Enable FP16 ACL kernels that accumulate into FP32

[renato-arantes]

Description

The ACL default accumulation for the fp16 kernels is to accumulate into fp16 as well, which causes some tests to fail in PyTorch. This PR enables fp16 ACL kernels that accumulate into fp32 when appropriate and consequently creates a path to fix the failing tests.

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?
• Have you submitted performance data that demonstrates performance improvements?

[jondea]

It's worth saying that we expect a regression here right, because the existing logic uses f16 accumulation? Equally, is there some logic in the benchdnn testing that allows bigger errors for ACL for f16 matmul? If there is, then it should be reverted in this commit too.

[renato-arantes]

It's worth saying that we expect a regression here right, because the existing logic uses f16 accumulation? Equally, is there some logic in the benchdnn testing that allows bigger errors for ACL for f16 matmul? If there is, then it should be reverted in this commit too.

I don't know any logic in benchdnn, including here test files, that contain any calibration to support bigger errors for ACL f16 kernels.

[jondea]

Thanks. So just to check, because benchdnn uses integers to test the matmul primitives, there were no roundoff errors caused by the f16 accumulation. This may be something worth looking into changing at some point (definitely not in the scope of this PR though).

[renato-arantes]

Thanks. So just to check, because benchdnn uses integers to test the matmul primitives, there were no roundoff errors caused by the f16 accumulation.

That is my assumption as well.

This may be something worth looking into changing at some point (definitely not in the scope of this PR though).

Agree.

[Sqvid]

Unfortunately there is nothing in the current test-set that hits this code. Could you add some tests that do? And ideally post some benchdnn lines that activate this path.

[Sqvid]

Waiting for testcases to be added.