Improve baseline comparison logic

[vedithal-amd]

rocprofiler-compute Pull Request

Related Issue

• Closes #

What type of PR is this? (check all that apply)

• Bug Fix
• Cherry Pick
• Continuous Integration
• Documentation Update
• Feature
• Optimization
• Refactor
• Other (please specify)

Technical Details

• Do not force unsupported metrics to be specified in older gpu
  architectures as None

• Remove metrics which are explicitly set to None

• Update CHANGELOG

• Fix analysis configuration to fix baseline comparisons across all gpu
  architectures

  • Add missing 1812 section for gfx908
  • Add missing 1812 section for gfx90a

We cannot guarantee common Metric ID index for metrics comparison across different GPU models.
The Metric ID is not very useful anyways and this is how it looks with and without Metric ID

With Metric ID

Without Metric ID

Have you added or updated tests to validate functionality?

• Yes
• No - does not apply to this PR
  Validated baseline comparison working across all gpu architectures

Added / Updated documentation?

• Yes
• No - does not apply to this PR

Have you updated CHANGELOG?

• Yes
• No - does not apply to this PR

[feizheng10]

1. Please separate gfx906 deprecation out of this pr.
2. Keep block IDs is important. When we are talking about "baseline comparison", the 1st run is used as baseline reference, which implies all the info from the 1st run are the baseline including block IDs.

[feizheng10]

And just for internal maintain/design purpose, we have meta info in the comments(temporally) for the reasons why we don't have the metrics available. It is better to provide solution when removing those.

[vedithal-amd]

1st run is used as baseline reference, which implies all the info from the 1st run are the baseline including block IDs.

My concern is, does it make sense to compare a metric when its only available in one run?
Lets say 1st run has Metric1 = 10, Metric2 = 10 and 2nd run does not have Metric1 but Metric2 = 5.
In above case, does it make sense to only compare Metric2 which is common?
Why will the user want to see Metric1 in baseline comparison when it is not there in run 2? Why cant the user just do analysis of run 1 to see Metric1?

Please separate gfx906 deprecation out of this pr.

OK

And just for internal maintain/design purpose, we have meta info in the comments(temporally) for the reasons why we don't have the metrics available. It is better to provide solution when removing those.

We can backport the metrics from MI350 config to older GPUs when someone ask for it in a different PR.
There is no point in having None metrics when you can look at the metric information in newer MI350 GPU config.

[vedithal-amd]

Please separate gfx906 deprecation out of this pr.

This will be tracked in #819

[vedithal-amd]

Keep block IDs is important.

There is no such thing as block IDs, I think you are referring to Metric ID and i dont think that adds any value to the analysis report. Metric ID will still be shown in non baseline comparison mode. Only in baseline comparison mode its not able to show Metric ID due to metrics being different.

[vedithal-amd]

As discussed on call, we can add Metric ID from first workload during baseline comparison.