Support IPEX ABI=0 build (#4910)

Author: guangyey

cmake/BuildFlags.cmake

@@ -13,12 +13,9 @@ if(BUILD_MODULE_TYPE STREQUAL "GPU")
 endif()
 
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${TORCH_CXX_FLAGS}")
-  if(NOT WINDOWS)
-    string(REGEX MATCH "-D_GLIBCXX_USE_CXX11_ABI=([0-9]+)" torch_cxx11 ${TORCH_CXX_FLAGS})
-    set(GLIBCXX_USE_CXX11_ABI ${CMAKE_MATCH_1})
-    if(BUILD_WITH_XPU AND NOT GLIBCXX_USE_CXX11_ABI)
-      message(FATAL_ERROR "Must set _GLIBCXX_USE_CXX11_ABI=1 for XPU build, but not is ${GLIBCXX_USE_CXX11_ABI}!")
-  endif()
+if(NOT WINDOWS)
+  string(REGEX MATCH "-D_GLIBCXX_USE_CXX11_ABI=([0-9]+)" torch_cxx11 ${TORCH_CXX_FLAGS})
+  set(GLIBCXX_USE_CXX11_ABI ${CMAKE_MATCH_1})
 endif()
 
 # Since 2016 Debian start using RUNPATH instead of normally RPATH, which gave the annoy effect that

cmake/Modules/FindoneDNN.cmake

@@ -106,7 +106,15 @@ ELSE()
   SET(DNNL_ENABLE_CONCURRENT_EXEC TRUE CACHE BOOL "multi-thread primitive execution" FORCE)
   SET(DNNL_EXPERIMENTAL TRUE CACHE BOOL "use one pass for oneDNN BatchNorm" FORCE)
 
+  IF(LINUX)
+    set(_original_cmake_cxx_flags "${CMAKE_CXX_FLAGS}")
+    # Workaround: when use icpx to build oneDNN, -fpreview-breaking-changes has to be explicitly passed.
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fpreview-breaking-changes")
+  ENDIF()
   ADD_SUBDIRECTORY(${ONEDNN_ROOT} oneDNN EXCLUDE_FROM_ALL)
+  IF(LINUX)
+    set(CMAKE_CXX_FLAGS "${_original_cmake_cxx_flags}")
+  ENDIF()
   SET(ONEDNN_LIBRARY ${DNNL_LIBRARY_NAME})
   IF(NOT TARGET ${ONEDNN_LIBRARY})
     MESSAGE(FATAL_ERROR "Failed to include oneDNN target")

cmake/gpu/SYCL.cmake

@@ -101,6 +101,11 @@ set(IPEX_SYCL_KERNEL_FLAGS ${IPEX_SYCL_KERNEL_FLAGS} -fhonor-nans)
 set(IPEX_SYCL_KERNEL_FLAGS ${IPEX_SYCL_KERNEL_FLAGS} -fhonor-infinities)
 set(IPEX_SYCL_KERNEL_FLAGS ${IPEX_SYCL_KERNEL_FLAGS} -fno-associative-math)
 set(IPEX_SYCL_KERNEL_FLAGS ${IPEX_SYCL_KERNEL_FLAGS} -fno-approx-func)
+# It informs compiler to link the ABI-neutral library.
+if(LINUX)
+  set(IPEX_SYCL_KERNEL_FLAGS ${IPEX_SYCL_KERNEL_FLAGS} -fpreview-breaking-changes)
+  set(IPEX_SYCL_LINK_FLAGS ${IPEX_SYCL_LINK_FLAGS} -fpreview-breaking-changes)
+endif()
 
 # Explicitly limit the index range (< Max int32) in kernel
 # set(IPEX_SYCL_KERNEL_FLAGS ${IPEX_SYCL_KERNEL_FLAGS} -fsycl-id-queries-fit-in-int)

csrc/gpu/aten/operators/xetla/kernels/compile_library.cmake

@@ -6,6 +6,9 @@ function(add_library_with_options TARGET USE_DOUBLE_GRF AOT_DEVLIST)
     -fsycl-device-code-split=per_kernel
     -fsycl-max-parallel-link-jobs=${SYCL_MAX_PARALLEL_LINK_JOBS}
   )
+  if(LINUX)
+    set(XETLA_KERNEL_FLAGS ${XETLA_KERNEL_FLAGS} -fpreview-breaking-changes)
+  endif()
 
   if (AOT_DEVLIST)
     set(XETLA_KERNEL_FLAGS ${XETLA_KERNEL_FLAGS} -fsycl-targets=spir64_gen)
@@ -67,6 +70,9 @@ function(add_library_with_options TARGET USE_DOUBLE_GRF AOT_DEVLIST)
 
   target_link_options(${TARGET} PRIVATE ${XETLA_KERNEL_FLAGS})
   target_compile_options(${TARGET} PRIVATE -fsycl)
+  if(LINUX)
+    target_compile_options(${TARGET} PRIVATE -fpreview-breaking-changes)
+  endif()
   if (AOT_DEVLIST)
     target_compile_options(${TARGET} PRIVATE -fsycl-targets=spir64_gen)
   endif()

docs/tutorials/known_issues.md

@@ -21,7 +21,7 @@ Troubleshooting
   - **Cause**: DPC++ does not support `_GLIBCXX_USE_CXX11_ABI=0`, Intel® Extension for PyTorch\* is always compiled with `_GLIBCXX_USE_CXX11_ABI=1`. This symbol undefined issue appears when PyTorch\* is
     compiled with `_GLIBCXX_USE_CXX11_ABI=0`.
   - **Solution**: Pass `export GLIBCXX_USE_CXX11_ABI=1` and compile PyTorch\* with particular compiler which supports `_GLIBCXX_USE_CXX11_ABI=1`. We recommend using prebuilt wheels
-    in [download server](https:// developer.intel.com/ipex-whl-stable-xpu) to avoid this issue.
+    in [download server](https:// developer.intel.com/ipex-whl-stable-xpu) to avoid this issue. Please **note** that starting from Intel® Extension for PyTorch\* 2.5, it supports both new and old C++ ABIs. Previously, it only supported the new C++ ABI.
 - **Problem**: `-997 runtime error` when running some AI models on Intel® Arc™ A-Series GPUs.
   - **Cause**:  Some of the `-997 runtime error` are actually out-of-memory errors. As Intel® Arc™ A-Series GPUs have less device memory than Intel® Data Center GPU Flex Series 170 and Intel® Data Center GPU
     Max  Series, running some AI models on them may trigger out-of-memory errors and cause them to report failure such as `-997 runtime error` most likely. This is expected. Memory usage optimization is a work in progress to allow Intel® Arc™ A-Series GPUs to support more AI models.

examples/gpu/inference/cpp/example-usm/CMakeLists.txt

@@ -7,5 +7,8 @@ set(target example-usm)
 add_executable(${target} example-usm.cpp)
 target_link_libraries(${target} ${TORCH_IPEX_LIBRARIES})
 list(APPEND CMAKE_CXX_FLAGS "-fsycl")
+if(LINUX)
+  list(APPEND CMAKE_CXX_FLAGS " -fpreview-breaking-changes")
+endif()
 
 set_property(TARGET ${target} PROPERTY CXX_STANDARD 17)

intel_extension_for_pytorch/xpu/cpp_extension.py

@@ -33,6 +33,7 @@
 CLIB_EXT = ".dll" if IS_WINDOWS else ".so"
 SHARED_FLAG = "/DLL" if IS_WINDOWS else "-shared"
 SYCL_FLAG = "-fsycl"
+ABI_FLAG = "-fpreview-breaking-changes" if IS_LINUX else ""
 
 MINIMUM_GCC_VERSION = (5, 0, 0)
 MINIMUM_MSVC_VERSION = (19, 0, 24215)
@@ -321,7 +322,7 @@ def _gen_link_lib_cmd_line(
 
             libraries_args = []
             libraries_args += [f"-l{x}" for x in libraries]
-            common_args = [SHARED_FLAG] + [SYCL_FLAG]
+            common_args = [SHARED_FLAG] + [SYCL_FLAG] + [ABI_FLAG]
 
             """
             link command formats:
@@ -452,7 +453,7 @@ def unix_wrap_ninja_link_shared_object(
             # create output directories avoid linker error.
             create_parent_dirs_by_path(output_libname)
 
-            ldflags = [SHARED_FLAG] + [SYCL_FLAG]
+            ldflags = [SHARED_FLAG] + [SYCL_FLAG] + [ABI_FLAG]
             ldflags += [f"-L{x}" for x in library_dirs]
             ldflags += [f"-L{x}" for x in runtime_library_dirs]
             ldflags += extra_postargs
@@ -510,7 +511,7 @@ def spawn(cmd):
 
                         if "-fPIC" in cflags:  # Windows does not support this argument
                             cflags.remove("-fPIC")
-                        cflags = cflags + ["-std=c++17", SYCL_FLAG]
+                        cflags = cflags + ["-std=c++17", SYCL_FLAG, ABI_FLAG]
 
                         cmd = [_bin, "-c", src, "-o", obj] + include_list + cflags
                     elif isinstance(self.cflags, dict):
@@ -1004,9 +1005,11 @@ def library_paths() -> List[str]:
 def _prepare_compile_flags(extra_compile_args):
     if isinstance(extra_compile_args, List):
         extra_compile_args.append(SYCL_FLAG)
+        extra_compile_args.append(ABI_FLAG)
     elif isinstance(extra_compile_args, dict):
         cl_flags = extra_compile_args.get("cxx", [])
         cl_flags.append(SYCL_FLAG)
+        cl_flags.append(ABI_FLAG)
         extra_compile_args["cxx"] = cl_flags
 
     return extra_compile_args
@@ -1048,7 +1051,7 @@ def _prepare_ldflags(extra_ldflags, verbose, is_standalone):
     oneapi_link_args += ["-Wl,--start-group"]
     oneapi_link_args += [f"{x}" for x in get_one_api_help().get_onemkl_libraries()]
     oneapi_link_args += ["-Wl,--end-group"]
-    oneapi_link_args += ["-lsycl", "-lOpenCL", "-lpthread", "-lm", "-ldl"]
+    oneapi_link_args += ["-lOpenCL", "-lpthread", "-lm", "-ldl"]
 
     # Append IPEX link parameters.
     oneapi_link_args += ["-lintel-ext-pt-gpu"]
@@ -1244,7 +1247,12 @@ def object_file_path(source_file: str) -> str:
         return target
 
     objects = [object_file_path(src) for src in sources]
-    ldflags = ([] if is_standalone else [SHARED_FLAG]) + [SYCL_FLAG] + extra_ldflags
+    ldflags = (
+        ([] if is_standalone else [SHARED_FLAG])
+        + [SYCL_FLAG]
+        + [ABI_FLAG]
+        + extra_ldflags
+    )
 
     # The darwin linker needs explicit consent to ignore unresolved symbols.
     if IS_MACOS: