[AMD] ElementwiseOpToLLVM: Do not convert types if they are equal

[joviliast]

This commit fixes failure in python/tutorials/03-matrix-multiplication.py for FMA cases,
also fixes mixed dot for FMA cases.
Tested on Navi31

[binarman]

Essentially, this patch enables triton::FPToFP operation to cast fp16 to fp32 and back, @joviliast is it correct?

@ptillet
Do we want FPToFP operation to be able to convert any float types?

[joviliast]

Essentially, this patch enables triton::FPToFP operation to cast fp16 to fp32 and back, @joviliast is it correct?

This patch disables casting. Did you mean "pass"?

[binarman]

This patch disables casting. Did you mean "pass"?

It disables only "intermediate" casting, the rest is in place:

• https://github.com/openai/triton/pull/3091/files#diff-c1b7645c56652fe811ae807f37bbea185af4221b1122fa095fbd43b80266182eR1686
• https://github.com/openai/triton/pull/3091/files#diff-c1b7645c56652fe811ae807f37bbea185af4221b1122fa095fbd43b80266182eR1700

[ptillet]

@binarman in principle yes, in practice there are many subtleties related to FP8 and rounding modes. Even our Nvidia backend isn't handling this super well today

[zhanglx13]

@joviliast @binarman Can you elaborate on the motivation of this PR? Why does converting between the same type cause FMA failures?

[joviliast]

@joviliast @binarman Can you elaborate on the motivation of this PR? Why does converting between the same type cause FMA failures?

Because the case of the same internal types just not supported and it fails on compile time as unsupported conversion.

[ThomasRaoux]

@joviliast @binarman Can you elaborate on the motivation of this PR? Why does converting between the same type cause FMA failures?

Because the case of the same internal types just not supported and it fails on compile time as unsupported conversion.

can you make a minimized lit test out of this failure. That will help everybody understand (and will be prevent regressions)

[joviliast]

can you make a minimized lit test out of this failure. That will help everybody understand (and will be prevent regressions)

@ThomasRaoux , done.
Added commit containing lit test for fixed case.
It can be ran after merging #3254

[ThomasRaoux]

LGTM, thanks for adding a test

[ThomasRaoux]

LGTM

[zhanglx13]

lit tests under amd dir is not tested

[ThomasRaoux]

lit tests under amd dir is not tested

Thanks for catching that. Is someone working on fixing this?

[joviliast]

lit tests under amd dir is not tested

I don't quite catch your point.

I can see it passed in CI logs:

[zhanglx13]

It actually works. Sorry for blocking the merge.