OneDNN rls-v3.6.2

.github/automation/build_acl.sh

@@ -28,7 +28,7 @@ source ${SCRIPT_DIR}/common_aarch64.sh
 
 ACL_CONFIG=${ACL_CONFIG:-"Release"}
 ACL_ROOT_DIR=${ACL_ROOT_DIR:-"${PWD}/ComputeLibrary"}
-ACL_VERSION=${ACL_VERSION:-v24.09}
+ACL_VERSION=${ACL_VERSION:-v24.11.1}
 ACL_ARCH=${ACL_ARCH:-"armv8.2-a"}
 ACL_REPO="https://github.com/ARM-software/ComputeLibrary.git"
 
@@ -61,13 +61,19 @@ fi
 echo "Compiler version:"
 $CC --version
 
-set -x
-git clone --branch $ACL_VERSION --depth 1 $ACL_REPO $ACL_ROOT_DIR
-
-cd $ACL_ROOT_DIR
-
-scons $MP Werror=0 debug=$ACL_DEBUG neon=1 opencl=0 embed_kernels=0 \
-    os=$ACL_OS arch=$ACL_ARCH build=native multi_isa=$ACL_MULTI_ISA_SUPPORT \
-    fixed_format_kernels=1 cppthreads=0 openmp=$ACL_OPENMP examples=0 \
-    validation_tests=0
-set +x
+if [[ "$ACL_ACTION" == "clone" ]]; then
+    set -x
+    git clone --branch $ACL_VERSION --depth 1 $ACL_REPO $ACL_ROOT_DIR
+    set +x
+elif [[ "$ACL_ACTION" == "build" ]]; then
+    cd $ACL_ROOT_DIR
+    set -x
+    scons $MP Werror=0 debug=$ACL_DEBUG neon=1 opencl=0 embed_kernels=0 \
+        os=$ACL_OS arch=$ACL_ARCH build=native multi_isa=$ACL_MULTI_ISA_SUPPORT \
+        fixed_format_kernels=1 cppthreads=0 openmp=$ACL_OPENMP examples=0 \
+        validation_tests=0
+    set +x
+else
+    echo "Unknown action: $ACL_ACTION"
+    exit 1
+fi

.github/automation/test_aarch64.sh

@@ -26,8 +26,8 @@ SCRIPT_DIR="$(dirname "$(readlink -f "$0")")"
 # Defines MP, CC, CXX and OS.
 source ${SCRIPT_DIR}/common_aarch64.sh
 
-# Skip tests for certain config to preserve resources, while maintaining 
-# coverage. Skip: 
+# Skip tests for certain config to preserve resources, while maintaining
+# coverage. Skip:
 # (SEQ,CLANG)
 # (OMP,CLANG,DEBUG)
 SKIP_TESTS=0
@@ -36,7 +36,7 @@ if [[ "$OS" == "Linux" ]]; then
         if [[ "$BUILD_TOOLSET" == "clang" ]]; then
             SKIP_TESTS=1
         fi
-    elif [[ "$ONEDNN_THREADING" == "OMP" ]]; then 
+    elif [[ "$ONEDNN_THREADING" == "OMP" ]]; then
         if [[ "$BUILD_TOOLSET" == "clang" ]]; then
             if [[ "$CMAKE_BUILD_TYPE" == "Debug" ]]; then
                 SKIP_TESTS=1
@@ -62,20 +62,38 @@ if [[ "$OS" == "Linux" ]]; then
         SKIPPED_TEST_FAILURES+="|test_benchdnn_modeC_conv_smoke_cpu"
         SKIPPED_TEST_FAILURES+="|test_benchdnn_modeC_deconv_smoke_cpu"
         SKIPPED_TEST_FAILURES+="|test_benchdnn_modeC_matmul_smoke_cpu"
+        SKIPPED_TEST_FAILURES+="|cpu-graph-gqa-cpp"
+        SKIPPED_TEST_FAILURES+="|cpu-graph-mqa-cpp"
+        SKIPPED_TEST_FAILURES+="|cpu-graph-sdpa-cpp"
+        SKIPPED_TEST_FAILURES+="|cpu-graph-sdpa-stacked-qkv-cpp"
+        SKIPPED_TEST_FAILURES+="|test_graph_unit_dnnl_large_partition_usm_cpu"
+        SKIPPED_TEST_FAILURES+="|test_graph_unit_dnnl_sdp_decomp_usm_cpu"
+        SKIPPED_TEST_FAILURES+="|test_graph_unit_dnnl_mqa_decomp_usm_cpu"
     elif [[ "$CMAKE_BUILD_TYPE" == "Release" ]]; then
         SKIPPED_TEST_FAILURES="cpu-primitives-deconvolution-cpp"
         SKIPPED_TEST_FAILURES+="|test_benchdnn_modeC_lnorm_smoke_cpu"
+        SKIPPED_TEST_FAILURES+="|cpu-graph-gqa-cpp"
+        SKIPPED_TEST_FAILURES+="|cpu-graph-mqa-cpp"
+        SKIPPED_TEST_FAILURES+="|cpu-graph-sdpa-cpp"
+        SKIPPED_TEST_FAILURES+="|cpu-graph-sdpa-stacked-qkv-cpp"
+        SKIPPED_TEST_FAILURES+="|test_graph_unit_dnnl_large_partition_usm_cpu"
+        SKIPPED_TEST_FAILURES+="|test_graph_unit_dnnl_sdp_decomp_usm_cpu"
+        SKIPPED_TEST_FAILURES+="|test_graph_unit_dnnl_mqa_decomp_usm_cpu"
     fi
 elif [[ "$OS" == "Darwin" ]]; then
     if [[ "$CMAKE_BUILD_TYPE" == "Debug" ]]; then
         SKIPPED_TEST_FAILURES="cpu-primitives-deconvolution-cpp"
         SKIPPED_TEST_FAILURES+="|test_benchdnn_modeC_lnorm_smoke_cpu"
         SKIPPED_TEST_FAILURES+="|test_benchdnn_modeC_brgemm_smoke_cpu"
         SKIPPED_TEST_FAILURES+="|test_benchdnn_modeC_brgemm_ci_cpu"
+        SKIPPED_TEST_FAILURES+="|test_graph_unit_dnnl_sdp_decomp_usm_cpu"
+        SKIPPED_TEST_FAILURES+="|test_graph_unit_dnnl_mqa_decomp_usm_cpu"
     elif [[ "$CMAKE_BUILD_TYPE" == "Release" ]]; then
         SKIPPED_TEST_FAILURES="cpu-primitives-deconvolution-cpp"
         SKIPPED_TEST_FAILURES+="|test_benchdnn_modeC_lnorm_smoke_cpu"
         SKIPPED_TEST_FAILURES+="|test_benchdnn_modeC_lnorm_ci_cpu"
+        SKIPPED_TEST_FAILURES+="|test_graph_unit_dnnl_sdp_decomp_usm_cpu"
+        SKIPPED_TEST_FAILURES+="|test_graph_unit_dnnl_mqa_decomp_usm_cpu"
     fi
 fi
 

.github/workflows/ci-aarch64.yml

@@ -27,24 +27,26 @@ on:
 #* Stop stale workflows when pull requests are updated: https://stackoverflow.com/a/70972844
 #* Does not apply to the main branch.
 concurrency:
-  group: ${{ github.ref }}
+  group: ${{ github.workflow }}-${{ github.ref }}
   cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
 
 # Declare default permissions as read only.
 permissions: read-all
 
 jobs:
   macos:
-    name: macOS
     runs-on: macos-14
     strategy:
       matrix:
-        toolset: [clang, gcc]
-        config: [Debug, Release]
+        config: [
+          { toolset: clang, build: Release },
+          { toolset: gcc, build: Release }
+        ]
 
+    name: MacOS, ${{ matrix.config.toolset }}, ${{ matrix.config.build }}
     steps:
       - name: Checkout oneDNN
-        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332 # v4.1.7
+        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
         with:
           path: oneDNN
 
@@ -53,31 +55,65 @@ jobs:
         with:
           packages: scons
 
+      - name: Clone ACL
+        run: ${{ github.workspace }}/oneDNN/.github/automation/build_acl.sh
+        env:
+          ACL_ACTION: clone
+          ACL_CONFIG: ${{ matrix.config.build }}
+          ACL_ROOT_DIR: ${{ github.workspace }}/ComputeLibrary
+          BUILD_TOOLSET: ${{ matrix.config.toolset }}
+          GCC_VERSION: 14
+
+      - name: Get ACL commit hash for cache key
+        id: get_acl_commit_hash
+        run: (cd ${{ github.workspace }}/ComputeLibrary && echo "ACLCommitHash=$(git rev-parse --short HEAD)") >> $GITHUB_OUTPUT
+
+      - name: Get system name
+        id: get_system_name
+        run: (echo "SystemName=$(uname)") >> $GITHUB_OUTPUT
+
+      - name: Restore cached ACL
+        id: cache-acl-restore
+        uses: actions/cache/restore@v4
+        with:
+          key: ${{ steps.get_system_name.outputs.SystemName }}-acl-${{ matrix.config.toolset }}-${{ matrix.config.build }}-${{ steps.get_acl_commit_hash.outputs.ACLCommitHash }}
+          path: ${{ github.workspace }}/ComputeLibrary/build
+
       - name: Build ACL
+        if:  ${{ steps.cache-acl-restore.outputs.cache-hit != 'true' }}
         run: ${{ github.workspace }}/oneDNN/.github/automation/build_acl.sh
         env:
+          ACL_ACTION: build
           ACL_ROOT_DIR: ${{ github.workspace }}/ComputeLibrary
-          BUILD_TOOLSET: ${{ matrix.toolset }}
-          ACL_CONFIG: ${{ matrix.config }}
+          BUILD_TOOLSET: ${{ matrix.config.toolset }}
+          ACL_CONFIG: ${{ matrix.config.build }}
           GCC_VERSION: 14
 
+      - name: Save ACL in cache
+        if:  ${{ steps.cache-acl-restore.outputs.cache-hit != 'true' }}
+        id: cache-acl_build-save
+        uses: actions/cache/save@v4
+        with:
+          key: ${{ steps.cache-acl-restore.outputs.cache-primary-key }}
+          path: ${{ github.workspace }}/ComputeLibrary/build
+
       - name: Build oneDNN
         run: ${{ github.workspace }}/oneDNN/.github/automation/build_aarch64.sh
         working-directory: ${{ github.workspace }}/oneDNN
         env:
           ACL_ROOT_DIR: ${{ github.workspace }}/ComputeLibrary
-          BUILD_TOOLSET: ${{ matrix.toolset }}
-          CMAKE_BUILD_TYPE: ${{ matrix.config }}
+          BUILD_TOOLSET: ${{ matrix.config.toolset }}
+          CMAKE_BUILD_TYPE: ${{ matrix.config.build }}
           GCC_VERSION: 14
-
-      - if: matrix.toolset == 'clang'
+    
+      - if: matrix.config.toolset == 'clang'
         name: Run oneDNN smoke tests
         run: ${{ github.workspace }}/oneDNN/.github/automation/test_aarch64.sh
         working-directory: ${{ github.workspace }}/oneDNN/build
         env:
-          CMAKE_BUILD_TYPE: ${{ matrix.config }}
+          CMAKE_BUILD_TYPE: ${{ matrix.config.build }}
           DYLD_LIBRARY_PATH: ${{ github.workspace }}/ComputeLibrary/build
-
+  
   # We only run the linux aarch64 runners if macos smoke tests pass.
   linux:
     needs: macos
@@ -87,15 +123,15 @@ jobs:
         toolset: [clang, gcc]
         config: [Debug, Release]
         host: [
-          { name: c6g, label: ah-ubuntu_22_04-c6g_2x-50 }, 
+          { name: c6g, label: ah-ubuntu_22_04-c6g_2x-50 },
           { name: c7g, label: ah-ubuntu_22_04-c7g_2x-50 }
         ]
 
     name: ${{ matrix.host.name }}, ${{ matrix.toolset }}, ${{ matrix.threading }}, ${{ matrix.config }}
     runs-on: ${{ matrix.host.label }}
     steps:
       - name: Checkout oneDNN
-        uses: actions/checkout@d632683dd7b4114ad314bca15554477dd762a938 # v4.2.0
+        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
         with:
           path: oneDNN
 
@@ -122,15 +158,49 @@ jobs:
         with:
           version: "17"
 
-      - name: Build ACL
+      - name: Clone ACL
         run: ${{ github.workspace }}/oneDNN/.github/automation/build_acl.sh
         env:
+          ACL_ACTION: clone
           ACL_CONFIG: ${{ matrix.config }}
           ACL_ROOT_DIR: ${{ github.workspace }}/ComputeLibrary
           BUILD_TOOLSET: ${{ matrix.toolset }}
           GCC_VERSION: 13
+
+      - name: Get ACL commit hash for cache key
+        id: get_acl_commit_hash
+        run: (cd ${{ github.workspace }}/ComputeLibrary && echo "ACLCommitHash=$(git rev-parse --short HEAD)") >> $GITHUB_OUTPUT
+
+      - name: Get system name
+        id: get_system_name
+        run: (echo "SystemName=$(uname)") >> $GITHUB_OUTPUT
+
+      - name: Restore cached ACL
+        id: cache-acl-restore
+        uses: actions/cache/restore@v4
+        with:
+          key: ${{ steps.get_system_name.outputs.SystemName }}-acl-${{ matrix.toolset }}-${{ matrix.config }}-${{ steps.get_acl_commit_hash.outputs.ACLCommitHash }}
+          path: ${{ github.workspace }}/ComputeLibrary/build
+
+      - name: Build ACL
+        if:  ${{ steps.cache-acl-restore.outputs.cache-hit != 'true' }}
+        run: ${{ github.workspace }}/oneDNN/.github/automation/build_acl.sh
+        env:
+          ACL_ACTION: build
+          ACL_ROOT_DIR: ${{ github.workspace }}/ComputeLibrary
+          BUILD_TOOLSET: ${{ matrix.toolset }}
+          ACL_CONFIG: ${{ matrix.config }}
+          GCC_VERSION: 13
           ACL_THREADING: ${{ matrix.threading }}
 
+      - name: Save ACL in cache
+        if:  ${{ steps.cache-acl-restore.outputs.cache-hit != 'true' }}
+        id: cache-acl_build-save
+        uses: actions/cache/save@v4
+        with:
+          key: ${{ steps.cache-acl-restore.outputs.cache-primary-key }}
+          path: ${{ github.workspace }}/ComputeLibrary/build
+
       - name: Build oneDNN
         run: ${{ github.workspace }}/oneDNN/.github/automation/build_aarch64.sh
         working-directory: ${{ github.workspace }}/oneDNN
@@ -148,4 +218,4 @@ jobs:
           BUILD_TOOLSET: ${{ matrix.toolset }}
           CMAKE_BUILD_TYPE: ${{ matrix.config }}
           DYLD_LIBRARY_PATH: ${{ github.workspace }}/ComputeLibrary/build
-          ONEDNN_THREADING: ${{ matrix.threading }}     
+          ONEDNN_THREADING: ${{ matrix.threading }}

.github/workflows/labeler.yml

@@ -28,7 +28,7 @@ jobs:
       pull-requests: write
 
     steps:
-      - uses: actions/[email protected]
+      - uses: actions/labeler@8558fd74291d67161a8a78ce36a881fa63b766a9 # v5.0.0
         with:
           sync-labels: true
           configuration-path: '.github/labels.yml'

.github/workflows/openssf-scorecard.yml

@@ -41,7 +41,7 @@ jobs:
 
     steps:
       - name: "Checkout code"
-        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332 # v4.1.7
+        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
         with:
           persist-credentials: false
 

.gitignore

@@ -15,7 +15,7 @@
 # limitations under the License.
 #===============================================================================
 
-build
+build*
 external
 .vs
 .vscode

README.md

@@ -173,7 +173,7 @@ On a CPU based on Arm AArch64 architecture, oneDNN CPU engine can be built with
 machine learning applications and provides AArch64 optimized implementations
 of core functions. This functionality currently requires that ACL is downloaded
 and built separately. See [Build from Source] section of the Developer Guide for
-details. oneDNN only supports Compute Library versions 24.08.1 or later.
+details. oneDNN only supports Compute Library versions 24.11.1 or later.
 
 [Arm Compute Library (ACL)]: https://github.com/arm-software/ComputeLibrary
 

cmake/ACL.cmake

@@ -31,7 +31,7 @@ endif()
 
 find_package(ACL REQUIRED)
 
-set(ACL_MINIMUM_VERSION "24.08.1")
+set(ACL_MINIMUM_VERSION "24.11.1")
 
 if(ACL_FOUND)
     file(GLOB_RECURSE ACL_VERSION_FILE ${ACL_INCLUDE_DIR}/*/arm_compute_version.embed)

cmake/platform.cmake

@@ -78,11 +78,12 @@ macro(platform_gnu_nowarn_ccxx_flags var gnu_version)
     append(${var} "-Wno-strict-overflow")
     # suppress false positive warnings about uninitialized variables
     append(${var} "-Wno-maybe-uninitialized")
-    # suppress false positive warnings with 10.x: GCC Bugzilla – Bug 96963
+    # suppress false positive warnings with 9.x+: GCC Bugzilla – Bug 96963
     # assume 0.0 is unknown version - always suppress the warning
     if(${gnu_version} VERSION_EQUAL 0.0 OR
-        (${gnu_version} VERSION_GREATER 10.0 AND ${gnu_version} VERSION_LESS 11.0))
+        ${gnu_version} VERSION_GREATER 9.0)
         append(${var} "-Wno-stringop-overflow")
+        append(${var} "-Wno-array-bounds")
     endif()
 endmacro()
 

src/common/compiler_workarounds.hpp

@@ -63,14 +63,6 @@
 #define GCC_WA_NO_TREE_DOMINATOR_OPTS 0
 #endif
 
-// Workaround 04: GCC
-//
-// GCC 10 & 11 && 12 (at least versiona 10.1, 10.3 & 11.1, 12.2) report false positives
-// in xbyak when -Warray-bounds build setting is on
-#if (defined NEED_GCC_WA_CHECK) && (__GNUC__ >= 10)
-#pragma GCC diagnostic ignored "-Warray-bounds"
-#endif
-
 // Workaround 05: GCC
 //
 // NOTE: inside lambda, type cast variables captured by reference using

src/common/utils.hpp

@@ -209,11 +209,30 @@ constexpr bool any_null(Args... ptrs) {
     return one_of(nullptr, ptrs...);
 }
 
+// For some unknown reason, GCC 11.x and beyond can't compile specific places
+// of the library that involve this routine. It's connected to the fact that
+// this function is inline and defined in a header.
+#if defined(__GNUC__) && __GNUC__ > 8 && !defined(__clang__)
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wrestrict"
+// /usr/include/bits/string_fortified.h:29:33: warning: ‘void* __builtin_memcpy(
+//     void*, const void*, long unsigned int)’ accessing 18446744056529682432 or
+//     more bytes at offsets 320 and 0 overlaps 9223372002495037441 bytes at
+//     offset -9223372019674906625 [-Wrestrict]
+#pragma GCC diagnostic ignored "-Wstringop-overflow"
+// warning: ‘void* __builtin_memcpy(void*, const void*, long unsigned int)’
+//     specified bound between 18446744056529682432 and 18446744073709551608
+//     exceeds maximum object size 9223372036854775807 [-Wstringop-overflow=]
+#endif
 template <typename T>
 inline void array_copy(T *dst, const T *src, size_t size) {
     for (size_t i = 0; i < size; ++i)
         dst[i] = src[i];
 }
+#if defined(__GNUC__) && __GNUC__ > 8 && !defined(__clang__)
+#pragma GCC diagnostic pop
+#endif
+
 template <typename T>
 inline bool array_cmp(const T *a1, const T *a2, size_t size) {
     for (size_t i = 0; i < size; ++i)