opelly/Stockfish
1
0
Fork 0
mirror of https://github.com/opelly27/Stockfish.git synced 2026-05-20 07:27:46 +00:00
Code Issues Packages Projects Releases Wiki Activity

Workaround using unaligned loads for gcc < 9

Browse Source 
despite usage of alignas, the generated (avx2/avx512) code with older compilers needs to use
unaligned loads with older gcc (e.g. confirmed crash with gcc 7.3/mingw on abrok).

Better performance thus requires gcc >= 9 on hardware supporting avx2/avx512

closes https://github.com/official-stockfish/Stockfish/pull/2969

No functional change
This commit is contained in:
mstembera
2020-08-09 23:50:59 -07:00
committed by Joost VandeVondele
parent a54f9011c3
commit 875183b310
4 changed files with 36 additions and 107 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/nnue/nnue_feature_transformer.h
+6 -36
View File
@@ -109,37 +109,11 @@ namespace Eval::NNUE {
#if defined(USE_AVX2)
auto out = reinterpret_cast<__m256i*>(&output[offset]);
for (IndexType j = 0; j < kNumChunks; ++j) {
__m256i sum0 =
#if defined(__MINGW32__) || defined(__MINGW64__)
// HACK: Use _mm256_loadu_si256() instead of _mm256_load_si256. Because the binary
// compiled with g++ in MSYS2 crashes here because the output memory is not aligned
// even though alignas is specified.
_mm256_loadu_si256
#else
_mm256_load_si256
#endif
(&reinterpret_cast<const __m256i*>(
accumulation[perspectives[p]][0])[j * 2 + 0]);
__m256i sum1 =
#if defined(__MINGW32__) || defined(__MINGW64__)
_mm256_loadu_si256
#else
_mm256_load_si256
#endif
(&reinterpret_cast<const __m256i*>(
accumulation[perspectives[p]][0])[j * 2 + 1]);
#if defined(__MINGW32__) || defined(__MINGW64__)
_mm256_storeu_si256
#else
_mm256_store_si256
#endif
(&out[j], _mm256_permute4x64_epi64(_mm256_max_epi8(
__m256i sum0 = _mm256_loadA_si256(
&reinterpret_cast<const __m256i*>(accumulation[perspectives[p]][0])[j * 2 + 0]);
__m256i sum1 = _mm256_loadA_si256(
&reinterpret_cast<const __m256i*>(accumulation[perspectives[p]][0])[j * 2 + 1]);
_mm256_storeA_si256(&out[j], _mm256_permute4x64_epi64(_mm256_max_epi8(
_mm256_packs_epi16(sum0, sum1), kZero), kControl));
}
@@ -202,11 +176,7 @@ namespace Eval::NNUE {
auto column = reinterpret_cast<const __m256i*>(&weights_[offset]);
constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
for (IndexType j = 0; j < kNumChunks; ++j) {
#if defined(__MINGW32__) || defined(__MINGW64__)
_mm256_storeu_si256(&accumulation[j], _mm256_add_epi16(_mm256_loadu_si256(&accumulation[j]), column[j]));
#else
accumulation[j] = _mm256_add_epi16(accumulation[j], column[j]);
#endif
_mm256_storeA_si256(&accumulation[j], _mm256_add_epi16(_mm256_loadA_si256(&accumulation[j]), column[j]));
}
#elif defined(USE_SSE2)