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

Split NEON dense affine accumulation two ways

Browse Source 
Speedup measured locally with apple-silicon M3 Pro:

```
Result of  50 runs
==================
base (./sf-master    ) =    1570076  +/- 27215
test (...-neon-2chain) =    1578923  +/- 29000
diff                   =      +8847  +/- 5700

speedup        = +0.0056
P(speedup > 0) =  0.9988

CPU: 11 x arm
Hyperthreading: off
```

Passed STC:
https://tests.stockfishchess.org/tests/view/6a49678cf97ff95f78795b38
LLR: 4.43 (-2.94,2.94) <0.00,2.00>
Total: 137696 W: 36041 L: 35477 D: 66178
Ptnml(0-2): 289, 14663, 38396, 15195, 305

closes https://github.com/official-stockfish/Stockfish/pull/6956

No functional change
This commit is contained in:
Linmiao Xu
2026-07-10 20:16:42 +02:00
committed by Joost VandeVondele
parent cd75d31127
commit 2d239de460
1 changed files with 8 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/nnue/layers/affine_transform.h
+8 -4
View File
@@ -255,7 +255,7 @@ class AffineTransform {
constexpr IndexType NumChunks = ceil_to_multiple<IndexType>(InputDimensions, 8) / 4;
constexpr IndexType NumAccums = OutputDimensions / OutputSimdWidth;
#if defined(USE_VNNI)
#if defined(USE_VNNI) || defined(USE_NEON_DOTPROD)
constexpr IndexType NumRegs = 2 * NumAccums;
#else
constexpr IndexType NumRegs = NumAccums;
@@ -269,7 +269,7 @@ class AffineTransform {
acc[k] = vec_set_32(0);
IndexType i = 0;
#if defined(USE_VNNI)
#if defined(USE_VNNI) || defined(USE_NEON_DOTPROD)
for (; i < NumChunks; i += 2)
{
const vec_t in0 = vec_set_32(load_as<i32>(input + i * sizeof(i32)));
@@ -287,7 +287,11 @@ class AffineTransform {
}
for (IndexType k = 0; k < NumAccums; ++k)
#if defined(USE_NEON_DOTPROD)
acc[k] = vaddq_s32(acc[k], acc[k + NumAccums]);
#else
acc[k] = vec_add_32(acc[k], acc[k + NumAccums]);
#endif
#endif
for (; i < NumChunks; ++i)
{
@@ -295,12 +299,12 @@ class AffineTransform {
const auto col0 =
reinterpret_cast<const vec_t*>(&weights[i * OutputDimensions * 4]);
for (IndexType k = 0; k < NumRegs; ++k)
for (IndexType k = 0; k < NumAccums; ++k)
vec_add_dpbusd_32(acc[k], in0, col0[k]);
}
vec_t* outptr = reinterpret_cast<vec_t*>(output);
for (IndexType k = 0; k < NumRegs; ++k)
for (IndexType k = 0; k < NumAccums; ++k)
outptr[k] = acc[k];
#undef vec_set_32