[BUG]: Validate tuning refactoring on all affected SM versions

[gevtushenko]

Is this a duplicate?

• I confirmed there appear to be no duplicate issues for this bug and that I agree to the Code of Conduct

Type of Bug

Performance

Component

CUB

Describe the bug

#3138 introduced 120% performance regression in cub::DeviceScan::Sum on Hopper.
Looking at PR description:

No changes in SASS for cub.test.device_scan.lid_0.types_0 except kernel symbol names

I assume that we checked SASS only on default set of architectures from our preset "60;70;80".

diff old.sass new.sass | wc -l 
88

Mentioned refactoring also touches SM90 policies, which results in the following diff:

diff old.90.sass new.90.sass | wc -l 
83618

We should verify SM90 SASS differences on other algorithms we refactored lately.

How to Reproduce

git checkout 52b4b671c111fc289b2f2247a16eff21d5a5a1a8
./ci/build_cub.sh -arch 90
cuobjdump --dump-sass build/cuda12.6-gcc13/cub-cpp17/bin/cub.bench.scan.exclusive.sum.base > new.90.sass
git checkout HEAD~
./ci/build_cub.sh -arch 90
cuobjdump --dump-sass build/cuda12.6-gcc13/cub-cpp17/bin/cub.bench.scan.exclusive.sum.base > old.90.sass
diff old.90.sass new.90.sass | wc -l 
83618

Expected behavior

No SASS difference on any architecture after tuning refactoring

Reproduction link

No response

Operating System

No response

nvidia-smi output

No response

NVCC version

No response

[gevtushenko]

Refactoring	Regression on SM90
#3138	Fixed by #3236
#3124	No
#3127	Fixed by #3239
#3125	Fixed by #3174
#3128	No
#3137	Fixed by #3240
#3139	No
#3140	No
#3145	No

[bernhardmgruber]

I assume that we checked SASS only on default set of architectures from our preset "60;70;80".

It may be even worse. I cannot find what exact cmake invocation I used, but it could have been just 86. I will revisit your table with all architectures touched by each PR.

[gevtushenko]

I will revisit your table with all architectures touched by each PR.

That's what I did to prepare the table above