yading@10: /* yading@10: * DSP utils mmx functions are compiled twice for rnd/no_rnd yading@10: * Copyright (c) 2000, 2001 Fabrice Bellard yading@10: * Copyright (c) 2003-2004 Michael Niedermayer yading@10: * yading@10: * MMX optimization by Nick Kurshev yading@10: * mostly rewritten by Michael Niedermayer yading@10: * and improved by Zdenek Kabelac yading@10: * yading@10: * This file is part of FFmpeg. yading@10: * yading@10: * FFmpeg is free software; you can redistribute it and/or yading@10: * modify it under the terms of the GNU Lesser General Public yading@10: * License as published by the Free Software Foundation; either yading@10: * version 2.1 of the License, or (at your option) any later version. yading@10: * yading@10: * FFmpeg is distributed in the hope that it will be useful, yading@10: * but WITHOUT ANY WARRANTY; without even the implied warranty of yading@10: * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU yading@10: * Lesser General Public License for more details. yading@10: * yading@10: * You should have received a copy of the GNU Lesser General Public yading@10: * License along with FFmpeg; if not, write to the Free Software yading@10: * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA yading@10: */ yading@10: yading@10: // put_pixels yading@10: static void DEF(put, pixels8_xy2)(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h) yading@10: { yading@10: MOVQ_ZERO(mm7); yading@10: SET_RND(mm6); // =2 for rnd and =1 for no_rnd version yading@10: __asm__ volatile( yading@10: "movq (%1), %%mm0 \n\t" yading@10: "movq 1(%1), %%mm4 \n\t" yading@10: "movq %%mm0, %%mm1 \n\t" yading@10: "movq %%mm4, %%mm5 \n\t" yading@10: "punpcklbw %%mm7, %%mm0 \n\t" yading@10: "punpcklbw %%mm7, %%mm4 \n\t" yading@10: "punpckhbw %%mm7, %%mm1 \n\t" yading@10: "punpckhbw %%mm7, %%mm5 \n\t" yading@10: "paddusw %%mm0, %%mm4 \n\t" yading@10: "paddusw %%mm1, %%mm5 \n\t" yading@10: "xor %%"REG_a", %%"REG_a" \n\t" yading@10: "add %3, %1 \n\t" yading@10: ".p2align 3 \n\t" yading@10: "1: \n\t" yading@10: "movq (%1, %%"REG_a"), %%mm0 \n\t" yading@10: "movq 1(%1, %%"REG_a"), %%mm2 \n\t" yading@10: "movq %%mm0, %%mm1 \n\t" yading@10: "movq %%mm2, %%mm3 \n\t" yading@10: "punpcklbw %%mm7, %%mm0 \n\t" yading@10: "punpcklbw %%mm7, %%mm2 \n\t" yading@10: "punpckhbw %%mm7, %%mm1 \n\t" yading@10: "punpckhbw %%mm7, %%mm3 \n\t" yading@10: "paddusw %%mm2, %%mm0 \n\t" yading@10: "paddusw %%mm3, %%mm1 \n\t" yading@10: "paddusw %%mm6, %%mm4 \n\t" yading@10: "paddusw %%mm6, %%mm5 \n\t" yading@10: "paddusw %%mm0, %%mm4 \n\t" yading@10: "paddusw %%mm1, %%mm5 \n\t" yading@10: "psrlw $2, %%mm4 \n\t" yading@10: "psrlw $2, %%mm5 \n\t" yading@10: "packuswb %%mm5, %%mm4 \n\t" yading@10: "movq %%mm4, (%2, %%"REG_a") \n\t" yading@10: "add %3, %%"REG_a" \n\t" yading@10: yading@10: "movq (%1, %%"REG_a"), %%mm2 \n\t" // 0 <-> 2 1 <-> 3 yading@10: "movq 1(%1, %%"REG_a"), %%mm4 \n\t" yading@10: "movq %%mm2, %%mm3 \n\t" yading@10: "movq %%mm4, %%mm5 \n\t" yading@10: "punpcklbw %%mm7, %%mm2 \n\t" yading@10: "punpcklbw %%mm7, %%mm4 \n\t" yading@10: "punpckhbw %%mm7, %%mm3 \n\t" yading@10: "punpckhbw %%mm7, %%mm5 \n\t" yading@10: "paddusw %%mm2, %%mm4 \n\t" yading@10: "paddusw %%mm3, %%mm5 \n\t" yading@10: "paddusw %%mm6, %%mm0 \n\t" yading@10: "paddusw %%mm6, %%mm1 \n\t" yading@10: "paddusw %%mm4, %%mm0 \n\t" yading@10: "paddusw %%mm5, %%mm1 \n\t" yading@10: "psrlw $2, %%mm0 \n\t" yading@10: "psrlw $2, %%mm1 \n\t" yading@10: "packuswb %%mm1, %%mm0 \n\t" yading@10: "movq %%mm0, (%2, %%"REG_a") \n\t" yading@10: "add %3, %%"REG_a" \n\t" yading@10: yading@10: "subl $2, %0 \n\t" yading@10: "jnz 1b \n\t" yading@10: :"+g"(h), "+S"(pixels) yading@10: :"D"(block), "r"((x86_reg)line_size) yading@10: :REG_a, "memory"); yading@10: } yading@10: yading@10: // in case more speed is needed - unroling would certainly help yading@10: static void DEF(avg, pixels8)(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h) yading@10: { yading@10: MOVQ_BFE(mm6); yading@10: JUMPALIGN(); yading@10: do { yading@10: __asm__ volatile( yading@10: "movq %0, %%mm0 \n\t" yading@10: "movq %1, %%mm1 \n\t" yading@10: OP_AVG(%%mm0, %%mm1, %%mm2, %%mm6) yading@10: "movq %%mm2, %0 \n\t" yading@10: :"+m"(*block) yading@10: :"m"(*pixels) yading@10: :"memory"); yading@10: pixels += line_size; yading@10: block += line_size; yading@10: } yading@10: while (--h); yading@10: } yading@10: yading@10: static void DEF(avg, pixels16)(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h) yading@10: { yading@10: MOVQ_BFE(mm6); yading@10: JUMPALIGN(); yading@10: do { yading@10: __asm__ volatile( yading@10: "movq %0, %%mm0 \n\t" yading@10: "movq %1, %%mm1 \n\t" yading@10: OP_AVG(%%mm0, %%mm1, %%mm2, %%mm6) yading@10: "movq %%mm2, %0 \n\t" yading@10: "movq 8%0, %%mm0 \n\t" yading@10: "movq 8%1, %%mm1 \n\t" yading@10: OP_AVG(%%mm0, %%mm1, %%mm2, %%mm6) yading@10: "movq %%mm2, 8%0 \n\t" yading@10: :"+m"(*block) yading@10: :"m"(*pixels) yading@10: :"memory"); yading@10: pixels += line_size; yading@10: block += line_size; yading@10: } yading@10: while (--h); yading@10: } yading@10: yading@10: // this routine is 'slightly' suboptimal but mostly unused yading@10: static void DEF(avg, pixels8_xy2)(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h) yading@10: { yading@10: MOVQ_ZERO(mm7); yading@10: SET_RND(mm6); // =2 for rnd and =1 for no_rnd version yading@10: __asm__ volatile( yading@10: "movq (%1), %%mm0 \n\t" yading@10: "movq 1(%1), %%mm4 \n\t" yading@10: "movq %%mm0, %%mm1 \n\t" yading@10: "movq %%mm4, %%mm5 \n\t" yading@10: "punpcklbw %%mm7, %%mm0 \n\t" yading@10: "punpcklbw %%mm7, %%mm4 \n\t" yading@10: "punpckhbw %%mm7, %%mm1 \n\t" yading@10: "punpckhbw %%mm7, %%mm5 \n\t" yading@10: "paddusw %%mm0, %%mm4 \n\t" yading@10: "paddusw %%mm1, %%mm5 \n\t" yading@10: "xor %%"REG_a", %%"REG_a" \n\t" yading@10: "add %3, %1 \n\t" yading@10: ".p2align 3 \n\t" yading@10: "1: \n\t" yading@10: "movq (%1, %%"REG_a"), %%mm0 \n\t" yading@10: "movq 1(%1, %%"REG_a"), %%mm2 \n\t" yading@10: "movq %%mm0, %%mm1 \n\t" yading@10: "movq %%mm2, %%mm3 \n\t" yading@10: "punpcklbw %%mm7, %%mm0 \n\t" yading@10: "punpcklbw %%mm7, %%mm2 \n\t" yading@10: "punpckhbw %%mm7, %%mm1 \n\t" yading@10: "punpckhbw %%mm7, %%mm3 \n\t" yading@10: "paddusw %%mm2, %%mm0 \n\t" yading@10: "paddusw %%mm3, %%mm1 \n\t" yading@10: "paddusw %%mm6, %%mm4 \n\t" yading@10: "paddusw %%mm6, %%mm5 \n\t" yading@10: "paddusw %%mm0, %%mm4 \n\t" yading@10: "paddusw %%mm1, %%mm5 \n\t" yading@10: "psrlw $2, %%mm4 \n\t" yading@10: "psrlw $2, %%mm5 \n\t" yading@10: "movq (%2, %%"REG_a"), %%mm3 \n\t" yading@10: "packuswb %%mm5, %%mm4 \n\t" yading@10: "pcmpeqd %%mm2, %%mm2 \n\t" yading@10: "paddb %%mm2, %%mm2 \n\t" yading@10: OP_AVG(%%mm3, %%mm4, %%mm5, %%mm2) yading@10: "movq %%mm5, (%2, %%"REG_a") \n\t" yading@10: "add %3, %%"REG_a" \n\t" yading@10: yading@10: "movq (%1, %%"REG_a"), %%mm2 \n\t" // 0 <-> 2 1 <-> 3 yading@10: "movq 1(%1, %%"REG_a"), %%mm4 \n\t" yading@10: "movq %%mm2, %%mm3 \n\t" yading@10: "movq %%mm4, %%mm5 \n\t" yading@10: "punpcklbw %%mm7, %%mm2 \n\t" yading@10: "punpcklbw %%mm7, %%mm4 \n\t" yading@10: "punpckhbw %%mm7, %%mm3 \n\t" yading@10: "punpckhbw %%mm7, %%mm5 \n\t" yading@10: "paddusw %%mm2, %%mm4 \n\t" yading@10: "paddusw %%mm3, %%mm5 \n\t" yading@10: "paddusw %%mm6, %%mm0 \n\t" yading@10: "paddusw %%mm6, %%mm1 \n\t" yading@10: "paddusw %%mm4, %%mm0 \n\t" yading@10: "paddusw %%mm5, %%mm1 \n\t" yading@10: "psrlw $2, %%mm0 \n\t" yading@10: "psrlw $2, %%mm1 \n\t" yading@10: "movq (%2, %%"REG_a"), %%mm3 \n\t" yading@10: "packuswb %%mm1, %%mm0 \n\t" yading@10: "pcmpeqd %%mm2, %%mm2 \n\t" yading@10: "paddb %%mm2, %%mm2 \n\t" yading@10: OP_AVG(%%mm3, %%mm0, %%mm1, %%mm2) yading@10: "movq %%mm1, (%2, %%"REG_a") \n\t" yading@10: "add %3, %%"REG_a" \n\t" yading@10: yading@10: "subl $2, %0 \n\t" yading@10: "jnz 1b \n\t" yading@10: :"+g"(h), "+S"(pixels) yading@10: :"D"(block), "r"((x86_reg)line_size) yading@10: :REG_a, "memory"); yading@10: } yading@10: yading@10: //FIXME optimize yading@10: static void DEF(put, pixels16_xy2)(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h){ yading@10: DEF(put, pixels8_xy2)(block , pixels , line_size, h); yading@10: DEF(put, pixels8_xy2)(block+8, pixels+8, line_size, h); yading@10: } yading@10: yading@10: static void DEF(avg, pixels16_xy2)(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h){ yading@10: DEF(avg, pixels8_xy2)(block , pixels , line_size, h); yading@10: DEF(avg, pixels8_xy2)(block+8, pixels+8, line_size, h); yading@10: }