Fix cmake for AMDGPU platform (#13801)

* HIP cmake.
Enable whole archieve build for pybind library.

Disable two warning.

Rollback to C++11.

Link RCCL to WA gpu kernel loading issue.

Update eigen to fix build failure.

Add more include directories.

Fix O3 build failure.

Update eigen.

fix tensor_util_test segment fault issue

add more macro check in hip.cmake.
we may consider refine hip.cmake to inherit all add_definitions() in parrent scope, in the future.

Fix rocRAND load.

Update eigen to fix gru_unit_op and reduce_op.

Add HIP support to testing.

Update eigen to support int16 and int8 in arg min and arg max.

* add rocprim as cub library used by nv implementation

* Reduce build time in rocprim.

* Add rocprim introduction, remove useless cmake code.

* Remove useless flags and format cmake file.

Author: sabreshao

CMakeLists.txt

@@ -204,6 +204,7 @@ include(external/eigen)     # download eigen3
 include(external/pybind11)  # download pybind11
 include(external/cares)
 include(external/cub)
+include(external/rocprim)
 include(external/xxhash)    # download xxhash
 include(external/dlpack)
 include(external/snappy)    # download snappy

cmake/external/eigen.cmake

@@ -17,7 +17,7 @@ if(WITH_AMD_GPU)
         extern_eigen3
         ${EXTERNAL_PROJECT_LOG_ARGS}
         GIT_REPOSITORY  "https://github.com/sabreshao/hipeigen.git"
-        GIT_TAG         0cba03ff9f8f9f70bbd92ac5857b031aa8fed6f9
+        GIT_TAG         7cb2b6e5a4b4a1efe658abb215cd866c6fb2275e
         PREFIX          ${EIGEN_SOURCE_DIR}
         UPDATE_COMMAND  ""
         CONFIGURE_COMMAND ""

cmake/external/rocprim.cmake

@@ -0,0 +1,44 @@
+if (NOT WITH_AMD_GPU)
+    return()
+endif()
+
+# rocprim is "ROCm Parallel Primitives" for short.
+# It is a header-only library providing HIP and HC parallel primitives
+# for developing performant GPU-accelerated code on AMD ROCm platform.
+
+if("x${HCC_HOME}" STREQUAL "x")
+  set(HCC_HOME "/opt/rocm/hcc")
+endif()
+
+INCLUDE(ExternalProject)
+
+SET(ROCPRIM_SOURCE_DIR ${THIRD_PARTY_PATH}/rocprim)
+SET(ROCPRIM_INSTALL_DIR  ${THIRD_PARTY_PATH}/install/rocprim)
+SET(ROCPRIM_INCLUDE_DIR ${ROCPRIM_INSTALL_DIR}/include)
+
+ExternalProject_Add(
+    extern_rocprim
+    GIT_REPOSITORY "https://github.com/ROCmSoftwarePlatform/rocPRIM.git"
+    GIT_TAG        5bd41b96ab8d8343330fb2c3e1b96775bde3b3fc 
+    PREFIX         ${ROCPRIM_SOURCE_DIR}
+    UPDATE_COMMAND  ""
+    CMAKE_ARGS     -DCMAKE_CXX_COMPILER=${HCC_HOME}/bin/hcc
+    CMAKE_ARGS     -DONLY_INSTALL=ON
+    CMAKE_ARGS     -DBUILD_TEST=OFF
+    CMAKE_ARGS     -DCMAKE_INSTALL_PREFIX=${ROCPRIM_INSTALL_DIR}
+
+    INSTALL_DIR    ${ROCPRIM_INSTALL_DIR}
+    ${EXTERNAL_PROJECT_LOG_ARGS}
+)
+
+INCLUDE_DIRECTORIES(${ROCPRIM_INCLUDE_DIR})
+
+if (${CMAKE_VERSION} VERSION_LESS "3.3.0")
+    set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/rocprim_dummy.c)
+    file(WRITE ${dummyfile} "const char *dummy_rocprim = \"${dummyfile}\";")
+    add_library(rocprim STATIC ${dummyfile})
+else()
+    add_library(rocprim INTERFACE)
+endif()
+
+add_dependencies(rocprim extern_rocprim)

cmake/flags.cmake

@@ -129,6 +129,9 @@ set(COMMON_FLAGS
     -Wno-error=parentheses-equality # Warnings in pybind11
     -Wno-error=ignored-attributes  # Warnings in Eigen, gcc 6.3
     -Wno-error=terminate  # Warning in PADDLE_ENFORCE
+    -Wno-error=int-in-bool-context # Warning in Eigen gcc 7.2
+    -Wimplicit-fallthrough=0 # Warning in tinyformat.h
+    -Wno-error=maybe-uninitialized # Warning in boost gcc 7.2
 )
 
 set(GPU_COMMON_FLAGS

cmake/generic.cmake

@@ -454,25 +454,29 @@ function(hip_library TARGET_NAME)
       else()
         add_library(${TARGET_NAME} STATIC ${_cmake_options} ${_generated_files} ${_sources})
         set_target_properties(${TARGET_NAME} PROPERTIES LINKER_LANGUAGE CXX)
-        target_link_libraries(${TARGET_NAME} /opt/rocm/hip/lib/libhip_hcc.so /opt/rocm/hip/lib/libhip_device.a)
-	find_fluid_modules(${TARGET_NAME})
+        target_link_libraries(${TARGET_NAME} /opt/rocm/hip/lib/libhip_hcc.so /opt/rocm/hip/lib/libhip_device.a /opt/rocm/rccl/lib/librccl.so /opt/rocm/hiprand/lib/libhiprand.so)
+        find_fluid_modules(${TARGET_NAME})
       endif()
-      if (hip_library_DEPS)
-	add_dependencies(${TARGET_NAME} ${hip_library_DEPS})
-	target_link_libraries(${TARGET_NAME} ${hip_library_DEPS})
+      if("${hip_library_DEPS}" MATCHES "ARCHIVE_START")
+        # Support linking flags: --whole-archive (Linux) / -force_load (MacOS).
+        # WARNING: Please don't use ARCHIVE_START&ARCHIVE_END if TARGET_NAME will be linked by other libraries.
+        target_circle_link_libraries(${TARGET_NAME} ${hip_library_DEPS})
+        list(REMOVE_ITEM hip_library_DEPS ARCHIVE_START ARCHIVE_END)
+      else()
+        target_link_libraries(${TARGET_NAME} ${hip_library_DEPS})
       endif()
       # cpplint code style
       foreach(source_file ${hip_library_SRCS})
-	string(REGEX REPLACE "\\.[^.]*$" "" source ${source_file})
-	if(EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h)
-	  list(APPEND hip_library_HEADERS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h)
-	endif()
+        string(REGEX REPLACE "\\.[^.]*$" "" source ${source_file})
+        if(EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h)
+          list(APPEND hip_library_HEADERS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h)
+        endif()
       endforeach()
     else(hip_library_SRCS)
       if (hip_library_DEPS)
-	merge_static_libs(${TARGET_NAME} ${hip_library_DEPS})
+        merge_static_libs(${TARGET_NAME} ${hip_library_DEPS})
       else()
-	message(FATAL "Please specify source file or library in nv_library.")
+        message(FATAL "Please specify source file or library in nv_library.")
       endif()
     endif(hip_library_SRCS)
   endif()

cmake/hip.cmake

@@ -3,6 +3,8 @@ if(NOT WITH_AMD_GPU)
 endif()
 
 include_directories("/opt/rocm/include")
+include_directories("/opt/rocm/hip/include")
+include_directories("/opt/rocm/miopen/include")
 include_directories("/opt/rocm/hipblas/include")
 include_directories("/opt/rocm/hiprand/include")
 include_directories("/opt/rocm/rocrand/include")
@@ -11,20 +13,40 @@ include_directories("/opt/rocm/thrust")
 
 list(APPEND EXTERNAL_LIBS "-L/opt/rocm/lib/ -lhip_hcc")
 
-set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -fPIC -DPADDLE_WITH_HIP -std=c++14" )
+set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -fPIC -DPADDLE_WITH_HIP -std=c++11" )
 
 if(WITH_DSO)
   set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_USE_DSO")
 endif(WITH_DSO)
 
-if(WITH_DOUBLE)
-  set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_TYPE_DOUBLE")
-endif(WITH_DOUBLE)
-
 if(WITH_TESTING)
   set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_WITH_TESTING")
 endif(WITH_TESTING)
 
+if(WITH_DISTRIBUTE)
+  set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_WITH_DISTRIBUTE")
+endif(WITH_DISTRIBUTE)
+
+if(WITH_GRPC)
+  set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_WITH_GRPC")
+endif(WITH_GRPC)
+
+if(NOT WITH_GOLANG)
+  set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_WITHOUT_GOLANG")
+endif(NOT WITH_GOLANG)
+
+if(WITH_MKLDNN)
+  set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_WITH_MKLDNN")
+endif(WITH_MKLDNN)
+
+set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DANY_IMPL_ANY_CAST_MOVEABLE")
+
+if(NOT WITH_RDMA)
+  set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_DISABLE_RDMA")
+endif(NOT WITH_RDMA)
+
+
+
 if(CMAKE_BUILD_TYPE  STREQUAL "Debug")
     list(APPEND HIP_HCC_FLAGS  ${CMAKE_CXX_FLAGS_DEBUG})
 elseif(CMAKE_BUILD_TYPE  STREQUAL "RelWithDebInfo")

paddle/fluid/pybind/CMakeLists.txt

@@ -5,8 +5,8 @@ if(WITH_PYTHON)
   if(WITH_AMD_GPU)
     hip_library(paddle_pybind SHARED
       SRCS ${PYBIND_SRCS}
-      DEPS ${PYBIND_DEPS}
-      ${GLOB_OP_LIB} ${GLOB_OPERATOR_DEPS})
+      DEPS ARCHIVE_START ${PYBIND_DEPS}
+      ${GLOB_OP_LIB} ${GLOB_OPERATOR_DEPS} ARCHIVE_END)
   else()
     cc_library(paddle_pybind SHARED
       SRCS ${PYBIND_SRCS}

paddle/testing/paddle_gtest_main.cc

@@ -28,7 +28,7 @@ int main(int argc, char** argv) {
   for (int i = 0; i < argc; ++i) {
     new_argv.push_back(argv[i]);
   }
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
   new_argv.push_back(
       strdup("--tryfromenv=fraction_of_gpu_memory_to_use,allocator_strategy"));
 #else