opelly/Stockfish
1
0
Fork 0
mirror of https://github.com/opelly27/Stockfish.git synced 2026-05-20 16:47:37 +00:00
Code Issues Packages Projects Releases Wiki Activity

Provide vectorized NNUE code for SSE2 and MMX targets

Browse Source 
This patch allows old x86 CPUs, from AMD K8 (which the x86-64 baseline
targets) all the way down to the Pentium MMX, to benefit from NNUE with
comparable performance hit versus hand-written eval as on more modern
processors.

NPS of the bench with NNUE enabled on a Pentium III 1.13 GHz (using the
MMX code):
  master: 38951
  this patch: 80586

NPS of the bench with NNUE enabled using baseline x86-64 arch, which is
how linux distros are likely to package stockfish, on a modern CPU
(using the SSE2 code):
  master: 882584
  this patch: 1203945

closes https://github.com/official-stockfish/Stockfish/pull/2956

No functional change.
This commit is contained in:
Fanael Linithien
2020-08-09 16:20:45 +02:00
committed by Joost VandeVondele
parent f948cd008d
commit 21df37d7fd
7 changed files with 150 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/nnue/layers/clipped_relu.h
+19 -1
View File
@@ -84,7 +84,7 @@ namespace Eval::NNUE::Layers {
}
constexpr IndexType kStart = kNumChunks * kSimdWidth;
#elif defined(USE_SSSE3)
#elif defined(USE_SSE2)
constexpr IndexType kNumChunks = kInputDimensions / kSimdWidth;
#ifdef USE_SSE41
@@ -115,6 +115,24 @@ namespace Eval::NNUE::Layers {
}
constexpr IndexType kStart = kNumChunks * kSimdWidth;
#elif defined(USE_MMX)
constexpr IndexType kNumChunks = kInputDimensions / kSimdWidth;
const __m64 k0x80s = _mm_set1_pi8(-128);
const auto in = reinterpret_cast<const __m64*>(input);
const auto out = reinterpret_cast<__m64*>(output);
for (IndexType i = 0; i < kNumChunks; ++i) {
const __m64 words0 = _mm_srai_pi16(
_mm_packs_pi32(in[i * 4 + 0], in[i * 4 + 1]),
kWeightScaleBits);
const __m64 words1 = _mm_srai_pi16(
_mm_packs_pi32(in[i * 4 + 2], in[i * 4 + 3]),
kWeightScaleBits);
const __m64 packedbytes = _mm_packs_pi16(words0, words1);
out[i] = _mm_subs_pi8(_mm_adds_pi8(packedbytes, k0x80s), k0x80s);
}
_mm_empty();
constexpr IndexType kStart = kNumChunks * kSimdWidth;
#elif defined(USE_NEON)
constexpr IndexType kNumChunks = kInputDimensions / (kSimdWidth / 2);
const int8x8_t kZero = {0};