compile cuda_ops.so for cuda10.1

Signed-off-by: zhoufeng <zhoufeng54@huawei.com>

cmake/package.cmake

@@ -184,9 +184,10 @@ if(ENABLE_GPU)
         endif()
     endif()
     install(
-            TARGETS cuda_ops
+            TARGETS cuda_ops LIBRARY
             DESTINATION ${INSTALL_PLUGIN_DIR}/gpu${CUDA_VERSION}
             COMPONENT mindspore
+            NAMELINK_SKIP
     )
 endif()
 

cmake/package_plugin.cmake

@@ -89,9 +89,10 @@ if(ENABLE_GPU)
         endif()
     endif()
     install(
-            TARGETS cuda_ops
+            TARGETS cuda_ops LIBRARY
             DESTINATION ${INSTALL_PLUGIN_DIR}/gpu${CUDA_VERSION}
             COMPONENT mindspore
+            NAMELINK_SKIP
     )
 endif()
 
@@ -117,7 +118,7 @@ if(ENABLE_SYM_FILE)
     install(CODE "\
       execute_process(COMMAND ${CMAKE_COMMAND} -DMS_PACK_ROOT_DIR=${CPACK_PACKAGE_DIRECTORY} \
         -DMS_INSTALL_DIR=${CPACK_PACKAGE_DIRECTORY}/_CPack_Packages/${CMAKE_HOST_SYSTEM_NAME}/${CPACK_GENERATOR} \
-        -DMS_PACKAGE_FILE_NAME=${CPACK_PACKAGE_FILE_NAME} -P${CMAKE_SOURCE_DIR}/cmake/plugin_debuginfo_script.cmake)"
+        -DMS_PACKAGE_FILE_NAME=${CPACK_PACKAGE_FILE_NAME} -P ${CMAKE_SOURCE_DIR}/cmake/plugin_debuginfo_script.cmake)"
     )
 endif()
 

mindspore/ccsrc/CMakeLists.txt

@@ -59,18 +59,16 @@ function(set_nvcc_flag CUDA_NVCC_FLAGS)
             list(APPEND cuda_archs_bin ${arch_bin})
         endforeach()
     else()
-        message("Failed to detect gpu arch automatically, build a base arch 5.3.")
+        message("Failed to detect gpu arch automatically, build a base arch 6.0.")
-        list(APPEND CUDA_NVCC_FLAGS -arch=sm_53)
+        list(APPEND CUDA_NVCC_FLAGS -arch=sm_60)
     endif()
     # Get build flag from env to choose common/auto build.
     set(NVCC_ARCH_FLAG_FROM_ENV $ENV{CUDA_ARCH})
     if(NVCC_ARCH_FLAG_FROM_ENV STREQUAL "common")
         message("Build common archs for release.")
-        list(APPEND CUDA_NVCC_FLAGS -gencode=arch=compute_53,code=sm_53
+        list(APPEND CUDA_NVCC_FLAGS -gencode=arch=compute_60,code=sm_60
-                                    -gencode=arch=compute_60,code=sm_60
                                     -gencode=arch=compute_61,code=sm_61
-                                    -gencode=arch=compute_70,code=sm_70
+                                    -gencode=arch=compute_70,code=sm_70)
-                                    -gencode=arch=compute_72,code=sm_72)
         if(${CUDA_VERSION} VERSION_GREATER "9.5")
             list(APPEND CUDA_NVCC_FLAGS -gencode=arch=compute_75,code=sm_75)
             if(${CUDA_VERSION} VERSION_LESS "11.0")
@@ -93,9 +91,9 @@ function(set_nvcc_flag CUDA_NVCC_FLAGS)
         foreach(arch ${cuda_archs_bin})
             list(APPEND CUDA_NVCC_FLAGS -gencode=arch=compute_${arch},code=sm_${arch})
         endforeach()
-        # For auto build, it only generate the code for indeed arch, so add sm_53 as a default arch
+        # For auto build, it only generate the code for indeed arch, so add sm_60 as a default arch
         # to avoid error in different archs. It may increase the compilation time.
-        list(APPEND CUDA_NVCC_FLAGS -arch=sm_53)
+        list(APPEND CUDA_NVCC_FLAGS -arch=sm_60)
     endif()
     list(REMOVE_DUPLICATES CUDA_NVCC_FLAGS)
     message("Final CUDA_NVCC_FLASG " ${CUDA_NVCC_FLAGS})

mindspore/ccsrc/plugin/device/gpu/kernel/cuda_impl/CMakeLists.txt

@@ -34,6 +34,9 @@ if(ENABLE_GPU)
     target_compile_options(cuda_common_obj PRIVATE "/std:c++17")
     endif()
     cuda_add_library(cuda_ops SHARED ${CUDA_OPS_SRC_LIST} $<TARGET_OBJECTS:cuda_common_obj>)
+    if(NOT BUILD_LITE)
+      set_target_properties(cuda_ops PROPERTIES SOVERSION ${CUDA_VERSION_MAJOR})
+    endif()
     if(NOT MSVC)
         target_link_options(cuda_ops PRIVATE "-Wl,--allow-shlib-undefined")
     else()

scripts/build/merge_whl_package.sh

@@ -28,6 +28,7 @@ for whl in "$@"; do
 done
 
 MAX_GPU_VERSION=0
+declare -A GPU_VERSION_MAP
 for ((i=1;i<$counter;i=$i+1))
 do
   echo "Rename $i dirname to mindspore ..."
@@ -50,18 +51,32 @@ do
   fi;
   CUR_GPU_VERSION=`find "./$i/mindspore/lib/plugin" -name 'gpu*' -exec sh -c 'echo ${0##*gpu}' {} \;`
   if [ -n "$CUR_GPU_VERSION" ]; then
+    GPU_VERSION_MAP[$CUR_GPU_VERSION]=$i
+  else
+    rm -rf $i
+  fi;
+done
+
+for key in $(for x in "${!GPU_VERSION_MAP[@]}"; do echo $x; done | sort)
+do
+  i=${GPU_VERSION_MAP[$key]}
+  CUR_GPU_VERSION=$key
+  CUDA_OPS_FILE=`basename ./$i/mindspore/lib/plugin/gpu$CUR_GPU_VERSION/libcuda_ops.so*`
+  if [ "`echo "$CUR_GPU_VERSION > $MAX_GPU_VERSION" | bc`" -eq 1 ]; then
     if [ "`echo "$CUR_GPU_VERSION > $MAX_GPU_VERSION" | bc`" -eq 1 ]; then
       MAX_GPU_VERSION=$CUR_GPU_VERSION
-      if [ ! -d "$BASE_PACKAGE_UNZIP_DIR/mindspore/lib/plugin/gpu" ]; then
-        mkdir -p $BASE_PACKAGE_UNZIP_DIR/mindspore/lib/plugin/gpu
-      fi;
-      \cp -rf ./$i/mindspore/lib/plugin/gpu$CUR_GPU_VERSION/libcuda_ops.so $BASE_PACKAGE_UNZIP_DIR/mindspore/lib/plugin/gpu
     fi;
-    rm -f $BASE_PACKAGE_UNZIP_DIR/mindspore/lib/plugin/gpu$CUR_GPU_VERSION/libcuda_ops.so
+    if [ ! -d "$BASE_PACKAGE_UNZIP_DIR/mindspore/lib/plugin/gpu" ]; then
+      mkdir -p $BASE_PACKAGE_UNZIP_DIR/mindspore/lib/plugin/gpu
+    fi;
+    \cp -rf ./$i/mindspore/lib/plugin/gpu$CUR_GPU_VERSION/$CUDA_OPS_FILE \
+      $BASE_PACKAGE_UNZIP_DIR/mindspore/lib/plugin/gpu/$CUDA_OPS_FILE
   fi;
+  rm -f $BASE_PACKAGE_UNZIP_DIR/mindspore/lib/plugin/gpu$CUR_GPU_VERSION/libcuda_ops.so*
   rm -rf $i
 done
 
+
 export COMMIT_ID=`cat $BASE_PACKAGE_UNZIP_DIR/mindspore/.commit_id | awk '{print $3}' | sed $'s/\'//g'`
 VERSION=`cat $BASE_PACKAGE_UNZIP_DIR/mindspore/version.py | awk '{print $3}' | sed $'s/\'//g'`
 echo -n "$VERSION" > $MINDSPORE_ROOT_PATH/version.txt