annotate arm/dsputil_iwmmxt_rnd_template.c @ 12512:58bd8cb86f5d libavcodec

Remove unused variable.
author rbultje
date Fri, 24 Sep 2010 15:31:46 +0000
parents 9281a8a9387a
children
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
3699
c537a97eec66 Add official LGPL license headers to the files that were missing them.
diego
parents: 2734
diff changeset
1 /*
c537a97eec66 Add official LGPL license headers to the files that were missing them.
diego
parents: 2734
diff changeset
2 * iWMMXt optimized DSP utils
c537a97eec66 Add official LGPL license headers to the files that were missing them.
diego
parents: 2734
diff changeset
3 * copyright (c) 2004 AGAWA Koji
c537a97eec66 Add official LGPL license headers to the files that were missing them.
diego
parents: 2734
diff changeset
4 *
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3699
diff changeset
5 * This file is part of FFmpeg.
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3699
diff changeset
6 *
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3699
diff changeset
7 * FFmpeg is free software; you can redistribute it and/or
3699
c537a97eec66 Add official LGPL license headers to the files that were missing them.
diego
parents: 2734
diff changeset
8 * modify it under the terms of the GNU Lesser General Public
c537a97eec66 Add official LGPL license headers to the files that were missing them.
diego
parents: 2734
diff changeset
9 * License as published by the Free Software Foundation; either
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3699
diff changeset
10 * version 2.1 of the License, or (at your option) any later version.
3699
c537a97eec66 Add official LGPL license headers to the files that were missing them.
diego
parents: 2734
diff changeset
11 *
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3699
diff changeset
12 * FFmpeg is distributed in the hope that it will be useful,
3699
c537a97eec66 Add official LGPL license headers to the files that were missing them.
diego
parents: 2734
diff changeset
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
c537a97eec66 Add official LGPL license headers to the files that were missing them.
diego
parents: 2734
diff changeset
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
c537a97eec66 Add official LGPL license headers to the files that were missing them.
diego
parents: 2734
diff changeset
15 * Lesser General Public License for more details.
c537a97eec66 Add official LGPL license headers to the files that were missing them.
diego
parents: 2734
diff changeset
16 *
c537a97eec66 Add official LGPL license headers to the files that were missing them.
diego
parents: 2734
diff changeset
17 * You should have received a copy of the GNU Lesser General Public
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3699
diff changeset
18 * License along with FFmpeg; if not, write to the Free Software
3699
c537a97eec66 Add official LGPL license headers to the files that were missing them.
diego
parents: 2734
diff changeset
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
c537a97eec66 Add official LGPL license headers to the files that were missing them.
diego
parents: 2734
diff changeset
20 */
c537a97eec66 Add official LGPL license headers to the files that were missing them.
diego
parents: 2734
diff changeset
21
2734
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
22 void DEF(put, pixels8)(uint8_t *block, const uint8_t *pixels, const int line_size, int h)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
23 {
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
24 int stride = line_size;
8031
eebc7209c47f Convert asm keyword into __asm__.
flameeyes
parents: 6636
diff changeset
25 __asm__ volatile (
2734
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
26 "and r12, %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
27 "bic %[pixels], %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
28 "tmcr wcgr1, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
29 "add r4, %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
30 "add r5, %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
31 "mov %[line_size], %[line_size], lsl #1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
32 "1: \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
33 "wldrd wr0, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
34 "subs %[h], %[h], #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
35 "wldrd wr1, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
36 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
37 "wldrd wr3, [r4] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
38 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
39 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
40 "wldrd wr4, [r4, #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
41 "add r4, r4, %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
42 "walignr1 wr8, wr0, wr1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
43 "pld [r4] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
44 "pld [r4, #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
45 "walignr1 wr10, wr3, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
46 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
47 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
48 "wstrd wr10, [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
49 "add r5, r5, %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
50 "bne 1b \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
51 : [block]"+r"(block), [pixels]"+r"(pixels), [line_size]"+r"(stride), [h]"+r"(h)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
52 :
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
53 : "memory", "r4", "r5", "r12");
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
54 }
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
55
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
56 void DEF(avg, pixels8)(uint8_t *block, const uint8_t *pixels, const int line_size, int h)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
57 {
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
58 int stride = line_size;
8031
eebc7209c47f Convert asm keyword into __asm__.
flameeyes
parents: 6636
diff changeset
59 __asm__ volatile (
2734
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
60 "and r12, %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
61 "bic %[pixels], %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
62 "tmcr wcgr1, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
63 "add r4, %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
64 "add r5, %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
65 "mov %[line_size], %[line_size], lsl #1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
66 "1: \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
67 "wldrd wr0, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
68 "subs %[h], %[h], #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
69 "wldrd wr1, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
70 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
71 "wldrd wr3, [r4] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
72 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
73 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
74 "wldrd wr4, [r4, #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
75 "add r4, r4, %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
76 "walignr1 wr8, wr0, wr1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
77 "wldrd wr0, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
78 "wldrd wr2, [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
79 "pld [r4] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
80 "pld [r4, #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
81 "walignr1 wr10, wr3, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
82 WAVG2B" wr8, wr8, wr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
83 WAVG2B" wr10, wr10, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
84 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
85 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
86 "wstrd wr10, [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
87 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
88 "pld [%[block], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
89 "add r5, r5, %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
90 "pld [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
91 "pld [r5, #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
92 "bne 1b \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
93 : [block]"+r"(block), [pixels]"+r"(pixels), [line_size]"+r"(stride), [h]"+r"(h)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
94 :
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
95 : "memory", "r4", "r5", "r12");
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
96 }
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
97
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
98 void DEF(put, pixels16)(uint8_t *block, const uint8_t *pixels, const int line_size, int h)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
99 {
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
100 int stride = line_size;
8031
eebc7209c47f Convert asm keyword into __asm__.
flameeyes
parents: 6636
diff changeset
101 __asm__ volatile (
2734
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
102 "and r12, %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
103 "bic %[pixels], %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
104 "tmcr wcgr1, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
105 "add r4, %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
106 "add r5, %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
107 "mov %[line_size], %[line_size], lsl #1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
108 "1: \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
109 "wldrd wr0, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
110 "wldrd wr1, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
111 "subs %[h], %[h], #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
112 "wldrd wr2, [%[pixels], #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
113 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
114 "wldrd wr3, [r4] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
115 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
116 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
117 "walignr1 wr8, wr0, wr1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
118 "wldrd wr4, [r4, #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
119 "walignr1 wr9, wr1, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
120 "wldrd wr5, [r4, #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
121 "add r4, r4, %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
122 "pld [r4] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
123 "pld [r4, #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
124 "walignr1 wr10, wr3, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
125 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
126 "walignr1 wr11, wr4, wr5 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
127 "wstrd wr9, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
128 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
129 "wstrd wr10, [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
130 "wstrd wr11, [r5, #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
131 "add r5, r5, %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
132 "bne 1b \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
133 : [block]"+r"(block), [pixels]"+r"(pixels), [line_size]"+r"(stride), [h]"+r"(h)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
134 :
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
135 : "memory", "r4", "r5", "r12");
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
136 }
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
137
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
138 void DEF(avg, pixels16)(uint8_t *block, const uint8_t *pixels, const int line_size, int h)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
139 {
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
140 int stride = line_size;
8031
eebc7209c47f Convert asm keyword into __asm__.
flameeyes
parents: 6636
diff changeset
141 __asm__ volatile (
2734
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
142 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
143 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
144 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
145 "pld [%[block], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
146 "and r12, %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
147 "bic %[pixels], %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
148 "tmcr wcgr1, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
149 "add r4, %[pixels], %[line_size]\n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
150 "add r5, %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
151 "mov %[line_size], %[line_size], lsl #1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
152 "1: \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
153 "wldrd wr0, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
154 "wldrd wr1, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
155 "subs %[h], %[h], #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
156 "wldrd wr2, [%[pixels], #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
157 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
158 "wldrd wr3, [r4] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
159 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
160 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
161 "walignr1 wr8, wr0, wr1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
162 "wldrd wr4, [r4, #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
163 "walignr1 wr9, wr1, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
164 "wldrd wr5, [r4, #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
165 "add r4, r4, %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
166 "wldrd wr0, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
167 "pld [r4] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
168 "wldrd wr1, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
169 "pld [r4, #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
170 "wldrd wr2, [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
171 "walignr1 wr10, wr3, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
172 "wldrd wr3, [r5, #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
173 WAVG2B" wr8, wr8, wr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
174 WAVG2B" wr9, wr9, wr1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
175 WAVG2B" wr10, wr10, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
176 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
177 "walignr1 wr11, wr4, wr5 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
178 WAVG2B" wr11, wr11, wr3 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
179 "wstrd wr9, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
180 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
181 "wstrd wr10, [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
182 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
183 "pld [%[block], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
184 "wstrd wr11, [r5, #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
185 "add r5, r5, %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
186 "pld [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
187 "pld [r5, #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
188 "bne 1b \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
189 : [block]"+r"(block), [pixels]"+r"(pixels), [line_size]"+r"(stride), [h]"+r"(h)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
190 :
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
191 : "memory", "r4", "r5", "r12");
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
192 }
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
193
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
194 void DEF(put, pixels8_x2)(uint8_t *block, const uint8_t *pixels, const int line_size, int h)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
195 {
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
196 int stride = line_size;
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
197 // [wr0 wr1 wr2 wr3] for previous line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
198 // [wr4 wr5 wr6 wr7] for current line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
199 SET_RND(wr15); // =2 for rnd and =1 for no_rnd version
8031
eebc7209c47f Convert asm keyword into __asm__.
flameeyes
parents: 6636
diff changeset
200 __asm__ volatile(
2734
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
201 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
202 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
203 "and r12, %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
204 "bic %[pixels], %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
205 "tmcr wcgr1, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
206 "add r12, r12, #1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
207 "add r4, %[pixels], %[line_size]\n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
208 "tmcr wcgr2, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
209 "add r5, %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
210 "mov %[line_size], %[line_size], lsl #1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
211
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
212 "1: \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
213 "wldrd wr10, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
214 "cmp r12, #8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
215 "wldrd wr11, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
216 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
217 "wldrd wr13, [r4] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
218 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
219 "wldrd wr14, [r4, #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
220 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
221 "add r4, r4, %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
222 "walignr1 wr0, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
223 "pld [r4] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
224 "pld [r4, #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
225 "walignr1 wr2, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
226 "wmoveq wr4, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
227 "wmoveq wr6, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
228 "walignr2ne wr4, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
229 "walignr2ne wr6, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
230 WAVG2B" wr0, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
231 WAVG2B" wr2, wr2, wr6 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
232 "wstrd wr0, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
233 "subs %[h], %[h], #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
234 "wstrd wr2, [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
235 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
236 "add r5, r5, %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
237 "bne 1b \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
238 : [h]"+r"(h), [pixels]"+r"(pixels), [block]"+r"(block), [line_size]"+r"(stride)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
239 :
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
240 : "r4", "r5", "r12", "memory");
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
241 }
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
242
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
243 void DEF(put, pixels16_x2)(uint8_t *block, const uint8_t *pixels, const int line_size, int h)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
244 {
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
245 int stride = line_size;
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
246 // [wr0 wr1 wr2 wr3] for previous line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
247 // [wr4 wr5 wr6 wr7] for current line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
248 SET_RND(wr15); // =2 for rnd and =1 for no_rnd version
8031
eebc7209c47f Convert asm keyword into __asm__.
flameeyes
parents: 6636
diff changeset
249 __asm__ volatile(
2734
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
250 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
251 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
252 "and r12, %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
253 "bic %[pixels], %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
254 "tmcr wcgr1, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
255 "add r12, r12, #1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
256 "add r4, %[pixels], %[line_size]\n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
257 "tmcr wcgr2, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
258 "add r5, %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
259 "mov %[line_size], %[line_size], lsl #1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
260
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
261 "1: \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
262 "wldrd wr10, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
263 "cmp r12, #8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
264 "wldrd wr11, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
265 "wldrd wr12, [%[pixels], #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
266 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
267 "wldrd wr13, [r4] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
268 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
269 "wldrd wr14, [r4, #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
270 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
271 "wldrd wr15, [r4, #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
272 "add r4, r4, %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
273 "walignr1 wr0, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
274 "pld [r4] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
275 "pld [r4, #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
276 "walignr1 wr1, wr11, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
277 "walignr1 wr2, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
278 "walignr1 wr3, wr14, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
279 "wmoveq wr4, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
280 "wmoveq wr5, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
281 "wmoveq wr6, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
282 "wmoveq wr7, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
283 "walignr2ne wr4, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
284 "walignr2ne wr5, wr11, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
285 "walignr2ne wr6, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
286 "walignr2ne wr7, wr14, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
287 WAVG2B" wr0, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
288 WAVG2B" wr1, wr1, wr5 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
289 "wstrd wr0, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
290 WAVG2B" wr2, wr2, wr6 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
291 "wstrd wr1, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
292 WAVG2B" wr3, wr3, wr7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
293 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
294 "wstrd wr2, [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
295 "subs %[h], %[h], #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
296 "wstrd wr3, [r5, #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
297 "add r5, r5, %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
298 "bne 1b \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
299 : [h]"+r"(h), [pixels]"+r"(pixels), [block]"+r"(block), [line_size]"+r"(stride)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
300 :
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
301 : "r4", "r5", "r12", "memory");
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
302 }
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
303
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
304 void DEF(avg, pixels8_x2)(uint8_t *block, const uint8_t *pixels, const int line_size, int h)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
305 {
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
306 int stride = line_size;
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
307 // [wr0 wr1 wr2 wr3] for previous line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
308 // [wr4 wr5 wr6 wr7] for current line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
309 SET_RND(wr15); // =2 for rnd and =1 for no_rnd version
8031
eebc7209c47f Convert asm keyword into __asm__.
flameeyes
parents: 6636
diff changeset
310 __asm__ volatile(
2734
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
311 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
312 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
313 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
314 "pld [%[block], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
315 "and r12, %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
316 "bic %[pixels], %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
317 "tmcr wcgr1, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
318 "add r12, r12, #1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
319 "add r4, %[pixels], %[line_size]\n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
320 "tmcr wcgr2, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
321 "add r5, %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
322 "mov %[line_size], %[line_size], lsl #1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
323 "pld [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
324 "pld [r5, #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
325
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
326 "1: \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
327 "wldrd wr10, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
328 "cmp r12, #8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
329 "wldrd wr11, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
330 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
331 "wldrd wr13, [r4] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
332 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
333 "wldrd wr14, [r4, #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
334 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
335 "add r4, r4, %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
336 "walignr1 wr0, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
337 "pld [r4] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
338 "pld [r4, #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
339 "walignr1 wr2, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
340 "wmoveq wr4, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
341 "wmoveq wr6, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
342 "walignr2ne wr4, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
343 "wldrd wr10, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
344 "walignr2ne wr6, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
345 "wldrd wr12, [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
346 WAVG2B" wr0, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
347 WAVG2B" wr2, wr2, wr6 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
348 WAVG2B" wr0, wr0, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
349 WAVG2B" wr2, wr2, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
350 "wstrd wr0, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
351 "subs %[h], %[h], #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
352 "wstrd wr2, [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
353 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
354 "add r5, r5, %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
355 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
356 "pld [%[block], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
357 "pld [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
358 "pld [r5, #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
359 "bne 1b \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
360 : [h]"+r"(h), [pixels]"+r"(pixels), [block]"+r"(block), [line_size]"+r"(stride)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
361 :
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
362 : "r4", "r5", "r12", "memory");
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
363 }
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
364
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
365 void DEF(avg, pixels16_x2)(uint8_t *block, const uint8_t *pixels, const int line_size, int h)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
366 {
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
367 int stride = line_size;
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
368 // [wr0 wr1 wr2 wr3] for previous line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
369 // [wr4 wr5 wr6 wr7] for current line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
370 SET_RND(wr15); // =2 for rnd and =1 for no_rnd version
8031
eebc7209c47f Convert asm keyword into __asm__.
flameeyes
parents: 6636
diff changeset
371 __asm__ volatile(
2734
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
372 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
373 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
374 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
375 "pld [%[block], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
376 "and r12, %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
377 "bic %[pixels], %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
378 "tmcr wcgr1, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
379 "add r12, r12, #1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
380 "add r4, %[pixels], %[line_size]\n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
381 "tmcr wcgr2, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
382 "add r5, %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
383 "mov %[line_size], %[line_size], lsl #1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
384 "pld [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
385 "pld [r5, #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
386
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
387 "1: \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
388 "wldrd wr10, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
389 "cmp r12, #8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
390 "wldrd wr11, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
391 "wldrd wr12, [%[pixels], #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
392 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
393 "wldrd wr13, [r4] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
394 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
395 "wldrd wr14, [r4, #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
396 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
397 "wldrd wr15, [r4, #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
398 "add r4, r4, %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
399 "walignr1 wr0, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
400 "pld [r4] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
401 "pld [r4, #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
402 "walignr1 wr1, wr11, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
403 "walignr1 wr2, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
404 "walignr1 wr3, wr14, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
405 "wmoveq wr4, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
406 "wmoveq wr5, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
407 "wmoveq wr6, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
408 "wmoveq wr7, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
409 "walignr2ne wr4, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
410 "walignr2ne wr5, wr11, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
411 "walignr2ne wr6, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
412 "walignr2ne wr7, wr14, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
413 "wldrd wr10, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
414 WAVG2B" wr0, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
415 "wldrd wr11, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
416 WAVG2B" wr1, wr1, wr5 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
417 "wldrd wr12, [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
418 WAVG2B" wr2, wr2, wr6 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
419 "wldrd wr13, [r5, #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
420 WAVG2B" wr3, wr3, wr7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
421 WAVG2B" wr0, wr0, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
422 WAVG2B" wr1, wr1, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
423 WAVG2B" wr2, wr2, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
424 WAVG2B" wr3, wr3, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
425 "wstrd wr0, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
426 "subs %[h], %[h], #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
427 "wstrd wr1, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
428 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
429 "wstrd wr2, [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
430 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
431 "wstrd wr3, [r5, #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
432 "add r5, r5, %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
433 "pld [%[block], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
434 "pld [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
435 "pld [r5, #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
436 "bne 1b \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
437 : [h]"+r"(h), [pixels]"+r"(pixels), [block]"+r"(block), [line_size]"+r"(stride)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
438 :
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
439 :"r4", "r5", "r12", "memory");
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
440 }
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
441
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
442 void DEF(avg, pixels8_y2)(uint8_t *block, const uint8_t *pixels, const int line_size, int h)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
443 {
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
444 int stride = line_size;
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
445 // [wr0 wr1 wr2 wr3] for previous line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
446 // [wr4 wr5 wr6 wr7] for current line
8031
eebc7209c47f Convert asm keyword into __asm__.
flameeyes
parents: 6636
diff changeset
447 __asm__ volatile(
2734
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
448 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
449 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
450 "and r12, %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
451 "tmcr wcgr1, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
452 "bic %[pixels], %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
453
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
454 "wldrd wr10, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
455 "wldrd wr11, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
456 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
457 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
458 "walignr1 wr0, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
459 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
460 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
461
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
462 "1: \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
463 "wldrd wr10, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
464 "wldrd wr11, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
465 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
466 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
467 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
468 "walignr1 wr4, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
469 "wldrd wr10, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
470 WAVG2B" wr8, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
471 WAVG2B" wr8, wr8, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
472 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
473 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
474
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
475 "wldrd wr10, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
476 "wldrd wr11, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
477 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
478 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
479 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
480 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
481 "walignr1 wr0, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
482 "wldrd wr10, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
483 WAVG2B" wr8, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
484 WAVG2B" wr8, wr8, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
485 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
486 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
487
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
488 "subs %[h], %[h], #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
489 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
490 "bne 1b \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
491 : [h]"+r"(h), [pixels]"+r"(pixels), [block]"+r"(block), [line_size]"+r"(stride)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
492 :
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
493 : "cc", "memory", "r12");
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
494 }
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
495
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
496 void DEF(put, pixels16_y2)(uint8_t *block, const uint8_t *pixels, const int line_size, int h)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
497 {
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
498 int stride = line_size;
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
499 // [wr0 wr1 wr2 wr3] for previous line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
500 // [wr4 wr5 wr6 wr7] for current line
8031
eebc7209c47f Convert asm keyword into __asm__.
flameeyes
parents: 6636
diff changeset
501 __asm__ volatile(
2734
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
502 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
503 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
504 "and r12, %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
505 "tmcr wcgr1, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
506 "bic %[pixels], %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
507
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
508 "wldrd wr10, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
509 "wldrd wr11, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
510 "wldrd wr12, [%[pixels], #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
511 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
512 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
513 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
514 "walignr1 wr0, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
515 "walignr1 wr1, wr11, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
516
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
517 "1: \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
518 "wldrd wr10, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
519 "wldrd wr11, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
520 "wldrd wr12, [%[pixels], #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
521 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
522 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
523 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
524 "walignr1 wr4, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
525 "walignr1 wr5, wr11, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
526 WAVG2B" wr8, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
527 WAVG2B" wr9, wr1, wr5 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
528 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
529 "wstrd wr9, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
530 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
531
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
532 "wldrd wr10, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
533 "wldrd wr11, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
534 "wldrd wr12, [%[pixels], #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
535 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
536 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
537 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
538 "walignr1 wr0, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
539 "walignr1 wr1, wr11, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
540 WAVG2B" wr8, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
541 WAVG2B" wr9, wr1, wr5 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
542 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
543 "wstrd wr9, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
544 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
545
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
546 "subs %[h], %[h], #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
547 "bne 1b \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
548 : [h]"+r"(h), [pixels]"+r"(pixels), [block]"+r"(block), [line_size]"+r"(stride)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
549 :
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
550 : "r4", "r5", "r12", "memory");
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
551 }
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
552
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
553 void DEF(avg, pixels16_y2)(uint8_t *block, const uint8_t *pixels, const int line_size, int h)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
554 {
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
555 int stride = line_size;
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
556 // [wr0 wr1 wr2 wr3] for previous line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
557 // [wr4 wr5 wr6 wr7] for current line
8031
eebc7209c47f Convert asm keyword into __asm__.
flameeyes
parents: 6636
diff changeset
558 __asm__ volatile(
2734
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
559 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
560 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
561 "and r12, %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
562 "tmcr wcgr1, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
563 "bic %[pixels], %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
564
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
565 "wldrd wr10, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
566 "wldrd wr11, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
567 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
568 "wldrd wr12, [%[pixels], #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
569 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
570 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
571 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
572 "walignr1 wr0, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
573 "walignr1 wr1, wr11, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
574
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
575 "1: \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
576 "wldrd wr10, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
577 "wldrd wr11, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
578 "wldrd wr12, [%[pixels], #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
579 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
580 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
581 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
582 "walignr1 wr4, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
583 "walignr1 wr5, wr11, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
584 "wldrd wr10, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
585 "wldrd wr11, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
586 WAVG2B" wr8, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
587 WAVG2B" wr9, wr1, wr5 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
588 WAVG2B" wr8, wr8, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
589 WAVG2B" wr9, wr9, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
590 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
591 "wstrd wr9, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
592 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
593
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
594 "wldrd wr10, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
595 "wldrd wr11, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
596 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
597 "wldrd wr12, [%[pixels], #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
598 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
599 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
600 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
601 "walignr1 wr0, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
602 "walignr1 wr1, wr11, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
603 "wldrd wr10, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
604 "wldrd wr11, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
605 WAVG2B" wr8, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
606 WAVG2B" wr9, wr1, wr5 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
607 WAVG2B" wr8, wr8, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
608 WAVG2B" wr9, wr9, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
609 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
610 "wstrd wr9, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
611 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
612
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
613 "subs %[h], %[h], #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
614 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
615 "bne 1b \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
616 : [h]"+r"(h), [pixels]"+r"(pixels), [block]"+r"(block), [line_size]"+r"(stride)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
617 :
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
618 : "r4", "r5", "r12", "memory");
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
619 }
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
620
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
621 void DEF(put, pixels8_xy2)(uint8_t *block, const uint8_t *pixels, const int line_size, int h)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
622 {
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
623 // [wr0 wr1 wr2 wr3] for previous line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
624 // [wr4 wr5 wr6 wr7] for current line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
625 SET_RND(wr15); // =2 for rnd and =1 for no_rnd version
8031
eebc7209c47f Convert asm keyword into __asm__.
flameeyes
parents: 6636
diff changeset
626 __asm__ volatile(
2734
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
627 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
628 "mov r12, #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
629 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
630 "tmcr wcgr0, r12 \n\t" /* for shift value */
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
631 "and r12, %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
632 "bic %[pixels], %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
633 "tmcr wcgr1, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
634
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
635 // [wr0 wr1 wr2 wr3] <= *
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
636 // [wr4 wr5 wr6 wr7]
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
637 "wldrd wr12, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
638 "add r12, r12, #1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
639 "wldrd wr13, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
640 "tmcr wcgr2, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
641 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
642 "cmp r12, #8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
643 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
644 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
645 "walignr1 wr2, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
646 "wmoveq wr10, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
647 "walignr2ne wr10, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
648 "wunpckelub wr0, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
649 "wunpckehub wr1, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
650 "wunpckelub wr8, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
651 "wunpckehub wr9, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
652 "waddhus wr0, wr0, wr8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
653 "waddhus wr1, wr1, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
654
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
655 "1: \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
656 // [wr0 wr1 wr2 wr3]
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
657 // [wr4 wr5 wr6 wr7] <= *
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
658 "wldrd wr12, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
659 "cmp r12, #8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
660 "wldrd wr13, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
661 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
662 "walignr1 wr6, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
663 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
664 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
665 "wmoveq wr10, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
666 "walignr2ne wr10, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
667 "wunpckelub wr4, wr6 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
668 "wunpckehub wr5, wr6 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
669 "wunpckelub wr8, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
670 "wunpckehub wr9, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
671 "waddhus wr4, wr4, wr8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
672 "waddhus wr5, wr5, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
673 "waddhus wr8, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
674 "waddhus wr9, wr1, wr5 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
675 "waddhus wr8, wr8, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
676 "waddhus wr9, wr9, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
677 "wsrlhg wr8, wr8, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
678 "wsrlhg wr9, wr9, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
679 "wpackhus wr8, wr8, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
680 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
681 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
682
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
683 // [wr0 wr1 wr2 wr3] <= *
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
684 // [wr4 wr5 wr6 wr7]
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
685 "wldrd wr12, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
686 "wldrd wr13, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
687 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
688 "walignr1 wr2, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
689 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
690 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
691 "wmoveq wr10, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
692 "walignr2ne wr10, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
693 "wunpckelub wr0, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
694 "wunpckehub wr1, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
695 "wunpckelub wr8, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
696 "wunpckehub wr9, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
697 "waddhus wr0, wr0, wr8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
698 "waddhus wr1, wr1, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
699 "waddhus wr8, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
700 "waddhus wr9, wr1, wr5 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
701 "waddhus wr8, wr8, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
702 "waddhus wr9, wr9, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
703 "wsrlhg wr8, wr8, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
704 "wsrlhg wr9, wr9, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
705 "wpackhus wr8, wr8, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
706 "subs %[h], %[h], #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
707 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
708 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
709 "bne 1b \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
710 : [h]"+r"(h), [pixels]"+r"(pixels), [block]"+r"(block)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
711 : [line_size]"r"(line_size)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
712 : "r12", "memory");
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
713 }
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
714
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
715 void DEF(put, pixels16_xy2)(uint8_t *block, const uint8_t *pixels, const int line_size, int h)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
716 {
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
717 // [wr0 wr1 wr2 wr3] for previous line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
718 // [wr4 wr5 wr6 wr7] for current line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
719 SET_RND(wr15); // =2 for rnd and =1 for no_rnd version
8031
eebc7209c47f Convert asm keyword into __asm__.
flameeyes
parents: 6636
diff changeset
720 __asm__ volatile(
2734
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
721 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
722 "mov r12, #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
723 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
724 "tmcr wcgr0, r12 \n\t" /* for shift value */
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
725 /* alignment */
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
726 "and r12, %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
727 "bic %[pixels], %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
728 "tmcr wcgr1, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
729 "add r12, r12, #1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
730 "tmcr wcgr2, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
731
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
732 // [wr0 wr1 wr2 wr3] <= *
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
733 // [wr4 wr5 wr6 wr7]
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
734 "wldrd wr12, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
735 "cmp r12, #8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
736 "wldrd wr13, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
737 "wldrd wr14, [%[pixels], #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
738 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
739 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
740 "walignr1 wr2, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
741 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
742 "walignr1 wr3, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
743 "wmoveq wr10, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
744 "wmoveq wr11, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
745 "walignr2ne wr10, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
746 "walignr2ne wr11, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
747 "wunpckelub wr0, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
748 "wunpckehub wr1, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
749 "wunpckelub wr2, wr3 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
750 "wunpckehub wr3, wr3 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
751 "wunpckelub wr8, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
752 "wunpckehub wr9, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
753 "wunpckelub wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
754 "wunpckehub wr11, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
755 "waddhus wr0, wr0, wr8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
756 "waddhus wr1, wr1, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
757 "waddhus wr2, wr2, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
758 "waddhus wr3, wr3, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
759
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
760 "1: \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
761 // [wr0 wr1 wr2 wr3]
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
762 // [wr4 wr5 wr6 wr7] <= *
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
763 "wldrd wr12, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
764 "cmp r12, #8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
765 "wldrd wr13, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
766 "wldrd wr14, [%[pixels], #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
767 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
768 "walignr1 wr6, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
769 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
770 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
771 "walignr1 wr7, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
772 "wmoveq wr10, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
773 "wmoveq wr11, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
774 "walignr2ne wr10, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
775 "walignr2ne wr11, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
776 "wunpckelub wr4, wr6 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
777 "wunpckehub wr5, wr6 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
778 "wunpckelub wr6, wr7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
779 "wunpckehub wr7, wr7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
780 "wunpckelub wr8, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
781 "wunpckehub wr9, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
782 "wunpckelub wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
783 "wunpckehub wr11, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
784 "waddhus wr4, wr4, wr8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
785 "waddhus wr5, wr5, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
786 "waddhus wr6, wr6, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
787 "waddhus wr7, wr7, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
788 "waddhus wr8, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
789 "waddhus wr9, wr1, wr5 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
790 "waddhus wr10, wr2, wr6 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
791 "waddhus wr11, wr3, wr7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
792 "waddhus wr8, wr8, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
793 "waddhus wr9, wr9, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
794 "waddhus wr10, wr10, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
795 "waddhus wr11, wr11, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
796 "wsrlhg wr8, wr8, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
797 "wsrlhg wr9, wr9, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
798 "wsrlhg wr10, wr10, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
799 "wsrlhg wr11, wr11, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
800 "wpackhus wr8, wr8, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
801 "wpackhus wr9, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
802 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
803 "wstrd wr9, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
804 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
805
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
806 // [wr0 wr1 wr2 wr3] <= *
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
807 // [wr4 wr5 wr6 wr7]
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
808 "wldrd wr12, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
809 "wldrd wr13, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
810 "wldrd wr14, [%[pixels], #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
811 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
812 "walignr1 wr2, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
813 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
814 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
815 "walignr1 wr3, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
816 "wmoveq wr10, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
817 "wmoveq wr11, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
818 "walignr2ne wr10, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
819 "walignr2ne wr11, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
820 "wunpckelub wr0, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
821 "wunpckehub wr1, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
822 "wunpckelub wr2, wr3 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
823 "wunpckehub wr3, wr3 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
824 "wunpckelub wr8, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
825 "wunpckehub wr9, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
826 "wunpckelub wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
827 "wunpckehub wr11, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
828 "waddhus wr0, wr0, wr8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
829 "waddhus wr1, wr1, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
830 "waddhus wr2, wr2, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
831 "waddhus wr3, wr3, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
832 "waddhus wr8, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
833 "waddhus wr9, wr1, wr5 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
834 "waddhus wr10, wr2, wr6 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
835 "waddhus wr11, wr3, wr7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
836 "waddhus wr8, wr8, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
837 "waddhus wr9, wr9, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
838 "waddhus wr10, wr10, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
839 "waddhus wr11, wr11, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
840 "wsrlhg wr8, wr8, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
841 "wsrlhg wr9, wr9, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
842 "wsrlhg wr10, wr10, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
843 "wsrlhg wr11, wr11, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
844 "wpackhus wr8, wr8, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
845 "wpackhus wr9, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
846 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
847 "wstrd wr9, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
848 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
849
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
850 "subs %[h], %[h], #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
851 "bne 1b \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
852 : [h]"+r"(h), [pixels]"+r"(pixels), [block]"+r"(block)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
853 : [line_size]"r"(line_size)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
854 : "r12", "memory");
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
855 }
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
856
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
857 void DEF(avg, pixels8_xy2)(uint8_t *block, const uint8_t *pixels, const int line_size, int h)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
858 {
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
859 // [wr0 wr1 wr2 wr3] for previous line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
860 // [wr4 wr5 wr6 wr7] for current line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
861 SET_RND(wr15); // =2 for rnd and =1 for no_rnd version
8031
eebc7209c47f Convert asm keyword into __asm__.
flameeyes
parents: 6636
diff changeset
862 __asm__ volatile(
2734
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
863 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
864 "pld [%[block], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
865 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
866 "mov r12, #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
867 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
868 "tmcr wcgr0, r12 \n\t" /* for shift value */
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
869 "and r12, %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
870 "bic %[pixels], %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
871 "tmcr wcgr1, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
872
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
873 // [wr0 wr1 wr2 wr3] <= *
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
874 // [wr4 wr5 wr6 wr7]
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
875 "wldrd wr12, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
876 "add r12, r12, #1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
877 "wldrd wr13, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
878 "tmcr wcgr2, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
879 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
880 "cmp r12, #8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
881 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
882 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
883 "walignr1 wr2, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
884 "wmoveq wr10, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
885 "walignr2ne wr10, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
886 "wunpckelub wr0, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
887 "wunpckehub wr1, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
888 "wunpckelub wr8, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
889 "wunpckehub wr9, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
890 "waddhus wr0, wr0, wr8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
891 "waddhus wr1, wr1, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
892
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
893 "1: \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
894 // [wr0 wr1 wr2 wr3]
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
895 // [wr4 wr5 wr6 wr7] <= *
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
896 "wldrd wr12, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
897 "cmp r12, #8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
898 "wldrd wr13, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
899 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
900 "walignr1 wr6, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
901 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
902 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
903 "wmoveq wr10, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
904 "walignr2ne wr10, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
905 "wunpckelub wr4, wr6 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
906 "wunpckehub wr5, wr6 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
907 "wunpckelub wr8, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
908 "wunpckehub wr9, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
909 "waddhus wr4, wr4, wr8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
910 "waddhus wr5, wr5, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
911 "waddhus wr8, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
912 "waddhus wr9, wr1, wr5 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
913 "waddhus wr8, wr8, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
914 "waddhus wr9, wr9, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
915 "wldrd wr12, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
916 "wsrlhg wr8, wr8, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
917 "wsrlhg wr9, wr9, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
918 "wpackhus wr8, wr8, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
919 WAVG2B" wr8, wr8, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
920 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
921 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
922 "wldrd wr12, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
923 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
924 "pld [%[block], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
925
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
926 // [wr0 wr1 wr2 wr3] <= *
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
927 // [wr4 wr5 wr6 wr7]
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
928 "wldrd wr13, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
929 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
930 "walignr1 wr2, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
931 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
932 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
933 "wmoveq wr10, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
934 "walignr2ne wr10, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
935 "wunpckelub wr0, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
936 "wunpckehub wr1, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
937 "wunpckelub wr8, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
938 "wunpckehub wr9, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
939 "waddhus wr0, wr0, wr8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
940 "waddhus wr1, wr1, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
941 "waddhus wr8, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
942 "waddhus wr9, wr1, wr5 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
943 "waddhus wr8, wr8, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
944 "waddhus wr9, wr9, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
945 "wldrd wr12, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
946 "wsrlhg wr8, wr8, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
947 "wsrlhg wr9, wr9, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
948 "wpackhus wr8, wr8, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
949 "subs %[h], %[h], #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
950 WAVG2B" wr8, wr8, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
951 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
952 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
953 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
954 "pld [%[block], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
955 "bne 1b \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
956 : [h]"+r"(h), [pixels]"+r"(pixels), [block]"+r"(block)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
957 : [line_size]"r"(line_size)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
958 : "r12", "memory");
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
959 }
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
960
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
961 void DEF(avg, pixels16_xy2)(uint8_t *block, const uint8_t *pixels, const int line_size, int h)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
962 {
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
963 // [wr0 wr1 wr2 wr3] for previous line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
964 // [wr4 wr5 wr6 wr7] for current line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
965 SET_RND(wr15); // =2 for rnd and =1 for no_rnd version
8031
eebc7209c47f Convert asm keyword into __asm__.
flameeyes
parents: 6636
diff changeset
966 __asm__ volatile(
2734
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
967 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
968 "pld [%[block], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
969 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
970 "mov r12, #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
971 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
972 "tmcr wcgr0, r12 \n\t" /* for shift value */
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
973 /* alignment */
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
974 "and r12, %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
975 "bic %[pixels], %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
976 "tmcr wcgr1, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
977 "add r12, r12, #1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
978 "tmcr wcgr2, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
979
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
980 // [wr0 wr1 wr2 wr3] <= *
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
981 // [wr4 wr5 wr6 wr7]
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
982 "wldrd wr12, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
983 "cmp r12, #8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
984 "wldrd wr13, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
985 "wldrd wr14, [%[pixels], #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
986 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
987 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
988 "walignr1 wr2, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
989 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
990 "walignr1 wr3, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
991 "wmoveq wr10, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
992 "wmoveq wr11, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
993 "walignr2ne wr10, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
994 "walignr2ne wr11, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
995 "wunpckelub wr0, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
996 "wunpckehub wr1, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
997 "wunpckelub wr2, wr3 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
998 "wunpckehub wr3, wr3 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
999 "wunpckelub wr8, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1000 "wunpckehub wr9, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1001 "wunpckelub wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1002 "wunpckehub wr11, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1003 "waddhus wr0, wr0, wr8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1004 "waddhus wr1, wr1, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1005 "waddhus wr2, wr2, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1006 "waddhus wr3, wr3, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1007
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1008 "1: \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1009 // [wr0 wr1 wr2 wr3]
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1010 // [wr4 wr5 wr6 wr7] <= *
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1011 "wldrd wr12, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1012 "cmp r12, #8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1013 "wldrd wr13, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1014 "wldrd wr14, [%[pixels], #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1015 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1016 "walignr1 wr6, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1017 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1018 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1019 "walignr1 wr7, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1020 "wmoveq wr10, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1021 "wmoveq wr11, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1022 "walignr2ne wr10, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1023 "walignr2ne wr11, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1024 "wunpckelub wr4, wr6 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1025 "wunpckehub wr5, wr6 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1026 "wunpckelub wr6, wr7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1027 "wunpckehub wr7, wr7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1028 "wunpckelub wr8, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1029 "wunpckehub wr9, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1030 "wunpckelub wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1031 "wunpckehub wr11, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1032 "waddhus wr4, wr4, wr8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1033 "waddhus wr5, wr5, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1034 "waddhus wr6, wr6, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1035 "waddhus wr7, wr7, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1036 "waddhus wr8, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1037 "waddhus wr9, wr1, wr5 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1038 "waddhus wr10, wr2, wr6 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1039 "waddhus wr11, wr3, wr7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1040 "waddhus wr8, wr8, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1041 "waddhus wr9, wr9, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1042 "waddhus wr10, wr10, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1043 "waddhus wr11, wr11, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1044 "wsrlhg wr8, wr8, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1045 "wsrlhg wr9, wr9, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1046 "wldrd wr12, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1047 "wldrd wr13, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1048 "wsrlhg wr10, wr10, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1049 "wsrlhg wr11, wr11, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1050 "wpackhus wr8, wr8, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1051 "wpackhus wr9, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1052 WAVG2B" wr8, wr8, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1053 WAVG2B" wr9, wr9, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1054 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1055 "wstrd wr9, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1056 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1057
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1058 // [wr0 wr1 wr2 wr3] <= *
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1059 // [wr4 wr5 wr6 wr7]
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1060 "wldrd wr12, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1061 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1062 "wldrd wr13, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1063 "pld [%[block], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1064 "wldrd wr14, [%[pixels], #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1065 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1066 "walignr1 wr2, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1067 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1068 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1069 "walignr1 wr3, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1070 "wmoveq wr10, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1071 "wmoveq wr11, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1072 "walignr2ne wr10, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1073 "walignr2ne wr11, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1074 "wunpckelub wr0, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1075 "wunpckehub wr1, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1076 "wunpckelub wr2, wr3 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1077 "wunpckehub wr3, wr3 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1078 "wunpckelub wr8, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1079 "wunpckehub wr9, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1080 "wunpckelub wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1081 "wunpckehub wr11, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1082 "waddhus wr0, wr0, wr8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1083 "waddhus wr1, wr1, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1084 "waddhus wr2, wr2, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1085 "waddhus wr3, wr3, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1086 "waddhus wr8, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1087 "waddhus wr9, wr1, wr5 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1088 "waddhus wr10, wr2, wr6 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1089 "waddhus wr11, wr3, wr7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1090 "waddhus wr8, wr8, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1091 "waddhus wr9, wr9, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1092 "waddhus wr10, wr10, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1093 "waddhus wr11, wr11, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1094 "wsrlhg wr8, wr8, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1095 "wsrlhg wr9, wr9, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1096 "wldrd wr12, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1097 "wldrd wr13, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1098 "wsrlhg wr10, wr10, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1099 "wsrlhg wr11, wr11, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1100 "wpackhus wr8, wr8, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1101 "wpackhus wr9, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1102 WAVG2B" wr8, wr8, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1103 WAVG2B" wr9, wr9, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1104 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1105 "wstrd wr9, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1106 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1107 "subs %[h], %[h], #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1108 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1109 "pld [%[block], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1110 "bne 1b \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1111 : [h]"+r"(h), [pixels]"+r"(pixels), [block]"+r"(block)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1112 : [line_size]"r"(line_size)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1113 : "r12", "memory");
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1114 }