Simplify out pext attacks

Passed STC non-regression on avx512icl (https://tests.stockfishchess.org/tests/view/6a421b46f97ff95f78795061)
LLR: 2.96 (-2.94,2.94) <-1.75,0.25>
Total: 81632 W: 21240 L: 21075 D: 39317
Ptnml(0-2): 184, 8907, 22464, 9082, 179

Passed STC non-regression on bmi2 (https://tests.stockfishchess.org/tests/view/6a41f6f9f97ff95f7879503e)
LLR: 2.93 (-2.94,2.94) <-1.75,0.25>
Total: 78112 W: 20411 L: 20245 D: 37456
Ptnml(0-2): 188, 8449, 21632, 8583, 204

After #6845 , pext attacks are slower or equal to parallel-hyperbola attacks which works on all architectures that pext supports. SVE2 supports pext/pdep in vector registers, but it's much slower than the `rbit`-based HQ that we have on ARM, so I don't foresee the code being useful there.

Ofc I don't think we should remove the `bmi2` build or anything, because it's nice for people to be able to freely use pext when testing ideas, and I'm sure we'll some day find another use for it!

closes https://github.com/official-stockfish/Stockfish/pull/6939

No functional change

Co-authored-by: Dubslow <bunslow@gmail.com>
This commit is contained in:
anematode
2026-07-03 20:27:33 +02:00
committed by Joost VandeVondele
co-authored by Dubslow
parent 46428b9346
commit 99489f57dd
3 changed files with 18 additions and 171 deletions
+11 -69
View File
@@ -38,12 +38,6 @@ alignas(64) Magic Magics[SQUARE_NB][2];
}
#ifdef USE_PEXT
using MagicMask = u16;
#else
using MagicMask = Bitboard;
#endif
[[maybe_unused]] static Bitboard line_mask(Square sq, Direction d1, Direction d2) {
Bitboard mask = 0, dest;
for (Direction d : {d1, d2})
@@ -76,7 +70,7 @@ static void init_magics(Magic magics[][2]) {
// Sliding attacks within a rank, indexed by the slider's file and the
// 8-bit rank occupancy, yielding the 8-bit attack set on that rank
constexpr auto RankAttacks = []() {
[[maybe_unused]] constexpr auto RankAttacks = []() {
std::array<std::array<u8, 256>, FILE_NB> table{};
for (int file = 0; file < 8; ++file)
for (int occ = 0; occ < 256; ++occ)
@@ -102,38 +96,15 @@ static void init_dual_magics(DualMagic magics[]) {
#else
namespace {
[[maybe_unused]] constexpr Bitboard constexpr_pext(Bitboard b, Bitboard m) {
Bitboard result = 0, bit = 0;
while (m)
{
Bitboard last = m & -m;
result |= bool(b & last) << bit++;
m ^= last;
}
return result;
}
void init_magics(PieceType pt, Bitboard table[], Magic magics[][2]) {
#ifdef USE_COMPTIME_ATTACKS
constexpr
#endif
void
init_magics(PieceType pt,
MagicMask table[],
Magic magics[][2],
[[maybe_unused]] bool tableAlreadyInit) {
#if !defined(USE_COMPTIME_ATTACKS)
tableAlreadyInit = false;
#endif
#ifndef USE_PEXT
int seeds[][RANK_NB] = {{8977, 44560, 54343, 38998, 5731, 95205, 104912, 17020},
{728, 10316, 55013, 32803, 12281, 15100, 16645, 255}};
Bitboard occupancy[4096];
int epoch[4096] = {}, cnt = 0;
Bitboard reference[4096] = {};
#endif
int size = 0;
int size = 0;
for (Square s = SQ_A1; s <= SQ_H8; ++s)
{
@@ -142,36 +113,21 @@ constexpr
Magic& m = magics[s][pt - BISHOP];
Bitboard attacks = sliding_attack(pt, s, 0);
m.mask = attacks & ~edges;
#ifdef USE_PEXT
m.pseudoAttacks = attacks;
#else
m.shift = (Is64Bit ? 64 : 32) - popcount(m.mask);
#endif
m.attacks = s == SQ_A1 ? table : magics[s - 1][pt - BISHOP].attacks + size;
size = 0;
m.shift = (Is64Bit ? 64 : 32) - popcount(m.mask);
m.attacks = s == SQ_A1 ? table : magics[s - 1][pt - BISHOP].attacks + size;
size = 0;
Bitboard b = 0;
[[maybe_unused]] Bitboard prevSliding = -1;
do
{
#ifdef USE_PEXT
if (!tableAlreadyInit)
{
Bitboard sliding = sliding_attack(pt, s, b);
m.attacks[size] =
sliding != prevSliding ? constexpr_pext(sliding, attacks) : m.attacks[size - 1];
prevSliding = sliding;
}
#else
occupancy[size] = b;
reference[size] = sliding_attack(pt, s, b);
#endif
size++;
b = (b - m.mask) & m.mask;
} while (b);
#ifndef USE_PEXT
PRNG rng(seeds[Is64Bit][rank_of(s)]);
for (int i = 0; i < size;)
@@ -192,26 +148,12 @@ constexpr
break;
}
}
#endif
}
}
#if defined(USE_COMPTIME_ATTACKS) && defined(USE_PEXT)
constexpr auto RookTable = []() {
std::array<u16, 0x19000> result{};
Magic magics[64][2] = {};
init_magics(ROOK, result.data(), magics, false);
return result;
}();
constexpr auto BishopTable = []() {
std::array<u16, 0x1480> result{};
Magic magics[64][2] = {};
init_magics(BISHOP, result.data(), magics, false);
return result;
}();
#elif !defined(USE_DUAL_HYPERBOLA_QUINT) && !defined(USE_HYPERBOLA_QUINT)
std::array<MagicMask, 0x19000> RookTable;
std::array<MagicMask, 0x1480> BishopTable;
#if !defined(USE_DUAL_HYPERBOLA_QUINT) && !defined(USE_HYPERBOLA_QUINT)
static std::array<Bitboard, 0x19000> RookTable;
static std::array<Bitboard, 0x1480> BishopTable;
#endif
}
@@ -224,8 +166,8 @@ void init() {
#elif defined(USE_DUAL_HYPERBOLA_QUINT)
init_dual_magics(DualMagics);
#else
init_magics(ROOK, const_cast<MagicMask*>(RookTable.data()), Magics, true);
init_magics(BISHOP, const_cast<MagicMask*>(BishopTable.data()), Magics, true);
init_magics(ROOK, RookTable.data(), Magics);
init_magics(BISHOP, BishopTable.data(), Magics);
#endif
for (Square s1 = SQ_A1; s1 <= SQ_H8; ++s1)
+3 -17
View File
@@ -31,7 +31,7 @@
#define USE_HYPERBOLA_QUINT
#elif defined(__loongarch__) && __loongarch_grlen == 64
#define USE_HYPERBOLA_QUINT
#elif defined(USE_AVX2) && !defined(USE_PEXT)
#elif defined(USE_AVX2)
#include <immintrin.h>
#define USE_DUAL_HYPERBOLA_QUINT
#endif
@@ -75,7 +75,7 @@ const Magic& magic(Square s, PieceType pt);
#elif defined(USE_DUAL_HYPERBOLA_QUINT)
struct DualMagic {
struct alignas(32) DualMagic {
// file, diagonal, unused, antidiagonal
Bitboard maskFile, maskDiag, maskNone, maskAntidiag;
// Precomputed 2 * square_bb(sq), 2 * reverse(square_bb(sq))
@@ -128,37 +128,23 @@ const DualMagic& dual_magic(Square s);
#else
// Magic holds all magic bitboards relevant data for a single square
struct Magic {
Bitboard mask;
#ifdef USE_PEXT
u16* attacks;
Bitboard pseudoAttacks;
#else
Bitboard mask;
Bitboard* attacks;
Bitboard magic;
unsigned shift;
#endif
// Compute the attack's index using the 'magic bitboards' approach
unsigned index(Bitboard occupied) const {
#ifdef USE_PEXT
return unsigned(pext(occupied, mask));
#else
if (Is64Bit)
return unsigned(((occupied & mask) * magic) >> shift);
unsigned lo = unsigned(occupied) & unsigned(mask);
unsigned hi = unsigned(occupied >> 32) & unsigned(mask >> 32);
return (lo * unsigned(magic) ^ hi * unsigned(magic >> 32)) >> shift;
#endif
}
Bitboard attacks_bb([[maybe_unused]] Square s, Bitboard occupied) const {
#ifdef USE_PEXT
return pdep(attacks[index(occupied)], pseudoAttacks);
#else
return attacks[index(occupied)];
#endif
}
};
+4 -85
View File
@@ -63,76 +63,6 @@ inline Move* splat_moves(Move* moveList, Square from, Bitboard to_bb) {
return moveList + popcount(to_bb);
}
// Rook/bishop, indexed by (Pt - BISHOP) and from sq
// Moves are provided in ascending order of the piece's attacks on an empty board
alignas(64) constexpr auto SliderMoves = []() {
std::array<std::array<std::array<Move, 16>, SQUARE_NB>, 2> arr{};
for (PieceType pt : {BISHOP, ROOK})
{
for (Square s = SQ_A1; s <= SQ_H8; ++s)
{
Bitboard bb = Attacks::PseudoAttacks[pt][s];
int i = 0;
while (bb)
{
arr[pt - BISHOP][s][i++] = Move(s, Square(constexpr_lsb(bb)));
bb &= bb - 1;
}
}
}
return arr;
}();
// Knight/king analog of the above
alignas(64) constexpr auto KnightKingMoves = []() {
std::array<std::array<std::array<Move, 8>, SQUARE_NB>, 2> arr{};
for (PieceType pt : {KNIGHT, KING})
{
for (Square s = SQ_A1; s <= SQ_H8; ++s)
{
Bitboard bb = Attacks::PseudoAttacks[pt][s];
int i = 0;
while (bb)
{
arr[pt == KING][s][i++] = Move(s, Square(constexpr_lsb(bb)));
bb &= bb - 1;
}
}
}
return arr;
}();
template<PieceType Pt>
inline Move*
splat_precomputed_moves(Move* moveList, Square from, Bitboard occupied, Bitboard target) {
static_assert(Pt != QUEEN && Pt != PAWN, "Unsupported piece type");
// The nth bit in the mask corresponds to the nth square in the piece's pseudo-attacks
u32 mask;
if constexpr (Pt == BISHOP || Pt == ROOK)
{
const Attacks::Magic& magic = Attacks::magic(from, Pt);
mask = magic.attacks[magic.index(occupied)];
mask &= pext(target, magic.pseudoAttacks);
const __m256i moves =
*reinterpret_cast<const __m256i*>(SliderMoves[Pt - BISHOP][from].data());
_mm256_storeu_si256(reinterpret_cast<__m256i*>(moveList),
_mm256_maskz_compress_epi16(mask, moves));
}
else
{
mask = pext(target, Attacks::PseudoAttacks[Pt][from]);
__m128i moves = *reinterpret_cast<const __m128i*>(KnightKingMoves[Pt == KING][from].data());
_mm_storeu_si128(reinterpret_cast<__m128i*>(moveList),
_mm_maskz_compress_epi16(mask, moves));
}
return moveList + popcount(mask);
}
#else
template<Direction offset>
@@ -263,15 +193,8 @@ Move* generate_moves(const Position& pos, Move* moveList, Bitboard target) {
while (bb)
{
Square from = pop_lsb(bb);
#ifdef USE_AVX512ICL
if constexpr (Pt != QUEEN)
{
moveList = splat_precomputed_moves<Pt>(moveList, from, pos.pieces(), target);
continue;
}
#endif
Bitboard b = Attacks::attacks_bb<Pt>(from, pos.pieces()) & target;
Square from = pop_lsb(bb);
Bitboard b = Attacks::attacks_bb<Pt>(from, pos.pieces()) & target;
moveList = splat_moves(moveList, from, b);
}
@@ -303,13 +226,9 @@ Move* generate_all(const Position& pos, Move* moveList) {
moveList = generate_moves<Us, QUEEN>(pos, moveList, target);
}
Bitboard b = Type == EVASIONS ? ~pos.pieces(Us) : target;
Bitboard b = Attacks::attacks_bb<KING>(ksq) & (Type == EVASIONS ? ~pos.pieces(Us) : target);
#ifdef USE_AVX512ICL
moveList = splat_precomputed_moves<KING>(moveList, ksq, 0ULL, b);
#else
moveList = splat_moves(moveList, ksq, Attacks::attacks_bb<KING>(ksq) & b);
#endif
moveList = splat_moves(moveList, ksq, b);
if ((Type == QUIETS || Type == NON_EVASIONS) && pos.can_castle(Us & ANY_CASTLING))
for (CastlingRights cr : {Us & KING_SIDE, Us & QUEEN_SIDE})