Mercurial > libavcodec.hg

annotate armv4l/dsputil_iwmmxt_rnd.h @ 3176:babf844e1308 libavcodec

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Init simplification and 2% faster wma_decode_block on amd64 with tables use instead of pow().
author banan
date Wed, 08 Mar 2006 09:26:57 +0000
parents aeea63c97878
children c537a97eec66
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
2734
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1 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
2 {
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
3 int stride = line_size;
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
4 __asm__ __volatile__ (
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
5 "and r12, %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
6 "bic %[pixels], %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
7 "tmcr wcgr1, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
8 "add r4, %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
9 "add r5, %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
10 "mov %[line_size], %[line_size], lsl #1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
11 "1: \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
12 "wldrd wr0, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
13 "subs %[h], %[h], #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
14 "wldrd wr1, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
15 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
16 "wldrd wr3, [r4] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
17 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
18 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
19 "wldrd wr4, [r4, #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
20 "add r4, r4, %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
21 "walignr1 wr8, wr0, wr1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
22 "pld [r4] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
23 "pld [r4, #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
24 "walignr1 wr10, wr3, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
25 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
26 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
27 "wstrd wr10, [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
28 "add r5, r5, %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
29 "bne 1b \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
30 : [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
31 :
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
32 : "memory", "r4", "r5", "r12");
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
33 }
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
34
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
35 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
36 {
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
37 int stride = line_size;
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
38 __asm__ __volatile__ (
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
39 "and r12, %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
40 "bic %[pixels], %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
41 "tmcr wcgr1, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
42 "add r4, %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
43 "add r5, %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
44 "mov %[line_size], %[line_size], lsl #1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
45 "1: \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
46 "wldrd wr0, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
47 "subs %[h], %[h], #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
48 "wldrd wr1, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
49 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
50 "wldrd wr3, [r4] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
51 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
52 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
53 "wldrd wr4, [r4, #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
54 "add r4, r4, %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
55 "walignr1 wr8, wr0, wr1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
56 "wldrd wr0, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
57 "wldrd wr2, [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
58 "pld [r4] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
59 "pld [r4, #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
60 "walignr1 wr10, wr3, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
61 WAVG2B" wr8, wr8, wr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
62 WAVG2B" wr10, wr10, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
63 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
64 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
65 "wstrd wr10, [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
66 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
67 "pld [%[block], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
68 "add r5, r5, %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
69 "pld [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
70 "pld [r5, #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
71 "bne 1b \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
72 : [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
73 :
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
74 : "memory", "r4", "r5", "r12");
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
75 }
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
76
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
77 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
78 {
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
79 int stride = line_size;
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
80 __asm__ __volatile__ (
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
81 "and r12, %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
82 "bic %[pixels], %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
83 "tmcr wcgr1, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
84 "add r4, %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
85 "add r5, %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
86 "mov %[line_size], %[line_size], lsl #1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
87 "1: \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
88 "wldrd wr0, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
89 "wldrd wr1, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
90 "subs %[h], %[h], #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
91 "wldrd wr2, [%[pixels], #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
92 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
93 "wldrd wr3, [r4] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
94 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
95 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
96 "walignr1 wr8, wr0, wr1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
97 "wldrd wr4, [r4, #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
98 "walignr1 wr9, wr1, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
99 "wldrd wr5, [r4, #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
100 "add r4, r4, %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
101 "pld [r4] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
102 "pld [r4, #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
103 "walignr1 wr10, wr3, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
104 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
105 "walignr1 wr11, wr4, wr5 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
106 "wstrd wr9, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
107 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
108 "wstrd wr10, [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
109 "wstrd wr11, [r5, #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
110 "add r5, r5, %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
111 "bne 1b \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
112 : [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
113 :
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
114 : "memory", "r4", "r5", "r12");
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
115 }
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
116
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
117 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
118 {
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
119 int stride = line_size;
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
120 __asm__ __volatile__ (
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
121 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
122 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
123 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
124 "pld [%[block], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
125 "and r12, %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
126 "bic %[pixels], %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
127 "tmcr wcgr1, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
128 "add r4, %[pixels], %[line_size]\n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
129 "add r5, %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
130 "mov %[line_size], %[line_size], lsl #1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
131 "1: \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
132 "wldrd wr0, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
133 "wldrd wr1, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
134 "subs %[h], %[h], #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
135 "wldrd wr2, [%[pixels], #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
136 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
137 "wldrd wr3, [r4] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
138 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
139 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
140 "walignr1 wr8, wr0, wr1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
141 "wldrd wr4, [r4, #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
142 "walignr1 wr9, wr1, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
143 "wldrd wr5, [r4, #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
144 "add r4, r4, %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
145 "wldrd wr0, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
146 "pld [r4] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
147 "wldrd wr1, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
148 "pld [r4, #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
149 "wldrd wr2, [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
150 "walignr1 wr10, wr3, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
151 "wldrd wr3, [r5, #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
152 WAVG2B" wr8, wr8, wr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
153 WAVG2B" wr9, wr9, wr1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
154 WAVG2B" wr10, wr10, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
155 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
156 "walignr1 wr11, wr4, wr5 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
157 WAVG2B" wr11, wr11, wr3 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
158 "wstrd wr9, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
159 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
160 "wstrd wr10, [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
161 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
162 "pld [%[block], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
163 "wstrd wr11, [r5, #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
164 "add r5, r5, %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
165 "pld [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
166 "pld [r5, #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
167 "bne 1b \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
168 : [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
169 :
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
170 : "memory", "r4", "r5", "r12");
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
171 }
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
172
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
173 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
174 {
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
175 int stride = line_size;
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
176 // [wr0 wr1 wr2 wr3] for previous line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
177 // [wr4 wr5 wr6 wr7] for current line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
178 SET_RND(wr15); // =2 for rnd and =1 for no_rnd version
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
179 __asm__ __volatile__(
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
180 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
181 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
182 "and r12, %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
183 "bic %[pixels], %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
184 "tmcr wcgr1, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
185 "add r12, r12, #1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
186 "add r4, %[pixels], %[line_size]\n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
187 "tmcr wcgr2, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
188 "add r5, %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
189 "mov %[line_size], %[line_size], lsl #1 \n\t"
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 "1: \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
192 "wldrd wr10, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
193 "cmp r12, #8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
194 "wldrd wr11, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
195 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
196 "wldrd wr13, [r4] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
197 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
198 "wldrd wr14, [r4, #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
199 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
200 "add r4, r4, %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
201 "walignr1 wr0, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
202 "pld [r4] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
203 "pld [r4, #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
204 "walignr1 wr2, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
205 "wmoveq wr4, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
206 "wmoveq wr6, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
207 "walignr2ne wr4, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
208 "walignr2ne wr6, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
209 WAVG2B" wr0, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
210 WAVG2B" wr2, wr2, wr6 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
211 "wstrd wr0, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
212 "subs %[h], %[h], #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
213 "wstrd wr2, [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
214 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
215 "add r5, r5, %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
216 "bne 1b \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
217 : [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
218 :
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
219 : "r4", "r5", "r12", "memory");
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
220 }
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
221
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
222 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
223 {
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
224 int stride = line_size;
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
225 // [wr0 wr1 wr2 wr3] for previous line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
226 // [wr4 wr5 wr6 wr7] for current line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
227 SET_RND(wr15); // =2 for rnd and =1 for no_rnd version
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
228 __asm__ __volatile__(
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
229 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
230 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
231 "and r12, %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
232 "bic %[pixels], %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
233 "tmcr wcgr1, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
234 "add r12, r12, #1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
235 "add r4, %[pixels], %[line_size]\n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
236 "tmcr wcgr2, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
237 "add r5, %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
238 "mov %[line_size], %[line_size], lsl #1 \n\t"
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 "1: \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
241 "wldrd wr10, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
242 "cmp r12, #8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
243 "wldrd wr11, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
244 "wldrd wr12, [%[pixels], #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
245 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
246 "wldrd wr13, [r4] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
247 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
248 "wldrd wr14, [r4, #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
249 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
250 "wldrd wr15, [r4, #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
251 "add r4, r4, %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
252 "walignr1 wr0, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
253 "pld [r4] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
254 "pld [r4, #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
255 "walignr1 wr1, wr11, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
256 "walignr1 wr2, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
257 "walignr1 wr3, wr14, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
258 "wmoveq wr4, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
259 "wmoveq wr5, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
260 "wmoveq wr6, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
261 "wmoveq wr7, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
262 "walignr2ne wr4, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
263 "walignr2ne wr5, wr11, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
264 "walignr2ne wr6, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
265 "walignr2ne wr7, wr14, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
266 WAVG2B" wr0, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
267 WAVG2B" wr1, wr1, wr5 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
268 "wstrd wr0, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
269 WAVG2B" wr2, wr2, wr6 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
270 "wstrd wr1, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
271 WAVG2B" wr3, wr3, wr7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
272 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
273 "wstrd wr2, [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
274 "subs %[h], %[h], #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
275 "wstrd wr3, [r5, #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
276 "add r5, r5, %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
277 "bne 1b \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
278 : [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
279 :
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
280 : "r4", "r5", "r12", "memory");
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
281 }
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
282
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
283 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
284 {
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
285 int stride = line_size;
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
286 // [wr0 wr1 wr2 wr3] for previous line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
287 // [wr4 wr5 wr6 wr7] for current line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
288 SET_RND(wr15); // =2 for rnd and =1 for no_rnd version
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
289 __asm__ __volatile__(
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
290 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
291 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
292 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
293 "pld [%[block], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
294 "and r12, %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
295 "bic %[pixels], %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
296 "tmcr wcgr1, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
297 "add r12, r12, #1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
298 "add r4, %[pixels], %[line_size]\n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
299 "tmcr wcgr2, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
300 "add r5, %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
301 "mov %[line_size], %[line_size], lsl #1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
302 "pld [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
303 "pld [r5, #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
304
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
305 "1: \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
306 "wldrd wr10, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
307 "cmp r12, #8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
308 "wldrd wr11, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
309 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
310 "wldrd wr13, [r4] \n\t"
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 "wldrd wr14, [r4, #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
313 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
314 "add r4, r4, %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
315 "walignr1 wr0, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
316 "pld [r4] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
317 "pld [r4, #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
318 "walignr1 wr2, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
319 "wmoveq wr4, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
320 "wmoveq wr6, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
321 "walignr2ne wr4, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
322 "wldrd wr10, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
323 "walignr2ne wr6, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
324 "wldrd wr12, [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
325 WAVG2B" wr0, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
326 WAVG2B" wr2, wr2, wr6 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
327 WAVG2B" wr0, wr0, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
328 WAVG2B" wr2, wr2, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
329 "wstrd wr0, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
330 "subs %[h], %[h], #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
331 "wstrd wr2, [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
332 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
333 "add r5, r5, %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
334 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
335 "pld [%[block], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
336 "pld [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
337 "pld [r5, #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
338 "bne 1b \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
339 : [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
340 :
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
341 : "r4", "r5", "r12", "memory");
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
342 }
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
343
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
344 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
345 {
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
346 int stride = line_size;
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
347 // [wr0 wr1 wr2 wr3] for previous line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
348 // [wr4 wr5 wr6 wr7] for current line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
349 SET_RND(wr15); // =2 for rnd and =1 for no_rnd version
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
350 __asm__ __volatile__(
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
351 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
352 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
353 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
354 "pld [%[block], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
355 "and r12, %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
356 "bic %[pixels], %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
357 "tmcr wcgr1, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
358 "add r12, r12, #1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
359 "add r4, %[pixels], %[line_size]\n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
360 "tmcr wcgr2, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
361 "add r5, %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
362 "mov %[line_size], %[line_size], lsl #1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
363 "pld [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
364 "pld [r5, #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
365
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
366 "1: \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
367 "wldrd wr10, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
368 "cmp r12, #8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
369 "wldrd wr11, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
370 "wldrd wr12, [%[pixels], #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
371 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
372 "wldrd wr13, [r4] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
373 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
374 "wldrd wr14, [r4, #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
375 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
376 "wldrd wr15, [r4, #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
377 "add r4, r4, %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
378 "walignr1 wr0, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
379 "pld [r4] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
380 "pld [r4, #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
381 "walignr1 wr1, wr11, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
382 "walignr1 wr2, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
383 "walignr1 wr3, wr14, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
384 "wmoveq wr4, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
385 "wmoveq wr5, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
386 "wmoveq wr6, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
387 "wmoveq wr7, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
388 "walignr2ne wr4, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
389 "walignr2ne wr5, wr11, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
390 "walignr2ne wr6, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
391 "walignr2ne wr7, wr14, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
392 "wldrd wr10, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
393 WAVG2B" wr0, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
394 "wldrd wr11, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
395 WAVG2B" wr1, wr1, wr5 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
396 "wldrd wr12, [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
397 WAVG2B" wr2, wr2, wr6 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
398 "wldrd wr13, [r5, #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
399 WAVG2B" wr3, wr3, wr7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
400 WAVG2B" wr0, wr0, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
401 WAVG2B" wr1, wr1, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
402 WAVG2B" wr2, wr2, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
403 WAVG2B" wr3, wr3, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
404 "wstrd wr0, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
405 "subs %[h], %[h], #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
406 "wstrd wr1, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
407 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
408 "wstrd wr2, [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
409 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
410 "wstrd wr3, [r5, #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
411 "add r5, r5, %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
412 "pld [%[block], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
413 "pld [r5] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
414 "pld [r5, #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
415 "bne 1b \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
416 : [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
417 :
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
418 :"r4", "r5", "r12", "memory");
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
419 }
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
420
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
421 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
422 {
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
423 int stride = line_size;
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
424 // [wr0 wr1 wr2 wr3] for previous line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
425 // [wr4 wr5 wr6 wr7] for current line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
426 __asm__ __volatile__(
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
427 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
428 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
429 "and r12, %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
430 "tmcr wcgr1, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
431 "bic %[pixels], %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
432
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
433 "wldrd wr10, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
434 "wldrd wr11, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
435 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
436 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
437 "walignr1 wr0, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
438 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
439 "pld [%[pixels], #32] \n\t"
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 "1: \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
442 "wldrd wr10, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
443 "wldrd wr11, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
444 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
445 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
446 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
447 "walignr1 wr4, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
448 "wldrd wr10, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
449 WAVG2B" wr8, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
450 WAVG2B" wr8, wr8, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
451 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
452 "add %[block], %[block], %[line_size] \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 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
459 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
460 "walignr1 wr0, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
461 "wldrd wr10, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
462 WAVG2B" wr8, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
463 WAVG2B" wr8, wr8, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
464 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
465 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
466
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
467 "subs %[h], %[h], #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
468 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
469 "bne 1b \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
470 : [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
471 :
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
472 : "cc", "memory", "r12");
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
473 }
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 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
476 {
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
477 int stride = line_size;
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
478 // [wr0 wr1 wr2 wr3] for previous line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
479 // [wr4 wr5 wr6 wr7] for current line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
480 __asm__ __volatile__(
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
481 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
482 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
483 "and r12, %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
484 "tmcr wcgr1, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
485 "bic %[pixels], %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
486
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
487 "wldrd wr10, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
488 "wldrd wr11, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
489 "wldrd wr12, [%[pixels], #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
490 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
491 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
492 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
493 "walignr1 wr0, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
494 "walignr1 wr1, wr11, wr12 \n\t"
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 "1: \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
497 "wldrd wr10, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
498 "wldrd wr11, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
499 "wldrd wr12, [%[pixels], #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
500 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
501 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
502 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
503 "walignr1 wr4, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
504 "walignr1 wr5, wr11, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
505 WAVG2B" wr8, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
506 WAVG2B" wr9, wr1, wr5 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
507 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
508 "wstrd wr9, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
509 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
510
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
511 "wldrd wr10, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
512 "wldrd wr11, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
513 "wldrd wr12, [%[pixels], #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
514 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
515 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
516 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
517 "walignr1 wr0, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
518 "walignr1 wr1, wr11, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
519 WAVG2B" wr8, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
520 WAVG2B" wr9, wr1, wr5 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
521 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
522 "wstrd wr9, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
523 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
524
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
525 "subs %[h], %[h], #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
526 "bne 1b \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
527 : [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
528 :
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
529 : "r4", "r5", "r12", "memory");
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
530 }
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 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
533 {
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
534 int stride = line_size;
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
535 // [wr0 wr1 wr2 wr3] for previous line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
536 // [wr4 wr5 wr6 wr7] for current line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
537 __asm__ __volatile__(
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
538 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
539 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
540 "and r12, %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
541 "tmcr wcgr1, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
542 "bic %[pixels], %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
543
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
544 "wldrd wr10, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
545 "wldrd wr11, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
546 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
547 "wldrd wr12, [%[pixels], #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
548 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
549 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
550 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
551 "walignr1 wr0, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
552 "walignr1 wr1, wr11, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
553
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
554 "1: \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
555 "wldrd wr10, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
556 "wldrd wr11, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
557 "wldrd wr12, [%[pixels], #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
558 "add %[pixels], %[pixels], %[line_size] \n\t"
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 "walignr1 wr4, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
562 "walignr1 wr5, wr11, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
563 "wldrd wr10, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
564 "wldrd wr11, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
565 WAVG2B" wr8, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
566 WAVG2B" wr9, wr1, wr5 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
567 WAVG2B" wr8, wr8, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
568 WAVG2B" wr9, wr9, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
569 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
570 "wstrd wr9, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
571 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
572
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
573 "wldrd wr10, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
574 "wldrd wr11, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
575 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
576 "wldrd wr12, [%[pixels], #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
577 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
578 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
579 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
580 "walignr1 wr0, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
581 "walignr1 wr1, wr11, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
582 "wldrd wr10, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
583 "wldrd wr11, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
584 WAVG2B" wr8, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
585 WAVG2B" wr9, wr1, wr5 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
586 WAVG2B" wr8, wr8, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
587 WAVG2B" wr9, wr9, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
588 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
589 "wstrd wr9, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
590 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
591
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
592 "subs %[h], %[h], #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
593 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
594 "bne 1b \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
595 : [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
596 :
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
597 : "r4", "r5", "r12", "memory");
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
598 }
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
599
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
600 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
601 {
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
602 // [wr0 wr1 wr2 wr3] for previous line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
603 // [wr4 wr5 wr6 wr7] for current line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
604 SET_RND(wr15); // =2 for rnd and =1 for no_rnd version
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
605 __asm__ __volatile__(
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
606 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
607 "mov r12, #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
608 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
609 "tmcr wcgr0, r12 \n\t" /* for shift value */
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
610 "and r12, %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
611 "bic %[pixels], %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
612 "tmcr wcgr1, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
613
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
614 // [wr0 wr1 wr2 wr3] <= *
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
615 // [wr4 wr5 wr6 wr7]
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
616 "wldrd wr12, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
617 "add r12, r12, #1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
618 "wldrd wr13, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
619 "tmcr wcgr2, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
620 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
621 "cmp r12, #8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
622 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
623 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
624 "walignr1 wr2, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
625 "wmoveq wr10, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
626 "walignr2ne wr10, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
627 "wunpckelub wr0, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
628 "wunpckehub wr1, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
629 "wunpckelub wr8, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
630 "wunpckehub wr9, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
631 "waddhus wr0, wr0, wr8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
632 "waddhus wr1, wr1, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
633
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
634 "1: \n\t"
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 "cmp r12, #8 \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 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
641 "walignr1 wr6, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
642 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
643 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
644 "wmoveq wr10, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
645 "walignr2ne wr10, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
646 "wunpckelub wr4, wr6 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
647 "wunpckehub wr5, wr6 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
648 "wunpckelub wr8, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
649 "wunpckehub wr9, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
650 "waddhus wr4, wr4, wr8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
651 "waddhus wr5, wr5, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
652 "waddhus wr8, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
653 "waddhus wr9, wr1, wr5 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
654 "waddhus wr8, wr8, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
655 "waddhus wr9, wr9, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
656 "wsrlhg wr8, wr8, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
657 "wsrlhg wr9, wr9, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
658 "wpackhus wr8, wr8, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
659 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
660 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
661
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
662 // [wr0 wr1 wr2 wr3] <= *
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
663 // [wr4 wr5 wr6 wr7]
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
664 "wldrd wr12, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
665 "wldrd wr13, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
666 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
667 "walignr1 wr2, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
668 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
669 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
670 "wmoveq wr10, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
671 "walignr2ne wr10, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
672 "wunpckelub wr0, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
673 "wunpckehub wr1, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
674 "wunpckelub wr8, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
675 "wunpckehub wr9, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
676 "waddhus wr0, wr0, wr8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
677 "waddhus wr1, wr1, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
678 "waddhus wr8, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
679 "waddhus wr9, wr1, wr5 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
680 "waddhus wr8, wr8, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
681 "waddhus wr9, wr9, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
682 "wsrlhg wr8, wr8, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
683 "wsrlhg wr9, wr9, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
684 "wpackhus wr8, wr8, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
685 "subs %[h], %[h], #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
686 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
687 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
688 "bne 1b \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
689 : [h]"+r"(h), [pixels]"+r"(pixels), [block]"+r"(block)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
690 : [line_size]"r"(line_size)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
691 : "r12", "memory");
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
692 }
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
693
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
694 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
695 {
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
696 // [wr0 wr1 wr2 wr3] for previous line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
697 // [wr4 wr5 wr6 wr7] for current line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
698 SET_RND(wr15); // =2 for rnd and =1 for no_rnd version
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
699 __asm__ __volatile__(
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
700 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
701 "mov r12, #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
702 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
703 "tmcr wcgr0, r12 \n\t" /* for shift value */
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
704 /* alignment */
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
705 "and r12, %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
706 "bic %[pixels], %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
707 "tmcr wcgr1, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
708 "add r12, r12, #1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
709 "tmcr wcgr2, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
710
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
711 // [wr0 wr1 wr2 wr3] <= *
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
712 // [wr4 wr5 wr6 wr7]
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
713 "wldrd wr12, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
714 "cmp r12, #8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
715 "wldrd wr13, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
716 "wldrd wr14, [%[pixels], #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
717 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
718 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
719 "walignr1 wr2, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
720 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
721 "walignr1 wr3, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
722 "wmoveq wr10, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
723 "wmoveq wr11, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
724 "walignr2ne wr10, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
725 "walignr2ne wr11, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
726 "wunpckelub wr0, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
727 "wunpckehub wr1, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
728 "wunpckelub wr2, wr3 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
729 "wunpckehub wr3, wr3 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
730 "wunpckelub wr8, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
731 "wunpckehub wr9, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
732 "wunpckelub wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
733 "wunpckehub wr11, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
734 "waddhus wr0, wr0, wr8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
735 "waddhus wr1, wr1, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
736 "waddhus wr2, wr2, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
737 "waddhus wr3, wr3, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
738
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
739 "1: \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
740 // [wr0 wr1 wr2 wr3]
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
741 // [wr4 wr5 wr6 wr7] <= *
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
742 "wldrd wr12, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
743 "cmp r12, #8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
744 "wldrd wr13, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
745 "wldrd wr14, [%[pixels], #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
746 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
747 "walignr1 wr6, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
748 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
749 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
750 "walignr1 wr7, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
751 "wmoveq wr10, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
752 "wmoveq wr11, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
753 "walignr2ne wr10, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
754 "walignr2ne wr11, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
755 "wunpckelub wr4, wr6 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
756 "wunpckehub wr5, wr6 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
757 "wunpckelub wr6, wr7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
758 "wunpckehub wr7, wr7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
759 "wunpckelub wr8, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
760 "wunpckehub wr9, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
761 "wunpckelub wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
762 "wunpckehub wr11, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
763 "waddhus wr4, wr4, wr8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
764 "waddhus wr5, wr5, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
765 "waddhus wr6, wr6, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
766 "waddhus wr7, wr7, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
767 "waddhus wr8, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
768 "waddhus wr9, wr1, wr5 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
769 "waddhus wr10, wr2, wr6 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
770 "waddhus wr11, wr3, wr7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
771 "waddhus wr8, wr8, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
772 "waddhus wr9, wr9, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
773 "waddhus wr10, wr10, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
774 "waddhus wr11, wr11, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
775 "wsrlhg wr8, wr8, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
776 "wsrlhg wr9, wr9, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
777 "wsrlhg wr10, wr10, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
778 "wsrlhg wr11, wr11, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
779 "wpackhus wr8, wr8, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
780 "wpackhus wr9, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
781 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
782 "wstrd wr9, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
783 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
784
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
785 // [wr0 wr1 wr2 wr3] <= *
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
786 // [wr4 wr5 wr6 wr7]
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
787 "wldrd wr12, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
788 "wldrd wr13, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
789 "wldrd wr14, [%[pixels], #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
790 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
791 "walignr1 wr2, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
792 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
793 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
794 "walignr1 wr3, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
795 "wmoveq wr10, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
796 "wmoveq wr11, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
797 "walignr2ne wr10, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
798 "walignr2ne wr11, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
799 "wunpckelub wr0, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
800 "wunpckehub wr1, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
801 "wunpckelub wr2, wr3 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
802 "wunpckehub wr3, wr3 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
803 "wunpckelub wr8, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
804 "wunpckehub wr9, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
805 "wunpckelub wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
806 "wunpckehub wr11, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
807 "waddhus wr0, wr0, wr8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
808 "waddhus wr1, wr1, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
809 "waddhus wr2, wr2, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
810 "waddhus wr3, wr3, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
811 "waddhus wr8, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
812 "waddhus wr9, wr1, wr5 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
813 "waddhus wr10, wr2, wr6 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
814 "waddhus wr11, wr3, wr7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
815 "waddhus wr8, wr8, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
816 "waddhus wr9, wr9, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
817 "waddhus wr10, wr10, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
818 "waddhus wr11, wr11, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
819 "wsrlhg wr8, wr8, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
820 "wsrlhg wr9, wr9, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
821 "wsrlhg wr10, wr10, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
822 "wsrlhg wr11, wr11, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
823 "wpackhus wr8, wr8, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
824 "wpackhus wr9, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
825 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
826 "wstrd wr9, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
827 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
828
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
829 "subs %[h], %[h], #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
830 "bne 1b \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
831 : [h]"+r"(h), [pixels]"+r"(pixels), [block]"+r"(block)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
832 : [line_size]"r"(line_size)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
833 : "r12", "memory");
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
834 }
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
835
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
836 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
837 {
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
838 // [wr0 wr1 wr2 wr3] for previous line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
839 // [wr4 wr5 wr6 wr7] for current line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
840 SET_RND(wr15); // =2 for rnd and =1 for no_rnd version
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
841 __asm__ __volatile__(
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
842 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
843 "pld [%[block], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
844 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
845 "mov r12, #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
846 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
847 "tmcr wcgr0, r12 \n\t" /* for shift value */
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
848 "and r12, %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
849 "bic %[pixels], %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
850 "tmcr wcgr1, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
851
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
852 // [wr0 wr1 wr2 wr3] <= *
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
853 // [wr4 wr5 wr6 wr7]
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
854 "wldrd wr12, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
855 "add r12, r12, #1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
856 "wldrd wr13, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
857 "tmcr wcgr2, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
858 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
859 "cmp r12, #8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
860 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
861 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
862 "walignr1 wr2, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
863 "wmoveq wr10, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
864 "walignr2ne wr10, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
865 "wunpckelub wr0, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
866 "wunpckehub wr1, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
867 "wunpckelub wr8, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
868 "wunpckehub wr9, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
869 "waddhus wr0, wr0, wr8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
870 "waddhus wr1, wr1, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
871
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
872 "1: \n\t"
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 "cmp r12, #8 \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 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
879 "walignr1 wr6, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
880 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
881 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
882 "wmoveq wr10, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
883 "walignr2ne wr10, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
884 "wunpckelub wr4, wr6 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
885 "wunpckehub wr5, wr6 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
886 "wunpckelub wr8, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
887 "wunpckehub wr9, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
888 "waddhus wr4, wr4, wr8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
889 "waddhus wr5, wr5, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
890 "waddhus wr8, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
891 "waddhus wr9, wr1, wr5 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
892 "waddhus wr8, wr8, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
893 "waddhus wr9, wr9, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
894 "wldrd wr12, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
895 "wsrlhg wr8, wr8, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
896 "wsrlhg wr9, wr9, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
897 "wpackhus wr8, wr8, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
898 WAVG2B" wr8, wr8, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
899 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
900 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
901 "wldrd wr12, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
902 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
903 "pld [%[block], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
904
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
905 // [wr0 wr1 wr2 wr3] <= *
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
906 // [wr4 wr5 wr6 wr7]
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
907 "wldrd wr13, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
908 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
909 "walignr1 wr2, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
910 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
911 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
912 "wmoveq wr10, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
913 "walignr2ne wr10, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
914 "wunpckelub wr0, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
915 "wunpckehub wr1, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
916 "wunpckelub wr8, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
917 "wunpckehub wr9, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
918 "waddhus wr0, wr0, wr8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
919 "waddhus wr1, wr1, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
920 "waddhus wr8, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
921 "waddhus wr9, wr1, wr5 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
922 "waddhus wr8, wr8, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
923 "waddhus wr9, wr9, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
924 "wldrd wr12, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
925 "wsrlhg wr8, wr8, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
926 "wsrlhg wr9, wr9, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
927 "wpackhus wr8, wr8, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
928 "subs %[h], %[h], #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
929 WAVG2B" wr8, wr8, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
930 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
931 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
932 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
933 "pld [%[block], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
934 "bne 1b \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
935 : [h]"+r"(h), [pixels]"+r"(pixels), [block]"+r"(block)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
936 : [line_size]"r"(line_size)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
937 : "r12", "memory");
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
938 }
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
939
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
940 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
941 {
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
942 // [wr0 wr1 wr2 wr3] for previous line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
943 // [wr4 wr5 wr6 wr7] for current line
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
944 SET_RND(wr15); // =2 for rnd and =1 for no_rnd version
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
945 __asm__ __volatile__(
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
946 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
947 "pld [%[block], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
948 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
949 "mov r12, #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
950 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
951 "tmcr wcgr0, r12 \n\t" /* for shift value */
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
952 /* alignment */
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
953 "and r12, %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
954 "bic %[pixels], %[pixels], #7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
955 "tmcr wcgr1, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
956 "add r12, r12, #1 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
957 "tmcr wcgr2, r12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
958
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
959 // [wr0 wr1 wr2 wr3] <= *
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
960 // [wr4 wr5 wr6 wr7]
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
961 "wldrd wr12, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
962 "cmp r12, #8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
963 "wldrd wr13, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
964 "wldrd wr14, [%[pixels], #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
965 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
966 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
967 "walignr1 wr2, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
968 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
969 "walignr1 wr3, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
970 "wmoveq wr10, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
971 "wmoveq wr11, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
972 "walignr2ne wr10, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
973 "walignr2ne wr11, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
974 "wunpckelub wr0, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
975 "wunpckehub wr1, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
976 "wunpckelub wr2, wr3 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
977 "wunpckehub wr3, wr3 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
978 "wunpckelub wr8, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
979 "wunpckehub wr9, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
980 "wunpckelub wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
981 "wunpckehub wr11, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
982 "waddhus wr0, wr0, wr8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
983 "waddhus wr1, wr1, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
984 "waddhus wr2, wr2, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
985 "waddhus wr3, wr3, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
986
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
987 "1: \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
988 // [wr0 wr1 wr2 wr3]
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
989 // [wr4 wr5 wr6 wr7] <= *
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
990 "wldrd wr12, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
991 "cmp r12, #8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
992 "wldrd wr13, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
993 "wldrd wr14, [%[pixels], #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
994 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
995 "walignr1 wr6, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
996 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
997 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
998 "walignr1 wr7, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
999 "wmoveq wr10, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1000 "wmoveq wr11, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1001 "walignr2ne wr10, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1002 "walignr2ne wr11, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1003 "wunpckelub wr4, wr6 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1004 "wunpckehub wr5, wr6 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1005 "wunpckelub wr6, wr7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1006 "wunpckehub wr7, wr7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1007 "wunpckelub wr8, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1008 "wunpckehub wr9, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1009 "wunpckelub wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1010 "wunpckehub wr11, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1011 "waddhus wr4, wr4, wr8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1012 "waddhus wr5, wr5, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1013 "waddhus wr6, wr6, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1014 "waddhus wr7, wr7, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1015 "waddhus wr8, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1016 "waddhus wr9, wr1, wr5 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1017 "waddhus wr10, wr2, wr6 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1018 "waddhus wr11, wr3, wr7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1019 "waddhus wr8, wr8, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1020 "waddhus wr9, wr9, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1021 "waddhus wr10, wr10, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1022 "waddhus wr11, wr11, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1023 "wsrlhg wr8, wr8, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1024 "wsrlhg wr9, wr9, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1025 "wldrd wr12, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1026 "wldrd wr13, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1027 "wsrlhg wr10, wr10, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1028 "wsrlhg wr11, wr11, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1029 "wpackhus wr8, wr8, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1030 "wpackhus wr9, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1031 WAVG2B" wr8, wr8, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1032 WAVG2B" wr9, wr9, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1033 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1034 "wstrd wr9, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1035 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1036
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1037 // [wr0 wr1 wr2 wr3] <= *
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1038 // [wr4 wr5 wr6 wr7]
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1039 "wldrd wr12, [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1040 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1041 "wldrd wr13, [%[pixels], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1042 "pld [%[block], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1043 "wldrd wr14, [%[pixels], #16] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1044 "add %[pixels], %[pixels], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1045 "walignr1 wr2, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1046 "pld [%[pixels]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1047 "pld [%[pixels], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1048 "walignr1 wr3, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1049 "wmoveq wr10, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1050 "wmoveq wr11, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1051 "walignr2ne wr10, wr12, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1052 "walignr2ne wr11, wr13, wr14 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1053 "wunpckelub wr0, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1054 "wunpckehub wr1, wr2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1055 "wunpckelub wr2, wr3 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1056 "wunpckehub wr3, wr3 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1057 "wunpckelub wr8, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1058 "wunpckehub wr9, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1059 "wunpckelub wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1060 "wunpckehub wr11, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1061 "waddhus wr0, wr0, wr8 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1062 "waddhus wr1, wr1, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1063 "waddhus wr2, wr2, wr10 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1064 "waddhus wr3, wr3, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1065 "waddhus wr8, wr0, wr4 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1066 "waddhus wr9, wr1, wr5 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1067 "waddhus wr10, wr2, wr6 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1068 "waddhus wr11, wr3, wr7 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1069 "waddhus wr8, wr8, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1070 "waddhus wr9, wr9, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1071 "waddhus wr10, wr10, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1072 "waddhus wr11, wr11, wr15 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1073 "wsrlhg wr8, wr8, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1074 "wsrlhg wr9, wr9, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1075 "wldrd wr12, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1076 "wldrd wr13, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1077 "wsrlhg wr10, wr10, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1078 "wsrlhg wr11, wr11, wcgr0 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1079 "wpackhus wr8, wr8, wr9 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1080 "wpackhus wr9, wr10, wr11 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1081 WAVG2B" wr8, wr8, wr12 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1082 WAVG2B" wr9, wr9, wr13 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1083 "wstrd wr8, [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1084 "wstrd wr9, [%[block], #8] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1085 "add %[block], %[block], %[line_size] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1086 "subs %[h], %[h], #2 \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1087 "pld [%[block]] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1088 "pld [%[block], #32] \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1089 "bne 1b \n\t"
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1090 : [h]"+r"(h), [pixels]"+r"(pixels), [block]"+r"(block)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1091 : [line_size]"r"(line_size)
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1092 : "r12", "memory");
aeea63c97878 Better ARM support for mplayer/ffmpeg, ported from atty fork
michael
parents:
diff changeset
1093 }