[TESTS][Vega][WORKAROUND] test_conv_embed_db fails with ROCm 4.4 tuning

[junliume]

[Symptom]
test_conv_embed_db fails with ROCm 4.4 tuning on the following stage:
http://micimaster.amd.com/blue/organizations/jenkins/MLLibs%2FMIOpen/detail/tuning-rocm-4.4-fordev/12/pipeline
http://micimaster.amd.com/blue/organizations/jenkins/MLLibs%2FMIOpen/detail/tuning-rocm-4.4-fordev/13/pipeline
http://micimaster.amd.com/blue/organizations/jenkins/MLLibs%2FMIOpen/detail/tuning-rocm-4.4-fordev/14/pipeline

• Fp32 Hip Debug Embedded Vega20

[Root Cause]
unknown yet, not reproducible on local tests

[Plan]
Workaround to unblock performance tuning updates

• test_conv_embed_db DISABLED on all VEGA platforms

[atamazov]

It seems like this issue is an "extension" of #874.

The CI logs show that some configs for ConvHipImplicitGemmV4R4Fwd are missing from the perf-db. The proper workaround should only disable this solver for the failing test cases instead of disabling the whole test (see #875 for example)

[junliume]

It seems like this issue is an "extension" of #874.

The CI logs show that some configs for ConvHipImplicitGemmV4R4Fwd are missing from perf-db. The proper workaround should only disable this solver for the failing test cases instead of disabling the whole test (see #875 for example)

The anticipation is that we will get this issue fixed within the next two weeks

[atamazov]

Note that if ConvHipImplicitGemmV4R4Fwd configs are really missing from perf-db (i.e. this issue is not specific to the "embedded" configuration of the library), then we most likely have performance drop. Why: CI logs show that this Solver is the fastest one, and tuning can make it faster.

[junliume]

Note that if ConvHipImplicitGemmV4R4Fwd configs are really missing from perf-db (i.e. this issue is not specific to the "embedded" configuration of the library), then we most likely have performance drop. Why: CI logs show that this Solver is the fastest one, and tuning can make it faster.

@JehandadKhan hopefully we can fix this issue with the coming 4.5 tuning?

[JehandadKhan]

@atamazov and @junliume I did tune the solver and add those entries to the db however, was unable to fix the issue when the test runs in the CI. I need to investigate further and figure out if I am doing the right thing.

The tuning PR going through staging would inform us whether there is perf regression elsewhere or not, if discovered we can cover it there.