combination the WITH_XPU and WITH_XPU2

PaddlePaddle · Nov 23, 2021 · 9c0a1dd · 9c0a1dd
1 parent a49660e
commit 9c0a1dd
Show file tree

Hide file tree

Showing 4 changed files with 91 additions and 52 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -270,13 +270,13 @@ if (NOT WITH_GPU AND WITH_NCCL)
         "Disable NCCL when compiling without GPU" FORCE)
 endif()
 
-# XPU XPU2 use BKCL
-# if (NOT (WITH_XPU OR WITH_XPU2) AND WITH_XPU_BKCL)
-#     MESSAGE(WARNING
-#         "Disable BKCL when compiling without XPU. Force WITH_XPU_BKCL=OFF.")
-#     set(WITH_XPU_BKCL OFF CACHE STRING
-#         "Disable BKCL when compiling without XPU" FORCE)
-# endif()
+# force XPU on when WITH_XPU2
+if (WITH_XPU2 AND NOT WITH_XPU)
+    MESSAGE(WARNING
+        "Enable XPU when compiling with XPU2. Force WITH_XPU=ON.")
+    set(WITH_XPU ON CACHE STRING
+        "Enable XPU when compiling with XPU2" FORCE)
+endif()
 
 #
 if (NOT WITH_XPU AND WITH_XPU_BKCL)

diff --git a/cmake/generic.cmake b/cmake/generic.cmake
@@ -401,8 +401,8 @@ function(cc_binary TARGET_NAME)
     target_link_libraries(${TARGET_NAME} ${ROCM_HIPRTC_LIB})
   endif()
   if(WITH_XPU2)
-    target_link_libraries(${TARGET_NAME} ${XPU2_CLANG_API_LIB})
-    target_link_libraries(${TARGET_NAME} ${XPU2_CLANG_RT_LIB})
+    # target_link_libraries(${TARGET_NAME} ${XPU2_CLANG_API_LIB})
+    # target_link_libraries(${TARGET_NAME} ${XPU2_CLANG_RT_LIB})
   endif()
 
   check_coverage_opt(${TARGET_NAME} ${cc_binary_SRCS})
@@ -430,7 +430,7 @@ function(cc_test_build TARGET_NAME)
     endif()
     # added by lxd
     if(WITH_XPU2)
-      target_link_libraries(${TARGET_NAME} ${XPU2_CLANGRTC_LIB})
+      #target_link_libraries(${TARGET_NAME} ${XPU2_CLANGRTC_LIB})
     endif()
     check_coverage_opt(${TARGET_NAME} ${cc_test_SRCS})
   endif()

diff --git a/cmake/operators.cmake b/cmake/operators.cmake
@@ -132,20 +132,53 @@ function(op_library TARGET)
                 list(APPEND cu_cc_srcs ${src})
             elseif(WITH_XPU AND ${src} MATCHES ".*_op_xpu.cc$")
                 list(APPEND xpu_cc_srcs ${src})
-            elseif(WITH_XPU2 AND ${src} MATCHES ".*_op_kps.cc$")
-                list(APPEND xpu2_cc_srcs ${src})
             elseif(WITH_XPU2 AND ${src} MATCHES ".*\\.xpu$")
                 list(APPEND xpu2_cc_srcs ${src})
             elseif(WITH_ASCEND_CL AND ${src} MATCHES ".*_op_npu.cc$")
                 list(APPEND npu_cc_srcs ${src})
             elseif(${src} MATCHES ".*\\.cc$")
                 list(APPEND cc_srcs ${src})
             else()
-                message(FATAL_ERROR "${TARGET} Source file ${src} should only be .cc or .cu")
+                message(FATAL_ERROR "${TARGET} Source file ${src} should only be .cc or .cu or .xpu")
+            endif()
+        endforeach()
+    endif()
+
+    list(LENGTH xpu_cc_srcs xpu_cc_srcs_len)
+    list(LENGTH xpu2_cc_srcs xpu2_cc_srcs_len)
+    if(WITH_XPU AND WITH_XPU2)
+        if(${xpu2_cc_srcs_len})
+            # message(STATUS "lxd_debug: >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>delete duplicate op in xpu")
+            # message(STATUS " xpu ${xpu_cc_srcs} xpu2 ${xpu2_cc_srcs}")
+        endif()
+    endif()
+
+    #TODO(liuxiandong) 
+    if(WITH_XPU2 AND ${xpu2_cc_srcs_len})
+        foreach(src ${xpu2_cc_srcs})
+            get_filename_component(op_name ${src} NAME_WE)
+            message(STATUS "lxd_debug op_name ${op_name}")
+            if(WITH_XPU)
+                #TODO
+                if(${xpu_cc_srcs} MATCHES "elementwise_add_op_xpu.cc")
+                    # delete it from the xpu_cc_srcs
+                    #list(REMOVE_ITEM ${xpu_cc_srcs} ${op_name}_xpu.cc)
+                    list(REMOVE_ITEM xpu_cc_srcs "elementwise_add_op_xpu.cc")
+                endif()
             endif()
         endforeach()
     endif()
 
+    if(WITH_XPU AND WITH_XPU2)
+        if(${xpu2_cc_srcs_len})
+            # message(STATUS "lxd_debug: <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<delete duplicate op in xpu")
+            # message(STATUS " xpu ${xpu_cc_srcs} xpu2 ${xpu2_cc_srcs}")
+        endif()
+    endif()
+    list(LENGTH xpu_cc_srcs xpu_cc_srcs_len)
+    list(LENGTH xpu2_cc_srcs xpu2_cc_srcs_len)
+    # message(STATUS "lxd_debug: !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! xpu ${xpu_cc_srcs} xpu2 ${xpu2_cc_srcs}")
+
     list(LENGTH cc_srcs cc_srcs_len)
     if (${cc_srcs_len} EQUAL 0)
         message(FATAL_ERROR "The op library ${TARGET} should contains at least one .cc file")
@@ -208,9 +241,12 @@ function(op_library TARGET)
         list(REMOVE_ITEM hip_srcs "decode_jpeg_op.cu")
         hip_library(${TARGET} SRCS ${cc_srcs} ${hip_cc_srcs} ${miopen_cu_cc_srcs} ${miopen_cu_srcs} ${mkldnn_cc_srcs} ${hip_srcs} DEPS ${op_library_DEPS}
                 ${op_common_deps})
-    elseif (WITH_XPU2)
+    elseif (WITH_XPU2 AND ${xpu2_cc_srcs_len} GREATER 0)
         xpu_library(${TARGET} SRCS ${cc_srcs} ${mkldnn_cc_srcs} ${xpu2_cc_srcs} DEPS ${op_library_DEPS} ${op_common_deps})
     else()
+        # if(WITH_XPU2 AND ${xpu2_cc_srcs_len} GREATER 0)
+        #     xpu_library(${TARGET} SRCS ${cc_srcs} ${mkldnn_cc_srcs} ${xpu2_cc_srcs} DEPS ${op_library_DEPS} ${op_common_deps})
+        # endif()
         # Unity Build relies on global option `WITH_UNITY_BUILD` and local option `UNITY`.
         if(WITH_UNITY_BUILD AND op_library_UNITY)
             # Combine the cc source files.
@@ -277,8 +313,8 @@ function(op_library TARGET)
     list(LENGTH cu_cc_srcs cu_cc_srcs_len)
     list(LENGTH hip_cc_srcs hip_cc_srcs_len)
     list(LENGTH mkldnn_cc_srcs mkldnn_cc_srcs_len)
-    list(LENGTH xpu_cc_srcs xpu_cc_srcs_len)
-    list(LENGTH xpu2_cc_srcs xpu2_cc_srcs_len)
+    #list(LENGTH xpu_cc_srcs xpu_cc_srcs_len)
+    #list(LENGTH xpu2_cc_srcs xpu2_cc_srcs_len)
     list(LENGTH miopen_cu_cc_srcs miopen_cu_cc_srcs_len)
     list(LENGTH npu_cc_srcs npu_cc_srcs_len)
     if (${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
@@ -320,10 +356,13 @@ function(op_library TARGET)
         file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, CUDNN);\n")
     endif()
 
-    if (WITH_XPU AND ${pybind_flag} EQUAL 0 AND ${xpu_cc_srcs_len} GREATER 0)
-        #file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, XPU);\n")
+    if (WITH_XPU AND ${pybind_flag} EQUAL 0 AND ${xpu_cc_srcs_len} GREATER 0 AND ${xpu2_cc_srcs_len} EQUAL 0)
+        #message(STATUS "lxd_debug: ${TARGET} op in XPU1")
+        file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, XPU);\n")
     endif()
 
+    #message(STATUS "lxd_debug: cmake source dir is: ${CMAKE_SOURCE_DIR}")
+
     if (WITH_XPU2 AND ${xpu2_cc_srcs_len} GREATER 0)
         file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, XPU);\n")
     endif()

diff --git a/cmake/xpu2.cmake b/cmake/xpu2.cmake
@@ -93,7 +93,7 @@ macro(compile_kernel COMPILE_ARGS)
   set(XTDK_DIR ${XPU_TOOLCHAIN})
   set(CXX_DIR ${HOST_SYSROOT})
   #-Wno-error=constant-conversion -Wno-error=c++11-narrowing -Wno-error=shift-count-overflow
-  set(XPU_CXX_INCLUDES  -I/workspace/paddle/Paddle/build -I/workspace/paddle/Paddle/paddle/fluid/framework/io -I/workspace/paddle/Paddle/build/third_party/install/zlib/include -I/workspace/paddle/Paddle/build/third_party/install -I/workspace/paddle/Paddle/build/third_party/install/gflags/include -I/workspace/paddle/Paddle/build/third_party/install/glog/include -I/workspace/paddle/Paddle/build/third_party/boost/src/extern_boost -I/workspace/paddle/Paddle/build/third_party/eigen3/src/extern_eigen3 -I/workspace/paddle/Paddle/build/third_party/threadpool/src/extern_threadpool -I/workspace/paddle/Paddle/build/third_party/dlpack/src/extern_dlpack/include -I/workspace/paddle/Paddle/build/third_party/install/xxhash/include -I/workspace/paddle/Paddle/build/third_party/install/warpctc/include -I/workspace/paddle/Paddle/build/third_party/install/openblas/include -I/workspace/paddle/Paddle/build/third_party/install/protobuf/include -I/usr/include/python3.7m -I/usr/local/lib/python3.7/dist-packages/numpy/core/include -I/workspace/paddle/Paddle/build/third_party/pybind/src/extern_pybind/include -I/workspace/paddle/Paddle/build/third_party/install/gtest/include -I/workspace/paddle/Paddle/build/third_party/install/gloo/include -I/workspace/paddle/Paddle/build/third_party/install/xbyak/include -I/workspace/paddle/Paddle/build/third_party/install/xbyak/include/xbyak -I/workspace/paddle/Paddle/build/third_party/install/cryptopp/include -I/workspace/paddle/Paddle/build/third_party/pocketfft/src -I/workspace/paddle/Paddle)
+  set(XPU_CXX_INCLUDES  -I${CMAKE_SOURCE_DIR}/build -I${CMAKE_SOURCE_DIR}/paddle/fluid/framework/io -I${CMAKE_SOURCE_DIR}/build/third_party/install/zlib/include -I${CMAKE_SOURCE_DIR}/build/third_party/install -I${CMAKE_SOURCE_DIR}/build/third_party/install/gflags/include -I${CMAKE_SOURCE_DIR}/build/third_party/install/glog/include -I${CMAKE_SOURCE_DIR}/build/third_party/boost/src/extern_boost -I${CMAKE_SOURCE_DIR}/build/third_party/eigen3/src/extern_eigen3 -I${CMAKE_SOURCE_DIR}/build/third_party/threadpool/src/extern_threadpool -I${CMAKE_SOURCE_DIR}/build/third_party/dlpack/src/extern_dlpack/include -I${CMAKE_SOURCE_DIR}/build/third_party/install/xxhash/include -I${CMAKE_SOURCE_DIR}/build/third_party/install/warpctc/include -I${CMAKE_SOURCE_DIR}/build/third_party/install/utf8proc/include -I${CMAKE_SOURCE_DIR}/build/third_party/install/openblas/include -I${CMAKE_SOURCE_DIR}/build/third_party/install/protobuf/include -I/usr/include/python3.7m -I/usr/local/lib/python3.7/dist-packages/numpy/core/include -I${CMAKE_SOURCE_DIR}/build/third_party/pybind/src/extern_pybind/include -I${CMAKE_SOURCE_DIR}/build/third_party/install/gtest/include -I${CMAKE_SOURCE_DIR}/build/third_party/install/xpu/include -I${CMAKE_SOURCE_DIR}/build/third_party/install/gloo/include -I${CMAKE_SOURCE_DIR}/build/third_party/install/xbyak/include -I${CMAKE_SOURCE_DIR}/build/third_party/install/xbyak/include/xbyak -I${CMAKE_SOURCE_DIR}/build/third_party/install/cryptopp/include -I${CMAKE_SOURCE_DIR}/build/third_party/pocketfft/src -I${CMAKE_SOURCE_DIR} -I${CMAKE_SOURCE_DIR}/paddle/fluid/platform)
   #set(XPU_CXX_FLAGS -Wno-error=deprecated-declarations -Wno-deprecated-declarations -std=c++14 -m64 -fPIC -fno-omit-frame-pointer -Werror -Wall -Wextra -Wnon-virtual-dtor -Wdelete-non-virtual-dtor -Wno-unused-parameter -Wno-unused-function -Wno-error=literal-suffix -Wno-error=unused-local-typedefs -Wno-error=ignored-attributes -Wno-error=terminate -Wno-error=int-in-bool-context -Wimplicit-fallthrough=0 -Wno-error=maybe-uninitialized -Wno-format-truncation -Wno-error=cast-function-type -Wno-error=parentheses -Wno-error=catch-value -Wno-error=nonnull-compare -Wno-error=address -Wno-ignored-qualifiers -Wno-ignored-attributes -Wno-parentheses -mavx -O3 -DNDEBUG )
   set(XPU_CXX_FLAGS  -Wno-error=pessimizing-move -Wno-error=constant-conversion -Wno-error=c++11-narrowing -Wno-error=shift-count-overflow -Wno-error=unused-local-typedef -Wno-error=deprecated-declarations -Wno-deprecated-declarations -std=c++14 -m64 -fPIC -fno-omit-frame-pointer  -Wall -Wno-inconsistent-missing-override -Wextra -Wnon-virtual-dtor -Wdelete-non-virtual-dtor -Wno-unused-parameter -Wno-unused-function  -Wno-error=unused-local-typedefs -Wno-error=ignored-attributes  -Wno-error=int-in-bool-context -Wno-error=parentheses -Wno-error=address -Wno-ignored-qualifiers -Wno-ignored-attributes -Wno-parentheses -O3 -DNDEBUG )
   #set(XPU_CXX_FLAGS -Wno-error=deprecated-declarations -Wno-deprecated-declarations -std=c++14 -m64 -fPIC -fno-omit-frame-pointer -Werror -Wall -Wextra -Wnon-virtual-dtor -Wdelete-non-virtual-dtor -Wno-unused-parameter -Wno-unused-function -Wno-error=literal-suffix -Wno-error=unused-local-typedefs -Wno-error=ignored-attributes -Wno-error=terminate -Wno-error=int-in-bool-context -Wimplicit-fallthrough=0 -Wno-error=maybe-uninitialized -Wno-format-truncation -Wno-error=cast-function-type -Wno-error=parentheses -Wno-error=catch-value -Wno-error=nonnull-compare -Wno-error=address -Wno-ignored-qualifiers -Wno-ignored-attributes -Wno-parentheses -O3 -DNDEBUG )
@@ -107,7 +107,7 @@ macro(compile_kernel COMPILE_ARGS)
     COMMAND
     # TODO(liuxiandong) xpu->kps -I${XTDK_DIR}/include -std=c++11 
     ${XPU_CLANG} --sysroot=${CXX_DIR}  -O2 -fno-builtin -g -mcpu=xpu2  -fPIC ${XPU_CXX_DEFINES}  ${XPU_CXX_FLAGS}  ${XPU_CXX_INCLUDES} 
-        -I${XTDK_DIR}/include -I.  -o kernel_build/${kernel_name}.bin.o.sec /workspace/paddle/Paddle/paddle/fluid/operators/elementwise/${kernel_name}.xpu
+        -I${XTDK_DIR}/include -I.  -o kernel_build/${kernel_name}.bin.o.sec ${CMAKE_SOURCE_DIR}/paddle/fluid/operators/elementwise/${kernel_name}.xpu
         --xpu-device-only -c -v 
     COMMAND
       ${XTDK_DIR}/bin/xpu2-elfconv kernel_build/${kernel_name}.bin.o.sec  kernel_build/${kernel_name}.bin.o ${XPU_CLANG} --sysroot=${CXX_DIR}
@@ -135,7 +135,7 @@ macro(compile_kernel COMPILE_ARGS)
     COMMAND
     # TODO(liuxiandong) xpu->kps -I${XTDK_DIR}/include -std=c++11 
     ${XPU_CLANG} --sysroot=${CXX_DIR}  -O2 -fno-builtin -g -mcpu=xpu2  -fPIC ${XPU_CXX_DEFINES}  ${XPU_CXX_FLAGS} ${XPU_CXX_INCLUDES} 
-        -I${XTDK_DIR}/include -I.  -o kernel_build/${kernel_name}.host.o /workspace/paddle/Paddle/paddle/fluid/operators/elementwise/${kernel_name}.xpu
+        -I${XTDK_DIR}/include -I.  -o kernel_build/${kernel_name}.host.o ${CMAKE_SOURCE_DIR}/paddle/fluid/operators/elementwise/${kernel_name}.xpu
         --xpu-host-only -c -v 
     # COMMAND
     #   ${CMAKE_COMMAND} -E cmake_depends "Unix Makefiles" ${CMAKE_CURRENT_BINARY_DIR} ${CMAKE_CURRENT_BINARY_DIR}
@@ -263,42 +263,42 @@ macro(xpu_add_library TARGET_NAME)
 endmacro()
 
 # XPU2 PATH
-if(NOT DEFINED ENV{XPU2_PATH})
-    set(XPU2_PATH "/workspace/paddle/xpu-demo/XTDK" CACHE PATH "Path to which XPU2 has been installed")
-    set(XPU_CLANG_PATH ${XPU2_PATH}/bin/clang CACHE PATH "Path to which XPU2 CLANG has been installed")
-else()
-    set(XPU2_PATH $ENV{XPU2_PATH} CACHE PATH "Path to which ROCm has been installed")
-    set(XPU_CLANG_PATH ${XPU2_PATH}/bin/clang CACHE PATH "Path to which XPU2 CLANG has been installed")
-endif()
-set(CMAKE_MODULE_PATH "${XPU2_CLANG_PATH}/cmake" ${CMAKE_MODULE_PATH})
+# if(NOT DEFINED ENV{XPU2_PATH})
+#     set(XPU2_PATH "/workspace/paddle/xpu-demo/XTDK" CACHE PATH "Path to which XPU2 has been installed")
+#     set(XPU_CLANG_PATH ${XPU2_PATH}/bin/clang CACHE PATH "Path to which XPU2 CLANG has been installed")
+# else()
+#     set(XPU2_PATH $ENV{XPU2_PATH} CACHE PATH "Path to which ROCm has been installed")
+#     set(XPU_CLANG_PATH ${XPU2_PATH}/bin/clang CACHE PATH "Path to which XPU2 CLANG has been installed")
+# endif()
+# set(CMAKE_MODULE_PATH "${XPU2_CLANG_PATH}/cmake" ${CMAKE_MODULE_PATH})
 
-# define XPU_CXX_FLAGS
-list(APPEND XPU_CFLAGS -fPIC)
-list(APPEND XPU_CFLAGS --sysroot = /opt/compiler/gcc-8.2)
-list(APPEND XPU_CFLAGS -std=c++11)
-list(APPEND XPU_CFLAGS -O2)
-list(APPEND XPU_CFLAGS -g)
-list(APPEND XPU_CFLAGS -mcpu=xpu2)
-list(APPEND XPU_CFLAGS --target=x86_64-linux-gnu)
-list(APPEND XPU_CFLAGS -v)
-list(APPEND XPU_CFLAGS --dyld-prefix=/opt/compiler/gcc-8.2)
-list(APPEND XPU_CFLAGS -fno-builtin)
-list(APPEND XPU_CFLAGS -Wno-dev)
+# # define XPU_CXX_FLAGS
+# list(APPEND XPU_CFLAGS -fPIC)
+# list(APPEND XPU_CFLAGS --sysroot = /opt/compiler/gcc-8.2)
+# list(APPEND XPU_CFLAGS -std=c++11)
+# list(APPEND XPU_CFLAGS -O2)
+# list(APPEND XPU_CFLAGS -g)
+# list(APPEND XPU_CFLAGS -mcpu=xpu2)
+# list(APPEND XPU_CFLAGS --target=x86_64-linux-gnu)
+# list(APPEND XPU_CFLAGS -v)
+# list(APPEND XPU_CFLAGS --dyld-prefix=/opt/compiler/gcc-8.2)
+# list(APPEND XPU_CFLAGS -fno-builtin)
+# list(APPEND XPU_CFLAGS -Wno-dev)
 
-set(XPU_XPUCC_FLAGS ${XPU_CFLAGS})
+# set(XPU_XPUCC_FLAGS ${XPU_CFLAGS})
 
 # set HIP link libs
-set(xpuapi_library_name xpuapi)
-message(STATUS "XPU API library name: ${xpuapi_library_name}")
-# link in the generic.cmake
-find_library(XPU2_CLANG_API_LIB ${xpuapi_library_name} HINTS ${XPU2_PATH}/shlib)
-message(STATUS "XPU2_CLANG_API_LIB: ${XPU2_CLANG_API_LIB}")
+# set(xpuapi_library_name xpuapi)
+# message(STATUS "XPU API library name: ${xpuapi_library_name}")
+# # link in the generic.cmake
+# find_library(XPU2_CLANG_API_LIB ${xpuapi_library_name} HINTS ${XPU2_PATH}/shlib)
+# message(STATUS "XPU2_CLANG_API_LIB: ${XPU2_CLANG_API_LIB}")
 
-set(xpurt_library_name xpurt)
-message(STATUS "XPU RT library name: ${xpurt_library_name}")
-# link in the generic.cmake
-find_library(XPU2_CLANG_RT_LIB ${xpurt_library_name} HINTS ${XPU2_PATH}/runtime/shlib)
-message(STATUS "XPU2_CLANG_RT_LIB: ${XPU2_CLANG_RT_LIB}")
+# set(xpurt_library_name xpurt)
+# message(STATUS "XPU RT library name: ${xpurt_library_name}")
+# # link in the generic.cmake
+# find_library(XPU2_CLANG_RT_LIB ${xpurt_library_name} HINTS ${XPU2_PATH}/runtime/shlib)
+# message(STATUS "XPU2_CLANG_RT_LIB: ${XPU2_CLANG_RT_LIB}")
 
 # # Ensure that xpu/api.h can be included without dependency errors.
 # file(GENERATE OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/.xpu_headers_dummy.cc CONTENT "")