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

Neon dotprod speedup in sparse-input affine transform

Browse Source 
Also apply VNNI accumulator splitting strategy to NEON dotprod.

Speedup measured locally with profile-build, apple-silicon M3 Pro:
```
Result of  20 runs
==================
base (...kfish-master) =    1582485  +/- 12985
test (...parse-affine) =    1605204  +/- 13801
diff                   =     +22720  +/- 1212

speedup        = +0.0144
P(speedup > 0) =  1.0000

CPU: 11 x arm
Hyperthreading: off
```

Passed STC:
https://tests.stockfishchess.org/tests/view/69c04c71f690a4b7f5fb0cde
LLR: 2.95 (-2.94,2.94) <0.00,2.00>
Total: 80576 W: 20748 L: 20391 D: 39437
Ptnml(0-2): 161, 8472, 22658, 8843, 154

closes https://github.com/official-stockfish/Stockfish/pull/6682

No functional change
This commit is contained in:
Linmiao Xu
2026-04-02 20:30:26 +02:00
committed by Disservin
parent b9f3273af1
commit f614ff90e2
1 changed files with 12 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/nnue/layers/affine_transform_sparse_input.h
+12 -2
View File
@@ -288,7 +288,7 @@ class AffineTransformSparseInput {
// If we're using high-latency dot product instructions, split the accumulators // If we're using high-latency dot product instructions, split the accumulators
// to create 3 separate dependency chains and merge at the end // to create 3 separate dependency chains and merge at the end
constexpr IndexType NumRegs = constexpr IndexType NumRegs =
#if defined(USE_VNNI) #if defined(USE_VNNI) || defined(USE_NEON_DOTPROD)
3 * NumAccums; 3 * NumAccums;
#else #else
NumAccums; NumAccums;
@@ -309,9 +309,14 @@ class AffineTransformSparseInput {
// convince GCC to not do weird pointer arithmetic in the following loop // convince GCC to not do weird pointer arithmetic in the following loop
const std::int8_t* weights_cp = weights; const std::int8_t* weights_cp = weights;
#if defined(USE_VNNI) #if defined(USE_VNNI) || defined(USE_NEON_DOTPROD)
#if defined(USE_VNNI)
for (IndexType k = NumAccums; k < NumRegs; ++k) for (IndexType k = NumAccums; k < NumRegs; ++k)
acc[k] = vec_zero(); acc[k] = vec_zero();
#else
for (IndexType k = NumAccums; k < NumRegs; ++k)
acc[k] = vdupq_n_s32(0);
#endif
while (start < end - 2) while (start < end - 2)
{ {
@@ -337,8 +342,13 @@ class AffineTransformSparseInput {
vec_add_dpbusd_32(acc[k + 2 * NumAccums], in2, col2[k]); vec_add_dpbusd_32(acc[k + 2 * NumAccums], in2, col2[k]);
} }
} }
#if defined(USE_VNNI)
for (IndexType k = 0; k < NumAccums; ++k) for (IndexType k = 0; k < NumAccums; ++k)
acc[k] = vec_add_32(vec_add_32(acc[k], acc[k + NumAccums]), acc[k + 2 * NumAccums]); acc[k] = vec_add_32(vec_add_32(acc[k], acc[k + NumAccums]), acc[k + 2 * NumAccums]);
#else
for (IndexType k = 0; k < NumAccums; ++k)
acc[k] = vaddq_s32(vaddq_s32(acc[k], acc[k + NumAccums]), acc[k + 2 * NumAccums]);
#endif
#endif #endif
while (start < end) while (start < end)
{ {