Add i8 in DATA_TYPES_ATTENTION

[dorde-antic]

Datatype i8 was missing in DATA_TYPES_ATTENTION in perfRunner.py

Resolves https://github.com/ROCm/rocMLIR-internal/issues/1836

[dhernandez0]

we don't have CI for perfRunner I think. Please, could you test this works manually?

[umangyadav]

LGTM but test locally and report back before merging

[dorde-antic]

@umangyadav @dhernandez0
Not sure if this fails because of the config and not because of the change that was made in this PR:

[dhernandez0]

Not sure if this fails because of the config and not because of the change that was made in this PR:

it looks like it fails f32, did you run this on a navi machine? that could be the reason.

Please, try i8 attention.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #1838      +/-   ##
===========================================
- Coverage    78.52%   78.05%   -0.47%     
===========================================
  Files          100      100              
  Lines        29907    30175     +268     
  Branches      4452     4675     +223     
===========================================
+ Hits         23482    23551      +69     
- Misses        4590     4602      +12     
- Partials      1835     2022     +187     

Flag	Coverage Δ
mfma	?
navi3x	78.01% <ø> (?)
navi4x	78.05% <ø> (?)

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

see 40 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[dhernandez0]

I understand the test was not done and it fails? Did we check that’s not the case before merging? @umangyadav @dorde-antic

[umangyadav]

I ran following on Navi3x and it passed and CI was green. But i have not ran it through perfRunner.py though. @dorde-antic would it be possible for you to check later ?

 rocmlir-gen -ph -operation attention -t i8 --arch gfx1100 --num_cu 48 -g 1 -seq_len_q 16384 -seq_len_k 16384 -head_dim_v 512 -head_dim_qk 512 -with-attn-scale=False -transQ=False -transK=True -transV=False -transO=False --kernel-repeats 5 | /home/umayadav/repo/rocMLIR/build/bin/rocmlir-driver -c | /home/umayadav/repo/rocMLIR/build/external/llvm-project/llvm/bin/mlir-runner -O2 --shared-libs=/home/umayadav/repo/rocMLIR/build/external/llvm-project/llvm/lib/libmlir_rocm_runtime.so,/home/umayadav/repo/rocMLIR/build/lib/libconv-validation-wrappers.so,/home/umayadav/repo/rocMLIR/build/external/llvm-project/llvm/lib/libmlir_runner_utils.so --entry-point-result=void

[dorde-antic]

@dhernandez0 @umangyadav
Last time I have tested it, perfRunner didn't try i8 as datatype even was added to datatypes for attention in this PR (uploading screenshot of one of the runs). It was left for me to discover why does it happen (why it doesn't also try i8 datatypes), but I haven't had time to do it yet. Anyway, perfRunner doesn't fail with this change and it worked correctly for other datatypes (as tested last time).

I will try perfRunner as mentioned in comment a little bit later and also I'll look up why is the first thing (not trying i8) happening, but for the second thing I think that I will be able to do it once I finish work related to other urgent issue which is bug.