[wip] HSA HAL

.github/workflows/ci-hsa-linux.yml

@@ -0,0 +1,299 @@
+name: CI Linux HSA
+
+on:
+  workflow_call:
+  workflow_dispatch:
+  pull_request:
+  merge_group:
+  push:
+    branches:
+      - main
+
+concurrency:
+  group: ci-build-test-cpp-linux-hsa-${{ github.event.number || github.sha }}
+  cancel-in-progress: true
+
+jobs:
+  build_hsa:
+    name: Build HSA (linux)
+    runs-on: nod-ai-shared-cpubuilder-manylinux-x86_64
+    steps:
+      - name: Configure local git mirrors
+        run: |
+          /gitmirror/scripts/trigger_update_mirrors.sh
+          /gitmirror/scripts/git_config.sh
+
+      - name: "Checking out repository"
+        env:
+          BRANCH_NAME: ${{ github.ref }}
+          REPO_ADDRESS: ${{ github.server_url }}/${{ github.repository }}
+        run: |
+          git init
+          git remote add origin $REPO_ADDRESS
+          git -c protocol.version=2 fetch --depth 1 origin $BRANCH_NAME
+          git reset --hard FETCH_HEAD
+          git -c submodule."third_party/torch-mlir".update=none -c submodule."third_party/stablehlo".update=none -c submodule."src/runtime_src/core/common/aiebu".update=none submodule update --init --recursive --depth 1 --single-branch -j 10
+
+      - name: Install deps
+        run: |
+          dnf install -y almalinux-release-devel
+          yum install -y elfutils-libelf-devel p7zip p7zip-plugins \
+                         sudo ncurses-compat-libs openssh vim-common
+
+      - name: Build and install libnuma
+        working-directory: ${{ github.workspace }}/third_party/ROCR-Runtime
+        run: |
+          curl --silent -L \
+            https://github.com/numactl/numactl/releases/download/v2.0.18/numactl-2.0.18.tar.gz \
+            -o numactl-2.0.18.tar.gz
+          tar -xf numactl-2.0.18.tar.gz
+          pushd numactl-2.0.18
+          ./configure
+          # i have no idea why this is necessary
+          # but without it you get something about "can't cd into dir"
+          sed -i '7563s/`cd "$dir" && pwd`/$dir/g' libtool
+          make install
+          popd
+
+      - name: Hack ROCR
+        working-directory: ${{ github.workspace }}/third_party/ROCR-Runtime
+        run: |
+          sed -i 's/amdgcn-amd-amdhsa/amdgcn-amd-amdhsa -nogpulib/g' runtime/hsa-runtime/core/runtime/blit_shaders/CMakeLists.txt
+          sed -i 's/amdgcn-amd-amdhsa/amdgcn-amd-amdhsa -nogpulib/g' runtime/hsa-runtime/core/runtime/trap_handler/CMakeLists.txt
+          sed -i 's/amdgcn-amd-amdhsa/amdgcn-amd-amdhsa -nogpulib/g' runtime/hsa-runtime/image/blit_src/CMakeLists.txt
+
+      - name: Get compatible Clang
+        working-directory: ${{ github.workspace }}/third_party/ROCR-Runtime
+        run: |
+          pip download mlir==20.0.0.2024090301+amdgpu.b6597f52 -f https://makslevental.github.io/wheels
+          unzip -q mlir-*.whl
+
+      - name: Build ROCR distro
+        working-directory: ${{ github.workspace }}/third_party/ROCR-Runtime
+        run: |
+          rocr_dir="$PWD"
+          build_rocr_dir="$PWD/rocr-build"
+          mkdir -p "$build_rocr_dir"
+          build_rocr_dir="$(cd $build_rocr_dir && pwd)"
+          rocr_install_dir="$PWD/rocr-install"
+
+          cmake -GNinja \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DCMAKE_INSTALL_PREFIX="$rocr_install_dir" \
+            -DClang_DIR=$PWD/mlir/lib/cmake/clang \
+            -DLLVM_DIR=$PWD/mlir/lib/cmake/mlir \
+            -DIMAGE_SUPPORT=OFF \
+            -S "$rocr_dir" -B "$build_rocr_dir"
+
+          cmake --build "$build_rocr_dir" --target install
+          tar -cf rocr-${GITHUB_SHA::8}.tar rocr-install
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v4
+        if: ${{ !cancelled() }}
+        with:
+          name: linux_hsa_x86_64_release_packages
+          path: ${{ github.workspace }}/third_party/ROCR-Runtime/rocr-*.tar
+          if-no-files-found: error
+
+  build_and_ctest:
+    name: Build and Test with HSA (linux, ASSERTIONS)
+    needs: [build_hsa]
+    runs-on: nod-ai-shared-cpubuilder-manylinux-x86_64
+    strategy:
+      fail-fast: true
+    env:
+      CACHE_DIR: ${{ github.workspace }}/.container-cache
+      # either the PR number or `branch-N` where N always increments
+      CACHE_KEY: linux-build-test-cpp-asserts-manylinux-v2-${{ format('{0}-{1}', github.ref_name, github.run_number) }}
+    steps:
+      - name: Set unified TZ
+        uses: szenius/[email protected]
+        with:
+          # this is an arbitrary choice
+          timezoneLinux: "Asia/Singapore"
+          timezoneMacos: "Asia/Singapore"
+          timezoneWindows: "Singapore Standard Time"
+
+      - name: Configure local git mirrors
+        run: |
+          /gitmirror/scripts/trigger_update_mirrors.sh
+          /gitmirror/scripts/git_config.sh
+
+      - name: "Checking out repository"
+        env:
+          BRANCH_NAME: ${{ github.ref }}
+          REPO_ADDRESS: ${{ github.server_url }}/${{ github.repository }}
+        run: |
+          git init
+          git remote add origin $REPO_ADDRESS
+          git -c protocol.version=2 fetch --depth 1 origin $BRANCH_NAME
+          git reset --hard FETCH_HEAD
+          git -c submodule."third_party/torch-mlir".update=none -c submodule."third_party/stablehlo".update=none -c submodule."src/runtime_src/core/common/aiebu".update=none submodule update --init --recursive --depth 1 --single-branch -j 10
+
+      - name: Install deps
+        run: |
+          dnf install -y almalinux-release-devel epel-release
+          yum remove -y openssl-devel zlib-devel || true
+          yum install -y protobuf-devel protobuf-compiler tmate
+
+      - name: Python deps
+        run: |
+          pip install "numpy<2" pyyaml "pybind11[global]==2.10.3" nanobind
+
+      - name: Enable cache
+        uses: actions/cache/restore@v3
+        with:
+          path: ${{ env.CACHE_DIR }}
+          key:  ${{ env.CACHE_KEY }}
+          restore-keys: linux-build-test-cpp-
+
+      - name: Download artifacts
+        uses: actions/download-artifact@v4
+        with:
+          name: linux_hsa_x86_64_release_packages
+
+      - name: Extract artifact
+        run: |
+          tar -xvf rocr-*.tar
+          echo "hsa-runtime64_ROOT=$PWD/rocr-install" >> $GITHUB_ENV
+          echo IREE_EXTERNAL_HAL_DRIVER=hsa >> $GITHUB_ENV
+
+      - name: Build packages
+        run: |
+          export cache_dir="${{ env.CACHE_DIR }}"
+          export CCACHE_COMPILERCHECK="string:$(clang --version)"
+          bash build_tools/ci/build_test_cpp.sh
+
+      - name: Create artifacts
+        if: ${{ !cancelled() }}
+        run: |
+          tar cf iree-dist-linux.tar iree-install
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v4
+        if: ${{ !cancelled() }}
+        with:
+          name: linux_x86_64_release_packages
+          path: iree-dist-linux.tar
+          if-no-files-found: warn
+
+      - name: Save cache
+        uses: actions/cache/save@v3
+        if: ${{ !cancelled() && github.event_name == 'push' && github.ref_name == 'main' }}
+        with:
+          path: ${{ env.CACHE_DIR }}
+          key: ${{ env.CACHE_KEY }}
+
+  test_linux:
+    name: E2E Test linux with HSA
+    needs: build_and_ctest
+    strategy:
+      fail-fast: false
+      matrix:
+        runs-on: [linux-phoenix]
+    runs-on: ${{ matrix.runs-on }}
+    env:
+      XILINXD_LICENSE_FILE: /opt/xilinx/Xilinx.lic
+    steps:
+      - name: "Checking out repository" # for test scripts
+        uses: actions/checkout@8f4b7f84864484a7bf31766abe9204da3cbe65b3 # v3.5.0
+        with:
+          submodules: false # not required for testbench
+
+      - name: Download artifacts
+        uses: actions/download-artifact@v4
+        with:
+          name: linux_x86_64_release_packages
+
+      - name: Extract artifact
+        run: |
+          tar -xvf iree-dist-linux.tar
+          bash build_tools/download_peano.sh
+
+      - name: Create venv and install dependencies
+        run: |
+          python3 -m venv .venv
+          source .venv/bin/activate
+          pip install -r tests/matmul/requirements.txt
+
+      - name: E2E correctness matmul test
+        run: |
+          source .venv/bin/activate
+          # Without this additional line an error like
+          #
+          # [XRT] ERROR: Failed to allocate host memory buffer (mmap(len=10616832, prot=3, flags=8193, offset=4294967296)
+          # failed (err=11): Resource temporarily unavailable), make sure host bank is enabled (see xbutil configure --host-mem)
+          # iree-amd-aie/runtime/src/iree-amd-aie/driver/xrt/direct_allocator.cc:179: RESOURCE_EXHAUSTED; could not allocate
+          # memory for buffer; while invoking C++ function matmul_test.generate_random_matrix; while calling import;
+          #
+          # might be observed when too much memory is allocated. For example this
+          # error was seen when running a bf16->f32 matmul with m=n=k=2304.
+          #
+          # This line was suggested at https://github.com/Xilinx/mlir-air/issues/566
+          #
+          # Note that this is only half of the fix. It is also necessary that
+          # the machine that CI is running on has permission to run this line.
+          #
+          # This permission can be adding by adding the line
+          # ```
+          # %github ALL=(ALL) NOPASSWD: /usr/bin/prlimit *
+          # ```
+          #
+          # to the file /etc/sudoers.d/github, which can be done by running
+          # ```
+          # sudo visudo -f /etc/sudoers.d/github
+          # ```
+          # on the guthub CI machine.
+          sudo prlimit -lunlimited --pid $$
+
+          source /opt/xilinx/xrt/setup.sh
+          bash build_tools/ci/run_matmul_test.sh \
+            test_matmuls \
+            iree-install \
+            $PWD/llvm-aie \
+            /opt/xilinx/xrt \
+            /opt/Xilinx/Vitis/2024.2
+
+
+      - name : Smoke E2E comparison flag test
+        run: |
+          source .venv/bin/activate
+          source /opt/xilinx/xrt/setup.sh
+          python3 build_tools/ci/cpu_comparison/run_test.py \
+            test_aie_vs_cpu \
+            iree-install \
+            $PWD/llvm-aie \
+            --xrt-dir /opt/xilinx/xrt \
+            --test-set='Smoke' \
+            --do-not-run-aie
+
+          # Assert that output.log is empty (because verbose=0)
+          if [ -s output.log ]; then
+            echo "output.log is not empty:"
+            cat output.log
+            exit 1
+          else
+            echo "output.log is empty"
+          fi
+
+      - name : E2E comparison of AIE to llvm-cpu
+        run: |
+          source .venv/bin/activate
+          source /opt/xilinx/xrt/setup.sh
+          python3 build_tools/ci/cpu_comparison/run_test.py \
+            test_aie_vs_cpu \
+            $PWD/iree-install \
+            $PWD/llvm-aie \
+            --xrt-dir /opt/xilinx/xrt \
+            --vitis-dir /opt/Xilinx/Vitis/2024.2 \
+            --reset-npu-between-runs -v
+
+      - name: Printing IR from aie2xclbin
+        run: |
+          source .venv/bin/activate
+          source /opt/xilinx/xrt/setup.sh
+          bash build_tools/ci/print_ir_aie2xclbin/print_ir_aie2xclbin.sh \
+            iree-install \
+            print_ir_aie2xclbin_results \
+            $PWD/llvm-aie

.gitmodules

@@ -29,3 +29,8 @@
   url = https://github.com/iree-org/iree.git
   shallow = true
   ignore = dirty
+[submodule "third_party/ROCR-Runtime"]
+  path = third_party/ROCR-Runtime
+  url = https://github.com/nod-ai/ROCR-Runtime.git
+  shallow = true
+  branch = iree-aie

build_tools/build_test_cpp.sh

@@ -100,7 +100,7 @@ if [[ "$OSTYPE" == "linux-gnu"* ]]; then
     -DCMAKE_CXX_COMPILER="${CXX}" \
     -DLLVM_TARGET_ARCH=X86 \
     -DLLVM_TARGETS_TO_BUILD=X86 \
-    -DIREE_EXTERNAL_HAL_DRIVERS=xrt \
+    -DIREE_EXTERNAL_HAL_DRIVERS=${IREE_EXTERNAL_HAL_DRIVER:-xrt} \
     -S $iree_dir -B $build_dir
 elif [[ "$OSTYPE" == "darwin"* ]]; then
   cmake $CMAKE_ARGS \

build_tools/ci/build_roct_rocr.sh

@@ -0,0 +1,54 @@
+#!/bin/bash
+
+set -eux -o errtrace
+
+this_dir="$(cd $(dirname $0) && pwd)"
+repo_root="$(cd $this_dir/../.. && pwd)"
+
+roct_dir="$(cd $repo_root/third_party/ROCT-Thunk-Interface && pwd)"
+rocr_dir="$(cd $repo_root/third_party/ROCR-Runtime && pwd)"
+
+build_roct_dir="$repo_root/roct-build"
+roct_install_dir="$repo_root/roct-install"
+mkdir -p "$build_roct_dir"
+build_roct_dir="$(cd $build_roct_dir && pwd)"
+
+build_rocr_dir="$repo_root/rocr-build"
+rocr_install_dir="$repo_root/rocr-install"
+mkdir -p "$build_rocr_dir"
+build_rocr_dir="$(cd $build_rocr_dir && pwd)"
+
+cache_dir="${cache_dir:-}"
+
+if [ -z "${cache_dir}" ]; then
+  cache_dir="${repo_root}/.build-cache"
+  mkdir -p "${cache_dir}"
+  cache_dir="$(cd ${cache_dir} && pwd)"
+fi
+echo "Caching to ${cache_dir}"
+mkdir -p "${cache_dir}/ccache"
+
+if [[ "$OSTYPE" == "msys"* ]]; then
+  export CC=clang-cl.exe
+  export CXX=clang-cl.exe
+fi
+export CCACHE_DIR="${cache_dir}/ccache"
+export CCACHE_MAXSIZE="700M"
+export CMAKE_C_COMPILER_LAUNCHER=ccache
+export CMAKE_CXX_COMPILER_LAUNCHER=ccache
+
+cd $roct_dir
+cmake -GNinja \
+      -DCMAKE_BUILD_TYPE=Release \
+      -DCMAKE_INSTALL_PREFIX="$roct_install_dir" \
+      -S "$roct_dir" -B "$build_roct_dir"
+cmake --build "$build_roct_dir" --target install
+
+cd $rocr_dir
+cmake -GNinja \
+      -DCMAKE_BUILD_TYPE=Release \
+      -DCMAKE_INSTALL_PREFIX="$rocr_install_dir" \
+      -DCMAKE_PREFIX_PATH="$roct_install_dir" \
+      -DIMAGE_SUPPORT=OFF \
+      -S "$rocr_dir/src" -B "$build_rocr_dir"
+cmake --build "$build_rocr_dir" --target install

iree_runtime_plugin.cmake

@@ -26,5 +26,16 @@ if(IREE_AMD_AIE_ENABLE_XRT_DRIVER)
   include(iree_aie_bootgen)
 endif()
 
+set(IREE_AMD_AIE_ENABLE_HSA_DRIVER OFF)
+if("hsa" IN_LIST IREE_EXTERNAL_HAL_DRIVERS)
+  message(STATUS "Enabling HSA build because it is an enabled HAL driver")
+  set(IREE_AMD_AIE_ENABLE_HSA_DRIVER ON)
+endif()
+
+if(IREE_AMD_AIE_ENABLE_HSA_DRIVER)
+  find_package(hsa-runtime64 CONFIG REQUIRED
+               NAMES hsa-runtime64 hsa_runtime64)
+endif()
+
 add_subdirectory(${CMAKE_CURRENT_LIST_DIR}/runtime/src AMD-AIE)
 add_subdirectory(${CMAKE_CURRENT_LIST_DIR}/experimental AMD-AIE-experimental)