1/* 2 * DSP utils : QNS functions are compiled 3 times for mmx/3dnow/ssse3 3 * Copyright (c) 2004 Michael Niedermayer 4 * 5 * MMX optimization by Michael Niedermayer <michaelni@gmx.at> 6 * 3DNow! and SSSE3 optimization by Zuxy Meng <zuxy.meng@gmail.com> 7 * 8 * This file is part of Libav. 9 * 10 * Libav is free software; you can redistribute it and/or 11 * modify it under the terms of the GNU Lesser General Public 12 * License as published by the Free Software Foundation; either 13 * version 2.1 of the License, or (at your option) any later version. 14 * 15 * Libav is distributed in the hope that it will be useful, 16 * but WITHOUT ANY WARRANTY; without even the implied warranty of 17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 18 * Lesser General Public License for more details. 19 * 20 * You should have received a copy of the GNU Lesser General Public 21 * License along with Libav; if not, write to the Free Software 22 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA 23 */ 24 25#define MAX_ABS (512 >> (SCALE_OFFSET>0 ? SCALE_OFFSET : 0)) 26 27static int DEF(try_8x8basis)(int16_t rem[64], int16_t weight[64], int16_t basis[64], int scale) 28{ 29 x86_reg i=0; 30 31 assert(FFABS(scale) < MAX_ABS); 32 scale<<= 16 + SCALE_OFFSET - BASIS_SHIFT + RECON_SHIFT; 33 34 SET_RND(mm6); 35 __asm__ volatile( 36 "pxor %%mm7, %%mm7 \n\t" 37 "movd %4, %%mm5 \n\t" 38 "punpcklwd %%mm5, %%mm5 \n\t" 39 "punpcklwd %%mm5, %%mm5 \n\t" 40 ".p2align 4 \n\t" 41 "1: \n\t" 42 "movq (%1, %0), %%mm0 \n\t" 43 "movq 8(%1, %0), %%mm1 \n\t" 44 PMULHRW(%%mm0, %%mm1, %%mm5, %%mm6) 45 "paddw (%2, %0), %%mm0 \n\t" 46 "paddw 8(%2, %0), %%mm1 \n\t" 47 "psraw $6, %%mm0 \n\t" 48 "psraw $6, %%mm1 \n\t" 49 "pmullw (%3, %0), %%mm0 \n\t" 50 "pmullw 8(%3, %0), %%mm1 \n\t" 51 "pmaddwd %%mm0, %%mm0 \n\t" 52 "pmaddwd %%mm1, %%mm1 \n\t" 53 "paddd %%mm1, %%mm0 \n\t" 54 "psrld $4, %%mm0 \n\t" 55 "paddd %%mm0, %%mm7 \n\t" 56 "add $16, %0 \n\t" 57 "cmp $128, %0 \n\t" //FIXME optimize & bench 58 " jb 1b \n\t" 59 PHADDD(%%mm7, %%mm6) 60 "psrld $2, %%mm7 \n\t" 61 "movd %%mm7, %0 \n\t" 62 63 : "+r" (i) 64 : "r"(basis), "r"(rem), "r"(weight), "g"(scale) 65 ); 66 return i; 67} 68 69static void DEF(add_8x8basis)(int16_t rem[64], int16_t basis[64], int scale) 70{ 71 x86_reg i=0; 72 73 if(FFABS(scale) < MAX_ABS){ 74 scale<<= 16 + SCALE_OFFSET - BASIS_SHIFT + RECON_SHIFT; 75 SET_RND(mm6); 76 __asm__ volatile( 77 "movd %3, %%mm5 \n\t" 78 "punpcklwd %%mm5, %%mm5 \n\t" 79 "punpcklwd %%mm5, %%mm5 \n\t" 80 ".p2align 4 \n\t" 81 "1: \n\t" 82 "movq (%1, %0), %%mm0 \n\t" 83 "movq 8(%1, %0), %%mm1 \n\t" 84 PMULHRW(%%mm0, %%mm1, %%mm5, %%mm6) 85 "paddw (%2, %0), %%mm0 \n\t" 86 "paddw 8(%2, %0), %%mm1 \n\t" 87 "movq %%mm0, (%2, %0) \n\t" 88 "movq %%mm1, 8(%2, %0) \n\t" 89 "add $16, %0 \n\t" 90 "cmp $128, %0 \n\t" // FIXME optimize & bench 91 " jb 1b \n\t" 92 93 : "+r" (i) 94 : "r"(basis), "r"(rem), "g"(scale) 95 ); 96 }else{ 97 for(i=0; i<8*8; i++){ 98 rem[i] += (basis[i]*scale + (1<<(BASIS_SHIFT - RECON_SHIFT-1)))>>(BASIS_SHIFT - RECON_SHIFT); 99 } 100 } 101} 102