Merge pull request #2824 from hedaoyuan/fix_nnpack

Fix some link errors about NNPACK.

CMakeLists.txt

@@ -137,7 +137,8 @@ if(WITH_GPU)
 endif(WITH_GPU)
 
 if(USE_NNPACK)
-  list(APPEND EXTERNAL_LIBS ${NNPACK_LIB} ${PTHREADPOOL_LIB} "rt")
+    include(external/nnpack)
+    list(APPEND EXTERNAL_LIBS ${NNPACK_LIBS})
 endif(USE_NNPACK)
 
 add_subdirectory(proto)

cmake/external/nnpack.cmake

@@ -7,10 +7,24 @@ set(NNPACK_ROOT $ENV{NNPACK_ROOT} CACHE PATH "Folder contains NNPACK")
 find_path(NNPACK_INC_DIR nnpack.h PATHS ${NNPACK_ROOT}/include)
 find_library(NNPACK_LIB NAMES nnpack PATHS ${NNPACK_ROOT}/lib)
 find_library(PTHREADPOOL_LIB NAMES pthreadpool PATHS ${NNPACK_ROOT}/lib)
+find_library(NNPACK_UKERNELS_LIB NAMES nnpack_ukernels PATHS ${NNPACK_ROOT}/lib)
+find_library(NNPACK_CPUFEATURES_LIB NAMES cpufeatures PATHS ${NNPACK_ROOT}/lib)
 
 if(NNPACK_INC_DIR AND NNPACK_LIB AND PTHREADPOOL_LIB)
   set(NNPACK_FOUND ON)
   INCLUDE_DIRECTORIES(${NNPACK_INC_DIR})
+
+  set(NNPACK_LIBS)
+  list(APPEND NNPACK_LIBS ${NNPACK_LIB} ${PTHREADPOOL_LIB})
+  if (NNPACK_UKERNELS_LIB)
+    list(APPEND NNPACK_LIBS ${NNPACK_UKERNELS_LIB})
+  endif()
+  if (NNPACK_CPUFEATURES_LIB)
+    list(APPEND NNPACK_LIBS ${NNPACK_CPUFEATURES_LIB})
+  endif()
+  if(NOT ANDROID)
+    list(APPEND NNPACK_LIBS "rt")
+  endif()
 else()
   message(FATAL_ERROR "Cannot find NNPACK in (${NNPACK_ROOT})")
 endif()

paddle/function/CMakeLists.txt

@@ -11,7 +11,6 @@ if(WITH_GPU)
 endif()
 
 if(USE_NNPACK)
-  include(nnpack/nnpack.cmake)
   list(APPEND cpp_files nnpack/NNPACKConvOp.cpp)
   if(WITH_TESTING)
     add_unittest(NNPACKConvOpTest nnpack/NNPACKConvOpTest.cpp)

paddle/function/nnpack/NNPACKConvOp.cpp

@@ -16,7 +16,7 @@ limitations under the License. */
 #include "paddle/function/ConvOp.h"
 
 DEFINE_bool(nnpack_allocate_outside,
-            false,
+            true,
             "Allocate and free workspace memory outside the NNPACK interface.");
 DEFINE_int32(nnpack_num_threads,
              0,
@@ -58,18 +58,10 @@ public:
     workspaceBuffer_ = nullptr;
     workspaceSize_ = 0;
 
-    threadpool_ = nullptr;
-    if (FLAGS_nnpack_num_threads) {
-      threadpool_ = pthreadpool_create(FLAGS_nnpack_num_threads);
-      VLOG(3) << "Number of threads "
-              << pthreadpool_get_threads_count(threadpool_);
-    }
+    create_nnpack_threadpool();
   }
 
   ~NNPACKConvFunction() {
-    if (threadpool_) {
-      pthreadpool_destroy(threadpool_);
-    }
     if (workspaceBuffer_) {
       free(workspaceBuffer_);
     }
@@ -225,14 +217,25 @@ public:
     }
   }
 
+  static void create_nnpack_threadpool() {
+    if (FLAGS_nnpack_num_threads && threadpool_ == nullptr) {
+      threadpool_ = pthreadpool_create(FLAGS_nnpack_num_threads);
+      VLOG(3) << "Number of threads "
+              << pthreadpool_get_threads_count(threadpool_);
+    }
+  }
+
 private:
   nnp_convolution_algorithm algorithm_;
   nnp_convolution_transform_strategy transform_strategy_;
   void* workspaceBuffer_;
   size_t workspaceSize_;
-  pthreadpool_t threadpool_;
+  static pthreadpool_t threadpool_;
 };
 
+template <DeviceType Device>
+pthreadpool_t NNPACKConvFunction<Device>::threadpool_ = nullptr;
+
 REGISTER_TYPED_FUNC(NNPACKConv, CPU, NNPACKConvFunction);
 
 }  // namespace paddle