[metal][macos]support metal on Mac OS(1/2) (PaddlePaddle#7347)

* support metal on arm Mac OS

* add load metallib error;
add metallib compile on arm macos;
build_macos.sh support metal

* Update CMakeLists.txt

* Update build_macos.sh

* remove new hardware support on MacOS;
add with_testing

lite/api/CMakeLists.txt

@@ -143,13 +143,15 @@ else()
         # 1. enable -flto compiling flag if toochain==gcc
         # TODO (hong19860320): Disable lto temporarily since it causes fail to catch the exceptions in android when toolchain is gcc.
         if (NOT (ARM_TARGET_LANG STREQUAL "clang"))
-            if (ARM_TARGET_OS STREQUAL "android" AND LITE_WITH_EXCEPTION)
+            if ((ARM_TARGET_OS STREQUAL "android" OR ARM_TARGET_OS STREQUAL "armmacos") AND LITE_WITH_EXCEPTION)
                 set(TARGET_COMIPILE_FLAGS "")
             else()
                 set(TARGET_COMIPILE_FLAGS "-fdata-sections -flto")
             endif()
         else()
-            set(TARGET_COMIPILE_FLAGS "-fdata-sections")
+            if (NOT LITE_WITH_METAL)
+                set(TARGET_COMIPILE_FLAGS "-fdata-sections")
+            endif()
         endif()
         #   1.1 enable -flto on PADDLELITE_OBJS
         set_target_properties(PADDLELITE_OBJS PROPERTIES COMPILE_FLAGS "${TARGET_COMIPILE_FLAGS}")
@@ -158,6 +160,9 @@ else()
         add_library(paddle_light_api_shared SHARED $<TARGET_OBJECTS:PADDLELITE_OBJS>)
         set_target_properties(paddle_light_api_shared PROPERTIES COMPILE_FLAGS "${TARGET_COMIPILE_FLAGS}")
         #   2.1 link `paddle_light_api_shared` to third-party libs
+        if (LITE_WITH_METAL)
+            target_link_libraries(paddle_light_api_shared ${METAL_LIBRARY} ${MPS_LIBRARY} ${FOUNDATION_LIBRARY})
+        endif()
         if (LITE_WITH_NPU)
             # Need to add HIAI runtime libs (libhiai.so) dependency
             target_link_libraries(paddle_light_api_shared ${npu_builder_libs} ${npu_runtime_libs})

lite/backends/metal/CMakeLists.txt

@@ -8,13 +8,14 @@ FILE(GLOB LITE_METAL_KERNELS_SRC ${CMAKE_SOURCE_DIR}/lite/backends/metal/metal_k
         ${CMAKE_SOURCE_DIR}/lite/backends/metal/metal_kernel/texture/*.metal
         )
 
-IF (DEFINED SDK_VERSION)
+IF (SDK_NAME STREQUAL "iphoneos")
     #Defined by iOS toolchain
     SET(SDK_NAME "iphoneos")
     SET(METAL_STD_OPTION "-std=ios-metal1.1")
     SET(TARGET_OPTION "-mios-version-min=${DEPLOYMENT_TARGET}")
 ELSE ()
     SET(SDK_NAME "macosx")
+    SET(METAL_STD_OPTION "-std=macos-metal1.1")
     SET(TARGET_OPTION "")
 ENDIF ()
 
@@ -41,6 +42,6 @@ lite_cc_library(metal_target_wrapper SRCS
         target_wrapper.mm
         )
 
-target_link_libraries(metal_target_wrapper ${METAL_LIBRARY} ${FOUNDATION_LIBRARY})
+target_link_libraries(metal_target_wrapper ${METAL_LIBRARY} ${MPS_LIBRARY} ${GRAPHIC} ${FOUNDATION_LIBRARY})
 
 add_dependencies(metal_target_wrapper LiteMetalLIB)

lite/backends/metal/metal_context_imp.mm

@@ -71,9 +71,10 @@ - (void)dealloc {
 
 - (void)setMetalPath:(std::string)path {
     NSString* pathStr = cString2NSString(path);
+    NSError* libraryErr = nil;
     if (pathStr) {
         self.libPath = pathStr;
-        self.library = [self.device newLibraryWithFile:pathStr error:NULL];
+        self.library = [self.device newLibraryWithFile:pathStr error:&libraryErr];
     }
     if (nil == _library) {
         LOG(INFO) << "Can't load metallib: " << [pathStr cStringUsingEncoding:NSUTF8StringEncoding];

lite/tools/build_macos.sh

@@ -11,8 +11,6 @@ ARCH=armv8
 
 CMAKE_EXTRA_OPTIONS=""
 BUILD_EXTRA=OFF
-BUILD_TRAIN=OFF
-BUILD_JAVA=ON
 BUILD_PYTHON=OFF
 BUILD_DIR=$(pwd)
 OPTMODEL_DIR=""
@@ -24,26 +22,13 @@ WITH_OPENCL=OFF
 WITH_STATIC_MKL=OFF
 WITH_AVX=ON
 WITH_EXCEPTION=OFF
+WITH_LIGHT_WEIGHT_FRAMEWORK=OFF
 WITH_PROFILE=OFF
 WITH_PRECISION_PROFILE=OFF
 WITH_BENCHMARK=OFF
 WITH_LTO=OFF
 BUILD_ARM82_FP16=OFF
 BUILD_ARM82_INT8_SDOT=OFF
-BUILD_NPU=OFF
-NPU_DDK_ROOT="$(pwd)/ai_ddk_lib/" # Download HiAI DDK from https://developer.huawei.com/consumer/cn/hiai/
-BUILD_XPU=OFF
-BUILD_XTCL=OFF
-XPU_SDK_ROOT=""
-XPU_SDK_URL=""
-XPU_SDK_ENV=""
-BUILD_APU=OFF
-APU_DDK_ROOT="$(pwd)/apu_sdk_lib/"
-BUILD_RKNPU=OFF
-RKNPU_DDK_ROOT="$(pwd)/rknpu/"
-WITH_HUAWEI_ASCEND_NPU=OFF # Huawei Ascend Builder/Runtime Libs on X86 host
-# default installation path, ensure acllib/atc/opp directories are all in this root dir
-HUAWEI_ASCEND_NPU_DDK_ROOT="/usr/local/Ascend/ascend-toolkit/latest/x86_64-linux_gcc4.8.5"
 PYTHON_EXECUTABLE_OPTION=""
 workspace=$PWD/$(dirname $0)/../../
 OPTMODEL_DIR=""
@@ -128,6 +113,21 @@ function make_armosx {
     fi
 
     build_dir=$workspace/build.macos.${os}.${arch}
+    if [ ${WITH_METAL} == "ON" ]; then
+      BUILD_EXTRA=ON
+      build_dir=${build_dir}.metal
+    fi
+
+    if [ ${WITH_OPENCL} == "ON" ]; then
+        build_dir=${build_dir}.opencl
+        prepare_opencl_source_code $workspace
+    fi
+
+    if [ ${WITH_TESTING} == "ON" ]; then
+      BUILD_EXTRA=ON
+      LITE_ON_TINY_PUBLISH=OFF
+    fi
+
     if [ -d $build_dir ]
     then
         rm -rf $build_dir
@@ -145,11 +145,17 @@ function make_armosx {
     touch ./${GEN_CODE_PATH_PREFIX}/__generated_code__.cc
     cmake $workspace \
             -DWITH_LITE=ON \
+            -DWITH_TESTING=${WITH_TESTING} \
             -DLITE_WITH_ARM=ON \
             -DLITE_ON_TINY_PUBLISH=ON \
+            -DLITE_WITH_METAL=${WITH_METAL} \
+            -DLITE_WITH_OPENCL=${WITH_OPENCL} \
+            -DLITE_ON_TINY_PUBLISH=${LITE_ON_TINY_PUBLISH} \
+            -DLITE_WITH_PROFILE=${WITH_PROFILE} \
+            -DLITE_WITH_LIGHT_WEIGHT_FRAMEWORK=ON \
+            -DLITE_WITH_PRECISION_PROFILE=${WITH_PRECISION_PROFILE} \
             -DLITE_WITH_OPENMP=OFF \
             -DWITH_ARM_DOTPROD=OFF \
-            -DLITE_WITH_LIGHT_WEIGHT_FRAMEWORK=ON \
             -DLITE_WITH_X86=OFF \
             -DLITE_WITH_LOG=$WITH_LOG \
             -DLITE_WITH_EXCEPTION=$WITH_EXCEPTION \
@@ -162,6 +168,13 @@ function make_armosx {
             -DARM_TARGET_OS=armmacos
 
     make publish_inference -j$NUM_PROC
+    if [ "${WITH_BENCHMARK}" == "ON" ]; then
+        make benchmark_bin -j$NUM_PROC
+    elif [ "${WITH_TESTING}" == "ON" ]; then
+        make lite_compile_deps -j$NUM_PROC
+    else
+        make publish_inference -j$NUM_PROC
+    fi
     cd -
 }
 
@@ -175,17 +188,22 @@ function make_x86 {
     set_benchmark_options
   fi
 
-  if [ ${WITH_HUAWEI_ASCEND_NPU} == "ON" ]; then
-    export CXX=g++ # Huawei Ascend NPU need g++
-    build_directory=$BUILD_DIR/build.lite.huawei_ascend_npu
-  fi
-
   if [ ${WITH_OPENCL} == "ON" ]; then
     BUILD_EXTRA=ON
     build_directory=$BUILD_DIR/build.lite.x86.opencl
     prepare_opencl_source_code $root_dir $build_directory
   fi
 
+  if [ ${WITH_METAL} == "ON" ]; then
+    BUILD_EXTRA=ON
+    build_directory=${build_directory}.metal
+  fi
+
+  if [ ${WITH_TESTING} == "ON" ]; then
+    BUILD_EXTRA=ON
+    LITE_ON_TINY_PUBLISH=OFF
+  fi
+
   if [ ${BUILD_PYTHON} == "ON" ]; then
     BUILD_EXTRA=ON
   fi
@@ -205,12 +223,16 @@ function make_x86 {
 
   cmake $root_dir  -DWITH_MKL=${WITH_MKL}  \
             -DWITH_STATIC_MKL=${WITH_STATIC_MKL}  \
-            -DWITH_TESTING=OFF \
+            -DWITH_TESTING=${WITH_TESTING} \
             -DWITH_AVX=${WITH_AVX} \
             -DWITH_MKLDNN=OFF    \
             -DLITE_WITH_X86=ON  \
             -DWITH_LITE=ON \
             -DLITE_WITH_LIGHT_WEIGHT_FRAMEWORK=OFF \
+            -DLITE_WITH_LIGHT_WEIGHT_FRAMEWORK=${WITH_LIGHT_WEIGHT_FRAMEWORK} \
+            -DLITE_ON_TINY_PUBLISH=${LITE_ON_TINY_PUBLISH} \
+            -DLITE_WITH_PROFILE=${WITH_PROFILE} \
+            -DLITE_WITH_PRECISION_PROFILE=${WITH_PRECISION_PROFILE} \
             -DLITE_WITH_ARM=OFF \
             -DLITE_WITH_OPENCL=${WITH_OPENCL} \
             -DWITH_GPU=OFF \
@@ -220,22 +242,15 @@ function make_x86 {
             -DLITE_OPTMODEL_DIR=${OPTMODEL_DIR} \
             -DLITE_WITH_LOG=${WITH_LOG} \
             -DLITE_WITH_EXCEPTION=$WITH_EXCEPTION \
-            -DLITE_WITH_PROFILE=${WITH_PROFILE} \
-            -DLITE_WITH_PRECISION_PROFILE=${WITH_PRECISION_PROFILE} \
             -DLITE_WITH_LTO=${WITH_LTO} \
-            -DLITE_WITH_XPU=$BUILD_XPU \
-            -DLITE_WITH_XTCL=$BUILD_XTCL \
-            -DXPU_SDK_ROOT=$XPU_SDK_ROOT \
-            -DXPU_SDK_URL=$XPU_SDK_URL \
-            -DXPU_SDK_ENV=$XPU_SDK_ENV \
-            -DLITE_WITH_HUAWEI_ASCEND_NPU=$WITH_HUAWEI_ASCEND_NPU \
-            -DHUAWEI_ASCEND_NPU_DDK_ROOT=$HUAWEI_ASCEND_NPU_DDK_ROOT \
             -DCMAKE_BUILD_TYPE=Release \
             -DPY_VERSION=$PY_VERSION \
             $PYTHON_EXECUTABLE_OPTION
 
   if [ "${WITH_BENCHMARK}" == "ON" ]; then
     make benchmark_bin -j$NUM_PROC
+  elif [ "${WITH_TESTING}" == "ON" ]; then
+    make lite_compile_deps -j$NUM_PROC
   else
     make publish_inference -j$NUM_PROC
   fi
@@ -258,6 +273,9 @@ function print_usage {
     echo -e "|     --with_exception: (OFF|ON); controls whether to throw the exception when error occurs, default is OFF                            |"
     echo -e "|     --build_extra: (OFF|ON); controls whether to publish extra operators and kernels for (sequence-related model such as OCR or NLP)  |"
     echo -e "|     --with_benchmark: (OFF|ON); controls whether to compile benchmark binary, default is OFF, only support macos x86                 |"
+    echo -e "|     --with_extra: (OFF|ON); controls whether to publish extra operators and kernels for (sequence-related model such as OCR or NLP) |"
+    echo -e "|     --with_benchmark: (OFF|ON); controls whether to compile benchmark binary, default is OFF                                         |"
+    echo -e "|     --with_testing: (OFF|ON); controls whether to compile unit test, default is OFF                                                  |"
     echo -e "|                                                                                                                                      |"
     echo -e "|  arguments of benchmark binary compiling for macos x86:                                                                              |"
     echo -e "|     ./lite/tools/build_macos.sh --with_benchmark=ON x86                                                                              |"
@@ -280,7 +298,15 @@ function main {
     # Parse command line.
     for i in "$@"; do
         case $i in
-            --build_extra=*)
+            --with_metal=*)
+                WITH_METAL="${i#*=}"
+                shift
+                ;;
+            --with_opencl=*)
+                WITH_OPENCL="${i#*=}"
+                shift
+                ;;
+            --with_extra=*)
                 BUILD_EXTRA="${i#*=}"
                 shift
                 ;;
@@ -375,12 +401,8 @@ function main {
                 fi
                 shift
                 ;;
-            --xpu_sdk_url=*)
-                XPU_SDK_URL="${i#*=}"
-                shift
-                ;;
-            --xpu_sdk_env=*)
-                XPU_SDK_ENV="${i#*=}"
+            --tiny_publish=*)
+                LITE_ON_TINY_PUBLISH="${i#*=}"
                 shift
                 ;;
             --python_executable=*)
@@ -391,30 +413,6 @@ function main {
                 PY_VERSION="${i#*=}"
                 shift
                 ;;
-            --build_apu=*)
-                BUILD_APU="${i#*=}"
-                shift
-                ;;
-           --apu_ddk_root=*)
-                APU_DDK_ROOT="${i#*=}"
-                shift
-                ;;
-            --build_rknpu=*)
-                BUILD_RKNPU="${i#*=}"
-                shift
-                ;;
-            --rknpu_ddk_root=*)
-                RKNPU_DDK_ROOT="${i#*=}"
-                shift
-                ;;
-            --with_huawei_ascend_npu=*)
-                WITH_HUAWEI_ASCEND_NPU="${i#*=}"
-                shift
-                ;;
-            --huawei_ascend_npu_ddk_root=*)
-                HUAWEI_ASCEND_NPU_DDK_ROOT="${i#*=}"
-                shift
-                ;;
             --ios_deployment_target=*)
                 IOS_DEPLOYMENT_TARGET="${i#*=}"
                 shift