Use specialised stores without mask in calc_loop_of_clusters about yask HOT 3 CLOSED

Thanks for the feedback and the detailed issue. The call to storeTo_masked was left in calc_loop_of_clusters because the Intel C++ compiler recognizes when a constant "all-ones" (from constexpr idx_t write_mask = idx_t (-1);) mask is used in calc_loop_of_clusters, and it automatically replaces the masked store with a regular store (i.e., vmovups), at least for avx512. I have confirmed that this optimization is not applied for avx2. This is likely because the int mask has to be transformed into a ymm mask (realv.hpp line 491), and the compiler doesn't recognize this idiom. It would not be advisable to put a run-time conditional test inside storeTo_masked, because branches also have a large negative impact on performance. The better solution would be for the YASK compiler to generate storeTo instead of storeTo_masked directly inside calc_loop_of_clusters. I will investigate this enhancement.

from yask.

A fix has been pushed to the develop branch if you want to test it.

from yask.

Thanks a lot for the quick action. I tested the fix, it works :)

from yask.