#define vec_store(a,b) _mm512_store_si512(a,b)
#define vec_add_16(a,b) _mm512_add_epi16(a,b)
#define vec_sub_16(a,b) _mm512_sub_epi16(a,b)
- static constexpr IndexType kNumRegs = 8; // only 8 are needed
+ static constexpr IndexType NumRegs = 8; // only 8 are needed
#elif USE_AVX2
typedef __m256i vec_t;
#define vec_store(a,b) _mm256_store_si256(a,b)
#define vec_add_16(a,b) _mm256_add_epi16(a,b)
#define vec_sub_16(a,b) _mm256_sub_epi16(a,b)
- static constexpr IndexType kNumRegs = 16;
+ static constexpr IndexType NumRegs = 16;
#elif USE_SSE2
typedef __m128i vec_t;
#define vec_store(a,b) *(a)=(b)
#define vec_add_16(a,b) _mm_add_epi16(a,b)
#define vec_sub_16(a,b) _mm_sub_epi16(a,b)
- static constexpr IndexType kNumRegs = Is64Bit ? 16 : 8;
+ static constexpr IndexType NumRegs = Is64Bit ? 16 : 8;
#elif USE_MMX
typedef __m64 vec_t;
#define vec_store(a,b) *(a)=(b)
#define vec_add_16(a,b) _mm_add_pi16(a,b)
#define vec_sub_16(a,b) _mm_sub_pi16(a,b)
- static constexpr IndexType kNumRegs = 8;
+ static constexpr IndexType NumRegs = 8;
#elif USE_NEON
typedef int16x8_t vec_t;
#define vec_store(a,b) *(a)=(b)
#define vec_add_16(a,b) vaddq_s16(a,b)
#define vec_sub_16(a,b) vsubq_s16(a,b)
- static constexpr IndexType kNumRegs = 16;
+ static constexpr IndexType NumRegs = 16;
#else
#undef VECTOR
private:
// Number of output dimensions for one side
- static constexpr IndexType kHalfDimensions = kTransformedFeatureDimensions;
+ static constexpr IndexType HalfDimensions = TransformedFeatureDimensions;
#ifdef VECTOR
- static constexpr IndexType kTileHeight = kNumRegs * sizeof(vec_t) / 2;
- static_assert(kHalfDimensions % kTileHeight == 0, "kTileHeight must divide kHalfDimensions");
+ static constexpr IndexType TileHeight = NumRegs * sizeof(vec_t) / 2;
+ static_assert(HalfDimensions % TileHeight == 0, "TileHeight must divide HalfDimensions");
#endif
public:
using OutputType = TransformedFeatureType;
// Number of input/output dimensions
- static constexpr IndexType kInputDimensions = RawFeatures::kDimensions;
- static constexpr IndexType kOutputDimensions = kHalfDimensions * 2;
+ static constexpr IndexType InputDimensions = RawFeatures::Dimensions;
+ static constexpr IndexType OutputDimensions = HalfDimensions * 2;
// Size of forward propagation buffer
- static constexpr std::size_t kBufferSize =
- kOutputDimensions * sizeof(OutputType);
+ static constexpr std::size_t BufferSize =
+ OutputDimensions * sizeof(OutputType);
// Hash value embedded in the evaluation file
- static constexpr std::uint32_t GetHashValue() {
-
- return RawFeatures::kHashValue ^ kOutputDimensions;
+ static constexpr std::uint32_t get_hash_value() {
+ return RawFeatures::HashValue ^ OutputDimensions;
}
// Read network parameters
- bool ReadParameters(std::istream& stream) {
-
- for (std::size_t i = 0; i < kHalfDimensions; ++i)
- biases_[i] = read_little_endian<BiasType>(stream);
- for (std::size_t i = 0; i < kHalfDimensions * kInputDimensions; ++i)
- weights_[i] = read_little_endian<WeightType>(stream);
+ bool read_parameters(std::istream& stream) {
+ for (std::size_t i = 0; i < HalfDimensions; ++i)
+ biases[i] = read_little_endian<BiasType>(stream);
+ for (std::size_t i = 0; i < HalfDimensions * InputDimensions; ++i)
+ weights[i] = read_little_endian<WeightType>(stream);
return !stream.fail();
}
// Convert input features
- void Transform(const Position& pos, OutputType* output) const {
-
- UpdateAccumulator(pos, WHITE);
- UpdateAccumulator(pos, BLACK);
+ void transform(const Position& pos, OutputType* output) const {
+ update_accumulator(pos, WHITE);
+ update_accumulator(pos, BLACK);
const auto& accumulation = pos.state()->accumulator.accumulation;
#if defined(USE_AVX512)
- constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth * 2);
- static_assert(kHalfDimensions % (kSimdWidth * 2) == 0);
- const __m512i kControl = _mm512_setr_epi64(0, 2, 4, 6, 1, 3, 5, 7);
- const __m512i kZero = _mm512_setzero_si512();
+ constexpr IndexType NumChunks = HalfDimensions / (SimdWidth * 2);
+ static_assert(HalfDimensions % (SimdWidth * 2) == 0);
+ const __m512i Control = _mm512_setr_epi64(0, 2, 4, 6, 1, 3, 5, 7);
+ const __m512i Zero = _mm512_setzero_si512();
#elif defined(USE_AVX2)
- constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth;
- constexpr int kControl = 0b11011000;
- const __m256i kZero = _mm256_setzero_si256();
+ constexpr IndexType NumChunks = HalfDimensions / SimdWidth;
+ constexpr int Control = 0b11011000;
+ const __m256i Zero = _mm256_setzero_si256();
#elif defined(USE_SSE2)
- constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth;
+ constexpr IndexType NumChunks = HalfDimensions / SimdWidth;
#ifdef USE_SSE41
- const __m128i kZero = _mm_setzero_si128();
+ const __m128i Zero = _mm_setzero_si128();
#else
const __m128i k0x80s = _mm_set1_epi8(-128);
#endif
#elif defined(USE_MMX)
- constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth;
+ constexpr IndexType NumChunks = HalfDimensions / SimdWidth;
const __m64 k0x80s = _mm_set1_pi8(-128);
#elif defined(USE_NEON)
- constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
- const int8x8_t kZero = {0};
+ constexpr IndexType NumChunks = HalfDimensions / (SimdWidth / 2);
+ const int8x8_t Zero = {0};
#endif
const Color perspectives[2] = {pos.side_to_move(), ~pos.side_to_move()};
for (IndexType p = 0; p < 2; ++p) {
- const IndexType offset = kHalfDimensions * p;
+ const IndexType offset = HalfDimensions * p;
#if defined(USE_AVX512)
auto out = reinterpret_cast<__m512i*>(&output[offset]);
- for (IndexType j = 0; j < kNumChunks; ++j) {
+ for (IndexType j = 0; j < NumChunks; ++j) {
__m512i sum0 = _mm512_load_si512(
&reinterpret_cast<const __m512i*>(accumulation[perspectives[p]][0])[j * 2 + 0]);
__m512i sum1 = _mm512_load_si512(
&reinterpret_cast<const __m512i*>(accumulation[perspectives[p]][0])[j * 2 + 1]);
- _mm512_store_si512(&out[j], _mm512_permutexvar_epi64(kControl,
- _mm512_max_epi8(_mm512_packs_epi16(sum0, sum1), kZero)));
+ _mm512_store_si512(&out[j], _mm512_permutexvar_epi64(Control,
+ _mm512_max_epi8(_mm512_packs_epi16(sum0, sum1), Zero)));
}
#elif defined(USE_AVX2)
auto out = reinterpret_cast<__m256i*>(&output[offset]);
- for (IndexType j = 0; j < kNumChunks; ++j) {
+ for (IndexType j = 0; j < NumChunks; ++j) {
__m256i sum0 = _mm256_load_si256(
&reinterpret_cast<const __m256i*>(accumulation[perspectives[p]][0])[j * 2 + 0]);
__m256i sum1 = _mm256_load_si256(
&reinterpret_cast<const __m256i*>(accumulation[perspectives[p]][0])[j * 2 + 1]);
_mm256_store_si256(&out[j], _mm256_permute4x64_epi64(_mm256_max_epi8(
- _mm256_packs_epi16(sum0, sum1), kZero), kControl));
+ _mm256_packs_epi16(sum0, sum1), Zero), Control));
}
#elif defined(USE_SSE2)
auto out = reinterpret_cast<__m128i*>(&output[offset]);
- for (IndexType j = 0; j < kNumChunks; ++j) {
+ for (IndexType j = 0; j < NumChunks; ++j) {
__m128i sum0 = _mm_load_si128(&reinterpret_cast<const __m128i*>(
accumulation[perspectives[p]][0])[j * 2 + 0]);
__m128i sum1 = _mm_load_si128(&reinterpret_cast<const __m128i*>(
_mm_store_si128(&out[j],
#ifdef USE_SSE41
- _mm_max_epi8(packedbytes, kZero)
+ _mm_max_epi8(packedbytes, Zero)
#else
_mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s)
#endif
#elif defined(USE_MMX)
auto out = reinterpret_cast<__m64*>(&output[offset]);
- for (IndexType j = 0; j < kNumChunks; ++j) {
+ for (IndexType j = 0; j < NumChunks; ++j) {
__m64 sum0 = *(&reinterpret_cast<const __m64*>(
accumulation[perspectives[p]][0])[j * 2 + 0]);
__m64 sum1 = *(&reinterpret_cast<const __m64*>(
#elif defined(USE_NEON)
const auto out = reinterpret_cast<int8x8_t*>(&output[offset]);
- for (IndexType j = 0; j < kNumChunks; ++j) {
+ for (IndexType j = 0; j < NumChunks; ++j) {
int16x8_t sum = reinterpret_cast<const int16x8_t*>(
accumulation[perspectives[p]][0])[j];
- out[j] = vmax_s8(vqmovn_s16(sum), kZero);
+ out[j] = vmax_s8(vqmovn_s16(sum), Zero);
}
#else
- for (IndexType j = 0; j < kHalfDimensions; ++j) {
+ for (IndexType j = 0; j < HalfDimensions; ++j) {
BiasType sum = accumulation[static_cast<int>(perspectives[p])][0][j];
output[offset + j] = static_cast<OutputType>(
std::max<int>(0, std::min<int>(127, sum)));
}
private:
- void UpdateAccumulator(const Position& pos, const Color c) const {
+ void update_accumulator(const Position& pos, const Color c) const {
#ifdef VECTOR
// Gcc-10.2 unnecessarily spills AVX2 registers if this array
// is defined in the VECTOR code below, once in each branch
- vec_t acc[kNumRegs];
+ vec_t acc[NumRegs];
#endif
// Look for a usable accumulator of an earlier position. We keep track
// The first condition tests whether an incremental update is
// possible at all: if this side's king has moved, it is not possible.
static_assert(std::is_same_v<RawFeatures::SortedTriggerSet,
- Features::CompileTimeList<Features::TriggerEvent, Features::TriggerEvent::kFriendKingMoved>>,
- "Current code assumes that only kFriendlyKingMoved refresh trigger is being used.");
+ Features::CompileTimeList<Features::TriggerEvent, Features::TriggerEvent::FriendKingMoved>>,
+ "Current code assumes that only FriendlyKingMoved refresh trigger is being used.");
if ( dp.piece[0] == make_piece(c, KING)
|| (gain -= dp.dirty_num + 1) < 0)
break;
// Gather all features to be updated. This code assumes HalfKP features
// only and doesn't support refresh triggers.
- static_assert(std::is_same_v<Features::FeatureSet<Features::HalfKP<Features::Side::kFriend>>,
+ static_assert(std::is_same_v<Features::FeatureSet<Features::HalfKP<Features::Side::Friend>>,
RawFeatures>);
Features::IndexList removed[2], added[2];
- Features::HalfKP<Features::Side::kFriend>::AppendChangedIndices(pos,
+ Features::HalfKP<Features::Side::Friend>::append_changed_indices(pos,
next->dirtyPiece, c, &removed[0], &added[0]);
for (StateInfo *st2 = pos.state(); st2 != next; st2 = st2->previous)
- Features::HalfKP<Features::Side::kFriend>::AppendChangedIndices(pos,
+ Features::HalfKP<Features::Side::Friend>::append_changed_indices(pos,
st2->dirtyPiece, c, &removed[1], &added[1]);
// Mark the accumulators as computed.
StateInfo *info[3] =
{ next, next == pos.state() ? nullptr : pos.state(), nullptr };
#ifdef VECTOR
- for (IndexType j = 0; j < kHalfDimensions / kTileHeight; ++j)
+ for (IndexType j = 0; j < HalfDimensions / TileHeight; ++j)
{
// Load accumulator
auto accTile = reinterpret_cast<vec_t*>(
- &st->accumulator.accumulation[c][0][j * kTileHeight]);
- for (IndexType k = 0; k < kNumRegs; ++k)
+ &st->accumulator.accumulation[c][0][j * TileHeight]);
+ for (IndexType k = 0; k < NumRegs; ++k)
acc[k] = vec_load(&accTile[k]);
for (IndexType i = 0; info[i]; ++i)
// Difference calculation for the deactivated features
for (const auto index : removed[i])
{
- const IndexType offset = kHalfDimensions * index + j * kTileHeight;
- auto column = reinterpret_cast<const vec_t*>(&weights_[offset]);
- for (IndexType k = 0; k < kNumRegs; ++k)
+ const IndexType offset = HalfDimensions * index + j * TileHeight;
+ auto column = reinterpret_cast<const vec_t*>(&weights[offset]);
+ for (IndexType k = 0; k < NumRegs; ++k)
acc[k] = vec_sub_16(acc[k], column[k]);
}
// Difference calculation for the activated features
for (const auto index : added[i])
{
- const IndexType offset = kHalfDimensions * index + j * kTileHeight;
- auto column = reinterpret_cast<const vec_t*>(&weights_[offset]);
- for (IndexType k = 0; k < kNumRegs; ++k)
+ const IndexType offset = HalfDimensions * index + j * TileHeight;
+ auto column = reinterpret_cast<const vec_t*>(&weights[offset]);
+ for (IndexType k = 0; k < NumRegs; ++k)
acc[k] = vec_add_16(acc[k], column[k]);
}
// Store accumulator
accTile = reinterpret_cast<vec_t*>(
- &info[i]->accumulator.accumulation[c][0][j * kTileHeight]);
- for (IndexType k = 0; k < kNumRegs; ++k)
+ &info[i]->accumulator.accumulation[c][0][j * TileHeight]);
+ for (IndexType k = 0; k < NumRegs; ++k)
vec_store(&accTile[k], acc[k]);
}
}
{
std::memcpy(info[i]->accumulator.accumulation[c][0],
st->accumulator.accumulation[c][0],
- kHalfDimensions * sizeof(BiasType));
+ HalfDimensions * sizeof(BiasType));
st = info[i];
// Difference calculation for the deactivated features
for (const auto index : removed[i])
{
- const IndexType offset = kHalfDimensions * index;
+ const IndexType offset = HalfDimensions * index;
- for (IndexType j = 0; j < kHalfDimensions; ++j)
- st->accumulator.accumulation[c][0][j] -= weights_[offset + j];
+ for (IndexType j = 0; j < HalfDimensions; ++j)
+ st->accumulator.accumulation[c][0][j] -= weights[offset + j];
}
// Difference calculation for the activated features
for (const auto index : added[i])
{
- const IndexType offset = kHalfDimensions * index;
+ const IndexType offset = HalfDimensions * index;
- for (IndexType j = 0; j < kHalfDimensions; ++j)
- st->accumulator.accumulation[c][0][j] += weights_[offset + j];
+ for (IndexType j = 0; j < HalfDimensions; ++j)
+ st->accumulator.accumulation[c][0][j] += weights[offset + j];
}
}
#endif
auto& accumulator = pos.state()->accumulator;
accumulator.state[c] = COMPUTED;
Features::IndexList active;
- Features::HalfKP<Features::Side::kFriend>::AppendActiveIndices(pos, c, &active);
+ Features::HalfKP<Features::Side::Friend>::append_active_indices(pos, c, &active);
#ifdef VECTOR
- for (IndexType j = 0; j < kHalfDimensions / kTileHeight; ++j)
+ for (IndexType j = 0; j < HalfDimensions / TileHeight; ++j)
{
auto biasesTile = reinterpret_cast<const vec_t*>(
- &biases_[j * kTileHeight]);
- for (IndexType k = 0; k < kNumRegs; ++k)
+ &biases[j * TileHeight]);
+ for (IndexType k = 0; k < NumRegs; ++k)
acc[k] = biasesTile[k];
for (const auto index : active)
{
- const IndexType offset = kHalfDimensions * index + j * kTileHeight;
- auto column = reinterpret_cast<const vec_t*>(&weights_[offset]);
+ const IndexType offset = HalfDimensions * index + j * TileHeight;
+ auto column = reinterpret_cast<const vec_t*>(&weights[offset]);
- for (unsigned k = 0; k < kNumRegs; ++k)
+ for (unsigned k = 0; k < NumRegs; ++k)
acc[k] = vec_add_16(acc[k], column[k]);
}
auto accTile = reinterpret_cast<vec_t*>(
- &accumulator.accumulation[c][0][j * kTileHeight]);
- for (unsigned k = 0; k < kNumRegs; k++)
+ &accumulator.accumulation[c][0][j * TileHeight]);
+ for (unsigned k = 0; k < NumRegs; k++)
vec_store(&accTile[k], acc[k]);
}
#else
- std::memcpy(accumulator.accumulation[c][0], biases_,
- kHalfDimensions * sizeof(BiasType));
+ std::memcpy(accumulator.accumulation[c][0], biases,
+ HalfDimensions * sizeof(BiasType));
for (const auto index : active)
{
- const IndexType offset = kHalfDimensions * index;
+ const IndexType offset = HalfDimensions * index;
- for (IndexType j = 0; j < kHalfDimensions; ++j)
- accumulator.accumulation[c][0][j] += weights_[offset + j];
+ for (IndexType j = 0; j < HalfDimensions; ++j)
+ accumulator.accumulation[c][0][j] += weights[offset + j];
}
#endif
}
using BiasType = std::int16_t;
using WeightType = std::int16_t;
- alignas(kCacheLineSize) BiasType biases_[kHalfDimensions];
- alignas(kCacheLineSize)
- WeightType weights_[kHalfDimensions * kInputDimensions];
+ alignas(CacheLineSize) BiasType biases[HalfDimensions];
+ alignas(CacheLineSize)
+ WeightType weights[HalfDimensions * InputDimensions];
};
} // namespace Stockfish::Eval::NNUE
projects / stockfish / blobdiff