Add a script for generating the quick-tuning perfconfigs list

[djramic]

This PR includes a script for finding quick-tuning perfconfigs in the following way:

• For each tuning problem, it finds the top n perfcofnigs based on a given threshold
• Using a linear programming model, it identifies the smallest subset of perfcofnigs that covers all problems.

Additionally, there are option to automatically generated the QuickTuningPerfconfigs.inc file that contains selected perfconfigs, and its integration into the codebase.

closes : ROCm/rocMLIR-internal#1641
closes : ROCm/rocMLIR-internal#1258
closes : ROCm/rocMLIR-internal#1518

[codecov]

Codecov Report

Attention: Patch coverage is 74.00000% with 13 lines in your changes missing coverage. Please review.

Project coverage is 77.70%. Comparing base (5f51701) to head (738bbaf).
Report is 15 commits behind head on develop.

Files with missing lines	Patch %	Lines
...lir/lib/Dialect/Rock/Tuning/GridwiseGemmParams.cpp	74.00%	11 Missing and 2 partials ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #1689      +/-   ##
===========================================
- Coverage    77.76%   77.70%   -0.07%     
===========================================
  Files          100      100              
  Lines        27866    27897      +31     
  Branches      4063     4072       +9     
===========================================
+ Hits         21671    21678       +7     
- Misses        4540     4552      +12     
- Partials      1655     1667      +12     

Flag	Coverage Δ
mfma	77.70% <74.00%> (-0.07%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[dhernandez0]

this would apply to both f8 and i8 here, is that the goal?

[djramic]

Here I 've just add select conv or gemm list based on opType. I think we're using non-accel configs for fp8

[dhernandez0]

There are fp8 mfmas and i8 mfmas, so here we are using the same params for both. Is that expected? or was this tuned for i8 only?

[djramic]

This is tuned only for the i8. When I run tuning for fp8, I get this output:
Tuning: -t fp8 -out_datatype f32 -transA false -transB false -g 3 -m 1024 -n 512 -k 769
v2:64,32,32,4,2,2,1 9.187230e+05
v2:64,32,32,4,2,4,1 7.604820e+05
v2:64,32,32,4,4,2,1 8.972820e+05
...
where it seems that the perfconfigs are in the format for non-xdlops

[dhernandez0]

Running this command (gfx942):

./bin/rocmlir-gen -operation gemm -t fp8 --arch gfx942:sramecc+:xnack- --num_cu 304 -g 1 -m 1 -k 768 -n 768 -transA=False -transB=False --kernel-repeats 1 --perf_config= | ./bin/rocmlir-tuning-driver --tuning-space=exhaustive

I get this:

v2:16,16,4,16,16,8,1,1,1        1.952500e+04
v2:16,16,4,16,16,8,3,1,1        5.493000e+03
v2:16,16,4,16,16,8,4,1,1        3.728000e+03
v2:16,16,4,16,16,16,1,1,1       8.179000e+03
v2:16,16,4,16,16,16,3,1,1       3.448000e+03
v2:16,16,4,16,16,16,4,1,1       2.887000e+03
v2:16,16,8,16,16,8,1,1,1        7.617000e+03
v2:16,16,8,16,16,8,3,1,1        1.128990e+05
...

We get 9 numbers which corresponds to xdlops.
What command are you running?

[dhernandez0]

So I think we either tune for f8 and bf8 or use the default params? So, instead of checking dataTypeA.getIntOrFloatBitWidth() we need to check for the actual type.

[djramic]

I ran python3 tuningRunner.py --op gemm --data-type fp8 -c toy-gemm-configs -d on gfx90a, but I get the same result with the rocmlir-tuning-driver on gfx90a. It seems that f8 mfma is supported only on gfx942?
As our quick-tuning list on develop is outdated, I think we should merge this first, and then investigate f8 in a new ticket?

[dhernandez0]

I think it's supported in navi4 as well. Ok, that's fair, let's merge this and open a new ticket for fp8 and bf8.

[causten]

@dhernandez0 ready for another review

[dhernandez0]

there are some comments not addressed in the review