Incremental compilation optimization of phi (PaddlePaddle#59811)

* mv phi/ops/.cc file to fluid

* fix phi/kernels/compile bug

* optimization cmake code

* optimiztion compiliation of phi

* tmp for gen fpaintb.py

* refine gen mxigemm py

* refine gen mxigemm py

* optimiztion compiliation of phi

* optimiztion compiliation of phi

* optimization compilation of phi

* optimization compilation of phi

---------

Co-authored-by: wwbitejotunn <wang_bojun@outlook.com>

.gitignore

@@ -82,7 +82,7 @@ paddle/fluid/operators/generated_op*.cc
 paddle/fluid/operators/generated_sparse_op.cc
 paddle/fluid/operators/generated_static_op.cc
 paddle/fluid/operators/generated_fused_op.cc
-paddle/phi/ops/compat/generated_*.cc
+paddle/fluid/operators/ops_signature/generated_*.cc
 paddle/phi/api/yaml/parsed_apis/
 paddle/fluid/operators/generator/parsed_ops/
 paddle/fluid/pybind/tmp_eager_op_function_impl.h
@@ -95,9 +95,11 @@ paddle/fluid/framework/__init__.py
 paddle/phi/api/profiler/__init__.py
 python/paddle/incubate/fleet/parameter_server/pslib/ps_pb2.py
 paddle/phi/kernels/fusion/cutlass/conv2d/generated/*
+paddle/phi/kernels/fusion/cutlass/conv2d/generated_tmp/*
 python/paddle/base/incubate/fleet/parameter_server/pslib/ps_pb2.py
 paddle/fluid/ir_adaptor/translator/op_compat_info.cc
 paddle/phi/kernels/fusion/cutlass/cutlass_kernels/fpA_intB_gemm/autogen/*
+paddle/phi/kernels/fusion/cutlass/cutlass_kernels/fpA_intB_gemm/autogen_tmp/*
 paddle/fluid/pybind/static_op_function.*
 paddle/fluid/pybind/ops_api.cc
 paddle/fluid/pir/dialect/operator/ir/pd_api.*

cmake/operators.cmake

@@ -734,3 +734,53 @@ function(prune_pybind_h)
     endif()
   endforeach()
 endfunction()
+
+function(append_op_util_declare TARGET)
+  file(READ ${TARGET} target_content)
+  string(REGEX MATCH "(PD_REGISTER_ARG_MAPPING_FN)\\([ \t\r\n]*[a-z0-9_]*"
+               util_registrar "${target_content}")
+  if(NOT ${util_registrar} EQUAL "")
+    string(REPLACE "PD_REGISTER_ARG_MAPPING_FN" "PD_DECLARE_ARG_MAPPING_FN"
+                   util_declare "${util_registrar}")
+    string(APPEND util_declare ");\n")
+    file(APPEND ${op_utils_header} "${util_declare}")
+  endif()
+endfunction()
+
+function(append_op_kernel_map_declare TARGET)
+  file(READ ${TARGET} target_content)
+  string(
+    REGEX
+      MATCH
+      "(PD_REGISTER_BASE_KERNEL_NAME)\\([ \t\r\n]*[a-z0-9_]*,[ \\\t\r\n]*[a-z0-9_]*"
+      kernel_mapping_registrar
+      "${target_content}")
+  if(NOT ${kernel_mapping_registrar} EQUAL "")
+    string(REPLACE "PD_REGISTER_BASE_KERNEL_NAME" "PD_DECLARE_BASE_KERNEL_NAME"
+                   kernel_mapping_declare "${kernel_mapping_registrar}")
+    string(APPEND kernel_mapping_declare ");\n")
+    file(APPEND ${op_utils_header} "${kernel_mapping_declare}")
+  endif()
+endfunction()
+
+function(register_op_utils TARGET_NAME)
+  set(utils_srcs)
+  set(options "")
+  set(oneValueArgs "")
+  set(multiValueArgs EXCLUDES DEPS)
+  cmake_parse_arguments(register_op_utils "${options}" "${oneValueArgs}"
+                        "${multiValueArgs}" ${ARGN})
+
+  file(GLOB SIGNATURES
+       "${PADDLE_SOURCE_DIR}/paddle/fluid/operators/ops_signature/*_sig.cc")
+  foreach(target ${SIGNATURES})
+    append_op_util_declare(${target})
+    append_op_kernel_map_declare(${target})
+    list(APPEND utils_srcs ${target})
+  endforeach()
+
+  cc_library(
+    ${TARGET_NAME}
+    SRCS ${utils_srcs}
+    DEPS ${register_op_utils_DEPS})
+endfunction()

cmake/phi.cmake

@@ -172,55 +172,6 @@ function(kernel_declare TARGET_LIST)
   endforeach()
 endfunction()
 
-function(append_op_util_declare TARGET)
-  file(READ ${TARGET} target_content)
-  string(REGEX MATCH "(PD_REGISTER_ARG_MAPPING_FN)\\([ \t\r\n]*[a-z0-9_]*"
-               util_registrar "${target_content}")
-  if(NOT ${util_registrar} EQUAL "")
-    string(REPLACE "PD_REGISTER_ARG_MAPPING_FN" "PD_DECLARE_ARG_MAPPING_FN"
-                   util_declare "${util_registrar}")
-    string(APPEND util_declare ");\n")
-    file(APPEND ${op_utils_header} "${util_declare}")
-  endif()
-endfunction()
-
-function(append_op_kernel_map_declare TARGET)
-  file(READ ${TARGET} target_content)
-  string(
-    REGEX
-      MATCH
-      "(PD_REGISTER_BASE_KERNEL_NAME)\\([ \t\r\n]*[a-z0-9_]*,[ \\\t\r\n]*[a-z0-9_]*"
-      kernel_mapping_registrar
-      "${target_content}")
-  if(NOT ${kernel_mapping_registrar} EQUAL "")
-    string(REPLACE "PD_REGISTER_BASE_KERNEL_NAME" "PD_DECLARE_BASE_KERNEL_NAME"
-                   kernel_mapping_declare "${kernel_mapping_registrar}")
-    string(APPEND kernel_mapping_declare ");\n")
-    file(APPEND ${op_utils_header} "${kernel_mapping_declare}")
-  endif()
-endfunction()
-
-function(register_op_utils TARGET_NAME)
-  set(utils_srcs)
-  set(options "")
-  set(oneValueArgs "")
-  set(multiValueArgs EXCLUDES DEPS)
-  cmake_parse_arguments(register_op_utils "${options}" "${oneValueArgs}"
-                        "${multiValueArgs}" ${ARGN})
-
-  file(GLOB SIGNATURES "${PADDLE_SOURCE_DIR}/paddle/phi/ops/compat/*_sig.cc")
-  foreach(target ${SIGNATURES})
-    append_op_util_declare(${target})
-    append_op_kernel_map_declare(${target})
-    list(APPEND utils_srcs ${target})
-  endforeach()
-
-  cc_library(
-    ${TARGET_NAME}
-    SRCS ${utils_srcs}
-    DEPS ${register_op_utils_DEPS})
-endfunction()
-
 function(prune_declaration_h)
   set(kernel_list ${KERNEL_LIST})
   file(STRINGS ${kernel_declare_file} kernel_registry_list)

paddle/fluid/framework/operator.cc

@@ -30,6 +30,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/var_type.h"
 #include "paddle/fluid/operators/isfinite_op.h"
 #include "paddle/fluid/operators/ops_extra_info.h"
+#include "paddle/fluid/operators/ops_signature/signatures.h"
 #include "paddle/fluid/platform/device/device_wrapper.h"
 #include "paddle/fluid/platform/enforce.h"
 #include "paddle/fluid/platform/profiler.h"
@@ -41,7 +42,6 @@ limitations under the License. */
 #include "paddle/phi/core/flags.h"
 #include "paddle/phi/core/kernel_context.h"
 #include "paddle/phi/core/kernel_factory.h"
-#include "paddle/phi/ops/compat/signatures.h"
 #include "paddle/utils/flags.h"
 
 namespace phi {

paddle/fluid/framework/phi_utils.cc

@@ -157,7 +157,7 @@ KernelArgsNameMakerByOpProto::GetInputArgsNames() {
       continue;
     }
     // If contains dispensable input, we should override the
-    // OpArgumentMapping method self in phi/ops/compat dir
+    // OpArgumentMapping method self in fluid/operators/ops_signature dir
     if (in.has_dispensable() && in.dispensable()) {
       continue;
     }

paddle/fluid/operators/generator/CMakeLists.txt

@@ -1,5 +1,5 @@
 # phi auto cmake utils
-include(phi)
+include(operators)
 
 # set yaml file path
 set(op_yaml_file ${CMAKE_SOURCE_DIR}/paddle/phi/api/yaml/ops.yaml)
@@ -115,13 +115,16 @@ set(generated_fused_op_path
 set(generated_sparse_ops_path
     ${CMAKE_SOURCE_DIR}/paddle/fluid/operators/generated_sparse_op.cc)
 set(generated_argument_mapping_path
-    ${CMAKE_SOURCE_DIR}/paddle/phi/ops/compat/generated_sig.cc)
+    ${CMAKE_SOURCE_DIR}/paddle/fluid/operators/ops_signature/generated_sig.cc)
 set(generated_fused_argument_mapping_path
-    ${CMAKE_SOURCE_DIR}/paddle/phi/ops/compat/generated_fused_sig.cc)
+    ${CMAKE_SOURCE_DIR}/paddle/fluid/operators/ops_signature/generated_fused_sig.cc
+)
 set(generated_static_argument_mapping_path
-    ${CMAKE_SOURCE_DIR}/paddle/phi/ops/compat/generated_static_sig.cc)
+    ${CMAKE_SOURCE_DIR}/paddle/fluid/operators/ops_signature/generated_static_sig.cc
+)
 set(generated_sparse_argument_mapping_path
-    ${CMAKE_SOURCE_DIR}/paddle/phi/ops/compat/generated_sparse_sig.cc)
+    ${CMAKE_SOURCE_DIR}/paddle/fluid/operators/ops_signature/generated_sparse_sig.cc
+)
 execute_process(
   WORKING_DIRECTORY ${CMAKE_SOURCE_DIR}/paddle/fluid/operators/generator
   COMMAND ${CMAKE_COMMAND} -E make_directory ${parsed_op_dir} RESULTS_VARIABLE
@@ -342,10 +345,10 @@ execute_process(
 message("generate ${ops_extra_info_file}")
 
 set(op_utils_header
-    ${PADDLE_BINARY_DIR}/paddle/phi/ops/compat/signatures.h.tmp
+    ${PADDLE_BINARY_DIR}/paddle/fluid/operators/ops_signature/signatures.h.tmp
     CACHE INTERNAL "op_args_fns.cc file")
 set(op_utils_header_final
-    ${PADDLE_BINARY_DIR}/paddle/phi/ops/compat/signatures.h)
+    ${PADDLE_BINARY_DIR}/paddle/fluid/operators/ops_signature/signatures.h)
 file(
   WRITE ${op_utils_header}
   "// Generated by the paddle/fluid/operators/generator/CMakeLists.txt.  DO NOT EDIT!\n\n"

paddle/phi/kernels/CMakeLists.txt

@@ -56,11 +56,33 @@ endif()
 if(WITH_CUTLASS)
   execute_process(
     COMMAND ${CMAKE_COMMAND} -E make_directory
-            "${CMAKE_CURRENT_SOURCE_DIR}/fusion/cutlass/conv2d/generated"
+            "${CMAKE_CURRENT_SOURCE_DIR}/fusion/cutlass/conv2d/generated_tmp"
     COMMAND ${PYTHON_EXECUTABLE} "conv2d_bias_act.py"
     COMMAND ${PYTHON_EXECUTABLE} "conv2d_bias_residual.py"
     COMMAND ${PYTHON_EXECUTABLE} "conv2d_depthwise_bias_act.py"
     WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/fusion/cutlass/conv2d")
+  set(generated_tmp_dir
+      ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/fusion/cutlass/conv2d/generated_tmp
+  )
+  set(generated_dir
+      ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/fusion/cutlass/conv2d/generated)
+  file(GLOB con2d_generated_files ${generated_tmp_dir}/*.cu)
+
+  if(EXISTS ${generated_dir})
+    foreach(gen_file ${con2d_generated_files})
+      string(REPLACE "generated_tmp" "generated" now_file ${gen_file})
+      execute_process(COMMAND ${CMAKE_COMMAND} -E copy_if_different
+                              "${gen_file}" "${now_file}")
+    endforeach()
+    message("copy if different ${generated_dir}")
+  else()
+    foreach(gen_file ${con2d_generated_files})
+      string(REPLACE "generated_tmp" "generated" now_file ${gen_file})
+      execute_process(COMMAND ${CMAKE_COMMAND} -E copy "${gen_file}"
+                              "${now_file}")
+    endforeach()
+    message("copy ${generated_dir}")
+  endif()
 
   execute_process(
     COMMAND
@@ -145,17 +167,39 @@ if(WITH_CUTLASS)
   )
 
   execute_process(
-    COMMAND
-      ${CMAKE_COMMAND} -E remove_directory
-      "${CMAKE_CURRENT_SOURCE_DIR}/fusion/cutlass/cutlass_kernels/fpA_intB_gemm/autogen"
     COMMAND
       ${CMAKE_COMMAND} -E make_directory
-      "${CMAKE_CURRENT_SOURCE_DIR}/fusion/cutlass/cutlass_kernels/fpA_intB_gemm/autogen"
+      "${CMAKE_CURRENT_SOURCE_DIR}/fusion/cutlass/cutlass_kernels/fpA_intB_gemm/autogen_tmp"
     COMMAND ${PYTHON_EXECUTABLE} generic_mixed_gemm_kernelLauncher.py
             --cuda_arch "${NVCC_ARCH_BIN}"
     WORKING_DIRECTORY
       "${CMAKE_CURRENT_SOURCE_DIR}/fusion/cutlass/cutlass_kernels/fpA_intB_gemm"
   )
+  set(fpA_intB_gemm_autogen_tmp_dir
+      ${CMAKE_CURRENT_SOURCE_DIR}/fusion/cutlass/cutlass_kernels/fpA_intB_gemm/autogen_tmp
+  )
+  set(fpA_intB_gemm_autogen_dir
+      ${CMAKE_CURRENT_SOURCE_DIR}/fusion/cutlass/cutlass_kernels/fpA_intB_gemm/autogen
+  )
+
+  file(GLOB fpA_intB_gemm_autogen_files ${fpA_intB_gemm_autogen_tmp_dir}/*.h
+       ${fpA_intB_gemm_autogen_tmp_dir}/*.cu)
+
+  if(EXISTS ${fpA_intB_gemm_autogen_dir})
+    foreach(gen_file ${fpA_intB_gemm_autogen_files})
+      string(REPLACE "autogen_tmp" "autogen" now_file ${gen_file})
+      execute_process(COMMAND ${CMAKE_COMMAND} -E copy_if_different
+                              "${gen_file}" "${now_file}")
+    endforeach()
+    message("copy if different ${fpA_intB_gemm_autogen_dir}")
+  else()
+    foreach(gen_file ${fpA_intB_gemm_autogen_files})
+      string(REPLACE "autogen_tmp" "autogen" now_file ${gen_file})
+      execute_process(COMMAND ${CMAKE_COMMAND} -E copy "${gen_file}"
+                              "${now_file}")
+    endforeach()
+    message("copy ${fpA_intB_gemm_autogen_dir}")
+  endif()
 
   file(
     GLOB cutlass_cu
@@ -167,6 +211,7 @@ if(WITH_CUTLASS)
     "fusion/cutlass/memory_efficient_attention/autogen_variable/impl/*.cu"
     "fusion/cutlass/cutlass_kernels/fpA_intB_gemm/autogen/*.cu"
     "fusion/cutlass/cutlass_kernels/fpA_intB_gemm/*.cu")
+
   list(APPEND kernel_cu ${cutlass_cu})
 endif()
 

paddle/phi/kernels/fusion/cutlass/conv2d/conv2d_bias_act.py

@@ -225,6 +225,6 @@ def generate_sm75_1688():
         sm_versions, SupportedAct, UnderScoreName, CamelName
     )
     all_code += CommonTail
-    with open("generated/conv2d_bias_act.cu", "w") as f:
+    with open("generated_tmp/conv2d_bias_act.cu", "w") as f:
         f.write(all_code)
         f.close()

paddle/phi/kernels/fusion/cutlass/conv2d/conv2d_bias_residual.py

@@ -210,6 +210,6 @@ def generate_sm75_1688():
         sm_versions, SupportedEpilogue, UnderScoreName, CamelName
     )
     all_code += CommonTail
-    with open("generated/conv2d_bias_residual.cu", "w") as f:
+    with open("generated_tmp/conv2d_bias_residual.cu", "w") as f:
         f.write(all_code)
         f.close()

paddle/phi/kernels/fusion/cutlass/conv2d/conv2d_depthwise_bias_act.py

@@ -225,6 +225,6 @@ def generate_conv2d_depthwise():
     all_code = cdba_header
     all_code += generate_conv2d_depthwise()
     all_code += CommonTail
-    with open("generated/conv2d_depthwise_bias_act.cu", "w") as f:
+    with open("generated_tmp/conv2d_depthwise_bias_act.cu", "w") as f:
         f.write(all_code)
         f.close()

paddle/phi/kernels/fusion/cutlass/cutlass_kernels/fpA_intB_gemm/generic_mixed_gemm_kernelLauncher.py

@@ -204,7 +204,7 @@ def generate_source_cu(
     args = parse_args()
     archs = args.cuda_arch
     header_all = DefineHeader
-    header_name = "autogen/arch_define.h"
+    header_name = "autogen_tmp/arch_define.h"
     if archs:
         for arch in archs:
             define_line = "#define USE_FPAINTB_GEMM_WITH_SM%s\n" % str(arch)
@@ -217,7 +217,7 @@ def generate_source_cu(
             for arch in archs:
                 for epilogue_tag in EpilogueTags.keys():
                     for stages in StagesList[arch]:
-                        file_name = "autogen/generic_mixed_gemm_kernelLauncher_{}_sm{}_stages{}_{}.cu".format(
+                        file_name = "autogen_tmp/generic_mixed_gemm_kernelLauncher_{}_sm{}_stages{}_{}.cu".format(
                             element_type, arch, stages, epilogue_tag
                         )
                         all_code = generate_source_cu(

test/cpp/phi/core/test_op_utils.cc

@@ -15,8 +15,8 @@ limitations under the License. */
 #include <iostream>
 
 #include "gtest/gtest.h"
+#include "paddle/fluid/operators/ops_signature/signatures.h"
 #include "paddle/phi/core/compat/op_utils.h"
-#include "paddle/phi/ops/compat/signatures.h"
 
 namespace phi {
 namespace tests {

test/cpp/phi/ops/test_op_signature.cc

@@ -19,8 +19,8 @@ limitations under the License. */
 #include <memory>
 #include <unordered_set>
 
+#include "paddle/fluid/operators/ops_signature/signatures.h"
 #include "paddle/phi/core/compat/op_utils.h"
-#include "paddle/phi/ops/compat/signatures.h"
 
 namespace phi {
 namespace tests {