SIMD Techniques

From Chessprogramming wiki
Jump to: navigation, search

Home * Board Representation * Bitboards * Sliding Piece Attacks * SIMD techniques

x86 MMX- or SSE2 SIMD-instruction sets provide a Packed Move Mask Byte, pmovmskb-instruction [1], available in C/C++ as _mm_movemask_epi intrinsic, which moves the most-significant bits of each byte of a MMX- or XMM-register to the lowest 8 or 16 bits of a general purpose register. Thus, this instruction may be used to map file- or diagonal occupancies to consecutive bits.

Bishop Attacks

For diagonals one may mask and compare byte-wise to get the occupancy to the sign-bits. With SSE2 and 128-bit XMM-registers one may process both diagonal- and anti-diagonal-occupancies in one run:

u64 fillRightAttacks[8][64]; // [file][occupiedIndex]
__m128i xmmBmask[64]; // antidiagonal::diagonal -  masks

U64 bishopAttacksSSE2(U64 occ, unsigned int sq) {
   __m128 mocc;
   mocc = _mm_cvtsi64x_si128(occ);            // gp to xmm, 0:occ
   mocc = _mm_unpacklo_epi64(mocc, mocc);     // occupancy to both xmm-halfs, occ:occ
   mocc = _mm_and_si128 (mocc, xmmBmask[sq]); // mask diagonal and antidiagonal
   mocc = _mm_cmpeq_epi8(mocc, xmmBmask[sq]); // cmp bytewise equal, FF if set, 00 otherwise
   unsigned int o = _mm_movemask_epi(mocc);   // get the 16 sign bits
   return (xmmBmask[sq].m128i_u64[0] & fillRightAttacks[sq>>3][(o>>1)&63])
        | (xmmBmask[sq].m128i_u64[1] & fillRightAttacks[sq>>3][(o>>9)&63]);

This sample code uses a shared 4KByte fill right lookup similar to fillUpAttacks of kindergarten bitboards. Of course one may use distinct lookup tables similar to rotated bitboards indexed by square and occupied-state without the trailing mask ands.

See also

Forum Posts

External Links


Up one Level