merge all phi kernel lib to several big static lib, reduce link command (#42185)

* merge all phi lib to several big static lib

* merge all phi lib to several big static lib

CMakeLists.txt

@@ -20,6 +20,9 @@ else(APPLE AND WITH_ARM)
     cmake_minimum_required(VERSION 3.15)
     cmake_policy(VERSION 3.10)
 endif(APPLE AND WITH_ARM)
+# use to get_property location of static lib
+# https://cmake.org/cmake/help/v3.0/policy/CMP0026.html?highlight=cmp0026
+cmake_policy(SET CMP0026 OLD)
 set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} "${CMAKE_CURRENT_SOURCE_DIR}/cmake")
 set(PADDLE_SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR})
 set(PADDLE_BINARY_DIR ${CMAKE_CURRENT_BINARY_DIR})

cmake/generic.cmake

@@ -176,6 +176,36 @@ function(create_static_lib TARGET_NAME)
   endif()
 endfunction()
 
+function(create_dummy_static_lib TARGET_NAME)
+  set(options "")
+  set(oneValueArgs "")
+  set(multiValueArgs LIBS DEPS LIMIT)
+  cmake_parse_arguments(merge "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+
+  list(REMOVE_DUPLICATES merge_LIBS)
+  set(index 1)
+  set(offset 1)
+  # the dummy target would be consisted of limit size libraries
+  set(limit ${merge_LIMIT})
+  list(LENGTH merge_LIBS libs_len)
+  foreach(lib ${merge_LIBS})
+    list(APPEND merge_list ${lib})
+    list(LENGTH merge_list listlen)
+    if ((${listlen} GREATER ${limit}) OR (${offset} EQUAL ${libs_len}))
+      message("Merge and generate static library: ${TARGET_NAME}_static_${index}")
+      merge_static_libs(${TARGET_NAME}_static_${index} ${merge_list})
+      if(merge_DEPS)
+        target_link_libraries(${TARGET_NAME}_static_${index} ${merge_DEPS})
+      endif()
+      set(merge_list)
+      list(APPEND ${TARGET_NAME}_list ${TARGET_NAME}_static_${index})
+      MATH(EXPR index "${index}+1")
+    endif()
+    MATH(EXPR offset "${offset}+1")
+  endforeach()
+  cc_library(${TARGET_NAME} DEPS ${${TARGET_NAME}_list})
+endfunction()
+
 function(merge_static_libs TARGET_NAME)
   set(libs ${ARGN})
   list(REMOVE_DUPLICATES libs)
@@ -193,7 +223,6 @@ function(merge_static_libs TARGET_NAME)
   # also help to track dependencies.
   set(target_SRCS ${CMAKE_CURRENT_BINARY_DIR}/${TARGET_NAME}_dummy.c)
 
-  if(APPLE) # Use OSX's libtool to merge archives
   # Make the generated dummy source file depended on all static input
   # libs. If input lib changes,the source file is touched
   # which causes the desired effect (relink).
@@ -201,84 +230,54 @@ function(merge_static_libs TARGET_NAME)
     COMMAND ${CMAKE_COMMAND} -E touch ${target_SRCS}
     DEPENDS ${libs})
   
-    # Generate dummy static lib
+    # Generate dummy staic lib
   generate_dummy_static_lib(LIB_NAME ${TARGET_NAME} FILE_PATH ${target_SRCS} GENERATOR "generic.cmake:merge_static_libs")
-
   target_link_libraries(${TARGET_NAME} ${libs_deps})
 
+  # OSX: use 'libtool' to merge archives
+  if(APPLE)
     foreach(lib ${libs})
       # Get the file names of the libraries to be merged
       set(libfiles ${libfiles} $<TARGET_FILE:${lib}>)
     endforeach()
     add_custom_command(TARGET ${TARGET_NAME} POST_BUILD
+      COMMENT "Merge and generate static lib: lib${TARGET_NAME}.a"
       COMMAND rm "${CMAKE_CURRENT_BINARY_DIR}/lib${TARGET_NAME}.a"
       COMMAND /usr/bin/libtool -static -o "${CMAKE_CURRENT_BINARY_DIR}/lib${TARGET_NAME}.a" ${libfiles}
       )
-  endif(APPLE)
-  if(LINUX) # general UNIX: use "ar" to extract objects and re-add to a common lib
-    set(target_DIR ${CMAKE_CURRENT_BINARY_DIR}/${TARGET_NAME}.dir)
-
-    foreach(lib ${libs})
-      set(objlistfile ${target_DIR}/${lib}.objlist) # list of objects in the input library
-      set(objdir ${target_DIR}/${lib}.objdir)
-
-      add_custom_command(OUTPUT ${objdir}
-        COMMAND ${CMAKE_COMMAND} -E make_directory ${objdir}
-        DEPENDS ${lib})
+  endif()
 
-      add_custom_command(OUTPUT ${objlistfile}
-        COMMAND ${CMAKE_AR} -x "$<TARGET_FILE:${lib}>"
-        COMMAND ${CMAKE_AR} -t "$<TARGET_FILE:${lib}>" > ${objlistfile}
-        DEPENDS ${lib} ${objdir}
-        WORKING_DIRECTORY ${objdir})
+  # LINUX: use "ar" to extract objects and re-add to a common lib
+  if(LINUX)
+    set(mri_file ${CMAKE_CURRENT_BINARY_DIR}/${TARGET_NAME}.mri CACHE INTERNAL "phi_static.mri file")
+    get_property(ABS_MERGE_LIB_PATH TARGET ${TARGET_NAME} PROPERTY LOCATION)
+    file(WRITE ${mri_file} "create ${ABS_MERGE_LIB_PATH}\n")
 
-      list(APPEND target_OBJS "${objlistfile}")
+    foreach(lib ${libs})
+      get_property(ABS_LIB_PATH TARGET ${lib} PROPERTY LOCATION)
+      file(APPEND ${mri_file} "addlib ${ABS_LIB_PATH}\n")
     endforeach()
-
-    # Make the generated dummy source file depended on all static input
-    # libs. If input lib changes,the source file is touched
-    # which causes the desired effect (relink).
-    add_custom_command(OUTPUT ${target_SRCS}
-      COMMAND ${CMAKE_COMMAND} -E touch ${target_SRCS}
-      DEPENDS ${libs} ${target_OBJS})
-
-    # Generate dummy staic lib
-    generate_dummy_static_lib(LIB_NAME ${TARGET_NAME} FILE_PATH ${target_SRCS} GENERATOR "generic.cmake:merge_static_libs")
-
-    target_link_libraries(${TARGET_NAME} ${libs_deps})
-
-    # Get the file name of the generated library
-    set(target_LIBNAME "$<TARGET_FILE:${TARGET_NAME}>")
+    file(APPEND ${mri_file} "save\nend\n")
 
     add_custom_command(TARGET ${TARGET_NAME} POST_BUILD
-        COMMAND ${CMAKE_AR} crs ${target_LIBNAME} `find ${target_DIR} -name '*.o'`
-        COMMAND ${CMAKE_RANLIB} ${target_LIBNAME}
-        WORKING_DIRECTORY ${target_DIR})
-  endif(LINUX)
-  if(WIN32) # windows do not support gcc/nvcc combined compiling. Use msvc lib.exe to merge libs.
-    # Make the generated dummy source file depended on all static input
-    # libs. If input lib changes,the source file is touched
-    # which causes the desired effect (relink).
-    add_custom_command(OUTPUT ${target_SRCS}
-      COMMAND ${CMAKE_COMMAND} -E touch ${target_SRCS}
-      DEPENDS ${libs})
-    # Generate dummy staic lib
-    generate_dummy_static_lib(LIB_NAME ${TARGET_NAME} FILE_PATH ${target_SRCS} GENERATOR "generic.cmake:merge_static_libs")
-
-    target_link_libraries(${TARGET_NAME} ${libs_deps})
+        COMMENT "Merge and generate static lib: lib${TARGET_NAME}.a"
+        COMMAND ${CMAKE_AR} -M < ${mri_file}
+        COMMAND ${CMAKE_RANLIB} "$<TARGET_FILE:${TARGET_NAME}>")
+  endif()
 
+  # Windows do not support gcc/nvcc combined compiling. Use msvc 'lib.exe' to merge libs.
+  if(WIN32)
     foreach(lib ${libs})
-      # Get the file names of the libraries to be merged
       set(libfiles ${libfiles} $<TARGET_FILE:${lib}>)
     endforeach()
-    # msvc will put libarary in directory of "/Release/xxxlib" by default
-    #       COMMAND cmake -E remove "${CMAKE_CURRENT_BINARY_DIR}/${CMAKE_BUILD_TYPE}/${TARGET_NAME}.lib"
+    # msvc compiler will put libarary in directory of "/Release/xxxlib" by default
     add_custom_command(TARGET ${TARGET_NAME} POST_BUILD
+      COMMENT "Merge and generate static lib: lib${TARGET_NAME}.lib"
       COMMAND cmake -E make_directory $<TARGET_FILE_DIR:${TARGET_NAME}>
       COMMAND lib /OUT:$<TARGET_FILE:${TARGET_NAME}> ${libfiles}
       )
-  endif(WIN32)
-endfunction(merge_static_libs)
+  endif()
+endfunction()
 
 function(check_coverage_opt TARGET_NAME SRCS)
   if(WITH_COVERAGE AND WITH_INCREMENTAL_COVERAGE)
@@ -1076,4 +1075,3 @@ function(math_library TARGET)
         cc_library(${TARGET} SRCS ${cc_srcs} DEPS ${math_library_DEPS} ${math_common_deps})
     endif()
 endfunction()
-

paddle/fluid/distributed/collective/CMakeLists.txt

-cc_library(processgroup SRCS ProcessGroup.cc DEPS phi phi_api eager_api)
-cc_library(eager_reducer SRCS reducer.cc DEPS eager_api processgroup phi phi_api string_helper)
+cc_library(processgroup SRCS ProcessGroup.cc DEPS phi_api eager_api)
+cc_library(eager_reducer SRCS reducer.cc DEPS eager_api processgroup phi_api string_helper)
 
 if (WITH_DISTRIBUTE)
-  cc_library(processgroup_gloo SRCS ProcessGroupGloo.cc DEPS phi phi_api eager_api gloo_wrapper)
+  cc_library(processgroup_gloo SRCS ProcessGroupGloo.cc DEPS phi_api eager_api gloo_wrapper)
 endif()
 
 if(WITH_NCCL)
-  cc_library(processgroup_nccl SRCS ProcessGroupNCCL.cc NCCLTools.cc Common.cc DEPS place cuda_stream enforce collective_helper device_context phi phi_api eager_api)
+  cc_library(processgroup_nccl SRCS ProcessGroupNCCL.cc NCCLTools.cc Common.cc DEPS place cuda_stream enforce collective_helper device_context phi_api eager_api)
   if (WITH_DISTRIBUTE AND WITH_PSCORE)
-    cc_library(processgroup_heter SRCS ProcessGroupHeter.cc NCCLTools.cc Common.cc DEPS place cuda_stream enforce collective_helper device_context phi phi_api eager_api)
+    cc_library(processgroup_heter SRCS ProcessGroupHeter.cc NCCLTools.cc Common.cc DEPS place cuda_stream enforce collective_helper device_context phi_api eager_api)
   endif()
 endif()
 
 if(WITH_ASCEND_CL)
-  cc_library(processgroup_hccl SRCS ProcessGroupHCCL.cc HCCLTools.cc Common.cc DEPS place npu_stream enforce collective_helper device_context phi phi_api eager_api)
+  cc_library(processgroup_hccl SRCS ProcessGroupHCCL.cc HCCLTools.cc Common.cc DEPS place npu_stream enforce collective_helper device_context phi_api eager_api)
   if (WITH_DISTRIBUTE AND WITH_PSCORE)
-    cc_library(processgroup_heter SRCS ProcessGroupHeter.cc HCCLTools.cc Common.cc DEPS place npu_stream enforce collective_helper device_context phi phi_api eager_api)
+    cc_library(processgroup_heter SRCS ProcessGroupHeter.cc HCCLTools.cc Common.cc DEPS place npu_stream enforce collective_helper device_context phi_api eager_api)
   endif()
 endif()

paddle/fluid/eager/accumulation/CMakeLists.txt

-cc_library(accumulation_node SRCS accumulation_node.cc DEPS gradient_accumulator phi phi_api grad_node_info)
+cc_library(accumulation_node SRCS accumulation_node.cc DEPS gradient_accumulator phi_api grad_node_info)

paddle/fluid/eager/api/utils/CMakeLists.txt

-cc_library(tensor_utils SRCS tensor_utils.cc DEPS phi phi_api autograd_meta grad_node_info accumulation_node)
+cc_library(tensor_utils SRCS tensor_utils.cc DEPS phi_api autograd_meta grad_node_info accumulation_node)
 cc_library(hook_utils SRCS hook_utils.cc DEPS phi tensor_utils autograd_meta grad_node_info utils accumulation_node)
 cc_library(global_utils SRCS global_utils.cc DEPS place tracer)

paddle/fluid/eager/pylayer/CMakeLists.txt

-cc_library(py_layer_node SRCS py_layer_node.cc DEPS pybind phi phi_api grad_node_info)
+cc_library(py_layer_node SRCS py_layer_node.cc DEPS pybind phi_api grad_node_info)

paddle/fluid/framework/CMakeLists.txt

@@ -206,11 +206,11 @@ ENDIF()
 IF(WITH_XPU)
 cc_library(operator SRCS operator.cc DEPS xpu_op_list op_info device_context tensor scope glog trainer_desc_proto data_feed_proto
     shape_inference data_transform lod_tensor profiler transfer_scope_cache op_kernel_type op_call_stack unused_var_check nan_inf_utils
-    phi phi_utils kernel_factory infershape_utils op_utils)
+    phi_utils kernel_factory infershape_utils op_utils)
 ELSE()
 cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog trainer_desc_proto data_feed_proto
     shape_inference data_transform lod_tensor profiler transfer_scope_cache op_kernel_type op_call_stack unused_var_check nan_inf_utils
-    phi phi_utils kernel_factory infershape_utils op_utils)
+    phi_utils kernel_factory infershape_utils op_utils)
 ENDIF()
 
 cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry device_context)
@@ -418,7 +418,7 @@ cc_library(save_load_util SRCS save_load_util.cc 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)
 
-cc_library(infershape_utils SRCS infershape_utils.cc DEPS lod_tensor selected_rows_utils attribute place phi var_type_traits phi phi_api_utils op_info shape_inference)
+cc_library(infershape_utils SRCS infershape_utils.cc DEPS lod_tensor selected_rows_utils attribute place var_type_traits phi phi_api_utils op_info shape_inference)
 cc_test(infershape_utils_test SRCS infershape_utils_test.cc DEPS infershape_utils infermeta_utils meta_tensor)
 
 # Get the current working branch

paddle/fluid/imperative/CMakeLists.txt

 cc_library(imperative_flag SRCS flags.cc DEPS gflags flags)
 cc_library(var_helper SRCS var_helper.cc DEPS tensor phi_api)
 IF(WITH_XPU)
-cc_library(prepared_operator SRCS prepared_operator.cc DEPS xpu_op_list proto_desc operator device_context lod_tensor selected_rows_utils var_type_traits op_kernel_type data_transform nan_inf_utils phi_api phi phi_utils var_helper)
+cc_library(prepared_operator SRCS prepared_operator.cc DEPS xpu_op_list proto_desc operator device_context lod_tensor selected_rows_utils var_type_traits op_kernel_type data_transform nan_inf_utils phi_api phi_utils var_helper)
 ELSE()
-cc_library(prepared_operator SRCS prepared_operator.cc DEPS proto_desc operator device_context lod_tensor selected_rows_utils var_type_traits op_kernel_type data_transform nan_inf_utils phi_api phi phi_utils var_helper)
+cc_library(prepared_operator SRCS prepared_operator.cc DEPS proto_desc operator device_context lod_tensor selected_rows_utils var_type_traits op_kernel_type data_transform nan_inf_utils phi_api phi_utils var_helper)
 ENDIF()
 cc_library(layer SRCS layer.cc DEPS prepared_operator math_function imperative_flag variable_helper op_registry var_helper phi_api)
 add_subdirectory(jit)

paddle/fluid/inference/CMakeLists.txt

@@ -36,7 +36,7 @@ endif()
 # fluid_modules exclude API-interface of inference/api and inference/capi_exp
 get_property(fluid_modules GLOBAL PROPERTY FLUID_MODULES)
 get_property(phi_modules GLOBAL PROPERTY PHI_MODULES)
-set(utils_modules stringpiece pretty_log string_helper)
+set(utils_modules stringpiece pretty_log string_helper benchmark)
 
 add_subdirectory(api)
 
@@ -50,9 +50,9 @@ if(WITH_ONNXRUNTIME)
   set(STATIC_INFERENCE_API ${STATIC_INFERENCE_API} onnxruntime_predictor)
 endif()
 
-#TODO(wilber, T8T9): Do we still need to support windows gpu static library?
+#windows GPU static library over the limit, so not create_static_lib, and cc_library is dummy
 if(WIN32 AND WITH_GPU)
-  cc_library(paddle_inference DEPS ${fluid_modules} ${phi_modules} ${STATIC_INFERENCE_API} ${utils_modules})
+  cc_library(paddle_inference DEPS ${fluid_modules} phi ${STATIC_INFERENCE_API} ${utils_modules})
 else()
   create_static_lib(paddle_inference ${fluid_modules} ${phi_modules} ${STATIC_INFERENCE_API} ${utils_modules})
 endif()
@@ -84,7 +84,7 @@ set(SHARED_INFERENCE_SRCS
     ${PADDLE_CUSTOM_OP_SRCS})
 
 # shared inference library deps
-set(SHARED_INFERENCE_DEPS ${fluid_modules} ${phi_modules} analysis_predictor)
+set(SHARED_INFERENCE_DEPS ${fluid_modules} phi analysis_predictor ${utils_modules})
 
 if (WITH_CRYPTO) 
     set(SHARED_INFERENCE_DEPS ${SHARED_INFERENCE_DEPS} paddle_crypto)

paddle/fluid/inference/tensorrt/CMakeLists.txt

 # Compiling with WITH_PYTHON=ON and WITH_TENSORRT=ON failed on windows. Temporarily add paddle_inference_api dependency to solve the problem
 if(WIN32)
-nv_library(tensorrt_engine SRCS engine.cc trt_int8_calibrator.cc DEPS ${GLOB_OPERATOR_DEPS} framework_proto device_context boost paddle_inference_api)
+    nv_library(tensorrt_engine SRCS engine.cc trt_int8_calibrator.cc DEPS ${GLOB_OPERATOR_DEPS} framework_proto device_context boost paddle_inference_api)
 else()
-nv_library(tensorrt_engine SRCS engine.cc trt_int8_calibrator.cc DEPS ${GLOB_OPERATOR_DEPS} framework_proto device_context boost)
+    nv_library(tensorrt_engine SRCS engine.cc trt_int8_calibrator.cc DEPS ${GLOB_OPERATOR_DEPS} framework_proto device_context boost)
 endif()
 nv_library(tensorrt_op_teller SRCS op_teller.cc DEPS framework_proto device_context boost)
 nv_test(test_tensorrt SRCS test_tensorrt.cc DEPS dynload_cuda device_context dynamic_loader)

paddle/phi/CMakeLists.txt

@@ -27,7 +27,7 @@ set(PHI_DEPS convert_utils dense_tensor phi_context kernel_factory kernel_contex
 get_property(phi_kernels GLOBAL PROPERTY PHI_KERNELS)
 set(PHI_DEPS ${PHI_DEPS} ${phi_kernels})
 
-cc_library(phi DEPS ${PHI_DEPS})
+create_dummy_static_lib(phi LIBS ${PHI_DEPS} LIMIT 100)
 
 set(phi_extension_header_file ${CMAKE_CURRENT_SOURCE_DIR}/extension.h CACHE INTERNAL "phi/extension.h file")
 file(WRITE ${phi_extension_header_file} "// Header file generated by paddle/phi/CMakeLists.txt for external users,\n// DO NOT edit or include it within paddle.\n\n#pragma once\n\n")

paddle/phi/kernels/CMakeLists.txt

@@ -36,7 +36,7 @@ set(MANUAL_BUILD_KERNELS ${AUTOTUNE_KERNELS} cross_entropy_kernel adam_kernel ad
     matrix_power_kernel matrix_power_grad_kernel maxout_kernel maxout_grad_kernel pool_kernel
     put_along_axis_kernel put_along_axis_grad_kernel segment_pool_kernel segment_pool_grad_kernel
     softmax_kernel softmax_grad_kernel take_along_axis_kernel take_along_axis_grad_kernel
-    triangular_solve_grad_kernel determinant_grad_kernel reduce_sum_kernel rnn_kernel rnn_grad_kernel warpctc_kernel warpctc_grad_kernel)
+    triangular_solve_grad_kernel determinant_grad_kernel reduce_sum_kernel reduce_mean_kernel rnn_kernel rnn_grad_kernel warpctc_kernel warpctc_grad_kernel)
 foreach(src ${AUTOTUNE_KERNELS})
   kernel_library(${src} DEPS ${COMMON_KERNEL_DEPS} switch_autotune)
 endforeach()
@@ -52,6 +52,7 @@ kernel_library(hierarchical_sigmoid_grad_kernel DEPS ${COMMON_KERNEL_DEPS} matri
 kernel_library(gumbel_softmax_kernel DEPS ${COMMON_KERNEL_DEPS} softmax)
 kernel_library(gumbel_softmax_grad_kernel DEPS ${COMMON_KERNEL_DEPS} softmax)
 kernel_library(reduce_sum_kernel DEPS ${COMMON_KERNEL_DEPS} cast_kernel)
+kernel_library(reduce_mean_kernel DEPS ${COMMON_KERNEL_DEPS} cast_kernel)
 kernel_library(matrix_power_kernel DEPS ${COMMON_KERNEL_DEPS} matrix_inverse)
 kernel_library(matrix_power_grad_kernel DEPS ${COMMON_KERNEL_DEPS} matrix_inverse)
 kernel_library(maxout_kernel DEPS ${COMMON_KERNEL_DEPS} maxouting)