Refine oneMKL Build (#1725)

This pull request refactors the handling of oneMKL-related code in the
XPU backend by improving conditional compilation and build
configuration. The primary changes involve:

- Filter oneMKL kernel files in CMAKE instead of wrapping the entire
file with macros
- Refine link group

Author: CuiYifeng

cmake/ONEMKL.cmake

@@ -19,5 +19,5 @@ set(TORCH_XPU_OPS_ONEMKL_INCLUDE_DIR ${ONEMKL_INCLUDE_DIR})
 
 set(TORCH_XPU_OPS_ONEMKL_LIBRARIES ${ONEMKL_LIBRARIES})
 
-list(INSERT TORCH_XPU_OPS_ONEMKL_LIBRARIES 1 "-Wl,--start-group")
+list(PREPEND TORCH_XPU_OPS_ONEMKL_LIBRARIES "-Wl,--start-group")
 list(APPEND TORCH_XPU_OPS_ONEMKL_LIBRARIES "-Wl,--end-group")

src/ATen/CMakeLists.txt

@@ -7,7 +7,9 @@ file(GLOB xpu_native_cpp "native/xpu/*.cpp" "native/sparse/*.cpp" "native/sparse
 file(GLOB xpu_sycl "native/xpu/sycl/*.cpp" "native/sparse/xpu/sycl/*.cpp" "native/nested/xpu/sycl/*.cpp" "native/transformers/sycl/*.cpp" "native/quantized/sycl/*.cpp")
 
 list(APPEND ATen_XPU_CPP_SRCS ${xpu_cpp})
-list(APPEND ATen_XPU_MKL_SRCS ${xpu_mkl})
+if(USE_ONEMKL_XPU)
+  list(APPEND ATen_XPU_MKL_SRCS ${xpu_mkl})
+endif()
 list(APPEND ATen_XPU_NATIVE_CPP_SRCS ${xpu_native_cpp})
 list(APPEND ATen_XPU_SYCL_SRCS ${xpu_sycl})
 

src/ATen/native/xpu/mkl/BatchLinearAlgebra.cpp

@@ -1,4 +1,3 @@
-#if defined(USE_ONEMKL_XPU)
 #include <ATen/ATen.h>
 #include <ATen/Dispatch.h>
 #include <ATen/ExpandUtils.h>
@@ -538,4 +537,3 @@ void lu_factor_mkl(
 }
 
 } // namespace at::native::xpu
-#endif // USE_ONEMKL_XPU

src/ATen/native/xpu/mkl/SpectralOps.cpp

@@ -1,4 +1,3 @@
-#if defined(USE_ONEMKL_XPU)
 #include <ATen/native/Resize.h>
 #include <ATen/native/SpectralOpsUtils.h>
 #include <ATen/native/xpu/mkl/SpectralOps.h>
@@ -591,4 +590,3 @@ Tensor& _fft_r2c_mkl_out(
 }
 
 } // namespace at::native::xpu
-#endif // USE_ONEMKL_XPU