Bnorm tests are mistrusted on aarch64

[Ryo-not-rio]

Summary

When running ./tests/benchdnn/benchdnn -v5 --bnorm --tag=axb --dir=FWD_I --flags=G,GR --attr-post-ops=relu:0.5,linear:3.4:5.0+tanh --inplace=true,false mb1ic19id4ih6iw6nbnorm_3d:2 on aarch64, 2 of the 8 tests are mistrusted.

The same occurs when running with --skip-impl=acl with the simple:any kernels being run, which suggests the cause to be outside of acl.

Version

v3.6.0 (commit 094cc1d)

Environment

oneDNN includes hardware-specific optimizations and may behave
differently on depending on the compiler and build environment. Include
the following information to help reproduce the issue:

• CPU make and model aarch64 graviton
  flags: fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp asimdhp cpuid asimdrdm jscvt fcma lr
  cpc dcpop sha3 sm3 sm4 asimddp sha512 sve asimdfhm dit uscat ilrcpc flagm ssbs paca pacg dcp
  odp svei8mm svebf16 i8mm bf16 dgh rng
• OS version - 20~22.04.1-Ubuntu
• Compiler version - 11.4.0

Steps to reproduce

Run ./tests/benchdnn/benchdnn -v5 --bnorm --tag=axb --dir=FWD_I --flags=G,GR --attr-post-ops=relu:0.5,linear:3.4:5.0+tanh --inplace=true,false mb1ic19id4ih6iw6nbnorm_3d:2

Observed behavior

tests:8 passed:6 skipped:0 mistrusted:2 unimplemented:0 invalid_arguments:0 failed:0 listed:0

Expected behavior

tests:8 passed:8 skipped:0 mistrusted:0 unimplemented:0 invalid_arguments:0 failed:0 listed:0

[dzarukin]

Hi @Ryo-not-rio, this is rather a benchdnn filling peculiarity rather than an issue. The mistrust status just tips that there are more zeros in the output than it should, but the benchmark decides what the number of zeros is expected. It can be treated as passed, it's a suggestion for developers to look into the problem as 100% of zeros can be in the output and it would be impossible to say if the test actually verifies the functionality.