[CPU] Support avx isa by default

build.sh

@@ -60,7 +60,7 @@ usage()
   echo "    -l Compile with python dependency, default on"
   echo "    -S Enable enable download cmake compile dependency from gitee , default off"
   echo "    -k Enable make clean, clean up compilation generated cache "
-  echo "    -W Enable x86_64 SSE or AVX instruction set, use [sse|avx|neon|off], default off"
+  echo "    -W Enable x86_64 SSE or AVX instruction set, use [sse|avx|neon|off], default off for lite and avx for CPU"
   echo "    -H Enable hidden"
   echo "    -L Link and specify Tensor-RT library path, default disable Tensor-RT lib linking"
 }
@@ -387,7 +387,7 @@ build_mindspore()
     echo "start build mindspore project."
     mkdir -pv "${BUILD_PATH}/mindspore"
     cd "${BUILD_PATH}/mindspore"
-    CMAKE_ARGS="-DDEBUG_MODE=$DEBUG_MODE -DBUILD_PATH=$BUILD_PATH -DX86_64_SIMD=${X86_64_SIMD}"
+    CMAKE_ARGS="-DDEBUG_MODE=$DEBUG_MODE -DBUILD_PATH=$BUILD_PATH"
     if [[ "X$ENABLE_COVERAGE" = "Xon" ]]; then
       CMAKE_ARGS="${CMAKE_ARGS} -DENABLE_COVERAGE=ON"
     fi

mindspore/ccsrc/CMakeLists.txt

@@ -6,18 +6,13 @@ include_directories(${CMAKE_SOURCE_DIR}/mindspore/core/mindrt/include)
 include_directories(${CMAKE_SOURCE_DIR}/mindspore/core/mindrt/src)
 
 if(ENABLE_CPU)
+    if(${CMAKE_HOST_SYSTEM_PROCESSOR} MATCHES "aarch64")
+        set(PLATFORM_ARM64 "on")
+        set(X86_64_SIMD "off")
+    elseif("${X86_64_SIMD}" STREQUAL "off")
+        set(X86_64_SIMD "avx")
+    endif()
     include_directories(${CMAKE_CURRENT_SOURCE_DIR}/backend/kernel_compiler/cpu)
-    if("${X86_64_SIMD}" STREQUAL "sse")
-        add_compile_definitions(ENABLE_SSE)
-        set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -msse4.1 -msse4.2")
-        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -msse4.1 -msse4.2")
-    endif()
-    if("${X86_64_SIMD}" STREQUAL "avx")
-        add_compile_definitions(ENABLE_SSE)
-        add_compile_definitions(ENABLE_AVX)
-        set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -msse4.1 -msse4.2 -mfma -mavx -mavx2")
-        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -msse4.1 -msse4.2 -mfma -mavx -mavx2")
-    endif()
     add_subdirectory(backend/kernel_compiler/cpu/nnacl)
 endif()
 

mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/CMakeLists.txt

@@ -13,9 +13,6 @@ if(PLATFORM_ARM32 OR PLATFORM_ARM64)
         -ffunction-sections -fdata-sections -ffast-math")
     endif()
 endif()
-if(ENABLE_CPU)
-    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -fPIC")
-endif()
 if("${X86_64_SIMD}" STREQUAL "avx")
     set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -msse4.1 -mavx -mavx2 -mfma")
 endif()
@@ -59,21 +56,27 @@ if(APPLE)
     set_source_files_properties(${ASSEMBLY_SRC} PROPERTIES COMPILE_FLAGS "-x assembler-with-cpp")
 endif()
 
-########################### build nnacl static library ########################
+########################### build nnacl library ########################
 string(REPLACE "-fvisibility=hidden" "-fvisibility=default" CMAKE_C_FLAGS "${CMAKE_C_FLAGS}")
-
-add_library(nnacl_mid OBJECT ${KERNEL_SRC} ${TRAIN_SRC} ${ASSEMBLY_SRC})
-
 if(ENABLE_CPU)
-    add_library(nnacl SHARED $<TARGET_OBJECTS:nnacl_mid>)
+    add_library(nnacl SHARED ${KERNEL_SRC} ${TRAIN_SRC} ${ASSEMBLY_SRC})
+    if(${CMAKE_HOST_SYSTEM_PROCESSOR} MATCHES "aarch64")
+        target_compile_definitions(nnacl PRIVATE ENABLE_ARM ENABLE_ARM64 ENABLE_NEON)
+        target_compile_options(nnacl PRIVATE -ffast-math -flax-vector-conversions)
+    elseif("${X86_64_SIMD}" STREQUAL "sse")
+        target_compile_definitions(nnacl PRIVATE ENABLE_SSE)
+    elseif("${X86_64_SIMD}" STREQUAL "avx")
+        target_compile_definitions(nnacl PRIVATE ENABLE_SSE ENABLE_AVX)
+    endif()
+    target_compile_options(nnacl PRIVATE -fPIC)
+    if(NOT CMAKE_SYSTEM_NAME MATCHES "Windows")
+        target_link_options(nnacl PRIVATE -Wl,-z,relro,-z,now)
+    endif()
 else()
+    add_library(nnacl_mid OBJECT ${KERNEL_SRC} ${TRAIN_SRC} ${ASSEMBLY_SRC})
     add_library(nnacl STATIC $<TARGET_OBJECTS:nnacl_mid>)
-endif()
-
-if(NOT ENABLE_CPU)
     add_dependencies(nnacl_mid fbs_src)
 endif()
-
 ########################### arm fp16 build optimize library ########################
 if(ENABLE_FP16)
     add_subdirectory(${NNACL_DIR}/optimize)

mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/fp32/activation_fp32.c

@@ -86,7 +86,11 @@ int LRelu(const float *src, int length, float *dst, float alpha) {
   for (; i < length - 4; i += 4) {
     MS_FLOAT32X4 src_tmp = MS_LDQ_F32(src + i);
     MS_FLOAT32X4 mul_tmp = MS_MULQ_N_F32(src_tmp, alpha);
+#ifdef ENABLE_ARM
+    MS_UINT32X4 mask = MS_CMPGTQ_F32(src_tmp, MS_MOVQ_F32(0.0f));
+#else
     MS_FLOAT32X4 mask = MS_CMPGTQ_F32(src_tmp, MS_MOVQ_F32(0.0f));
+#endif
     MS_STQ_F32(dst + i, MS_BLENDQ_F32(mul_tmp, src_tmp, mask));
   }
 #endif

mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/intrinsics/ms_simd_instructions.h

@@ -29,6 +29,7 @@
 #ifdef ENABLE_ARM
 #define MS_FLOAT32X4 float32x4_t
 #define MS_INT32X4 int32x4_t
+#define MS_UINT32X4 uint32x4_t
 #define MS_LDQ_F32 vld1q_f32
 #define MS_LDQ_EPI32 vld1q_s32
 #define MS_ADDQ_F32 vaddq_f32