cpu: aarch64: enable ACL's inner-product for BF16

[fadara01]

Description

cpu: aarch64: enable ACL's inner-product for BF16

This benchdnn repro with ACL built with ARM-software/ComputeLibrary#1279 now goes to ACL which is > 100x faster:

ONEDNN_VERBOSE=all ./tests/benchdnn/benchdnn --ip --mode=C --dir=FWD_I  --bia-dt=bf16  --dt=bf16 mb1024ic1024oc1024

This PR + ARM-software/ComputeLibrary#1279 fix: pytorch/pytorch#180447

Fixes # (github issue)

Checklist

General

• Do all unit and benchdnn tests (make test and make test_benchdnn_*) pass locally for each commit?
• Have you formatted the code using clang-format?

Performance improvements

• Have you submitted performance data that demonstrates performance improvements?

New features

• Have you published an RFC for the new feature?
• Was the RFC approved?
• Have you added relevant tests?

Bug fixes

• Have you included information on how to reproduce the issue (either in a github issue or in this PR)?
• Have you added relevant regression tests?

RFC PR

• Does RFC document follow the template?
• Have you added a link to the rendered document?

[fadara01]

cc: @jondea @Sqvid

[aditew01]

do you need to upgrade ACL version once this is merged? ARM-software/ComputeLibrary#1279

[fadara01]

do you need to upgrade ACL version once this is merged? ARM-software/ComputeLibrary#1279

Nope, I think we can just merge this without updating ACL version - in this case oneDNN will just bail out a bit later when doing ACL validate.

[aditew01]

Nope, I think we can just merge this without updating ACL version - in this case oneDNN will just bail out a bit later when doing ACL validate.

Yeah, i was referring to this:

This benchdnn repro now goes to ACL which is > 100x faster

[fadara01]

@aditew01 - I see your point, I updated the description.

[fadara01]

I assume we have CI here to test oneDNN main vs ACL main? In which case let's wait for the ACL PR ARM-software/ComputeLibrary#1279 to hit main before we merge this.

[aditew01]

I'm not sure if this extra check is needed here / is covered in ACL.

Suggested change

	&& attr()->has_default_values(smask_t::post_ops, bf16);
	&& attr()->has_default_values(smask_t::post_ops, bf16) && platform::has_data_type_support(data_type::bf16);

[fadara01]

I'm not sure tbh, I think most of these checks are redundant anyways because the same checks will happen by calls to validate / has_opt_impl across ACL interfaces from CpuFullyConnected all the way to arm_gemm.

We have c6g CI here, we can test this in CI once the ACL PR is merged

[jondea]

I assume we have CI here to test oneDNN main vs ACL main

I don't think we have this currently

[fadara01]

I don't think we have this currently

okay, I'll run the nightly test suite locally with both the oneDNN and ACL PRs.

[fadara01]

OK, I built:

• ACL with main + feat: Enable BF16 I/O for CpuFullyConnected in the experimental Operator API ARM-software/ComputeLibrary#1279 and
• oneDNN with main + this PR + above ACL build, with -DDNNL_TEST_SET=NIGHTLY

then I ran the full nightly suite (246 tests) with ctest on Neoverse-V2 and everything passed

100% tests passed, 0 tests failed out of 246

[Sqvid]

@fadara01 If you want this change in the oneDNN 3.12 release then please raise a backport PR.

[fadara01]

@Sqvid - I raised this backport PR against 3.12: #5061