annotate i386/dsputil_mmx_avg.h @ 1625:c910ff78ef80 libavcodec

av_d2q() documentation
author michael
date Sun, 16 Nov 2003 19:23:39 +0000
parents b32afefe7d33
children 07a484280a82
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1 /*
986e461dc072 Initial revision
glantau
parents:
diff changeset
2 * DSP utils : average functions are compiled twice for 3dnow/mmx2
429
718a22dc121f license/copyright change
glantau
parents: 416
diff changeset
3 * Copyright (c) 2000, 2001 Fabrice Bellard.
386
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
4 * Copyright (c) 2002 Michael Niedermayer
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
5 *
429
718a22dc121f license/copyright change
glantau
parents: 416
diff changeset
6 * This library is free software; you can redistribute it and/or
718a22dc121f license/copyright change
glantau
parents: 416
diff changeset
7 * modify it under the terms of the GNU Lesser General Public
718a22dc121f license/copyright change
glantau
parents: 416
diff changeset
8 * License as published by the Free Software Foundation; either
718a22dc121f license/copyright change
glantau
parents: 416
diff changeset
9 * version 2 of the License, or (at your option) any later version.
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
10 *
429
718a22dc121f license/copyright change
glantau
parents: 416
diff changeset
11 * This library is distributed in the hope that it will be useful,
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
429
718a22dc121f license/copyright change
glantau
parents: 416
diff changeset
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
718a22dc121f license/copyright change
glantau
parents: 416
diff changeset
14 * Lesser General Public License for more details.
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
15 *
429
718a22dc121f license/copyright change
glantau
parents: 416
diff changeset
16 * You should have received a copy of the GNU Lesser General Public
718a22dc121f license/copyright change
glantau
parents: 416
diff changeset
17 * License along with this library; if not, write to the Free Software
718a22dc121f license/copyright change
glantau
parents: 416
diff changeset
18 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
19 *
986e461dc072 Initial revision
glantau
parents:
diff changeset
20 * MMX optimization by Nick Kurshev <nickols_k@mail.ru>
386
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
21 * mostly rewritten by Michael Niedermayer <michaelni@gmx.at>
441
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
22 * and improved by Zdenek Kabelac <kabi@users.sf.net>
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
23 */
387
b8f3affeb8e1 shared lib support (req by kabi) ...
michaelni
parents: 386
diff changeset
24
389
f874493a1970 tried to avoid gcc 2.95.2 bug by puting explicit register constraints - added comment about rounding bug in some functions (need to correct or suppress them for regression tests)
glantau
parents: 387
diff changeset
25 /* XXX: we use explicit registers to avoid a gcc 2.95.2 register asm
413
1548abb7bbed * fix for -fPIC compilation - compiles with 2.95.2 as well - any
kabi
parents: 402
diff changeset
26 clobber bug - now it will work with 2.95.2 and also with -fPIC
1548abb7bbed * fix for -fPIC compilation - compiles with 2.95.2 as well - any
kabi
parents: 402
diff changeset
27 */
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 954
diff changeset
28 static void DEF(put_pixels8_x2)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
29 {
386
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
30 __asm __volatile(
441
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
31 "lea (%3, %3), %%eax \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
32 "1: \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
33 "movq (%1), %%mm0 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
34 "movq (%1, %3), %%mm1 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
35 PAVGB" 1(%1), %%mm0 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
36 PAVGB" 1(%1, %3), %%mm1 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
37 "movq %%mm0, (%2) \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
38 "movq %%mm1, (%2, %3) \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
39 "addl %%eax, %1 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
40 "addl %%eax, %2 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
41 "movq (%1), %%mm0 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
42 "movq (%1, %3), %%mm1 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
43 PAVGB" 1(%1), %%mm0 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
44 PAVGB" 1(%1, %3), %%mm1 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
45 "addl %%eax, %1 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
46 "movq %%mm0, (%2) \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
47 "movq %%mm1, (%2, %3) \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
48 "addl %%eax, %2 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
49 "subl $4, %0 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
50 "jnz 1b \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
51 :"+g"(h), "+S"(pixels), "+D"(block)
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
52 :"r" (line_size)
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
53 :"%eax", "memory");
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
54 }
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
55
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
56 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
57 {
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
58 __asm __volatile(
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
59 "1: \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
60 "movq (%1), %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
61 "addl %4, %1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
62 "movq (%1), %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
63 "addl %4, %1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
64 PAVGB" (%2), %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
65 PAVGB" 8(%2), %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
66 "movq %%mm0, (%3) \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
67 "addl %5, %3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
68 "movq %%mm1, (%3) \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
69 "addl %5, %3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
70 "movq (%1), %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
71 "addl %4, %1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
72 "movq (%1), %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
73 "addl %4, %1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
74 PAVGB" 16(%2), %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
75 PAVGB" 24(%2), %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
76 "movq %%mm0, (%3) \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
77 "addl %5, %3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
78 "movq %%mm1, (%3) \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
79 "addl %5, %3 \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 "subl $4, %0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
82 "jnz 1b \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
83 :"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst)
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
84 :"r"(src1Stride), "r"(dstStride)
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
85 :"memory");
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
86 }
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
87
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 954
diff changeset
88 static void DEF(put_pixels16_x2)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
89 {
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
90 __asm __volatile(
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
91 "lea (%3, %3), %%eax \n\t"
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
92 "1: \n\t"
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
93 "movq (%1), %%mm0 \n\t"
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
94 "movq (%1, %3), %%mm1 \n\t"
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
95 "movq 8(%1), %%mm2 \n\t"
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
96 "movq 8(%1, %3), %%mm3 \n\t"
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
97 PAVGB" 1(%1), %%mm0 \n\t"
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
98 PAVGB" 1(%1, %3), %%mm1 \n\t"
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
99 PAVGB" 9(%1), %%mm2 \n\t"
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
100 PAVGB" 9(%1, %3), %%mm3 \n\t"
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
101 "movq %%mm0, (%2) \n\t"
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
102 "movq %%mm1, (%2, %3) \n\t"
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
103 "movq %%mm2, 8(%2) \n\t"
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
104 "movq %%mm3, 8(%2, %3) \n\t"
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
105 "addl %%eax, %1 \n\t"
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
106 "addl %%eax, %2 \n\t"
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
107 "movq (%1), %%mm0 \n\t"
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
108 "movq (%1, %3), %%mm1 \n\t"
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
109 "movq 8(%1), %%mm2 \n\t"
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
110 "movq 8(%1, %3), %%mm3 \n\t"
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
111 PAVGB" 1(%1), %%mm0 \n\t"
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
112 PAVGB" 1(%1, %3), %%mm1 \n\t"
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
113 PAVGB" 9(%1), %%mm2 \n\t"
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
114 PAVGB" 9(%1, %3), %%mm3 \n\t"
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
115 "addl %%eax, %1 \n\t"
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
116 "movq %%mm0, (%2) \n\t"
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
117 "movq %%mm1, (%2, %3) \n\t"
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
118 "movq %%mm2, 8(%2) \n\t"
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
119 "movq %%mm3, 8(%2, %3) \n\t"
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
120 "addl %%eax, %2 \n\t"
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
121 "subl $4, %0 \n\t"
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
122 "jnz 1b \n\t"
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
123 :"+g"(h), "+S"(pixels), "+D"(block)
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
124 :"r" (line_size)
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
125 :"%eax", "memory");
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
126 }
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
127
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
128 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
129 {
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
130 __asm __volatile(
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
131 "1: \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
132 "movq (%1), %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
133 "movq 8(%1), %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
134 "addl %4, %1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
135 PAVGB" (%2), %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
136 PAVGB" 8(%2), %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
137 "movq %%mm0, (%3) \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
138 "movq %%mm1, 8(%3) \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
139 "addl %5, %3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
140 "movq (%1), %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
141 "movq 8(%1), %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
142 "addl %4, %1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
143 PAVGB" 16(%2), %%mm0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
144 PAVGB" 24(%2), %%mm1 \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
145 "movq %%mm0, (%3) \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
146 "movq %%mm1, 8(%3) \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
147 "addl %5, %3 \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
148 "addl $32, %2 \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
149 "subl $2, %0 \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
150 "jnz 1b \n\t"
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
151 :"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst)
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
152 :"r"(src1Stride), "r"(dstStride)
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
153 :"memory");
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 651
diff changeset
154 }
441
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
155
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
156 /* GL: this function does incorrect rounding if overflow */
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 954
diff changeset
157 static void DEF(put_no_rnd_pixels8_x2)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
441
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
158 {
448
e8c8ca9106aa * removed MANGLE from macros for setting constants
kabi
parents: 443
diff changeset
159 MOVQ_BONE(mm6);
441
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
160 __asm __volatile(
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
161 "lea (%3, %3), %%eax \n\t"
439
6ae275655a23 * more PIC friendly and faster code
kabi
parents: 429
diff changeset
162 "1: \n\t"
6ae275655a23 * more PIC friendly and faster code
kabi
parents: 429
diff changeset
163 "movq (%1), %%mm0 \n\t"
6ae275655a23 * more PIC friendly and faster code
kabi
parents: 429
diff changeset
164 "movq (%1, %3), %%mm2 \n\t"
6ae275655a23 * more PIC friendly and faster code
kabi
parents: 429
diff changeset
165 "movq 1(%1), %%mm1 \n\t"
6ae275655a23 * more PIC friendly and faster code
kabi
parents: 429
diff changeset
166 "movq 1(%1, %3), %%mm3 \n\t"
441
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
167 "addl %%eax, %1 \n\t"
448
e8c8ca9106aa * removed MANGLE from macros for setting constants
kabi
parents: 443
diff changeset
168 "psubusb %%mm6, %%mm0 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants
kabi
parents: 443
diff changeset
169 "psubusb %%mm6, %%mm2 \n\t"
386
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
170 PAVGB" %%mm1, %%mm0 \n\t"
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
171 PAVGB" %%mm3, %%mm2 \n\t"
439
6ae275655a23 * more PIC friendly and faster code
kabi
parents: 429
diff changeset
172 "movq %%mm0, (%2) \n\t"
6ae275655a23 * more PIC friendly and faster code
kabi
parents: 429
diff changeset
173 "movq %%mm2, (%2, %3) \n\t"
6ae275655a23 * more PIC friendly and faster code
kabi
parents: 429
diff changeset
174 "movq (%1), %%mm0 \n\t"
6ae275655a23 * more PIC friendly and faster code
kabi
parents: 429
diff changeset
175 "movq 1(%1), %%mm1 \n\t"
6ae275655a23 * more PIC friendly and faster code
kabi
parents: 429
diff changeset
176 "movq (%1, %3), %%mm2 \n\t"
6ae275655a23 * more PIC friendly and faster code
kabi
parents: 429
diff changeset
177 "movq 1(%1, %3), %%mm3 \n\t"
441
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
178 "addl %%eax, %2 \n\t"
439
6ae275655a23 * more PIC friendly and faster code
kabi
parents: 429
diff changeset
179 "addl %%eax, %1 \n\t"
448
e8c8ca9106aa * removed MANGLE from macros for setting constants
kabi
parents: 443
diff changeset
180 "psubusb %%mm6, %%mm0 \n\t"
e8c8ca9106aa * removed MANGLE from macros for setting constants
kabi
parents: 443
diff changeset
181 "psubusb %%mm6, %%mm2 \n\t"
386
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
182 PAVGB" %%mm1, %%mm0 \n\t"
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
183 PAVGB" %%mm3, %%mm2 \n\t"
439
6ae275655a23 * more PIC friendly and faster code
kabi
parents: 429
diff changeset
184 "movq %%mm0, (%2) \n\t"
6ae275655a23 * more PIC friendly and faster code
kabi
parents: 429
diff changeset
185 "movq %%mm2, (%2, %3) \n\t"
441
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
186 "addl %%eax, %2 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
187 "subl $4, %0 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
188 "jnz 1b \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
189 :"+g"(h), "+S"(pixels), "+D"(block)
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
190 :"r" (line_size)
386
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
191 :"%eax", "memory");
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
192 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
193
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 954
diff changeset
194 static void DEF(put_pixels8_y2)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
195 {
413
1548abb7bbed * fix for -fPIC compilation - compiles with 2.95.2 as well - any
kabi
parents: 402
diff changeset
196 __asm __volatile(
441
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
197 "lea (%3, %3), %%eax \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
198 "movq (%1), %%mm0 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
199 "subl %3, %2 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
200 "1: \n\t"
413
1548abb7bbed * fix for -fPIC compilation - compiles with 2.95.2 as well - any
kabi
parents: 402
diff changeset
201 "movq (%1, %3), %%mm1 \n\t"
416
ca1f2c0e44ef * fixed contrains and avoid usage of scale index access
kabi
parents: 414
diff changeset
202 "movq (%1, %%eax), %%mm2 \n\t"
441
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
203 "addl %%eax, %1 \n\t"
413
1548abb7bbed * fix for -fPIC compilation - compiles with 2.95.2 as well - any
kabi
parents: 402
diff changeset
204 PAVGB" %%mm1, %%mm0 \n\t"
1548abb7bbed * fix for -fPIC compilation - compiles with 2.95.2 as well - any
kabi
parents: 402
diff changeset
205 PAVGB" %%mm2, %%mm1 \n\t"
1548abb7bbed * fix for -fPIC compilation - compiles with 2.95.2 as well - any
kabi
parents: 402
diff changeset
206 "movq %%mm0, (%2, %3) \n\t"
416
ca1f2c0e44ef * fixed contrains and avoid usage of scale index access
kabi
parents: 414
diff changeset
207 "movq %%mm1, (%2, %%eax) \n\t"
413
1548abb7bbed * fix for -fPIC compilation - compiles with 2.95.2 as well - any
kabi
parents: 402
diff changeset
208 "movq (%1, %3), %%mm1 \n\t"
416
ca1f2c0e44ef * fixed contrains and avoid usage of scale index access
kabi
parents: 414
diff changeset
209 "movq (%1, %%eax), %%mm0 \n\t"
441
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
210 "addl %%eax, %2 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
211 "addl %%eax, %1 \n\t"
413
1548abb7bbed * fix for -fPIC compilation - compiles with 2.95.2 as well - any
kabi
parents: 402
diff changeset
212 PAVGB" %%mm1, %%mm2 \n\t"
1548abb7bbed * fix for -fPIC compilation - compiles with 2.95.2 as well - any
kabi
parents: 402
diff changeset
213 PAVGB" %%mm0, %%mm1 \n\t"
1548abb7bbed * fix for -fPIC compilation - compiles with 2.95.2 as well - any
kabi
parents: 402
diff changeset
214 "movq %%mm2, (%2, %3) \n\t"
416
ca1f2c0e44ef * fixed contrains and avoid usage of scale index access
kabi
parents: 414
diff changeset
215 "movq %%mm1, (%2, %%eax) \n\t"
441
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
216 "addl %%eax, %2 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
217 "subl $4, %0 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
218 "jnz 1b \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
219 :"+g"(h), "+S"(pixels), "+D" (block)
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
220 :"r" (line_size)
413
1548abb7bbed * fix for -fPIC compilation - compiles with 2.95.2 as well - any
kabi
parents: 402
diff changeset
221 :"%eax", "memory");
386
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
222 }
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
223
389
f874493a1970 tried to avoid gcc 2.95.2 bug by puting explicit register constraints - added comment about rounding bug in some functions (need to correct or suppress them for regression tests)
glantau
parents: 387
diff changeset
224 /* GL: this function does incorrect rounding if overflow */
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 954
diff changeset
225 static void DEF(put_no_rnd_pixels8_y2)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
386
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
226 {
448
e8c8ca9106aa * removed MANGLE from macros for setting constants
kabi
parents: 443
diff changeset
227 MOVQ_BONE(mm6);
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
228 __asm __volatile(
441
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
229 "lea (%3, %3), %%eax \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
230 "movq (%1), %%mm0 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
231 "subl %3, %2 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
232 "1: \n\t"
439
6ae275655a23 * more PIC friendly and faster code
kabi
parents: 429
diff changeset
233 "movq (%1, %3), %%mm1 \n\t"
6ae275655a23 * more PIC friendly and faster code
kabi
parents: 429
diff changeset
234 "movq (%1, %%eax), %%mm2 \n\t"
441
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
235 "addl %%eax, %1 \n\t"
448
e8c8ca9106aa * removed MANGLE from macros for setting constants
kabi
parents: 443
diff changeset
236 "psubusb %%mm6, %%mm1 \n\t"
386
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
237 PAVGB" %%mm1, %%mm0 \n\t"
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
238 PAVGB" %%mm2, %%mm1 \n\t"
439
6ae275655a23 * more PIC friendly and faster code
kabi
parents: 429
diff changeset
239 "movq %%mm0, (%2, %3) \n\t"
6ae275655a23 * more PIC friendly and faster code
kabi
parents: 429
diff changeset
240 "movq %%mm1, (%2, %%eax) \n\t"
6ae275655a23 * more PIC friendly and faster code
kabi
parents: 429
diff changeset
241 "movq (%1, %3), %%mm1 \n\t"
6ae275655a23 * more PIC friendly and faster code
kabi
parents: 429
diff changeset
242 "movq (%1, %%eax), %%mm0 \n\t"
441
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
243 "addl %%eax, %2 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
244 "addl %%eax, %1 \n\t"
448
e8c8ca9106aa * removed MANGLE from macros for setting constants
kabi
parents: 443
diff changeset
245 "psubusb %%mm6, %%mm1 \n\t"
386
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
246 PAVGB" %%mm1, %%mm2 \n\t"
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
247 PAVGB" %%mm0, %%mm1 \n\t"
439
6ae275655a23 * more PIC friendly and faster code
kabi
parents: 429
diff changeset
248 "movq %%mm2, (%2, %3) \n\t"
6ae275655a23 * more PIC friendly and faster code
kabi
parents: 429
diff changeset
249 "movq %%mm1, (%2, %%eax) \n\t"
441
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
250 "addl %%eax, %2 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
251 "subl $4, %0 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
252 "jnz 1b \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
253 :"+g"(h), "+S"(pixels), "+D" (block)
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
254 :"r" (line_size)
439
6ae275655a23 * more PIC friendly and faster code
kabi
parents: 429
diff changeset
255 :"%eax", "memory");
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
256 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
257
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 954
diff changeset
258 static void DEF(avg_pixels8)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
259 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
260 __asm __volatile(
441
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
261 "lea (%3, %3), %%eax \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
262 "1: \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
263 "movq (%2), %%mm0 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
264 "movq (%2, %3), %%mm1 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
265 PAVGB" (%1), %%mm0 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
266 PAVGB" (%1, %3), %%mm1 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
267 "movq %%mm0, (%2) \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
268 "movq %%mm1, (%2, %3) \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
269 "addl %%eax, %1 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
270 "addl %%eax, %2 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
271 "movq (%2), %%mm0 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
272 "movq (%2, %3), %%mm1 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
273 PAVGB" (%1), %%mm0 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
274 PAVGB" (%1, %3), %%mm1 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
275 "addl %%eax, %1 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
276 "movq %%mm0, (%2) \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
277 "movq %%mm1, (%2, %3) \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
278 "addl %%eax, %2 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
279 "subl $4, %0 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
280 "jnz 1b \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
281 :"+g"(h), "+S"(pixels), "+D"(block)
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
282 :"r" (line_size)
386
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
283 :"%eax", "memory");
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
284 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
285
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 954
diff changeset
286 static void DEF(avg_pixels8_x2)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
287 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
288 __asm __volatile(
441
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
289 "lea (%3, %3), %%eax \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
290 "1: \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
291 "movq (%1), %%mm0 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
292 "movq (%1, %3), %%mm2 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
293 PAVGB" 1(%1), %%mm0 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
294 PAVGB" 1(%1, %3), %%mm2 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
295 PAVGB" (%2), %%mm0 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
296 PAVGB" (%2, %3), %%mm2 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
297 "addl %%eax, %1 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
298 "movq %%mm0, (%2) \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
299 "movq %%mm2, (%2, %3) \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
300 "movq (%1), %%mm0 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
301 "movq (%1, %3), %%mm2 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
302 PAVGB" 1(%1), %%mm0 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
303 PAVGB" 1(%1, %3), %%mm2 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
304 "addl %%eax, %2 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
305 "addl %%eax, %1 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
306 PAVGB" (%2), %%mm0 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
307 PAVGB" (%2, %3), %%mm2 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
308 "movq %%mm0, (%2) \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
309 "movq %%mm2, (%2, %3) \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
310 "addl %%eax, %2 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
311 "subl $4, %0 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
312 "jnz 1b \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
313 :"+g"(h), "+S"(pixels), "+D"(block)
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
314 :"r" (line_size)
386
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
315 :"%eax", "memory");
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
316 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
317
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 954
diff changeset
318 static void DEF(avg_pixels8_y2)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
319 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
320 __asm __volatile(
441
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
321 "lea (%3, %3), %%eax \n\t"
386
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
322 "movq (%1), %%mm0 \n\t"
441
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
323 "subl %3, %2 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
324 "1: \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
325 "movq (%1, %3), %%mm1 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
326 "movq (%1, %%eax), %%mm2 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
327 "addl %%eax, %1 \n\t"
386
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
328 PAVGB" %%mm1, %%mm0 \n\t"
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
329 PAVGB" %%mm2, %%mm1 \n\t"
441
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
330 "movq (%2, %3), %%mm3 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
331 "movq (%2, %%eax), %%mm4 \n\t"
386
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
332 PAVGB" %%mm3, %%mm0 \n\t"
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
333 PAVGB" %%mm4, %%mm1 \n\t"
441
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
334 "movq %%mm0, (%2, %3) \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
335 "movq %%mm1, (%2, %%eax) \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
336 "movq (%1, %3), %%mm1 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
337 "movq (%1, %%eax), %%mm0 \n\t"
386
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
338 PAVGB" %%mm1, %%mm2 \n\t"
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
339 PAVGB" %%mm0, %%mm1 \n\t"
441
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
340 "addl %%eax, %2 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
341 "addl %%eax, %1 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
342 "movq (%2, %3), %%mm3 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
343 "movq (%2, %%eax), %%mm4 \n\t"
386
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
344 PAVGB" %%mm3, %%mm2 \n\t"
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
345 PAVGB" %%mm4, %%mm1 \n\t"
441
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
346 "movq %%mm2, (%2, %3) \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
347 "movq %%mm1, (%2, %%eax) \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
348 "addl %%eax, %2 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
349 "subl $4, %0 \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
350 "jnz 1b \n\t"
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
351 :"+g"(h), "+S"(pixels), "+D"(block)
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
352 :"r" (line_size)
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
353 :"%eax", "memory");
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
354 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
355
386
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
356 // Note this is not correctly rounded, but this function is only used for b frames so it doesnt matter
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 954
diff changeset
357 static void DEF(avg_pixels8_xy2)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
358 {
448
e8c8ca9106aa * removed MANGLE from macros for setting constants
kabi
parents: 443
diff changeset
359 MOVQ_BONE(mm6);
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
360 __asm __volatile(
442
006965950f49 * optimized remaing avg_pixels_xy2
kabi
parents: 441
diff changeset
361 "lea (%3, %3), %%eax \n\t"
386
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
362 "movq (%1), %%mm0 \n\t"
442
006965950f49 * optimized remaing avg_pixels_xy2
kabi
parents: 441
diff changeset
363 PAVGB" 1(%1), %%mm0 \n\t"
006965950f49 * optimized remaing avg_pixels_xy2
kabi
parents: 441
diff changeset
364 ".balign 8 \n\t"
441
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
365 "1: \n\t"
442
006965950f49 * optimized remaing avg_pixels_xy2
kabi
parents: 441
diff changeset
366 "movq (%1, %%eax), %%mm2 \n\t"
006965950f49 * optimized remaing avg_pixels_xy2
kabi
parents: 441
diff changeset
367 "movq (%1, %3), %%mm1 \n\t"
448
e8c8ca9106aa * removed MANGLE from macros for setting constants
kabi
parents: 443
diff changeset
368 "psubusb %%mm6, %%mm2 \n\t"
442
006965950f49 * optimized remaing avg_pixels_xy2
kabi
parents: 441
diff changeset
369 PAVGB" 1(%1, %3), %%mm1 \n\t"
006965950f49 * optimized remaing avg_pixels_xy2
kabi
parents: 441
diff changeset
370 PAVGB" 1(%1, %%eax), %%mm2 \n\t"
006965950f49 * optimized remaing avg_pixels_xy2
kabi
parents: 441
diff changeset
371 "addl %%eax, %1 \n\t"
386
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
372 PAVGB" %%mm1, %%mm0 \n\t"
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
373 PAVGB" %%mm2, %%mm1 \n\t"
442
006965950f49 * optimized remaing avg_pixels_xy2
kabi
parents: 441
diff changeset
374 PAVGB" (%2), %%mm0 \n\t"
006965950f49 * optimized remaing avg_pixels_xy2
kabi
parents: 441
diff changeset
375 PAVGB" (%2, %3), %%mm1 \n\t"
006965950f49 * optimized remaing avg_pixels_xy2
kabi
parents: 441
diff changeset
376 "movq %%mm0, (%2) \n\t"
006965950f49 * optimized remaing avg_pixels_xy2
kabi
parents: 441
diff changeset
377 "movq %%mm1, (%2, %3) \n\t"
006965950f49 * optimized remaing avg_pixels_xy2
kabi
parents: 441
diff changeset
378 "movq (%1, %3), %%mm1 \n\t"
006965950f49 * optimized remaing avg_pixels_xy2
kabi
parents: 441
diff changeset
379 "movq (%1, %%eax), %%mm0 \n\t"
006965950f49 * optimized remaing avg_pixels_xy2
kabi
parents: 441
diff changeset
380 PAVGB" 1(%1, %3), %%mm1 \n\t"
006965950f49 * optimized remaing avg_pixels_xy2
kabi
parents: 441
diff changeset
381 PAVGB" 1(%1, %%eax), %%mm0 \n\t"
006965950f49 * optimized remaing avg_pixels_xy2
kabi
parents: 441
diff changeset
382 "addl %%eax, %2 \n\t"
006965950f49 * optimized remaing avg_pixels_xy2
kabi
parents: 441
diff changeset
383 "addl %%eax, %1 \n\t"
386
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
384 PAVGB" %%mm1, %%mm2 \n\t"
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
385 PAVGB" %%mm0, %%mm1 \n\t"
442
006965950f49 * optimized remaing avg_pixels_xy2
kabi
parents: 441
diff changeset
386 PAVGB" (%2), %%mm2 \n\t"
006965950f49 * optimized remaing avg_pixels_xy2
kabi
parents: 441
diff changeset
387 PAVGB" (%2, %3), %%mm1 \n\t"
006965950f49 * optimized remaing avg_pixels_xy2
kabi
parents: 441
diff changeset
388 "movq %%mm2, (%2) \n\t"
006965950f49 * optimized remaing avg_pixels_xy2
kabi
parents: 441
diff changeset
389 "movq %%mm1, (%2, %3) \n\t"
006965950f49 * optimized remaing avg_pixels_xy2
kabi
parents: 441
diff changeset
390 "addl %%eax, %2 \n\t"
441
c0de4d3c7d3c * optimized avg_* functions (except xy2)
kabi
parents: 439
diff changeset
391 "subl $4, %0 \n\t"
442
006965950f49 * optimized remaing avg_pixels_xy2
kabi
parents: 441
diff changeset
392 "jnz 1b \n\t"
443
63467327c06c * cosmetic minor change
kabi
parents: 442
diff changeset
393 :"+g"(h), "+S"(pixels), "+D"(block)
442
006965950f49 * optimized remaing avg_pixels_xy2
kabi
parents: 441
diff changeset
394 :"r" (line_size)
386
f49629bab18d hopefully faster mmx2&3dnow MC
michaelni
parents: 151
diff changeset
395 :"%eax", "memory");
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
396 }
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
397
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
398 //FIXME the following could be optimized too ...
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 954
diff changeset
399 static void DEF(put_no_rnd_pixels16_x2)(uint8_t *block, const uint8_t *pixels, int line_size, int h){
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
400 DEF(put_no_rnd_pixels8_x2)(block , pixels , line_size, h);
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
401 DEF(put_no_rnd_pixels8_x2)(block+8, pixels+8, line_size, h);
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
402 }
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 954
diff changeset
403 static void DEF(put_pixels16_y2)(uint8_t *block, const uint8_t *pixels, int line_size, int h){
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
404 DEF(put_pixels8_y2)(block , pixels , line_size, h);
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
405 DEF(put_pixels8_y2)(block+8, pixels+8, line_size, h);
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
406 }
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 954
diff changeset
407 static void DEF(put_no_rnd_pixels16_y2)(uint8_t *block, const uint8_t *pixels, int line_size, int h){
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
408 DEF(put_no_rnd_pixels8_y2)(block , pixels , line_size, h);
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
409 DEF(put_no_rnd_pixels8_y2)(block+8, pixels+8, line_size, h);
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
410 }
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 954
diff changeset
411 static void DEF(avg_pixels16)(uint8_t *block, const uint8_t *pixels, int line_size, int h){
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
412 DEF(avg_pixels8)(block , pixels , line_size, h);
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
413 DEF(avg_pixels8)(block+8, pixels+8, line_size, h);
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
414 }
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 954
diff changeset
415 static void DEF(avg_pixels16_x2)(uint8_t *block, const uint8_t *pixels, int line_size, int h){
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
416 DEF(avg_pixels8_x2)(block , pixels , line_size, h);
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
417 DEF(avg_pixels8_x2)(block+8, pixels+8, line_size, h);
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
418 }
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 954
diff changeset
419 static void DEF(avg_pixels16_y2)(uint8_t *block, const uint8_t *pixels, int line_size, int h){
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
420 DEF(avg_pixels8_y2)(block , pixels , line_size, h);
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
421 DEF(avg_pixels8_y2)(block+8, pixels+8, line_size, h);
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
422 }
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 954
diff changeset
423 static void DEF(avg_pixels16_xy2)(uint8_t *block, const uint8_t *pixels, int line_size, int h){
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
424 DEF(avg_pixels8_xy2)(block , pixels , line_size, h);
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
425 DEF(avg_pixels8_xy2)(block+8, pixels+8, line_size, h);
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
426 }
45e8f39fda50 put/avg_pixels16
michaelni
parents: 448
diff changeset
427