Aarch64 jit depthwise convolution kernels

[nishith-fujitsu]

Description

This commit expands ARM SVE support for forward and backward JIT SVE Depth-wise convolution in FP32 for, introducing compatibility with various vector lengths. The changes made are for implementing different ARM ISA.
Major code changes:

1. Added common files jit_uni_dw_convolution.cpp, jit_uni_dw_convolution.hpp, jit_uni_dw_conv_kernel_utils.hpp, jit_uni_dw_conv_kernel_f32.cpp, jit_uni_dw_conv_kernel_f32.hpp to accommodate the extended ARM SVE ISA for depth-wise convolution operators.
2. Set data format tags according to the ARM ISA being used for forward and backward operator in JIT depth-wise convolution.
3. Replaced ldr, and str instructions for vector registers with ld1w and st1w to utilize predication.
4. Made changes to reducer.cpp to support Vector agnostic approach.

Checklist

General

• [✓] Do all unit and benchdnn tests (make test and make test_benchdnn_*) pass locally for each commit? Yes
  Test output is same with and without this commit.

99% tests passed, 2 tests failed out of 196
 
Total Test time (real) = 525.25 sec
 
The following tests FAILED:
        155 - test_graph_unit_dnnl_large_partition_usm_cpu (Failed)
        177 - test_benchdnn_modeC_graph_ci_cpu (Failed)
Errors while running CTest
Output from these tests are in: /home/nishith/oss/oneDNN/build/Testing/Temporary/LastTest.log
Use "--rerun-failed --output-on-failure" to re-run the failed cases verbosely.
make: *** [Makefile:71: test] Error 8

• [ ✓] Have you formatted the code using clang-format? Yes

[abhijain1204fujitsu]

@vpirogov , Can you please share your feedback for the PR & Support us for merger

[vpirogov]

@oneapi-src/onednn-cpu-aarch64, @snadampal, @kawakami-k, could you please help reviewing and validating this one?

[jondea]

This is really great work, thank you!

My only important comment is that I think it is better than the current ACL depthwise implementation and it should go ahead of it in the dispatch list.

I have also done some testing with this on a C7g with benchdnn and with PyTorch, and I didn't hit any issues.

[jondea]

Code generally looks good, but can you remove the merge commit from the history by rebasing onto main please? https://github.com/oneapi-src/oneDNN/blob/main/CONTRIBUTING.md#code-contribution-guidelines

[abhijain1204fujitsu]

@vpirogov , @jondea the requested changes has been completed, kindly check and support to merge the PR

[abhijain1204fujitsu]

@vpirogov kindly support to merge the changes
In case there is any more feedback kindly let us know.
Thank you !

[mgouicem]

I believe a missing closing brace should make that code break the build on ARM platforms.

[mgouicem]

please remove the merge commit. Instead rebase your branch on top of the main branch.

[nishith-fujitsu]

please remove the merge commit. Instead rebase your branch on top of the main branch.

Hi @mgouicem, removed merge commits and rebased to latest main branch.

[abhijain1204fujitsu]

@vpirogov , @mgouicem please support to merge the PR