Merge pull request #1 from PaddlePaddle/develop

Update forked PaddlePaddle

CMakeLists.txt

@@ -28,7 +28,10 @@ include(generic)            # simplify cmake module
 # TODO(Shibo Tao): remove find_package(CUDA) completely.
 find_package(CUDA QUIET)
 option(WITH_GPU         "Compile PaddlePaddle with NVIDIA GPU"          ${CUDA_FOUND})
-
+option(WITH_XPU         "Compile PaddlePaddle with BAIDU KUNLUN"        OFF)
+if (WITH_GPU  AND WITH_XPU)
+    message(FATAL_ERROR "Error when compile GPU and XPU at the same time")
+endif()
 # cmake 3.12, 3.13, 3.14 will append gcc link options to nvcc, and nvcc doesn't recognize them.
 if(WITH_GPU AND (${CMAKE_VERSION} VERSION_GREATER_EQUAL 3.12) AND (${CMAKE_VERSION} VERSION_LESS 3.15))
     message(FATAL_ERROR "cmake ${CMAKE_VERSION} is not supported when WITH_GPU=ON because of bug https://cmake.org/pipermail/cmake/2018-September/068195.html. "

Dockerfile

@@ -11,7 +11,6 @@ RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ub
 ARG WITH_GPU
 ARG WITH_AVX
 
-ENV WOBOQ OFF
 ENV WITH_GPU=${WITH_GPU:-ON}
 ENV WITH_AVX=${WITH_AVX:-ON}
 
@@ -149,21 +148,11 @@ RUN localedef -i en_US -f UTF-8 en_US.UTF-8
 # FIXME: due to temporary ipykernel dependency issue, specify ipykernel jupyter
 # version util jupyter fixes this issue.
 
-# specify sphinx version as 1.5.6 and remove -U option for [pip install -U
-# sphinx-rtd-theme] since -U option will cause sphinx being updated to newest
-# version(1.7.1 for now), which causes building documentation failed.
+
 RUN pip3 --no-cache-dir install -U wheel py-cpuinfo==5.0.0 && \
-    pip3 --no-cache-dir install -U docopt PyYAML sphinx==1.5.6 && \
-    pip3 --no-cache-dir install sphinx-rtd-theme==0.1.9 recommonmark && \
     pip3.6 --no-cache-dir install -U wheel py-cpuinfo==5.0.0 && \
-    pip3.6 --no-cache-dir install -U docopt PyYAML sphinx==1.5.6 && \
-    pip3.6 --no-cache-dir install sphinx-rtd-theme==0.1.9 recommonmark && \
     pip3.7 --no-cache-dir install -U wheel py-cpuinfo==5.0.0 && \
-    pip3.7 --no-cache-dir install -U docopt PyYAML sphinx==1.5.6 && \
-    pip3.7 --no-cache-dir install sphinx-rtd-theme==0.1.9 recommonmark && \
     pip --no-cache-dir install -U wheel py-cpuinfo==5.0.0 && \
-    pip --no-cache-dir install -U docopt PyYAML sphinx==1.5.6 && \
-    pip --no-cache-dir install sphinx-rtd-theme==0.1.9 recommonmark
 
 RUN pip3 --no-cache-dir install 'pre-commit==1.10.4' 'ipython==5.3.0' && \
     pip3 --no-cache-dir install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \
@@ -184,9 +173,9 @@ RUN pip3.6 --no-cache-dir install pylint pytest astroid isort
 RUN pip3.7 --no-cache-dir install pylint pytest astroid isort
 RUN pip --no-cache-dir install pylint pytest astroid isort LinkChecker
 
-RUN pip3 --no-cache-dir install coverage                
-RUN pip3.6 --no-cache-dir install coverage             
-RUN pip3.7 --no-cache-dir install coverage            
+RUN pip3 --no-cache-dir install coverage
+RUN pip3.6 --no-cache-dir install coverage
+RUN pip3.7 --no-cache-dir install coverage
 RUN pip --no-cache-dir install coverage
 
 COPY ./python/requirements.txt /root/
@@ -204,12 +193,6 @@ RUN pip3.7 --no-cache-dir install certifi urllib3[secure]
 RUN pip --no-cache-dir install certifi urllib3[secure]
 
 
-# Install woboq_codebrowser to /woboq
-RUN git clone https://github.com/woboq/woboq_codebrowser /woboq && \
-    (cd /woboq \
-     cmake -DLLVM_CONFIG_EXECUTABLE=/usr/bin/llvm-config-3.8 \
-           -DCMAKE_BUILD_TYPE=Release . \
-     make)
 
 # ar mishandles 4GB files
 # https://sourceware.org/bugzilla/show_bug.cgi?id=14625

README.md

@@ -33,7 +33,7 @@ pip install paddlepaddle
 # Linux GPU cuda10cudnn7
 pip install paddlepaddle-gpu
 # Linux GPU cuda9cudnn7
-pip install paddlepaddle-gpu==1.8.3.post97
+pip install paddlepaddle-gpu==1.8.4.post97
 
 ```
 It is recommended to read [this doc](https://www.paddlepaddle.org.cn/documentation/docs/en/beginners_guide/install/index_en.html) on our website.

README_cn.md

@@ -30,7 +30,7 @@ pip install paddlepaddle
 # Linux GPU cuda10cudnn7
 pip install paddlepaddle-gpu
 # Linux GPU cuda9cudnn7
-pip install paddlepaddle-gpu==1.8.3.post97
+pip install paddlepaddle-gpu==1.8.4.post97
 
 ```
 更多安装信息详见官网 [安装说明](http://www.paddlepaddle.org.cn/documentation/docs/zh/1.8/beginners_guide/install/index_cn.html)

cmake/configure.cmake

@@ -63,6 +63,11 @@ if(WITH_BOX_PS)
     add_definitions(-DPADDLE_WITH_BOX_PS)
 endif()
 
+if(WITH_XPU)
+    message(STATUS "Compile with XPU!")
+    add_definitions(-DPADDLE_WITH_XPU)
+endif()
+
 if(WITH_GPU)
     add_definitions(-DPADDLE_WITH_CUDA)
     add_definitions(-DEIGEN_USE_GPU)

cmake/cuda.cmake

@@ -16,6 +16,7 @@ else()
   set(paddle_known_gpu_archs8 "30 35 50 52 60 61")
   set(paddle_known_gpu_archs9 "30 35 50 52 60 61 70")
   set(paddle_known_gpu_archs10 "30 35 50 52 60 61 70 75")
+  set(paddle_known_gpu_archs11 "35 50 52 60 61 70 75 80")
 endif()
 
 ######################################################################################
@@ -61,6 +62,10 @@ function(detect_installed_gpus out_variable)
   if(NOT CUDA_gpu_detect_output)
     message(STATUS "Automatic GPU detection failed. Building for all known architectures.")
     set(${out_variable} ${paddle_known_gpu_archs} PARENT_SCOPE)
+    #Todo: fix Automatic GPU detection failed on windows
+    if(WIN32)
+      set(${out_variable} "61 75" PARENT_SCOPE)
+    endif()
   else()
     set(${out_variable} ${CUDA_gpu_detect_output} PARENT_SCOPE)
   endif()
@@ -184,6 +189,10 @@ elseif (${CMAKE_CUDA_COMPILER_VERSION} LESS 11.0) # CUDA 10.x
   set(paddle_known_gpu_archs ${paddle_known_gpu_archs10})
   set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -D_MWAITXINTRIN_H_INCLUDED")
   set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -D__STRICT_ANSI__")
+elseif (${CMAKE_CUDA_COMPILER_VERSION} LESS 12.0) # CUDA 11.x
+  set(paddle_known_gpu_archs ${paddle_known_gpu_archs11})
+  set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -D_MWAITXINTRIN_H_INCLUDED")
+  set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -D__STRICT_ANSI__")
 endif()
 
 add_definitions("-DPADDLE_CUDA_BINVER=\"${CUDA_VERSION_MAJOR}${CUDA_VERSION_MINOR}\"")

cmake/external/cub.cmake

@@ -17,7 +17,7 @@ include(ExternalProject)
 set(CUB_PREFIX_DIR ${THIRD_PARTY_PATH}/cub)
 set(CUB_SOURCE_DIR ${THIRD_PARTY_PATH}/cub/src/extern_cub)
 set(CUB_REPOSITORY https://github.com/NVlabs/cub.git)
-set(CUB_TAG        1.9.8)
+set(CUB_TAG        1.8.0)
 
 cache_third_party(extern_cub
     REPOSITORY    ${CUB_REPOSITORY}

cmake/external/lite.cmake

@@ -34,7 +34,7 @@ if (NOT LITE_SOURCE_DIR OR NOT LITE_BINARY_DIR)
   set(LITE_INSTALL_DIR ${THIRD_PARTY_PATH}/install/lite)
 
   if(NOT LITE_GIT_TAG)
-    set(LITE_GIT_TAG 42ab4d559f6659edfc35040fb30fdcec3dc3f8aa)
+    set(LITE_GIT_TAG dfdfa6440c83bf0b415f9f5a9ff84842ce0bb0fa)
   endif()
 
   if(NOT CUDA_ARCH_NAME)

cmake/external/xpu.cmake

@@ -0,0 +1,54 @@
+if (NOT WITH_XPU)
+    return()
+endif()
+
+INCLUDE(ExternalProject)
+SET(XPU_PROJECT                 "extern_xpu")
+SET(XPU_URL    "https://kunlun1.su.bcebos.com/xpu.tar.gz" CACHE STRING "" FORCE)
+SET(XPU_SOURCE_DIR              "${THIRD_PARTY_PATH}/xpu")
+SET(XPU_DOWNLOAD_DIR            "${XPU_SOURCE_DIR}/src/${XPU_PROJECT}")
+SET(XPU_INSTALL_DIR             "${THIRD_PARTY_PATH}/install/xpu")
+SET(XPU_API_INC_DIR             "${THIRD_PARTY_PATH}/install/xpu/api/include")
+SET(XPU_RUNTIME_INC_DIR         "${THIRD_PARTY_PATH}/install/xpu/runtime/include")
+SET(XPU_LIB_DIR                 "${THIRD_PARTY_PATH}/install/xpu/lib")
+
+SET(XPU_API_LIB_NAME            "libxpuapi.so")
+SET(XPU_RT_LIB_NAME             "libxpurt.so")
+SET(XPU_SIM_LIB_NAME            "libxpusim.so")
+SET(XPU_API_LIB                 "${XPU_LIB_DIR}/${XPU_API_LIB_NAME}")
+SET(XPU_RT_LIB                  "${XPU_LIB_DIR}/${XPU_RT_LIB_NAME}")
+SET(XPU_SIM_LIB                 "${XPU_LIB_DIR}/${XPU_SIM_LIB_NAME}")
+
+SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${XPU_INSTALL_DIR}/lib")
+
+INCLUDE_DIRECTORIES(${XPU_API_INC_DIR})
+INCLUDE_DIRECTORIES(${XPU_RUNTIME_INC_DIR})
+
+FILE(WRITE ${XPU_DOWNLOAD_DIR}/CMakeLists.txt
+  "PROJECT(XPU)\n"
+  "cmake_minimum_required(VERSION 3.0)\n"
+  "install(DIRECTORY xpu/api xpu/runtime xpu/lib \n"
+  "        DESTINATION ${XPU_INSTALL_DIR})\n")
+
+ExternalProject_Add(
+    ${XPU_PROJECT}
+    ${EXTERNAL_PROJECT_LOG_ARGS}
+    PREFIX                ${XPU_SOURCE_DIR}
+    DOWNLOAD_DIR          ${XPU_DOWNLOAD_DIR}
+    DOWNLOAD_COMMAND      wget --no-check-certificate ${XPU_URL} -c -q -O xpu.tar.gz
+                          && tar xvf xpu.tar.gz
+    DOWNLOAD_NO_PROGRESS  1
+    UPDATE_COMMAND        ""
+    CMAKE_ARGS            -DCMAKE_INSTALL_PREFIX=${XPU_INSTALL_ROOT}
+    CMAKE_CACHE_ARGS      -DCMAKE_INSTALL_PREFIX:PATH=${XPU_INSTALL_ROOT}
+)
+
+ADD_LIBRARY(shared_xpuapi SHARED IMPORTED GLOBAL)
+set_property(TARGET shared_xpuapi PROPERTY IMPORTED_LOCATION "${XPU_API_LIB}")
+
+# generate a static dummy target to track xpulib dependencies
+# for cc_library(xxx SRCS xxx.c DEPS xpulib)
+generate_dummy_static_lib(LIB_NAME "xpulib" GENERATOR "xpu.cmake")
+
+TARGET_LINK_LIBRARIES(xpulib ${XPU_API_LIB} ${XPU_RT_LIB} ${XPU_SIM_LIB})
+ADD_DEPENDENCIES(xpulib ${XPU_PROJECT})

cmake/generic.cmake

@@ -384,8 +384,12 @@ function(cc_test_run TARGET_NAME)
     set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cpu_deterministic=true)
     set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_init_allocated_mem=true)
     set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cudnn_deterministic=true)
-    # No unit test should exceed 10 minutes.
-    set_tests_properties(${TARGET_NAME} PROPERTIES TIMEOUT 600)
+    # No unit test should exceed 2 minutes.
+    if (APPLE OR WIN32)
+        set_tests_properties(${TARGET_NAME} PROPERTIES TIMEOUT 600)
+    else()
+        set_tests_properties(${TARGET_NAME} PROPERTIES TIMEOUT 120)
+    endif()
   endif()
 endfunction()
 
@@ -742,9 +746,14 @@ function(py_test TARGET_NAME)
                ${PYTHON_EXECUTABLE} -u ${py_test_SRCS} ${py_test_ARGS}
                WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
     endif()
+
+    if (APPLE OR WIN32)
+        set_tests_properties(${TARGET_NAME} PROPERTIES TIMEOUT 600)
+    else()
+        # No unit test should exceed 2 minutes in Linux.
+        set_tests_properties(${TARGET_NAME} PROPERTIES TIMEOUT 120)
+    endif()
 
-    # No unit test should exceed 10 minutes.
-    set_tests_properties(${TARGET_NAME} PROPERTIES TIMEOUT 600)
   endif()
 endfunction()
 

cmake/inference_lib.cmake

@@ -110,10 +110,12 @@ function(copy_part_of_thrid_party TARGET DST)
             SRCS ${GLOG_INCLUDE_DIR} ${GLOG_LIBRARIES}
             DSTS ${dst_dir} ${dst_dir}/lib)
 
+    if (WITH_CRYPTO)
         set(dst_dir "${DST}/third_party/install/cryptopp")
         copy(${TARGET}
-        SRCS ${CRYPTOPP_INCLUDE_DIR} ${CRYPTOPP_LIBRARIES}
-        DSTS ${dst_dir} ${dst_dir}/lib)
+            SRCS ${CRYPTOPP_INCLUDE_DIR} ${CRYPTOPP_LIBRARIES}
+            DSTS ${dst_dir} ${dst_dir}/lib)
+    endif()
 
     set(dst_dir "${DST}/third_party/install/xxhash")
     copy(${TARGET}
@@ -187,7 +189,7 @@ copy(inference_lib_dist
         SRCS  ${CMAKE_BINARY_DIR}/paddle/fluid/framework/framework.pb.h
         DSTS  ${FLUID_INFERENCE_INSTALL_DIR}/paddle/include/internal)
 copy(inference_lib_dist
-        SRCS  ${CMAKE_BINARY_DIR}/../paddle/fluid/framework/io/crypto/cipher.h
+        SRCS  ${PADDLE_SOURCE_DIR}/paddle/fluid/framework/io/crypto/cipher.h
         DSTS  ${FLUID_INFERENCE_INSTALL_DIR}/paddle/include/crypto/)
 include_directories(${CMAKE_BINARY_DIR}/../paddle/fluid/framework/io)
 

cmake/operators.cmake

@@ -8,6 +8,7 @@ function(op_library TARGET)
     set(hip_cu_srcs)
     set(miopen_hip_cc_srcs)
     set(cu_cc_srcs)
+    set(xpu_cc_srcs)
     set(cudnn_cu_cc_srcs)
     set(cudnn_cu_srcs)
     set(CUDNN_FILE)
@@ -60,6 +61,12 @@ function(op_library TARGET)
                 list(APPEND mkldnn_cc_srcs mkldnn/${MKLDNN_FILE}.cc)
             endif()
         endif()
+        if(WITH_XPU)
+            string(REPLACE "_op" "_xpu_op" XPU_FILE "${TARGET}")
+            if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/xpu/${XPU_FILE}.cc)
+                list(APPEND xpu_cc_srcs xpu/${XPU_FILE}.cc)
+            endif()
+        endif()
     else()
         foreach(src ${op_library_SRCS})
             if (${src} MATCHES ".*\\.hip.cu$")
@@ -76,6 +83,8 @@ function(op_library TARGET)
                 list(APPEND mkldnn_cc_srcs ${src})
             elseif(${src} MATCHES ".*\\.cu.cc$")
                 list(APPEND cu_cc_srcs ${src})
+            elseif(WITH_XPU AND ${src} MATCHES ".*_xpu_op.cc$")
+                list(APPEND xpu_cc_srcs ${src})
             elseif(${src} MATCHES ".*\\.cc$")
                 list(APPEND cc_srcs ${src})
             else()
@@ -109,7 +118,7 @@ function(op_library TARGET)
         hip_library(${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 ${op_library_DEPS}
+        cc_library(${TARGET} SRCS ${cc_srcs} ${mkldnn_cc_srcs} ${xpu_cc_srcs} DEPS ${op_library_DEPS}
             ${op_common_deps})
     endif()
 
@@ -150,10 +159,11 @@ 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)
+    list(LENGTH xpu_cc_srcs xpu_cc_srcs_len)
     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)
+        ${hip_cu_srcs_len} EQUAL 0 AND ${miopen_hip_cc_srcs_len} EQUAL 0 AND ${xpu_cc_srcs_len} EQUAL 0)
         file(APPEND ${pybind_file} "USE_CPU_ONLY_OP(${TARGET});\n")
         set(pybind_flag 1)
     endif()
@@ -179,6 +189,9 @@ function(op_library TARGET)
         file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, MIOPEN);\n")
     endif()
 
+    if (WITH_XPU AND ${xpu_cc_srcs_len} GREATER 0)
+        file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, XPU);\n")
+    endif()
     # pybind USE_OP_DEVICE_KERNEL for MKLDNN
     if (WITH_MKLDNN AND ${mkldnn_cc_srcs_len} GREATER 0)
       # Append first implemented MKLDNN activation operator
@@ -228,6 +241,7 @@ function(register_operators)
 
     file(GLOB OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*_op.cc")
     string(REPLACE "_mkldnn" "" OPS "${OPS}")
+    string(REPLACE "_xpu" "" OPS "${OPS}")
     string(REPLACE ".cc" "" OPS "${OPS}")
     list(REMOVE_DUPLICATES OPS)
     list(LENGTH register_operators_DEPS register_operators_DEPS_len)

cmake/third_party.cmake

@@ -250,6 +250,11 @@ if(WITH_GPU)
     file_download_and_uncompress(${CUDAERROR_URL} "cudaerror") # download file cudaErrorMessage
 endif(WITH_GPU)
 
+if(WITH_XPU)
+    include(external/xpu)          # download, build, install xpu
+    list(APPEND third_party_deps extern_xpu)
+endif(WITH_XPU)
+
 if(WITH_PSLIB)
     include(external/pslib)          # download, build, install pslib
     list(APPEND third_party_deps extern_pslib)

go/paddle/config.go

@@ -154,10 +154,17 @@ func (config *AnalysisConfig) EnableMkldnnQuantizer() {
 	C.PD_EnableMkldnnQuantizer(config.c)
 }
 
+func (config *AnalysisConfig) EnableMkldnnBfloat16() {
+	C.PD_EnableMkldnnBfloat16(config.c)
+}
+
 func (config *AnalysisConfig) MkldnnQuantizerEnabled() bool {
 	return ConvertCBooleanToGo(C.PD_MkldnnQuantizerEnabled(config.c))
 }
 
+func (config *AnalysisConfig) MkldnnBfloat16Enabled() bool {
+	return ConvertCBooleanToGo(C.PD_MkldnnBfloat16Enabled(config.c))
+}
 // SetModelBuffer
 // ModelFromMemory
 

paddle/fluid/framework/CMakeLists.txt

@@ -119,9 +119,13 @@ cc_test(data_layout_transform_test SRCS data_layout_transform_test.cc DEPS data_
 cc_library(data_transform SRCS data_transform.cc DEPS math_function tensor
         framework_proto selected_rows data_device_transform data_type_transform data_layout_transform)
 
-cc_library(attribute SRCS attribute.cc DEPS framework_proto boost)
+cc_library(attribute SRCS attribute.cc DEPS framework_proto boost enforce)
 cc_test(program_desc_test SRCS program_desc_test.cc DEPS proto_desc
 device_context)
+
+cc_library(op_version_registry SRCS op_version_registry.cc DEPS framework_proto boost)
+cc_test(op_version_registry_test SRCS op_version_registry_test.cc DEPS op_version_registry)
+
 cc_library(op_proto_maker SRCS op_proto_maker.cc DEPS framework_proto attribute glog)
 cc_test(op_proto_maker_test SRCS op_proto_maker_test.cc DEPS op_proto_maker)
 cc_library(no_need_buffer_vars_inference SRCS no_need_buffer_vars_inference.cc DEPS attribute device_context)
@@ -268,6 +272,7 @@ cc_test(op_compatible_info_test SRCS op_compatible_info_test.cc DEPS op_compatib
 
 cc_library(save_load_util SRCS save_load_util DEPS tensor scope layer)
 cc_test(save_load_util_test SRCS save_load_util_test.cc DEPS save_load_util tensor scope layer)
+cc_library(generator SRCS generator.cc DEPS enforce place)
 
 # Get the current working branch
 execute_process(

paddle/fluid/framework/data_layout_transform.cc

@@ -116,6 +116,8 @@ void* GetDataFromTensor(const Tensor& tensor, mkldnn::memory::data_type type) {
       return platform::to_void_cast(tensor.data<unsigned char>());
     case mkldnn::memory::data_type::s32:
       return platform::to_void_cast(tensor.data<int32_t>());
+    case mkldnn::memory::data_type::bf16:
+      return platform::to_void_cast(tensor.data<paddle::platform::bfloat16>());
     default:
       PADDLE_THROW(
           platform::errors::InvalidArgument("Wrong mkldnn type provided."));