Upgrade oneDNN (mkl-dnn) to v1.7 (#47853)

Summary:
Bump oneDNN (mkl-dnn) to 1.7 for bug fixes and performance optimizations
- Fixes https://github.com/pytorch/pytorch/issues/42115. Fixed build issue on Windows for the case when oneDNN is built as submodule
- Fixes https://github.com/pytorch/pytorch/issues/45746. Fixed segmentation fault for convolution weight gradient on systems with Intel AVX512 support

This PR also contains a few changes in ideep for follow-up update (not enabled in current PR yet):
- Performance improvements for the CPU path of Convolution
- Channel-last support

Pull Request resolved: https://github.com/pytorch/pytorch/pull/47853

Reviewed By: bdhirsh

Differential Revision: D25275268

Pulled By: VitalyFedyunin

fbshipit-source-id: 75a589d57e3d19a7f23272a67045ad7494f1bdbe

.jenkins/pytorch/win-test-helpers/installation-helpers/install_mkl.bat

@@ -7,4 +7,4 @@ if "%REBUILD%"=="" (
   7z x -aoa %TMP_DIR_WIN%\mkl.7z -o%TMP_DIR_WIN%\mkl
 )
 set CMAKE_INCLUDE_PATH=%TMP_DIR_WIN%\mkl\include
-set LIB=%TMP_DIR_WIN%\mkl\lib;%LIB
+set LIB=%TMP_DIR_WIN%\mkl\lib;%LIB%

aten/src/ATen/native/CompositeRandomAccessorCommon.h

@@ -129,7 +129,7 @@ public:
 
   // Pointer-like operations {
   C10_HOST_DEVICE
-  reference operator*() {
+  reference operator*() const {
     return TupleInfo::tie(*keys, *values);
   }
 

caffe2/ideep/ideep_utils.h

@@ -1,7 +1,7 @@
 #pragma once
 
 #include <caffe2/core/macros.h>  // For caffe2 macros.
-
+#include <caffe2/utils/eigen_utils.h>
 // All caffe2 ideep related headers
 #include <ideep.hpp>
 #include <caffe2/ideep/utils/ideep_context.h>

third_party/ideep

@@ -1 +1 @@
-Subproject commit ba885200dbbc1f144c7b58eba487378eb324f281
+Subproject commit f0280bb805c2dedd4bb5dd4765cda7dfcd30989f

third_party/mkl-dnn.BUILD

@@ -7,9 +7,9 @@ template_rule(
     out = "include/dnnl_version.h",
     substitutions = {
         "@DNNL_VERSION_MAJOR@": "1",
-        "@DNNL_VERSION_MINOR@": "5",
+        "@DNNL_VERSION_MINOR@": "7",
         "@DNNL_VERSION_PATCH@": "0",
-        "@DNNL_VERSION_HASH@": "e2ac1fac44c5078ca927cb9b90e1b3066a0b2ed0",
+        "@DNNL_VERSION_HASH@": "2e4732679f0211bb311780d0f383cf2dce9baca7",
     },
 )
 
@@ -30,6 +30,8 @@ cc_library(
     srcs = glob([
         "src/common/*.cpp",
         "src/cpu/**/*.cpp",
+    ], exclude=[
+        "src/cpu/aarch64/*.cpp",
     ]),
     hdrs = glob([
         "include/*.h",
@@ -38,7 +40,8 @@ cc_library(
         "src/cpu/**/*.hpp",
         "src/cpu/**/*.h",
         "src/common/*.hpp",
-        "src/cpu/rnn/*.hpp",
+    ], exclude=[
+        "src/cpu/aarch64/*.hpp",
     ]) + [
         "include/dnnl_version.h",
         "include/dnnl_config.h",

torch/csrc/python_headers.h

@@ -1,5 +1,5 @@
 #pragma once
-
+#include <math.h>
 // workaround for Python 2 issue: https://bugs.python.org/issue17120
 // NOTE: It looks like this affects Python 3 as well.
 #pragma push_macro("_XOPEN_SOURCE")