Merge branch 'main' into penghuic/update_matmul_cuda_cases

intel · Mar 7, 2025 · d25714a · d25714a
2 parents d919c33 + 3f93cf8
commit d25714a
Show file tree

Hide file tree

Showing 54 changed files with 435 additions and 275 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -40,7 +40,7 @@ include(${TORCH_XPU_OPS_ROOT}/cmake/ONEMKL.cmake)
 include(${TORCH_XPU_OPS_ROOT}/cmake/BuildFlags.cmake)
 
 option(USE_XCCL "Build with XCCL support" OFF)
-option(USE_C10D_XCCL "Build with XCCL support for C10D" OFF)
+option(USE_C10D_XCCL "Build with XCCL support for C10D" ON)
 
 # -- [ Re-generate the macros file for https://github.com/pytorch/pytorch/pull/147161
 macro(update_caffe2_macros_file)

diff --git a/cmake/BuildFlags.cmake b/cmake/BuildFlags.cmake
@@ -120,6 +120,7 @@ if(CMAKE_CXX_COMPILER_ID STREQUAL "GNU" OR CMAKE_CXX_COMPILER_ID STREQUAL "MSVC"
   set(SYCL_OFFLINE_COMPILER_CG_OPTIONS "${SYCL_OFFLINE_COMPILER_CG_OPTIONS} -cl-poison-unsupported-fp64-kernels")
   set(SYCL_OFFLINE_COMPILER_CG_OPTIONS "${SYCL_OFFLINE_COMPILER_CG_OPTIONS} -cl-intel-enable-auto-large-GRF-mode")
   set(SYCL_OFFLINE_COMPILER_CG_OPTIONS "${SYCL_OFFLINE_COMPILER_CG_OPTIONS} -cl-fp32-correctly-rounded-divide-sqrt")
+  set(SYCL_OFFLINE_COMPILER_CG_OPTIONS "${SYCL_OFFLINE_COMPILER_CG_OPTIONS} -cl-intel-greater-than-4GB-buffer-required")
   set(SYCL_OFFLINE_COMPILER_CG_OPTIONS "-options '${SYCL_OFFLINE_COMPILER_CG_OPTIONS}'")
 
   # LNL and BMG share the same compatibility name, which is BMG. BMG is defined as the base platform.

diff --git a/cmake/Codegen.cmake b/cmake/Codegen.cmake
@@ -1,89 +1,93 @@
-if(Codegen_GPU_cmake_included)
+if(Codegen_XPU_cmake_included)
   return()
 endif()
-set(Codegen_GPU_cmake_included true)
+set(Codegen_XPU_cmake_included true)
 
-set(BUILD_TORCH_XPU_ATEN_GENERATED "${CMAKE_BINARY_DIR}/xpu/ATen/")
+set(BUILD_TORCH_XPU_ATEN_GENERATED "${CMAKE_BINARY_DIR}/xpu/ATen")
+set(BUILD_TORCH_ATEN_GENERATED "${CMAKE_BINARY_DIR}/aten/src/ATen")
 file(MAKE_DIRECTORY ${BUILD_TORCH_XPU_ATEN_GENERATED})
 
-set(RegisterXPU_PATH ${BUILD_TORCH_XPU_ATEN_GENERATED}/RegisterXPU_0.cpp)
-set(RegisterSparseXPU_PATH ${BUILD_TORCH_XPU_ATEN_GENERATED}/RegisterSparseXPU_0.cpp)
-set(RegisterSparseCsrXPU_PATH ${BUILD_TORCH_XPU_ATEN_GENERATED}/RegisterSparseCsrXPU_0.cpp)
-set(RegisterNestedTensorXPU_PATH ${BUILD_TORCH_XPU_ATEN_GENERATED}/RegisterNestedTensorXPU_0.cpp)
-set(XPUFallback_PATH ${TORCH_XPU_OPS_ROOT}/src/ATen/native/xpu/XPUFallback.template)
+set(RegisterXPU_GENERATED ${BUILD_TORCH_XPU_ATEN_GENERATED}/RegisterXPU_0.cpp)
+set(RegisterSparseXPU_GENERATED ${BUILD_TORCH_XPU_ATEN_GENERATED}/RegisterSparseXPU_0.cpp)
+set(RegisterSparseCsrXPU_GENERATED ${BUILD_TORCH_XPU_ATEN_GENERATED}/RegisterSparseCsrXPU_0.cpp)
+set(RegisterNestedTensorXPU_GENERATED ${BUILD_TORCH_XPU_ATEN_GENERATED}/RegisterNestedTensorXPU_0.cpp)
+set(XPUFallback_TEMPLATE ${TORCH_XPU_OPS_ROOT}/src/ATen/native/xpu/XPUFallback.template)
+set(XPU_AOTI_INSTALL_DIR ${TORCH_ROOT}/torch/csrc/inductor/aoti_torch/generated/extend)
+set(XPU_AOTI_SHIM_HEADER ${XPU_AOTI_INSTALL_DIR}/c_shim_xpu.h)
+set(XPU_AOTI_SHIM_SOURCE ${XPU_AOTI_INSTALL_DIR}/c_shim_xpu.cpp)
 
 if(WIN32)
   set(FILE_DISPLAY_CMD type)
-  # replace forward slash with back slash for compatibility with 'type' command on Windows
-  string(REPLACE "/" "\\" RegisterXPU_PATH_BACKSLASH "${RegisterXPU_PATH}")
-  string(REPLACE "/" "\\" XPUFallback_PATH_BACKSLASH "${XPUFallback_PATH}")
-  set(REGISTER_FALLBACK_CMD ${FILE_DISPLAY_CMD} ${XPUFallback_PATH_BACKSLASH} ">>" ${RegisterXPU_PATH_BACKSLASH})
 else()
   set(FILE_DISPLAY_CMD cat)
-  set(REGISTER_FALLBACK_CMD ${FILE_DISPLAY_CMD} ${XPUFallback_PATH} ">>" ${RegisterXPU_PATH})
 endif()
+file(TO_NATIVE_PATH "${RegisterXPU_GENERATED}" RegisterXPU_GENERATED_NATIVE)
+file(TO_NATIVE_PATH "${XPUFallback_TEMPLATE}" XPUFallback_TEMPLATE_NATIVE)
+set(REGISTER_FALLBACK_CMD ${FILE_DISPLAY_CMD} ${XPUFallback_TEMPLATE_NATIVE} ">>" ${RegisterXPU_GENERATED_NATIVE})
 
-function(GEN_BACKEND file_yaml)
-  set(generated_files "")
-  foreach(f ${ARGN})
-    list(APPEND generated_files "${BUILD_TORCH_XPU_ATEN_GENERATED}/${f}")
-  endforeach()
-  file(GLOB_RECURSE depended_files ${TORCH_XPU_OPS_ROOT}/yaml/${file_yaml})
-  add_custom_command(
-    OUTPUT ${generated_files}
-    COMMAND
-    "${PYTHON_EXECUTABLE}" -m torchgen.gen_backend_stubs
-    --output_dir ${BUILD_TORCH_XPU_ATEN_GENERATED}
-    --source_yaml ${TORCH_XPU_OPS_ROOT}/yaml/${file_yaml}
-    COMMAND
-    ${REGISTER_FALLBACK_CMD}
-    ${SIMPLE_TRACE}
-    WORKING_DIRECTORY ${TORCH_ROOT}
-    DEPENDS
-    ${depended_files}
-    ${TORCH_XPU_OPS_ROOT}/yaml/${file_yaml}
-    ${XPUFallback_PATH}
-    )
-endfunction(GEN_BACKEND)
-
-
-set(RegisterXPU_PATH ${BUILD_TORCH_XPU_ATEN_GENERATED}/RegisterXPU_0.cpp)
-set(RegisterSparseXPU_PATH ${BUILD_TORCH_XPU_ATEN_GENERATED}/RegisterSparseXPU_0.cpp)
-set(RegisterSparseCsrXPU_PATH ${BUILD_TORCH_XPU_ATEN_GENERATED}/RegisterSparseCsrXPU_0.cpp)
-set(RegisterNestedTensorXPU_PATH ${BUILD_TORCH_XPU_ATEN_GENERATED}/RegisterNestedTensorXPU_0.cpp)
-set(XPUFallback_PATH ${TORCH_XPU_OPS_ROOT}/src/ATen/native/xpu/XPUFallback.template)
-set(XPU_AOTI_INSTALL_DIR ${TORCH_ROOT}/torch/csrc/inductor/aoti_torch/generated/extend)
 function(GEN_XPU file_yaml)
   set(generated_files "")
   foreach(f ${ARGN})
     list(APPEND generated_files "${f}")
   endforeach()
-  file(GLOB_RECURSE depend_files ${TORCH_XPU_OPS_ROOT}/yaml/${file_yaml})
-  set(CODEGEN_TEMPLATE ${TORCH_XPU_OPS_ROOT}/yaml/)
+  set(CODEGEN_XPU_YAML_DIR ${TORCH_XPU_OPS_ROOT}/yaml)
 
   # Codegen prepare process
   if(WIN32)
-    string(REPLACE "/" "\\" DestPATH "${CODEGEN_TEMPLATE}templates")
-    string(REPLACE "/" "\\" SrcPATH "${CMAKE_SOURCE_DIR}/aten/src/ATen/templates")
+    file(TO_NATIVE_PATH "${CODEGEN_XPU_YAML_DIR}/templates" DestPATH)
+    file(TO_NATIVE_PATH "${CMAKE_SOURCE_DIR}/aten/src/ATen/templates" SrcPATH)
     execute_process(COMMAND cmd /c xcopy ${SrcPATH} ${DestPATH} /E /H /C /I /Y > nul)
-    string(REPLACE "/" "\\" RegisterXPU_PATH_BACKSLASH "${RegisterXPU_PATH}")
-    string(REPLACE "/" "\\" XPUFallback_PATH_BACKSLASH "${XPUFallback_PATH}")
-    set(REGISTER_FALLBACK_CMD ${FILE_DISPLAY_CMD} ${XPUFallback_PATH_BACKSLASH} ">>" ${RegisterXPU_PATH_BACKSLASH})
   else()
-    execute_process(COMMAND ln -s ${CMAKE_SOURCE_DIR}/aten/src/ATen/templates ${CODEGEN_TEMPLATE}) # soft link to pytorch templates
-    set(REGISTER_FALLBACK_CMD ${FILE_DISPLAY_CMD} ${XPUFallback_PATH} ">>" ${RegisterXPU_PATH})
+    execute_process(COMMAND ln -s ${CMAKE_SOURCE_DIR}/aten/src/ATen/templates ${CODEGEN_XPU_YAML_DIR}) # soft link to pytorch templates
   endif()
-  add_custom_command(
-    OUTPUT ${generated_files}
-    COMMAND
+
+  set(XPU_CODEGEN_COMMAND
     "${PYTHON_EXECUTABLE}" -m torchgen.gen
-    --source-path ${TORCH_XPU_OPS_ROOT}/yaml/
+    --source-path ${CODEGEN_XPU_YAML_DIR}
     --install-dir ${BUILD_TORCH_XPU_ATEN_GENERATED}
     --per-operator-headers
-    --static-dispatch-backend
     --backend-whitelist XPU SparseXPU SparseCsrXPU NestedTensorXPU
-    # --xpu: generate in-tree RegisterXPU_0.cpp for in-tree OPs
     --xpu
+  )
+
+  set(XPU_INSTALL_HEADER_COMMAND
+    "${PYTHON_EXECUTABLE}" ${TORCH_XPU_OPS_ROOT}/tools/codegen/install_xpu_headers.py
+    --src-header-dir ${BUILD_TORCH_XPU_ATEN_GENERATED}
+    --dst-header-dir ${BUILD_TORCH_ATEN_GENERATED}
+  )
+
+  execute_process(
+    COMMAND
+    ${XPU_CODEGEN_COMMAND}
+    --generate headers
+    --dry-run
+    --output-dependencies ${BUILD_TORCH_XPU_ATEN_GENERATED}/generated_headers.cmake
+    RESULT_VARIABLE RETURN_VALUE
+    WORKING_DIRECTORY ${TORCH_ROOT}
+  )
+
+  if(NOT RETURN_VALUE EQUAL 0)
+    message(FATAL_ERROR "Failed to get generated_headers list")
+  endif()
+
+  execute_process(
+    COMMAND
+    ${XPU_INSTALL_HEADER_COMMAND}
+    --dry-run
+    RESULT_VARIABLE RETURN_VALUE
+    WORKING_DIRECTORY ${TORCH_ROOT}
+  )
+
+  if(NOT RETURN_VALUE EQUAL 0)
+    message(FATAL_ERROR "Failed to get XPU header list to install")
+  endif()
+
+  add_custom_command(
+    COMMENT "Generating XPU ATen Codegen..."
+    OUTPUT ${generated_files}
+    COMMAND
+    ${XPU_CODEGEN_COMMAND}
+    --static-dispatch-backend
     # --update-aoti-c-shim: generate extend/c_shim_xpu.h
     --update-aoti-c-shim
     # --exten-aoti-c-shim: specifiy the extend/c_shim_xpu
@@ -95,16 +99,13 @@ function(GEN_XPU file_yaml)
     COMMAND
     ${REGISTER_FALLBACK_CMD}
     # Codegen post-process
-    COMMAND "${PYTHON_EXECUTABLE}" ${TORCH_XPU_OPS_ROOT}/tools/codegen/remove_headers.py --register_xpu_path ${RegisterXPU_PATH}
-    COMMAND "${PYTHON_EXECUTABLE}" ${TORCH_XPU_OPS_ROOT}/tools/codegen/remove_headers.py --register_xpu_path ${RegisterSparseXPU_PATH}
-    COMMAND "${PYTHON_EXECUTABLE}" ${TORCH_XPU_OPS_ROOT}/tools/codegen/remove_headers.py --register_xpu_path ${RegisterSparseCsrXPU_PATH}
-    COMMAND "${PYTHON_EXECUTABLE}" ${TORCH_XPU_OPS_ROOT}/tools/codegen/remove_headers.py --register_xpu_path ${RegisterNestedTensorXPU_PATH}
-    ${SIMPLE_TRACE}
+    COMMAND
+    ${XPU_INSTALL_HEADER_COMMAND}
     WORKING_DIRECTORY ${TORCH_ROOT}
     DEPENDS
-    ${depended_files}
-    ${TORCH_XPU_OPS_ROOT}/yaml/native/${file_yaml}
-    ${XPUFallback_PATH}
+    ${CODEGEN_XPU_YAML_DIR}/native/${file_yaml}
+    ${XPUFallback_TEMPLATE}
+    ${TORCH_XPU_OPS_ROOT}/tools/codegen/install_xpu_headers.py
   )
 
   # Post codegen delete the copied templates folder only on Windows.
@@ -118,30 +119,25 @@ function(GEN_XPU file_yaml)
   endif()
 endfunction(GEN_XPU)
 
-# GEN_BACKEND(
-#   xpu_functions.yaml
-#   XPUNativeFunctions.h
-#   RegisterXPU_0.cpp)
-
 GEN_XPU(
   native_functions.yaml
   ${BUILD_TORCH_XPU_ATEN_GENERATED}/XPUFunctions.h
-  ${BUILD_TORCH_XPU_ATEN_GENERATED}/RegisterXPU_0.cpp
-  ${BUILD_TORCH_XPU_ATEN_GENERATED}/RegisterSparseXPU_0.cpp
-  ${BUILD_TORCH_XPU_ATEN_GENERATED}/RegisterSparseCsrXPU_0.cpp
-  ${BUILD_TORCH_XPU_ATEN_GENERATED}/RegisterNestedTensorXPU_0.cpp
-  ${XPU_AOTI_INSTALL_DIR}/c_shim_xpu.h
-  ${XPU_AOTI_INSTALL_DIR}/c_shim_xpu.cpp
+  ${BUILD_TORCH_XPU_ATEN_GENERATED}/XPUFunctions_inl.h
+  ${RegisterXPU_GENERATED}
+  ${RegisterSparseXPU_GENERATED}
+  ${RegisterSparseCsrXPU_GENERATED}
+  ${RegisterNestedTensorXPU_GENERATED}
+  ${XPU_AOTI_SHIM_HEADER}
+  ${XPU_AOTI_SHIM_SOURCE}
 )
 
+include(${BUILD_TORCH_XPU_ATEN_GENERATED}/xpu_ops_generated_headers.cmake)
 
-# The c_shim_xpu.cpp needs include files in ${CMAKE_BINARY_DIR}/xpu/ATen/ops/*.h)
-# The include path is auto generated as "#include <ATen/ops/*.h">
-# To follow the design of aoti codegen, here ${CMAKE_BINARY_DIR}/xpu is added to
-# $TORCH_XPU_OPS_INCLUDE_DIRS, so that "#include <ATen/ops/*.h>" works.
-list(APPEND TORCH_XPU_OPS_INCLUDE_DIRS ${CMAKE_BINARY_DIR}/xpu)
-
-list(APPEND xpu_generated_src ${RegisterXPU_PATH} ${RegisterSparseXPU_PATH} ${RegisterSparseCsrXPU_PATH} ${RegisterNestedTensorXPU_PATH})
-list(APPEND xpu_generated_src ${XPU_AOTI_INSTALL_DIR}/c_shim_xpu.cpp)
-add_custom_target(TORCH_XPU_GEN_TARGET DEPENDS ${xpu_generated_src})
+list(APPEND xpu_generated_src
+  ${RegisterXPU_GENERATED}
+  ${RegisterSparseXPU_GENERATED}
+  ${RegisterSparseCsrXPU_GENERATED}
+  ${RegisterNestedTensorXPU_GENERATED}
+  ${XPU_AOTI_SHIM_SOURCE}
+)
 set(ATen_XPU_GEN_SRCS ${xpu_generated_src})
diff --git a/src/ATen/CMakeLists.txt b/src/ATen/CMakeLists.txt
@@ -19,3 +19,7 @@ set(ATen_XPU_SYCL_SRCS ${ATen_XPU_SYCL_SRCS} PARENT_SCOPE)
 foreach(HEADER  ${xpu_h})
   install(FILES ${HEADER} DESTINATION "${AT_INSTALL_INCLUDE_DIR}/ATen/xpu")
 endforeach()
+
+foreach(HEADER  ${xpu_ops_generated_headers})
+  install(FILES ${HEADER} DESTINATION ${AT_INSTALL_INCLUDE_DIR}/ATen/ops)
+endforeach()
diff --git a/src/ATen/native/sparse/xpu/SparseCsrTensorMath.cpp b/src/ATen/native/sparse/xpu/SparseCsrTensorMath.cpp
@@ -1,7 +1,7 @@
 #include <ATen/native/sparse/SparseStubs.h>
 #include <ATen/native/sparse/xpu/sycl/SparseCsrTensorMathKernels.h>
-#include <xpu/ATen/ops/_convert_indices_from_coo_to_csr_native.h>
-#include <xpu/ATen/ops/_convert_indices_from_csr_to_coo_native.h>
+#include <ATen/ops/_convert_indices_from_coo_to_csr_native.h>
+#include <ATen/ops/_convert_indices_from_csr_to_coo_native.h>
 
 namespace at::native {
 

diff --git a/src/ATen/native/xpu/Activation.cpp b/src/ATen/native/xpu/Activation.cpp
@@ -7,9 +7,9 @@
 #include <ATen/native/TensorIterator.h>
 
 #include <ATen/ops/empty_like.h>
-#include <xpu/ATen/ops/empty.h>
-#include <xpu/ATen/ops/gelu_backward_native.h>
-#include <xpu/ATen/ops/gelu_native.h>
+#include <ATen/ops/empty.h>
+#include <ATen/ops/gelu_backward_native.h>
+#include <ATen/ops/gelu_native.h>
 
 #include <ATen/native/xpu/sycl/ActivationEluKernels.h>
 #include <ATen/native/xpu/sycl/ActivationGeluKernel.h>

diff --git a/src/ATen/native/xpu/AdaptiveAveragePooling2d.cpp b/src/ATen/native/xpu/AdaptiveAveragePooling2d.cpp
@@ -7,8 +7,8 @@
 
 #include <ATen/ops/mean.h>
 #include <ATen/ops/zeros_like.h>
-#include <xpu/ATen/ops/_adaptive_avg_pool2d_backward_native.h>
-#include <xpu/ATen/ops/_adaptive_avg_pool2d_native.h>
+#include <ATen/ops/_adaptive_avg_pool2d_backward_native.h>
+#include <ATen/ops/_adaptive_avg_pool2d_native.h>
 
 #include <ATen/native/xpu/sycl/AdaptiveAveragePooling2dKernels.h>
 

diff --git a/src/ATen/native/xpu/AdaptiveAveragePooling3d.cpp b/src/ATen/native/xpu/AdaptiveAveragePooling3d.cpp
@@ -4,8 +4,8 @@
 
 #include <ATen/ops/empty.h>
 #include <ATen/ops/empty_like.h>
-#include <xpu/ATen/ops/adaptive_avg_pool3d_backward_native.h>
-#include <xpu/ATen/ops/adaptive_avg_pool3d_native.h>
+#include <ATen/ops/adaptive_avg_pool3d_backward_native.h>
+#include <ATen/ops/adaptive_avg_pool3d_native.h>
 
 namespace at::native {
 

diff --git a/src/ATen/native/xpu/AdaptiveMaxPooling2d.cpp b/src/ATen/native/xpu/AdaptiveMaxPooling2d.cpp
@@ -4,8 +4,8 @@
 #include <ATen/native/xpu/sycl/AdaptiveMaxPooling2dKernels.h>
 #include <comm/RegisterUtils.h>
 
-#include <xpu/ATen/ops/adaptive_max_pool2d_backward_native.h>
-#include <xpu/ATen/ops/adaptive_max_pool2d_native.h>
+#include <ATen/ops/adaptive_max_pool2d_backward_native.h>
+#include <ATen/ops/adaptive_max_pool2d_native.h>
 
 namespace at {
 namespace native {

diff --git a/src/ATen/native/xpu/AdaptiveMaxPooling3d.cpp b/src/ATen/native/xpu/AdaptiveMaxPooling3d.cpp
@@ -4,8 +4,8 @@
 #include <ATen/native/xpu/sycl/AdaptiveMaxPooling3dKernels.h>
 
 #include <ATen/ops/empty.h>
-#include <xpu/ATen/ops/adaptive_max_pool3d_backward_native.h>
-#include <xpu/ATen/ops/adaptive_max_pool3d_native.h>
+#include <ATen/ops/adaptive_max_pool3d_backward_native.h>
+#include <ATen/ops/adaptive_max_pool3d_native.h>
 
 namespace at {
 namespace native {

diff --git a/src/ATen/native/xpu/AveragePool2d.cpp b/src/ATen/native/xpu/AveragePool2d.cpp
@@ -5,8 +5,8 @@
 #include <ATen/native/xpu/sycl/AveragePool2dKernels.h>
 #include <comm/RegisterUtils.h>
 
-#include <xpu/ATen/ops/avg_pool2d_backward_native.h>
-#include <xpu/ATen/ops/avg_pool2d_native.h>
+#include <ATen/ops/avg_pool2d_backward_native.h>
+#include <ATen/ops/avg_pool2d_native.h>
 
 namespace at {
 namespace native {

diff --git a/src/ATen/native/xpu/AveragePool3d.cpp b/src/ATen/native/xpu/AveragePool3d.cpp
@@ -1,8 +1,8 @@
 #include <ATen/core/Tensor.h>
 #include <ATen/native/xpu/sycl/AveragePool3dKernels.h>
 
-#include <xpu/ATen/ops/avg_pool3d_backward_native.h>
-#include <xpu/ATen/ops/avg_pool3d_native.h>
+#include <ATen/ops/avg_pool3d_backward_native.h>
+#include <ATen/ops/avg_pool3d_native.h>
 
 namespace at {
 namespace native {

diff --git a/src/ATen/native/xpu/BinaryOps.cpp b/src/ATen/native/xpu/BinaryOps.cpp
@@ -4,7 +4,7 @@
 #include <ATen/native/DispatchStub.h>
 #include <ATen/native/TensorIterator.h>
 
-#include <xpu/ATen/ops/add_native.h>
+#include <ATen/ops/add_native.h>
 
 #include <ATen/native/xpu/sycl/BinaryBitwiseOpsKernels.h>
 #include <ATen/native/xpu/sycl/BinaryGeometricKernels.h>

diff --git a/src/ATen/native/xpu/Col2Im.cpp b/src/ATen/native/xpu/Col2Im.cpp
@@ -7,7 +7,7 @@
 #include <ATen/native/xpu/sycl/Col2ImKernel.h>
 
 #include <comm/xpu_aten.h>
-#include <xpu/ATen/ops/col2im_native.h>
+#include <ATen/ops/col2im_native.h>
 
 namespace at::native {
 

diff --git a/src/ATen/native/xpu/DilatedMaxPool2d.cpp b/src/ATen/native/xpu/DilatedMaxPool2d.cpp
@@ -4,9 +4,9 @@
 #include <ATen/native/xpu/sycl/DilatedMaxPool2d.h>
 #include <comm/RegisterUtils.h>
 
-#include <xpu/ATen/ops/max.h>
-#include <xpu/ATen/ops/max_pool2d_with_indices_backward_native.h>
-#include <xpu/ATen/ops/max_pool2d_with_indices_native.h>
+#include <ATen/ops/max.h>
+#include <ATen/ops/max_pool2d_with_indices_backward_native.h>
+#include <ATen/ops/max_pool2d_with_indices_native.h>
 
 namespace at {
 namespace native {

diff --git a/src/ATen/native/xpu/DilatedMaxPool3d.cpp b/src/ATen/native/xpu/DilatedMaxPool3d.cpp
@@ -2,8 +2,8 @@
 #include <ATen/native/xpu/sycl/DilatedMaxPool3d.h>
 
 #include <ATen/ops/empty.h>
-#include <xpu/ATen/ops/max_pool3d_with_indices_backward_native.h>
-#include <xpu/ATen/ops/max_pool3d_with_indices_native.h>
+#include <ATen/ops/max_pool3d_with_indices_backward_native.h>
+#include <ATen/ops/max_pool3d_with_indices_native.h>
 namespace at {
 namespace native {
 

diff --git a/src/ATen/native/xpu/Dropout.cpp b/src/ATen/native/xpu/Dropout.cpp
@@ -3,8 +3,8 @@
 #include <ATen/native/TensorIterator.h>
 #include <ATen/native/xpu/sycl/DropoutKernels.h>
 
-#include <xpu/ATen/ops/native_dropout_backward_native.h>
-#include <xpu/ATen/ops/native_dropout_native.h>
+#include <ATen/ops/native_dropout_backward_native.h>
+#include <ATen/ops/native_dropout_native.h>
 
 #include <comm/xpu_aten.h>
 

diff --git a/src/ATen/native/xpu/Embedding.cpp b/src/ATen/native/xpu/Embedding.cpp
@@ -1,6 +1,6 @@
 #include <ATen/core/op_registration/adaption.h>
 
-#include <xpu/ATen/ops/embedding_dense_backward_native.h>
+#include <ATen/ops/embedding_dense_backward_native.h>
 
 #include <ATen/native/xpu/sycl/EmbeddingKernels.h>
 #include <comm/xpu_aten.h>