CMake refine for HIP support.

1. Add option WITH_AMD_GPU.
2. Add cmake/hip.cmake for HIP toolchain.
3. Some external module such as eigen may need HIP port.
4. Add macro hip_library/hip_binary/hip_test to cmake/generic.cmake.
5. Add one HIP source concat.hip.cu as an example. Each .cu may have its corresponding .hip.cu.

CMakeLists.txt

@@ -70,9 +70,6 @@ if(NOT CMAKE_BUILD_TYPE)
       FORCE)
 endif()
 
-if(WITH_AMD_GPU)
-endif()
-
 if(ANDROID OR IOS)
     if(ANDROID)
         if(${CMAKE_SYSTEM_VERSION} VERSION_LESS "16")

cmake/external/eigen.cmake

@@ -1,8 +1,8 @@
 INCLUDE(ExternalProject)
 
 SET(EIGEN_SOURCE_DIR ${THIRD_PARTY_PATH}/eigen3)
-
-INCLUDE_DIRECTORIES(${EIGEN_SOURCE_DIR}/src/extern_eigen3)
+SET(EIGEN_INCLUDE_DIR ${EIGEN_SOURCE_DIR}/src/extern_eigen3)
+INCLUDE_DIRECTORIES(${EIGEN_INCLUDE_DIR})
 
 if(WITH_AMD_GPU)
     ExternalProject_Add(

cmake/hip.cmake

@@ -27,9 +27,6 @@ endif(WITH_TESTING)
 
 if(CMAKE_BUILD_TYPE  STREQUAL "Debug")
     list(APPEND HIP_HCC_FLAGS  ${CMAKE_CXX_FLAGS_DEBUG})
-elseif(CMAKE_BUILD_TYPE  STREQUAL "Release")
-# Disable optimization since one eigen symbol will be removed in math_function.cu
-    #list(APPEND HIP_HCC_FLAGS  ${CMAKE_CXX_FLAGS_RELEASE})
 elseif(CMAKE_BUILD_TYPE  STREQUAL "RelWithDebInfo")
     list(APPEND HIP_HCC_FLAGS  ${CMAKE_CXX_FLAGS_RELWITHDEBINFO})
 elseif(CMAKE_BUILD_TYPE  STREQUAL "MinSizeRel")

paddle/fluid/operators/CMakeLists.txt

@@ -12,6 +12,8 @@ function(op_library TARGET)
     set(OP_LIBRARY ${TARGET} ${OP_LIBRARY} PARENT_SCOPE)
     set(cc_srcs)
     set(cu_srcs)
+    set(hip_cu_srcs)
+    set(miopen_hip_cc_srcs)
     set(cu_cc_srcs)
     set(cudnn_cu_cc_srcs)
     set(CUDNN_FILE)
@@ -36,10 +38,19 @@ function(op_library TARGET)
         if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cu)
             list(APPEND cu_srcs ${TARGET}.cu)
         endif()
+        if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.hip.cu)
+            list(APPEND hip_cu_srcs ${TARGET}.hip.cu)
+        endif()
         string(REPLACE "_op" "_cudnn_op" CUDNN_FILE "${TARGET}")
         if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${CUDNN_FILE}.cu.cc)
             list(APPEND cudnn_cu_cc_srcs ${CUDNN_FILE}.cu.cc)
         endif()
+        if(WITH_AMD_GPU)
+            string(REPLACE "_op" "_miopen_op" MIOPEN_FILE "${TARGET}")
+            if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${MIOPEN_FILE}.hip.cc)
+                list(APPEND miopen_hip_cc_srcs ${MIOPEN_FILE}.hip.cc)
+            endif()
+        endif()
         if(WITH_MKLDNN)
             string(REPLACE "_op" "_mkldnn_op" MKLDNN_FILE "${TARGET}")
             if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${MKLDNN_FILE}.cc)
@@ -48,10 +59,14 @@ function(op_library TARGET)
         endif()
     else()
         foreach(src ${op_library_SRCS})
-            if (${src} MATCHES ".*\\.cu$")
+            if (${src} MATCHES ".*\\.hip.cu$")
+                list(APPEND hip_cu_srcs ${src})
+            elseif (${src} MATCHES ".*\\.cu$")
                 list(APPEND cu_srcs ${src})
             elseif(${src} MATCHES ".*_cudnn_op.cu.cc$")
                 list(APPEND cudnn_cu_cc_srcs ${src})
+            elseif(WITH_AMD_GPU AND ${src} MATCHES ".*_miopen_op.hip.cc$")
+                list(APPEND miopen_hip_cc_srcs ${src})
             elseif(WITH_MKLDNN AND ${src} MATCHES ".*_mkldnn_op.cc$")
                 list(APPEND mkldnn_cc_srcs ${src})
             elseif(${src} MATCHES ".*\\.cu.cc$")
@@ -77,8 +92,8 @@ function(op_library TARGET)
         nv_library(${TARGET} SRCS ${cc_srcs} ${cu_cc_srcs} ${cudnn_cu_cc_srcs} ${mkldnn_cc_srcs} ${cu_srcs} DEPS ${op_library_DEPS}
                 ${op_common_deps})
     elseif (WITH_AMD_GPU)
-        hip_library(${TARGET} SRCS ${cc_srcs} ${hip_cc_srcs} ${miopen_cu_cc_srcs} ${mkldnn_cc_srcs} ${hip_srcs} DEPS
-                ${op_library_DEPS} ${op_common_deps})
+        hip_library(${TARGET} SRCS ${cc_srcs} ${hip_cu_srcs} ${miopen_hip_cc_srcs} ${mkldnn_cc_srcs} DEPS ${op_library_DEPS}
+                ${op_common_deps})
     else()
         cc_library(${TARGET} SRCS ${cc_srcs} ${mkldnn_cc_srcs} DEPS ${op_library_DEPS}
             ${op_common_deps})
@@ -91,7 +106,7 @@ function(op_library TARGET)
         endif()
     endforeach()
 
-    # The registration of USE_OP, please refer to paddle/framework/op_registry.h.
+    # The registration of USE_OP, please refer to paddle/fluid/framework/op_registry.h.
     # Note that it's enough to just adding one operator to pybind in a *_op.cc file.
     # And for detail pybind information, please see generated paddle/pybind/pybind.h.
     file(READ ${TARGET}.cc TARGET_CONTENT)
@@ -117,7 +132,10 @@ function(op_library TARGET)
     list(LENGTH cu_srcs cu_srcs_len)
     list(LENGTH cu_cc_srcs cu_cc_srcs_len)
     list(LENGTH mkldnn_cc_srcs mkldnn_cc_srcs_len)
-    if (${pybind_flag} EQUAL 0 AND ${mkldnn_cc_srcs_len} EQUAL 0 AND ${cu_srcs_len} EQUAL 0 AND ${cu_cc_srcs_len} EQUAL 0)
+    list(LENGTH hip_cu_srcs hip_cu_srcs_len)
+    list(LENGTH miopen_hip_cc_srcs miopen_hip_cc_srcs_len)
+    if (${pybind_flag} EQUAL 0 AND ${mkldnn_cc_srcs_len} EQUAL 0 AND ${cu_srcs_len} EQUAL 0 AND ${cu_cc_srcs_len} EQUAL 0 AND
+        ${hip_cu_srcs_len} EQUAL 0 AND ${miopen_hip_cc_srcs_len} EQUAL 0)
         file(APPEND ${pybind_file} "USE_CPU_ONLY_OP(${TARGET});\n")
         set(pybind_flag 1)
     endif()
@@ -128,6 +146,11 @@ function(op_library TARGET)
         file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, CUDNN);\n")
     endif()
 
+    # pybind USE_OP_DEVICE_KERNEL for MIOPEN
+    if (WITH_AMD_GPU AND ${miopen_hip_cc_srcs_len} GREATER 0)
+        file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, MIOPEN);\n")
+    endif()
+
     # pybind USE_OP_DEVICE_KERNEL for MKLDNN
     if (WITH_MKLDNN AND ${mkldnn_cc_srcs_len} GREATER 0)
         file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, MKLDNN);\n")