ARM Compute Library flags (#17640)

.ci/azure/linux_arm64.yml

@@ -46,8 +46,7 @@ jobs:
     system.debug: true
     VSTS_HTTP_RETRY: 5
     VSTS_HTTP_TIMEOUT: 200
-    OPENVINO_ARCH: 'aarch64'
+    NUM_PROC: 2
-    NUM_PROC: 1
     BUILD_TYPE: Release
     OPENVINO_REPO_DIR: $(Build.Repository.LocalPath)
     BUILD_OPENVINO: $(WORK_DIR)/build
@@ -116,14 +115,9 @@ jobs:
       python3 -m pip install -r $(OPENVINO_REPO_DIR)/src/bindings/python/requirements.txt
       python3 -m pip install -r $(OPENVINO_REPO_DIR)/src/bindings/python/wheel/requirements-dev.txt
       # install dependencies needed to build CPU plugin for ARM
-      sudo -E apt --assume-yes install scons crossbuild-essential-arm64
+      sudo -E apt --assume-yes install scons gcc-10-aarch64-linux-gnu g++-10-aarch64-linux-gnu
       # generic dependencies
-      sudo -E apt --assume-yes install cmake ccache
+      sudo -E apt --assume-yes install cmake ccache ninja-build unzip
-      # Speed up build
-      sudo -E apt -y --no-install-recommends install unzip
-      wget https://github.com/ninja-build/ninja/releases/download/v1.10.2/ninja-linux.zip
-      unzip ninja-linux.zip
-      sudo cp -v ninja /usr/local/bin/
     displayName: 'Install dependencies'
 
   - script: |
@@ -133,13 +127,15 @@ jobs:
 
   - script: |
       python3 -m pip install conan
+      # install build profile compilers
+      sudo -E apt --assume-yes install gcc g++
       # generate build profile
       conan profile detect
       # generate host profile for linux_arm64
       echo "include(default)" > $(BUILD_OPENVINO)/linux_arm64
       echo "[buildenv]" >> $(BUILD_OPENVINO)/linux_arm64
-      echo "CC=aarch64-linux-gnu-gcc" >> $(BUILD_OPENVINO)/linux_arm64
+      echo "CC=aarch64-linux-gnu-gcc-10" >> $(BUILD_OPENVINO)/linux_arm64
-      echo "CXX=aarch64-linux-gnu-g++" >> $(BUILD_OPENVINO)/linux_arm64
+      echo "CXX=aarch64-linux-gnu-g++-10" >> $(BUILD_OPENVINO)/linux_arm64
       # install OpenVINO dependencies
       export CMAKE_CXX_COMPILER_LAUNCHER=ccache
       export CMAKE_C_COMPILER_LAUNCHER=ccache

src/plugins/intel_cpu/thirdparty/ACLConfig.cmake

@@ -87,6 +87,42 @@ elseif(ENABLE_ARM_COMPUTE_CMAKE)
     # required by oneDNN to attempt to parse ACL version
     set(ENV{ACL_ROOT_DIR} "${ARM_COMPUTE_SOURCE_DIR}")
 elseif(NOT TARGET arm_compute::arm_compute)
+    #
+    # Options
+    #
+
+    set(ARM_COMPUTE_SCONS_JOBS "8" CACHE STRING "Number of parallel threads to build ARM Compute Library")
+
+    set(ARM_COMPUTE_TARGET_GENERIC_ARCHS armv8a
+                                         armv8.2-a
+                                         armv8.6-a armv8.6-a-sve armv8.6-a-sve2 armv8.6-a-sve2-sme2
+                                         armv8r64 # the same as armv8.4-a
+    )
+    if(ARM)
+        set(ARM_COMPUTE_TARGET_ARCH_DEFAULT armv7a)
+        set(ARM_COMPUTE_TARGET_ARCHS armv7a armv7a-hf
+                                     # requires estate=32
+                                     ${ARM_COMPUTE_TARGET_GENERIC_ARCHS})
+    else()
+        if(APPLE)
+            # Apple M1 / M2 is assumed
+            set(ARM_COMPUTE_TARGET_ARCH_DEFAULT arm64-v8.2-a)
+        else()
+            set(ARM_COMPUTE_TARGET_ARCH_DEFAULT arm64-v8a)
+        endif()
+        set(ARM_COMPUTE_TARGET_ARCHS arm64-v8a
+                                     arm64-v8.2-a arm64-v8.2-a-sve arm64-v8.2-a-sve2
+                                     # used with estate=64
+                                     ${ARM_COMPUTE_TARGET_GENERIC_ARCHS})
+    endif()
+
+    set(ARM_COMPUTE_TARGET_ARCH "${ARM_COMPUTE_TARGET_ARCH_DEFAULT}" CACHE STRING "Architecture for ARM ComputeLibrary")
+    set_property(CACHE ARM_COMPUTE_TARGET_ARCH PROPERTY STRINGS ${ARM_COMPUTE_TARGET_ARCHS})
+
+    #
+    # Configure & build
+    #
+
     set(ARM_COMPUTE_SOURCE_DIR "${intel_cpu_thirdparty_SOURCE_DIR}/ComputeLibrary")
     set(ARM_COMPUTE_BINARY_DIR "${intel_cpu_thirdparty_BINARY_DIR}/ComputeLibrary")
 
@@ -128,8 +164,6 @@ elseif(NOT TARGET arm_compute::arm_compute)
         reference_openmp=0
         validation_tests=0
         benchmark_tests=0
-        # TODO: check this for Apple Silicon
-        # multi_isa=1
         # TODO: use CC for ARM compute library to minimize binary size
         # build_config=<file>
         # TODO: use data_type_support to disable useless kernels
@@ -137,6 +171,16 @@ elseif(NOT TARGET arm_compute::arm_compute)
         arch=${ARM_COMPUTE_TARGET_ARCH}
     )
 
+    if(ARM)
+        list(APPEND ARM_COMPUTE_OPTIONS estate=32)
+    else()
+        list(APPEND ARM_COMPUTE_OPTIONS estate=64)
+        if(NOT APPLE AND CMAKE_COMPILER_IS_GNUCXX AND CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 10.2)
+            # arm_sve.h header is not available on gcc older 10.2
+            list(APPEND ARM_COMPUTE_OPTIONS multi_isa=1)
+        endif()
+    endif()
+
     if(NOT MSVC64)
         list(APPEND ARM_COMPUTE_OPTIONS
             build_dir=${ARM_COMPUTE_BINARY_DIR}
@@ -342,8 +386,7 @@ elseif(NOT TARGET arm_compute::arm_compute)
 
     add_library(arm_compute::half INTERFACE IMPORTED GLOBAL)
     set_target_properties(arm_compute::half PROPERTIES
-        INTERFACE_INCLUDE_DIRECTORIES ${ARM_COMPUTE_SOURCE_DIR}/include
+        INTERFACE_INCLUDE_DIRECTORIES ${ARM_COMPUTE_SOURCE_DIR}/include)
-        OSX_ARCHITECTURES arm64)
 
     # Helpers for oneDNN intergation
 

src/plugins/intel_cpu/thirdparty/CMakeLists.txt

@@ -4,9 +4,9 @@
 
 project(intel_cpu_thirdparty)
 
-if((CMAKE_CXX_COMPILER_ID STREQUAL "MSVC") AND (MSVC_VERSION VERSION_GREATER_EQUAL "1910"))
+if(CMAKE_CXX_COMPILER_ID STREQUAL "MSVC" AND MSVC_TOOLSET_VERSION GREATER_EQUAL 141)
-    # 1910 version of Visual Studio 2017
+    # Visual Studio 2017 (v141 toolset)
-    # This flagis needed for enabling SIMD vectorization with command '#pragma omp simd'.
+    # This flag is needed for enabling SIMD vectorization with command '#pragma omp simd'.
     # Compilation with '/openmp:experimental' key allow us to enable vectorizatikon capability in MSVC.
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /openmp:experimental")
 endif()
@@ -38,13 +38,8 @@ function(ov_add_onednn)
     set(DNNL_ENABLE_WORKLOAD "INFERENCE" CACHE STRING "" FORCE)
 
     # Allow to enable oneDNN verbose with CPU_DEBUG_CAPS and rely on oneDNN default configuration otherwise
-    if (ENABLE_CPU_DEBUG_CAPS)
+    if(ENABLE_CPU_DEBUG_CAPS)
-        set(DNNL_VERBOSE "ON" CACHE STRING "" FORCE)
+        set(DNNL_VERBOSE ON CACHE STRING "" FORCE)
-    endif()
-
-    set(SDL_cmake_included ON)  ## to skip internal SDL flags. SDL flags are already set on IE level
-    if (ANDROID OR ((CMAKE_CXX_COMPILER_ID STREQUAL "MSVC" OR OV_COMPILER_IS_CLANG) AND NOT (THREADING STREQUAL "OMP")))
-        set(OpenMP_cmake_included ON) ## to skip "omp simd" inside a code. Lead to some crashes inside NDK LLVM..
     endif()
 
     if(X86_64)
@@ -53,43 +48,23 @@ function(ov_add_onednn)
         set(DNNL_TARGET_ARCH "X86" CACHE STRING "" FORCE)
     elseif(RISCV64)
         set(DNNL_TARGET_ARCH "RV64" CACHE STRING "" FORCE)
-    elseif(AARCH64 OR ARM)
+    elseif(ARM)
-        # TODO: fix warning
-        if(CMAKE_COMPILER_IS_GNUCXX OR OV_COMPILER_IS_CLANG)
-            ie_add_compiler_flags(-Wno-macro-redefined)
-        endif()
-
-        set(ARM_COMPUTE_SCONS_JOBS "8" CACHE STRING "Number of parallel threads to build ARM Compute Library")
-        set(DNNL_USE_ACL ON CACHE BOOL "" FORCE)
-
-        if(ARM)
         set(DNNL_TARGET_ARCH "ARM" CACHE STRING "" FORCE)
-            set(ARM_COMPUTE_TARGET_ARCH_DEFAULT armv7a)
+    elseif(AARCH64)
-            set(ARM_COMPUTE_TARGET_ARCHS armv7a armv7a-hf)
-        else()
         set(DNNL_TARGET_ARCH "AARCH64" CACHE STRING "" FORCE)
-            # move to separate ACL cmake
-            if(APPLE)
-                # Apple M1 / M2 is assumed
-                set(ARM_COMPUTE_TARGET_ARCH_DEFAULT armv8.2-a)
-            else()
-                set(ARM_COMPUTE_TARGET_ARCH_DEFAULT arm64-v8a)
-            endif()
-            set(ARM_COMPUTE_TARGET_ARCHS arm64-v8a
-                                         arm64-v8.2-a arm64-v8.2-a-sve arm64-v8.2-a-sve2
-                                         armv8a
-                                         armv8.2-a armv8.2-a-sve
-                                         armv8.6-a armv8.6-a-sve armv8.6-a-sve2 armv8.6-a-sve2-sme2
-                                         armv8r64 # the same as armv8.4-a
-                                         )
-        endif()
-
-        set(ARM_COMPUTE_TARGET_ARCH "${ARM_COMPUTE_TARGET_ARCH_DEFAULT}" CACHE STRING "Architecture for ARM ComputeLibrary")
-        set_property(CACHE ARM_COMPUTE_TARGET_ARCH PROPERTY STRINGS ${ARM_COMPUTE_TARGET_ARCHS})
     else()
         message(FATAL_ERROR "Unsupported system processor ${CMAKE_SYSTEM_PROCESSOR}")
     endif()
 
+    if(AARCH64 OR ARM)
+        set(DNNL_USE_ACL ON CACHE BOOL "Use ARM Conpute Library kernels in oneDNN" FORCE)
+    endif()
+
+    set(SDL_cmake_included ON)  ## to skip internal SDL flags. SDL flags are already set on IE level
+    if (ANDROID OR ((CMAKE_CXX_COMPILER_ID STREQUAL "MSVC" OR OV_COMPILER_IS_CLANG) AND NOT (THREADING STREQUAL "OMP")))
+        set(OpenMP_cmake_included ON) ## to skip "omp simd" inside a code. Lead to some crashes inside NDK LLVM..
+    endif()
+
     # WA for old TBBConfig.cmake like tbb2019_20180718oss
     # they don't check that imported target is already created
     if(TBB_FOUND)
@@ -101,15 +76,12 @@ function(ov_add_onednn)
         link_libraries(TBB::tbb)
     endif()
 
-    if(CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64")
-        set(DNNL_TARGET_ARCH "X64")
-    elseif(CMAKE_OSX_ARCHITECTURES STREQUAL "arm64")
-        set(DNNL_TARGET_ARCH "AARCH64")
-    endif()
-
     if(CMAKE_COMPILER_IS_GNUCXX OR OV_COMPILER_IS_CLANG)
         ie_add_compiler_flags(-Wno-undef)
         ie_add_compiler_flags(-Wno-missing-declarations)
+        if(ARM OR AARCH64)
+            ie_add_compiler_flags(-Wno-macro-redefined)
+        endif()
         if(CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 11 AND CMAKE_COMPILER_IS_GNUCXX)
             ie_add_compiler_flags(-Wno-array-bounds)
             ie_add_compiler_flags(-Wno-stringop-overflow)
@@ -137,7 +109,9 @@ function(ov_add_onednn)
     endif()
 
     # to find our FindACL.cmake
+    if(DNNL_USE_ACL)
         list(APPEND CMAKE_MODULE_PATH "${intel_cpu_thirdparty_SOURCE_DIR}")
+    endif()
 
     add_subdirectory(onednn EXCLUDE_FROM_ALL)
 
@@ -157,6 +131,4 @@ function(ov_add_onednn)
     endif()
 endfunction()
 
-if(ENABLE_INTEL_CPU)
+ov_add_onednn()
-    ov_add_onednn()
-endif()