enable compiled into one library on windows

intel · Dec 25, 2024 · 89af91f · 89af91f
1 parent 5bfd384
commit 89af91f
Show file tree

Hide file tree

Showing 3 changed files with 54 additions and 25 deletions.
diff --git a/cmake/BuildFlags.cmake b/cmake/BuildFlags.cmake
@@ -123,9 +123,9 @@ if(CMAKE_CXX_COMPILER_ID STREQUAL "GNU" OR CMAKE_CXX_COMPILER_ID STREQUAL "MSVC"
   set(SYCL_OFFLINE_COMPILER_CG_OPTIONS "-options '${SYCL_OFFLINE_COMPILER_CG_OPTIONS}'")
 
   if(WIN32)
-    set(AOT_TARGETS "ats-m150,mtl-u,mtl-h,xe2-lpg,xe2-hpg")
+    set(AOT_TARGETS "dg2-g10,mtl-u,mtl-h,xe2-lpg,xe2-hpg")
   else()
-    set(AOT_TARGETS "pvc,xe-lpg,ats-m150")
+    set(AOT_TARGETS "pvc,xe-lpg,dg2-g10,xe2-lpg,xe2-hpg")
   endif()
   if((DEFINED ENV{TORCH_XPU_ARCH_LIST}) AND NOT ("$ENV{TORCH_XPU_ARCH_LIST}" STREQUAL ""))
     set(AOT_TARGETS "$ENV{TORCH_XPU_ARCH_LIST}")

diff --git a/cmake/Modules/FindSYCL.cmake b/cmake/Modules/FindSYCL.cmake
@@ -215,10 +215,17 @@ macro(SYCL_WRAP_SRCS sycl_target generated_files)
   set(SYCL_flags "")
   set(generated_extension ${CMAKE_${SYCL_C_OR_CXX}_OUTPUT_EXTENSION})
 
+  set(library_name)
+  if(${sycl_target} STREQUAL "xpu_ops_kernels")
+    set(library_name torch_xpu_ops_sycl_kernels)
+  else()
+    set(library_name ${sycl_target})
+  endif()
+
   set(SYCL_include_dirs "${SYCL_INCLUDE_DIR}")
-  list(APPEND SYCL_include_dirs "$<TARGET_PROPERTY:${sycl_target},INCLUDE_DIRECTORIES>")
+  list(APPEND SYCL_include_dirs "$<TARGET_PROPERTY:${library_name},INCLUDE_DIRECTORIES>")
 
-  set(SYCL_compile_definitions "$<TARGET_PROPERTY:${sycl_target},COMPILE_DEFINITIONS>")
+  set(SYCL_compile_definitions "$<TARGET_PROPERTY:${library_name},COMPILE_DEFINITIONS>")
 
   SYCL_GET_SOURCES_AND_OPTIONS(
     _sycl_sources
@@ -264,7 +271,7 @@ macro(SYCL_WRAP_SRCS sycl_target generated_files)
 
       get_filename_component( basename ${file} NAME )
       set(generated_file_path "${SYCL_compile_output_dir}/${CMAKE_CFG_INTDIR}")
-      set(generated_file_basename "${sycl_target}_generated_${basename}${generated_extension}")
+      set(generated_file_basename "${sycl_target}_gen_${basename}${generated_extension}")
       set(generated_file "${generated_file_path}/${generated_file_basename}")
       set(SYCL_generated_dependency_file "${SYCL_compile_intermediate_directory}/${generated_file_basename}.SYCL-depend") # generate by compiler options -M -MF
       set(cmake_dependency_file "${SYCL_compile_intermediate_directory}/${generated_file_basename}.depend") # parse and convert SYCL_generated_dependency_file(compiler format) to cmake format
@@ -434,6 +441,16 @@ macro(SYCL_ADD_LIBRARY sycl_target)
     ${ARGN})
 
   SYCL_BUILD_SHARED_LIBRARY(_sycl_shared_flag ${ARGN})
+
+  # For one library, in which case the sycl_target should be shorter to fit the 
+  # character limit of Windows batch files, and the library_name is set to be 
+  # more recognizable.
+  set(library_name)
+  if(${sycl_target} STREQUAL "xpu_ops_kernels")
+    set(library_name torch_xpu_ops_sycl_kernels)
+  else()
+    set(library_name ${sycl_target})
+  endif()
 
   if(_sycl_sources)
     # Compile sycl sources
@@ -455,24 +472,24 @@ macro(SYCL_ADD_LIBRARY sycl_target)
       ${${sycl_target}_sycl_objects})
 
     add_library(
-      ${sycl_target}
+      ${library_name}
       ${_cmake_options}
       ${_cxx_sources}
       ${${sycl_target}_sycl_objects}
       ${device_object})
   else()
     add_library(
-      ${sycl_target}
+      ${library_name}
       ${_cmake_options}
       ${_cxx_sources})
   endif()
 
   target_link_libraries(
-    ${sycl_target}
+    ${library_name}
     ${SYCL_LINK_LIBRARIES_KEYWORD}
     ${SYCL_LIBRARY})
 
-  set_target_properties(${sycl_target}
+  set_target_properties(${library_name}
     PROPERTIES
     LINKER_LANGUAGE ${SYCL_C_OR_CXX})
 

diff --git a/src/BuildOnWindows.cmake b/src/BuildOnWindows.cmake
@@ -35,7 +35,7 @@ if(BUILD_SEPARATE_OPS)
     # Decouple with PyTorch cmake definition.
     install(TARGETS ${sycl_lib} DESTINATION "${TORCH_INSTALL_LIB_DIR}")
   endforeach()
-else()
+elseif(BUILD_SPLIT_KERNEL_LIB)
   # Split SYCL kernels into 2 libraries as categories 1) Unary+Binary 2) Others.
   set(ATen_XPU_SYCL_BINARY_SRCS)
   set(ATen_XPU_SYCL_UNARY_SRCS)
@@ -60,13 +60,13 @@ else()
     string(REGEX MATCH "Activation" IS_ACTIVATION ${sycl_src})
     string(REGEX MATCH "Foreach" IS_FOREACH ${sycl_src})
     string(REGEX MATCH "Reduce" IS_REDUCE ${sycl_src})
-	string(REGEX MATCH "Tensor" IS_TENSOR ${sycl_src})
-	string(REGEX MATCH "Norm" IS_NORM ${sycl_src})
-	string(REGEX MATCH "Loss" IS_LOSS ${sycl_src})
-	string(REGEX MATCH "Polynomial" IS_POLY ${sycl_src})
-	#Move resize kernel to Norm and Loss lib, to resolve symbol.
-	string(REGEX MATCH "Resize" IS_RESIZE ${sycl_src})
-	string(REGEX MATCH "Distribution" IS_DISTRIBUTION ${sycl_src})
+	  string(REGEX MATCH "Tensor" IS_TENSOR ${sycl_src})
+	  string(REGEX MATCH "Norm" IS_NORM ${sycl_src})
+	  string(REGEX MATCH "Loss" IS_LOSS ${sycl_src})
+	  string(REGEX MATCH "Polynomial" IS_POLY ${sycl_src})
+	  #Move resize kernel to Norm and Loss lib, to resolve symbol.
+	  string(REGEX MATCH "Resize" IS_RESIZE ${sycl_src})
+	  string(REGEX MATCH "Distribution" IS_DISTRIBUTION ${sycl_src})
 
     if(NOT IS_FOREACH STREQUAL "")
       list(APPEND ATen_XPU_SYCL_FOREACH_SRCS ${sycl_src})
@@ -78,14 +78,14 @@ else()
       list(APPEND ATen_XPU_SYCL_REDUCE_SRCS ${sycl_src})
     elseif(NOT IS_ACTIVATION STREQUAL "")
       list(APPEND ATen_XPU_SYCL_ACTIVATION_SRCS ${sycl_src})
-	elseif(NOT IS_TENSOR STREQUAL "")
-	  list(APPEND ATen_XPU_SYCL_TENSOR_SRCS ${sycl_src})
-	elseif(NOT IS_DISTRIBUTION STREQUAL "")
-	  list(APPEND ATen_XPU_SYCL_DISTRIBUTION_SRCS ${sycl_src})
-	elseif(NOT IS_NORM STREQUAL "" OR NOT IS_LOSS STREQUAL "" OR NOT IS_RESIZE STREQUAL "")
-	  list(APPEND ATen_XPU_SYCL_NORM_LOSS_SRCS ${sycl_src})
-	elseif(NOT IS_POLY STREQUAL "")
-	  list(APPEND ATen_XPU_SYCL_POLY_SRCS ${sycl_src})
+    elseif(NOT IS_TENSOR STREQUAL "")
+      list(APPEND ATen_XPU_SYCL_TENSOR_SRCS ${sycl_src})
+    elseif(NOT IS_DISTRIBUTION STREQUAL "")
+      list(APPEND ATen_XPU_SYCL_DISTRIBUTION_SRCS ${sycl_src})
+    elseif(NOT IS_NORM STREQUAL "" OR NOT IS_LOSS STREQUAL "" OR NOT IS_RESIZE STREQUAL "")
+      list(APPEND ATen_XPU_SYCL_NORM_LOSS_SRCS ${sycl_src})
+    elseif(NOT IS_POLY STREQUAL "")
+      list(APPEND ATen_XPU_SYCL_POLY_SRCS ${sycl_src})
     else()
       list(APPEND ATen_XPU_SYCL_OTHERS_SRCS ${sycl_src})
     endif()
@@ -229,6 +229,18 @@ else()
 
   # Decouple with PyTorch cmake definition.
   install(TARGETS ${sycl_lib} DESTINATION "${TORCH_INSTALL_LIB_DIR}")
+else()
+  sycl_add_library(
+    xpu_ops_kernels
+    SHARED
+    SYCL_SOURCES ${ATen_XPU_SYCL_SRCS})
+  target_compile_definitions(torch_xpu_ops_sycl_kernels PRIVATE TORCH_XPU_BUILD_MAIN_LIB)
+  target_link_libraries(torch_xpu_ops_aten PUBLIC torch_xpu_ops_sycl_kernels)
+  target_link_libraries(torch_xpu_ops_sycl_kernels PUBLIC torch_xpu)
+  list(APPEND TORCH_XPU_OPS_LIBRARIES torch_xpu_ops_sycl_kernels)
+
+  # Decouple with PyTorch cmake definition.
+  install(TARGETS torch_xpu_ops_sycl_kernels DESTINATION "${TORCH_INSTALL_LIB_DIR}")
 endif()
 set(SYCL_LINK_LIBRARIES_KEYWORD)