jirayu_sri/stockfish
1
0
Fork 0
mirror of https://github.com/official-stockfish/Stockfish.git synced 2026-07-22 12:47:08 +00:00
Code Issues Packages Projects Releases Wiki Activity

More loongarch optimizations

Browse Source 
Passed STC:
LLR: 2.93 (-2.94,2.94) <0.00,2.00>
Total: 113120 W: 29094 L: 28674 D: 55352
Ptnml(0-2): 362, 12619, 30248, 12899, 432
https://tests.stockfishchess.org/tests/view/6a1486f7818cacc1db0ac832

closes https://github.com/official-stockfish/Stockfish/pull/6862

No functional change
This commit is contained in:
anematode
2026-06-06 10:03:25 +02:00
committed by Joost VandeVondele
parent 8416ccafff
commit db6cba47d6
2 changed files with 16 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/nnue/nnue_feature_transformer.h
+10 -7
View File
@@ -308,21 +308,23 @@ class FeatureTransformer {
vec_t acc1a = vec_add_16(in1[i + 0], tin1[i + 0]);
vec_t acc1b = vec_add_16(in1[i + 1], tin1[i + 1]);
#if defined(USE_NEON)
// The NEON path relies on unsigned saturation for crelu
static_assert(FtMaxVal == 255);
#if defined(USE_NEON)
uint16x8_t mul0 = vmull_u8(vqmovun_s16(acc0a), vqmovun_s16(acc1a));
uint16x8_t mul1 = vmull_u8(vqmovun_s16(acc0b), vqmovun_s16(acc1b));
uint8x16x2_t uzp =
vuzpq_u8(vreinterpretq_u8_u16(mul0), vreinterpretq_u8_u16(mul1));
uint8x16_t pab = vshrq_n_u8(uzp.val[1], 1);
packed[k] = out[j + k] = reinterpret_cast<vec_t>(pab);
vec_t result = reinterpret_cast<vec_t>(pab);
#elif defined(USE_LSX) || defined(USE_LASX)
vec_t pa = vec_packus_16(acc0a, acc0b);
vec_t pb = vec_packus_16(acc1a, acc1b);
vec_t hi = vec_mulhi_8(pa, pb);
vec_t result = vec_srli_8(hi, 1);
#else
vec_t sum0a = vec_slli_16(vec_max_16(vec_min_16(acc0a, FtMax), Zero), shift);
vec_t sum0b = vec_slli_16(vec_max_16(vec_min_16(acc0b, FtMax), Zero), shift);
vec_t sum1a = vec_min_16(acc1a, FtMax);
@@ -331,9 +333,10 @@ class FeatureTransformer {
vec_t pa = vec_mulhi_16(sum0a, sum1a);
vec_t pb = vec_mulhi_16(sum0b, sum1b);
packed[k] = out[j + k] = vec_packus_16(pa, pb);
vec_t result = vec_packus_16(pa, pb);
#endif
packed[k] = out[j + k] = result;
}
cursor.record2(packed[0], packed[1]);
src/nnue/simd.h
+5
View File
@@ -282,6 +282,8 @@ inline __m256i lasx_packus_32(__m256i a, __m256i b) {
#define vec_zero_psqt() __lasx_xvldi(0)
#define vec_nnz(a) lasx_vec_nnz(a)
#define vec_convert_8_16(a) lasx_cvtepi8_epi16(a)
#define vec_mulhi_8 __lasx_xvmuh_bu
#define vec_srli_8 __lasx_xvsrli_b
#define vec128_zero __lsx_vldi(0)
#define vec128_set_16(a) __lsx_vreplgr2vr_h(a)
@@ -368,6 +370,9 @@ inline __m128i vec_convert_8_16(std::uint64_t x) {
return __lsx_vsllwil_h_b(v, 0);
}
#define vec_mulhi_8 __lsx_vmuh_bu
#define vec_srli_8 __lsx_vsrli_b
#define vec128_zero __lsx_vldi(0)
#define vec128_set_16(a) __lsx_vreplgr2vr_h(a)
#define vec128_load(a) (*(a))