link static library with cuda, test=develop (#2228)

* add static libraries of cuda, test=develop

* update cuda make

cmake/cuda.cmake

@@ -174,6 +174,16 @@ if(NOT WITH_DSO)
     endif(WIN32)
 endif(NOT WITH_DSO)
 
+get_filename_component(CUDA_LIB_PATH ${CUDA_curand_LIBRARY} DIRECTORY)
+function(import_static_library alias path)
+    add_library(${alias} STATIC IMPORTED GLOBAL)
+    set_property(TARGET ${alias} PROPERTY IMPORTED_LOCATION ${path})
+endfunction()
+import_static_library(cudart_static ${CUDA_LIB_PATH}/libcudart_static.a)
+import_static_library(cublas_static ${CUDA_LIB_PATH}/libcublas_static.a)
+import_static_library(curand_static ${CUDA_LIB_PATH}/libcurand_static.a)
+import_static_library(culibos_static ${CUDA_LIB_PATH}/libculibos.a)
+
 # setting nvcc arch flags
 select_nvcc_arch_flags(NVCC_FLAGS_EXTRA)
 list(APPEND CUDA_NVCC_FLAGS ${NVCC_FLAGS_EXTRA})

cmake/cudnn.cmake

@@ -53,11 +53,10 @@ if(APPLE)
 set(CUDNN_LIB_NAME "libcudnn.dylib" "libcudnn.so")
 endif(APPLE)
 
-find_library(CUDNN_LIBRARY NAMES ${CUDNN_LIB_NAME} # libcudnn_static.a
+find_library(CUDNN_LIBRARY NAMES ${CUDNN_LIB_NAME}
     PATHS ${CUDNN_CHECK_LIBRARY_DIRS} ${CUDNN_INCLUDE_DIR} ${__libpath_hist}
           NO_DEFAULT_PATH
-    DOC "Path to cuDNN library.")
-
+    DOC "Path to cuDNN dynamic library.")
 
 if(CUDNN_INCLUDE_DIR AND CUDNN_LIBRARY)
     set(CUDNN_FOUND ON)
@@ -69,6 +68,9 @@ if(CUDNN_FOUND)
     file(READ ${CUDNN_INCLUDE_DIR}/cudnn.h CUDNN_VERSION_FILE_CONTENTS)
 
     get_filename_component(CUDNN_LIB_PATH ${CUDNN_LIBRARY} DIRECTORY)
+    add_library(cudnn_static STATIC IMPORTED GLOBAL)
+    set_property(TARGET cudnn_static PROPERTY IMPORTED_LOCATION
+               "${CUDNN_LIB_PATH}/libcudnn_static.a")
 
     string(REGEX MATCH "define CUDNN_VERSION +([0-9]+)"
         CUDNN_VERSION "${CUDNN_VERSION_FILE_CONTENTS}")

lite/api/CMakeLists.txt

@@ -68,6 +68,8 @@ set(light_api_deps
     scope target_wrapper_host model_parser program)
 if(LITE_WITH_CUDA)
     set(light_api_deps ${light_api_deps} target_wrapper_cuda)
+    set(cuda_static_deps cudart_static cublas_static curand_static
+        cudnn_static culibos_static)
 endif()
 lite_cc_library(light_api SRCS light_api.cc
         DEPS scope target_wrapper_host model_parser
@@ -195,7 +197,7 @@ if (NOT LITE_ON_TINY_PUBLISH)
     # The final inference library for just MobileConfig.
     bundle_static_library(paddle_api_full paddle_api_full_bundled bundle_full_api)
     get_property(fluid_modules GLOBAL PROPERTY FLUID_MODULES)
-    cc_library(api_full_static SRCS DEPS paddle_api_full cxx_api paddle_api light_api  ${cxx_api_deps} ${ops} ${host_kernels} ${cuda_kernels} program tensor memory naive_buffer types ${fluid_modules} protobuf)
+    cc_library(api_full_static SRCS DEPS paddle_api_full cxx_api paddle_api light_api  ${cxx_api_deps} ${ops} ${host_kernels} ${cuda_kernels} program tensor memory naive_buffer types ${fluid_modules} protobuf ${cuda_static_deps})
 endif()
 bundle_static_library(paddle_api_light paddle_api_light_bundled bundle_light_api)
 #-----------------------------------------------------------------------------------------------------

lite/backends/cuda/math/CMakeLists.txt

@@ -2,13 +2,16 @@ if(NOT LITE_WITH_CUDA)
     return()
 endif()
 
-nv_library(cuda_activation SRCS activation.cu)
-nv_library(cuda_scale SRCS scale.cu)
-nv_library(cuda_type_trans SRCS type_trans.cu)
-nv_library(cuda_transpose SRCS transpose.cu )
+set(cuda_static_deps cudnn_static cublas_static curand_static
+    culibos_static cudart_static)
+
+nv_library(cuda_activation SRCS activation.cu DEPS ${cuda_static_deps})
+nv_library(cuda_scale SRCS scale.cu DEPS ${cuda_static_deps})
+nv_library(cuda_type_trans SRCS type_trans.cu DEPS ${cuda_static_deps}) 
+nv_library(cuda_transpose SRCS transpose.cu DEPS ${cuda_static_deps})
 nv_library(cudnn_conv SRCS cudnn_conv.cc DEPS cuda_activation cuda_scale
-cuda_type_trans)
-nv_library(cuda_elementwise SRCS elementwise.cu )
+cuda_type_trans ${cuda_static_deps})
+nv_library(cuda_elementwise SRCS elementwise.cu DEPS ${cuda_static_deps})
 
 set (
  math_cuda