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

Speed up feature transform on neon

Browse Source 
STC:
LLR: 2.95 (-2.94,2.94) <0.00,2.00>
Total: 109344 W: 27919 L: 27520 D: 53905
Ptnml(0-2): 200, 11718, 30434, 12123, 197
https://tests.stockfishchess.org/tests/view/6a123ed0818cacc1db0ac39e

Updated to anematode’s suggestion - apparently clang already optimized it to this.

Apple M1

Result of 100 runs
base (...kfish.master) = 1307276 +/- 10930
test (./stockfish ) = 1320940 +/- 11279
diff = +13664 +/- 2223

speedup = +0.0105
P(speedup > 0) = 1.0000

closes https://github.com/official-stockfish/Stockfish/pull/6855

No functional change
This commit is contained in:
Ryan Hirsch
2026-05-29 19:05:08 +02:00
committed by Joost VandeVondele
parent 313ea4ab04
commit 24a551899e
1 changed files with 22 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/nnue/nnue_feature_transformer.h
+20 -13
View File
@@ -234,8 +234,11 @@ class FeatureTransformer {
static_assert((HalfDimensions / 2) % OutputChunkSize == 0);
constexpr IndexType NumOutputChunks = HalfDimensions / 2 / OutputChunkSize;
#if !defined(USE_NEON)
const vec_t Zero = vec_zero();
const vec_t FtMax = vec_set_16(FtMaxVal);
constexpr int shift = 7;
#endif
const vec_t* in0 = reinterpret_cast<const vec_t*>(&(accumulation[perspectives[p]][0]));
const vec_t* in1 =
@@ -289,19 +292,6 @@ class FeatureTransformer {
// 8 bits. Shifting it by 7 bits left will no longer occupy the
// signed bit, so we are safe.
// Note that on NEON processors, we shift left by 6 instead
// because the instruction "vqdmulhq_s16" also doubles the
// return value after the multiplication, adding an extra shift
// to the left by 1, so we compensate by shifting less before
// the multiplication.
constexpr int shift =
#if defined(USE_SSE2) || defined(USE_LASX) || defined(USE_LSX)
7;
#else
6;
#endif
const vec_t* tin0 =
reinterpret_cast<const vec_t*>(&(threatAccumulation[perspectives[p]][0]));
const vec_t* tin1 = reinterpret_cast<const vec_t*>(
@@ -318,6 +308,21 @@ 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);
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);
#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);
@@ -327,6 +332,8 @@ class FeatureTransformer {
vec_t pb = vec_mulhi_16(sum0b, sum1b);
packed[k] = out[j + k] = vec_packus_16(pa, pb);
#endif
}
cursor.record2(packed[0], packed[1]);