Reduce FP8_E5M2 -> FP16 conversion overhead

[sanchitintel]

Summary

• Line-encoding was changed to Unix, so in order to view the actual diff (not as distracting), please go to https://github.com/codeplaysoftware/cutlass-sycl/pull/396/files?diff=split&w=1.

• Even if we would use shared memory for caching FP8 -> FP16 converted elements, faster conversion would still be helpful on GPUs that lack native FP8 support.

• Since FP8_E5M2 -> FP16 conversion is using 32-bit registers & int32 ALUs, we might as well convert 4 FP8 elements to FP16 in one iteration. This novel implementation uses 3 shift instructions for every 4 FP8 elements (besides 4 and, 2 or instructions but shift has quite higher latency than them), so it should perform better than the projection, which uses 1 shift instruction for each FP8 element (apart from other lower latency instructions).

• Also moved convert_FP8_to_FP16 to fp8_to_fp16.h, since it needs to be reused in FP8 Grouped GEMM.

Performance improvement

Manual loop unrolling can improve performance a bit more. For example, if 8 elements would be converted inside the loop, the performance would be slightly better. With 16, a bit even more.

Converting 8 elements in a loop iteration

With the default problem size in the example, GEMM performance improved by ~18.37% on Intel GPU Max 1550.

On BMG, the performance improvement was ~15.28% for the default input shape.

Converting 4 elements in a loop iteration (current implementation in this PR)

By only converting 4 elements inside the loop & using a compile-time selected unroll factor equal to the number of iterations, there's a small regression compared to converting 8 elements in an iteration.

Intel GPU Max 1550:

BMG:

DPCPP: https://github.com/intel/llvm/releases/tag/nightly-2025-03-24

cc @jiyang1011 @pengzhao-intel

[sanchitintel]

Hi @t4c1 @jiyang1011,

Can you please review this PR as per your convenience? Thanks!
The CI failures are unrelated, BTW.

[sanchitintel]

Hi @aacostadiaz, thank you for triggering the CI!
All CI checks have now passed. Can you please help land this PR, as I lack the permissions to do so at my end?

Thanks!