Add way to control SIMD width

[axeldavy]

For some reasons, a few changes in my kernels can lead the compiler to pick a different choice for SIMD width (8, 16 or 32).

For some reason I don't explain, some of my kernels using shared memory and barriers perform very significantly better (several times faster) when the selected SIMD width is 8.

Thus I'd like a way to enforce the choice of SIMD width of 8. (Also ideally compiler would make better choices).

Maybe make __attribute__((vec_type_hint(float8))) trigger SIMD-8 for example ?

[paigeale]

Hello Axel,

For forcing SIMD-8 please use the following attribute....
__attribute__((intel_reqd_sub_group_size(8)))

This works on any of the three SIMD sizes.

[axeldavy]

It works indeed, thank you !