Speedup write_moves using _mm512_mask_compressstoreu_epi16

[KierenP]

We can use the _mm512_mask_compressstoreu_epi16 instruction which does a compress + store in one instruction rather than two. It's part of the same AVX512_VBMI2 instruction set.

Baseline:

$ ./stockfish_master.exe speedtest
Stockfish dev-20250816-169737a9 by the Stockfish developers (see AUTHORS file)
info string Using 32 threads
Warmup position 3/3
Position 258/258
===========================
Version                    : Stockfish dev-20250816-169737a9
Compiled by                : g++ (GNUC) 15.1.0 on MinGW64
Compilation architecture   : x86-64-avx512icl
Compilation settings       : 64bit AVX512ICL VNNI AVX512 BMI2 AVX2 SSE41 SSSE3 SSE2 POPCNT
Compiler __VERSION__ macro : 15.1.0
Large pages                : no
User invocation            : speedtest
Filled invocation          : speedtest 32 4096 150
Available processors       : 0-31
Thread count               : 32
Thread binding             : none
TT size [MiB]              : 4096
Hash max, avg [per mille]  :
    single search          : 48, 23
    single game            : 647, 456
Total nodes searched       : 5322766982
Total search time [s]      : 153.515
Nodes/second               : 34672618

With my change:

$ ./stockfish.exe speedtest
Stockfish dev-20250816-169737a9 by the Stockfish developers (see AUTHORS file)
info string Using 32 threads
Warmup position 3/3
Position 258/258
===========================
Version                    : Stockfish dev-20250816-169737a9
Compiled by                : g++ (GNUC) 15.1.0 on MinGW64
Compilation architecture   : x86-64-avx512icl
Compilation settings       : 64bit AVX512ICL VNNI AVX512 BMI2 AVX2 SSE41 SSSE3 SSE2 POPCNT
Compiler __VERSION__ macro : 15.1.0
Large pages                : no
User invocation            : speedtest
Filled invocation          : speedtest 32 4096 150
Available processors       : 0-31
Thread count               : 32
Thread binding             : none
TT size [MiB]              : 4096
Hash max, avg [per mille]  :
    single search          : 47, 22
    single game            : 655, 448
Total nodes searched       : 5403131641
Total search time [s]      : 153.519
Nodes/second               : 35195198

Speedup: 1.51%

[mstembera]

See #6153 (comment)

[mstembera]

We should probably put a comment here just like

Stockfish/src/nnue/layers/affine_transform_sparse_input.h

Line 105 in 169737a

// Avoid _mm512_mask_compressstoreu_epi16() as it's 256 uOps on Zen4

[KierenP]

Ah thanks, I didn't realise. That being said, it is a speedup on Zen5 so in the future maybe we can have a target taking advantage of this.

[Disservin]

how much of a speedup is this for zen5? if it is only minor id simply avoid this and not introduce a new target, nor start parsing cpuid..

[KierenP]

It's 1.5% (as seen in the above speedtest comparison). That being said, the different would be bigger if I also applied _mm512_mask_compressstoreu_epi16 in the NNUE inference.

This PR can probably be closed.