sdpa fma f16

[syurkevi]

This PR adds FMA support to SDPA for the fp16 data type. This allows SDPA to work on older platforms that don't have systolic support.

• Do all unit and benchdnn tests (make test and make test_benchdnn_*) pass locally for each commit?
• Have you formatted the code using clang-format?

Performance improvements

Performance was measured and tuned relative to a primitives only based implementation:

[TaoLv]

Hi @syurkevi what does this mean from user perspective? Is it f16 acc automatically or users will need to turn on any flag to enable? With this, do we expect the same numerical behavior between micro-kernel implementation and matmul primitive based implementation? Thanks!

[syurkevi]

Hi @syurkevi what does this mean from user perspective? Is it f16 acc automatically or users will need to turn on any flag to enable? With this, do we expect the same numerical behavior between micro-kernel implementation and matmul primitive based implementation? Thanks!

F16 support is now automatically available on MTL where it was previously unsupported. The accumulation mode matches the behavior of the systolic SDPA kernel (f32 acc) so results should match our current implementation.

[syurkevi]

make test
disable benchdnn_all
enable benchdnn_graph
enable test_device_gpu
enable arch_gpu_xe-hpc
enable arch_gpu_xe-hpg-atsm
enable arch_gpu_xe-hpg-dg2
enable arch_gpu_xe-lp
enable arch_gpu_xe-lpg
enable arch_gpu_xe-lpg+
enable arch_gpu_xe2-hpg-bmg
enable arch_gpu_xe2-lpg

[syurkevi]

make test
disable benchdnn_all
enable benchdnn_graph
enable test_device_gpu
enable arch_gpu_xe-hpc
enable arch_gpu_xe-hpg-atsm
enable arch_gpu_xe-hpg-dg2
enable arch_gpu_xe-lp
enable arch_gpu_xe-lpg
enable arch_gpu_xe-lpg+
enable arch_gpu_xe2-hpg-bmg
enable arch_gpu_xe2-lpg