#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)
+ #define vec_mul_16(a,b) _mm512_mullo_epi16(a,b)
+ #define vec_zero() _mm512_setzero_epi32()
+ #define vec_set_16(a) _mm512_set1_epi16(a)
+ #define vec_max_16(a,b) _mm512_max_epi16(a,b)
+ #define vec_min_16(a,b) _mm512_min_epi16(a,b)
+ inline vec_t vec_msb_pack_16(vec_t a, vec_t b){
+ vec_t compacted = _mm512_packs_epi16(_mm512_srli_epi16(a,7),_mm512_srli_epi16(b,7));
+ return _mm512_permutexvar_epi64(_mm512_setr_epi64(0, 2, 4, 6, 1, 3, 5, 7), compacted);
+ }
#define vec_load_psqt(a) _mm256_load_si256(a)
#define vec_store_psqt(a,b) _mm256_store_si256(a,b)
#define vec_add_psqt_32(a,b) _mm256_add_epi32(a,b)
#define vec_sub_psqt_32(a,b) _mm256_sub_epi32(a,b)
#define vec_zero_psqt() _mm256_setzero_si256()
#define NumRegistersSIMD 32
+ #define MaxChunkSize 64
#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)
+ #define vec_mul_16(a,b) _mm256_mullo_epi16(a,b)
+ #define vec_zero() _mm256_setzero_si256()
+ #define vec_set_16(a) _mm256_set1_epi16(a)
+ #define vec_max_16(a,b) _mm256_max_epi16(a,b)
+ #define vec_min_16(a,b) _mm256_min_epi16(a,b)
+ inline vec_t vec_msb_pack_16(vec_t a, vec_t b){
+ vec_t compacted = _mm256_packs_epi16(_mm256_srli_epi16(a,7), _mm256_srli_epi16(b,7));
+ return _mm256_permute4x64_epi64(compacted, 0b11011000);
+ }
#define vec_load_psqt(a) _mm256_load_si256(a)
#define vec_store_psqt(a,b) _mm256_store_si256(a,b)
#define vec_add_psqt_32(a,b) _mm256_add_epi32(a,b)
#define vec_sub_psqt_32(a,b) _mm256_sub_epi32(a,b)
#define vec_zero_psqt() _mm256_setzero_si256()
#define NumRegistersSIMD 16
+ #define MaxChunkSize 32
#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)
+ #define vec_mul_16(a,b) _mm_mullo_epi16(a,b)
+ #define vec_zero() _mm_setzero_si128()
+ #define vec_set_16(a) _mm_set1_epi16(a)
+ #define vec_max_16(a,b) _mm_max_epi16(a,b)
+ #define vec_min_16(a,b) _mm_min_epi16(a,b)
+ #define vec_msb_pack_16(a,b) _mm_packs_epi16(_mm_srli_epi16(a,7),_mm_srli_epi16(b,7))
#define vec_load_psqt(a) (*(a))
#define vec_store_psqt(a,b) *(a)=(b)
#define vec_add_psqt_32(a,b) _mm_add_epi32(a,b)
#define vec_sub_psqt_32(a,b) _mm_sub_epi32(a,b)
#define vec_zero_psqt() _mm_setzero_si128()
#define NumRegistersSIMD (Is64Bit ? 16 : 8)
+ #define MaxChunkSize 16
#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)
+ #define vec_mul_16(a,b) _mm_mullo_pi16(a,b)
+ #define vec_zero() _mm_setzero_si64()
+ #define vec_set_16(a) _mm_set1_pi16(a)
+ inline vec_t vec_max_16(vec_t a,vec_t b){
+ vec_t comparison = _mm_cmpgt_pi16(a,b);
+ return _mm_or_si64(_mm_and_si64(comparison, a), _mm_andnot_si64(comparison, b));
+ }
+ inline vec_t vec_min_16(vec_t a,vec_t b){
+ vec_t comparison = _mm_cmpgt_pi16(a,b);
+ return _mm_or_si64(_mm_and_si64(comparison, b), _mm_andnot_si64(comparison, a));
+ }
+ #define vec_msb_pack_16(a,b) _mm_packs_pi16(_mm_srli_pi16(a,7),_mm_srli_pi16(b,7))
#define vec_load_psqt(a) (*(a))
#define vec_store_psqt(a,b) *(a)=(b)
#define vec_add_psqt_32(a,b) _mm_add_pi32(a,b)
#define vec_sub_psqt_32(a,b) _mm_sub_pi32(a,b)
#define vec_zero_psqt() _mm_setzero_si64()
+ #define vec_cleanup() _mm_empty()
#define NumRegistersSIMD 8
+ #define MaxChunkSize 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)
+ #define vec_mul_16(a,b) vmulq_s16(a,b)
+ #define vec_zero() vec_t{0}
+ #define vec_set_16(a) vdupq_n_s16(a)
+ #define vec_max_16(a,b) vmaxq_s16(a,b)
+ #define vec_min_16(a,b) vminq_s16(a,b)
+ inline vec_t vec_msb_pack_16(vec_t a, vec_t b){
+ const int8x8_t shifta = vshrn_n_s16(a, 7);
+ const int8x8_t shiftb = vshrn_n_s16(b, 7);
+ const int8x16_t compacted = vcombine_s8(shifta,shiftb);
+ return *reinterpret_cast<const vec_t*> (&compacted);
+ }
#define vec_load_psqt(a) (*(a))
#define vec_store_psqt(a,b) *(a)=(b)
#define vec_add_psqt_32(a,b) vaddq_s32(a,b)
#define vec_sub_psqt_32(a,b) vsubq_s32(a,b)
#define vec_zero_psqt() psqt_vec_t{0}
#define NumRegistersSIMD 16
+ #define MaxChunkSize 16
#else
#undef VECTOR
// We use __m* types as template arguments, which causes GCC to emit warnings
// about losing some attribute information. This is irrelevant to us as we
// only take their size, so the following pragma are harmless.
+ #if defined(__GNUC__)
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wignored-attributes"
+ #endif
template <typename SIMDRegisterType,
typename LaneType,
static constexpr int NumRegs = BestRegisterCount<vec_t, WeightType, TransformedFeatureDimensions, NumRegistersSIMD>();
static constexpr int NumPsqtRegs = BestRegisterCount<psqt_vec_t, PSQTWeightType, PSQTBuckets, NumRegistersSIMD>();
-
+ #if defined(__GNUC__)
#pragma GCC diagnostic pop
-
+ #endif
#endif
// Convert input features
std::int32_t transform(const Position& pos, OutputType* output, int bucket) const {
- update_accumulator(pos, WHITE);
- update_accumulator(pos, BLACK);
+ update_accumulator<WHITE>(pos);
+ update_accumulator<BLACK>(pos);
const Color perspectives[2] = {pos.side_to_move(), ~pos.side_to_move()};
const auto& accumulation = pos.state()->accumulator.accumulation;
{
const IndexType offset = (HalfDimensions / 2) * p;
-#if defined(USE_AVX512)
-
- constexpr IndexType OutputChunkSize = 512 / 8;
- static_assert((HalfDimensions / 2) % OutputChunkSize == 0);
- constexpr IndexType NumOutputChunks = HalfDimensions / 2 / OutputChunkSize;
-
- const __m512i Zero = _mm512_setzero_si512();
- const __m512i One = _mm512_set1_epi16(127);
- const __m512i Control = _mm512_setr_epi64(0, 2, 4, 6, 1, 3, 5, 7);
-
- const __m512i* in0 = reinterpret_cast<const __m512i*>(&(accumulation[perspectives[p]][0]));
- const __m512i* in1 = reinterpret_cast<const __m512i*>(&(accumulation[perspectives[p]][HalfDimensions / 2]));
- __m512i* out = reinterpret_cast< __m512i*>(output + offset);
-
- for (IndexType j = 0; j < NumOutputChunks; j += 1)
- {
- const __m512i sum0a = _mm512_max_epi16(_mm512_min_epi16(in0[j * 2 + 0], One), Zero);
- const __m512i sum0b = _mm512_max_epi16(_mm512_min_epi16(in0[j * 2 + 1], One), Zero);
- const __m512i sum1a = _mm512_max_epi16(_mm512_min_epi16(in1[j * 2 + 0], One), Zero);
- const __m512i sum1b = _mm512_max_epi16(_mm512_min_epi16(in1[j * 2 + 1], One), Zero);
-
- const __m512i pa = _mm512_srli_epi16(_mm512_mullo_epi16(sum0a, sum1a), 7);
- const __m512i pb = _mm512_srli_epi16(_mm512_mullo_epi16(sum0b, sum1b), 7);
-
- out[j] = _mm512_permutexvar_epi64(Control, _mm512_packs_epi16(pa, pb));
- }
-
-#elif defined(USE_AVX2)
-
- constexpr IndexType OutputChunkSize = 256 / 8;
- static_assert((HalfDimensions / 2) % OutputChunkSize == 0);
- constexpr IndexType NumOutputChunks = HalfDimensions / 2 / OutputChunkSize;
-
- const __m256i Zero = _mm256_setzero_si256();
- const __m256i One = _mm256_set1_epi16(127);
- constexpr int Control = 0b11011000;
-
- const __m256i* in0 = reinterpret_cast<const __m256i*>(&(accumulation[perspectives[p]][0]));
- const __m256i* in1 = reinterpret_cast<const __m256i*>(&(accumulation[perspectives[p]][HalfDimensions / 2]));
- __m256i* out = reinterpret_cast< __m256i*>(output + offset);
-
- for (IndexType j = 0; j < NumOutputChunks; j += 1)
- {
- const __m256i sum0a = _mm256_max_epi16(_mm256_min_epi16(in0[j * 2 + 0], One), Zero);
- const __m256i sum0b = _mm256_max_epi16(_mm256_min_epi16(in0[j * 2 + 1], One), Zero);
- const __m256i sum1a = _mm256_max_epi16(_mm256_min_epi16(in1[j * 2 + 0], One), Zero);
- const __m256i sum1b = _mm256_max_epi16(_mm256_min_epi16(in1[j * 2 + 1], One), Zero);
-
- const __m256i pa = _mm256_srli_epi16(_mm256_mullo_epi16(sum0a, sum1a), 7);
- const __m256i pb = _mm256_srli_epi16(_mm256_mullo_epi16(sum0b, sum1b), 7);
-
- out[j] = _mm256_permute4x64_epi64(_mm256_packs_epi16(pa, pb), Control);
- }
-
-#elif defined(USE_SSE2)
-
- constexpr IndexType OutputChunkSize = 128 / 8;
- static_assert((HalfDimensions / 2) % OutputChunkSize == 0);
- constexpr IndexType NumOutputChunks = HalfDimensions / 2 / OutputChunkSize;
-
- const __m128i Zero = _mm_setzero_si128();
- const __m128i One = _mm_set1_epi16(127);
+#if defined(VECTOR)
- const __m128i* in0 = reinterpret_cast<const __m128i*>(&(accumulation[perspectives[p]][0]));
- const __m128i* in1 = reinterpret_cast<const __m128i*>(&(accumulation[perspectives[p]][HalfDimensions / 2]));
- __m128i* out = reinterpret_cast< __m128i*>(output + offset);
-
- for (IndexType j = 0; j < NumOutputChunks; j += 1)
- {
- const __m128i sum0a = _mm_max_epi16(_mm_min_epi16(in0[j * 2 + 0], One), Zero);
- const __m128i sum0b = _mm_max_epi16(_mm_min_epi16(in0[j * 2 + 1], One), Zero);
- const __m128i sum1a = _mm_max_epi16(_mm_min_epi16(in1[j * 2 + 0], One), Zero);
- const __m128i sum1b = _mm_max_epi16(_mm_min_epi16(in1[j * 2 + 1], One), Zero);
-
- const __m128i pa = _mm_srli_epi16(_mm_mullo_epi16(sum0a, sum1a), 7);
- const __m128i pb = _mm_srli_epi16(_mm_mullo_epi16(sum0b, sum1b), 7);
-
- out[j] = _mm_packs_epi16(pa, pb);
- }
-
-#elif defined(USE_NEON)
-
- constexpr IndexType OutputChunkSize = 128 / 8;
+ constexpr IndexType OutputChunkSize = MaxChunkSize;
static_assert((HalfDimensions / 2) % OutputChunkSize == 0);
constexpr IndexType NumOutputChunks = HalfDimensions / 2 / OutputChunkSize;
- const int16x8_t Zero = vdupq_n_s16(0);
- const int16x8_t One = vdupq_n_s16(127);
+ vec_t Zero = vec_zero();
+ vec_t One = vec_set_16(127);
- const int16x8_t* in0 = reinterpret_cast<const int16x8_t*>(&(accumulation[perspectives[p]][0]));
- const int16x8_t* in1 = reinterpret_cast<const int16x8_t*>(&(accumulation[perspectives[p]][HalfDimensions / 2]));
- int8x16_t* out = reinterpret_cast< int8x16_t*>(output + offset);
+ const vec_t* in0 = reinterpret_cast<const vec_t*>(&(accumulation[perspectives[p]][0]));
+ const vec_t* in1 = reinterpret_cast<const vec_t*>(&(accumulation[perspectives[p]][HalfDimensions / 2]));
+ vec_t* out = reinterpret_cast< vec_t*>(output + offset);
for (IndexType j = 0; j < NumOutputChunks; j += 1)
{
- const int16x8_t sum0a = vmaxq_s16(vminq_s16(in0[j * 2 + 0], One), Zero);
- const int16x8_t sum0b = vmaxq_s16(vminq_s16(in0[j * 2 + 1], One), Zero);
- const int16x8_t sum1a = vmaxq_s16(vminq_s16(in1[j * 2 + 0], One), Zero);
- const int16x8_t sum1b = vmaxq_s16(vminq_s16(in1[j * 2 + 1], One), Zero);
+ const vec_t sum0a = vec_max_16(vec_min_16(in0[j * 2 + 0], One), Zero);
+ const vec_t sum0b = vec_max_16(vec_min_16(in0[j * 2 + 1], One), Zero);
+ const vec_t sum1a = vec_max_16(vec_min_16(in1[j * 2 + 0], One), Zero);
+ const vec_t sum1b = vec_max_16(vec_min_16(in1[j * 2 + 1], One), Zero);
- const int8x8_t pa = vshrn_n_s16(vmulq_s16(sum0a, sum1a), 7);
- const int8x8_t pb = vshrn_n_s16(vmulq_s16(sum0b, sum1b), 7);
+ const vec_t pa = vec_mul_16(sum0a, sum1a);
+ const vec_t pb = vec_mul_16(sum0b, sum1b);
- out[j] = vcombine_s8(pa, pb);
+ out[j] = vec_msb_pack_16(pa, pb);
}
#else
#endif
}
+#if defined(vec_cleanup)
+ vec_cleanup();
+#endif
+
return psqt;
} // end of function transform()
private:
- void update_accumulator(const Position& pos, const Color perspective) const {
+ template<Color Perspective>
+ void update_accumulator(const Position& pos) const {
// The size must be enough to contain the largest possible update.
// That might depend on the feature set and generally relies on the
// of the estimated gain in terms of features to be added/subtracted.
StateInfo *st = pos.state(), *next = nullptr;
int gain = FeatureSet::refresh_cost(pos);
- while (st->previous && !st->accumulator.computed[perspective])
+ while (st->previous && !st->accumulator.computed[Perspective])
{
// This governs when a full feature refresh is needed and how many
// updates are better than just one full refresh.
- if ( FeatureSet::requires_refresh(st, perspective)
+ if ( FeatureSet::requires_refresh(st, Perspective)
|| (gain -= FeatureSet::update_cost(st) + 1) < 0)
break;
next = st;
st = st->previous;
}
- if (st->accumulator.computed[perspective])
+ if (st->accumulator.computed[Perspective])
{
if (next == nullptr)
return;
// accumulator. Then, we update the current accumulator (pos.state()).
// Gather all features to be updated.
- const Square ksq = pos.square<KING>(perspective);
+ const Square ksq = pos.square<KING>(Perspective);
FeatureSet::IndexList removed[2], added[2];
- FeatureSet::append_changed_indices(
- ksq, next->dirtyPiece, perspective, removed[0], added[0]);
+ FeatureSet::append_changed_indices<Perspective>(
+ ksq, next->dirtyPiece, removed[0], added[0]);
for (StateInfo *st2 = pos.state(); st2 != next; st2 = st2->previous)
- FeatureSet::append_changed_indices(
- ksq, st2->dirtyPiece, perspective, removed[1], added[1]);
+ FeatureSet::append_changed_indices<Perspective>(
+ ksq, st2->dirtyPiece, removed[1], added[1]);
// Mark the accumulators as computed.
- next->accumulator.computed[perspective] = true;
- pos.state()->accumulator.computed[perspective] = true;
+ next->accumulator.computed[Perspective] = true;
+ pos.state()->accumulator.computed[Perspective] = true;
// Now update the accumulators listed in states_to_update[], where the last element is a sentinel.
StateInfo *states_to_update[3] =
{
// Load accumulator
auto accTile = reinterpret_cast<vec_t*>(
- &st->accumulator.accumulation[perspective][j * TileHeight]);
+ &st->accumulator.accumulation[Perspective][j * TileHeight]);
for (IndexType k = 0; k < NumRegs; ++k)
acc[k] = vec_load(&accTile[k]);
// Store accumulator
accTile = reinterpret_cast<vec_t*>(
- &states_to_update[i]->accumulator.accumulation[perspective][j * TileHeight]);
+ &states_to_update[i]->accumulator.accumulation[Perspective][j * TileHeight]);
for (IndexType k = 0; k < NumRegs; ++k)
vec_store(&accTile[k], acc[k]);
}
{
// Load accumulator
auto accTilePsqt = reinterpret_cast<psqt_vec_t*>(
- &st->accumulator.psqtAccumulation[perspective][j * PsqtTileHeight]);
+ &st->accumulator.psqtAccumulation[Perspective][j * PsqtTileHeight]);
for (std::size_t k = 0; k < NumPsqtRegs; ++k)
psqt[k] = vec_load_psqt(&accTilePsqt[k]);
// Store accumulator
accTilePsqt = reinterpret_cast<psqt_vec_t*>(
- &states_to_update[i]->accumulator.psqtAccumulation[perspective][j * PsqtTileHeight]);
+ &states_to_update[i]->accumulator.psqtAccumulation[Perspective][j * PsqtTileHeight]);
for (std::size_t k = 0; k < NumPsqtRegs; ++k)
vec_store_psqt(&accTilePsqt[k], psqt[k]);
}
#else
for (IndexType i = 0; states_to_update[i]; ++i)
{
- std::memcpy(states_to_update[i]->accumulator.accumulation[perspective],
- st->accumulator.accumulation[perspective],
+ std::memcpy(states_to_update[i]->accumulator.accumulation[Perspective],
+ st->accumulator.accumulation[Perspective],
HalfDimensions * sizeof(BiasType));
for (std::size_t k = 0; k < PSQTBuckets; ++k)
- states_to_update[i]->accumulator.psqtAccumulation[perspective][k] = st->accumulator.psqtAccumulation[perspective][k];
+ states_to_update[i]->accumulator.psqtAccumulation[Perspective][k] = st->accumulator.psqtAccumulation[Perspective][k];
st = states_to_update[i];
const IndexType offset = HalfDimensions * index;
for (IndexType j = 0; j < HalfDimensions; ++j)
- st->accumulator.accumulation[perspective][j] -= weights[offset + j];
+ st->accumulator.accumulation[Perspective][j] -= weights[offset + j];
for (std::size_t k = 0; k < PSQTBuckets; ++k)
- st->accumulator.psqtAccumulation[perspective][k] -= psqtWeights[index * PSQTBuckets + k];
+ st->accumulator.psqtAccumulation[Perspective][k] -= psqtWeights[index * PSQTBuckets + k];
}
// Difference calculation for the activated features
const IndexType offset = HalfDimensions * index;
for (IndexType j = 0; j < HalfDimensions; ++j)
- st->accumulator.accumulation[perspective][j] += weights[offset + j];
+ st->accumulator.accumulation[Perspective][j] += weights[offset + j];
for (std::size_t k = 0; k < PSQTBuckets; ++k)
- st->accumulator.psqtAccumulation[perspective][k] += psqtWeights[index * PSQTBuckets + k];
+ st->accumulator.psqtAccumulation[Perspective][k] += psqtWeights[index * PSQTBuckets + k];
}
}
#endif
{
// Refresh the accumulator
auto& accumulator = pos.state()->accumulator;
- accumulator.computed[perspective] = true;
+ accumulator.computed[Perspective] = true;
FeatureSet::IndexList active;
- FeatureSet::append_active_indices(pos, perspective, active);
+ FeatureSet::append_active_indices<Perspective>(pos, active);
#ifdef VECTOR
for (IndexType j = 0; j < HalfDimensions / TileHeight; ++j)
}
auto accTile = reinterpret_cast<vec_t*>(
- &accumulator.accumulation[perspective][j * TileHeight]);
+ &accumulator.accumulation[Perspective][j * TileHeight]);
for (unsigned k = 0; k < NumRegs; k++)
vec_store(&accTile[k], acc[k]);
}
}
auto accTilePsqt = reinterpret_cast<psqt_vec_t*>(
- &accumulator.psqtAccumulation[perspective][j * PsqtTileHeight]);
+ &accumulator.psqtAccumulation[Perspective][j * PsqtTileHeight]);
for (std::size_t k = 0; k < NumPsqtRegs; ++k)
vec_store_psqt(&accTilePsqt[k], psqt[k]);
}
#else
- std::memcpy(accumulator.accumulation[perspective], biases,
+ std::memcpy(accumulator.accumulation[Perspective], biases,
HalfDimensions * sizeof(BiasType));
for (std::size_t k = 0; k < PSQTBuckets; ++k)
- accumulator.psqtAccumulation[perspective][k] = 0;
+ accumulator.psqtAccumulation[Perspective][k] = 0;
for (const auto index : active)
{
const IndexType offset = HalfDimensions * index;
for (IndexType j = 0; j < HalfDimensions; ++j)
- accumulator.accumulation[perspective][j] += weights[offset + j];
+ accumulator.accumulation[Perspective][j] += weights[offset + j];
for (std::size_t k = 0; k < PSQTBuckets; ++k)
- accumulator.psqtAccumulation[perspective][k] += psqtWeights[index * PSQTBuckets + k];
+ accumulator.psqtAccumulation[Perspective][k] += psqtWeights[index * PSQTBuckets + k];
}
#endif
}