libavcodec.hg: i386/dsputil_mmx

annotate i386/dsputil_mmx_rnd.h @ 967:274b518c4ecb libavcodec

PIC / ebx fix

author	michaelni
date	Tue, 07 Jan 2003 13:00:38 +0000
parents	7ef9226f430e
children	e162c09efbe7

rev	line source
445 62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	1 /*
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	2 * DSP utils mmx functions are compiled twice for rnd/no_rnd
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	3 * Copyright (c) 2000, 2001 Fabrice Bellard.
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	4 *
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	5 * This library is free software; you can redistribute it and/or
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	6 * modify it under the terms of the GNU Lesser General Public
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	7 * License as published by the Free Software Foundation; either
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	8 * version 2 of the License, or (at your option) any later version.
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	9 *
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	10 * This library is distributed in the hope that it will be useful,
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	13 * Lesser General Public License for more details.
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	14 *
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	15 * You should have received a copy of the GNU Lesser General Public
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	16 * License along with this library; if not, write to the Free Software
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	18 *
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	19 * MMX optimization by Nick Kurshev <nickols_k@mail.ru>
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	20 * mostly rewritten by Michael Niedermayer <michaelni@gmx.at>
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	21 * and improved by Zdenek Kabelac <kabi@users.sf.net>
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	22 */
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	23
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	24 // put_pixels
651 45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	25 static void DEF(put, pixels8_x2)(UINT8 block, const UINT8 pixels, int line_size, int h)
445 62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	26 {
448 e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	27 MOVQ_BFE(mm6);
445 62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	28 __asm __volatile(
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	29 "lea (%3, %3), %%eax \n\t"
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	30 ".balign 8 \n\t"
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	31 "1: \n\t"
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	32 "movq (%1), %%mm0 \n\t"
447 810f726ee3cc * using unrolled loops seems to help to MMX Celerons kabi parents: 445 diff changeset	33 "movq 1(%1), %%mm1 \n\t"
445 62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	34 "movq (%1, %3), %%mm2 \n\t"
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	35 "movq 1(%1, %3), %%mm3 \n\t"
448 e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	36 PAVGBP(%%mm0, %%mm1, %%mm4, %%mm2, %%mm3, %%mm5)
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	37 "movq %%mm4, (%2) \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	38 "movq %%mm5, (%2, %3) \n\t"
445 62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	39 "addl %%eax, %1 \n\t"
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	40 "addl %%eax, %2 \n\t"
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	41 "movq (%1), %%mm0 \n\t"
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	42 "movq 1(%1), %%mm1 \n\t"
447 810f726ee3cc * using unrolled loops seems to help to MMX Celerons kabi parents: 445 diff changeset	43 "movq (%1, %3), %%mm2 \n\t"
445 62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	44 "movq 1(%1, %3), %%mm3 \n\t"
448 e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	45 PAVGBP(%%mm0, %%mm1, %%mm4, %%mm2, %%mm3, %%mm5)
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	46 "movq %%mm4, (%2) \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	47 "movq %%mm5, (%2, %3) \n\t"
445 62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	48 "addl %%eax, %1 \n\t"
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	49 "addl %%eax, %2 \n\t"
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	50 "subl $4, %0 \n\t"
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	51 "jnz 1b \n\t"
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	52 :"+g"(h), "+S"(pixels), "+D"(block)
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	53 :"r"(line_size)
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	54 :"eax", "memory");
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	55 }
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	56
954 13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	57 static void DEF(put, pixels8_l2)(uint8_t dst, uint8_t src1, uint8_t *src2, int dstStride, int src1Stride, int h)
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	58 {
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	59 MOVQ_BFE(mm6);
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	60 __asm __volatile(
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	61 ".balign 8 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	62 "1: \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	63 "movq (%1), %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	64 "movq (%2), %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	65 "addl %4, %1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	66 "movq (%1), %%mm2 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	67 "movq 8(%2), %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	68 "addl %4, %1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	69 PAVGBP(%%mm0, %%mm1, %%mm4, %%mm2, %%mm3, %%mm5)
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	70 "movq %%mm4, (%3) \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	71 "addl %5, %3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	72 "movq %%mm5, (%3) \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	73 "addl %5, %3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	74 "movq (%1), %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	75 "movq 16(%2), %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	76 "addl %4, %1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	77 "movq (%1), %%mm2 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	78 "movq 24(%2), %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	79 "addl %4, %1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	80 "addl $32, %2 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	81 PAVGBP(%%mm0, %%mm1, %%mm4, %%mm2, %%mm3, %%mm5)
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	82 "movq %%mm4, (%3) \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	83 "addl %5, %3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	84 "movq %%mm5, (%3) \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	85 "addl %5, %3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	86 "subl $4, %0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	87 "jnz 1b \n\t"
967 274b518c4ecb PIC / ebx fix michaelni parents: 966 diff changeset	88 #ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cant be used
274b518c4ecb PIC / ebx fix michaelni parents: 966 diff changeset	89 :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst)
274b518c4ecb PIC / ebx fix michaelni parents: 966 diff changeset	90 #else
274b518c4ecb PIC / ebx fix michaelni parents: 966 diff changeset	91 :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst)
274b518c4ecb PIC / ebx fix michaelni parents: 966 diff changeset	92 #endif
966 7ef9226f430e more gcc bug workarounds michaelni parents: 965 diff changeset	93 :"S"(src1Stride), "D"(dstStride)
954 13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	94 :"memory");
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	95 }
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	96
651 45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	97 static void DEF(put, pixels16_x2)(UINT8 block, const UINT8 pixels, int line_size, int h)
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	98 {
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	99 MOVQ_BFE(mm6);
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	100 __asm __volatile(
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	101 "lea (%3, %3), %%eax \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	102 ".balign 8 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	103 "1: \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	104 "movq (%1), %%mm0 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	105 "movq 1(%1), %%mm1 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	106 "movq (%1, %3), %%mm2 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	107 "movq 1(%1, %3), %%mm3 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	108 PAVGBP(%%mm0, %%mm1, %%mm4, %%mm2, %%mm3, %%mm5)
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	109 "movq %%mm4, (%2) \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	110 "movq %%mm5, (%2, %3) \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	111 "movq 8(%1), %%mm0 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	112 "movq 9(%1), %%mm1 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	113 "movq 8(%1, %3), %%mm2 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	114 "movq 9(%1, %3), %%mm3 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	115 PAVGBP(%%mm0, %%mm1, %%mm4, %%mm2, %%mm3, %%mm5)
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	116 "movq %%mm4, 8(%2) \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	117 "movq %%mm5, 8(%2, %3) \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	118 "addl %%eax, %1 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	119 "addl %%eax, %2 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	120 "movq (%1), %%mm0 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	121 "movq 1(%1), %%mm1 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	122 "movq (%1, %3), %%mm2 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	123 "movq 1(%1, %3), %%mm3 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	124 PAVGBP(%%mm0, %%mm1, %%mm4, %%mm2, %%mm3, %%mm5)
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	125 "movq %%mm4, (%2) \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	126 "movq %%mm5, (%2, %3) \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	127 "movq 8(%1), %%mm0 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	128 "movq 9(%1), %%mm1 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	129 "movq 8(%1, %3), %%mm2 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	130 "movq 9(%1, %3), %%mm3 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	131 PAVGBP(%%mm0, %%mm1, %%mm4, %%mm2, %%mm3, %%mm5)
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	132 "movq %%mm4, 8(%2) \n\t"
954 13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	133 "movq %%mm5, 8(%2, %3) \n\t"
651 45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	134 "addl %%eax, %1 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	135 "addl %%eax, %2 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	136 "subl $4, %0 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	137 "jnz 1b \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	138 :"+g"(h), "+S"(pixels), "+D"(block)
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	139 :"r"(line_size)
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	140 :"eax", "memory");
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	141 }
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	142
954 13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	143 static void DEF(put, pixels16_l2)(uint8_t dst, uint8_t src1, uint8_t *src2, int dstStride, int src1Stride, int h)
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	144 {
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	145 MOVQ_BFE(mm6);
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	146 __asm __volatile(
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	147 ".balign 8 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	148 "1: \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	149 "movq (%1), %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	150 "movq (%2), %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	151 "movq 8(%1), %%mm2 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	152 "movq 8(%2), %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	153 "addl %4, %1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	154 PAVGBP(%%mm0, %%mm1, %%mm4, %%mm2, %%mm3, %%mm5)
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	155 "movq %%mm4, (%3) \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	156 "movq %%mm5, 8(%3) \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	157 "addl %5, %3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	158 "movq (%1), %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	159 "movq 16(%2), %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	160 "movq 8(%1), %%mm2 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	161 "movq 24(%2), %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	162 "addl %4, %1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	163 PAVGBP(%%mm0, %%mm1, %%mm4, %%mm2, %%mm3, %%mm5)
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	164 "movq %%mm4, (%3) \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	165 "movq %%mm5, 8(%3) \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	166 "addl %5, %3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	167 "addl $32, %2 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	168 "subl $2, %0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	169 "jnz 1b \n\t"
967 274b518c4ecb PIC / ebx fix michaelni parents: 966 diff changeset	170 #ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cant be used
274b518c4ecb PIC / ebx fix michaelni parents: 966 diff changeset	171 :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst)
274b518c4ecb PIC / ebx fix michaelni parents: 966 diff changeset	172 #else
274b518c4ecb PIC / ebx fix michaelni parents: 966 diff changeset	173 :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst)
274b518c4ecb PIC / ebx fix michaelni parents: 966 diff changeset	174 #endif
965 e149203f4a0d gcc 3.2.2 -O3 bug workaround (older gcc are very likely affected too but didnt check) michaelni parents: 959 diff changeset	175 :"S"(src1Stride), "D"(dstStride)
e149203f4a0d gcc 3.2.2 -O3 bug workaround (older gcc are very likely affected too but didnt check) michaelni parents: 959 diff changeset	176 :"memory");
954 13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	177 }
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	178
651 45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	179 static void DEF(put, pixels8_y2)(UINT8 block, const UINT8 pixels, int line_size, int h)
445 62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	180 {
448 e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	181 MOVQ_BFE(mm6);
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	182 __asm __volatile(
445 62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	183 "lea (%3, %3), %%eax \n\t"
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	184 "movq (%1), %%mm0 \n\t"
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	185 ".balign 8 \n\t"
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	186 "1: \n\t"
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	187 "movq (%1, %3), %%mm1 \n\t"
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	188 "movq (%1, %%eax),%%mm2 \n\t"
448 e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	189 PAVGBP(%%mm1, %%mm0, %%mm4, %%mm2, %%mm1, %%mm5)
470 b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	190 "movq %%mm4, (%2) \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	191 "movq %%mm5, (%2, %3) \n\t"
445 62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	192 "addl %%eax, %1 \n\t"
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	193 "addl %%eax, %2 \n\t"
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	194 "movq (%1, %3), %%mm1 \n\t"
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	195 "movq (%1, %%eax),%%mm0 \n\t"
448 e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	196 PAVGBP(%%mm1, %%mm2, %%mm4, %%mm0, %%mm1, %%mm5)
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	197 "movq %%mm4, (%2) \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	198 "movq %%mm5, (%2, %3) \n\t"
445 62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	199 "addl %%eax, %1 \n\t"
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	200 "addl %%eax, %2 \n\t"
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	201 "subl $4, %0 \n\t"
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	202 "jnz 1b \n\t"
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	203 :"+g"(h), "+S"(pixels), "+D"(block)
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	204 :"r"(line_size)
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	205 :"eax", "memory");
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	206 }
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	207
651 45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	208 static void DEF(put, pixels8_xy2)(UINT8 block, const UINT8 pixels, int line_size, int h)
445 62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	209 {
448 e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	210 MOVQ_ZERO(mm7);
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	211 SET_RND(mm6); // =2 for rnd and =1 for no_rnd version
445 62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	212 __asm __volatile(
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	213 "movq (%1), %%mm0 \n\t"
448 e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	214 "movq 1(%1), %%mm4 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	215 "movq %%mm0, %%mm1 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	216 "movq %%mm4, %%mm5 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	217 "punpcklbw %%mm7, %%mm0 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	218 "punpcklbw %%mm7, %%mm4 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	219 "punpckhbw %%mm7, %%mm1 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	220 "punpckhbw %%mm7, %%mm5 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	221 "paddusw %%mm0, %%mm4 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	222 "paddusw %%mm1, %%mm5 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	223 "xorl %%eax, %%eax \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	224 "addl %3, %1 \n\t"
470 b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	225 ".balign 8 \n\t"
445 62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	226 "1: \n\t"
448 e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	227 "movq (%1, %%eax), %%mm0 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	228 "movq 1(%1, %%eax), %%mm2 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	229 "movq %%mm0, %%mm1 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	230 "movq %%mm2, %%mm3 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	231 "punpcklbw %%mm7, %%mm0 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	232 "punpcklbw %%mm7, %%mm2 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	233 "punpckhbw %%mm7, %%mm1 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	234 "punpckhbw %%mm7, %%mm3 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	235 "paddusw %%mm2, %%mm0 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	236 "paddusw %%mm3, %%mm1 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	237 "paddusw %%mm6, %%mm4 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	238 "paddusw %%mm6, %%mm5 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	239 "paddusw %%mm0, %%mm4 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	240 "paddusw %%mm1, %%mm5 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	241 "psrlw $2, %%mm4 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	242 "psrlw $2, %%mm5 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	243 "packuswb %%mm5, %%mm4 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	244 "movq %%mm4, (%2, %%eax) \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	245 "addl %3, %%eax \n\t"
445 62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	246
448 e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	247 "movq (%1, %%eax), %%mm2 \n\t" // 0 <-> 2 1 <-> 3
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	248 "movq 1(%1, %%eax), %%mm4 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	249 "movq %%mm2, %%mm3 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	250 "movq %%mm4, %%mm5 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	251 "punpcklbw %%mm7, %%mm2 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	252 "punpcklbw %%mm7, %%mm4 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	253 "punpckhbw %%mm7, %%mm3 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	254 "punpckhbw %%mm7, %%mm5 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	255 "paddusw %%mm2, %%mm4 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	256 "paddusw %%mm3, %%mm5 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	257 "paddusw %%mm6, %%mm0 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	258 "paddusw %%mm6, %%mm1 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	259 "paddusw %%mm4, %%mm0 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	260 "paddusw %%mm5, %%mm1 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	261 "psrlw $2, %%mm0 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	262 "psrlw $2, %%mm1 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	263 "packuswb %%mm1, %%mm0 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	264 "movq %%mm0, (%2, %%eax) \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	265 "addl %3, %%eax \n\t"
445 62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	266
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	267 "subl $2, %0 \n\t"
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	268 "jnz 1b \n\t"
448 e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	269 :"+g"(h), "+S"(pixels)
e8c8ca9106aa * removed MANGLE from macros for setting constants kabi parents: 447 diff changeset	270 :"D"(block), "r"(line_size)
445 62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	271 :"eax", "memory");
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	272 }
62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	273
954 13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	274 static void DEF(put, pixels8_l4)(uint8_t dst, uint8_t src1, uint8_t *src2, int stride, int h)
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	275 {
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	276 MOVQ_ZERO(mm7);
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	277 SET_RND(mm6); // =2 for rnd and =1 for no_rnd version
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	278 __asm __volatile(
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	279 ".balign 8 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	280 "1: \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	281 "movq (%1), %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	282 "movq (%2), %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	283 "movq 64(%2), %%mm2 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	284 "movq 136(%2), %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	285 "punpcklbw %%mm7, %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	286 "punpcklbw %%mm7, %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	287 "punpcklbw %%mm7, %%mm2 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	288 "punpcklbw %%mm7, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	289 "paddusw %%mm6, %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	290 "paddusw %%mm0, %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	291 "paddusw %%mm2, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	292 "paddusw %%mm1, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	293 "psrlw $2, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	294 "movq (%1), %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	295 "movq (%2), %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	296 "movq 64(%2), %%mm2 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	297 "movq 136(%2), %%mm4 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	298 "punpckhbw %%mm7, %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	299 "punpckhbw %%mm7, %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	300 "punpckhbw %%mm7, %%mm2 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	301 "punpckhbw %%mm7, %%mm4 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	302 "paddusw %%mm6, %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	303 "paddusw %%mm0, %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	304 "paddusw %%mm2, %%mm4 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	305 "paddusw %%mm1, %%mm4 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	306 "psrlw $2, %%mm4 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	307 "packuswb %%mm4, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	308 "movq %%mm3, (%0) \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	309 "addl %4, %0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	310 "addl %4, %1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	311 "addl $8, %2 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	312 "decl %3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	313 "jnz 1b \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	314 :"+r"(dst), "+r"(src1), "+r"(src2), "+r"(h)
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	315 :"r"(stride)
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	316 :"memory");
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	317 }
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	318
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	319 static void DEF(put, pixels16_l4)(uint8_t dst, uint8_t src1, uint8_t *src2, int stride, int h)
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	320 {
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	321 MOVQ_ZERO(mm7);
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	322 SET_RND(mm6); // =2 for rnd and =1 for no_rnd version
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	323 __asm __volatile(
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	324 ".balign 8 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	325 "1: \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	326 "movq (%1), %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	327 "movq (%2), %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	328 "movq 256(%2), %%mm2 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	329 "movq 528(%2), %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	330 "punpcklbw %%mm7, %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	331 "punpcklbw %%mm7, %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	332 "punpcklbw %%mm7, %%mm2 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	333 "punpcklbw %%mm7, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	334 "paddusw %%mm6, %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	335 "paddusw %%mm0, %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	336 "paddusw %%mm2, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	337 "paddusw %%mm1, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	338 "psrlw $2, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	339 "movq (%1), %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	340 "movq (%2), %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	341 "movq 256(%2), %%mm2 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	342 "movq 528(%2), %%mm4 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	343 "punpckhbw %%mm7, %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	344 "punpckhbw %%mm7, %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	345 "punpckhbw %%mm7, %%mm2 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	346 "punpckhbw %%mm7, %%mm4 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	347 "paddusw %%mm6, %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	348 "paddusw %%mm0, %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	349 "paddusw %%mm2, %%mm4 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	350 "paddusw %%mm1, %%mm4 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	351 "psrlw $2, %%mm4 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	352 "packuswb %%mm4, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	353 "movq %%mm3, (%0) \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	354 "movq 8(%1), %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	355 "movq 8(%2), %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	356 "movq 264(%2), %%mm2 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	357 "movq 536(%2), %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	358 "punpcklbw %%mm7, %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	359 "punpcklbw %%mm7, %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	360 "punpcklbw %%mm7, %%mm2 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	361 "punpcklbw %%mm7, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	362 "paddusw %%mm6, %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	363 "paddusw %%mm0, %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	364 "paddusw %%mm2, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	365 "paddusw %%mm1, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	366 "psrlw $2, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	367 "movq 8(%1), %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	368 "movq 8(%2), %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	369 "movq 264(%2), %%mm2 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	370 "movq 536(%2), %%mm4 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	371 "punpckhbw %%mm7, %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	372 "punpckhbw %%mm7, %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	373 "punpckhbw %%mm7, %%mm2 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	374 "punpckhbw %%mm7, %%mm4 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	375 "paddusw %%mm6, %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	376 "paddusw %%mm0, %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	377 "paddusw %%mm2, %%mm4 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	378 "paddusw %%mm1, %%mm4 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	379 "psrlw $2, %%mm4 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	380 "packuswb %%mm4, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	381 "movq %%mm3, 8(%0) \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	382 "addl %4, %0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	383 "addl %4, %1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	384 "addl $16, %2 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	385 "decl %3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	386 "jnz 1b \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	387 :"+r"(dst), "+r"(src1), "+r"(src2), "+r"(h)
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	388 :"r"(stride)
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	389 :"memory");
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	390 }
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	391
445 62c01dbdc1e0 * code with new PAVGB for MMX only CPU splited into separate file kabi parents: diff changeset	392 // avg_pixels
470 b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	393 // in case more speed is needed - unroling would certainly help
651 45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	394 static void DEF(avg, pixels8)(UINT8 block, const UINT8 pixels, int line_size, int h)
470 b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	395 {
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	396 MOVQ_BFE(mm6);
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	397 JUMPALIGN();
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	398 do {
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	399 __asm __volatile(
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	400 "movq %0, %%mm0 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	401 "movq %1, %%mm1 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	402 PAVGB(%%mm0, %%mm1, %%mm2, %%mm6)
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	403 "movq %%mm2, %0 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	404 :"+m"(*block)
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	405 :"m"(*pixels)
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	406 :"memory");
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	407 pixels += line_size;
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	408 block += line_size;
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	409 }
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	410 while (--h);
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	411 }
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	412
651 45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	413 static void DEF(avg, pixels16)(UINT8 block, const UINT8 pixels, int line_size, int h)
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	414 {
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	415 MOVQ_BFE(mm6);
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	416 JUMPALIGN();
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	417 do {
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	418 __asm __volatile(
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	419 "movq %0, %%mm0 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	420 "movq %1, %%mm1 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	421 PAVGB(%%mm0, %%mm1, %%mm2, %%mm6)
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	422 "movq %%mm2, %0 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	423 "movq 8%0, %%mm0 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	424 "movq 8%1, %%mm1 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	425 PAVGB(%%mm0, %%mm1, %%mm2, %%mm6)
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	426 "movq %%mm2, 8%0 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	427 :"+m"(*block)
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	428 :"m"(*pixels)
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	429 :"memory");
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	430 pixels += line_size;
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	431 block += line_size;
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	432 }
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	433 while (--h);
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	434 }
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	435
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	436 static void DEF(avg, pixels8_x2)(UINT8 block, const UINT8 pixels, int line_size, int h)
470 b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	437 {
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	438 MOVQ_BFE(mm6);
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	439 JUMPALIGN();
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	440 do {
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	441 __asm __volatile(
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	442 "movq %1, %%mm0 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	443 "movq 1%1, %%mm1 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	444 "movq %0, %%mm3 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	445 PAVGB(%%mm0, %%mm1, %%mm2, %%mm6)
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	446 PAVGB(%%mm3, %%mm2, %%mm0, %%mm6)
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	447 "movq %%mm0, %0 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	448 :"+m"(*block)
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	449 :"m"(*pixels)
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	450 :"memory");
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	451 pixels += line_size;
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	452 block += line_size;
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	453 } while (--h);
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	454 }
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	455
954 13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	456 static void DEF(avg, pixels8_l2)(uint8_t dst, uint8_t src1, uint8_t *src2, int dstStride, int src1Stride, int h)
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	457 {
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	458 MOVQ_BFE(mm6);
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	459 JUMPALIGN();
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	460 do {
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	461 __asm __volatile(
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	462 "movq %1, %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	463 "movq %2, %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	464 "movq %0, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	465 PAVGB(%%mm0, %%mm1, %%mm2, %%mm6)
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	466 PAVGB(%%mm3, %%mm2, %%mm0, %%mm6)
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	467 "movq %%mm0, %0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	468 :"+m"(*dst)
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	469 :"m"(src1), "m"(src2)
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	470 :"memory");
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	471 dst += dstStride;
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	472 src1 += src1Stride;
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	473 src2 += 8;
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	474 } while (--h);
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	475 }
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	476
651 45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	477 static void DEF(avg, pixels16_x2)(UINT8 block, const UINT8 pixels, int line_size, int h)
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	478 {
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	479 MOVQ_BFE(mm6);
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	480 JUMPALIGN();
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	481 do {
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	482 __asm __volatile(
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	483 "movq %1, %%mm0 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	484 "movq 1%1, %%mm1 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	485 "movq %0, %%mm3 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	486 PAVGB(%%mm0, %%mm1, %%mm2, %%mm6)
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	487 PAVGB(%%mm3, %%mm2, %%mm0, %%mm6)
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	488 "movq %%mm0, %0 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	489 "movq 8%1, %%mm0 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	490 "movq 9%1, %%mm1 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	491 "movq 8%0, %%mm3 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	492 PAVGB(%%mm0, %%mm1, %%mm2, %%mm6)
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	493 PAVGB(%%mm3, %%mm2, %%mm0, %%mm6)
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	494 "movq %%mm0, 8%0 \n\t"
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	495 :"+m"(*block)
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	496 :"m"(*pixels)
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	497 :"memory");
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	498 pixels += line_size;
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	499 block += line_size;
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	500 } while (--h);
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	501 }
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	502
954 13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	503 static void DEF(avg, pixels16_l2)(uint8_t dst, uint8_t src1, uint8_t *src2, int dstStride, int src1Stride, int h)
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	504 {
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	505 MOVQ_BFE(mm6);
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	506 JUMPALIGN();
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	507 do {
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	508 __asm __volatile(
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	509 "movq %1, %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	510 "movq %2, %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	511 "movq %0, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	512 PAVGB(%%mm0, %%mm1, %%mm2, %%mm6)
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	513 PAVGB(%%mm3, %%mm2, %%mm0, %%mm6)
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	514 "movq %%mm0, %0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	515 "movq 8%1, %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	516 "movq 8%2, %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	517 "movq 8%0, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	518 PAVGB(%%mm0, %%mm1, %%mm2, %%mm6)
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	519 PAVGB(%%mm3, %%mm2, %%mm0, %%mm6)
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	520 "movq %%mm0, 8%0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	521 :"+m"(*dst)
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	522 :"m"(src1), "m"(src2)
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	523 :"memory");
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	524 dst += dstStride;
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	525 src1 += src1Stride;
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	526 src2 += 16;
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	527 } while (--h);
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	528 }
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	529
651 45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	530 static void DEF(avg, pixels8_y2)(UINT8 block, const UINT8 pixels, int line_size, int h)
470 b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	531 {
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	532 MOVQ_BFE(mm6);
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	533 __asm __volatile(
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	534 "lea (%3, %3), %%eax \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	535 "movq (%1), %%mm0 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	536 ".balign 8 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	537 "1: \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	538 "movq (%1, %3), %%mm1 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	539 "movq (%1, %%eax), %%mm2 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	540 PAVGBP(%%mm1, %%mm0, %%mm4, %%mm2, %%mm1, %%mm5)
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	541 "movq (%2), %%mm3 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	542 PAVGB(%%mm3, %%mm4, %%mm0, %%mm6)
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	543 "movq (%2, %3), %%mm3 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	544 PAVGB(%%mm3, %%mm5, %%mm1, %%mm6)
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	545 "movq %%mm0, (%2) \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	546 "movq %%mm1, (%2, %3) \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	547 "addl %%eax, %1 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	548 "addl %%eax, %2 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	549
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	550 "movq (%1, %3), %%mm1 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	551 "movq (%1, %%eax), %%mm0 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	552 PAVGBP(%%mm1, %%mm2, %%mm4, %%mm0, %%mm1, %%mm5)
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	553 "movq (%2), %%mm3 \n\t"
606 72d9661683f2 dont trash mm0, its still used in the next iteration of the loop (fixes strange horizontal lines on some blocks in b frames on mmx only cpus) michaelni parents: 470 diff changeset	554 PAVGB(%%mm3, %%mm4, %%mm2, %%mm6)
470 b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	555 "movq (%2, %3), %%mm3 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	556 PAVGB(%%mm3, %%mm5, %%mm1, %%mm6)
606 72d9661683f2 dont trash mm0, its still used in the next iteration of the loop (fixes strange horizontal lines on some blocks in b frames on mmx only cpus) michaelni parents: 470 diff changeset	557 "movq %%mm2, (%2) \n\t"
470 b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	558 "movq %%mm1, (%2, %3) \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	559 "addl %%eax, %1 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	560 "addl %%eax, %2 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	561
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	562 "subl $4, %0 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	563 "jnz 1b \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	564 :"+g"(h), "+S"(pixels), "+D"(block)
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	565 :"r"(line_size)
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	566 :"eax", "memory");
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	567 }
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	568
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	569 // this routine is 'slightly' suboptimal but mostly unused
651 45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	570 static void DEF(avg, pixels8_xy2)(UINT8 block, const UINT8 pixels, int line_size, int h)
470 b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	571 {
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	572 MOVQ_ZERO(mm7);
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	573 SET_RND(mm6); // =2 for rnd and =1 for no_rnd version
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	574 __asm __volatile(
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	575 "movq (%1), %%mm0 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	576 "movq 1(%1), %%mm4 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	577 "movq %%mm0, %%mm1 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	578 "movq %%mm4, %%mm5 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	579 "punpcklbw %%mm7, %%mm0 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	580 "punpcklbw %%mm7, %%mm4 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	581 "punpckhbw %%mm7, %%mm1 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	582 "punpckhbw %%mm7, %%mm5 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	583 "paddusw %%mm0, %%mm4 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	584 "paddusw %%mm1, %%mm5 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	585 "xorl %%eax, %%eax \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	586 "addl %3, %1 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	587 ".balign 8 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	588 "1: \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	589 "movq (%1, %%eax), %%mm0 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	590 "movq 1(%1, %%eax), %%mm2 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	591 "movq %%mm0, %%mm1 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	592 "movq %%mm2, %%mm3 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	593 "punpcklbw %%mm7, %%mm0 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	594 "punpcklbw %%mm7, %%mm2 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	595 "punpckhbw %%mm7, %%mm1 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	596 "punpckhbw %%mm7, %%mm3 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	597 "paddusw %%mm2, %%mm0 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	598 "paddusw %%mm3, %%mm1 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	599 "paddusw %%mm6, %%mm4 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	600 "paddusw %%mm6, %%mm5 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	601 "paddusw %%mm0, %%mm4 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	602 "paddusw %%mm1, %%mm5 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	603 "psrlw $2, %%mm4 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	604 "psrlw $2, %%mm5 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	605 "movq (%2, %%eax), %%mm3 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	606 "packuswb %%mm5, %%mm4 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	607 "pcmpeqd %%mm2, %%mm2 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	608 "paddb %%mm2, %%mm2 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	609 PAVGB(%%mm3, %%mm4, %%mm5, %%mm2)
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	610 "movq %%mm5, (%2, %%eax) \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	611 "addl %3, %%eax \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	612
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	613 "movq (%1, %%eax), %%mm2 \n\t" // 0 <-> 2 1 <-> 3
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	614 "movq 1(%1, %%eax), %%mm4 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	615 "movq %%mm2, %%mm3 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	616 "movq %%mm4, %%mm5 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	617 "punpcklbw %%mm7, %%mm2 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	618 "punpcklbw %%mm7, %%mm4 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	619 "punpckhbw %%mm7, %%mm3 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	620 "punpckhbw %%mm7, %%mm5 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	621 "paddusw %%mm2, %%mm4 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	622 "paddusw %%mm3, %%mm5 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	623 "paddusw %%mm6, %%mm0 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	624 "paddusw %%mm6, %%mm1 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	625 "paddusw %%mm4, %%mm0 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	626 "paddusw %%mm5, %%mm1 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	627 "psrlw $2, %%mm0 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	628 "psrlw $2, %%mm1 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	629 "movq (%2, %%eax), %%mm3 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	630 "packuswb %%mm1, %%mm0 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	631 "pcmpeqd %%mm2, %%mm2 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	632 "paddb %%mm2, %%mm2 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	633 PAVGB(%%mm3, %%mm0, %%mm1, %%mm2)
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	634 "movq %%mm1, (%2, %%eax) \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	635 "addl %3, %%eax \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	636
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	637 "subl $2, %0 \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	638 "jnz 1b \n\t"
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	639 :"+g"(h), "+S"(pixels)
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	640 :"D"(block), "r"(line_size)
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	641 :"eax", "memory");
b94e82d31b06 * implemented remaing avg_ pixel functions (these are not used offen) kabi parents: 448 diff changeset	642 }
651 45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	643
954 13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	644 static void DEF(avg, pixels8_l4)(uint8_t dst, uint8_t src1, uint8_t *src2, int stride, int h)
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	645 {
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	646 MOVQ_ZERO(mm7);
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	647 SET_RND(mm6); // =2 for rnd and =1 for no_rnd version
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	648 MOVQ_BFE(mm5);
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	649 __asm __volatile(
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	650 ".balign 8 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	651 "1: \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	652 "movq (%1), %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	653 "movq (%2), %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	654 "movq 64(%2), %%mm2 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	655 "movq 136(%2), %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	656 "punpcklbw %%mm7, %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	657 "punpcklbw %%mm7, %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	658 "punpcklbw %%mm7, %%mm2 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	659 "punpcklbw %%mm7, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	660 "paddusw %%mm6, %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	661 "paddusw %%mm0, %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	662 "paddusw %%mm2, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	663 "paddusw %%mm1, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	664 "psrlw $2, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	665 "movq (%1), %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	666 "movq (%2), %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	667 "movq 64(%2), %%mm2 \n\t"
959 3ec070eef24a qpel in b frames bugfixes michaelni parents: 954 diff changeset	668 "movq 136(%2), %%mm4 \n\t"
954 13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	669 "punpckhbw %%mm7, %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	670 "punpckhbw %%mm7, %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	671 "punpckhbw %%mm7, %%mm2 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	672 "punpckhbw %%mm7, %%mm4 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	673 "paddusw %%mm6, %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	674 "paddusw %%mm0, %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	675 "paddusw %%mm2, %%mm4 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	676 "paddusw %%mm1, %%mm4 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	677 "psrlw $2, %%mm4 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	678 "packuswb %%mm4, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	679 "movq (%0), %%mm4 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	680 PAVGB(%%mm3, %%mm4, %%mm0, %%mm5)
959 3ec070eef24a qpel in b frames bugfixes michaelni parents: 954 diff changeset	681 "movq %%mm0, (%0) \n\t"
954 13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	682 "addl %4, %0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	683 "addl %4, %1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	684 "addl $8, %2 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	685 "decl %3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	686 "jnz 1b \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	687 :"+r"(dst), "+r"(src1), "+r"(src2), "+r"(h)
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	688 :"r"(stride)
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	689 :"memory");
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	690 }
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	691
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	692 static void DEF(avg, pixels16_l4)(uint8_t dst, uint8_t src1, uint8_t *src2, int stride, int h)
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	693 {
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	694 MOVQ_ZERO(mm7);
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	695 SET_RND(mm6); // =2 for rnd and =1 for no_rnd version
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	696 MOVQ_BFE(mm5);
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	697 __asm __volatile(
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	698 ".balign 8 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	699 "1: \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	700 "movq (%1), %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	701 "movq (%2), %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	702 "movq 256(%2), %%mm2 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	703 "movq 528(%2), %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	704 "punpcklbw %%mm7, %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	705 "punpcklbw %%mm7, %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	706 "punpcklbw %%mm7, %%mm2 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	707 "punpcklbw %%mm7, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	708 "paddusw %%mm6, %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	709 "paddusw %%mm0, %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	710 "paddusw %%mm2, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	711 "paddusw %%mm1, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	712 "psrlw $2, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	713 "movq (%1), %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	714 "movq (%2), %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	715 "movq 256(%2), %%mm2 \n\t"
959 3ec070eef24a qpel in b frames bugfixes michaelni parents: 954 diff changeset	716 "movq 528(%2), %%mm4 \n\t"
954 13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	717 "punpckhbw %%mm7, %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	718 "punpckhbw %%mm7, %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	719 "punpckhbw %%mm7, %%mm2 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	720 "punpckhbw %%mm7, %%mm4 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	721 "paddusw %%mm6, %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	722 "paddusw %%mm0, %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	723 "paddusw %%mm2, %%mm4 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	724 "paddusw %%mm1, %%mm4 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	725 "psrlw $2, %%mm4 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	726 "packuswb %%mm4, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	727 "movq (%0), %%mm4 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	728 PAVGB(%%mm3, %%mm4, %%mm0, %%mm5)
959 3ec070eef24a qpel in b frames bugfixes michaelni parents: 954 diff changeset	729 "movq %%mm0, (%0) \n\t"
954 13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	730 "movq 8(%1), %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	731 "movq 8(%2), %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	732 "movq 264(%2), %%mm2 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	733 "movq 536(%2), %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	734 "punpcklbw %%mm7, %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	735 "punpcklbw %%mm7, %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	736 "punpcklbw %%mm7, %%mm2 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	737 "punpcklbw %%mm7, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	738 "paddusw %%mm6, %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	739 "paddusw %%mm0, %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	740 "paddusw %%mm2, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	741 "paddusw %%mm1, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	742 "psrlw $2, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	743 "movq 8(%1), %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	744 "movq 8(%2), %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	745 "movq 264(%2), %%mm2 \n\t"
959 3ec070eef24a qpel in b frames bugfixes michaelni parents: 954 diff changeset	746 "movq 536(%2), %%mm4 \n\t"
954 13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	747 "punpckhbw %%mm7, %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	748 "punpckhbw %%mm7, %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	749 "punpckhbw %%mm7, %%mm2 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	750 "punpckhbw %%mm7, %%mm4 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	751 "paddusw %%mm6, %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	752 "paddusw %%mm0, %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	753 "paddusw %%mm2, %%mm4 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	754 "paddusw %%mm1, %%mm4 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	755 "psrlw $2, %%mm4 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	756 "packuswb %%mm4, %%mm3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	757 "movq 8(%0), %%mm4 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	758 PAVGB(%%mm3, %%mm4, %%mm0, %%mm5)
959 3ec070eef24a qpel in b frames bugfixes michaelni parents: 954 diff changeset	759 "movq %%mm0, 8(%0) \n\t"
954 13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	760 "addl %4, %0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	761 "addl %4, %1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	762 "addl $16, %2 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	763 "decl %3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	764 "jnz 1b \n\t"
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	765 :"+r"(dst), "+r"(src1), "+r"(src2), "+r"(h)
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	766 :"r"(stride)
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	767 :"memory");
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	768 }
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	769
13aec7e50c52 qpel in mmx2/3dnow michaelni parents: 651 diff changeset	770
651 45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	771 //FIXME optimize
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	772 static void DEF(put, pixels16_y2)(UINT8 block, const UINT8 pixels, int line_size, int h){
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	773 DEF(put, pixels8_y2)(block , pixels , line_size, h);
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	774 DEF(put, pixels8_y2)(block+8, pixels+8, line_size, h);
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	775 }
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	776
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	777 static void DEF(put, pixels16_xy2)(UINT8 block, const UINT8 pixels, int line_size, int h){
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	778 DEF(put, pixels8_xy2)(block , pixels , line_size, h);
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	779 DEF(put, pixels8_xy2)(block+8, pixels+8, line_size, h);
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	780 }
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	781
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	782 static void DEF(avg, pixels16_y2)(UINT8 block, const UINT8 pixels, int line_size, int h){
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	783 DEF(avg, pixels8_y2)(block , pixels , line_size, h);
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	784 DEF(avg, pixels8_y2)(block+8, pixels+8, line_size, h);
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	785 }
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	786
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	787 static void DEF(avg, pixels16_xy2)(UINT8 block, const UINT8 pixels, int line_size, int h){
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	788 DEF(avg, pixels8_xy2)(block , pixels , line_size, h);
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	789 DEF(avg, pixels8_xy2)(block+8, pixels+8, line_size, h);
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	790 }
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	791
45e8f39fda50 put/avg_pixels16 michaelni parents: 606 diff changeset	792

Mercurial > libavcodec.hg

annotate i386/dsputil_mmx_rnd.h @ 967:274b518c4ecb libavcodec