[HACKATHON 6th] Refactor phi module structure

cmake/generic.cmake

@@ -603,7 +603,7 @@ function(paddle_test_build TARGET_NAME)
       add_dependencies(${TARGET_NAME} onednn)
     endif()
     if(WITH_SHARED_PHI)
-      target_link_libraries(${TARGET_NAME} $<TARGET_LINKER_FILE:phi>)
+      target_link_libraries(${TARGET_NAME} phi)
       add_dependencies(${TARGET_NAME} phi)
     endif()
     if(WITH_SHARED_IR)

cmake/inference_lib.cmake

@@ -280,10 +280,10 @@ else()
     DSTS ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include
          ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/lib)
   if(WITH_SHARED_PHI)
-    set(paddle_phi_lib ${PADDLE_BINARY_DIR}/paddle/phi/libphi.*)
+    set(paddle_phi_libs ${PADDLE_BINARY_DIR}/paddle/phi/libphi*)
     copy(
       inference_lib_dist
-      SRCS ${paddle_phi_lib}
+      SRCS ${paddle_phi_libs}
       DSTS ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/lib)
   endif()
 endif()

paddle/fluid/eager/CMakeLists.txt

@@ -1,5 +1,6 @@
 set(eager_deps
     phi
+    phi_core
     common
     hook_utils
     utils
@@ -13,6 +14,10 @@ set(eager_deps
     grad_tensor_holder
     custom_operator_node)
 
+if(WITH_GPU OR WITH_ROCM)
+  set(eager_deps ${eager_deps} phi_gpu)
+endif()
+
 if(NOT (NOT WITH_PYTHON AND ON_INFER))
   set(eager_deps ${eager_deps} accumulation_node prim_utils)
 endif()

paddle/fluid/eager/auto_code_generator/CMakeLists.txt

@@ -58,6 +58,16 @@ if(WIN32)
       OUTPUT ${eager_generator_path}/phi.dll
       COMMAND ${CMAKE_COMMAND} -E copy ${PHI_LIB} ${eager_generator_path}
       DEPENDS phi)
+    add_custom_command(
+      OUTPUT ${eager_generator_path}/phi_core.dll
+      COMMAND ${CMAKE_COMMAND} -E copy ${PHI_CORE_LIB} ${eager_generator_path}
+      DEPENDS phi)
+    if(WITH_GPU OR WITH_ROCM)
+      add_custom_command(
+        OUTPUT ${eager_generator_path}/phi_gpu.dll
+        COMMAND ${CMAKE_COMMAND} -E copy ${PHI_GPU_LIB} ${eager_generator_path}
+        DEPENDS phi)
+    endif()
     list(APPEND EAGER_CODEGEN_DEPS ${eager_generator_path}/phi.dll)
   endif()
 

paddle/fluid/inference/CMakeLists.txt

@@ -96,8 +96,11 @@ list(REMOVE_ITEM fluid_modules cinn_op_dialect)
 # shared library to prune library size.
 # list(REMOVE_ITEM fluid_modules ${not_infer_modules})
 
-set(SHARED_INFERENCE_DEPS phi common ${fluid_modules} analysis_predictor
-                          ${utils_modules})
+set(SHARED_INFERENCE_DEPS phi phi_core common ${fluid_modules}
+                          analysis_predictor ${utils_modules})
+if(WITH_GPU OR WITH_ROCM)
+  list(APPEND SHARED_INFERENCE_DEPS phi_gpu)
+endif()
 if(NOT WIN32)
   list(APPEND SHARED_INFERENCE_DEPS ${ir_targets})
 endif()

paddle/fluid/inference/api/demo_ci/CMakeLists.txt

@@ -211,7 +211,13 @@ if(NOT WIN32)
   if(WITH_SHARED_PHI)
     set(DEPS
         ${DEPS} ${PADDLE_LIB}/paddle/lib/libphi${CMAKE_SHARED_LIBRARY_SUFFIX}
+        ${PADDLE_LIB}/paddle/lib/libphi_core${CMAKE_SHARED_LIBRARY_SUFFIX}
         ${PADDLE_LIB}/paddle/lib/libcommon${CMAKE_SHARED_LIBRARY_SUFFIX})
+    if(WITH_GPU OR WITH_ROCM)
+      set(DEPS
+          ${DEPS}
+          ${PADDLE_LIB}/paddle/lib/libphi_gpu${CMAKE_SHARED_LIBRARY_SUFFIX})
+    endif()
   endif()
 else()
   set(DEPS

paddle/fluid/pybind/CMakeLists.txt

@@ -370,6 +370,16 @@ if(WITH_PYTHON)
         OUTPUT ${op_impl_path}/phi.dll
         COMMAND ${CMAKE_COMMAND} -E copy ${PHI_LIB} ${op_impl_path}
         DEPENDS phi)
+      add_custom_command(
+        OUTPUT ${op_impl_path}/phi_core.dll
+        COMMAND ${CMAKE_COMMAND} -E copy ${PHI_CORE_LIB} ${op_impl_path}
+        DEPENDS phi)
+      if(WITH_GPU OR WITH_ROCM)
+        add_custom_command(
+          OUTPUT ${op_impl_path}/phi_gpu.dll
+          COMMAND ${CMAKE_COMMAND} -E copy ${PHI_GPU_LIB} ${op_impl_path}
+          DEPENDS phi)
+      endif()
       list(APPEND EAGER_OP_IMPL_DEPS ${op_impl_path}/phi.dll)
     endif()
 

paddle/phi/CMakeLists.txt

@@ -103,7 +103,7 @@ if(WITH_DGC)
   list(APPEND PHI_DEPS dgc)
 endif()
 
-set(PHI_SRCS
+set(PHI_CORE_SRCS
     ${common_srcs}
     ${api_srcs}
     ${core_srcs}
@@ -112,6 +112,8 @@ set(PHI_SRCS
     ${infermeta_srcs}
     ${capi_srcs})
 
+set(PHI_GPU_SRCS ${kernels_gpu_srcs})
+
 if(WITH_SHARED_PHI)
   set(PHI_BUILD_TYPE
       SHARED
@@ -140,37 +142,67 @@ if(WITH_GPU)
     PROPERTIES COMPILE_FLAGS
                "-DCUDA_REAL_ARCHS=\"${NVCC_FLAGS_EXTRA_real_archs}\"")
   nv_library(
-    phi ${PHI_BUILD_TYPE}
-    SRCS ${PHI_SRCS}
+    phi_core ${PHI_BUILD_TYPE}
+    SRCS ${PHI_CORE_SRCS}
     DEPS ${PHI_DEPS})
-
+  nv_library(
+    phi_gpu ${PHI_BUILD_TYPE}
+    SRCS ${PHI_GPU_SRCS}
+    DEPS ${PHI_DEPS} phi_core)
 elseif(WITH_ROCM)
   hip_library(
-    phi ${PHI_BUILD_TYPE}
-    SRCS ${PHI_SRCS}
+    phi_core ${PHI_BUILD_TYPE}
+    SRCS ${PHI_CORE_SRCS}
     DEPS ${PHI_DEPS})
-
+  hip_library(
+    phi_gpu ${PHI_BUILD_TYPE}
+    SRCS ${PHI_GPU_SRCS}
+    DEPS ${PHI_DEPS} phi_core)
 elseif(WITH_XPU_KP)
   xpu_library(
-    phi ${PHI_BUILD_TYPE}
-    SRCS ${PHI_SRCS}
+    phi_core ${PHI_BUILD_TYPE}
+    SRCS ${PHI_CORE_SRCS}
     DEPS ${PHI_DEPS})
 else()
   cc_library(
-    phi ${PHI_BUILD_TYPE}
-    SRCS ${PHI_SRCS}
+    phi_core ${PHI_BUILD_TYPE}
+    SRCS ${PHI_CORE_SRCS}
     DEPS ${PHI_DEPS})
 endif()
 
-target_compile_definitions(phi PUBLIC PHI_INNER)
+set(PHI_DUMMY_FILE ${CMAKE_CURRENT_BINARY_DIR}/phi_dummy.cpp)
+if(MSVC)
+  set(PHI_DUMMY_FILE_CONTENT
+      "__declspec(dllexport) int phi_dummy_placeholder_function(){ return 0\\; }"
+  )
+else()
+  set(PHI_DUMMY_FILE_CONTENT "")
+endif()
+file(WRITE ${PHI_DUMMY_FILE} ${PHI_DUMMY_FILE_CONTENT})
+
+add_library(phi ${PHI_BUILD_TYPE} ${PHI_DUMMY_FILE})
+target_link_libraries(phi phi_core)
+if(WITH_GPU OR WITH_ROCM)
+  target_link_libraries(phi phi_gpu)
+endif()
+
+# Note(silverling): some functions in phi_core depend on phi_gpu,
+# when phi is built to dynamic library, it's fine. But when phi
+# is built to static library, phi_gpu should be linked to phi_core.
+# By the way, cyclic dependency is allowed in static library.
+if(WITH_ROCM AND NOT WITH_SHARED_PHI)
+  target_link_libraries(phi_core phi_gpu)
+endif()
+
+target_compile_definitions(phi_core PUBLIC PHI_INNER)
 
 if(WIN32)
-  target_link_libraries(phi shlwapi.lib)
+  target_link_libraries(phi_core shlwapi.lib)
 endif()
 
 if(WIN32)
   if(WITH_SHARED_PHI)
-    set_property(TARGET phi PROPERTY WINDOWS_EXPORT_ALL_SYMBOLS ON)
+    set_property(TARGET phi_core PROPERTY WINDOWS_EXPORT_ALL_SYMBOLS ON)
     set(PHI_NAME
         phi.dll
         CACHE INTERNAL "" FORCE)
@@ -203,10 +235,28 @@ endif()
 
 set(PHI_LIB
     "${CMAKE_CURRENT_BINARY_DIR}/${PHI_NAME}"
-    CACHE FILEPATH "PHI Library" FORCE)
+    CACHE FILEPATH "PHI Dummy Library" FORCE)
+
+string(REPLACE "phi" "phi_core" PHI_CORE_NAME ${PHI_NAME})
+set(PHI_CORE_NAME
+    ${PHI_CORE_NAME}
+    CACHE INTERNAL "" FORCE)
+set(PHI_CORE_LIB
+    "${CMAKE_CURRENT_BINARY_DIR}/${PHI_CORE_NAME}"
+    CACHE FILEPATH "PHI CPU Library" FORCE)
+
+if(WITH_GPU OR WITH_ROCM)
+  string(REPLACE "phi" "phi_gpu" PHI_GPU_NAME ${PHI_NAME})
+  set(PHI_GPU_NAME
+      ${PHI_GPU_NAME}
+      CACHE INTERNAL "" FORCE)
+  set(PHI_GPU_LIB
+      "${CMAKE_CURRENT_BINARY_DIR}/${PHI_GPU_NAME}"
+      CACHE FILEPATH "PHI GPU Library" FORCE)
+endif()
 
 if(MKL_FOUND AND WITH_ONEMKL)
-  target_include_directories(phi PRIVATE ${MKL_INCLUDE})
+  target_include_directories(phi_core PRIVATE ${MKL_INCLUDE})
 endif()
 
 add_dependencies(phi extern_lapack)

paddle/phi/kernels/CMakeLists.txt

@@ -29,8 +29,6 @@ file(GLOB kernel_primitive_h "primitive/*.h")
 file(
   GLOB kernel_cu
   RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}"
-  "gpu/*.cu"
-  "gpu/*.cu.cc"
   "gpudnn/*.cu"
   "kps/*.cu"
   "legacy/kps/*.cu"
@@ -40,18 +38,25 @@ file(
   "strings/gpu/*.cu"
   "fusion/gpu/*.cu")
 
+file(
+  GLOB kernel_gpu
+  RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}"
+  "gpu/*.cu" "gpu/*.cu.cc")
+
 if(APPLE OR WIN32)
   list(REMOVE_ITEM kernel_cu "fusion/gpu/fusion_group_kernel.cu")
   list(REMOVE_ITEM kernel_cu "sparse/gpu/conv_kernel_igemm.cu")
 endif()
 
 if(NOT WITH_DGC)
-  list(REMOVE_ITEM kernel_cu "gpu/dgc_kernel.cu")
+  list(REMOVE_ITEM kernel_gpu "gpu/dgc_kernel.cu")
 endif()
 
 if(DEFINED REDUCE_INFERENCE_LIB_SIZE)
   list(FILTER kernel_cu EXCLUDE REGEX ".*_grad_kernel\\.cc$")
   list(FILTER kernel_cu EXCLUDE REGEX ".*_grad_kernel\\.cu$")
+  list(FILTER kernel_gpu EXCLUDE REGEX ".*_grad_kernel\\.cc$")
+  list(FILTER kernel_gpu EXCLUDE REGEX ".*_grad_kernel\\.cu$")
 endif()
 
 if(WITH_CUTLASS)
@@ -201,6 +206,15 @@ if(WITH_ROCM)
   list(
     REMOVE_ITEM
     kernel_cu
+    "gpudnn/mha_cudnn_frontend.cu"
+    "fusion/gpu/blha_get_max_len.cu"
+    "fusion/gpu/block_multi_head_attention_kernel.cu"
+    "fusion/gpu/fused_bn_add_activation_grad_kernel.cu"
+    "fusion/gpu/fused_bn_add_activation_kernel.cu"
+    "fusion/gpu/fusion_transpose_flatten_concat_kernel.cu")
+  list(
+    REMOVE_ITEM
+    kernel_gpu
     "gpu/affine_grid_grad_kernel.cu"
     "gpu/apply_per_channel_scale_kernel.cu"
     "gpu/cholesky_solve_kernel.cu"
@@ -213,13 +227,7 @@ if(WITH_ROCM)
     "gpu/put_along_axis_grad_kernel.cu"
     "gpu/put_along_axis_kernel.cu"
     "gpu/qr_kernel.cu"
-    "gpu/svd_kernel.cu"
-    "gpudnn/mha_cudnn_frontend.cu"
-    "fusion/gpu/blha_get_max_len.cu"
-    "fusion/gpu/block_multi_head_attention_kernel.cu"
-    "fusion/gpu/fused_bn_add_activation_grad_kernel.cu"
-    "fusion/gpu/fused_bn_add_activation_kernel.cu"
-    "fusion/gpu/fusion_transpose_flatten_concat_kernel.cu")
+    "gpu/svd_kernel.cu")
 endif()
 
 set(cc_search_pattern
@@ -276,6 +284,8 @@ file(
 if(WITH_GPU OR WITH_ROCM)
   collect_srcs(kernels_srcs SRCS ${kernel_cu})
   kernel_declare("${kernel_cu}")
+  collect_srcs(kernels_gpu_srcs SRCS ${kernel_gpu})
+  kernel_declare("${kernel_gpu}")
 endif()
 
 if(WITH_XPU)

python/env_dict.py.in

@@ -14,6 +14,10 @@ env_dict={
     'FLUID_CORE_NAME':'@FLUID_CORE_NAME@',
     'PHI_LIB':'@PHI_LIB@',
     'PHI_NAME':'@PHI_NAME@',
+    'PHI_CORE_LIB':'@PHI_CORE_LIB@',
+    'PHI_CORE_NAME':'@PHI_CORE_NAME@',
+    'PHI_GPU_LIB':'@PHI_GPU_LIB@',
+    'PHI_GPU_NAME':'@PHI_GPU_NAME@',
     'WITH_SHARED_PHI':'@WITH_SHARED_PHI@',
     'IR_LIB':'@IR_LIB@',
     'IR_NAME':'@IR_NAME@',

python/setup.py.in

@@ -688,6 +688,11 @@ package_data['paddle.libs']= []
 if('${WITH_SHARED_PHI}' == 'ON'):
     package_data['paddle.libs'] += [('libphi' if os.name != 'nt' else 'phi') + ext_name]
     shutil.copy('${PHI_LIB}', libs_path)
+    package_data['paddle.libs'] += [('libphi_core' if os.name != 'nt' else 'phi_core') + ext_name]
+    shutil.copy('${PHI_CORE_LIB}', libs_path)
+    if('${WTIH_GPU}' == 'ON' or '${WTIH_ROCM}' == 'ON'):
+        package_data['paddle.libs'] += [('libphi_gpu' if os.name != 'nt' else 'phi_gpu') + ext_name]
+        shutil.copy('${PHI_GPU_LIB}', libs_path)
 
 if('${WITH_SHARED_IR}' == 'ON'):
     package_data['paddle.libs'] += [('libpir' if os.name != 'nt' else 'pir') + ext_name]
@@ -840,6 +845,9 @@ if '${CMAKE_BUILD_TYPE}' == 'Release':
             if('${WITH_SHARED_PHI}' == 'ON'):
                 # change rpath of phi.ext for loading 3rd party libb
                 commands.append("install_name_tool -add_rpath '@loader_path' ${PADDLE_BINARY_DIR}/python/paddle/libs/${PHI_NAME}")
+                commands.append("install_name_tool -add_rpath '@loader_path' ${PADDLE_BINARY_DIR}/python/paddle/libs/${PHI_CORE_NAME}")
+                if('${WITH_GPU}' == 'ON' or '${WITH_ROCM}' == 'ON'):
+                    commands.append("install_name_tool -add_rpath '@loader_path' ${PADDLE_BINARY_DIR}/python/paddle/libs/${PHI_GPU_NAME}")
             if('${WITH_SHARED_IR}' == 'ON'):
                 # change rpath of pir.ext for loading 3rd party libb
                 commands.append("install_name_tool -add_rpath '@loader_path' ${PADDLE_BINARY_DIR}/python/paddle/libs/${IR_NAME}")
@@ -848,6 +856,10 @@ if '${CMAKE_BUILD_TYPE}' == 'Release':
             if('${WITH_SHARED_PHI}' == 'ON'):
                 # change rpath of phi.ext for loading 3rd party lib
                 commands.append("patchelf --set-rpath '$ORIGIN/../../nvidia/cuda_runtime/lib:$ORIGIN:$ORIGIN/../libs' ${PADDLE_BINARY_DIR}/python/paddle/libs/${PHI_NAME}")
+                commands.append("patchelf --set-rpath '$ORIGIN/../../nvidia/cuda_runtime/lib:$ORIGIN:$ORIGIN/../libs' ${PADDLE_BINARY_DIR}/python/paddle/libs/${PHI_CORE_NAME}")
+                if('${WITH_GPU}' == 'ON' or '${WITH_ROCM}' == 'ON'):
+                    commands.append("patchelf --set-rpath '$ORIGIN/../../nvidia/cuda_runtime/lib:$ORIGIN:$ORIGIN/../libs' ${PADDLE_BINARY_DIR}/python/paddle/libs/${PHI_GPU_NAME}")
+
             if('${WITH_SHARED_IR}' == 'ON'):
                 # change rpath of pir.ext for loading 3rd party lib
                 commands.append("patchelf --set-rpath '$ORIGIN:$ORIGIN/../libs' ${PADDLE_BINARY_DIR}/python/paddle/libs/${IR_NAME}")

python/setup_cinn.py.in

@@ -142,6 +142,9 @@ if '${WITH_ONEDNN}' == 'ON':
     cinnlibs.append('${ONEDNN_SHARED_LIB}')
 
 cinnlibs.append('${PHI_LIB}')
+cinnlibs.append('${PHI_CORE_LIB}')
+if '${WITH_GPU}' == 'ON' or '${WITH_ROCM}' == 'ON':
+    cinnlibs.append('${PHI_GPU_LIB}')
 cinnlibs.append('${IR_LIB}')
 cinnlibs.append('${COMMON_LIB}')