PR #77927: [oneDNN] upgrading oneDNN version to 3.6

Imported from GitHub PR tensorflow/tensorflow#77927

This PR upgrades oneDNN version from v3.5 to v3.6, this PR has been tested on several models across different platforms including cascade-lake, sapphire-rapids, and granite-rapids

Several bug fixes have been resolved in this version. Details can be found here https://github.com/oneapi-src/oneDNN/releases
Copybara import of the project:

--
748ad53782ed53252ce154ba18e8259ca6a282c3 by Ashiq Imran <[email protected]>:

upgrading oneDNN version to 3.6

--
ed9992b5c01483118639e012c8f7fe161e964aae by Ashiq Imran <[email protected]>:

adding suggested change

--
367c8c8ba54b6c89852c0829cd918178837b0039 by Ashiq Imran <[email protected]>:

Update mkldnn_v1.BUILD
--
232890c357ab49da3dac93bd91b71fa1cb6bd2a6 by Ashiq Imran <[email protected]>:

Update mkldnn_v1.BUILD

adding spdlog into the include and textual hdrs list

Merging this change closes #77927

FUTURE_COPYBARA_INTEGRATE_REVIEW=tensorflow/tensorflow#77927 from Intel-tensorflow:aimran/oneDNN_3.6 232890c357ab49da3dac93bd91b71fa1cb6bd2a6
PiperOrigin-RevId: 696254899

third_party/tsl/third_party/mkl_dnn/mkldnn_v1.BUILD

@@ -6,15 +6,19 @@ exports_files(["LICENSE"])
 
 _CMAKE_COMMON_LIST = {
     "#cmakedefine DNNL_GPU_RUNTIME DNNL_RUNTIME_${DNNL_GPU_RUNTIME}": "#define DNNL_GPU_RUNTIME DNNL_RUNTIME_NONE",
+    "#cmakedefine DNNL_GPU_VENDOR DNNL_VENDOR_${DNNL_GPU_VENDOR}": "#define DNNL_VENDOR_NONE",
     "#cmakedefine DNNL_USE_RT_OBJECTS_IN_PRIMITIVE_CACHE": "#undef DNNL_USE_RT_OBJECTS_IN_PRIMITIVE_CACHE",
     "#cmakedefine DNNL_WITH_SYCL": "#undef DNNL_WITH_SYCL",
+    "#cmakedefine DNNL_SYCL_GENERIC": "#define DNNL_SYCL_GENERIC",
     "#cmakedefine DNNL_WITH_LEVEL_ZERO": "#undef DNNL_WITH_LEVEL_ZERO",
     "#cmakedefine DNNL_SYCL_CUDA": "#undef DNNL_SYCL_CUDA",
     "#cmakedefine DNNL_SYCL_HIP": "#undef DNNL_SYCL_HIP",
     "#cmakedefine DNNL_ENABLE_STACK_CHECKER": "#undef DNNL_ENABLE_STACK_CHECKER",
     "#cmakedefine ONEDNN_BUILD_GRAPH": "#define ONEDNN_BUILD_GRAPH",
+    "#cmakedefine DNNL_DISABLE_GPU_REF_KERNELS": "#define DNNL_DISABLE_GPU_REF_KERNELS",
     "#cmakedefine DNNL_EXPERIMENTAL_SPARSE": "#define DNNL_EXPERIMENTAL_SPARSE",
     "#cmakedefine DNNL_EXPERIMENTAL": "#undef DNNL_EXPERIMENTAL",
+    "#cmakedefine DNNL_EXPERIMENTAL_LOGGING": "#undef DNNL_EXPERIMENTAL_LOGGING",
     "#cmakedefine01 BUILD_TRAINING": "#define BUILD_TRAINING 1",
     "#cmakedefine01 BUILD_INFERENCE": "#define BUILD_INFERENCE 0",
     "#cmakedefine01 BUILD_PRIMITIVE_ALL": "#define BUILD_PRIMITIVE_ALL 1",
@@ -51,6 +55,7 @@ _CMAKE_COMMON_LIST = {
     "#cmakedefine01 BUILD_PRIMITIVE_GPU_ISA_ALL": "#define BUILD_PRIMITIVE_GPU_ISA_ALL 0",
     "#cmakedefine01 BUILD_GEN9": "#define BUILD_GEN9 0",
     "#cmakedefine01 BUILD_GEN11": "#define BUILD_GEN11 0",
+    "#cmakedefine01 BUILD_SDPA": "#define BUILD_SDPA 1",
     "#cmakedefine01 BUILD_XE2": "#define BUILD_XE2 0",
     "#cmakedefine01 BUILD_XELP": "#define BUILD_XELP 0",
     "#cmakedefine01 BUILD_XEHPG": "#define BUILD_XEHPG 0",
@@ -94,13 +99,21 @@ expand_template(
     out = "include/oneapi/dnnl/dnnl_version.h",
     substitutions = {
         "@DNNL_VERSION_MAJOR@": "3",
-        "@DNNL_VERSION_MINOR@": "5",
+        "@DNNL_VERSION_MINOR@": "6",
         "@DNNL_VERSION_PATCH@": "0",
-        "@DNNL_VERSION_HASH@": "N/A",
     },
     template = "include/oneapi/dnnl/dnnl_version.h.in",
 )
 
+expand_template(
+    name = "dnnl_version_hash_h",
+    out = "include/oneapi/dnnl/dnnl_version_hash.h",
+    substitutions = {
+        "@DNNL_VERSION_HASH@": "N/A",
+    },
+    template = "include/oneapi/dnnl/dnnl_version_hash.h.in",
+)
+
 _COPTS_LIST = select({
     "@xla//xla/tsl:windows": [],
     "//conditions:default": ["-fexceptions"],
@@ -117,6 +130,11 @@ _INCLUDES_LIST = [
     "src",
     "src/common",
     "src/common/ittnotify",
+    "src/common/spdlog",
+    "src/common/spdlog/details",
+    "src/common/spdlog/fmt",
+    "src/common/spdlog/fmt/bundled",
+    "src/common/spdlog/sinks",
     "src/cpu",
     "src/cpu/gemm",
     "src/cpu/x64/xbyak",
@@ -127,6 +145,10 @@ _TEXTUAL_HDRS_LIST = glob([
     "include/**/*",
     "src/common/*.hpp",
     "src/common/ittnotify/**/*.h",
+    "src/common/spdlog/*.h",
+    "src/common/spdlog/details/*.h",
+    "src/common/spdlog/fmt/**/*.h",
+    "src/common/spdlog/sinks/*.h",
     "src/cpu/*.hpp",
     "src/cpu/**/*.hpp",
     "src/cpu/jit_utils/**/*.hpp",
@@ -143,6 +165,7 @@ _TEXTUAL_HDRS_LIST = glob([
 ]) + [
     ":dnnl_config_h",
     ":dnnl_version_h",
+    ":dnnl_version_hash_h",
 ]
 
 # Large autogen files take too long time to compile with usual optimization
@@ -186,6 +209,7 @@ cc_library(
             "src/cpu/aarch64/**",
             "src/cpu/rv64/**",
             "src/cpu/x64/gemm/**/*_kern_autogen.cpp",
+            "src/cpu/sycl/**",
         ],
     ),
     copts = _COPTS_LIST,