[1], 0.5),
+      v_cvt_pk_norm_u16_f32(v[2], v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][2]
+    lo = result & 0xffff
+    hi = (result >> 16) & 0xffff
+    self.assertEqual(lo, 65535)
+    self.assertAlmostEqual(hi, 32768, delta=1)
+
+
+class TestCvtPkInt(unittest.TestCase):
+  """Tests for V_CVT_PK_I16_I32, V_CVT_PK_U16_U32, V_CVT_PK_I16_F32, V_CVT_PK_U16_F32."""
+
+  def test_cvt_pk_i16_i32_basic(self):
+    """V_CVT_PK_I16_I32: pack two i32 to i16."""
+    instructions = [
+      s_mov_b32(s[0], 100),
+      s_mov_b32(s[1], -100 & 0xffffffff),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_cvt_pk_i16_i32(v[2], v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][2]
+    lo = result & 0xffff
+    hi = (result >> 16) & 0xffff
+    lo_signed = lo if lo < 32768 else lo - 65536
+    hi_signed = hi if hi < 32768 else hi - 65536
+    self.assertEqual(lo_signed, 100)
+    self.assertEqual(hi_signed, -100)
+
+  def test_cvt_pk_u16_u32_basic(self):
+    """V_CVT_PK_U16_U32: pack two u32 to u16."""
+    instructions = [
+      s_mov_b32(s[0], 1000),
+      s_mov_b32(s[1], 2000),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_cvt_pk_u16_u32(v[2], v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][2]
+    lo = result & 0xffff
+    hi = (result >> 16) & 0xffff
+    self.assertEqual(lo, 1000)
+    self.assertEqual(hi, 2000)
+
+  def test_cvt_pk_i16_f32_basic(self):
+    """V_CVT_PK_I16_F32: convert two f32 to packed i16."""
+    instructions = [
+      v_mov_b32_e32(v[0], 100.5),
+      v_mov_b32_e32(v[1], -50.7),
+      v_cvt_pk_i16_f32(v[2], v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][2]
+    lo = result & 0xffff
+    hi = (result >> 16) & 0xffff
+    lo_signed = lo if lo < 32768 else lo - 65536
+    hi_signed = hi if hi < 32768 else hi - 65536
+    self.assertEqual(lo_signed, 100)
+    self.assertEqual(hi_signed, -50)
+
+  def test_cvt_pk_u16_f32_basic(self):
+    """V_CVT_PK_U16_F32: convert two f32 to packed u16."""
+    instructions = [
+      v_mov_b32_e32(v[0], 100.9),
+      v_mov_b32_e32(v[1], 200.1),
+      v_cvt_pk_u16_f32(v[2], v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][2]
+    lo = result & 0xffff
+    hi = (result >> 16) & 0xffff
+    self.assertEqual(lo, 100)
+    self.assertEqual(hi, 200)
+
+  def test_cvt_pk_u8_f32_basic(self):
+    """V_CVT_PK_U8_F32: convert f32 to u8 and pack at byte position."""
+    instructions = [
+      v_mov_b32_e32(v[0], 128.5),
+      v_mov_b32_e32(v[1], 0),
+      v_mov_b32_e32(v[2], 0),
+      v_cvt_pk_u8_f32(v[2], v[0], v[1], v[2]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][2]
+    byte0 = result & 0xff
+    self.assertEqual(byte0, 128)
+
+
+class TestDotProduct(unittest.TestCase):
+  """Tests for dot product instructions V_DOT4_U32_U8, V_DOT8_U32_U4."""
+
+  def test_v_dot4_u32_u8_basic(self):
+    """V_DOT4_U32_U8: 4-element dot product of u8 vectors."""
+    src0 = 0x04030201  # {4, 3, 2, 1}
+    src1 = 0x01010101  # {1, 1, 1, 1}
+    instructions = [
+      s_mov_b32(s[0], src0),
+      s_mov_b32(s[1], src1),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_mov_b32_e32(v[2], 0),
+      v_dot4_u32_u8(v[2], v[0], v[1], v[2]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][2]
+    self.assertEqual(result, 10)
+
+  def test_v_dot4_u32_u8_with_accumulator(self):
+    """V_DOT4_U32_U8 with non-zero accumulator."""
+    src0 = 0x02020202  # {2, 2, 2, 2}
+    src1 = 0x03030303  # {3, 3, 3, 3}
+    instructions = [
+      s_mov_b32(s[0], src0),
+      s_mov_b32(s[1], src1),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_mov_b32_e32(v[2], 100),
+      v_dot4_u32_u8(v[2], v[0], v[1], v[2]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][2]
+    self.assertEqual(result, 124)
+
+  def test_v_dot8_u32_u4_basic(self):
+    """V_DOT8_U32_U4: 8-element dot product of u4 vectors."""
+    # src0 = 8 nibbles: {1,2,3,4,5,6,7,8} packed as 0x87654321
+    # src1 = 8 nibbles: {1,1,1,1,1,1,1,1} packed as 0x11111111
+    # result = 1+2+3+4+5+6+7+8 = 36
+    src0 = 0x87654321
+    src1 = 0x11111111
+    instructions = [
+      s_mov_b32(s[0], src0),
+      s_mov_b32(s[1], src1),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_mov_b32_e32(v[2], 0),
+      v_dot8_u32_u4(v[2], v[0], v[1], v[2]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][2]
+    self.assertEqual(result, 36)
+
+
+class TestMinMaxF16Vop3(unittest.TestCase):
+  """Tests for V_MIN3_F16, V_MAX3_F16, V_MED3_F16, V_MINMAX_F16, V_MAXMIN_F16."""
+
+  def test_v_min3_f16_basic(self):
+    """V_MIN3_F16: minimum of three f16 values."""
+    instructions = [
+      s_mov_b32(s[0], f32_to_f16(3.0)),
+      s_mov_b32(s[1], f32_to_f16(1.0)),
+      s_mov_b32(s[2], f32_to_f16(2.0)),
+      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_min3_f16(v[3], v[0], v[1], v[2]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = f16(st.vgpr[0][3] & 0xffff)
+    self.assertAlmostEqual(result, 1.0, delta=0.01)
+
+  def test_v_max3_f16_basic(self):
+    """V_MAX3_F16: maximum of three f16 values."""
+    instructions = [
+      s_mov_b32(s[0], f32_to_f16(1.0)),
+      s_mov_b32(s[1], f32_to_f16(3.0)),
+      s_mov_b32(s[2], f32_to_f16(2.0)),
+      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_max3_f16(v[3], v[0], v[1], v[2]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = f16(st.vgpr[0][3] & 0xffff)
+    self.assertAlmostEqual(result, 3.0, delta=0.01)
+
+  def test_v_med3_f16_basic(self):
+    """V_MED3_F16: median of three f16 values."""
+    instructions = [
+      s_mov_b32(s[0], f32_to_f16(3.0)),
+      s_mov_b32(s[1], f32_to_f16(1.0)),
+      s_mov_b32(s[2], f32_to_f16(2.0)),
+      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_med3_f16(v[3], v[0], v[1], v[2]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = f16(st.vgpr[0][3] & 0xffff)
+    self.assertAlmostEqual(result, 2.0, delta=0.01)
+
+  def test_v_minmax_f16_basic(self):
+    """V_MINMAX_F16: clamp(src0, min=src1, max=src2)."""
+    instructions = [
+      s_mov_b32(s[0], f32_to_f16(2.5)),
+      s_mov_b32(s[1], f32_to_f16(1.0)),
+      s_mov_b32(s[2], f32_to_f16(2.0)),
+      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_minmax_f16(v[3], v[0], v[1], v[2]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = f16(st.vgpr[0][3] & 0xffff)
+    self.assertAlmostEqual(result, 2.0, delta=0.01)
+
+  def test_v_maxmin_f16_basic(self):
+    """V_MAXMIN_F16: clamp(src0, min=src2, max=src1)."""
+    instructions = [
+      s_mov_b32(s[0], f32_to_f16(0.5)),
+      s_mov_b32(s[1], f32_to_f16(2.0)),
+      s_mov_b32(s[2], f32_to_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_maxmin_f16(v[3], v[0], v[1], v[2]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = f16(st.vgpr[0][3] & 0xffff)
+    self.assertAlmostEqual(result, 1.0, delta=0.01)
+
+  def test_v_min3_f16_with_neg(self):
+    """V_MIN3_F16 with neg modifier: min(-3, 1, 2) = -3."""
+    instructions = [
+      s_mov_b32(s[0], f32_to_f16(3.0)),
+      s_mov_b32(s[1], f32_to_f16(1.0)),
+      s_mov_b32(s[2], f32_to_f16(2.0)),
+      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_min3_f16(v[3], -v[0], v[1], v[2]),  # neg on first operand
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = f16(st.vgpr[0][3] & 0xffff)
+    self.assertAlmostEqual(result, -3.0, delta=0.01)
+
+  def test_v_max3_f16_with_abs(self):
+    """V_MAX3_F16 with abs modifier: max(|-3|, 1, 2) = 3."""
+    instructions = [
+      s_mov_b32(s[0], f32_to_f16(-3.0)),
+      s_mov_b32(s[1], f32_to_f16(1.0)),
+      s_mov_b32(s[2], f32_to_f16(2.0)),
+      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_max3_f16(v[3], abs(v[0]), v[1], v[2]),  # abs on first operand
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = f16(st.vgpr[0][3] & 0xffff)
+    self.assertAlmostEqual(result, 3.0, delta=0.01)
+
+  def test_v_med3_f16_opsel_hi(self):
+    """V_MED3_F16 with opsel reading from hi half."""
+    # Pack two f16 values: hi=5.0, lo=1.0
+    packed = (f32_to_f16(5.0) << 16) | f32_to_f16(1.0)
+    instructions = [
+      s_mov_b32(s[0], packed),
+      s_mov_b32(s[1], f32_to_f16(3.0)),
+      s_mov_b32(s[2], f32_to_f16(4.0)),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_mov_b32_e32(v[2], s[2]),
+      # Read hi half of v[0] (5.0), med3(5, 3, 4) = 4
+      v_med3_f16(v[3], v[0].h, v[1], v[2]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = f16(st.vgpr[0][3] & 0xffff)
+    self.assertAlmostEqual(result, 4.0, delta=0.01)
+
+
+class TestSadHi(unittest.TestCase):
+  """Tests for V_SAD_HI_U8 instruction."""
+
+  def test_v_sad_hi_u8_basic(self):
+    """V_SAD_HI_U8: (sad << 16) + acc."""
+    # |1-5| + |2-6| + |3-7| + |4-8| = 16, << 16 = 0x100000, + 100 = 0x100064
+    instructions = [
+      v_mov_b32_e32(v[0], 0x04030201),
+      v_mov_b32_e32(v[1], 0x08070605),
+      v_mov_b32_e32(v[2], 100),
+      v_sad_hi_u8(v[3], v[0], v[1], v[2]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][3], (16 << 16) + 100)
+
+  def test_v_sad_hi_u8_zero_diff(self):
+    """V_SAD_HI_U8: identical inputs gives acc only."""
+    instructions = [
+      v_mov_b32_e32(v[0], 0x12345678),
+      v_mov_b32_e32(v[2], 50),
+      v_sad_hi_u8(v[3], v[0], v[0], v[2]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][3], 50)
+
+
+class TestPermlane(unittest.TestCase):
+  """Tests for V_PERMLANE16_B32 and V_PERMLANEX16_B32 instructions."""
+
+  def test_v_permlane16_b32_identity(self):
+    """V_PERMLANE16_B32 with identity permutation (lane i reads from lane i within row)."""
+    # lanesel encodes 4 bits per position: position i gets lanesel[i*4+3:i*4]
+    # Identity: position 0->0, 1->1, ..., 15->15
+    # lanesel = 0xFEDCBA9876543210 (positions 15-0 in nibbles)
+    instructions = [
+      v_mov_b32_e32(v[0], 0xDEADBEEF),  # source data
+      s_mov_b32(s[0], 0x76543210),       # lanesel low (positions 0-7)
+      s_mov_b32(s[1], 0xFEDCBA98),       # lanesel high (positions 8-15)
+      v_permlane16_b32(v[1], v[0], s[0], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    # Lane 0 reads from lane 0 (position 0 -> lanesel[3:0] = 0)
+    self.assertEqual(st.vgpr[0][1], 0xDEADBEEF)
+
+  def test_v_permlane16_b32_broadcast(self):
+    """V_PERMLANE16_B32 broadcast lane 0 to all lanes in row."""
+    # lanesel = all zeros -> all positions read from lane 0 within row
+    instructions = [
+      v_mov_b32_e32(v[0], 0xCAFEBABE),  # source data
+      s_mov_b32(s[0], 0),                # lanesel low = 0 (all read lane 0)
+      s_mov_b32(s[1], 0),                # lanesel high = 0
+      v_permlane16_b32(v[1], v[0], s[0], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=4)
+    # All lanes read from lane 0 of their row
+    for lane in range(4):
+      self.assertEqual(st.vgpr[lane][1], 0xCAFEBABE)
+
+  def test_v_permlanex16_b32_identity(self):
+    """V_PERMLANEX16_B32 cross-row read with identity selection."""
+    # In wave32: row 0 (lanes 0-15) reads from row 1 (lanes 16-31) and vice versa
+    # With single lane in row 0, it reads from lane 0 of row 1 (lane 16)
+    # But lane 16 doesn't exist in 1-lane test, so use 32 lanes
+    instructions = [
+      v_mov_b32_e32(v[0], 0x11111111),  # All lanes have this initially
+      s_mov_b32(s[0], 0x76543210),       # lanesel low
+      s_mov_b32(s[1], 0xFEDCBA98),       # lanesel high
+      v_permlanex16_b32(v[1], v[0], s[0], s[1]),
+    ]
+    st = run_program(instructions, n_lanes=32)
+    # Lane 0 in row 0 reads from lane 0 of row 1 (lane 16)
+    self.assertEqual(st.vgpr[0][1], 0x11111111)
+    # Lane 16 in row 1 reads from lane 0 of row 0 (lane 0)
+    self.assertEqual(st.vgpr[16][1], 0x11111111)
+
+
 if __name__ == '__main__':
  unittest.main()
@@ -149,7 +149,6 @@ class TestFmaMix(unittest.TestCase):

  def test_v_fma_mix_f32_src2_f16_lo(self):
    """V_FMA_MIX_F32 with src2 as f16 from lo bits."""
-    from extra.assembly.amd.test.hw.helpers import f32_to_f16
    f16_2 = f32_to_f16(2.0)
    instructions = [
      s_mov_b32(s[0], f2i(1.0)),
@@ -166,7 +165,6 @@ class TestFmaMix(unittest.TestCase):

  def test_v_fma_mix_f32_src2_f16_hi(self):
    """V_FMA_MIX_F32 with src2 as f16 from hi bits."""
-    from extra.assembly.amd.test.hw.helpers import f32_to_f16
    f16_2 = f32_to_f16(2.0)
    val = (f16_2 << 16) | 0
    instructions = [
@@ -199,7 +197,6 @@ class TestFmaMix(unittest.TestCase):

  def test_v_fma_mix_f32_with_abs_f16_src2_lo(self):
    """V_FMA_MIX_F32 with abs modifier on f16 src2 (lo half). Regression test for sin(1.0) bug."""
-    from extra.assembly.amd.test.hw.helpers import f32_to_f16
    f16_neg1 = f32_to_f16(-1.0)  # 0xbc00
    instructions = [
      s_mov_b32(s[0], f2i(0.0)),  # src0 = 0.0 (f32)
@@ -217,7 +214,6 @@ class TestFmaMix(unittest.TestCase):

  def test_v_fma_mix_f32_with_neg_f16_src2_lo(self):
    """V_FMA_MIX_F32 with neg modifier on f16 src2 (lo half)."""
-    from extra.assembly.amd.test.hw.helpers import f32_to_f16
    f16_1 = f32_to_f16(1.0)  # 0x3c00
    instructions = [
      s_mov_b32(s[0], f2i(0.0)),  # src0 = 0.0 (f32)
@@ -235,7 +231,6 @@ class TestFmaMix(unittest.TestCase):

  def test_v_fma_mix_f32_with_abs_f16_src2_hi(self):
    """V_FMA_MIX_F32 with abs modifier on f16 src2 (hi half)."""
-    from extra.assembly.amd.test.hw.helpers import f32_to_f16
    f16_neg1 = f32_to_f16(-1.0)  # 0xbc00
    val = (f16_neg1 << 16) | 0  # -1.0 in hi, 0 in lo
    instructions = [
@@ -254,7 +249,6 @@ class TestFmaMix(unittest.TestCase):

  def test_v_fma_mixlo_f16(self):
    """V_FMA_MIXLO_F16 writes to low 16 bits of destination."""
-    from extra.assembly.amd.test.hw.helpers import _f16
    instructions = [
      s_mov_b32(s[0], f2i(2.0)),
      v_mov_b32_e32(v[0], s[0]),
@@ -267,14 +261,13 @@ class TestFmaMix(unittest.TestCase):
      VOP3P(VOP3POp.V_FMA_MIXLO_F16, vdst=v[3], src0=v[0], src1=v[1], src2=v[2], opsel=0, opsel_hi=0, opsel_hi2=0),
    ]
    st = run_program(instructions, n_lanes=1)
-    lo = _f16(st.vgpr[0][3] & 0xffff)
+    lo = f16(st.vgpr[0][3] & 0xffff)
    hi = (st.vgpr[0][3] >> 16) & 0xffff
    self.assertAlmostEqual(lo, 7.0, places=1)
    self.assertEqual(hi, 0xdead, f"hi should be preserved, got 0x{hi:04x}")

  def test_v_fma_mixlo_f16_all_f32_sources(self):
    """V_FMA_MIXLO_F16 with all f32 sources."""
-    from extra.assembly.amd.test.hw.helpers import _f16
    instructions = [
      s_mov_b32(s[0], f2i(1.0)),
      v_mov_b32_e32(v[0], s[0]),
@@ -286,13 +279,12 @@ class TestFmaMix(unittest.TestCase):
      VOP3P(VOP3POp.V_FMA_MIXLO_F16, vdst=v[3], src0=v[0], src1=v[1], src2=v[2], opsel=0, opsel_hi=0, opsel_hi2=0),
    ]
    st = run_program(instructions, n_lanes=1)
-    lo = _f16(st.vgpr[0][3] & 0xffff)
+    lo = f16(st.vgpr[0][3] & 0xffff)
    # 1*2+3 = 5
    self.assertAlmostEqual(lo, 5.0, places=1)

  def test_v_fma_mixlo_f16_sin_case(self):
    """V_FMA_MIXLO_F16 case from sin kernel."""
-    from extra.assembly.amd.test.hw.helpers import _f16
    instructions = [
      s_mov_b32(s[0], 0x3f800000),  # f32 1.0
      v_mov_b32_e32(v[3], s[0]),
@@ -305,7 +297,7 @@ class TestFmaMix(unittest.TestCase):
      VOP3P(VOP3POp.V_FMA_MIXLO_F16, vdst=v[3], src0=v[3], src1=s[6], src2=v[5], opsel=0, opsel_hi=0, opsel_hi2=0),
    ]
    st = run_program(instructions, n_lanes=1)
-    lo = _f16(st.vgpr[0][3] & 0xffff)
+    lo = f16(st.vgpr[0][3] & 0xffff)
    self.assertAlmostEqual(lo, -3.14159, delta=0.01)


@@ -314,7 +306,6 @@ class TestVOP3P(unittest.TestCase):

  def test_v_pk_add_f16_basic(self):
    """V_PK_ADD_F16 adds two packed f16 values."""
-    from extra.assembly.amd.test.hw.helpers import _f16
    instructions = [
      s_mov_b32(s[0], 0x40003c00),  # hi=2.0, lo=1.0
      s_mov_b32(s[1], 0x44004200),  # hi=4.0, lo=3.0
@@ -324,14 +315,13 @@ class TestVOP3P(unittest.TestCase):
    ]
    st = run_program(instructions, n_lanes=1)
    result = st.vgpr[0][2]
-    lo = _f16(result & 0xffff)
-    hi = _f16((result >> 16) & 0xffff)
+    lo = f16(result & 0xffff)
+    hi = f16((result >> 16) & 0xffff)
    self.assertAlmostEqual(lo, 4.0, places=2)
    self.assertAlmostEqual(hi, 6.0, places=2)

  def test_v_pk_mul_f16_basic(self):
    """V_PK_MUL_F16 multiplies two packed f16 values."""
-    from extra.assembly.amd.test.hw.helpers import _f16
    instructions = [
      s_mov_b32(s[0], 0x42004000),  # hi=3.0, lo=2.0
      s_mov_b32(s[1], 0x45004400),  # hi=5.0, lo=4.0
@@ -341,14 +331,13 @@ class TestVOP3P(unittest.TestCase):
    ]
    st = run_program(instructions, n_lanes=1)
    result = st.vgpr[0][2]
-    lo = _f16(result & 0xffff)
-    hi = _f16((result >> 16) & 0xffff)
+    lo = f16(result & 0xffff)
+    hi = f16((result >> 16) & 0xffff)
    self.assertAlmostEqual(lo, 8.0, places=1)
    self.assertAlmostEqual(hi, 15.0, places=1)

  def test_v_pk_fma_f16_basic(self):
    """V_PK_FMA_F16: D = A * B + C for packed f16."""
-    from extra.assembly.amd.test.hw.helpers import _f16
    instructions = [
      s_mov_b32(s[0], 0x42004000),  # A: hi=3.0, lo=2.0
      s_mov_b32(s[1], 0x45004400),  # B: hi=5.0, lo=4.0
@@ -360,8 +349,8 @@ class TestVOP3P(unittest.TestCase):
    ]
    st = run_program(instructions, n_lanes=1)
    result = st.vgpr[0][3]
-    lo = _f16(result & 0xffff)
-    hi = _f16((result >> 16) & 0xffff)
+    lo = f16(result & 0xffff)
+    hi = f16((result >> 16) & 0xffff)
    self.assertAlmostEqual(lo, 9.0, places=1)   # 2*4+1
    self.assertAlmostEqual(hi, 16.0, places=0)  # 3*5+1

@@ -370,7 +359,6 @@ class TestVOP3P(unittest.TestCase):
    Inline constants for VOP3P are f16 values in the low 16 bits only.
    hi half of inline constant is 0, so hi result = v0.hi + 0 = 1.0.
    """
-    from extra.assembly.amd.test.hw.helpers import _f16
    instructions = [
      s_mov_b32(s[0], 0x3c003c00),  # packed f16: hi=1.0, lo=1.0
      v_mov_b32_e32(v[0], s[0]),
@@ -378,8 +366,8 @@ class TestVOP3P(unittest.TestCase):
    ]
    st = run_program(instructions, n_lanes=1)
    result = st.vgpr[0][1]
-    lo = _f16(result & 0xffff)
-    hi = _f16((result >> 16) & 0xffff)
+    lo = f16(result & 0xffff)
+    hi = f16((result >> 16) & 0xffff)
    # lo = 1.0 + 1.0 = 2.0, hi = 1.0 + 0.0 = 1.0 (inline const hi half is 0)
    self.assertAlmostEqual(lo, 2.0, places=2)
    self.assertAlmostEqual(hi, 1.0, places=2)
@@ -388,7 +376,6 @@ class TestVOP3P(unittest.TestCase):
    """V_PK_MUL_F16 with inline constant POS_TWO (2.0).
    Inline constant has value only in low 16 bits, hi is 0.
    """
-    from extra.assembly.amd.test.hw.helpers import _f16
    # v0 = packed (3.0, 4.0), multiply by POS_TWO
    # lo = 3.0 * 2.0 = 6.0, hi = 4.0 * 0.0 = 0.0 (inline const hi is 0)
    instructions = [
@@ -398,11 +385,29 @@ class TestVOP3P(unittest.TestCase):
    ]
    st = run_program(instructions, n_lanes=1)
    result = st.vgpr[0][1]
-    lo = _f16(result & 0xffff)
-    hi = _f16((result >> 16) & 0xffff)
+    lo = f16(result & 0xffff)
+    hi = f16((result >> 16) & 0xffff)
    self.assertAlmostEqual(lo, 6.0, places=1)
    self.assertAlmostEqual(hi, 0.0, places=1)

+  def test_v_pk_add_u16_float_inline_const_opsel(self):
+    """V_PK_ADD_U16 with float inline constant 2.0
+    Regression test: for integer packed ops, do not perform the f32->f16 conversion.
+    """
+    # src1 = inline float constant 2.0
+    instructions = [
+      s_mov_b32(s[0], 0x00030005),  # packed u16: hi=3, lo=5
+      v_mov_b32_e32(v[0], s[0]),
+      v_pk_add_u16(v[1], v[0], SrcEnum.POS_TWO, opsel_hi=3, opsel_hi2=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][1]
+    lo = result & 0xffff
+    hi = (result >> 16) & 0xffff
+    # lo = 5 + 0x0000 = 0x0005, hi = 3 + 0x4000 = 0x4003
+    self.assertEqual(lo, 0x0005, f"lo: expected 0x0005, got 0x{lo:04x}")
+    self.assertEqual(hi, 0x4003, f"hi: expected 0x4003, got 0x{hi:04x}")
+

 class TestWMMAF16(unittest.TestCase):
  """Tests for WMMA F16 output variant (V_WMMA_F16_16X16X16_F16).
@@ -413,7 +418,6 @@ class TestWMMAF16(unittest.TestCase):

  def test_v_wmma_f16_16x16x16_f16_all_ones(self):
    """V_WMMA_F16_16X16X16_F16 with all ones produces 16.0 in f16."""
-    from extra.assembly.amd.test.hw.helpers import _f16
    instructions = []
    instructions.append(s_mov_b32(s[0], 0x3c003c00))  # packed f16 1.0
    # Initialize A matrix in v[16:23] (8 regs)
@@ -432,13 +436,12 @@ class TestWMMAF16(unittest.TestCase):
    for lane in range(32):
      for reg in range(8):
        result = st.vgpr[lane][reg]
-        lo = _f16(result & 0xffff)
+        lo = f16(result & 0xffff)
        self.assertAlmostEqual(lo, 16.0, places=1, msg=f"v[{reg}] lane {lane}: expected 16.0, got {lo}")
        self.assertEqual(result >> 16, 0, msg=f"v[{reg}] lane {lane}: hi bits should be 0")

  def test_v_wmma_f16_16x16x16_f16_with_accumulator(self):
    """V_WMMA_F16_16X16X16_F16 with non-zero accumulator."""
-    from extra.assembly.amd.test.hw.helpers import _f16
    instructions = []
    instructions.append(s_mov_b32(s[0], 0x3c003c00))  # packed f16 1.0
    instructions.append(s_mov_b32(s[1], 0x4500))  # f16 5.0 in lo bits only
@@ -458,7 +461,7 @@ class TestWMMAF16(unittest.TestCase):
    for lane in range(32):
      for reg in range(8):
        result = st.vgpr[lane][reg]
-        lo = _f16(result & 0xffff)
+        lo = f16(result & 0xffff)
        self.assertAlmostEqual(lo, 21.0, places=0, msg=f"v[{reg}] lane {lane}: expected 21.0, got {lo}")
        self.assertEqual(result >> 16, 0, msg=f"v[{reg}] lane {lane}: hi bits should be 0")

@@ -468,7 +471,6 @@ class TestWMMAF16(unittest.TestCase):
    Regression test: WMMA was using static register indices instead of dynamic.
    This test uses v[64:71] for A, v[80:87] for B, v[96:103] for C/D.
    """
-    from extra.assembly.amd.test.hw.helpers import _f16
    instructions = []
    instructions.append(s_mov_b32(s[0], 0x3c003c00))  # packed f16 1.0
    # Initialize A matrix in v[64:71] (8 regs)
@@ -490,7 +492,7 @@ class TestWMMAF16(unittest.TestCase):
    for lane in range(32):
      for reg in range(8):
        result = st.vgpr[lane][reg]
-        lo = _f16(result & 0xffff)
+        lo = f16(result & 0xffff)
        self.assertAlmostEqual(lo, 16.0, places=1, msg=f"v[{reg}] lane {lane}: expected 16.0, got {lo}")
        self.assertEqual(result >> 16, 0, msg=f"v[{reg}] lane {lane}: hi bits should be 0")

@@ -713,7 +715,6 @@ class TestPackedMixedSigns(unittest.TestCase):

  def test_pk_add_f16_mixed_signs(self):
    """V_PK_ADD_F16 with mixed positive/negative values."""
-    from extra.assembly.amd.test.hw.helpers import _f16
    instructions = [
      s_mov_b32(s[0], 0xc0003c00),  # packed: hi=-2.0, lo=1.0
      s_mov_b32(s[1], 0x3c003c00),  # packed: hi=1.0, lo=1.0
@@ -723,14 +724,13 @@ class TestPackedMixedSigns(unittest.TestCase):
    ]
    st = run_program(instructions, n_lanes=1)
    result = st.vgpr[0][2]
-    lo = _f16(result & 0xffff)
-    hi = _f16((result >> 16) & 0xffff)
+    lo = f16(result & 0xffff)
+    hi = f16((result >> 16) & 0xffff)
    self.assertAlmostEqual(lo, 2.0, places=2)   # 1.0 + 1.0
    self.assertAlmostEqual(hi, -1.0, places=2)  # -2.0 + 1.0

  def test_pk_mul_f16_zero(self):
    """V_PK_MUL_F16 with zero."""
-    from extra.assembly.amd.test.hw.helpers import _f16
    instructions = [
      s_mov_b32(s[0], 0x40004000),  # packed: 2.0, 2.0
      s_mov_b32(s[1], 0x00000000),  # packed: 0.0, 0.0
@@ -743,5 +743,277 @@ class TestPackedMixedSigns(unittest.TestCase):
    self.assertEqual(result, 0x00000000, "2.0 * 0.0 should be 0.0")


+class TestDot2F32F16(unittest.TestCase):
+  """Tests for V_DOT2_F32_F16 - dot product of f16 pairs producing f32."""
+
+  def test_v_dot2_f32_f16_basic(self):
+    """V_DOT2_F32_F16: dot product of two packed f16 pairs -> f32."""
+    # src0 = {hi=2.0, lo=1.0}, src1 = {hi=4.0, lo=3.0}
+    # result = 1.0*3.0 + 2.0*4.0 + 0 = 3 + 8 = 11.0
+    src0 = (f32_to_f16(2.0) << 16) | f32_to_f16(1.0)
+    src1 = (f32_to_f16(4.0) << 16) | f32_to_f16(3.0)
+    instructions = [
+      s_mov_b32(s[0], src0),
+      s_mov_b32(s[1], src1),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_mov_b32_e32(v[2], 0),
+      v_dot2_f32_f16(v[3], v[0], v[1], v[2], opsel_hi=3, opsel_hi2=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = i2f(st.vgpr[0][3])
+    self.assertAlmostEqual(result, 11.0, places=2)
+
+  def test_v_dot2_f32_f16_with_accumulator(self):
+    """V_DOT2_F32_F16 with non-zero f32 accumulator."""
+    # src0 = {hi=1.0, lo=1.0}, src1 = {hi=1.0, lo=1.0}, acc = 5.0
+    # result = 1.0*1.0 + 1.0*1.0 + 5.0 = 7.0
+    src0 = (f32_to_f16(1.0) << 16) | f32_to_f16(1.0)
+    instructions = [
+      s_mov_b32(s[0], src0),
+      s_mov_b32(s[1], f2i(5.0)),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[0]),  # same as src0
+      v_mov_b32_e32(v[2], s[1]),
+      v_dot2_f32_f16(v[3], v[0], v[1], v[2], opsel_hi=3, opsel_hi2=1),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = i2f(st.vgpr[0][3])
+    self.assertAlmostEqual(result, 7.0, places=2)
+
+  def test_v_dot2_f32_f16_negative_values(self):
+    """V_DOT2_F32_F16 with negative f16 values."""
+    # src0 = {hi=-2.0, lo=3.0}, src1 = {hi=1.0, lo=2.0}
+    # result = 3.0*2.0 + (-2.0)*1.0 + 0 = 6 - 2 = 4.0
+    # NOTE: Hardware DOT2 may have up to 1 ULP difference due to internal implementation
+    src0 = (f32_to_f16(-2.0) << 16) | f32_to_f16(3.0)
+    src1 = (f32_to_f16(1.0) << 16) | f32_to_f16(2.0)
+    instructions = [
+      s_mov_b32(s[0], src0),
+      s_mov_b32(s[1], src1),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_mov_b32_e32(v[2], 0),
+      v_dot2_f32_f16(v[3], v[0], v[1], v[2], opsel_hi=3, opsel_hi2=1),
+    ]
+    st = run_program(instructions, n_lanes=1, ulp_tolerance=1)
+    result = i2f(st.vgpr[0][3])
+    self.assertAlmostEqual(result, 4.0, places=2)
+
+
+class TestDot2F16F16(unittest.TestCase):
+  """Tests for V_DOT2_F16_F16 - dot product of f16 pairs producing f16."""
+
+  def test_v_dot2_f16_f16_basic(self):
+    """V_DOT2_F16_F16: dot product of two packed f16 pairs -> f16."""
+    # src0 = {hi=2.0, lo=1.0}, src1 = {hi=3.0, lo=2.0}
+    # result = 1.0*2.0 + 2.0*3.0 + 0 = 2 + 6 = 8.0 (f16)
+    src0 = (f32_to_f16(2.0) << 16) | f32_to_f16(1.0)
+    src1 = (f32_to_f16(3.0) << 16) | f32_to_f16(2.0)
+    instructions = [
+      s_mov_b32(s[0], src0),
+      s_mov_b32(s[1], src1),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_mov_b32_e32(v[2], 0),
+      v_dot2_f16_f16(v[3], v[0], v[1], v[2]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = f16(st.vgpr[0][3] & 0xffff)
+    self.assertAlmostEqual(result, 8.0, places=1)
+
+  def test_v_dot2_f16_f16_with_accumulator(self):
+    """V_DOT2_F16_F16 with non-zero f16 accumulator."""
+    # src0 = {hi=1.0, lo=1.0}, src1 = {hi=1.0, lo=1.0}, acc = 3.0 (f16)
+    # result = 1.0*1.0 + 1.0*1.0 + 3.0 = 5.0 (f16)
+    src0 = (f32_to_f16(1.0) << 16) | f32_to_f16(1.0)
+    acc = f32_to_f16(3.0)
+    instructions = [
+      s_mov_b32(s[0], src0),
+      s_mov_b32(s[2], acc),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[0]),  # same as src0
+      v_mov_b32_e32(v[2], s[2]),
+      v_dot2_f16_f16(v[3], v[0], v[1], v[2]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = f16(st.vgpr[0][3] & 0xffff)
+    self.assertAlmostEqual(result, 5.0, places=1)
+
+
+class TestSignedDotProducts(unittest.TestCase):
+  """Tests for V_DOT4_I32_IU8 and V_DOT8_I32_IU4 with signed inputs."""
+
+  def test_v_dot4_i32_iu8_signed_both(self):
+    """V_DOT4_I32_IU8 with both inputs signed (neg=0b011)."""
+    # src0 = {-1, -2, 3, 4} as i8 = {0xff, 0xfe, 0x03, 0x04}
+    # src1 = {1, 1, 1, 1} as i8
+    # result = (-1)*1 + (-2)*1 + 3*1 + 4*1 = -1 - 2 + 3 + 4 = 4
+    src0 = (0xff << 24) | (0xfe << 16) | (0x03 << 8) | 0x04  # -1, -2, 3, 4
+    src1 = 0x01010101  # 1, 1, 1, 1
+    instructions = [
+      s_mov_b32(s[0], src0),
+      s_mov_b32(s[1], src1),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_mov_b32_e32(v[2], 0),
+      v_dot4_i32_iu8(v[3], v[0], v[1], v[2], neg=0b011),  # both signed
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][3]
+    # Result is i32, interpret as signed
+    if result >= 0x80000000:
+      result = result - 0x100000000
+    self.assertEqual(result, 4)
+
+  def test_v_dot4_i32_iu8_src0_signed(self):
+    """V_DOT4_I32_IU8 with only src0 signed (neg=0b001)."""
+    # src0 = {-1, -1, -1, -1} as i8 = {0xff, 0xff, 0xff, 0xff}
+    # src1 = {2, 2, 2, 2} as u8
+    # result = (-1)*2 + (-1)*2 + (-1)*2 + (-1)*2 = -8
+    src0 = 0xffffffff  # -1, -1, -1, -1 (as i8)
+    src1 = 0x02020202  # 2, 2, 2, 2 (as u8)
+    instructions = [
+      s_mov_b32(s[0], src0),
+      s_mov_b32(s[1], src1),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_mov_b32_e32(v[2], 0),
+      v_dot4_i32_iu8(v[3], v[0], v[1], v[2], neg=0b001),  # src0 signed
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][3]
+    if result >= 0x80000000:
+      result = result - 0x100000000
+    self.assertEqual(result, -8)
+
+  def test_v_dot4_i32_iu8_src1_signed(self):
+    """V_DOT4_I32_IU8 with only src1 signed (neg=0b010)."""
+    # src0 = {2, 2, 2, 2} as u8
+    # src1 = {-1, -1, -1, -1} as i8 = {0xff, 0xff, 0xff, 0xff}
+    # result = 2*(-1) + 2*(-1) + 2*(-1) + 2*(-1) = -8
+    src0 = 0x02020202  # 2, 2, 2, 2 (as u8)
+    src1 = 0xffffffff  # -1, -1, -1, -1 (as i8)
+    instructions = [
+      s_mov_b32(s[0], src0),
+      s_mov_b32(s[1], src1),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_mov_b32_e32(v[2], 0),
+      v_dot4_i32_iu8(v[3], v[0], v[1], v[2], neg=0b010),  # src1 signed
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][3]
+    if result >= 0x80000000:
+      result = result - 0x100000000
+    self.assertEqual(result, -8)
+
+  def test_v_dot4_i32_iu8_unsigned_as_reference(self):
+    """V_DOT4_I32_IU8 with both unsigned (neg=0) - same as V_DOT4_U32_U8."""
+    # src0 = {0xff, 0xff, 0xff, 0xff} = 255 each as u8
+    # src1 = {1, 1, 1, 1}
+    # result = 255*1 + 255*1 + 255*1 + 255*1 = 1020
+    src0 = 0xffffffff
+    src1 = 0x01010101
+    instructions = [
+      s_mov_b32(s[0], src0),
+      s_mov_b32(s[1], src1),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_mov_b32_e32(v[2], 0),
+      v_dot4_i32_iu8(v[3], v[0], v[1], v[2], neg=0),  # both unsigned
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vgpr[0][3], 1020)
+
+  def test_v_dot8_i32_iu4_signed_both(self):
+    """V_DOT8_I32_IU4 with both inputs signed (neg=0b011)."""
+    # src0 = 8 nibbles: {-1, -2, 3, 4, -1, -2, 3, 4} as i4
+    # i4 -1 = 0xf, -2 = 0xe, 3 = 0x3, 4 = 0x4
+    # src0 = 0xfe34fe34
+    # src1 = {1, 1, 1, 1, 1, 1, 1, 1} as i4 = 0x11111111
+    # result = 2 * ((-1)*1 + (-2)*1 + 3*1 + 4*1) = 2 * 4 = 8
+    src0 = 0xfe34fe34
+    src1 = 0x11111111
+    instructions = [
+      s_mov_b32(s[0], src0),
+      s_mov_b32(s[1], src1),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_mov_b32_e32(v[2], 0),
+      v_dot8_i32_iu4(v[3], v[0], v[1], v[2], neg=0b011),  # both signed
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][3]
+    if result >= 0x80000000:
+      result = result - 0x100000000
+    self.assertEqual(result, 8)
+
+  def test_v_dot8_i32_iu4_all_negative(self):
+    """V_DOT8_I32_IU4 with all negative signed values."""
+    # src0 = 8 nibbles all -1 (0xf) = 0xffffffff
+    # src1 = 8 nibbles all 1 = 0x11111111
+    # result = 8 * ((-1)*1) = -8
+    src0 = 0xffffffff  # all -1 as i4
+    src1 = 0x11111111  # all 1
+    instructions = [
+      s_mov_b32(s[0], src0),
+      s_mov_b32(s[1], src1),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_mov_b32_e32(v[2], 0),
+      v_dot8_i32_iu4(v[3], v[0], v[1], v[2], neg=0b011),  # both signed
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][3]
+    if result >= 0x80000000:
+      result = result - 0x100000000
+    self.assertEqual(result, -8)
+
+
+class TestPkMinMaxF16(unittest.TestCase):
+  """Tests for V_PK_MIN_F16 and V_PK_MAX_F16."""
+
+  def test_v_pk_min_f16_basic(self):
+    """V_PK_MIN_F16: packed min of two f16 pairs."""
+    # src0 = {hi=3.0, lo=1.0}, src1 = {hi=2.0, lo=4.0}
+    # result = {min(3,2)=2, min(1,4)=1}
+    src0 = (f32_to_f16(3.0) << 16) | f32_to_f16(1.0)
+    src1 = (f32_to_f16(2.0) << 16) | f32_to_f16(4.0)
+    instructions = [
+      s_mov_b32(s[0], src0),
+      s_mov_b32(s[1], src1),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_pk_min_f16(v[2], v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][2]
+    lo = f16(result & 0xffff)
+    hi = f16((result >> 16) & 0xffff)
+    self.assertAlmostEqual(lo, 1.0, delta=0.01)
+    self.assertAlmostEqual(hi, 2.0, delta=0.01)
+
+  def test_v_pk_max_f16_basic(self):
+    """V_PK_MAX_F16: packed max of two f16 pairs."""
+    # src0 = {hi=3.0, lo=1.0}, src1 = {hi=2.0, lo=4.0}
+    # result = {max(3,2)=3, max(1,4)=4}
+    src0 = (f32_to_f16(3.0) << 16) | f32_to_f16(1.0)
+    src1 = (f32_to_f16(2.0) << 16) | f32_to_f16(4.0)
+    instructions = [
+      s_mov_b32(s[0], src0),
+      s_mov_b32(s[1], src1),
+      v_mov_b32_e32(v[0], s[0]),
+      v_mov_b32_e32(v[1], s[1]),
+      v_pk_max_f16(v[2], v[0], v[1]),
+    ]
+    st = run_program(instructions, n_lanes=1)
+    result = st.vgpr[0][2]
+    lo = f16(result & 0xffff)
+    hi = f16((result >> 16) & 0xffff)
+    self.assertAlmostEqual(lo, 4.0, delta=0.01)
+    self.assertAlmostEqual(hi, 3.0, delta=0.01)
+
+
 if __name__ == '__main__':
  unittest.main()
@@ -180,7 +180,7 @@ class TestDSPcodePatterns(unittest.TestCase):
  def test_mem_read_parsing(self):
    """Test MEM[addr].type read expression parsing."""
    # Create a mock LDS buffer
-    lds = UOp(Ops.DEFINE_GLOBAL, dtypes.uint32.ptr(16384), arg=3)
+    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)}

@@ -213,7 +213,7 @@ class TestDSPcodePatterns(unittest.TestCase):
    """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.DEFINE_GLOBAL, dtypes.uint32.ptr(16384), arg=3)
+    lds = UOp(Ops.PARAM, dtypes.uint32.ptr(16384), arg=3)
    srcs = {
      'ADDR': UOp.const(dtypes.uint32, 0),
      'OFFSET0': UOp.const(dtypes.uint32, 0),
@@ -286,7 +286,7 @@ class TestAllPcode(unittest.TestCase):
  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.DEFINE_GLOBAL, dtypes.uint32.ptr(16384), arg=3)
+    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(),
@@ -294,7 +294,7 @@ class TestAllPcode(unittest.TestCase):
            '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(), 'WAVE_STATUS': 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)}

@@ -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()
@@ -203,12 +203,12 @@ class SQTTExamplesTestBase(unittest.TestCase):
 class TestSQTTExamplesRDNA3(SQTTExamplesTestBase):
  target = "gfx1100"
  expected = {
-    "profile_empty_run_0": [1803, 1908, 1928, 1979, 2006, 1912],
-    "profile_empty_run_1": [1803, 1908, 1928, 1979, 2006, 1912],
-    "profile_gemm_run_0": [2531, 1844, 1864, 1915, 1942, 1848, 3074, 1919, 1939, 1990, 2017, 1923, 19026, 1919, 1939, 1990, 2017, 1929],
-    "profile_gemm_run_1": [2554, 1844, 1864, 1915, 1942, 1848, 3084, 1919, 1939, 1990, 2017, 1923, 19010, 1919, 1939, 1990, 2017, 1923],
-    "profile_plus_run_0": [1900, 1908, 1928, 1979, 2006, 1912],
-    "profile_plus_run_1": [1856, 1908, 1928, 1979, 2006, 1912],
+    "profile_empty_run_0": [1844, 1885, 1905, 1956, 1983, 1889],
+    "profile_empty_run_1": [1780, 1885, 1905, 1956, 1983, 1889],
+    "profile_gemm_run_0": [2656, 2025, 2045, 2096, 2123, 2029, 3183, 2019, 2039, 2090, 2117, 2023, 19119, 2013, 2033, 2084, 2111, 2017],
+    "profile_gemm_run_1": [2662, 2025, 2045, 2096, 2123, 2029, 3179, 2019, 2039, 2090, 2117, 2023, 19113, 2071, 2091, 2142, 2169, 2075],
+    "profile_plus_run_0": [1886, 2013, 2033, 2084, 2111, 2017],
+    "profile_plus_run_1": [1988, 2071, 2091, 2142, 2169, 2075],
  }

 class TestSQTTExamplesRDNA4(SQTTExamplesTestBase): target = "gfx1200"
@@ -471,7 +471,7 @@ THREADS = 128

 def test_matmul():
  dev = Device[Device.DEFAULT]
-  print(f"Device arch: {dev.arch}")
+  print(f"Device arch: {dev.renderer.arch}")

  if getenv("STOCK", 0):
    # Load the stock kernel from amd_seb/kernel8_batched_gmem.s
@@ -479,7 +479,7 @@ def test_matmul():
    asm = stock_path.read_text()
    print(f"Loaded stock kernel from {stock_path}")
  else:
-    asm = build_kernel(dev.arch)
+    asm = build_kernel(dev.renderer.arch)

  binary = dev.compiler.compile(asm)
  print(f"Compiled! Binary size: {len(binary)} bytes")
@@ -0,0 +1,96 @@
+import atexit, functools
+from tinygrad.runtime.support.compiler_amd import HIPCompiler
+from tinygrad import Tensor, Device, dtypes
+from tinygrad.uop.ops import UOp, Ops, KernelInfo, AxisType
+from tinygrad.renderer import Estimates
+from tinygrad.helpers import getenv, all_same, dedup
+from extra.gemm.asm.cdna.asm import build_kernel, GEMM_ARGS
+
+# ** CDNA4 assembly gemm
+
+WORKGROUP_SIZE = 256
+
+def custom_asm_gemm(C:UOp, A:UOp, B:UOp, dname:str, arch:str, wg:int) -> UOp:
+  batch, M, K = A.shape
+  K2, N = B.shape[(1 if B.ndim == 3 else 0):]
+  assert K == K2
+  lidx = UOp.special(WORKGROUP_SIZE, "lidx0")
+  gidx = UOp.special(wg, "gidx0")
+  k = build_kernel(batch, M, N, K, A.dtype.base)
+  sink = UOp.sink(C.base, A.base, B.base, lidx, gidx,
+                  arg=KernelInfo(name=k.name, estimates=Estimates(ops=2*batch*M*N*K, mem=(batch*M*K + K*N + batch*M*N)*2)))
+  binary = HIPCompiler(arch).compile(k.to_asm())
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)),
+                               UOp(Ops.SOURCE, arg=k.to_text()), UOp(Ops.BINARY, arg=binary)))
+
+counters = {"used":0, "todos":[]}
+def todo(msg:str) -> bool: counters["todos"].append(msg); return False
+atexit.register(lambda: print(f'asm_gemm: {counters["used"]} used, {len(counters["todos"])} not used'))
+
+def can_use_asm_gemm(a:Tensor, b:Tensor) -> bool:
+  if a.dtype != b.dtype: return todo(f"dtypes must match {a.dtype} != {b.dtype}")
+  if a.dtype not in {dtypes.bfloat16, dtypes.float16}: return todo(f"only bfloat16/float16, got {a.dtype}")
+  # only sharding on the batch is tested, others might work too
+  if isinstance(a.device, tuple) and not (a.ndim == 3 and a.uop.axis == 0 and b.uop.axis is None):
+    return todo(f"sharding mismatch a.ndim={a.ndim} a.uop.axis={a.uop.axis} b.uop.axis={b.uop.axis}")
+  batch, M, K = (1, *a.shape) if a.ndim == 2 else a.shape
+  N = b.shape[1]
+  if isinstance(a.device, tuple):
+    batch //= len(a.device)
+    dname = a.device[0]
+  else: dname = a.device
+  arch = getattr(Device[dname].renderer, "arch", "")
+  if batch not in {1, 2}: return todo(f"GEMM batch size {batch}")
+  if (key:=(M, N, K)) not in GEMM_ARGS and arch == "gfx950": return todo(f"GEMM shape not supported {key} on {arch}")
+  return True
+
+# ** UOp gemm to test Tensor.custom_kernel multi and backward correctness on non cdna4
+# note: this can be removed after we have GEMM on mixins
+
+def custom_uop_gemm(C:UOp, A:UOp, B:UOp) -> UOp:
+  M, K = A.shape[0]*A.shape[1], A.shape[2]
+  K2, N = B.shape[(1 if B.ndim == 3 else 0):]
+  assert K == K2
+  m = UOp.range(M, 1, AxisType.LOOP)
+  n = UOp.range(N, 2, AxisType.LOOP)
+  k = UOp.range(K, 0, AxisType.REDUCE)
+  mul = (A.index((m*UOp.const(dtypes.index, K)+k))*B.index((k*UOp.const(dtypes.index, N)+n))).cast(dtypes.float32)
+  red = mul.reduce(k, arg=Ops.ADD, dtype=dtypes.float32).cast(C.dtype.base)
+  store = C.index((m*UOp.const(dtypes.index, N)+n), ptr=True).store(red).end(m, n)
+  return store.sink(arg=KernelInfo(name=f'uop_gemm_{M}_{N}_{K}'))
+
+# ** backward gemm, might use the asm gemm
+
+def custom_gemm_bw(gradient:UOp, kernel:UOp):
+  out, a, b = kernel.src
+  assert all_same([gradient.device, a.device, b.device, out.device])
+  a_t, b_t, g_t = Tensor(a, device=a.device), Tensor(b, device=a.device), Tensor(gradient, device=a.device)
+  grad_a = (g_t @ b_t.T).uop
+  grad_b = (a_t.permute(2, 0, 1).reshape(a_t.shape[2], -1) @ g_t.reshape(-1, g_t.shape[-1])).uop
+  return (None, grad_a, grad_b)
+
+# ** main gemm function
+
+def asm_gemm(a:Tensor, b:Tensor) -> Tensor:
+  assert can_use_asm_gemm(a, b), f"{counters['todos'][-1]}"
+  counters["used"] += 1
+  squeeze = a.ndim == 2
+  if squeeze: a = a.unsqueeze(0)
+
+  batch, M, K = a.shape
+  N = b.shape[1]
+  is_multi = isinstance(a.device, tuple)
+
+  if is_multi:
+    out = Tensor(Tensor.empty(batch//len(a.device), M, N, dtype=a.dtype, device=a.device).uop.multi(0), device=a.device)
+  else:
+    out = Tensor.empty(batch, M, N, dtype=a.dtype, device=a.device)
+
+  dname = a.device[0] if is_multi else a.device
+  arch = getattr(Device[dname].renderer, "arch", "")
+  if arch.startswith("gfx950") and getenv("USE_ASM", 1):
+    numWG = GEMM_ARGS[(M, N, K)][0]
+    out = Tensor.custom_kernel(out, a, b, fxn=functools.partial(custom_asm_gemm, dname=dname, wg=numWG, arch=arch), grad_fxn=custom_gemm_bw)[0]
+  else:
+    out = Tensor.custom_kernel(out, a, b, fxn=custom_uop_gemm, grad_fxn=custom_gemm_bw)[0]
+  return out.squeeze(0) if squeeze else out
@@ -1,78 +0,0 @@
-.text
-.section	.text.
-.global	gemm
-.p2align	8
-.type	gemm,@function
-
-gemm:
-INSTRUCTIONS
-
-.section .rodata,"a",@progbits
-.p2align 6, 0x0
-.amdhsa_kernel gemm
-  # basic memory requirements
-  .amdhsa_group_segment_fixed_size 133120
-  .amdhsa_private_segment_fixed_size 0
-  .amdhsa_kernarg_size 28
-  # register usage (RSRC1)
-  .amdhsa_next_free_vgpr 504
-  .amdhsa_next_free_sgpr 96
-  # workgroup / workitem IDs (RSRC2)
-  .amdhsa_system_sgpr_workgroup_id_x 1
-  .amdhsa_system_sgpr_workgroup_id_y 1
-  .amdhsa_system_sgpr_workgroup_id_z 1
-  # user SGPRs, we only specify the kernel args ptr in s[0:1]
-  .amdhsa_user_sgpr_kernarg_segment_ptr 1
-  .amdhsa_user_sgpr_count 2
-  .amdhsa_user_sgpr_kernarg_preload_length 0
-  .amdhsa_user_sgpr_kernarg_preload_offset 0
-  # gfx90a / gfx940 specifics (RSRC3)
-  .amdhsa_accum_offset 248
-  .amdhsa_uses_dynamic_stack 0
-  .amdhsa_tg_split 0
-.end_amdhsa_kernel
-
-.amdgpu_metadata
---
-amdhsa.kernels:
-  - .name: gemm
-    .symbol: gemm.kd
-    .args:
-      - .name: C
-        .address_space: global
-        .offset: 0
-        .size: 8
-        .value_kind: global_buffer
-        .value_type: bf16
-      - .name: B
-        .address_space: global
-        .offset: 8
-        .size: 8
-        .value_kind: global_buffer
-        .value_type: bf16
-      - .name: A
-        .address_space: global
-        .offset: 16
-        .size: 8
-        .value_kind: global_buffer
-        .value_type: bf16
-      - .name: sz
-        .offset: 24
-        .size: 4
-        .value_kind: by_value
-        .value_type: u32
-    .group_segment_fixed_size: 133120
-    .private_segment_fixed_size: 0
-    .kernarg_segment_align: 8
-    .kernarg_segment_size: 28
-    .max_flat_workgroup_size: 256
-    .sgpr_count: 88
-    .sgpr_spill_count: 0
-    .vgpr_count: 248
-    .vgpr_spill_count: 0
-    .wavefront_size: 64
-amdhsa.version:
-  - 1
-  - 0
-...
-.end_amdgpu_metadata
@@ -1,73 +0,0 @@
-# Run assembly on the AMD runtime and check correctness
-# VIZ=2 to profile
-import pathlib
-from tinygrad import Tensor, Device, dtypes, Context
-from tinygrad.uop.ops import UOp, Ops, KernelInfo
-from tinygrad.engine.realize import Estimates
-from tinygrad.helpers import getenv
-
-fp = pathlib.Path(__file__).parent/"gemm.s"
-
-N = getenv("N", 8192)
-THREADS_PER_WG = 256
-NUM_WG = N//THREADS_PER_WG * N//THREADS_PER_WG
-
-assert N % THREADS_PER_WG == 0, "N must be divisible by THREADS_PER_WG"
-
-# ** generate inputs on CPU
-
-scale = 10.0
-
-import torch
-torch.manual_seed(0)
-A = (torch.randn(N, N, dtype=torch.float32, device="cpu") / scale).to(torch.bfloat16).contiguous()
-B = (torch.randn(N, N, dtype=torch.float32, device="cpu") / scale).to(torch.bfloat16).contiguous()
-Bt = B.t().contiguous() # transpose B for the asm gemm
-C_torch = A@B
-
-# ** copy buffers to AMD
-
-# input creation and validation run on the copy engine for simpler tracing
-
-def from_torch(t:torch.Tensor) -> Tensor:
-  return Tensor.from_blob(t.data_ptr(), t.shape, dtype=dtypes.bfloat16, device="cpu").to(Device.DEFAULT).realize()
-
-C_tiny = from_torch(A) @ from_torch(B)
-C_asm = Tensor.empty_like(C_tiny)
-
-# ** assembly custom kernel
-
-def custom_asm_gemm(C:UOp, A:UOp, B:UOp) -> UOp:
-  lidx = UOp.special(THREADS_PER_WG, "lidx0")
-  gidx = UOp.special(NUM_WG, "gidx0")
-
-  src = (pathlib.Path(__file__).parent/"template.s").read_text().replace("INSTRUCTIONS", fp.read_text())
-
-  sz = UOp.variable("SZ", 256, 8192)
-
-  sink = UOp.sink(C.base, A.base, B.base, sz, lidx, gidx, arg=KernelInfo(name="gemm", estimates=Estimates(ops=N*N*N*2, mem=N*N*4*3)))
-  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=Device.DEFAULT), UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=src)))
-
-C_asm = Tensor.custom_kernel(C_asm, from_torch(A), from_torch(Bt), fxn=custom_asm_gemm)[0]
-
-# ** run gemms
-
-sched = Tensor.schedule(C_tiny, C_asm)
-eis = [si.lower() for si in sched]
-
-with Context(DEBUG=2):
-  for ei in eis:
-    et = ei.run({"SZ":N}, wait=True)
-    print(f"{(N*N*N*2 / et)*1e-12:.2f} REAL TFLOPS")
-
-# ** correctness
-
-import ctypes
-
-def torch_bf16(t:Tensor) -> torch.tensor:
-  asm_out = t.to("cpu").realize().uop.buffer._buf
-  buf = (ctypes.c_uint16*C_asm.uop.size).from_address(asm_out.va_addr)
-  return torch.frombuffer(buf, dtype=torch.bfloat16, count=C_asm.uop.size).reshape(C_asm.shape)
-
-assert torch.allclose(torch_bf16(C_asm), C_torch, rtol=1e-2, atol=1e-3)
-assert torch.allclose(torch_bf16(C_tiny), C_torch, rtol=1e-2, atol=1e-3)
@@ -37,7 +37,7 @@ b.copyin(row.data)
 c.copyin(mat.data)
 ret = prog(a._buf, b._buf, c._buf, global_size=[1,1,1], local_size=[8,1,1], wait=True)
 print(ret)
-out = np.frombuffer(a.as_buffer(), np.float32)
+out = np.frombuffer(a.as_memoryview(), np.float32)
 real = row.astype(np.float32)@mat.T.astype(np.float32)
 print("out:", out)
 print("real", real)
@@ -98,10 +98,10 @@ if __name__ == "__main__":
  # check correctness
  if getenv("VERIFY"):
    from tinygrad.engine.realize import run_schedule
-    triton_buf = np.frombuffer(si.bufs[0].as_buffer(), np.float16).reshape(M,N)
+    triton_buf = np.frombuffer(si.bufs[0].as_memoryview(), np.float16).reshape(M,N)
    print(triton_buf)
    run_schedule(sched)
-    tinygrad_buf = np.frombuffer(si.bufs[0].as_buffer(), np.float16).reshape(M,N)
+    tinygrad_buf = np.frombuffer(si.bufs[0].as_memoryview(), np.float16).reshape(M,N)
    print(tinygrad_buf)
    np.testing.assert_allclose(triton_buf, tinygrad_buf)
    print("correct!")
@@ -1,14 +1,15 @@
 #!/usr/bin/env python3

 import argparse, glob, os, time, subprocess, sys
+from tinygrad.helpers import temp

 def scan_devs_based_on_lock(prefix:str, args) -> list[str]:
  target_dev = args.pci_bus if 'pci_bus' in args.__dir__() else ""

  devs = []
-  for dev in glob.glob(f'/tmp/{prefix}_*.lock'):
-    dev_id = dev[8:-5]
-    if os.path.exists(f"/sys/bus/pci/devices/{dev_id}") and dev_id.startswith(target_dev): devs.append(dev_id)
+  for dev in glob.glob(temp(f'{prefix}_*.lock')):
+    dev_id = dev.split('/')[-1][len(prefix)+1:-5]
+    if dev_id.startswith(target_dev): devs.append(dev_id)
  return devs

 def _do_reset_device(pci_bus): os.system(f"sudo sh -c 'echo 1 > /sys/bus/pci/devices/{pci_bus}/reset'")
@@ -53,16 +54,7 @@ def cmd_show_pids(args):

  for dev in devs:
    try:
-      pid = subprocess.check_output(['sudo', 'lsof', f'/tmp/{prefix}_{dev}.lock']).decode('utf-8').strip().split('\n')[1].split()[1]
-      print(f"{dev}: {pid}")
-    except subprocess.CalledProcessError: print(f"{dev}: No processes found using this device")
-
-def cmd_kill_pids(args):
-  devs = scan_devs_based_on_lock(prefix:={"amd":"am", "nv":"nv"}[args.backend], args)
-
-  for dev in devs:
-    try:
-      pid = subprocess.check_output(['sudo', 'lsof', f'/tmp/{prefix}_{dev}.lock']).decode('utf-8').strip().split('\n')[1].split()[1]
+      pid = subprocess.check_output(['sudo', 'lsof', temp(f'{prefix}_{dev}.lock')]).decode('utf-8').strip().split('\n')[1].split()[1]
      print(f"{dev}: {pid}")
    except subprocess.CalledProcessError: print(f"{dev}: No processes found using this device")

@@ -74,7 +66,7 @@ def cmd_kill_pids(args):
      if i > 0: time.sleep(0.2)

      try:
-        try: pid = subprocess.check_output(['sudo', 'lsof', f'/tmp/{prefix}_{dev}.lock']).decode('utf-8').strip().split('\n')[1].split()[1]
+        try: pid = subprocess.check_output(['sudo', 'lsof', temp(f'{prefix}_{dev}.lock')]).decode('utf-8').strip().split('\n')[1].split()[1]
        except subprocess.CalledProcessError: break

        print(f"Killing process {pid} (which uses {dev})")
@@ -18,7 +18,7 @@ prg = dev.runtime("write_ones", mbin)
 prg(buf0._buf, global_size=(1,65537,1), local_size=(1,1,1), wait=True)

 import numpy as np
-def to_np(buf): return np.frombuffer(buf.as_buffer().cast(buf.dtype.base.fmt), dtype=_to_np_dtype(buf.dtype.base))
+def to_np(buf): return np.frombuffer(buf.as_memoryview().cast(buf.dtype.base.fmt), dtype=_to_np_dtype(buf.dtype.base))

 big = to_np(buf0)
 print(big)
@@ -8,14 +8,14 @@ from tinygrad.helpers import _ensure_downloads_dir
 DOWNLOADS_DIR = _ensure_downloads_dir() / "models"
 from tinygrad.helpers import tqdm

-def snapshot_download_with_retry(*, repo_id: str, allow_patterns: list[str]|tuple[str, ...]|None=None, cache_dir: str|Path|None=None,
+def snapshot_download_with_retry(*, repo_id: str, allow_patterns: list[str]|tuple[str, ...]|None=None, local_dir: str|Path|None=None,
                                 tries: int=2, **kwargs) -> Path:
  for attempt in range(tries):
    try:
      return Path(snapshot_download(
        repo_id=repo_id,
        allow_patterns=allow_patterns,
-        cache_dir=str(cache_dir) if cache_dir is not None else None,
+        local_dir=str(local_dir) if local_dir is not None else None,
        **kwargs
      ))
    except Exception as e:
@@ -144,14 +144,14 @@ class HuggingFaceONNXManager:
        root_path = snapshot_download_with_retry(
          repo_id=model_id,
          allow_patterns=allow_patterns,
-          cache_dir=str(self.models_dir)
+          local_dir=str(self.models_dir / model_id)
        )

        # Download config files (usually small)
        snapshot_download_with_retry(
          repo_id=model_id,
          allow_patterns=["*config.json"],
-          cache_dir=str(self.models_dir)
+          local_dir=str(self.models_dir / model_id)
        )

        model_data["download_path"] = str(root_path)
@@ -88,8 +88,8 @@ if __name__ == "__main__":
      # repo id
      # validates all onnx models inside repo
      repo_id = "/".join(path)
-      root_path = snapshot_download_with_retry(repo_id=repo_id, allow_patterns=["*.onnx", "*.onnx_data"], cache_dir=DOWNLOADS_DIR)
-      snapshot_download_with_retry(repo_id=repo_id, allow_patterns=["*config.json"], cache_dir=DOWNLOADS_DIR)
+      root_path = snapshot_download_with_retry(repo_id=repo_id, allow_patterns=["*.onnx", "*.onnx_data"], local_dir=DOWNLOADS_DIR / repo_id)
+      snapshot_download_with_retry(repo_id=repo_id, allow_patterns=["*config.json"], local_dir=DOWNLOADS_DIR / repo_id)
      config = get_config(root_path)
      for onnx_model in root_path.rglob("*.onnx"):
        rtol, atol = get_tolerances(onnx_model.name)
@@ -101,8 +101,8 @@ if __name__ == "__main__":
      onnx_model = path[-1]
      assert path[-1].endswith(".onnx")
      repo_id, relative_path = "/".join(path[:2]), "/".join(path[2:])
-      root_path = snapshot_download_with_retry(repo_id=repo_id, allow_patterns=[relative_path], cache_dir=DOWNLOADS_DIR)
-      snapshot_download_with_retry(repo_id=repo_id, allow_patterns=["*config.json"], cache_dir=DOWNLOADS_DIR)
+      root_path = snapshot_download_with_retry(repo_id=repo_id, allow_patterns=[relative_path], local_dir=DOWNLOADS_DIR / repo_id)
+      snapshot_download_with_retry(repo_id=repo_id, allow_patterns=["*config.json"], local_dir=DOWNLOADS_DIR / repo_id)
      config = get_config(root_path)
      rtol, atol = get_tolerances(onnx_model)
      print(f"validating {relative_path} with truncate={args.truncate}, {rtol=}, {atol=}")
@@ -5,7 +5,6 @@ from tinygrad.nn import Linear, LayerNorm, Embedding, Conv2d
 from typing import List, Optional, Union, Tuple, Dict
 from abc import ABC, abstractmethod
 from functools import lru_cache
-from PIL import Image
 import numpy as np
 import re, gzip

@@ -444,7 +443,8 @@ class OpenClipEncoder:
  # TODO:
  # Should be doable in pure tinygrad, would just require some work and verification.
  # This is very desirable since it would allow for full generation->evaluation in a single JIT call.
-  def prepare_image(self, image:Image.Image) -> Tensor:
+  def prepare_image(self, image) -> Tensor:
+    from PIL import Image
    SIZE = 224
    w, h = image.size
    scale = min(SIZE / h, SIZE / w)
@@ -61,6 +61,9 @@ class Attention:
      xq = self.q_norm(xq)
      xk = self.k_norm(xk)

+    # cast_float_to_bf16 is expensive in reduction loops, break it out
+    if x.dtype == dtypes.bfloat16: xq, xk = xq.contiguous_backward(), xk.contiguous_backward()
+
    xq = xq.reshape(xq.shape[0], xq.shape[1], self.n_heads, self.head_dim)
    xk = xk.reshape(xk.shape[0], xk.shape[1], self.n_kv_heads, self.head_dim)
    xv = xv.reshape(xv.shape[0], xv.shape[1], self.n_kv_heads, self.head_dim)
@@ -86,11 +89,14 @@ class Attention:
      assert start_pos == 0
      keys, values = xk, xv

-    keys, values = repeat_kv(keys, self.n_rep), repeat_kv(values, self.n_rep)
-    xq, keys, values = xq.transpose(1, 2), keys.transpose(1, 2), values.transpose(1, 2)
-    attn = xq.scaled_dot_product_attention(keys, values, mask).transpose(1, 2)
+    if Tensor.training:
+      xq, keys, values = xq.transpose(1, 2), keys.transpose(1, 2), values.transpose(1, 2)
+      attn = xq.scaled_dot_product_attention(keys, values, is_causal=True, enable_gqa=True).transpose(1, 2)
+    else:
+      keys, values = repeat_kv(keys, self.n_rep), repeat_kv(values, self.n_rep)
+      xq, keys, values = xq.transpose(1, 2), keys.transpose(1, 2), values.transpose(1, 2)
+      attn = xq.scaled_dot_product_attention(keys, values, mask).transpose(1, 2)
    if getenv("STUB_ATTENTION"):
-      # TODO: do we need mask?
      from tinygrad.uop.ops import UOp, KernelInfo
      def fa_custom_forward(attn:UOp, q:UOp, k:UOp, v:UOp) -> UOp:
        return UOp.sink(arg=KernelInfo(name="fa_custom_forward"))
@@ -197,7 +203,9 @@ class Transformer:
    h = self.tok_embeddings(tokens)
    freqs_cis = self.freqs_cis.cast(h.dtype)[:, start_pos:start_pos+seqlen, :, :, :]

-    mask = Tensor.full((1, 1, seqlen, start_pos+seqlen), float("-inf"), dtype=h.dtype, device=h.device).triu(start_pos+1) if seqlen > 1 else None
+    if not Tensor.training and seqlen > 1:
+      mask = Tensor.full((1, 1, seqlen, start_pos+seqlen), float("-inf"), dtype=h.dtype, device=h.device).triu(start_pos+1)
+    else: mask = None
    for layer in self.layers: h = layer(h, start_pos, freqs_cis, mask)
    logits = self.output(self.norm(h))
    if math.isnan(temperature): return logits
@@ -202,7 +202,7 @@ def ioctl(fd, request, argp):
        if s.hClass == nv_gpu.NV1_MEMORY_SYSTEM: dump_struct(get_struct(s.pAllocParms, nv_gpu.NV_MEMORY_ALLOCATION_PARAMS))
        if s.hClass == nv_gpu.GT200_DEBUGGER: dump_struct(get_struct(s.pAllocParms, nv_gpu.NV83DE_ALLOC_PARAMETERS))
        if s.hClass == nv_gpu.MAXWELL_PROFILER_DEVICE: dump_struct(get_struct(s.pAllocParms, nv_gpu.NVB2CC_ALLOC_PARAMETERS))
-        if s.hClass == nv_gpu.AMPERE_CHANNEL_GPFIFO_A:
+        if s.hClass in {nv_gpu.AMPERE_CHANNEL_GPFIFO_A, nv_gpu.BLACKWELL_CHANNEL_GPFIFO_A}:
          sx = get_struct(s.pAllocParms, nv_gpu.NV_CHANNELGPFIFO_ALLOCATION_PARAMETERS)
          dump_struct(sx)
          gpus_fifo.append((sx.gpFifoOffset, sx.gpFifoEntries))
@@ -0,0 +1,193 @@
+#!/usr/bin/env python3
+from __future__ import annotations
+import enum, collections
+from typing import Iterator
+from tinygrad.helpers import colored
+from extra.assembly.amd.sqtt import PacketType, bits
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# STALL REASONS
+# ═══════════════════════════════════════════════════════════════════════════════
+
+class StallReason(enum.IntEnum):
+  # Based on CUpti_ActivityPCSamplingStallReason
+  INVALID = 0
+  NONE = 1              # selected, selected_not_issued
+  INST_FETCH = 2        # branch_resolving, no_instructions
+  EXEC_DEPENDENCY = 3   # short_scoreboard, wait
+  MEMORY_DEPENDENCY = 4 # long_scoreboard
+  TEXTURE = 5           # tex_throttle
+  SYNC = 6              # barrier, membar
+  CONSTANT_MEMORY = 7   # imc_miss
+  PIPE_BUSY = 8         # mio_throttle, math_pipe_throttle
+  MEMORY_THROTTLE = 9   # drain, lg_throttle
+  NOT_SELECTED = 10     # not_selected
+  OTHER = 11            # misc, dispatch_stall
+  SLEEPING = 12         # sleeping
+
+STALL_KEY_MAP_AMPERE: dict[int, StallReason] = {
+  1: StallReason.MEMORY_THROTTLE, 15: StallReason.MEMORY_THROTTLE,
+  2: StallReason.CONSTANT_MEMORY,
+  3: StallReason.SYNC,
+  6: StallReason.INST_FETCH, 11: StallReason.INST_FETCH,
+  7: StallReason.EXEC_DEPENDENCY, 10: StallReason.EXEC_DEPENDENCY,
+  9: StallReason.MEMORY_DEPENDENCY,
+  12: StallReason.PIPE_BUSY,
+  17: StallReason.OTHER, 20: StallReason.OTHER,
+  18: StallReason.NONE,
+}
+
+STALL_KEY_MAP_BLACKWELL: dict[int, StallReason] = {
+  0x01: StallReason.MEMORY_THROTTLE, 0x0e: StallReason.MEMORY_THROTTLE,
+  0x02: StallReason.SYNC,
+  0x05: StallReason.INST_FETCH, 0x0a: StallReason.INST_FETCH,
+  0x06: StallReason.EXEC_DEPENDENCY, 0x09: StallReason.EXEC_DEPENDENCY,
+  0x08: StallReason.MEMORY_DEPENDENCY,
+  0x0b: StallReason.PIPE_BUSY, 0x0f: StallReason.PIPE_BUSY,
+  0x10: StallReason.OTHER, 0x13: StallReason.OTHER,
+  0x11: StallReason.NONE,
+}
+
+# Lookup table for extracting sample bytes from 32-byte packet (bytes 0-3, 8-31, skipping header at 4-7)
+LOOKUP_28B = [0, 1, 2, 3, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31]
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# PACKET HEADER
+# ═══════════════════════════════════════════════════════════════════════════════
+
+class PMAHeader(PacketType):
+  num_bytes  = bits[4:0]    # number of sample bytes in this packet
+  tpc_id_lo  = bits[15:8]   # TPC identifier low 8 bits
+  tpc_id_hi  = bits[27:25]  # TPC identifier high 3 bits
+  dropped    = bits[28:28]  # dropped flag (resets byte accumulator)
+  @property
+  def tpc_id(self) -> int: return self.tpc_id_lo | (self.tpc_id_hi << 8)
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# 8-BYTE SAMPLE FORMAT (Ampere/Ada/Hopper)
+# ═══════════════════════════════════════════════════════════════════════════════
+
+class PMASampleAmpere8B(PacketType):
+  pc_raw     = bits[44:0]   # raw PC value (pc_offset = pc_raw << 4)
+  stall_key  = bits[49:45]  # stall reason key
+  wave_id    = bits[55:50]  # warp/wave identifier
+  active     = bits[62:62]  # 1 if warp was executing, 0 if scheduled but not issued
+  @property
+  def pc_offset(self) -> int: return self.pc_raw << 4
+  @property
+  def stall_reason(self) -> StallReason: return STALL_KEY_MAP_AMPERE.get(self.stall_key, StallReason.OTHER)
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# 9-BYTE SAMPLE FORMAT (Blackwell+)
+# ═══════════════════════════════════════════════════════════════════════════════
+
+class PMASampleBlackwell9B(PacketType):
+  stall_key  = bits[5:0]    # stall reason key
+  pc_raw     = bits[60:8]   # raw PC value (pc_offset = pc_raw << 4)
+  wave_hi    = bits[7:6]    # wave_id high 2 bits
+  wave_lo    = bits[71:68]  # wave_id low 4 bits
+  active     = bits[67:67]  # 1 if warp was executing, 0 if scheduled but not issued
+  @property
+  def pc_offset(self) -> int: return self.pc_raw << 4
+  @property
+  def stall_reason(self) -> StallReason: return STALL_KEY_MAP_BLACKWELL.get(self.stall_key, StallReason.OTHER)
+  @property
+  def wave_id(self) -> int: return (self.wave_hi << 4) | self.wave_lo
+
+PMASample = PMASampleAmpere8B|PMASampleBlackwell9B
+
+def decode(data: bytes, sm_version: int = 0x800) -> Iterator[tuple[PMASample, int]]:
+  use_9byte = sm_version >= 0xa04
+  record_size = 9 if use_9byte else 8
+  sample_cls = PMASampleBlackwell9B if use_9byte else PMASampleAmpere8B
+
+  tpc_state: dict[int, list[int]] = collections.defaultdict(list)
+  for pkt_idx in range(len(data) // 32):
+    pkt = data[pkt_idx * 32:(pkt_idx + 1) * 32]
+    hdr = PMAHeader.from_raw(int.from_bytes(pkt[4:8], 'little'))
+
+    if hdr.dropped: tpc_state[hdr.tpc_id].clear()
+
+    for i in range(hdr.num_bytes):
+      tpc_state[hdr.tpc_id].append(pkt[LOOKUP_28B[i]])
+
+    while len(tpc_state[hdr.tpc_id]) >= record_size:
+      yield sample_cls.from_raw(int.from_bytes(bytes(tpc_state[hdr.tpc_id][:record_size]), 'little')), hdr.tpc_id
+      del tpc_state[hdr.tpc_id][:record_size]
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# CLI
+# ═══════════════════════════════════════════════════════════════════════════════
+
+STALL_COLORS = {
+  StallReason.NONE: "green", StallReason.INST_FETCH: "yellow", StallReason.EXEC_DEPENDENCY: "cyan",
+  StallReason.MEMORY_DEPENDENCY: "red", StallReason.SYNC: "magenta", StallReason.CONSTANT_MEMORY: "blue",
+  StallReason.PIPE_BUSY: "yellow", StallReason.MEMORY_THROTTLE: "RED", StallReason.OTHER: "white",
+}
+
+def decode_tpc_id(tpc_id:int) -> tuple[int, int, int]:
+  # NOTE: valid only for ops_nv, cuda encoding is different
+  return (tpc_id >> 5, (tpc_id >> 1) & 0xf, tpc_id & 1)
+
+def print_samples(samples:list[tuple[PMASample, int]]) -> None:
+  if not samples: return
+  base_pc = min(s.pc_offset for s, _ in samples)
+  for s, tpc_id in samples:
+    gpc, tpc, sm = decode_tpc_id(tpc_id)
+    stall_str = colored(f"{s.stall_reason.name:17}", STALL_COLORS.get(s.stall_reason, "white"))
+    print(f"pc=0x{s.pc_offset - base_pc:06x} {stall_str} ev={s.stall_key:2d} active={s.active} wave={s.wave_id:2d} gpc={gpc} tpc={tpc} sm={sm}")
+
+def print_packets(data:bytes, sm_version:int=0x800) -> None:
+  record_size = 9 if sm_version >= 0x890 else 8
+  tpc_state: dict[int, list[int]] = collections.defaultdict(list)
+  for i in range(len(data) // 32):
+    pkt = data[i * 32:(i + 1) * 32]
+    hdr = PMAHeader.from_raw(int.from_bytes(pkt[4:8], 'little'))
+    if hdr.dropped: tpc_state[hdr.tpc_id].clear()
+    for j in range(hdr.num_bytes): tpc_state[hdr.tpc_id].append(pkt[LOOKUP_28B[j]])
+    # Show complete records extracted from this packet
+    records = []
+    while len(tpc_state[hdr.tpc_id]) >= record_size:
+      records.append(bytes(tpc_state[hdr.tpc_id][:record_size]).hex())
+      del tpc_state[hdr.tpc_id][:record_size]
+    leftover = len(tpc_state[hdr.tpc_id])
+    print(f"Pkt {i:3d}: tpc={hdr.tpc_id:4d} n={hdr.num_bytes:2d} drop={hdr.dropped} left={leftover} | {' '.join(records)}")
+
+def print_aggregated(samples:list[tuple[PMASample, int]]) -> None:
+  if not samples: return
+  base_pc = min(s.pc_offset for s, _ in samples)
+  counter: collections.Counter[tuple[int, StallReason]] = collections.Counter((s.pc_offset, s.stall_reason) for s, _ in samples)
+  print(f"\nAggregated samples (base_pc=0x{base_pc:x}):")
+  for (pc, reason), cnt in sorted(counter.items()):
+    stall_str = colored(f"{reason.name:17}", STALL_COLORS.get(reason, "white"))
+    print(f"  pc=0x{pc - base_pc:06x} {stall_str} samples={cnt:4d}")
+
+if __name__ == "__main__":
+  import sys, pickle
+
+  if len(sys.argv) < 2:
+    print("Usage: python decode.py <pkl_file> [--raw] [--sm=0xNNN]")
+    sys.exit(1)
+
+  with open(sys.argv[1], "rb") as f:
+    data = pickle.load(f)
+
+  if isinstance(data, dict):
+    sm_version = 0x800  # default to Ampere
+    for arg in sys.argv:
+      if arg.startswith("--sm="): sm_version = int(arg[5:], 0)
+    dumps = [(i, x, sm_version) for i, x in enumerate(data["pma_raw_dumps"])]
+  else:
+    devs = {e.device: e for e in data if type(e).__name__ == "ProfileDeviceEvent"}
+    dumps = []
+    for i, e in enumerate(e for e in data if type(e).__name__ == "ProfilePMAEvent"):
+      dumps.append((i, e.blob, devs[e.device].props.get('sm_version', 0x800)))
+
+  for dump_idx, raw, sm_ver in dumps:
+    print(f"\n{'='*60}\nDump {dump_idx} ({len(raw)} bytes, {len(raw)//32} packets)\n{'='*60}")
+    if "--raw" in sys.argv: print_packets(raw, sm_ver)
+    else:
+      samples = list(decode(raw, sm_ver))
+      print(f"\nDecoded {len(samples)} samples:")
+      print_samples(samples)
+      print_aggregated(samples)
@@ -0,0 +1,76 @@
+import pickle, unittest
+from collections import Counter
+from pathlib import Path
+
+from extra.nv_pma.decode import decode
+from tinygrad.helpers import DEBUG
+
+EXAMPLES_DIR = Path(__file__).parent.parent / "examples"
+EXAMPLES_5090_DIR = Path(__file__).parent.parent / "examples_5090"
+
+def decode_and_aggregate(raw_dumps: list[bytes], sm_version: int = 0x800) -> Counter[tuple[int, int]]:
+  """Decode all PMA buffers and aggregate by (relative_pc, stall_reason). Each dump is normalized separately."""
+  result: Counter[tuple[int, int]] = Counter()
+  for raw in raw_dumps:
+    samples = [s for s, _ in decode(raw, sm_version)]
+    if not samples: continue
+    base_pc = min(s.pc_offset for s in samples)
+    result += Counter((s.pc_offset - base_pc, int(s.stall_reason)) for s in samples)
+  return result
+
+def cupti_to_counter(cupti_records: list[dict]) -> Counter[tuple[int, int]]:
+  """Convert CUPTI records to Counter[(pcOffset, stallReason)]."""
+  counter: Counter[tuple[int, int]] = Counter()
+  for r in cupti_records:
+    counter[(r['pcOffset'], r['stallReason'])] += r['samples']
+  return counter
+
+class TestNVProf(unittest.TestCase):
+  def _test_example(self, name: str, sm_version: int = 0x800, examples_dir: Path = EXAMPLES_DIR):
+    pkl_file = examples_dir / f"{name}.pkl"
+    if not pkl_file.exists():
+      self.skipTest(f"Example data not found: {pkl_file}. Run collect.py first.")
+
+    with open(pkl_file, "rb") as f:
+      data = pickle.load(f)
+
+    self.assertEqual(data["test_name"], name)
+    pma_agg = decode_and_aggregate(data["pma_raw_dumps"], sm_version)
+    cupti_agg = cupti_to_counter(data["cupti_pc_samples"])
+
+    if DEBUG >= 2:
+      total = sum(cupti_agg.values())
+      mismatched = sum(abs(pma_agg.get(k, 0) - v) for k, v in cupti_agg.items())
+      mismatched += sum(v for k, v in pma_agg.items() if k not in cupti_agg)
+      mismatched //= 2
+
+      print(f"\n=== Test: {name} ===")
+      print(f"Total samples: {total}, Mismatched: {mismatched} ({mismatched/total*100 if total else 0:.1f}%)")
+
+    self.assertEqual(pma_agg, cupti_agg, f"PMA: {dict(pma_agg)}\nCUPTI: {dict(cupti_agg)}")
+
+  # Ampere tests (8-byte format)
+  def test_decode_test_plus(self): self._test_example("test_plus")
+  def test_decode_test_reduce_sum(self): self._test_example("test_reduce_sum")
+  def test_decode_test_broadcast(self): self._test_example("test_broadcast")
+  def test_decode_test_matmul(self): self._test_example("test_matmul")
+  def test_decode_test_plus_big(self): self._test_example("test_plus_big")
+  def test_decode_test_elementwise_chain(self): self._test_example("test_elementwise_chain")
+  def test_decode_test_conv2d(self): self._test_example("test_conv2d")
+  def test_decode_test_large_matmul(self): self._test_example("test_large_matmul")
+
+  # Blackwell/5090 tests (9-byte format)
+  def test_5090_test_plus(self): self._test_example("test_plus", 0xa04, EXAMPLES_5090_DIR)
+  def test_5090_test_plus_big(self): self._test_example("test_plus_big", 0xa04, EXAMPLES_5090_DIR)
+  def test_5090_test_broadcast(self): self._test_example("test_broadcast", 0xa04, EXAMPLES_5090_DIR)
+  def test_5090_test_matmul(self): self._test_example("test_matmul", 0xa04, EXAMPLES_5090_DIR)
+  def test_5090_test_large_matmul(self): self._test_example("test_large_matmul", 0xa04, EXAMPLES_5090_DIR)
+  def test_5090_test_reduce_sum(self): self._test_example("test_reduce_sum", 0xa04, EXAMPLES_5090_DIR)
+  def test_5090_test_reduce_max(self): self._test_example("test_reduce_max", 0xa04, EXAMPLES_5090_DIR)
+  def test_5090_test_elementwise_chain(self): self._test_example("test_elementwise_chain", 0xa04, EXAMPLES_5090_DIR)
+  def test_5090_test_conv2d(self): self._test_example("test_conv2d", 0xa04, EXAMPLES_5090_DIR)
+  def test_5090_test_exp(self): self._test_example("test_exp", 0xa04, EXAMPLES_5090_DIR)
+  def test_5090_test_softmax(self): self._test_example("test_softmax", 0xa04, EXAMPLES_5090_DIR)
+
+if __name__ == "__main__":
+  unittest.main()
@@ -7,7 +7,7 @@ export CAPTURE_PROCESS_REPLAY=1
 rm "$LOGOPS" 2>/dev/null || true
 test/external/process_replay/reset.py

-CI=1 python3 -m pytest -n=auto test/test_ops.py test/test_nn.py test/unit/test_winograd.py test/models/test_real_world.py --durations=20
+CI=1 python3 -m pytest -n=auto test/test_ops.py test/test_nn.py test/unit/test_winograd.py test/null/test_real_world.py --durations=20
 CL=1 python3 -m pytest test/test_tiny.py

 # extract, sort and uniq
@@ -1,31 +0,0 @@
-import random
-from extra.optimization.helpers import load_worlds, ast_str_to_lin
-from tinygrad.codegen.opt.search import actions
-from tinygrad.codegen.opt.kernel import Kernel
-from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
-from tinygrad.helpers import tqdm
-
-tactions = set()
-def test_rebuild(lin):
-  linr = Kernel(lin.ast)
-  for o in lin.applied_opts:
-    assert o in actions, f"{o} is not in actions"
-    tactions.add(o)
-    linr.apply_opt(o)
-
-  assert len(lin.sts) == len(linr.sts)
-  for st1,st2 in zip(lin.sts, linr.sts):
-    assert st1 == st2, f"{st1} != {st2}"
-
-if __name__ == "__main__":
-  ast_strs = load_worlds(False, False, False)
-  random.shuffle(ast_strs)
-  ast_strs = ast_strs[:2000]
-  for ast_str in tqdm(ast_strs):
-    lin = ast_str_to_lin(ast_str)
-    #if not lin.apply_tensor_cores():
-    lin.apply_opts(hand_coded_optimizations(lin))
-    test_rebuild(lin)
-
-  print(len(tactions), len(actions))
-  print(sorted(list(tactions)))
@@ -1,76 +0,0 @@
-import os
-import numpy as np
-import math, random
-from tinygrad.tensor import Tensor
-from tinygrad.nn.state import get_parameters, get_state_dict, safe_save, safe_load, load_state_dict
-from tinygrad.codegen.opt.search import actions, bufs_from_lin, get_kernel_actions
-from tinygrad.nn.optim import Adam
-from extra.optimization.extract_policynet import PolicyNet
-from extra.optimization.helpers import load_worlds, ast_str_to_lin, lin_to_feats, time_linearizer
-
-if __name__ == "__main__":
-  net = PolicyNet()
-  if os.path.isfile("/tmp/policynet.safetensors"): load_state_dict(net, safe_load("/tmp/policynet.safetensors"))
-  optim = Adam(get_parameters(net))
-
-  ast_strs = load_worlds()
-
-  # select a world
-  all_feats, all_acts, all_rews = [], [], []
-  while 1:
-    Tensor.training = False
-    lin = ast_str_to_lin(random.choice(ast_strs))
-    rawbufs = bufs_from_lin(lin)
-    tm = last_tm = base_tm = time_linearizer(lin, rawbufs)
-
-    # take actions
-    feats, acts, rews = [], [], []
-    while 1:
-      feat = lin_to_feats(lin)
-      feats.append(feat)
-      probs = net(Tensor([feat])).exp()[0].numpy()
-
-      # mask valid actions
-      valid_action_mask = np.zeros((len(actions)+1), dtype=np.float32)
-      for x in get_kernel_actions(lin): valid_action_mask[x] = 1
-      probs *= valid_action_mask
-      probs /= sum(probs)
-
-      act = np.random.choice(len(probs), p=probs)
-      acts.append(act)
-      if act == 0:
-        rews.append(0)
-        break
-      try:
-        lin.apply_opt(actions[act-1])
-        tm = time_linearizer(lin, rawbufs)
-        if math.isinf(tm): raise Exception("failed")
-        rews.append(((last_tm-tm)/base_tm))
-        last_tm = tm
-      except Exception:
-        rews.append(-0.5)
-        break
-      #print(f"{tm*1e6:10.2f}", lin.colored_shape())
-
-    assert len(feats) == len(acts) and len(acts) == len(rews)
-    #print(rews)
-    print(f"***** EPISODE {len(rews)} steps, {sum(rews):5.2f} reward, {base_tm*1e6:12.2f} -> {tm*1e6:12.2f} : {lin.colored_shape()}")
-    all_feats += feats
-    all_acts += acts
-    # rewards to go
-    for i in range(len(rews)-2, -1, -1): rews[i] += rews[i+1]
-    all_rews += rews
-
-    BS = 32
-    if len(all_feats) >= BS:
-      Tensor.training = True
-      x = Tensor(all_feats[:BS])
-      mask = np.zeros((BS, len(actions)+1), dtype=np.float32)
-      mask[range(BS), all_acts[:BS]] = all_rews[:BS]
-      loss = -(net(x) * Tensor(mask)).mean()
-      optim.zero_grad()
-      loss.backward()
-      optim.step()
-      all_feats = all_feats[BS:]
-      all_acts = all_acts[BS:]
-      all_rews = all_rews[BS:]
@@ -1,32 +0,0 @@
-from typing import List, Tuple
-from tinygrad.codegen.opt.kernel import Kernel
-from tinygrad.codegen.opt.search import get_kernel_actions, actions
-
-_net = None
-def beam_q_estimate(beam:List[Tuple[Kernel, float]]) -> List[Tuple[Kernel, float]]:
-  global _net
-  if _net is None:
-    from tinygrad.nn.state import load_state_dict, safe_load
-    from extra.optimization.pretrain_valuenet import ValueNet
-    _net = ValueNet(1021+len(actions), 2)
-    load_state_dict(_net, safe_load("/tmp/qnet.safetensors"), verbose=False)
-  from tinygrad.tensor import Tensor
-  from tinygrad.helpers import Context
-  from extra.optimization.helpers import lin_to_feats
-  import numpy as np
-  feats = []
-  lins = []
-  base_tms = []
-  for lin,tm in beam:
-    lin_feats = lin_to_feats(lin)
-    for a,v in get_kernel_actions(lin, include_0=False).items():
-      acts = np.zeros(len(actions))
-      acts[a-1] = 1.0
-      feats.append(np.concatenate([lin_feats, acts]))
-      lins.append(v)
-      base_tms.append(tm)
-  with Context(BEAM=0):
-    with Tensor.train(False):
-      preds = _net(Tensor(feats)).numpy()
-  pred_time = np.array(base_tms) / np.exp(preds[:, 0])
-  return sorted(zip(lins, pred_time), key=lambda x: x[1])
@@ -1,34 +0,0 @@
-import argparse
-from extra.optimization.helpers import ast_str_to_lin, time_linearizer
-
-from tinygrad import dtypes
-from tinygrad.helpers import BEAM, getenv
-from tinygrad.device import Device, Compiled
-from tinygrad.codegen.opt.kernel import Kernel
-from tinygrad.codegen.opt.search import beam_search, bufs_from_lin
-
-
-if __name__ == '__main__':
-  parser = argparse.ArgumentParser(description="Run a search for the optimal opts for a kernel", formatter_class=argparse.ArgumentDefaultsHelpFormatter)
-  parser.add_argument("--ast", type=str, default=None, help="the ast for the kernel to be optimized")
-  parser.add_argument("--file", type=str, default=None, help="a file containing asts to be optimized, one per line")
-  args = parser.parse_args()
-
-  device: Compiled = Device[Device.DEFAULT]
-  print(f"optimizing for {Device.DEFAULT}")
-
-  if args.ast is not None:
-    ast_strs = [args.ast]
-  elif args.file is not None:
-    with open(args.file, 'r') as file:
-      ast_strs = file.readlines()
-
-  for i, ast_str in enumerate(ast_strs):
-    print(f"optimizing {i}/{len(ast_strs)}\nast={ast_str}")
-    lin = ast_str_to_lin(ast_str, opts=device.renderer)
-    rawbufs = bufs_from_lin(lin)
-    lin = beam_search(lin, rawbufs, getenv("BEAM", 8), bool(getenv("BEAM_ESTIMATE", 1)))
-
-    tm = time_linearizer(lin, rawbufs, allow_test_size=False, cnt=10)
-    print(f"final time {tm*1e6:9.0f} us: {lin.colored_shape()}")
-    print(lin.applied_opts)
@@ -1,19 +0,0 @@
-import unittest
-
-from extra.optimization.helpers import load_worlds
-
-class TestKernelDataset(unittest.TestCase):
-  def test_load_worlds_filters(self):
-    all_kernels = load_worlds(filter_reduce=False, filter_noimage=False, filter_novariable=False)
-
-    reduce_kernels = load_worlds(filter_reduce=True, filter_noimage=False, filter_novariable=False)
-    self.assertGreater(len(all_kernels), len(reduce_kernels))
-
-    image_kernels = load_worlds(filter_reduce=False, filter_noimage=True, filter_novariable=False)
-    self.assertGreater(len(all_kernels), len(image_kernels))
-
-    variable_kernels = load_worlds(filter_reduce=False, filter_noimage=False, filter_novariable=True)
-    self.assertGreater(len(all_kernels), len(variable_kernels))
-
-if __name__ == '__main__':
-  unittest.main()
@@ -1,67 +0,0 @@
-import numpy as np
-import math
-import random
-np.set_printoptions(suppress=True)
-from copy import deepcopy
-from tinygrad.helpers import getenv, colored
-from tinygrad.tensor import Tensor
-from tinygrad.nn.state import get_parameters, get_state_dict, safe_save, safe_load, load_state_dict
-from tinygrad.codegen.opt.search import bufs_from_lin, actions, get_kernel_actions
-from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
-from extra.optimization.helpers import load_worlds, ast_str_to_lin, lin_to_feats, time_linearizer
-from extra.optimization.extract_policynet import PolicyNet
-from extra.optimization.pretrain_valuenet import ValueNet
-
-VALUE = getenv("VALUE")
-
-if __name__ == "__main__":
-  if VALUE:
-    net = ValueNet()
-    load_state_dict(net, safe_load("/tmp/valuenet.safetensors"))
-  else:
-    net = PolicyNet()
-    load_state_dict(net, safe_load("/tmp/policynet.safetensors"))
-
-  ast_strs = load_worlds()
-
-  # real randomness
-  random.seed()
-  random.shuffle(ast_strs)
-
-  wins = 0
-  for ep_num,ast_str in enumerate(ast_strs):
-    print("\nEPISODE", ep_num, f"win {wins*100/max(1,ep_num):.2f}%")
-    lin = ast_str_to_lin(ast_str)
-    rawbufs = bufs_from_lin(lin)
-
-    linhc = deepcopy(lin)
-    linhc.applied_opts(hand_coded_optimizations(linhc))
-    tmhc = time_linearizer(linhc, rawbufs)
-    print(f"{tmhc*1e6:10.2f}     HC    ", linhc.colored_shape())
-
-    pred_time = float('nan')
-    tm = float('inf')
-    while 1:
-      if VALUE:
-        acts,feats = [], []
-        for k,v in get_kernel_actions(lin).items():
-          acts.append(k)
-          feats.append(lin_to_feats(v))
-        preds = net(Tensor(feats))
-        pred_time = math.exp(preds.numpy().min())
-        act = acts[preds.numpy().argmin()]
-      else:
-        probs = net(Tensor([lin_to_feats(lin)]))
-        dist = probs.exp().numpy()
-        act = dist.argmax()
-      if act == 0: break
-      try:
-        lin.apply_opt(actions[act-1])
-      except Exception:
-        print("FAILED")
-        break
-      tm = time_linearizer(lin, rawbufs)
-      print(f"{tm*1e6:10.2f} {pred_time*1e6:10.2f}", lin.colored_shape())
-
-    print(f"{colored('BEAT', 'green') if tm < tmhc else colored('lost', 'red')} hand coded {tmhc/tm:5.2f}x")
-    wins += int(tm < tmhc)
@@ -1,21 +0,0 @@
-from extra.optimization.helpers import load_worlds, ast_str_to_lin, time_linearizer
-from tinygrad.codegen.opt.search import bufs_from_lin, get_kernel_actions
-
-if __name__ == "__main__":
-  ast_strs = load_worlds()
-  for i, ast_str in enumerate(ast_strs):
-    lin = ast_str_to_lin(ast_str)
-    rawbufs = bufs_from_lin(lin)
-    test_tm = time_linearizer(lin, rawbufs)
-    if test_tm < 1e-2: continue
-    print(f"EXAMPLE {i}")
-    acted_lins = get_kernel_actions(lin)
-    ok_avg, short_avg = 0, 0
-    for k,v in acted_lins.items():
-      tm1 = time_linearizer(v, rawbufs)
-      tm2 = time_linearizer(v, rawbufs)
-      tm3 = time_linearizer(v, rawbufs, False)
-      print(v.colored_shape(50), f"{tm1*1e3:10.2f} {tm2*1e3:10.2f} {tm3*1e3:10.2f}  :  {((tm1-tm2)/tm1)*100:5.2f}%  vs  {((tm1-tm3)/tm1)*100:5.2f}%")
-      ok_avg += (tm1-tm2)/tm1
-      short_avg += (tm1-tm3)/tm1
-    print(f"{ok_avg/len(acted_lins)*100:5.2f}% vs {short_avg/len(acted_lins)*100:5.2f}%")
@@ -2,15 +2,15 @@ import math

 from tinygrad import Tensor, dtypes
 from tinygrad.helpers import DEBUG
-from tinygrad.uop.ops import UOp
+from tinygrad.uop.ops import UOp, Ops

 from extra.thunder.tiny.tk import WARP_THREADS
 from extra.thunder.tiny.tk.kernel import Kernel
 from extra.thunder.tiny.tk.tiles import GL, TileLayout

 NUM_WORKERS = 1
-Q_BLOCK_SIZE = 16
-KV_BLOCK_SIZE = 16
+Q_BLOCK_SIZE = 32
+KV_BLOCK_SIZE = 32

 def _sharded_empty(shape:Tensor, ref:Tensor, axis:int|None) -> Tensor:
  if not isinstance(ref.device, tuple): return Tensor.empty(*shape, dtype=ref.dtype, device=ref.device)
@@ -43,20 +43,18 @@ def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False
  B_local = B // num_devices
  if DEBUG >= 2: print(f"Flash Attention {B=} {B_local=} {N=} {H=} {D=} {H_KV=} {GROUP_SIZE=}")

-  def custom_forward(ou:UOp, l_vecu:UOp, qu:UOp, ku:UOp, vu:UOp, masku:UOp) -> UOp:
+  def _custom_forward_impl(ou:UOp, l_vecu:UOp, qu:UOp, ku:UOp, vu:UOp, masku:UOp|None) -> UOp:
    with Kernel("fa_custom_forward", (H, N // (Q_BLOCK_SIZE*NUM_WORKERS), B_local), NUM_WORKERS * WARP_THREADS) as ker:
      warp = ker.warp

-      o, q, k, v, mask, l_vec = GL(ou, ker), GL(qu, ker), GL(ku, ker), GL(vu, ker), GL(masku, ker), GL(l_vecu, ker)
+      o, q, k, v, l_vec = GL(ou, ker), GL(qu, ker), GL(ku, ker), GL(vu, ker), GL(l_vecu, ker)
+      mask = GL(masku, ker) if masku is not None else None

      head = ker.blockIdx_x
      head_kv = head // GROUP_SIZE
      batch = ker.blockIdx_z
      q_seq = ker.blockIdx_y * NUM_WORKERS + ker.warpid

-      k_smem = ker.st((KV_BLOCK_SIZE, D), dtypes.bfloat16)
-      v_smem = ker.st((KV_BLOCK_SIZE, D), dtypes.bfloat16)
-
      q_reg_fl = ker.rt((Q_BLOCK_SIZE, D), dtypes.float32)
      q_reg = ker.rt((Q_BLOCK_SIZE, D), dtypes.bfloat16)
      q_reg_transposed = ker.rt((D, Q_BLOCK_SIZE), dtypes.bfloat16, TileLayout.COL)
@@ -70,10 +68,10 @@ def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False
      mask_reg = ker.rt((Q_BLOCK_SIZE, KV_BLOCK_SIZE), dtypes.float32)
      mask_reg_transposed = ker.rt((KV_BLOCK_SIZE, Q_BLOCK_SIZE), dtypes.float32, TileLayout.COL)

-      max_vec_last = ker.rv(KV_BLOCK_SIZE, dtypes.float32)
-      max_vec = ker.rv(KV_BLOCK_SIZE, dtypes.float32)
-      norm_vec = ker.rv(KV_BLOCK_SIZE, dtypes.float32)
-      scale_vec = ker.rv(KV_BLOCK_SIZE, dtypes.float32)
+      max_vec_last = ker.rv(Q_BLOCK_SIZE, dtypes.float32)
+      max_vec = ker.rv(Q_BLOCK_SIZE, dtypes.float32)
+      norm_vec = ker.rv(Q_BLOCK_SIZE, dtypes.float32)
+      scale_vec = ker.rv(Q_BLOCK_SIZE, dtypes.float32)

      max_vec = warp.neg_inf(max_vec)
      norm_vec = warp.zero(norm_vec)
@@ -86,12 +84,10 @@ def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False
      q_reg = warp.copy(q_reg, q_reg_fl)
      q_reg_transposed = warp.transpose(q_reg_transposed, q_reg)

-      for kv_idx in ker.range(N // KV_BLOCK_SIZE):
-        k_smem = warp.load(k_smem, k, (), (batch, kv_idx, head_kv, 0), axis=1)
-        v_smem = warp.load(v_smem, v, (), (batch, kv_idx, head_kv, 0), axis=1)
-
-        k_reg = warp.load(k_reg, k_smem)
-        v_reg = warp.load(v_reg, v_smem)
+      num_kv_blocks = (q_seq + 1) if is_causal else (N // KV_BLOCK_SIZE)
+      for kv_idx in ker.range(num_kv_blocks):
+        k_reg = warp.load(k_reg, k, (), (batch, kv_idx, head_kv, 0), axis=1)
+        v_reg = warp.load(v_reg, v, (), (batch, kv_idx, head_kv, 0), axis=1)

        # mma qk^t
        att_block = warp.zero(att_block.after(kv_idx))
@@ -99,13 +95,20 @@ def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False
        att_block = warp.mma_AtB(att_block, k_reg_transposed, q_reg_transposed)

        # apply attention mask
-        mask_reg = warp.load(mask_reg, mask, (), (batch, 0, q_seq, kv_idx), axis=2)
-        mask_reg_transposed = warp.transpose(mask_reg_transposed, mask_reg)
-        att_block += mask_reg_transposed
+        if is_causal:
+          bs_rows, bs_cols, bs_stride = att_block.base_shape.rows, att_block.base_shape.cols, att_block.base_shape.stride
+          q_base = q_seq * Q_BLOCK_SIZE + (warp.laneid % bs_cols)
+          kv_base = kv_idx * KV_BLOCK_SIZE + (warp.laneid // bs_cols) * bs_stride
+          att_block = warp.map(att_block,
+            lambda x, idx: ((kv_base + idx[0]*bs_rows + idx[2]) > (q_base + idx[1]*bs_cols)).alu(Ops.WHERE, UOp.ufix(x._uop, -math.inf), x))
+        elif mask is not None:
+          mask_reg = warp.load(mask_reg, mask, (), (batch, 0, q_seq, kv_idx), axis=2)
+          mask_reg_transposed = warp.transpose(mask_reg_transposed, mask_reg)
+          att_block += mask_reg_transposed

        # softmax
        max_vec_last = warp.copy(max_vec_last.after(kv_idx), max_vec)
-        max_vec = warp.row_reduce(max_vec.after(max_vec_last), att_block, lambda a, b: a.maximum(b), init_value=-math.inf)
+        max_vec = warp.col_reduce(max_vec.after(max_vec_last), att_block, lambda a, b: a.maximum(b), init_value=-math.inf)

        scale_vec = warp.map(scale_vec.after(max_vec_last, max_vec), lambda _, idx: max_vec_last[*idx] - max_vec[*idx])
        scale_vec = scale_vec.exp2()
@@ -116,7 +119,7 @@ def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False
        att_block -= max_vec
        att_block = att_block.exp2()

-        norm_vec = warp.row_reduce(norm_vec.after(scale_vec), att_block, lambda a, b: a + b)
+        norm_vec = warp.col_reduce(norm_vec.after(scale_vec), att_block, lambda a, b: a + b)

        # mma av
        att_block_mma = warp.copy(att_block_mma.after(kv_idx, norm_vec), att_block)
@@ -141,11 +144,18 @@ def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False

      return ker.finish()

-  def custom_backward_q(dqu:UOp, dou:UOp, qu:UOp, ku:UOp, vu:UOp, masku:UOp, l_vecu:UOp, delta_vecu:UOp) -> UOp:
+  def custom_forward_causal(ou:UOp, l_vecu:UOp, qu:UOp, ku:UOp, vu:UOp) -> UOp:
+    return _custom_forward_impl(ou, l_vecu, qu, ku, vu, None)
+
+  def custom_forward_masked(ou:UOp, l_vecu:UOp, qu:UOp, ku:UOp, vu:UOp, masku:UOp) -> UOp:
+    return _custom_forward_impl(ou, l_vecu, qu, ku, vu, masku)
+
+  def _custom_backward_q_impl(dqu:UOp, dou:UOp, qu:UOp, ku:UOp, vu:UOp, masku:UOp|None, l_vecu:UOp, delta_vecu:UOp) -> UOp:
    with Kernel("fa_custom_backward_q", (H, N // (Q_BLOCK_SIZE*NUM_WORKERS), B_local), NUM_WORKERS * WARP_THREADS) as ker:
      warp = ker.warp

-      dq, do, q, k, v, mask = GL(dqu, ker), GL(dou, ker), GL(qu, ker), GL(ku, ker), GL(vu, ker), GL(masku, ker)
+      dq, do, q, k, v = GL(dqu, ker), GL(dou, ker), GL(qu, ker), GL(ku, ker), GL(vu, ker)
+      mask = GL(masku, ker) if masku is not None else None
      l_vec, delta_vec = GL(l_vecu, ker), GL(delta_vecu, ker)

      head = ker.blockIdx_x
@@ -153,9 +163,6 @@ def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False
      batch = ker.blockIdx_z
      q_seq = ker.blockIdx_y * NUM_WORKERS + ker.warpid

-      k_smem = ker.st((KV_BLOCK_SIZE, D), dtypes.bfloat16)
-      v_smem = ker.st((KV_BLOCK_SIZE, D), dtypes.bfloat16)
-
      q_reg_fl = ker.rt((Q_BLOCK_SIZE, D), dtypes.float32)
      q_reg = ker.rt((Q_BLOCK_SIZE, D), dtypes.bfloat16)
      q_reg_t = ker.rt((D, Q_BLOCK_SIZE), dtypes.bfloat16, TileLayout.COL)
@@ -194,24 +201,30 @@ def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False
      l_vec_reg *= 1.0 / math.log(2)
      delta_vec_reg = warp.load(delta_vec_reg, delta_vec, (), (batch, head, 0, q_seq), axis=2)

-      for kv_idx in ker.range(N // KV_BLOCK_SIZE):
-        k_smem = warp.load(k_smem, k, (), (batch, kv_idx, head_kv, 0), axis=1)
-        v_smem = warp.load(v_smem, v, (), (batch, kv_idx, head_kv, 0), axis=1)
+      num_kv_blocks = (q_seq + 1) if is_causal else (N // KV_BLOCK_SIZE)
+      for kv_idx in ker.range(num_kv_blocks):
+        k_reg = warp.load(k_reg, k, (), (batch, kv_idx, head_kv, 0), axis=1)
+        k_reg_col = warp.load(k_reg_col, k, (), (batch, kv_idx, head_kv, 0), axis=1)
+        v_reg = warp.load(v_reg, v, (), (batch, kv_idx, head_kv, 0), axis=1)

-        k_reg = warp.load(k_reg, k_smem)
        k_reg_t = warp.transpose(k_reg_t, k_reg)
-        k_reg_col = warp.load(k_reg_col, k_smem)
        k_reg_col_t = warp.transpose(k_reg_col_t, k_reg_col)
-        v_reg = warp.load(v_reg, v_smem)

        # mma qk^t
        att_block = warp.zero(att_block.after(kv_idx))
        att_block = warp.mma_AtB(att_block, k_reg_t, q_reg_t)

        # apply attention mask
-        mask_reg = warp.load(mask_reg, mask, (), (batch, 0, q_seq, kv_idx), axis=2)
-        mask_reg_transposed = warp.transpose(mask_reg_transposed, mask_reg)
-        att_block += mask_reg_transposed
+        if is_causal:
+          bs_rows, bs_cols, bs_stride = att_block.base_shape.rows, att_block.base_shape.cols, att_block.base_shape.stride
+          q_base = q_seq * Q_BLOCK_SIZE + (warp.laneid % bs_cols)
+          kv_base = kv_idx * KV_BLOCK_SIZE + (warp.laneid // bs_cols) * bs_stride
+          att_block = warp.map(att_block,
+            lambda x, idx: ((kv_base + idx[0]*bs_rows + idx[2]) > (q_base + idx[1]*bs_cols)).alu(Ops.WHERE, UOp.ufix(x._uop, -math.inf), x))
+        elif mask is not None:
+          mask_reg = warp.load(mask_reg, mask, (), (batch, 0, q_seq, kv_idx), axis=2)
+          mask_reg_transposed = warp.transpose(mask_reg_transposed, mask_reg)
+          att_block += mask_reg_transposed

        att_block -= l_vec_reg
        att_block = att_block.exp2()
@@ -231,19 +244,24 @@ def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False

      return ker.finish()

-  def custom_backward_kv(dku:UOp, dvu:UOp, dou:UOp, qu:UOp, ku:UOp, vu:UOp, masku:UOp, l_vecu:UOp, delta_vecu:UOp):
+  def custom_backward_q_causal(dqu:UOp, dou:UOp, qu:UOp, ku:UOp, vu:UOp, l_vecu:UOp, delta_vecu:UOp) -> UOp:
+    return _custom_backward_q_impl(dqu, dou, qu, ku, vu, None, l_vecu, delta_vecu)
+
+  def custom_backward_q_masked(dqu:UOp, dou:UOp, qu:UOp, ku:UOp, vu:UOp, masku:UOp, l_vecu:UOp, delta_vecu:UOp) -> UOp:
+    return _custom_backward_q_impl(dqu, dou, qu, ku, vu, masku, l_vecu, delta_vecu)
+
+  def _custom_backward_kv_impl(dku:UOp, dvu:UOp, dou:UOp, qu:UOp, ku:UOp, vu:UOp, masku:UOp|None, l_vecu:UOp, delta_vecu:UOp):
    with Kernel("fa_custom_backward_kv", (H_KV, N // (KV_BLOCK_SIZE*NUM_WORKERS), B_local), NUM_WORKERS * WARP_THREADS) as ker:
      warp = ker.warp

-      dk, dv, do, q, k, v, mask = GL(dku, ker), GL(dvu, ker), GL(dou, ker), GL(qu, ker), GL(ku, ker), GL(vu, ker), GL(masku, ker)
+      dk, dv, do, q, k, v = GL(dku, ker), GL(dvu, ker), GL(dou, ker), GL(qu, ker), GL(ku, ker), GL(vu, ker)
+      mask = GL(masku, ker) if masku is not None else None
      l_vec, delta_vec = GL(l_vecu, ker), GL(delta_vecu, ker)

      head_kv = ker.blockIdx_x
      batch = ker.blockIdx_z
      kv_seq = ker.blockIdx_y * NUM_WORKERS + ker.warpid

-      q_smem = ker.st((Q_BLOCK_SIZE, D), dtypes.bfloat16)
-      do_smem = ker.st((Q_BLOCK_SIZE, D), dtypes.bfloat16)
      att_smem = ker.st((Q_BLOCK_SIZE, KV_BLOCK_SIZE), dtypes.bfloat16)

      q_reg = ker.rt((Q_BLOCK_SIZE, D), dtypes.bfloat16)
@@ -277,19 +295,17 @@ def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False
      k_reg_t = warp.transpose(k_reg_t, k_reg)
      v_reg = warp.load(v_reg, v, (), (batch, kv_seq, head_kv, 0), axis=1)

-      for q_idx in ker.range(N // Q_BLOCK_SIZE):
+      q_start = kv_seq if is_causal else 0
+      for q_idx in ker.range(q_start, N // Q_BLOCK_SIZE):
        for g in ker.range(GROUP_SIZE):
          head_q = head_kv * GROUP_SIZE + g

-          # load q and do
-          q_smem = warp.load(q_smem, q, (), (batch, q_idx, head_q, 0), axis=1)
-          do_smem = warp.load(do_smem, do, (), (batch, q_idx, head_q, 0), axis=1)
+          q_reg = warp.load(q_reg, q, (), (batch, q_idx, head_q, 0), axis=1)
+          q_reg_col = warp.load(q_reg_col, q, (), (batch, q_idx, head_q, 0), axis=1)
+          do_reg = warp.load(do_reg, do, (), (batch, q_idx, head_q, 0), axis=1)
+          do_reg_col = warp.load(do_reg_col, do, (), (batch, q_idx, head_q, 0), axis=1)

-          q_reg = warp.load(q_reg, q_smem)
          q_reg_t = warp.transpose(q_reg_t, q_reg)
-          q_reg_col = warp.load(q_reg_col, q_smem)
-          do_reg = warp.load(do_reg, do_smem)
-          do_reg_col = warp.load(do_reg_col, do_smem)

          # load l_vec and delta_vec
          l_vec_reg = warp.load(l_vec_reg, l_vec, (), (batch, head_q, 0, q_idx), axis=2)
@@ -302,9 +318,16 @@ def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False
          att_block *= (1.0 / math.sqrt(D)) * (1.0 / math.log(2))

          # apply attention mask
-          mask_reg = warp.load(mask_reg, mask, (), (batch, 0, q_idx, kv_seq), axis=2)
-          mask_reg_transposed = warp.transpose(mask_reg_transposed, mask_reg)
-          att_block += mask_reg_transposed
+          if is_causal:
+            bs_rows, bs_cols, bs_stride = att_block.base_shape.rows, att_block.base_shape.cols, att_block.base_shape.stride
+            q_base = q_idx * Q_BLOCK_SIZE + (warp.laneid % bs_cols)
+            kv_base = kv_seq * KV_BLOCK_SIZE + (warp.laneid // bs_cols) * bs_stride
+            att_block = warp.map(att_block,
+              lambda x, idx: ((kv_base + idx[0]*bs_rows + idx[2]) > (q_base + idx[1]*bs_cols)).alu(Ops.WHERE, UOp.ufix(x._uop, -math.inf), x))
+          elif mask is not None:
+            mask_reg = warp.load(mask_reg, mask, (), (batch, 0, q_idx, kv_seq), axis=2)
+            mask_reg_transposed = warp.transpose(mask_reg_transposed, mask_reg)
+            att_block += mask_reg_transposed

          att_block -= l_vec_reg
          att_block = att_block.exp2()
@@ -313,7 +336,7 @@ def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False
          att_block_transposed = warp.transpose(att_block_transposed, att_block_mma)
          att_smem = warp.store(att_smem, att_block_transposed)
          att_block_row = warp.load(att_block_row, att_smem)
-          dv_reg_ = warp.mma_AB(dv_reg, att_block_row, do_reg_col)
+          dv_reg_ = warp.mma_AtB(dv_reg, att_block_row, do_reg_col)

          dp_block = warp.zero(dp_block.after(g, q_idx, dv_reg_))
          dp_block = warp.mma_ABt(dp_block, v_reg, do_reg)
@@ -325,7 +348,7 @@ def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False
          att_block_transposed = warp.transpose(att_block_transposed, att_block_mma)
          att_smem = warp.store(att_smem, att_block_transposed)
          att_block_row = warp.load(att_block_row, att_smem)
-          dk_reg = warp.mma_AB(dk_reg, att_block_row, q_reg_col)
+          dk_reg = warp.mma_AtB(dk_reg, att_block_row, q_reg_col)
      dk_reg = ker.endrange(2)
      dv_reg = dv_reg.after(dk_reg)

@@ -336,24 +359,31 @@ def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False

      return ker.finish(2)

+  def custom_backward_kv_causal(dku:UOp, dvu:UOp, dou:UOp, qu:UOp, ku:UOp, vu:UOp, l_vecu:UOp, delta_vecu:UOp):
+    return _custom_backward_kv_impl(dku, dvu, dou, qu, ku, vu, None, l_vecu, delta_vecu)
+
+  def custom_backward_kv_masked(dku:UOp, dvu:UOp, dou:UOp, qu:UOp, ku:UOp, vu:UOp, masku:UOp, l_vecu:UOp, delta_vecu:UOp):
+    return _custom_backward_kv_impl(dku, dvu, dou, qu, ku, vu, masku, l_vecu, delta_vecu)
+
  single_device = xq.device[0] if isinstance(xq.device, tuple) else xq.device

  if is_causal:
    if attn_mask is not None: raise RuntimeError("cannot set attn_mask when is_causal=True")
-    attn_mask = Tensor.ones((B, 1, N, N), requires_grad=False, device=single_device, dtype=dtypes.bool).tril()
-  if attn_mask is not None:
+  elif attn_mask is not None:
    if attn_mask.dtype == dtypes.bool: attn_mask = attn_mask.where(0, -float("inf"))
+    if attn_mask.shape != (B, 1, N, N):
+      attn_mask = attn_mask.expand(B, 1, N, N)
+    if isinstance(xq.device, tuple) and not isinstance(attn_mask.device, tuple):
+      attn_mask = attn_mask.shard(xq.device, axis=0)
  else:
    attn_mask = Tensor.zeros((B, 1, N, N), requires_grad=False, device=single_device, dtype=dtypes.float32)
-  if attn_mask.shape != (B, 1, N, N):
-    attn_mask = attn_mask.expand(B, 1, N, N)
-  if isinstance(xq.device, tuple) and not isinstance(attn_mask.device, tuple):
-    attn_mask = attn_mask.shard(xq.device, axis=0)
+    if isinstance(xq.device, tuple):
+      attn_mask = attn_mask.shard(xq.device, axis=0)

  attn = _sharded_empty_like(xq, axis=0)
  l_vec = _sharded_empty((B, H, 1, N), xq, axis=0)

-  def grad(gradu:UOp, _) -> tuple[None, None, UOp, UOp, UOp, None]:
+  def grad_causal(gradu:UOp, _) -> tuple[None, None, UOp, UOp, UOp]:
    grad = Tensor(gradu, device=gradu.device)
    grad_q = _sharded_empty_like(xq, axis=0)
    grad_k = _sharded_empty_like(xk, axis=0)
@@ -361,11 +391,26 @@ def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False

    delta_vec = (grad * attn).sum(-1, dtype=dtypes.float32).transpose(1, 2).unsqueeze(-2).detach()

-    grad_q = Tensor.custom_kernel(grad_q, grad, xq, xk, xv, attn_mask, l_vec, delta_vec, fxn=custom_backward_q)[0]
-    grad_k, grad_v = Tensor.custom_kernel(grad_k, grad_v, grad, xq, xk, xv, attn_mask, l_vec, delta_vec, fxn=custom_backward_kv)[:2]
+    grad_q = Tensor.custom_kernel(grad_q, grad, xq, xk, xv, l_vec, delta_vec, fxn=custom_backward_q_causal)[0]
+    grad_k, grad_v = Tensor.custom_kernel(grad_k, grad_v, grad, xq, xk, xv, l_vec, delta_vec, fxn=custom_backward_kv_causal)[:2]
+    return (None, None, grad_q.uop, grad_k.uop, grad_v.uop)
+
+  def grad_masked(gradu:UOp, _) -> tuple[None, None, UOp, UOp, UOp, None]:
+    grad = Tensor(gradu, device=gradu.device)
+    grad_q = _sharded_empty_like(xq, axis=0)
+    grad_k = _sharded_empty_like(xk, axis=0)
+    grad_v = _sharded_empty_like(xv, axis=0)
+
+    delta_vec = (grad * attn).sum(-1, dtype=dtypes.float32).transpose(1, 2).unsqueeze(-2).detach()
+
+    grad_q = Tensor.custom_kernel(grad_q, grad, xq, xk, xv, attn_mask, l_vec, delta_vec, fxn=custom_backward_q_masked)[0]
+    grad_k, grad_v = Tensor.custom_kernel(grad_k, grad_v, grad, xq, xk, xv, attn_mask, l_vec, delta_vec, fxn=custom_backward_kv_masked)[:2]
    return (None, None, grad_q.uop, grad_k.uop, grad_v.uop, None)

-  attn, l_vec = Tensor.custom_kernel(attn, l_vec, xq, xk, xv, attn_mask, fxn=custom_forward, grad_fxn=grad)[:2]
+  if is_causal:
+    attn, l_vec = Tensor.custom_kernel(attn, l_vec, xq, xk, xv, fxn=custom_forward_causal, grad_fxn=grad_causal)[:2]
+  else:
+    attn, l_vec = Tensor.custom_kernel(attn, l_vec, xq, xk, xv, attn_mask, fxn=custom_forward_masked, grad_fxn=grad_masked)[:2]
  attn_ = attn[:, :N_, :, :D_]

  return attn_.transpose(1, 2).cast(odtype)
@@ -12,7 +12,7 @@ class _tk_range:
  def __next__(self):
    if not self.done:
      self.done = True
-      self._rng = UOp.range(self.end // self.step, self.rid, axis_type=self.axis_type) * self.step + self.start
+      self._rng = UOp.range((self.end - self.start) // self.step, self.rid, axis_type=self.axis_type) * self.step + self.start
      return self._rng
    raise StopIteration

@@ -29,6 +29,10 @@ def wrap(x: Tensor) -> torch.Tensor:
  x._strides = strides_for_shape(x.shape) # always recalculate
  if (not hasattr(x, '_storage_offset')) or (not x.uop.is_realized): x._storage_offset = calculate_storage_offset(x)
  return mod.wrap(x, _to_torch_dtype(x.dtype), _to_torch_device(x.device).index)
+def _update_torch_metadata(tensor: torch.Tensor, tiny: Tensor) -> None:
+  tiny._strides = strides_for_shape(tiny.shape)
+  tiny._storage_offset = calculate_storage_offset(tiny)
+  mod.update_metadata(tensor, tiny.shape, tiny._strides, tiny._storage_offset)
 def unwrap(x:torch.Tensor) -> Tensor:
  assert isinstance(x, torch.Tensor), f"x isn't {type(x)}"
  return mod.unwrap(x)
@@ -344,7 +348,7 @@ def scatter_add(self, dim, index, src, out):
 def _copy_between_devices(src, dest, cast_dtype, to_device, non_blocking=False):
  if src.is_tiny and dest.is_tiny:
    src_t, dest_t = unwrap(src), unwrap(dest)
-    if dest_t.uop.is_contiguous() or dest_t.uop.is_realized: src_t = src_t.contiguous()
+    if dest_t.uop.has_buffer_identity() or dest_t.uop.is_realized: src_t = src_t.contiguous()
    _apply_inplace(dest_t, src_t.cast(cast_dtype).to(to_device))
  elif src.is_tiny and dest.is_cpu:
    dest.resize_(src.numel()).resize_(src.shape)
@@ -611,7 +615,10 @@ tiny_backend = {**{k:wrap_out(v) for k,v in tiny_backend_out.items()}, **{
  "aten.fill_.Tensor": lambda self, value: Tensor.full(self.shape, value.reshape(()).item(), device=self.device, dtype=self.dtype),
  "aten.flip": Tensor.flip,
  "aten.scatter_reduce.two": Tensor.scatter_reduce,
-  "aten.squeeze_.dim": lambda self, dim: self.replace(self.squeeze(dim), allow_shape_mismatch=True), # TODO: inplace view op, here?
+  "aten.squeeze_.dim": Tensor.squeeze,
+  "aten.unsqueeze_": Tensor.unsqueeze,
+  "aten.transpose_": Tensor.transpose,
+  "aten.t_": Tensor.transpose,
  "aten.add.Tensor": lambda input,other,alpha=1: input+alpha*other,
  "aten.linspace": lambda start, stop, steps, dtype=None, **kwargs:
    Tensor.linspace(start, stop, steps, **({"dtype": _from_torch_dtype(dtype)} if dtype is not None else {})),
@@ -655,6 +662,13 @@ inplace_ops = {
  "aten.masked_fill_.Tensor",
 }

+inplace_view_ops = {
+  "aten.squeeze_.dim",
+  "aten.unsqueeze_",
+  "aten.transpose_",
+  "aten.t_",
+}
+
 def wrap_fxn(k,f):
  def nf(*args, **kwargs):
    if TORCH_DEBUG:
@@ -675,8 +689,42 @@ def wrap_inplace(k,f):
    return orig
  return nf

+def wrap_inplace_view_op(k,f):
+  def nf(*args, **kwargs):
+    orig = args[0]
+    args, kwargs = unwrap_args(args, kwargs)
+    target = args[0]
+    new_view = f(*args, **kwargs)
+    if new_view is target or new_view.uop is target.uop:
+      _update_torch_metadata(orig, target)
+      return orig
+    base = canonical_base(target)
+    op = (f, args[1:], kwargs)
+    if target is base:
+      views = derived_views(base)
+      if views:
+        old_base = Tensor(base.uop, device=base.device)
+        old_base.requires_grad = base.requires_grad
+        old_base._views = getattr(base, "_views", set())
+        for v in views: v._view_base = old_base
+        base._views = set()
+        base._view_base = old_base
+        base._view_ops = [op]
+        old_base._views.add(weakref.ref(base))
+    else:
+      target._view_base = base
+      base._views = getattr(base, "_views", set())
+      base._views.add(weakref.ref(target))
+      target._view_ops = _get_view_ops(target) + [op]
+    target.uop = new_view.uop
+    _update_torch_metadata(orig, target)
+    return orig
+  return nf
+
 for k,v in tiny_backend.items():
-  wrapper = wrap_inplace if k in inplace_ops else wrap_fxn
+  if k in inplace_view_ops: wrapper = wrap_inplace_view_op
+  elif k in inplace_ops: wrapper = wrap_inplace
+  else: wrapper = wrap_fxn
  torch.library.impl(k.replace("aten.", "aten::"), "privateuseone")(wrapper(k,v))

@torch.library.impl("aten::equal", "privateuseone")
@@ -67,5 +67,37 @@ class TestTorchBackendInplace(unittest.TestCase):
    d += torch.arange(4)
    np.testing.assert_array_equal(a.cpu(), torch.arange(4).cpu())

+  def test_inplace_view_metadata(self):
+    a = torch.arange(6, dtype=torch.float32).reshape(1, 2, 3)
+    ret = a.squeeze_(0)
+    self.assertIs(ret, a)
+    self.assertEqual(a.shape, torch.Size([2, 3]))
+    ret = a.unsqueeze_(1)
+    self.assertIs(ret, a)
+    self.assertEqual(a.shape, torch.Size([2, 1, 3]))
+    ret = a.transpose_(0, 2)
+    self.assertIs(ret, a)
+    self.assertEqual(a.shape, torch.Size([3, 1, 2]))
+
+  def test_t_inplace_metadata(self):
+    a = torch.arange(6, dtype=torch.float32).reshape(2, 3)
+    ret = a.t_()
+    self.assertIs(ret, a)
+    self.assertEqual(a.shape, torch.Size([3, 2]))
+    expected = torch.arange(6, dtype=torch.float32).reshape(2, 3).t()
+    np.testing.assert_array_equal(a.cpu().numpy(), expected.cpu().numpy())
+
+  def test_squeeze_matmul(self):
+    # squeeze_ is used internally by PyTorch for vector-matrix matmul (unsqueeze -> mm -> squeeze_)
+    a = torch.arange(65, dtype=torch.float32)
+    b = torch.arange(65*45, dtype=torch.float32).reshape(65, 45)
+    result = a.matmul(b)
+    self.assertEqual(result.shape, torch.Size([45]))
+    # verify correctness
+    a_cpu = torch.arange(65, dtype=torch.float32, device='cpu')
+    b_cpu = torch.arange(65*45, dtype=torch.float32, device='cpu').reshape(65, 45)
+    expected = a_cpu.matmul(b_cpu)
+    np.testing.assert_allclose(result.cpu().numpy(), expected.numpy(), rtol=1e-4, atol=1e-4)
+
 if __name__ == "__main__":
  unittest.main()
@@ -131,7 +131,15 @@ py::object unwrap_tensor(const at::Tensor &tensor) {
  return py::reinterpret_borrow<py::object>(tiny->ptr(getPyInterpreter()));
 }

+void update_metadata(const at::Tensor &tensor, const std::vector<int64_t> &sizes,
+                     const std::vector<int64_t> &strides, int64_t storage_offset) {
+  auto* impl = tensor.unsafeGetTensorImpl();
+  impl->set_allow_tensor_metadata_change(true);
+  impl->set_sizes_and_strides(sizes, strides, storage_offset);
+}
+
 PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
  m.def("wrap", &wrap_tensor);
  m.def("unwrap", &unwrap_tensor);
+  m.def("update_metadata", &update_metadata);
 }
@@ -1,6 +1,6 @@
 #!/usr/bin/env python3
-import sys, os, zlib, struct, hashlib
-from tinygrad.helpers import DEBUG, getenv, fetch
+import os, zlib, struct, hashlib
+from tinygrad.helpers import getenv
 from tinygrad.runtime.support.usb import USB3

 SUPPORTED_CONTROLLERS = [
@@ -50,7 +50,7 @@ patched_fw = patch(file_path, file_hash, patches)
 dev = None
 for vendor, device in SUPPORTED_CONTROLLERS:
  try:
-    dev = USB3(vendor, device, 0x81, 0x83, 0x02, 0x04)
+    dev = USB3(vendor, device, 0x81, 0x83, 0x02, 0x04, use_bot=True)
    break
  except RuntimeError: pass
 if dev is None:
@@ -3,14 +3,19 @@ import argparse, pathlib
 from typing import Iterator
 from tinygrad.viz import serve as viz
 from tinygrad.uop.ops import RewriteTrace
-from tinygrad.helpers import temp, ansistrip, colored
+from tinygrad.helpers import temp, ansistrip, colored, time_to_str, ansilen
+from test.null.test_viz import load_profile

 def optional_eq(val:dict, arg:str|None) -> bool: return arg is None or ansistrip(val["name"]) == arg

 def print_data(data:dict) -> None:
  if isinstance(data.get("value"), Iterator):
    for m in data["value"]:
+      if m.get("uop"):
+        print("Input UOp:")
+        print(m["uop"])
      if not m["diff"]: continue
+      print("Rewrites:")
      fp = pathlib.Path(m["upat"][0][0])
      print(f"{fp.parent.name}/{fp.name}:{m['upat'][0][1]}")
      print(m["upat"][1])
@@ -23,11 +28,46 @@ if __name__ == "__main__":
  parser = argparse.ArgumentParser()
  parser.add_argument('--kernel', type=str, default=None, metavar="NAME", help='Select a kernel by name (optional name, default: only list names)')
  parser.add_argument('--select', type=str, default=None, metavar="NAME",
-                      help='Select an item within the chosen kernel (optional name, default: only list names)')
+                      help='Rewrites: Select an item within the chosen kernel (optional name, default: only list names)')
+  parser.add_argument('--profile', action="store_true", help="View profiling trace (default: views rewrites)")
+  parser.add_argument('--device', type=str, default=None, metavar="NAME", help="Profile only: Select a device (default: prints all devices)")
+  parser.add_argument('--profile-path', 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('--rewrites-path', type=pathlib.Path, metavar="PATH", help='Path to rewrites (optional file, default: latest rewrites)',
+                      default=pathlib.Path(temp("rewrites.pkl", append_user=True)))
  args = parser.parse_args()

-  viz.trace = viz.load_pickle(pathlib.Path(temp("rewrites.pkl", append_user=True)), default=RewriteTrace([], [], {}))
+  viz.trace = viz.load_pickle(args.rewrites_path, default=RewriteTrace([], [], {}))
  viz.ctxs = viz.get_rewrites(viz.trace)
+
+  if args.profile:
+    from tabulate import tabulate
+    profile = load_profile(viz.load_pickle(args.profile_path, default=[]))
+    agg, total, n = {}, 0, 0
+    for k,v in profile["layout"].items():
+      if not optional_eq({"name":k}, args.device): continue
+      print(k)
+      if args.device is None: continue
+      for e in v.get("events", []):
+        et = e["dur"]*1e-6
+        if args.kernel is not None:
+          if ansistrip(e["name"]) == args.kernel and n < 10:
+            ptm = colored(time_to_str(et, w=9), "yellow" if et > 0.01 else None) if et is not None else ""
+            name = e["name"]+(" " * (46 - ansilen(e["name"])))
+            print(f"{name} {ptm}/{(et or 0)*1e3:9.2f}ms  "+e['fmt'].replace('\n', ' | ')+"  ")
+            n += 1
+        else:
+          a = agg.setdefault(e["name"], [0.0, 0])
+          a[0] += et
+          a[1] += 1
+          total += et
+
+    if agg:
+      rows = [[n, t, time_to_str(t, w=9), t / c if c else 0.0, c, (t / total * 100.0) if total else 0.0] for n, (t, c) in agg.items()]
+      rows.sort(key=lambda r: r[1], reverse=True)
+      print(tabulate([[r[0], r[2], r[4], f"{r[5]:.2f}%"] for r in rows[:30]], headers=["name", "total", "count", "pct"], tablefmt="github"))
+    exit(0)
+
  for k in viz.ctxs:
    if not optional_eq(k, args.kernel): continue
    print(k["name"])
@@ -119,7 +119,7 @@ class TestHCQ(unittest.TestCase):
    TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
    TestHCQ.d0.timeline_value += 1

-    val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[0]
+    val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[0]
    assert val == 1.0, f"got val {val}"

  def test_exec_2_kernels_100_times(self):
@@ -135,7 +135,7 @@ class TestHCQ(unittest.TestCase):
      q.submit(TestHCQ.d0, {virt_val.expr: TestHCQ.d0.timeline_value})
      TestHCQ.d0.timeline_value += 1

-    val = TestHCQ.a.uop.buffer.as_buffer().cast("f")[0]
+    val = TestHCQ.a.uop.buffer.as_memoryview().cast("f")[0]
    assert val == 200.0, f"got val {val}"

  @unittest.skipIf(Device.DEFAULT in {"CPU"}, "No globals/locals on LLVM/CPU")
@@ -151,9 +151,9 @@ class TestHCQ(unittest.TestCase):
    TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
    TestHCQ.d0.timeline_value += 1

-    val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[0]
+    val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[0]
    assert val == 1.0, f"got val {val}"
-    val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[1]
+    val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[1]
    assert val == 0.0, f"got val {val}, should not be updated"

  @unittest.skipIf(Device.DEFAULT in {"CPU"}, "No globals/locals on LLVM/CPU")
@@ -186,7 +186,7 @@ class TestHCQ(unittest.TestCase):
          TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
          TestHCQ.d0.timeline_value += 1

-          res_sum = sum(x for x in zt.as_buffer().cast("I"))
+          res_sum = sum(x for x in zt.as_memoryview().cast("I"))
          assert x * y * z == res_sum, f"want {x * y * z}, got {res_sum}"

  # Test copy
@@ -200,7 +200,7 @@ class TestHCQ(unittest.TestCase):
    TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
    TestHCQ.d0.timeline_value += 1

-    val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[1]
+    val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[1]
    assert val == 1.0, f"got val {val}"

  def test_copy_long(self):
@@ -218,7 +218,7 @@ class TestHCQ(unittest.TestCase):
    TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
    TestHCQ.d0.timeline_value += 1

-    mv_buf1 = buf1.as_buffer().cast('Q')
+    mv_buf1 = buf1.as_memoryview().cast('Q')
    assert libc.memcmp(mv_address(mv_buf1), buf2._buf.va_addr, sz) == 0

  @slow
@@ -242,7 +242,7 @@ class TestHCQ(unittest.TestCase):
      TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
      TestHCQ.d0.timeline_value += 1

-      mv_buf1 = buf1.as_buffer()
+      mv_buf1 = buf1.as_memoryview()
      assert libc.memcmp(mv_address(mv_buf1), buf2._buf.va_addr, sz) == 0

  def test_update_copy(self):
@@ -260,7 +260,7 @@ class TestHCQ(unittest.TestCase):
    TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
    TestHCQ.d0.timeline_value += 1

-    val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[1]
+    val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[1]
    assert val == 1.0, f"got val {val}"

  def test_update_copy_long(self):
@@ -283,7 +283,7 @@ class TestHCQ(unittest.TestCase):
    TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
    TestHCQ.d0.timeline_value += 1

-    mv_buf1 = buf1.as_buffer().cast('Q')
+    mv_buf1 = buf1.as_memoryview().cast('Q')
    for i in range(sz//8): assert mv_buf1[i] == 0x0101010101010101, f"offset {i*8} differs, not all copied, got {hex(mv_buf1[i])}"

  # Test bind api
@@ -421,7 +421,7 @@ class TestHCQ(unittest.TestCase):
      TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
      TestHCQ.d0.timeline_value += 1

-      assert buf1.as_buffer()[0] == i
+      assert buf1.as_memoryview()[0] == i

  def test_small_copies_from_host_buf_intercopy(self):
    if TestHCQ.d0.hw_copy_queue_t is None: self.skipTest("device does not support copy queue")
@@ -440,7 +440,7 @@ class TestHCQ(unittest.TestCase):
      TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
      TestHCQ.d0.timeline_value += 1

-      assert buf2.as_buffer()[0] == i
+      assert buf2.as_memoryview()[0] == i

  def test_small_copies_from_host_buf_transfer(self):
    if TestHCQ.d0.hw_copy_queue_t is None: self.skipTest("device does not support copy queue")
@@ -463,7 +463,7 @@ class TestHCQ(unittest.TestCase):
      TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
      TestHCQ.d0.timeline_value += 1

-      assert buf2.as_buffer()[0] == i
+      assert buf2.as_memoryview()[0] == i

  def test_memory_barrier(self):
    a = Tensor([0, 1], device=Device.DEFAULT, dtype=dtypes.int8).realize()
@@ -486,7 +486,7 @@ class TestHCQ(unittest.TestCase):
      TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
      TestHCQ.d0.timeline_value += 1

-      assert buf1.as_buffer()[0] == (i + 1), f"has {buf1.as_buffer()[0]}, need {i + 1}"
+      assert buf1.as_memoryview()[0] == (i + 1), f"has {buf1.as_memoryview()[0]}, need {i + 1}"

  def test_memory_barrier_before_copy(self):
    if TestHCQ.d0.hw_copy_queue_t is None: self.skipTest("device does not support copy queue")
@@ -511,7 +511,7 @@ class TestHCQ(unittest.TestCase):
      TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
      TestHCQ.d0.timeline_value += 1

-      assert buf2.as_buffer()[0] == i
+      assert buf2.as_memoryview()[0] == i

  def test_map_cpu_buffer_to_device(self):
    if Device[Device.DEFAULT].hw_copy_queue_t is None: self.skipTest("skip device without copy queue")
@@ -1,43 +0,0 @@
-import random
-from tinygrad.helpers import getenv
-from tinygrad.codegen.opt.search import beam_search, bufs_from_lin
-from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
-from extra.optimization.helpers import load_worlds, ast_str_to_lin, time_linearizer
-
-def optimize_kernel(k):
-  # TODO: update this
-  return hand_coded_optimizations(k)
-
-if __name__ == '__main__':
-  hcopt_wins = beam_wins = tie = 0
-  hcopt_total = beam_total = 0.0
-
-  worlds = load_worlds(filter_reduce=False, filter_noimage=True, filter_novariable=False)
-  random.seed(0)
-  random.shuffle(worlds)
-
-  for world in worlds[:500]:
-    k = ast_str_to_lin(world)
-    rawbufs = bufs_from_lin(k)
-
-    k_hcopt = k.copy()
-    k_hcopt.apply_opts(optimize_kernel(k_hcopt))
-    k_beam = beam_search(k.copy(), rawbufs, getenv("BEAM", 2))
-
-    disable_cache = bool(getenv("NOCACHE", 0))
-    t_hcopt = time_linearizer(k_hcopt, rawbufs, allow_test_size=False, cnt=10, disable_cache=disable_cache, clear_l2=True) * 1e6
-    t_beam = time_linearizer(k_beam, rawbufs, allow_test_size=False, cnt=10, disable_cache=disable_cache, clear_l2=True) * 1e6
-
-    if t_hcopt == t_beam: tie += 1
-    elif t_hcopt < t_beam: hcopt_wins += 1
-    else: beam_wins += 1
-    hcopt_total += t_hcopt
-    beam_total += t_beam
-
-    print(f"{t_hcopt=:5.2f} {k_hcopt.applied_opts=}")
-    print("")
-    print(f"{t_beam=:5.2f} {k_beam.applied_opts=}")
-    print("*"*20)
-
-  print(f"{hcopt_wins=}, {beam_wins=}, {tie=}")
-  print(f"{hcopt_total=:.2f}, {beam_total=:.2f}")
@@ -11,7 +11,7 @@ from tinygrad.dtype import ImageDType, Invalid
 # PYTHONPATH="." DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx

 def vision_conv_143():
-  c0 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((16, 1024, 4)), (), 0)
+  c0 = UOp(Ops.PARAM, dtypes.imageh((16, 1024, 4)), (), 0)
  c2 = UOp.range(32, 3, AxisType.LOOP)
  c5 = UOp.range(128, 4, AxisType.LOOP)
  c8 = UOp.range(16, 2, AxisType.LOOP)
@@ -21,13 +21,13 @@ def vision_conv_143():
  c26 = UOp.range(7, 1, AxisType.REDUCE)
  c27 = c2*2+c26
  c32 = ((c27<3)!=True)&(c27<67)
-  c34 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((32, 1024, 4)), (), 1)
+  c34 = UOp(Ops.PARAM, dtypes.imageh((32, 1024, 4)), (), 1)
  c38 = c5//2
  c45 = (c32&c24).where((c27*64+c38+c17*4096+-12480), UOp.const(dtypes.index, Invalid))
  c48 = (c24&c32).where(c34.index(c45), UOp.const(dtypes.float, 0.0))
-  c49 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((64, 49, 4)), (), 2)
+  c49 = UOp(Ops.PARAM, dtypes.imageh((64, 49, 4)), (), 2)
  c61 = c48*c49.index((c26*4+c5%2+c16*28+c38*196))
-  c63 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(128), (), 3)
+  c63 = UOp(Ops.PARAM, dtypes.float.ptr(128), (), 3)
  c65 = c61.reduce(c16, c26, arg=Ops.ADD)+c63.index(c5)
  c67 = c0.index((c2*128+c5+c8*4096), ptr=True).store(c65).end(c8, c2, c5)

@@ -37,7 +37,7 @@ def vision_conv_143():
  return c67.sink(arg=KernelInfo(name="conv", opts_to_apply=opts))

 def vision_conv_153():
-  c0 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((8, 1024, 4)), (), 0)
+  c0 = UOp(Ops.PARAM, dtypes.imageh((8, 1024, 4)), (), 0)
  c2 = UOp.range(16, 3, AxisType.LOOP)
  c5 = UOp.range(256, 4, AxisType.LOOP)
  c8 = UOp.range(8, 2, AxisType.LOOP)
@@ -47,13 +47,13 @@ def vision_conv_153():
  c26 = UOp.range(7, 1, AxisType.REDUCE)
  c27 = c2*2+c26
  c32 = ((c27<3)!=True)&(c27<35)
-  c34 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((16, 1024, 4)), (), 1)
+  c34 = UOp(Ops.PARAM, dtypes.imageh((16, 1024, 4)), (), 1)
  c38 = c5//2
  c45 = (c32&c24).where((c27*128+c38+c17*4096+-12672), UOp.const(dtypes.index, Invalid))
  c48 = (c24&c32).where(c34.index(c45), UOp.const(dtypes.float, 0.0))
-  c49 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((128, 49, 4)), (), 2)
+  c49 = UOp(Ops.PARAM, dtypes.imageh((128, 49, 4)), (), 2)
  c61 = c48*c49.index((c26*4+c5%2+c16*28+c38*196))
-  c63 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(256), (), 3)
+  c63 = UOp(Ops.PARAM, dtypes.float.ptr(256), (), 3)
  c65 = c61.reduce(c16, c26, arg=Ops.ADD)+c63.index(c5)
  c67 = c0.index((c2*256+c5+c8*4096), ptr=True).store(c65).end(c8, c2, c5)

@@ -63,16 +63,16 @@ def vision_conv_153():
  return c67.sink(arg=KernelInfo(name="conv", opts_to_apply=opts))

 def dm_conv_172():
-  c0 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((1, 240, 4)), (), 0)
+  c0 = UOp(Ops.PARAM, dtypes.imageh((1, 240, 4)), (), 0)
  c2 = UOp.range(960, 4, AxisType.LOOP)
-  c5 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((8, 384, 4)), (), 1)
+  c5 = UOp(Ops.PARAM, dtypes.imageh((8, 384, 4)), (), 1)
  c7 = UOp.range(32, 0, AxisType.REDUCE)
  c10 = UOp.range(4, 1, AxisType.REDUCE)
  c13 = UOp.range(12, 3, AxisType.REDUCE)
  c18 = UOp.range(8, 2, AxisType.REDUCE)
-  c23 = UOp(Ops.DEFINE_GLOBAL, dtypes.imageh((240, 128, 4)), (), 2)
+  c23 = UOp(Ops.PARAM, dtypes.imageh((240, 128, 4)), (), 2)
  c35 = c5.index((c7*4+c10+c13*128+c18*1536))*c23.index((c10*4+c2%4+c7*16+c2//4*512))
-  c37 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(960), (), 3)
+  c37 = UOp(Ops.PARAM, dtypes.float.ptr(960), (), 3)
  c39 = c35.reduce(c7, c10, arg=Ops.ADD)+c37.index(c2)
  c50 = (1.0+((c39+0.044708251953125*(c39*(c39*c39)))*-2.3021129851685216).exp2()).reciprocal()*c39
  c53 = c50.reduce(c18, c13, arg=Ops.ADD)*0.010416666666666666
@@ -91,7 +91,7 @@ allocator = Device.default.allocator
 ps = get_program(ast, renderer)
 cr = CompiledRunner(replace(ps, device=Device.DEFAULT))

-gs = sorted(dedup([u for u in ast.toposort() if u.op is Ops.DEFINE_GLOBAL]), key=lambda u: u.arg)
+gs = sorted(dedup([u for u in ast.toposort() if u.op is Ops.PARAM]), key=lambda u: u.arg)
 # print(len(gs))
 # print([g.dtype for g in gs])
 bufs = [Buffer(ps.device, g.size, g.dtype if isinstance(g.dtype, ImageDType) else g.dtype._base).ensure_allocated() for g in gs]
@@ -0,0 +1,11 @@
+#!/usr/bin/env python3
+from tinygrad.tensor import Tensor
+import numpy as np
+
+while True:
+  arr = np.ones(1000000, dtype=np.uint8)
+  print(f"numpy: {(arr + 1)[:10]}")
+
+  ptr = arr.ctypes.data
+  tensor = Tensor.from_blob(ptr, arr.shape, dtype='uint8', device='QCOM').realize() + 1
+  print(f"from_blob: {tensor.numpy()[:10]}")
@@ -0,0 +1,137 @@
+# ruff: noqa: F405
+import unittest, subprocess, os
+from extra.assembly.amd.autogen.rdna3.ins import *  # noqa: F403
+from extra.assembly.amd.dsl import s, v, Inst, NULL
+
+def assemble_kernel(insts:list[Inst], name:str="test") -> str:
+  kd = {"next_free_vgpr": 8, "next_free_sgpr": 8, "wavefront_size32": 1, "user_sgpr_kernarg_segment_ptr": 1, "kernarg_size": 8}
+  disasm = "\n".join(inst.disasm() for inst in insts)
+  hsasrc = f".text\n.globl {name}\n.p2align 8\n.type {name},@function\n{name}:\n{disasm}\n"
+  return 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"
+
+def _run(code:str, timeout:float=15.0) -> subprocess.CompletedProcess:
+  # TODO: AM_RESET is required for now, so subprocesses
+  return subprocess.run(["python", "-c", code], env={**os.environ, "AMD": "1"}, capture_output=True, text=True, timeout=timeout)
+
+def _run_asm(asm_src:str) -> subprocess.CompletedProcess:
+  return _run('from tinygrad.device import Device; from tinygrad.runtime.ops_amd import AMDProgram; '
+              'from tinygrad.runtime.support.compiler_amd import HIPCompiler; dev = Device["AMD"]; '
+              f'AMDProgram(dev, "test", HIPCompiler(dev.arch).compile("""{asm_src}"""))('
+              'dev.allocator.alloc(64), global_size=(1,1,1), local_size=(1,1,1), wait=True)')
+
+def _verify_recovery() -> subprocess.CompletedProcess:
+  return _run('from tinygrad import Tensor; t = Tensor([1.0, 2.0], device="AMD").realize(); assert (t + 1).numpy().tolist() == [2.0, 3.0]')
+
+_ILLEGAL_INST_ASM = ".text\n.globl test\n.p2align 8\n.type test,@function\ntest:\n.byte 0xff,0xff,0xff,0xff\ns_endpgm\n" \
+  ".rodata\n.p2align 6\n.amdhsa_kernel test\n.amdhsa_next_free_vgpr 8\n.amdhsa_next_free_sgpr 8\n" \
+  ".amdhsa_wavefront_size32 1\n.amdhsa_user_sgpr_kernarg_segment_ptr 1\n.amdhsa_kernarg_size 8\n.end_amdhsa_kernel"
+
+@unittest.skipIf(os.environ.get("AMD") != "1" or os.environ.get("MOCKGPU") == "1", "AMD with AM driver required")
+class TestAMFaultRecovery(unittest.TestCase):
+  def _run_kernel(self, insts: list[Inst]) -> subprocess.CompletedProcess: return _run_asm(assemble_kernel(insts))
+
+  def _assert_fault_and_recovery(self, result:subprocess.CompletedProcess):
+    if result.stdout.strip(): print(f"\nstdout: {result.stdout.strip()}")
+    if result.stderr.strip(): print(f"\nstderr: {result.stderr.strip()}")
+    self.assertNotEqual(result.returncode, 0, f"Expected fault but succeeded: {result.stdout}")
+    self.assertEqual(_verify_recovery().returncode, 0)
+
+
+class TestGlobalMemoryFaults(TestAMFaultRecovery):
+  def test_global_load_unmapped(self):
+    insts = [v_mov_b32_e32(v[0], 0xBEEF0000), v_mov_b32_e32(v[1], 0xDEAD),
+             global_load_b32(v[2], addr=v[0:1], saddr=NULL, offset=0), s_waitcnt(vmcnt=0), s_endpgm()]
+    self._assert_fault_and_recovery(self._run_kernel(insts))
+
+  def test_global_store_unmapped(self):
+    insts = [v_mov_b32_e32(v[0], 0xBEEF0000), v_mov_b32_e32(v[1], 0xDEAD), v_mov_b32_e32(v[2], 0x12345678),
+             global_store_b32(addr=v[0:1], data=v[2], saddr=NULL, offset=0), s_waitcnt(vmcnt=0), s_endpgm()]
+    self._assert_fault_and_recovery(self._run_kernel(insts))
+
+  def test_global_null_ptr(self):
+    insts = [v_mov_b32_e32(v[0], 0), v_mov_b32_e32(v[1], 0),
+             global_load_b32(v[2], addr=v[0:1], saddr=NULL, offset=0), s_waitcnt(vmcnt=0), s_endpgm()]
+    self._assert_fault_and_recovery(self._run_kernel(insts))
+
+  def test_global_misaligned_b64(self):
+    insts = [v_mov_b32_e32(v[0], 0xBEEF0001), v_mov_b32_e32(v[1], 0xDEAD),
+             global_load_b64(v[2:3], addr=v[0:1], saddr=NULL, offset=0), s_waitcnt(vmcnt=0), s_endpgm()]
+    self._assert_fault_and_recovery(self._run_kernel(insts))
+
+  def test_global_misaligned_b128(self):
+    insts = [v_mov_b32_e32(v[0], 0xBEEF0004), v_mov_b32_e32(v[1], 0xDEAD),
+             global_load_b128(v[2:5], addr=v[0:1], saddr=NULL, offset=0), s_waitcnt(vmcnt=0), s_endpgm()]
+    self._assert_fault_and_recovery(self._run_kernel(insts))
+
+
+class TestSMEMFaults(TestAMFaultRecovery):
+  def test_smem_null_base(self):
+    insts = [s_mov_b32(s[2], 0), s_mov_b32(s[3], 0),
+             s_load_b32(s[4], s[2:3], 0, soffset=NULL), s_waitcnt(lgkmcnt=0), s_endpgm()]
+    self._assert_fault_and_recovery(self._run_kernel(insts))
+
+  def test_smem_unmapped_address(self):
+    insts = [s_mov_b32(s[2], 0xBEEF0000), s_mov_b32(s[3], 0xDEAD),
+             s_load_b32(s[4], s[2:3], 0, soffset=NULL), s_waitcnt(lgkmcnt=0), s_endpgm()]
+    self._assert_fault_and_recovery(self._run_kernel(insts))
+
+  def test_smem_misaligned_b64(self):
+    insts = [s_mov_b32(s[2], 0xBEEF0004), s_mov_b32(s[3], 0xDEAD),
+             s_load_b64(s[4:5], s[2:3], 0, soffset=NULL), s_waitcnt(lgkmcnt=0), s_endpgm()]
+    self._assert_fault_and_recovery(self._run_kernel(insts))
+
+  def test_smem_misaligned_b128(self):
+    insts = [s_mov_b32(s[2], 0xBEEF0004), s_mov_b32(s[3], 0xDEAD),
+             s_load_b128(s[4:7], s[2:3], 0, soffset=NULL), s_waitcnt(lgkmcnt=0), s_endpgm()]
+    self._assert_fault_and_recovery(self._run_kernel(insts))
+
+
+class TestIllegalInstruction(TestAMFaultRecovery):
+  def test_malformed_encoding(self):
+    self._assert_fault_and_recovery(_run_asm(_ILLEGAL_INST_ASM))
+
+
+class TestFlatFaults(TestAMFaultRecovery):
+  def test_flat_load_unmapped(self):
+    insts = [v_mov_b32_e32(v[0], 0xBEEF0000), v_mov_b32_e32(v[1], 0xDEAD),
+             flat_load_b32(v[2], addr=v[0:1], saddr=NULL, offset=0), s_waitcnt(vmcnt=0, lgkmcnt=0), s_endpgm()]
+    self._assert_fault_and_recovery(self._run_kernel(insts))
+
+  def test_flat_store_unmapped(self):
+    insts = [v_mov_b32_e32(v[0], 0xBEEF0000), v_mov_b32_e32(v[1], 0xDEAD), v_mov_b32_e32(v[2], 0x12345678),
+             flat_store_b32(addr=v[0:1], data=v[2], saddr=NULL, offset=0), s_waitcnt(vmcnt=0, lgkmcnt=0), s_endpgm()]
+    self._assert_fault_and_recovery(self._run_kernel(insts))
+
+
+class TestAtomicFaults(TestAMFaultRecovery):
+  def test_global_atomic_unmapped(self):
+    insts = [v_mov_b32_e32(v[0], 0xBEEF0000), v_mov_b32_e32(v[1], 0xDEAD), v_mov_b32_e32(v[2], 1),
+             global_atomic_add_u32(addr=v[0:1], data=v[2], saddr=NULL, offset=0), s_waitcnt(vmcnt=0), s_endpgm()]
+    self._assert_fault_and_recovery(self._run_kernel(insts))
+
+  def test_flat_atomic_unmapped(self):
+    insts = [v_mov_b32_e32(v[0], 0xBEEF0000), v_mov_b32_e32(v[1], 0xDEAD), v_mov_b32_e32(v[2], 1),
+             flat_atomic_add_u32(addr=v[0:1], data=v[2], saddr=NULL, offset=0), s_waitcnt(vmcnt=0, lgkmcnt=0), s_endpgm()]
+    self._assert_fault_and_recovery(self._run_kernel(insts))
+
+
+class TestRecovery(TestAMFaultRecovery):
+  def test_recovery_after_memviol(self):
+    insts = [v_mov_b32_e32(v[0], 0xBEEF0000), v_mov_b32_e32(v[1], 0xDEAD),
+             global_load_b32(v[2], addr=v[0:1], saddr=NULL, offset=0), s_waitcnt(vmcnt=0), s_endpgm()]
+    self.assertNotEqual(self._run_kernel(insts).returncode, 0)
+    self.assertEqual(_verify_recovery().returncode, 0)
+
+  def test_recovery_after_illegal_inst(self):
+    self.assertNotEqual(_run_asm(_ILLEGAL_INST_ASM).returncode, 0)
+    self.assertEqual(_verify_recovery().returncode, 0)
+
+  def test_multiple_faults_recovery(self):
+    insts = [v_mov_b32_e32(v[0], 0xBEEF0000), v_mov_b32_e32(v[1], 0xDEAD),
+             global_load_b32(v[2], addr=v[0:1], saddr=NULL, offset=0), s_waitcnt(vmcnt=0), s_endpgm()]
+    for _ in range(3):
+      self.assertNotEqual(self._run_kernel(insts).returncode, 0)
+      self.assertEqual(_verify_recovery().returncode, 0)
+
+if __name__ == "__main__":
+  unittest.main()
@@ -20,7 +20,7 @@ class TestAMD(unittest.TestCase):
                              global_size=TestAMD.d0_runner.global_size, local_size=TestAMD.d0_runner.local_size)
      TestAMD.d0_runner.clprg(TestAMD.a.uop.buffer._buf, TestAMD.b.uop.buffer._buf,
                              global_size=TestAMD.d0_runner.global_size, local_size=TestAMD.d0_runner.local_size)
-    val = TestAMD.a.uop.buffer.as_buffer().cast("f")[0]
+    val = TestAMD.a.uop.buffer.as_memoryview().cast("f")[0]
    assert val == 4000.0, f"got val {val}"

 if __name__ == "__main__":
@@ -0,0 +1,126 @@
+# ruff: noqa: F405
+"""Tests for GPU crash scenarios using AMD assembly to trigger invalid operations.
+
+These tests intentionally cause GPU faults to verify error handling.
+Run with: AMD=1 python -m pytest test/external/external_test_gpu_crash.py -v
+"""
+import unittest, re
+from tinygrad.device import Device
+from extra.assembly.amd.autogen.rdna3.ins import *  # noqa: F403
+from extra.assembly.amd.dsl import s, v, Inst, NULL
+
+def assemble(code:str, name:str="test") -> str:
+  kd = {"next_free_vgpr": 8, "next_free_sgpr": 8, "wavefront_size32": 1, "user_sgpr_kernarg_segment_ptr": 1, "kernarg_size": 8}
+  return f".text\n.globl {name}\n.p2align 8\n.type {name},@function\n{name}:\n{code}\n.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"
+
+@unittest.skipIf(Device.DEFAULT != "AMD", "AMD required")
+class TestGPUCrash(unittest.TestCase):
+  @classmethod
+  def setUpClass(cls):
+    from tinygrad.runtime.support.compiler_amd import HIPCompiler
+    cls.dev = Device["AMD"]
+    cls.compiler = HIPCompiler(cls.dev.arch)
+
+  def setUp(self):
+    # Verify device works before each test
+    from tinygrad import Tensor
+    try:
+      t = Tensor([1.0, 2.0], device="AMD").realize()
+      assert (t + 1).numpy().tolist() == [2.0, 3.0]
+    except Exception:
+      self.fail("Device not working before test")
+
+  def _run(self, code: str):
+    from tinygrad.runtime.ops_amd import AMDProgram
+    prg = AMDProgram(self.dev, "test", self.compiler.compile(assemble(code)))
+    prg(self.dev.allocator.alloc(64), global_size=(1,1,1), local_size=(1,1,1), wait=True)
+
+  def _run_insts(self, insts: list[Inst]): self._run("\n".join(i.disasm() for i in insts))
+
+  def _assert_gpu_fault(self, func):
+    """Assert that func raises a RuntimeError indicating a GPU fault (not a setup error)."""
+    with self.assertRaises(RuntimeError) as cm:
+      func()
+    err_msg = str(cm.exception).lower()
+    # Verify it's a GPU fault, not a setup/device initialization error
+    self.assertTrue(
+      re.search(r'fault|hang|timeout|illegal|memviol', err_msg),
+      f"Expected GPU fault error, got: {cm.exception}"
+    )
+
+
+class TestOutOfBoundsMemoryAccess(TestGPUCrash):
+  """Tests for out-of-bounds memory accesses."""
+
+  def test_global_load_null_ptr(self):
+    """Global load from NULL pointer."""
+    insts = [v_mov_b32_e32(v[0], 0), v_mov_b32_e32(v[1], 0),
+             global_load_b32(v[2], addr=v[0:1], saddr=NULL, offset=0), s_waitcnt(0), s_endpgm()]
+    self._assert_gpu_fault(lambda: self._run_insts(insts))
+
+  def test_global_store_null_ptr(self):
+    """Global store to NULL pointer."""
+    insts = [v_mov_b32_e32(v[0], 0), v_mov_b32_e32(v[1], 0), v_mov_b32_e32(v[2], 0xDEADBEEF),
+             global_store_b32(addr=v[0:1], data=v[2], saddr=NULL, offset=0), s_waitcnt(0), s_endpgm()]
+    self._assert_gpu_fault(lambda: self._run_insts(insts))
+
+  def test_global_load_unmapped_high_address(self):
+    """Global load from high unmapped address (0xDEAD00000000)."""
+    insts = [v_mov_b32_e32(v[0], 0x00000000), v_mov_b32_e32(v[1], 0xDEAD),
+             global_load_b32(v[2], addr=v[0:1], saddr=NULL, offset=0), s_waitcnt(0), s_endpgm()]
+    self._assert_gpu_fault(lambda: self._run_insts(insts))
+
+  def test_global_store_unmapped_high_address(self):
+    """Global store to high unmapped address."""
+    insts = [v_mov_b32_e32(v[0], 0x00000000), v_mov_b32_e32(v[1], 0xDEAD), v_mov_b32_e32(v[2], 0x12345678),
+             global_store_b32(addr=v[0:1], data=v[2], saddr=NULL, offset=0), s_waitcnt(0), s_endpgm()]
+    self._assert_gpu_fault(lambda: self._run_insts(insts))
+
+  def test_global_atomic_unmapped(self):
+    """Atomic operation on unmapped memory."""
+    insts = [v_mov_b32_e32(v[0], 0xBEEF0000), v_mov_b32_e32(v[1], 0xDEAD), v_mov_b32_e32(v[2], 1),
+             global_atomic_add_u32(addr=v[0:1], data=v[2], saddr=NULL, offset=0), s_waitcnt(0), s_endpgm()]
+    self._assert_gpu_fault(lambda: self._run_insts(insts))
+
+
+class TestSMEMFaults(TestGPUCrash):
+  """Tests for scalar memory (SMEM) faults."""
+
+  def test_smem_load_null(self):
+    """SMEM load from NULL base."""
+    insts = [s_mov_b32(s[2], 0), s_mov_b32(s[3], 0),
+             s_load_b32(s[4], s[2:3], 0, soffset=NULL), s_waitcnt(0), s_endpgm()]
+    self._assert_gpu_fault(lambda: self._run_insts(insts))
+
+  def test_smem_load_unmapped(self):
+    """SMEM load from unmapped address."""
+    insts = [s_mov_b32(s[2], 0xBEEF0000), s_mov_b32(s[3], 0xDEAD),
+             s_load_b32(s[4], s[2:3], 0, soffset=NULL), s_waitcnt(0), s_endpgm()]
+    self._assert_gpu_fault(lambda: self._run_insts(insts))
+
+
+class TestFlatMemoryFaults(TestGPUCrash):
+  """Tests for FLAT memory instruction faults."""
+
+  def test_flat_load_null(self):
+    """FLAT load from NULL address."""
+    insts = [v_mov_b32_e32(v[0], 0), v_mov_b32_e32(v[1], 0),
+             flat_load_b32(v[2], addr=v[0:1], saddr=NULL, offset=0), s_waitcnt(0), s_endpgm()]
+    self._assert_gpu_fault(lambda: self._run_insts(insts))
+
+  def test_flat_store_null(self):
+    """FLAT store to NULL address."""
+    insts = [v_mov_b32_e32(v[0], 0), v_mov_b32_e32(v[1], 0), v_mov_b32_e32(v[2], 0xDEADBEEF),
+             flat_store_b32(addr=v[0:1], data=v[2], saddr=NULL, offset=0), s_waitcnt(0), s_endpgm()]
+    self._assert_gpu_fault(lambda: self._run_insts(insts))
+
+  def test_flat_atomic_null(self):
+    """FLAT atomic on NULL address."""
+    insts = [v_mov_b32_e32(v[0], 0), v_mov_b32_e32(v[1], 0), v_mov_b32_e32(v[2], 1),
+             flat_atomic_add_u32(addr=v[0:1], data=v[2], saddr=NULL, offset=0), s_waitcnt(0), s_endpgm()]
+    self._assert_gpu_fault(lambda: self._run_insts(insts))
+
+
+if __name__ == "__main__":
+  unittest.main()
@@ -65,7 +65,7 @@ class TestHCQ(unittest.TestCase):
    q.submit(TestHCQ.d0)
    TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
    TestHCQ.d0.timeline_value += 1
-    val = TestHCQ.a.uop.buffer.as_buffer().cast("f")[0]
+    val = TestHCQ.a.uop.buffer.as_memoryview().cast("f")[0]
    assert val == 2000.0, f"got val {val}"

  def test_run_1000_times(self):
@@ -81,7 +81,7 @@ class TestHCQ(unittest.TestCase):
      TestHCQ.compute_queue().signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value).submit(TestHCQ.d0)
      TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
      TestHCQ.d0.timeline_value += 1
-    val = TestHCQ.a.uop.buffer.as_buffer().cast("f")[0]
+    val = TestHCQ.a.uop.buffer.as_memoryview().cast("f")[0]
    assert val == 2000.0, f"got val {val}"

  def test_run_to_3(self):
@@ -95,7 +95,7 @@ class TestHCQ(unittest.TestCase):
    q.signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value).submit(TestHCQ.d0)
    TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
    TestHCQ.d0.timeline_value += 1
-    val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[0]
+    val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[0]
    assert val == 3.0, f"got val {val}"

  def test_update_exec(self):
@@ -106,9 +106,9 @@ class TestHCQ(unittest.TestCase):
    q.signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value).submit(TestHCQ.d0)
    TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
    TestHCQ.d0.timeline_value += 1
-    val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[0]
+    val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[0]
    assert val == 1.0, f"got val {val}"
-    val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[1]
+    val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[1]
    assert val == 0.0, f"got val {val}, should not be updated"

  @unittest.skipUnless(Device.DEFAULT == "NV", "Only NV supports bind")
@@ -126,7 +126,7 @@ class TestHCQ(unittest.TestCase):
      TestHCQ.compute_queue().signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value).submit(TestHCQ.d0)
      TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
      TestHCQ.d0.timeline_value += 1
-    val = TestHCQ.a.uop.buffer.as_buffer().cast("f")[0]
+    val = TestHCQ.a.uop.buffer.as_memoryview().cast("f")[0]
    assert val == 2000.0, f"got val {val}"

  @unittest.skipUnless(Device.DEFAULT == "NV", "Only NV supports bind")
@@ -141,9 +141,9 @@ class TestHCQ(unittest.TestCase):
    q.submit(TestHCQ.d0)
    TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
    TestHCQ.d0.timeline_value += 1
-    val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[0]
+    val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[0]
    assert val == 1.0, f"got val {val}"
-    val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[1]
+    val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[1]
    assert val == 0.0, f"got val {val}, should not be updated"

  @unittest.skipIf(CI, "Can't handle async update on CPU")
@@ -174,7 +174,7 @@ class TestHCQ(unittest.TestCase):
    q.signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value).submit(TestHCQ.d0)
    TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
    TestHCQ.d0.timeline_value += 1
-    val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[0]
+    val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[0]
    assert val == 1.0, f"got val {val}"

  def test_submit_empty_queues(self):
@@ -206,7 +206,7 @@ class TestHCQ(unittest.TestCase):
    q.submit(TestHCQ.d0)
    TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
    TestHCQ.d0.timeline_value += 1
-    val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[0]
+    val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[0]
    assert val == 1.0, f"got val {val}"

  def test_copy_1000_times(self):
@@ -221,7 +221,7 @@ class TestHCQ(unittest.TestCase):
    # confirm the signal didn't exceed the put value
    with self.assertRaises(RuntimeError):
      TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value + 1, timeout=50)
-    val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[1]
+    val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[1]
    assert val == 0.0, f"got val {val}"

  def test_copy(self):
@@ -231,7 +231,7 @@ class TestHCQ(unittest.TestCase):
    q.submit(TestHCQ.d0)
    TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
    TestHCQ.d0.timeline_value += 1
-    val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[1]
+    val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[1]
    assert val == 1.0, f"got val {val}"

  @unittest.skipUnless(Device.DEFAULT == "NV", "Only NV supports bind")
@@ -248,7 +248,7 @@ class TestHCQ(unittest.TestCase):
    # confirm the signal didn't exceed the put value
    with self.assertRaises(RuntimeError):
      TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value + 1, timeout=50)
-    val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[1]
+    val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[1]
    assert val == 0.0, f"got val {val}"

  def test_copy_bandwidth(self):
@@ -288,7 +288,7 @@ class TestHCQ(unittest.TestCase):
    q.submit(TestHCQ.d0)
    TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
    TestHCQ.d0.timeline_value += 1
-    val = TestHCQ.a.uop.buffer.as_buffer().cast("f")[0]
+    val = TestHCQ.a.uop.buffer.as_memoryview().cast("f")[0]
    assert val == 1.0, f"got val {val}"

  def test_cross_device_signal(self):
@@ -319,7 +319,7 @@ class TestHCQ(unittest.TestCase):
      q.signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value).submit(TestHCQ.d0)
      TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
      TestHCQ.d0.timeline_value += 1
-      val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[0]
+      val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[0]
      assert val == 1.0, f"got val {val}"

 if __name__ == "__main__":
@@ -30,7 +30,7 @@ def alloc_rawbuffer(device, fill=False):
  if fill:
    with Context(DEBUG=0):
      data = np.random.randint(-10000, 10000, size=rawbuf.size, dtype=_to_np_dtype(rawbuf.dtype))
-      rawbuf.copyin(Tensor(data).realize().uop.base.realized.as_buffer())
+      rawbuf.copyin(Tensor(data).realize().uop.base.realized.as_memoryview())
  return rawbuf

 def gen_kernel_ji(device, deps):
@@ -93,7 +93,7 @@ def run_jit(jis, all_buffers, input_buffers, var_vals):

  with Context(DEBUG=0):
    res_buffers = []
-    for rawbuf in all_buffers: res_buffers.append(rawbuf.as_buffer())
+    for rawbuf in all_buffers: res_buffers.append(rawbuf.as_memoryview())
    return res_buffers

 def fuzz_graph(jis, all_buffers, input_buffers):
@@ -1,339 +0,0 @@
-import random, traceback, ctypes, argparse, os
-from typing import Any
-import numpy as np
-from collections import defaultdict
-from extra.optimization.helpers import load_worlds, ast_str_to_lin, kern_str_to_lin
-from tinygrad.engine.realize import get_program
-
-# We need to insert ioctl before opening devices.
-if os.getenv("VALIDATE_HCQ", 0) != 0:
-  try:
-    import extra.nv_gpu_driver.nv_ioctl
-    from tinygrad import Device
-    _, _ = Device["NV"], Device["CUDA"]
-  except Exception: pass
-
-  try:
-    import extra.qcom_gpu_driver.opencl_ioctl
-    from tinygrad import Device
-    _, _ = Device["QCOM"], Device["CL"]
-  except Exception: pass
-
-from tinygrad import Tensor, Device, dtypes
-from tinygrad.tensor import _to_np_dtype
-from tinygrad.codegen.opt.kernel import Kernel
-from tinygrad.codegen.opt import Opt, OptOps
-from tinygrad.codegen.opt.search import get_kernel_actions, bufs_from_lin
-from tinygrad.engine.realize import CompiledRunner
-from tinygrad.helpers import getenv, from_mv, prod, colored, Context, DEBUG, Timing
-from tinygrad.uop.ops import UOp, Ops
-from tinygrad.device import is_dtype_supported
-
-def on_linearizer_will_run(): pass
-def on_linearizer_did_run(): pass
-def compare_states(x, y): return (True, "")
-
-if getenv("VALIDATE_HCQ"):
-  if Device.DEFAULT == "NV":
-    print("VALIDATE_HCQ: Comparing NV to CUDA")
-    import extra.nv_gpu_driver.nv_ioctl
-    validate_device = Device["CUDA"]
-    on_linearizer_will_run = extra.nv_gpu_driver.nv_ioctl.before_launch
-    on_linearizer_did_run = extra.nv_gpu_driver.nv_ioctl.collect_last_launch_state
-    compare_states = extra.nv_gpu_driver.nv_ioctl.compare_launch_state
-  elif Device.DEFAULT == "QCOM":
-    print("VALIDATE_HCQ: Comparing QCOM to CL")
-    import extra.qcom_gpu_driver.opencl_ioctl
-    validate_device = Device["CL"]
-    on_linearizer_will_run = extra.qcom_gpu_driver.opencl_ioctl.before_launch
-    on_linearizer_did_run = extra.qcom_gpu_driver.opencl_ioctl.collect_last_launch_state
-    compare_states = extra.qcom_gpu_driver.opencl_ioctl.compare_launch_state
-  else:
-    print(colored("VALIDATE_HCQ options is ignored", 'red'))
-
-def tuplize_uops(uops:list[UOp]) -> tuple:
-  return tuple([(x.op, x.dtype, tuple(uops.index(x) for x in x.src), x.arg) for x in uops])
-
-def get_fuzz_rawbufs(lin):
-  rawbufs = bufs_from_lin(lin)
-
-  # Reallocate output buffer with additional area to detect out-of-bounds writes.
-  RED_AREA_SIZE = 1024
-  # setting output  # TODO: multi-output kernel
-  rawbufs[0] = get_fuzz_rawbuf_like(rawbufs[0], zero=True, size=rawbufs[0].size+RED_AREA_SIZE)
-  # setting inputs
-  with Context(DEBUG=0):
-    for rawbuf in rawbufs[1:]:
-      if dtypes.is_unsigned(rawbuf.dtype):
-        data = np.random.randint(0, 100, size=rawbuf.size, dtype=_to_np_dtype(rawbuf.dtype))
-      elif dtypes.is_int(rawbuf.dtype):
-        data = np.random.randint(-100, 100, size=rawbuf.size, dtype=_to_np_dtype(rawbuf.dtype))
-      elif rawbuf.dtype == dtypes.bool:
-        data = np.random.choice([True, False], size=rawbuf.size)
-      elif rawbuf.dtype == dtypes.half:
-        data = np.random.uniform(-1, 1, size=rawbuf.size).astype(dtype=_to_np_dtype(rawbuf.dtype))
-      else:
-        data = np.random.uniform(-10, 10, size=rawbuf.size).astype(dtype=_to_np_dtype(rawbuf.dtype))
-      rawbuf.copyin(Tensor(data, device=lin.opts.device).realize().uop.base.realized.as_buffer())
-  return rawbufs
-
-def get_fuzz_rawbuf_like(old_rawbuf, zero=False, copy=False, size=None, force_device=None):
-  rawbuf = type(old_rawbuf)(force_device or old_rawbuf.device, old_rawbuf.size if size is None else size, old_rawbuf.dtype).allocate()
-  if copy:
-    with Context(DEBUG=0): rawbuf.copyin(old_rawbuf.as_buffer())
-  elif zero:
-    with Context(DEBUG=0):
-      mv = memoryview(bytearray(rawbuf.size * rawbuf.dtype.itemsize))
-      ctypes.memset(from_mv(mv), 0, len(mv))
-      rawbuf.copyin(mv)
-  return rawbuf
-
-def run_linearizer(lin: Kernel, rawbufs=None, var_vals=None) -> tuple[str, Any]: # (error msg, run state)
-  if rawbufs is None: rawbufs = bufs_from_lin(lin)
-  if var_vals is None: var_vals = {v.expr: v.min for v in lin.vars}
-
-  # TODO: images needs required_optimization
-  try:
-    prg = CompiledRunner(get_program(lin.get_optimized_ast(), lin.opts))
-  except KeyboardInterrupt: raise
-  except Exception:
-    traceback.print_exc()
-    return "COMPILE_ERROR", None
-
-  if getenv("VALIDATE_HCQ"): on_linearizer_will_run()
-  try:
-    prg(rawbufs, var_vals, wait=True)
-  except KeyboardInterrupt: raise
-  except Exception:
-    traceback.print_exc()
-    return "EXEC_ERROR", None
-
-  if getenv("VALIDATE_HCQ"): run_state = on_linearizer_did_run()
-  else: run_state = None
-
-  return "PASS", run_state
-
-def compare_linearizer(lin: Kernel, rawbufs=None, var_vals=None, ground_truth=None, rtol=1e-2, atol=1e-2):
-  # TODO: for bfloat16 it compiles linearizer, but it does not run because numpy cannot generate bf16 buffer.
-  has_bf16 = any(b.dtype.base == dtypes.bfloat16 for b in lin.bufs)
-
-  # TODO: raise specific fuzzing errors instead of str, and propagate the error message
-  try:
-    if rawbufs is None:
-      rawbufs = get_fuzz_rawbufs(lin)
-    else:
-      rawbufs[0] = get_fuzz_rawbuf_like(rawbufs[0], zero=True) # get a new output buffer
-  except KeyboardInterrupt: raise
-  except BaseException:
-    return ("RAWBUFS_ERROR", rawbufs, var_vals, ground_truth, None)
-
-  if var_vals is None:
-    # TODO: handle symbolic max case
-    var_vals = {v.expr: random.randint(v.vmin, v.vmax) for v in lin.ast.variables()}
-
-  if ground_truth is None and not has_bf16:
-    unoptimized = Kernel(lin.ast)
-    if run_linearizer(unoptimized, rawbufs, var_vals)[0] != "PASS":
-      return ("BASELINE_ERROR", rawbufs, var_vals, ground_truth, None)
-    ground_truth = np.frombuffer(rawbufs[0].as_buffer(), _to_np_dtype(rawbufs[0].dtype)).copy()
-
-  rawbufs[0] = get_fuzz_rawbuf_like(rawbufs[0], zero=True) # get a new output buffer
-  run_msg, run_state = run_linearizer(lin, rawbufs, var_vals)
-  if run_msg != "PASS": return (run_msg, rawbufs, var_vals, ground_truth, run_state)
-
-  try:
-    if not has_bf16:
-      result = np.frombuffer(rawbufs[0].as_buffer(), _to_np_dtype(rawbufs[0].dtype))
-      np.testing.assert_allclose(result, ground_truth, rtol=rtol, atol=atol)
-  except KeyboardInterrupt: raise
-  except AssertionError as e:
-    if DEBUG >= 2:
-      print(f"COMPARE_ERROR details: {e}")
-      if getenv("DEBUG_VALUES") > 0:
-        mismatch_indices = np.where(~np.isclose(result, ground_truth, rtol=rtol, atol=atol))
-        mismatched_result = result[mismatch_indices]
-        mismatched_ground_truth = ground_truth[mismatch_indices]
-        for i, idx in enumerate(mismatch_indices[0]):
-          print(f"mismatch at {idx=}: result={mismatched_result[i]} <> ground_truth={mismatched_ground_truth[i]}")
-    return ("COMPARE_ERROR", rawbufs, var_vals, ground_truth, run_state)
-
-  return ("PASS", rawbufs, var_vals, ground_truth, run_state)
-
-def fuzz_linearizer(lin: Kernel, rtol=1e-2, atol=1e-2, opts_list=None):
-  SEED = getenv("SEED", 42)
-  random.seed(SEED)
-  np.random.seed(SEED)
-  print(lin.ast)
-  print(lin.colored_shape())
-  seen_uops = {}
-  last_lins = [lin]
-  failures:defaultdict[str, list[tuple[tuple[UOp, ...], list[Opt]]]] = defaultdict(list)
-  rawbufs, var_vals, ground_truth, validate_rawbufs = None, None, None, None
-
-  FUZZ_ALL_ACTIONS = getenv("FUZZ_ALL_ACTIONS", 0)
-  FUZZ_MAX_SIZE = getenv("FUZZ_MAX_SIZE", 0)
-  FUZZ_IGNORE_SIMPLE_OPS = getenv("FUZZ_IGNORE_SIMPLE_OPS", 1)
-
-  if FUZZ_MAX_SIZE > 0 and prod(lin.full_shape) > FUZZ_MAX_SIZE:
-    print("skipping large kernel")
-    return failures
-  if FUZZ_IGNORE_SIMPLE_OPS and _is_simple(lin):
-    print("skipping simple kernel")
-    return failures
-
-  test_depth = 1 if opts_list is not None else getenv("DEPTH", 1 if FUZZ_ALL_ACTIONS else 10)
-  for depth in range(test_depth):
-    next_lins = []
-    for lin in last_lins:
-      if opts_list is None: actions = get_kernel_actions(lin, include_0=False)
-      else:
-        actions = {}
-        for oi,opts in enumerate(opts_list):
-          lin2 = lin.copy()
-          for o in opts: lin2.apply_opt(o)
-          actions[oi] = lin2
-
-      if not actions: continue
-      if depth == 0 and getenv("FUZZ_REQUIRE_TC", 0):
-        tc_acts = {i: k for k in actions.values() if k.applied_opts[0].op == OptOps.TC}
-        if len(tc_acts) == 0: return failures
-        else: actions = tc_acts
-
-      test_lins = list(actions.values())
-      if FUZZ_ALL_ACTIONS: print(f"testing {lin.applied_opts=} with {len(actions)} actions")
-      elif opts_list is None: test_lins = [random.choice(test_lins)]
-
-      for test_lin in test_lins:
-        if not FUZZ_ALL_ACTIONS and test_lin.applied_opts: print(f"applied opts: {test_lin.applied_opts}")
-
-        # stop if kernel uops repeat
-        try: tuops = tuplize_uops(get_program(test_lin.get_optimized_ast(), test_lin.ren).uops)
-        except KeyboardInterrupt: raise
-        except BaseException as e:
-          print(test_lin.ast)
-          print(test_lin.applied_opts)
-          print(e)
-          failures["LINEARIZE_ERROR"].append((test_lin.ast, test_lin.applied_opts))
-          continue
-
-        if tuops in seen_uops: continue
-        seen_uops[tuops] = tuple(test_lin.applied_opts)
-
-        if not FUZZ_ALL_ACTIONS: print(test_lin.colored_shape())
-
-        (msg, rawbufs, var_vals, ground_truth, state1) = compare_linearizer(test_lin, rawbufs, var_vals, ground_truth, rtol=rtol, atol=atol)
-        if state1 is not None and validate_device is not None:
-          validate_lin = test_lin.copy()
-          validate_lin.ren = validate_device.renderer
-          if validate_rawbufs is None:
-            validate_rawbufs = [get_fuzz_rawbuf_like(x, copy=True, force_device=validate_device.device) for x in rawbufs]
-          (_msg, _, _, _, state2) = compare_linearizer(validate_lin, validate_rawbufs, var_vals, ground_truth, rtol=rtol, atol=atol)
-
-          if _msg != "PASS": failures[f"VALIDATE_DEV_{_msg}"].append((validate_lin.ast, validate_lin.applied_opts))
-
-          ok, err_msg = compare_states(state1, state2)
-          if not ok: failures["HCQ_COMPARE_FAILURE"].append((err_msg, test_lin.ast, test_lin.applied_opts, state1, state2))
-
-        if msg != "PASS":
-          print(test_lin.ast)
-          print(test_lin.applied_opts)
-          print(msg)
-          failures[msg].append((test_lin.ast, test_lin.applied_opts))
-          continue
-
-        next_lins.append(test_lin)
-
-    last_lins = next_lins
-    if FUZZ_ALL_ACTIONS: print(f"depth={depth} total_lins={len(last_lins)} {failures=}")
-  return failures
-
-def _is_simple(lin: Kernel) -> bool:
-  if len(lin.ast.src) > 1: return False
-  ast:UOp = lin.ast.src[0]
-  if ast.src[0].op is Ops.CAST and ast.src[0].src[0].op is Ops.LOAD: return True
-  return False
-
-if __name__ == "__main__":
-  parser = argparse.ArgumentParser(description="Run a fuzz testing on one or more kernels", formatter_class=argparse.ArgumentDefaultsHelpFormatter)
-  parser.add_argument("--ast", type=str, default=None, help="the ast for the kernel to be optimized")
-  parser.add_argument("--file", type=str, default=None, help="a file containing asts to be optimized, one per line")
-  parser.add_argument("--beamreplay", type=str, default=None, help="replay asts and opts got from beam with CAPTURE_BEAM")
-  parser.add_argument("--logfile", type=str, default=None, help="a file containing a tuple of ast and applied_opts, one per line")
-  parser.add_argument("--expected-failures", type=int, default=0, help="the number of expected failed kernels")
-  parser.add_argument("--rtol", type=float, default=1e-2, help="relative tolerance for numerical comparison")
-  parser.add_argument("--atol", type=float, default=1e-2, help="absolute tolerance for numerical comparison")
-  args = parser.parse_args()
-
-  opts_list = None
-  if args.ast is not None:
-    print("loaded AST from CLI")
-    ast_strs = [args.ast]
-  elif args.file is not None:
-    print(f"loading ASTs from file '{args.file}'")
-    with open(args.file, 'r') as file:
-      ast_strs = file.readlines()
-  elif args.beamreplay is not None:
-    print(f"loading BEAM replay from file '{args.beamreplay}'")
-    with open(args.beamreplay, 'r') as file: fdata = file.readlines()
-    ast_strs, opts_list = [x.split(' :: ')[0] for x in fdata if not x.startswith("#")], [x.split(' :: ')[1] for x in fdata if not x.startswith("#")]
-
-    # dedup ast_strs and opts_list
-    dct = defaultdict(list)
-    for i in range(len(ast_strs)): dct[ast_strs[i]].append(eval(opts_list[i]))
-    ast_strs_items = list(dct.keys())
-    opts_list = [dct[c] for c in ast_strs_items]
-  elif args.logfile is not None:
-    print(f"loading ASTs from LOGKERNS file '{args.file}'")
-    with open(args.logfile, 'r') as file:
-      kern_strs = file.readlines()
-      test_lins = [kern_str_to_lin(kern_str) for kern_str in kern_strs]
-      ast_strs = [f"{lin.ast}" for lin in test_lins]
-  else:
-    print("loading ASTs from world")
-    ast_strs = load_worlds(filter_reduce=False, filter_novariable=False)
-
-  print(f"{len(ast_strs)=}")
-  tested = 0
-  failed_ids = []
-  failures = defaultdict(list)
-  seen_ast_strs = set()
-
-  try:
-    for i, ast in enumerate(ast_strs[:getenv("FUZZ_N", len(ast_strs))]):
-      if (nth := getenv("FUZZ_NTH", -1)) != -1 and i != nth: continue
-      if getenv("FUZZ_IMAGEONLY") and "dtypes.image" not in ast: continue
-      if "dtypes.image" in ast and Device.DEFAULT not in {"CL", "QCOM"}: continue  # IMAGE is only for CL
-      if ast in seen_ast_strs: continue
-      seen_ast_strs.add(ast)
-
-      lin = ast_str_to_lin(ast)
-      if not all(is_dtype_supported(buf.dtype) for buf in lin.bufs):
-        print("skipping kernel due to not supported dtype")
-        continue
-
-      with Timing(f"tested ast {i}: "):
-        tested += 1
-        fuzz_failures = fuzz_linearizer(lin, rtol=args.rtol, atol=args.atol, opts_list=(opts_list[i] if opts_list else None))
-        if fuzz_failures: failed_ids.append(i)
-        for k, v in fuzz_failures.items():
-          for f in v:
-            failures[k].append(f)
-  except KeyboardInterrupt: print(colored("STOPPING...", 'red'))
-
-  for msg, errors in failures.items():
-    for i, payload in enumerate(errors):
-      print(f"{msg} {i} kernel: {payload}") # easier to use with output with verify_kernel.py
-
-  print(f"{tested=}")
-  if failures:
-    print(f"{failed_ids=}")
-    for msg, errors in failures.items():
-      print(f"{msg}: {len(errors)}")
-    if len(failed_ids) == args.expected_failures:
-      print(colored(f"{len(failed_ids)} failed as expected", "yellow"))
-  if len(failed_ids) != args.expected_failures:
-    print(colored(f"failed on {len(failed_ids)} kernels, expected {args.expected_failures}", "red"))
-    # TODO: fix this
-    # raise RuntimeError(f"failed on {len(failed_ids)} kernels, expected {args.expected_failures}")
-  else:
-    print(colored("all passed", "green"))
@@ -1,157 +0,0 @@
-from tinygrad import Device, dtypes
-from tinygrad.helpers import getenv, colorize_float, DEBUG
-from extra.optimization.helpers import load_worlds, ast_str_to_lin
-from test.external.fuzz_linearizer import get_fuzz_rawbufs
-from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
-from tinygrad.codegen.opt.search import bufs_from_lin
-from tinygrad.engine.realize import CompiledRunner, get_program
-from tinygrad.tensor import _to_np_dtype
-from tinygrad.runtime.ops_amd import AMDDevice
-from contextlib import contextmanager
-import numpy as np
-import os, random, statistics
-
-am_signal_pages, am_signal_pool, am_devices = [], [], []
-amd_signal_pages, amd_signal_pool, amd_devices = [], [], []
-
-def rebind_vfio(pcibus="0000:44:00.0"):
-  print("rebind ", pcibus)
-  os.system("sudo rmmod amdgpu")
-  os.system("sudo modprobe vfio-pci")
-
-  base = f"/sys/bus/pci/devices/{pcibus}"
-  if os.path.exists(f"{base}/driver"):
-    with open(f"{base}/driver/unbind", "w") as f: f.write(pcibus)
-  with open(f"{base}/driver_override", "w") as f: f.write("vfio-pci")
-  with open("/sys/bus/pci/drivers_probe", "w") as f: f.write(pcibus)
-
-  os.system("sudo modprobe amdgpu")
-  os.system("rocm-smi --setprofile compute")
-  os.system("rocm-smi --setperflevel high")
-
-@contextmanager
-def run_amd():
-  global amd_signal_pages, amd_signal_pool, amd_devices
-  AMDDevice.driverless = False
-  AMDDevice.signal_pages, AMDDevice.signal_pool, AMDDevice.devices = amd_signal_pages, amd_signal_pool, amd_devices
-  yield
-  amd_signal_pages, amd_signal_pool, amd_devices = AMDDevice.signal_pages, AMDDevice.signal_pool, AMDDevice.devices
-  AMDDevice.signal_pages, AMDDevice.signal_pool, AMDDevice.devices = [], [], []
-
-@contextmanager
-def run_am():
-  global am_signal_pages, am_signal_pool, am_devices
-  AMDDevice.driverless = True
-  AMDDevice.signal_pages, AMDDevice.signal_pool, AMDDevice.devices = am_signal_pages, am_signal_pool, am_devices
-  yield
-  am_signal_pages, am_signal_pool, am_devices = AMDDevice.signal_pages, AMDDevice.signal_pool, AMDDevice.devices
-  AMDDevice.signal_pages, AMDDevice.signal_pool, AMDDevice.devices = [], [], []
-
-if __name__ == "__main__":
-  CHECK_CPU = getenv("CHECK_CPU", 0)
-  SEED = getenv("SEED", 42)
-  CNT = getenv("CNT", 7)
-  random.seed(SEED)
-  np.random.seed(SEED)
-
-  # TODO: NUM=780 is super slow
-  # NUM=1907 is broken on AMD and AM have some mismatches (0 vs 1)
-  # kfd feels so bad when taking gpu out while it's running... Need hacks to rebind it before running.
-  rebind_vfio(pcibus="0000:44:00.0")
-
-  ast_strs = load_worlds(filter_reduce=False, filter_novariable=True)
-
-  with run_am():
-    amdev = Device["AMD:1"]
-
-  with run_amd():
-    amddev = Device["AMD"]
-
-  if CHECK_CPU: cpudev = Device["CPU"]
-
-  single = getenv("NUM", -1)
-  if single != -1: ast_strs = ast_strs[single:single+1]
-
-  average_tm_amd, average_tm_am = 0, 0
-  for num,ast in enumerate(ast_strs):
-    with run_amd():
-      amdlin = ast_str_to_lin(ast, opts=amddev.renderer)
-      amdlin.apply_opts(hand_coded_optimizations(amdlin))
-      has_bf16 = any(b.dtype == dtypes.bfloat16 for b in amdlin.bufs)
-
-      amd_prg = CompiledRunner(get_program(amdlin.get_optimized_ast(), amdlin.opts))
-      amdbufs = bufs_from_lin(amdlin)
-      test_amdbufs = get_fuzz_rawbufs(amdlin) if not has_bf16 else amdbufs
-      if not has_bf16: contents = [buf.as_buffer() for buf in test_amdbufs]
-
-    with run_am():
-      rdr = amdev.renderer
-      rdr.device = "AMD:1"
-      amlin = ast_str_to_lin(ast, opts=amdev.renderer)
-      amlin.apply_opts(hand_coded_optimizations(amlin))
-      am_prg = CompiledRunner(get_program(amlin.get_optimized_ast(), amlin.opts))
-      ambufs = bufs_from_lin(amlin)
-      test_ambufs = get_fuzz_rawbufs(amlin) if not has_bf16 else ambufs
-      if not has_bf16:
-        for i,rawbuf in enumerate(test_ambufs): rawbuf.copyin(contents[i])
-
-    if CHECK_CPU:
-      cpu_rdr = cpudev.renderer
-      cpu_rdr.device = "CPU"
-      cpulin = ast_str_to_lin(ast, opts=cpu_rdr)
-      cpulin.apply_opts(hand_coded_optimizations(cpulin))
-      cpu_prg = CompiledRunner(get_program(cpulin.get_optimized_ast(), cpulin.opts))
-      cpubufs = bufs_from_lin(cpulin)
-      test_cpubufs = get_fuzz_rawbufs(cpulin) if not has_bf16 else ambufs
-      if not has_bf16:
-        for i,rawbuf in enumerate(test_cpubufs): rawbuf.copyin(contents[i])
-
-    # warmup
-    tm_amd, tm_am, failed = [], [], False
-    with run_amd():
-      try:
-        amd_prg(test_amdbufs, {}, wait=True)
-        for i in range(CNT): tm_amd.append(amd_prg(amdbufs, {}, wait=True))
-      except RuntimeError:
-        print("AMD FAILED")
-        tm_amd = [1e9]
-        failed = True
-
-    with run_am():
-      try:
-        am_prg(test_ambufs, {}, wait=True)
-        for i in range(CNT): tm_am.append(am_prg(ambufs, {}, wait=True))
-      except RuntimeError:
-        print("AM FAILED")
-        tm_am = [1e9]
-        failed = True
-
-    if CHECK_CPU:
-      cpu_prg(test_cpubufs, {}, wait=True)
-      for i in range(1): cpu_prg(cpubufs, {}, wait=True)
-
-    if not failed and not has_bf16:
-      with run_amd():
-        curesult = np.frombuffer(test_amdbufs[0].as_buffer(), _to_np_dtype(test_amdbufs[0].dtype))
-
-      with run_am():
-        amresult = np.frombuffer(test_ambufs[0].as_buffer(), _to_np_dtype(test_ambufs[0].dtype))
-
-      if CHECK_CPU:
-        cpuresult = np.frombuffer(test_cpubufs[0].as_buffer(), _to_np_dtype(test_cpubufs[0].dtype))
-        np.testing.assert_allclose(amresult, cpuresult, rtol=1e-2, atol=1e-2)
-        np.testing.assert_allclose(curesult, cpuresult, rtol=1e-2, atol=1e-2)
-
-      try:
-        np.testing.assert_allclose(curesult, amresult, rtol=1e-2, atol=1e-2)
-      except AssertionError as e:
-        print("AM and AMD results do not match")
-        print(e)
-
-    bam = statistics.median(tm_am)
-    bamd = statistics.median(tm_amd)
-    average_tm_amd += bamd
-    average_tm_am += bam
-    ratio = bam/bamd
-    print(f"{average_tm_am/average_tm_amd:5.2f}x -- {num:4d} {colorize_float(ratio)} {bam*1e6:7.2f} vs {bamd*1e6:7.2f} us", amlin.name)
-    if DEBUG > 3 and ratio > 1.04: print(f"AM slower {ratio}", amlin.ast, amlin.applied_opts)
@@ -1,69 +0,0 @@
-from tinygrad import Device, dtypes
-from tinygrad.helpers import getenv, colorize_float
-from extra.optimization.helpers import load_worlds, ast_str_to_lin
-from test.external.fuzz_linearizer import get_fuzz_rawbufs
-from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
-from tinygrad.codegen.opt.search import bufs_from_lin
-from tinygrad.engine.realize import CompiledRunner, get_program
-from tinygrad.tensor import _to_np_dtype
-import numpy as np
-
-if __name__ == "__main__":
-  ast_strs = load_worlds(filter_reduce=False, filter_novariable=True)
-  cudev = Device["CUDA"]
-  nvdev = Device["NV"]
-
-  # NUM=112 python3 test/external/speed_compare_cuda_nv.py
-
-  single = getenv("NUM", -1)
-  if single != -1: ast_strs = ast_strs[single:single+1]
-
-  average_tm_cuda, average_tm_nv = 0, 0
-  for num,ast in enumerate(ast_strs):
-    # cuda compile
-    culin = ast_str_to_lin(ast, opts=cudev.renderer)
-    culin.apply_opts(hand_coded_optimizations(culin))
-    has_bf16 = any(b.dtype == dtypes.bfloat16 for b in culin.bufs)
-
-    cuda_prg = CompiledRunner(get_program(culin.get_optimized_ast(), culin.opts))
-    cubufs = bufs_from_lin(culin)
-    test_cubufs = get_fuzz_rawbufs(culin) if not has_bf16 else cubufs
-
-    rdr = nvdev.renderer
-    rdr.device = "NV"
-    nvlin = ast_str_to_lin(ast, opts=rdr)
-    nvlin.apply_opts(hand_coded_optimizations(nvlin))
-    nv_prg = CompiledRunner(get_program(nvlin.get_optimized_ast(), nvlin.opts))
-    nvbufs = bufs_from_lin(nvlin)
-    test_nvbufs = get_fuzz_rawbufs(nvlin) if not has_bf16 else nvbufs
-    if not has_bf16:
-      for i,rawbuf in enumerate(test_nvbufs): rawbuf.copyin(test_cubufs[i].as_buffer())
-
-    # warmup
-    tm_cuda, tm_nv, failed = [], [], False
-    try:
-      cuda_prg(test_cubufs, {}, wait=True)
-      for i in range(5): tm_cuda.append(cuda_prg(cubufs, {}, wait=True))
-    except RuntimeError:
-      print("CUDA FAILED")
-      tm_cuda = [1e9]
-      failed = True
-
-    try:
-      nv_prg(test_nvbufs, {}, wait=True)
-      for i in range(5): tm_nv.append(nv_prg(nvbufs, {}, wait=True))
-    except RuntimeError:
-      print("NV FAILED")
-      tm_nv = [1e9]
-      failed = True
-
-    if not failed and not has_bf16:
-      curesult = np.frombuffer(test_cubufs[0].as_buffer(), _to_np_dtype(test_cubufs[0].dtype))
-      nvresult = np.frombuffer(test_nvbufs[0].as_buffer(), _to_np_dtype(test_nvbufs[0].dtype))
-      np.testing.assert_allclose(curesult, nvresult, rtol=1e-2, atol=1e-2)
-
-    average_tm_cuda += min(tm_cuda)
-    average_tm_nv += min(tm_nv)
-    ratio = min(tm_nv)/min(tm_cuda)
-    print(f"{average_tm_nv/average_tm_cuda:5.2f}x -- {num:4d} {colorize_float(ratio)}  {min(tm_nv)*1e6:7.2f} us", nvlin.name)
-    if ratio > 1.04: print(f"NV slower {ratio}", nvlin.ast, nvlin.applied_opts)
@@ -1,60 +0,0 @@
-import itertools
-from tinygrad import Device
-from tinygrad.engine.realize import CompiledRunner, get_program
-from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
-from tinygrad.helpers import getenv, colorize_float
-from extra.optimization.helpers import load_worlds, ast_str_to_lin
-from tinygrad.codegen.opt.search import bufs_from_lin
-from tinygrad.runtime.ops_cuda import PTXCompiler, PTXRenderer, CUDACompiler
-
-if __name__ == "__main__":
-  ast_strs = load_worlds(filter_reduce=False, filter_novariable=True)
-  # no bfloat16 for ptx at the moment
-  ast_strs = [x for x in ast_strs if "dtypes.bfloat16" not in x]
-  dev = Device["CUDA"]
-  ptx = PTXRenderer(dev.arch)
-
-  # NUM=112 python3 test/external/speed_compare_cuda_ptx.py
-
-  single = getenv("NUM", -1)
-  if single != -1: ast_strs = ast_strs[single:single+1]
-
-  average_tm_cuda, average_tm_ptx = 0, 0
-  for num,ast in enumerate(ast_strs):
-    # cuda compile
-    dev.compiler = CUDACompiler(dev.arch)
-    lin = ast_str_to_lin(ast, opts=dev.renderer)
-    lin.apply_opts(hand_coded_optimizations(lin))
-    cuda_prg = CompiledRunner(get_program(lin.get_optimized_ast(), lin.opts))
-
-    bufs = bufs_from_lin(lin)
-
-    # ptx compile
-    dev.compiler = PTXCompiler(dev.arch)
-    lin = ast_str_to_lin(ast, opts=ptx)
-    lin.apply_opts(hand_coded_optimizations(lin))
-    ptx_prg = CompiledRunner(get_program(lin.get_optimized_ast(), lin.opts))
-
-    # warmup
-    try:
-      cuda_prg(bufs, {}, wait=True)
-    except RuntimeError:
-      print("cuda failed ast:", num)
-      continue
-    ptx_prg(bufs, {}, wait=True)
-
-    tm_cuda, tm_ptx = [], []
-    for i in range(5):
-      tm_cuda.append(cuda_prg(bufs, {}, wait=True))
-      tm_ptx.append(ptx_prg(bufs, {}, wait=True))
-    average_tm_cuda += min(tm_cuda)
-    average_tm_ptx += min(tm_ptx)
-    ratio = min(tm_ptx)/min(tm_cuda)
-    print(f"{average_tm_ptx/average_tm_cuda:5.2f}x -- {num:4d} {colorize_float(ratio)}  {min(tm_ptx)*1e6:7.2f} us", lin.name)
-    if ratio > 1.5:
-      def fix(x): return x.replace('\t', ' ').strip()
-      ll1, ll2 = cuda_prg.lib.decode().split('\n'), ptx_prg.lib.decode().split('\n')
-      if single != -1:
-        for ln, (l1, l2) in enumerate(itertools.zip_longest(ll1, ll2, fillvalue='')):
-          print(f"{ln:5d} | {fix(l1):80s} | {fix(l2):80s}")
-      print(len(ll1), len(ll2), "RATIO", ratio, "us", min(tm_ptx)*1e6)
@@ -1,78 +0,0 @@
-import argparse
-from collections import defaultdict
-from extra.optimization.helpers import kern_str_to_lin, time_linearizer
-from test.external.fuzz_linearizer import compare_linearizer
-from tinygrad.helpers import colored
-from tinygrad.codegen.opt.kernel import Kernel
-
-# Use this with the LOGKERNS options to verify that all executed kernels are valid and evaluate to the same ground truth results
-
-# Example for GPT2:
-# 1) Run the model to log all kernels: `PYTHONPATH=. LOGKERNS=/tmp/gpt2_kerns.txt JIT=1 HALF=1 BEAM=2 CACHELEVEL=0 python3 examples/gpt2.py --count 10 --temperature 0 --timing`   # noqa: E501
-# 2) Validate the kernel correctness: `PYTHONPATH=. python3 ./test/external/verify_kernel.py --file /tmp/gpt2_kerns.txt`
-
-if __name__ == "__main__":
-  parser = argparse.ArgumentParser(description="Verify the correctness of one or more kernel", formatter_class=argparse.ArgumentDefaultsHelpFormatter)    # noqa: E501
-  parser.add_argument("--kernel", type=str, default=None, help="a string of a tuple of (ast, applied_opts,)")
-  parser.add_argument("--file", type=str, default=None, help="a file containing a tuple of ast and applied_opts, one per line")
-  parser.add_argument("--pkl", type=str, default=None, help="a pickle file containing a single tuple of ast and applied_opts")
-  parser.add_argument("--rtol", type=float, default=1e-2, help="relative tolerance for numerical comparison")
-  parser.add_argument("--atol", type=float, default=1e-2, help="absolute tolerance for numerical comparison")
-  parser.add_argument("--timing", action='store_true', help="show final timing for the kernel")
-  parser.add_argument("--expected-failures", type=int, default=0, help="the number of expected failed kernels")
-  args = parser.parse_args()
-
-  if args.kernel is not None:
-    print("loading kernel from args")
-    test_lins = [kern_str_to_lin(args.kernel)]
-  elif args.file is not None:
-    print(f"loading kernel from file '{args.file}'")
-    with open(args.file, 'r') as file:
-      kern_strs = file.readlines()
-      test_lins = [kern_str_to_lin(kern_str) for kern_str in kern_strs]
-  elif args.pkl is not None:
-    print(f"loading kernel from pickle file '{args.file}'")
-    import pickle
-    with open(args.pkl, 'rb') as file:
-      (ast, applied_opts,) = pickle.load(file)
-      lin = Kernel(ast)
-      lin.apply_opts(applied_opts)
-      test_lins = [lin]
-
-  else:
-    raise RuntimeError("no kernel specified; use --kernel, --file, or --pkl options")
-
-  print(f"verifying {len(test_lins)} kernels")
-
-  failed_ids = []
-  failures = defaultdict(list)
-  for i, test_lin in enumerate(test_lins):
-    print(f"testing kernel {i}")
-    print(test_lin.ast)
-    print(test_lin.applied_opts)
-    unoptimized_lin = Kernel(test_lin.ast)
-    print(f"{unoptimized_lin.colored_shape()} -> {test_lin.colored_shape()}")
-    (msg,rb,vv,gt) = compare_linearizer(test_lin, None, None, None, rtol=args.rtol, atol=args.atol)
-    if msg != "PASS":
-      failed_ids.append(i)
-      failures[msg].append((test_lin.ast, test_lin.applied_opts))
-    if args.timing:
-      tm = time_linearizer(test_lin, rb, allow_test_size=False, cnt=10)
-      print(f"final time {tm*1e6:9.0f} us")
-
-  for msg, errors in failures.items():
-    for i, (ast, opts) in enumerate(errors):
-      print(f"{msg} {i} AST: {ast}")
-      print(f"{msg} {i} OPTS: {opts}\n")
-
-  print(f"tested {len(test_lins)} kernels")
-  if failures:
-    print(f"{failed_ids=}")
-    for msg, errors in failures.items():
-      print(f"{msg}: {len(errors)}")
-    if len(failed_ids) == args.expected_failures:
-      print(colored(f"{len(failed_ids)} failed as expected", "yellow"))
-  if len(failed_ids) != args.expected_failures:
-    raise RuntimeError(f"failed on {len(failed_ids)} kernels, expected {args.expected_failures}")
-  else:
-    print(colored("all passed", "green"))
@@ -61,7 +61,7 @@ def eval_uop(uop:UOp, inputs:list[tuple[DType, list[Any]]]|None=None):
  for buf_dt, data in inputs or []:
    bufs.append(buf:=allocator.alloc(len(data) * buf_dt.itemsize))
    allocator._copyin(buf, memoryview(struct.pack(str(len(data)) + (buf_dt.fmt or ""), *data)))
-  g = UOp(Ops.DEFINE_GLOBAL, uop.dtype.ptr(), arg=0, src=())
+  g = UOp(Ops.PARAM, uop.dtype.ptr(), arg=0, src=())
  prg = get_program(UOp.store(g.index(UOp.const(dtypes.int, 0)), uop).sink(), PythonRenderer())
  prog = PythonProgram("run", PythonCompiler().compile(prg.src))
  prog(out_buf:=allocator.alloc(uop.dtype.itemsize), *bufs)
@@ -4,7 +4,7 @@ import tinygrad.runtime.autogen.am.am as am
 import tinygrad.runtime.autogen.amdgpu_drm as amdgpu_drm
 from tinygrad.helpers import from_mv
 from test.mockgpu.driver import VirtDriver, VirtFileDesc, TextFileDesc, DirFileDesc, VirtFile
-from test.mockgpu.amd.amdgpu import AMDGPU, gpu_props
+from test.mockgpu.amd.amdgpu import AMDGPU, gpu_props, GFX_TARGET_VERSION, MOCKGPU_ARCH

 libc = ctypes.CDLL(ctypes.util.find_library("c"))
 libc.mmap.argtypes = [ctypes.c_void_p, ctypes.c_size_t, ctypes.c_int, ctypes.c_int, ctypes.c_int, ctypes.c_long]
@@ -90,35 +90,30 @@ class AMDDriver(VirtDriver):
  def _prepare_gpu(self, gpu_id):
    self.doorbells[gpu_id] = memoryview(bytearray(0x2000))
    self.gpus[gpu_id] = AMDGPU(gpu_id)
+    # IP versions: rdna3 = GC 11.0.0, NBIF 4.3.0; rdna4 = GC 12.0.0, NBIF 6.3.1
+    ip_versions = {"rdna3": {"gc": (11, 0, 0), "sdma": (6, 0, 0), "nbif": (4, 3, 0)},
+                   "rdna4": {"gc": (12, 0, 0), "sdma": (6, 0, 0), "nbif": (6, 3, 1)}}[MOCKGPU_ARCH]
+    def ip_discovery_files(hwid, ver, base_addr):
+      p = f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{hwid}/0'
+      return [VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{hwid}', functools.partial(DirFileDesc, child_names=['0'])),
+              VirtFile(f'{p}/major', functools.partial(TextFileDesc, text=str(ver[0]))),
+              VirtFile(f'{p}/minor', functools.partial(TextFileDesc, text=str(ver[1]))),
+              VirtFile(f'{p}/revision', functools.partial(TextFileDesc, text=str(ver[2]))),
+              VirtFile(f'{p}/base_addr', functools.partial(TextFileDesc, text=base_addr))]
    self.tracked_files += [
      VirtFile('/sys/module/amdgpu', functools.partial(TextFileDesc, text="1")),
      VirtFile('/sys/module/amdgpu/parameters/ppfeaturemask', functools.partial(TextFileDesc, text="0xffff3fff")),
      VirtFile(f'/sys/devices/virtual/kfd/kfd/topology/nodes/{gpu_id}', functools.partial(DirFileDesc, child_names=['gpu_id', 'properties'])),
      VirtFile(f'/sys/devices/virtual/kfd/kfd/topology/nodes/{gpu_id}/gpu_id', functools.partial(TextFileDesc, text=f"{gpu_id}")),
      VirtFile(f'/sys/devices/virtual/kfd/kfd/topology/nodes/{gpu_id}/properties',
-        functools.partial(TextFileDesc, text=gpu_props.format(drm_render_minor=gpu_id))),
+        functools.partial(TextFileDesc, text=gpu_props.format(drm_render_minor=gpu_id, gfx_target_version=GFX_TARGET_VERSION))),
      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/power_dpm_force_performance_level',
               functools.partial(TextFileDesc, text='profile_standard\n')),
      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0',
               functools.partial(DirFileDesc, child_names=[str(am.GC_HWID), str(am.SDMA0_HWID), str(am.NBIF_HWID)])),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.GC_HWID}', functools.partial(DirFileDesc, child_names=['0'])),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.GC_HWID}/0/major', functools.partial(TextFileDesc, text='11')),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.GC_HWID}/0/minor', functools.partial(TextFileDesc, text='0')),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.GC_HWID}/0/revision', functools.partial(TextFileDesc, text='0')),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.GC_HWID}/0/base_addr',
-               functools.partial(TextFileDesc, text='0x00001260\n0x0000A000\n0x0001C000\n0x02402C00')),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.SDMA0_HWID}', functools.partial(DirFileDesc, child_names=['0'])),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.SDMA0_HWID}/0/major', functools.partial(TextFileDesc, text='6')),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.SDMA0_HWID}/0/minor', functools.partial(TextFileDesc, text='0')),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.SDMA0_HWID}/0/revision', functools.partial(TextFileDesc, text='0')),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.SDMA0_HWID}/0/base_addr',
-               functools.partial(TextFileDesc, text='0x00001260\n0x0000A000\n0x0001C000\n0x02402C00')),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.NBIF_HWID}', functools.partial(DirFileDesc, child_names=['0'])),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.NBIF_HWID}/0/major', functools.partial(TextFileDesc, text='4')),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.NBIF_HWID}/0/minor', functools.partial(TextFileDesc, text='3')),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.NBIF_HWID}/0/revision', functools.partial(TextFileDesc, text='0')),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.NBIF_HWID}/0/base_addr',
-               functools.partial(TextFileDesc, text='0x00000000\n0x00000014\n0x00000D20\n0x00010400\n0x0241B000\n0x04040000')),
+      *ip_discovery_files(am.GC_HWID, ip_versions["gc"], '0x00001260\n0x0000A000\n0x0001C000\n0x02402C00'),
+      *ip_discovery_files(am.SDMA0_HWID, ip_versions["sdma"], '0x00001260\n0x0000A000\n0x0001C000\n0x02402C00'),
+      *ip_discovery_files(am.NBIF_HWID, ip_versions["nbif"], '0x00000000\n0x00000014\n0x00000D20\n0x00010400\n0x0241B000\n0x04040000'),
      VirtFile(f'/dev/dri/renderD{gpu_id}', functools.partial(DRMFileDesc, driver=self, gpu=f"{self.gpus[gpu_id]}")),
    ]

@@ -1,8 +1,11 @@
 import ctypes, time
 from test.mockgpu.gpu import VirtGPU
 from test.mockgpu.helpers import _try_dlopen_remu
-from tinygrad.helpers import getbits, to_mv
+from tinygrad.helpers import getbits, to_mv, getenv
 from tinygrad.runtime.support import c
+
+MOCKGPU_ARCH = getenv("MOCKGPU_ARCH", "rdna3")
+GFX_TARGET_VERSION = {"rdna3": 110000, "rdna4": 120000}[MOCKGPU_ARCH]
 import tinygrad.runtime.autogen.amd_gpu as amd_gpu, tinygrad.runtime.autogen.am.pm4_nv as pm4

 SDMA_MAX_COPY_SIZE = 0x400000
@@ -194,10 +197,11 @@ class PM4Executor(AMDQueue):
    scratch_size = wavesize * 4  # This gives the scratch size per thread (lane)

    assert prg_sz > 0, "Invalid prg ptr (not found in mapped ranges)"
-    # Pass valid memory ranges, rsrc2, and scratch_size to Python emulator
+    # Pass valid memory ranges, rsrc2, scratch_size and arch to Python emulator
    if hasattr(remu, 'valid_mem_ranges'): remu.valid_mem_ranges = self.gpu.mapped_ranges
    if hasattr(remu, 'rsrc2'): remu.rsrc2 = rsrc2
    if hasattr(remu, 'scratch_size'): remu.scratch_size = scratch_size
+    if hasattr(remu, 'arch'): remu.arch = self.gpu.arch
    err = remu.run_asm(prg_addr, prg_sz, *gl, *lc, args_addr)
    if err != 0: raise RuntimeError("remu does not support the new instruction introduced in this kernel")

@@ -314,6 +318,7 @@ class AMDGPU(VirtGPU):
    self.regs = AMDGPURegisters()
    self.mapped_ranges = set()
    self.queues = []
+    self.arch = MOCKGPU_ARCH

  def map_range(self, vaddr, size): self.mapped_ranges.add((vaddr, size))
  def unmap_range(self, vaddr, size): self.mapped_ranges.remove((vaddr, size))
@@ -342,7 +347,7 @@ simd_arrays_per_engine 2
 cu_per_simd_array 8
 simd_per_cu 2
 max_slots_scratch_cu 32
-gfx_target_version 110000
+gfx_target_version {gfx_target_version}
 vendor_id 4098
 device_id 29772
 location_id 34304
@@ -16,14 +16,15 @@ def _try_dlopen_gpuocelot():
  return None

 class PythonRemu:
-  """Python RDNA3 emulator wrapper that matches the libremu.so interface."""
+  """Python RDNA3/RDNA4 emulator wrapper that matches the libremu.so interface."""
  valid_mem_ranges: set[tuple[int, int]] = set()
  rsrc2: int = 0x19c  # Default: USER_SGPR_COUNT=14, enable X and Y workgroup IDs
  scratch_size: int = 0  # private_segment_fixed_size from kernel descriptor
+  arch: str = "rdna3"  # Architecture: rdna3 or rdna4

  def run_asm(self, lib: int, lib_sz: int, gx: int, gy: int, gz: int, lx: int, ly: int, lz: int, args_ptr: int) -> int:
    from extra.assembly.amd.emu import run_asm
-    return run_asm(lib, lib_sz, gx, gy, gz, lx, ly, lz, args_ptr, self.rsrc2, self.scratch_size)
+    return run_asm(lib, lib_sz, gx, gy, gz, lx, ly, lz, args_ptr, self.rsrc2, self.scratch_size, self.arch)

 def _try_dlopen_remu():
  # Use Python emulator only if PYTHON_REMU=1
@@ -42,7 +42,9 @@ def _memoryview(cls, mem):
      for st,en,rcb,wcb in d.tracked_addresses:
        if st <= addr <= en: return TrackedMemoryView(mem, rcb, wcb)
  return original_memoryview(mem)
-builtins.memoryview = type("memoryview", (), {'__new__': _memoryview}) # type: ignore
+class _MockMemoryviewMeta(type):
+  def __instancecheck__(cls, instance): return isinstance(instance, (original_memoryview, TrackedMemoryView))
+builtins.memoryview = _MockMemoryviewMeta("memoryview", (), {'__new__': _memoryview}) # type: ignore

 def _open(path, flags):
  for d in drivers:
@@ -73,7 +73,7 @@ class TestHuggingFaceOnnxModels(unittest.TestCase):
    onnx_model_path = snapshot_download_with_retry(
      repo_id=repo_id,
      allow_patterns=["*.onnx", "*.onnx_data"],
-      cache_dir=str(DOWNLOADS_DIR)
+      local_dir=DOWNLOADS_DIR / repo_id
    )
    onnx_model_path = onnx_model_path / model_file
    file_size = onnx_model_path.stat().st_size
@@ -0,0 +1,39 @@
+import unittest
+from tinygrad import Tensor, dtypes, TinyJit, UOp
+from tinygrad.apps.llm import apply_rope as apply_rope_new, precompute_freqs_cis
+
+def apply_rope(x:Tensor, start_pos:int):
+  B, H, T, Hd = x.shape
+  precompute_freqs_cis.cache_clear()
+  freqs_cis = precompute_freqs_cis(Hd, start_pos+T)[start_pos:start_pos+T]
+  return apply_rope_new(x, freqs_cis)
+
+class TestAttention(unittest.TestCase):
+  def test_half_qkv_buffers(self):
+    BS, seqlen, dim = 10, 4, 100
+    q = Tensor.ones(BS, seqlen, dim, dtype=dtypes.half).contiguous().realize()
+    k = Tensor.ones(BS, seqlen, dim, dtype=dtypes.half).contiguous().realize()
+    v = Tensor.ones(BS, seqlen, dim, dtype=dtypes.half).contiguous().realize()
+    attn = q.scaled_dot_product_attention(k, v)
+    sched = attn.schedule()
+    # attention has 4 kernels now
+    self.assertEqual(len(sched), 4)
+
+  def test_apply_rope_jit_prune(self):
+    def rope_fn(x_in, pos): return apply_rope(x_in, pos)
+    rope_noprune = TinyJit(rope_fn)
+    rope_prune = TinyJit(rope_fn, prune=True)
+
+    v_pos = UOp.variable("start_pos", 0, 100)
+    for _ in range(3):
+      rope_noprune(Tensor.randn(1, 2, 4, 8, dtype=dtypes.float32), v_pos.bind(1))
+      rope_prune(Tensor.randn(1, 2, 4, 8, dtype=dtypes.float32), v_pos.bind(1))
+    noprune_size = len(rope_noprune.captured.jit_cache)
+    prune_size = len(rope_prune.captured.jit_cache)
+
+    self.assertGreater(noprune_size, prune_size)
+    self.assertGreaterEqual(noprune_size, 2)
+    self.assertEqual(prune_size, 1)
+
+if __name__ == '__main__':
+  unittest.main()
@@ -5,24 +5,13 @@ from tinygrad.runtime.support.c import DLL, record, init_records
 from tinygrad.runtime.support import c
 from tinygrad.runtime.support.autogen import gen

-class TestAutogen(unittest.TestCase):
+@unittest.skipIf(WIN, "doesn't compile on windows")
+class TestC(unittest.TestCase):
  def compile(self, src):
    with tempfile.NamedTemporaryFile(suffix=".so") as f:
      subprocess.check_output(('clang', '-x', 'c', '-fPIC', '-shared', '-', '-o', f.name), input=src.encode())
      return DLL("test", f.name)

-  def run_gen(self, contents):
-    with tempfile.NamedTemporaryFile(mode='w', suffix='.h') as f:
-      f.write(contents)
-      f.flush()
-
-      generated_code = gen(name="test_header", dll=None, files=[f.name])
-
-      namespace = {}
-      exec(generated_code, namespace)
-      return namespace
-
-  @unittest.skipIf(WIN, "doesn't compile on windows")
  def test_packed_struct(self):
    @record
    class Baz:
@@ -45,7 +34,6 @@ class TestAutogen(unittest.TestCase):
    assert b.c == 1
    assert b.d == 0

-  @unittest.skipIf(WIN, "doesn't compile on windows")
  def test_packed_struct_interop(self):
    @record
    class Baz:
@@ -75,7 +63,6 @@ class TestAutogen(unittest.TestCase):
    self.assertEqual(test(b), b.a + b.b + b.c + b.d)

  # https://github.com/python/cpython/issues/90914
-  @unittest.skipIf(WIN, "doesn't compile on windows")
  def test_bitfield_interop(self):
    @record
    class Baz:
@@ -103,7 +90,6 @@ class TestAutogen(unittest.TestCase):
    def test(x:Baz) -> ctypes.c_int: ...
    for i in range(8): self.assertEqual(test(Baz(*(j==i for j in range(8)))), i==2)

-  @unittest.skipIf(WIN, "doesn't compile on windows")
  def test_struct_interop(self):
    @record
    class Baz:
@@ -131,7 +117,6 @@ class TestAutogen(unittest.TestCase):
    def test(x:Baz) -> Baz: ...
    self.assertEqual(bytes(test(Baz(*range(8)))), struct.pack("8i", *range(7, -1, -1)))

-  @unittest.skipIf(WIN, "doesn't compile on windows")
  def test_aos_interop(self):
    @record
    class Item:
@@ -151,7 +136,6 @@ class TestAutogen(unittest.TestCase):
    def test(arr:(Item * 3)) -> ctypes.c_int: ...
    self.assertEqual(test((Item * 3)(Item(10), Item(20), Item(30))), 60)

-  @unittest.skipIf(WIN, "doesn't compile on windows")
  def test_soa_interop(self):
    @record
    class Row:
@@ -173,7 +157,6 @@ class TestAutogen(unittest.TestCase):
    self.assertEqual(r.data[1], 20)
    self.assertEqual(r.data[2], 10)

-  @unittest.skipIf(WIN, "doesn't compile on windows")
  def test_soa_ptr_interop(self):
    @record
    class Row:
@@ -191,7 +174,6 @@ class TestAutogen(unittest.TestCase):
    def test(x:Row) -> ctypes.c_int: ...
    assert test(Row((ctypes.c_int * 3)(10, 20, 30))) == 60

-  @unittest.skipIf(WIN, "doesn't compile on windows")
  def test_nested_struct_interop(self):
    @record
    class Inner:
@@ -217,7 +199,6 @@ class TestAutogen(unittest.TestCase):
    self.assertEqual(o.inner.a, 20)
    self.assertEqual(o.b, 10)

-  @unittest.skipIf(WIN, "doesn't compile on windows")
  def test_struct_pointer_interop(self):
    @record
    class Foo:
@@ -242,7 +223,88 @@ class TestAutogen(unittest.TestCase):
    self.assertEqual(out.contents.a, 20)
    self.assertEqual(out.contents.b, 10)

-  @unittest.skipIf(WIN, "doesn't compile on windows")
+  def test_pointer_field_roundtrip(self):
+    # This tests storing a pointer in a record struct field and passing it to C
+    # Mimics how mesa.struct_lp_build_tgsi_params.mask is used
+    from tinygrad.runtime.support.c import POINTER
+    @record
+    class Inner:
+      SIZE = 8
+      value: Annotated[ctypes.c_int, 0]
+      flag: Annotated[ctypes.c_int, 4]
+    @record
+    class Outer:
+      SIZE = 16
+      x: Annotated[ctypes.c_int, 0]
+      inner_ptr: Annotated[POINTER[Inner], 8]
+    init_records()
+
+    src = """
+      struct inner { int value; int flag; };
+      struct outer { int x; struct inner *inner_ptr; };
+      int test(struct inner *p) {
+        return p->value + p->flag;
+      }
+    """
+    dll = self.compile(src)
+    @dll.bind
+    def test(p:POINTER[Inner]) -> ctypes.c_int: ...
+
+    inner = Inner(value=42, flag=10)
+    outer = Outer(x=1, inner_ptr=ctypes.pointer(inner))
+    # Retrieve pointer from struct field and pass to C
+    self.assertEqual(test(outer.inner_ptr), 52)
+
+  def test_pointer_field_loses_reference(self):
+    # BUG: When a pointer is stored in a record struct field, only the address bytes are saved.
+    # The pointer's _objects dict (which prevents GC of the pointed-to object) is lost.
+    # This causes the pointed-to object to be garbage collected, leading to use-after-free.
+    from tinygrad.runtime.support.c import POINTER
+    @record
+    class MaskContext:
+      SIZE = 16
+      value: Annotated[ctypes.c_int, 0]
+      initialized: Annotated[ctypes.c_int, 4]
+      ptr: Annotated[ctypes.c_void_p, 8]
+    @record
+    class Params:
+      SIZE = 16
+      x: Annotated[ctypes.c_int, 0]
+      mask: Annotated[POINTER[MaskContext], 8]
+    init_records()
+
+    src = """
+      struct mask_ctx { int value; int initialized; void *ptr; };
+      void mask_begin(struct mask_ctx *m, int val) { m->value = val; m->initialized = 1; }
+      int mask_end(struct mask_ctx *m) { return m->value + m->initialized; }
+    """
+    dll = self.compile(src)
+    @dll.bind
+    def mask_begin(m:POINTER[MaskContext], val:ctypes.c_int) -> None: ...
+    @dll.bind
+    def mask_end(m:POINTER[MaskContext]) -> ctypes.c_int: ...
+
+    # When MaskContext() is created inline, it gets garbage collected after the pointer
+    # is stored because only the address bytes are saved, not the _objects reference.
+    params = Params(x=1, mask=ctypes.pointer(MaskContext()))
+    mask_begin(params.mask, 42)
+    result = mask_end(params.mask)
+    self.assertEqual(result, 43)  # 42 + 1
+
+@unittest.skipIf(OSX and ('MTLCompiler' in DLL._loaded_ or 'llvm' in DLL._loaded_), "libclang can't be loaded after MTLCompiler or llvm on OSX")
+@unittest.skipIf(WIN, "doesn't compile on windows")
+class TestAutogen(unittest.TestCase):
+  def run_gen(self, contents):
+    with tempfile.NamedTemporaryFile(mode='w', suffix='.h') as f:
+      f.write(contents)
+      f.flush()
+
+      generated_code = gen(name="test_header", dll=None, files=[f.name])
+
+      namespace = {}
+      exec(generated_code, namespace)
+      return namespace
+
  def test_packed_structs(self):
    ns = self.run_gen("""
 typedef unsigned NvU32;
@@ -292,47 +354,6 @@ typedef struct
    assert frts_cmd.readVbiosDesc.__class__ is FWSECLIC_READ_VBIOS_DESC
    assert frts_cmd.frtsRegionDesc.__class__ is FWSECLIC_FRTS_REGION_DESC

-  @unittest.skipIf(WIN, "doesn't compile on windows")
-  @unittest.skipIf(OSX, "can't find stdint?")
-  def test_packed_fields(self):
-    ns = self.run_gen("""#include <stdint.h>
-typedef struct die_info
- {
-	 uint16_t die_id;
-	 uint16_t die_offset; /* Points to the corresponding die_header structure */
- } die_info;
-
-typedef struct ip_discovery_header
- {
-	 uint32_t signature;    /* Table Signature */
-	 uint16_t version;      /* Table Version */
-	 uint16_t size;         /* Table Size */
-	 uint32_t id;           /* Table ID */
-	 uint16_t num_dies;     /* Number of Dies */
-	 die_info die_info[16]; /* list die information for up to 16 dies */
-	 union {
-		 uint16_t padding[1];	/* version <= 3 */
-		 struct {		/* version == 4 */
-			 uint8_t base_addr_64_bit : 1; /* ip structures are using 64 bit base address */
-			 uint8_t reserved : 7;
-			 uint8_t reserved2;
-		 };
-	 };
- } ip_discovery_header;
-""")
-
-    ip_discovery_header = ns['ip_discovery_header']
-
-    hdr = b'IPDS\x04\x00|\x1d\x80\x1a\xffd\x01\x00\x00\x00\x8c\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01\x00' # noqa: E501
-    ihdr = ip_discovery_header.from_buffer_copy(hdr)
-
-    assert ctypes.sizeof(ihdr) == 80
-    assert ihdr.signature == 0x53445049
-    assert ihdr.version == 0x0004
-    assert ihdr.num_dies == 1
-    assert ihdr.base_addr_64_bit == 1
-
-  @unittest.skipIf(WIN, "doesn't compile on windows")
  def test_gen_from_header(self):
    namespace = self.run_gen("""
    typedef struct {
@@ -378,7 +399,6 @@ typedef struct ip_discovery_header
    self.assertTrue(hasattr(rect, 'height'))
    self.assertTrue(hasattr(rect, 'color'))

-  @unittest.skipIf(WIN, "doesn't compile on windows")
  def test_struct_ordering(self):
    namespace = self.run_gen("""
    struct A;
@@ -408,77 +428,6 @@ typedef struct ip_discovery_header
    self.assertTrue(hasattr(b, 'c_ptr'))
    self.assertTrue(hasattr(c, 'a_ptr'))

-  @unittest.skipIf(WIN, "doesn't compile on windows")
-  def test_pointer_field_roundtrip(self):
-    # This tests storing a pointer in a record struct field and passing it to C
-    # Mimics how mesa.struct_lp_build_tgsi_params.mask is used
-    from tinygrad.runtime.support.c import POINTER
-    @record
-    class Inner:
-      SIZE = 8
-      value: Annotated[ctypes.c_int, 0]
-      flag: Annotated[ctypes.c_int, 4]
-    @record
-    class Outer:
-      SIZE = 16
-      x: Annotated[ctypes.c_int, 0]
-      inner_ptr: Annotated[POINTER[Inner], 8]
-    init_records()
-
-    src = """
-      struct inner { int value; int flag; };
-      struct outer { int x; struct inner *inner_ptr; };
-      int test(struct inner *p) {
-        return p->value + p->flag;
-      }
-    """
-    dll = self.compile(src)
-    @dll.bind
-    def test(p:POINTER[Inner]) -> ctypes.c_int: ...
-
-    inner = Inner(value=42, flag=10)
-    outer = Outer(x=1, inner_ptr=ctypes.pointer(inner))
-    # Retrieve pointer from struct field and pass to C
-    self.assertEqual(test(outer.inner_ptr), 52)
-
-  @unittest.skipIf(WIN, "doesn't compile on windows")
-  def test_pointer_field_loses_reference(self):
-    # BUG: When a pointer is stored in a record struct field, only the address bytes are saved.
-    # The pointer's _objects dict (which prevents GC of the pointed-to object) is lost.
-    # This causes the pointed-to object to be garbage collected, leading to use-after-free.
-    from tinygrad.runtime.support.c import POINTER
-    @record
-    class MaskContext:
-      SIZE = 16
-      value: Annotated[ctypes.c_int, 0]
-      initialized: Annotated[ctypes.c_int, 4]
-      ptr: Annotated[ctypes.c_void_p, 8]
-    @record
-    class Params:
-      SIZE = 16
-      x: Annotated[ctypes.c_int, 0]
-      mask: Annotated[POINTER[MaskContext], 8]
-    init_records()
-
-    src = """
-      struct mask_ctx { int value; int initialized; void *ptr; };
-      void mask_begin(struct mask_ctx *m, int val) { m->value = val; m->initialized = 1; }
-      int mask_end(struct mask_ctx *m) { return m->value + m->initialized; }
-    """
-    dll = self.compile(src)
-    @dll.bind
-    def mask_begin(m:POINTER[MaskContext], val:ctypes.c_int) -> None: ...
-    @dll.bind
-    def mask_end(m:POINTER[MaskContext]) -> ctypes.c_int: ...
-
-    # When MaskContext() is created inline, it gets garbage collected after the pointer
-    # is stored because only the address bytes are saved, not the _objects reference.
-    params = Params(x=1, mask=ctypes.pointer(MaskContext()))
-    mask_begin(params.mask, 42)
-    result = mask_end(params.mask)
-    self.assertEqual(result, 43)  # 42 + 1
-
-  @unittest.skipIf(WIN, "doesn't compile on windows")
  def test_anonymous_children(self):
    namespace = self.run_gen("""
      struct foo {
@@ -491,7 +440,6 @@ typedef struct ip_discovery_header
    self.assertIn('struct_foo', namespace)
    self.assertIn('struct_foo_bar', namespace)

-  @unittest.skipIf(WIN, "doesn't compile on windows")
  def test_enums(self):
    namespace = self.run_gen("""
      enum Foo { A, B, C };
@@ -511,4 +459,43 @@ typedef struct ip_discovery_header
    assert namespace["enum_Bar"].get(1) == "Y"
    assert namespace["enum_Bar"].get(2) == "Z"

+  @unittest.skipIf(OSX, "can't find stdint?")
+  def test_packed_fields(self):
+    ns = self.run_gen("""#include <stdint.h>
+typedef struct die_info
+ {
+	 uint16_t die_id;
+	 uint16_t die_offset; /* Points to the corresponding die_header structure */
+ } die_info;
+
+typedef struct ip_discovery_header
+ {
+	 uint32_t signature;    /* Table Signature */
+	 uint16_t version;      /* Table Version */
+	 uint16_t size;         /* Table Size */
+	 uint32_t id;           /* Table ID */
+	 uint16_t num_dies;     /* Number of Dies */
+	 die_info die_info[16]; /* list die information for up to 16 dies */
+	 union {
+		 uint16_t padding[1];	/* version <= 3 */
+		 struct {		/* version == 4 */
+			 uint8_t base_addr_64_bit : 1; /* ip structures are using 64 bit base address */
+			 uint8_t reserved : 7;
+			 uint8_t reserved2;
+		 };
+	 };
+ } ip_discovery_header;
+""")
+
+    ip_discovery_header = ns['ip_discovery_header']
+
+    hdr = b'IPDS\x04\x00|\x1d\x80\x1a\xffd\x01\x00\x00\x00\x8c\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01\x00' # noqa: E501
+    ihdr = ip_discovery_header.from_buffer_copy(hdr)
+
+    assert ctypes.sizeof(ihdr) == 80
+    assert ihdr.signature == 0x53445049
+    assert ihdr.version == 0x0004
+    assert ihdr.num_dies == 1
+    assert ihdr.base_addr_64_bit == 1
+
 if __name__ == "__main__": unittest.main()
--- a/Show More
+++ b/Show More