[Whatcookie]

• Code is about 4 times as fast on my zen4 machine
• Should be twice as fast on zen5 machines with full width AVX-512
• Code operates without a separate remainder-masking loop, since the instruction_mask bits can be used to avoid loading out of bounds data.

This code uses vpermt2b, and needs full length AVX-512 to function, so this won't work for skylake-x nor avx10-256.

Before: (194 avg)

After: (204 avg)

Before:

After:

[Whatcookie]

Pushed another commit which resolves an issue where the compiler was loading a value from memory every loop. Somehow the compiler inexplicably failed to calculate the loop conditions before the loop, and just placing them in a "length" variable resolves this.

Adds another few fps:

[cipherxof]

Comically faster in some situations. Haven't noticed much of a difference in actual rendered scenes for MGS4/MGO2, even in RSX bottlenecked areas.

Before:

After:

Before:

After:

[kd-11]

I can't prove or disprove the correctness of the avx512 stuff without an avx512 cpu and we have no unit tests, so I guess just yolo and bulk-revert if a regression occurs.

[Jonathan44062]

7 FPS Extra in Sonic Unleashed
Master

PR

[elad335]

@cipherxof Can you disable locally this pr and apply #16797 and test it? it applies a similar improvement to non-avx512 users as well.