MVN test extensions

[chenhu-wang]

Details:

• create MVN1+FQ+MVN2 subgraph test. This pattern is transformed to MVN1_fused_with_quantize + MVN2_fused_with_dequantize. MVN1_fused_with_quantize will produce and store lp data. MVN2_fused_with_dequantize will load lp data. This is to cover load/store emitters which is responsible data load/store in MVN. Meanwhile this can cover round type of precision convert in MVN_fused_with_quantize

Tickets:

• 56733
• 57688 test cover

[dmitry-gorokhov]

@antonvor could you please review? We also need to understand do we have better way to validate MVN with fused FQ and i8/u8 output precisions.

[antonvor]

@dmitry-gorokhov yes, the functionality here #6546 allows you to test similar cases. We can not write a separate subgraph test, but expand the existing fusing test.

[antonvor]

LGTM. I think it is possible to try to extend the existing test using mechanism in #6546. Perhaps there will be comments and we will improve it. @chenhu-wang and @dmitry-gorokhov what do you think?

[chenhu-wang]

LGTM. I think it is possible to try to extend the existing test using mechanism in #6546. Perhaps there will be comments and we will improve it. @chenhu-wang and @dmitry-gorokhov what do you think?

1. This is a subgraph(x + FQ + MVN) that reproduce and cover bug in real model.
2. Currently MVN reference implementation do not have best accuracy for low precision data calculation. With MVN(fp32) + FQ (fp32) + MVN(fp32), ng ref can return accurate result with fp32 execution and pass the comparison, and meanwhile CPU plugin is transformed executed with low precision.
   Above is my thought to keep this test.

[dmitry-gorokhov]

Outdated PR