annotate alpha/motion_est_alpha.c @ 12201:c4b53914f286 libavcodec

vp8: add do { } while(0) around XCHG() macro to avoid confusing if/else This is the correct solution to the warning "fixed" in the previous commit.
author mru
date Tue, 20 Jul 2010 17:54:25 +0000
parents 9e7d38743146
children
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
586
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
1 /*
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
2 * Alpha optimized DSP utils
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
3 * Copyright (c) 2002 Falk Hueffner <falk@debian.org>
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
4 *
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3036
diff changeset
5 * This file is part of FFmpeg.
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3036
diff changeset
6 *
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3036
diff changeset
7 * FFmpeg is free software; you can redistribute it and/or
586
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
8 * modify it under the terms of the GNU Lesser General Public
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
9 * License as published by the Free Software Foundation; either
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3036
diff changeset
10 * version 2.1 of the License, or (at your option) any later version.
586
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
11 *
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3036
diff changeset
12 * FFmpeg is distributed in the hope that it will be useful,
586
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
15 * Lesser General Public License for more details.
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
16 *
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
17 * You should have received a copy of the GNU Lesser General Public
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3036
diff changeset
18 * License along with FFmpeg; if not, write to the Free Software
3036
0b546eab515d Update licensing information: The FSF changed postal address.
diego
parents: 2979
diff changeset
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
586
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
20 */
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
21
6763
f7cbb7733146 Use full path for #includes from another directory.
diego
parents: 5010
diff changeset
22 #include "libavcodec/dsputil.h"
11396
9e7d38743146 Alpha: move dsputil prototypes to a header file
mru
parents: 6763
diff changeset
23 #include "dsputil_alpha.h"
586
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
24 #include "asm.h"
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
25
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
26 void get_pixels_mvi(DCTELEM *restrict block,
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
27 const uint8_t *restrict pixels, int line_size)
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
28 {
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
29 int h = 8;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
30
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
31 do {
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
32 uint64_t p;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
33
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
34 p = ldq(pixels);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
35 stq(unpkbw(p), block);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 1708
diff changeset
36 stq(unpkbw(p >> 32), block + 4);
586
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
37
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
38 pixels += line_size;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
39 block += 8;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
40 } while (--h);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
41 }
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
42
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
43 void diff_pixels_mvi(DCTELEM *block, const uint8_t *s1, const uint8_t *s2,
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
44 int stride) {
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
45 int h = 8;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
46 uint64_t mask = 0x4040;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
47
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
48 mask |= mask << 16;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
49 mask |= mask << 32;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
50 do {
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
51 uint64_t x, y, c, d, a;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
52 uint64_t signs;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
53
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
54 x = ldq(s1);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
55 y = ldq(s2);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
56 c = cmpbge(x, y);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
57 d = x - y;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
58 a = zap(mask, c); /* We use 0x4040404040404040 here... */
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
59 d += 4 * a; /* ...so we can use s4addq here. */
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
60 signs = zap(-1, c);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
61
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
62 stq(unpkbw(d) | (unpkbw(signs) << 8), block);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
63 stq(unpkbw(d >> 32) | (unpkbw(signs >> 32) << 8), block + 4);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
64
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
65 s1 += stride;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
66 s2 += stride;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
67 block += 8;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
68 } while (--h);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
69 }
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
70
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
71 static inline uint64_t avg2(uint64_t a, uint64_t b)
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
72 {
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
73 return (a | b) - (((a ^ b) & BYTE_VEC(0xfe)) >> 1);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
74 }
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
75
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
76 static inline uint64_t avg4(uint64_t l1, uint64_t l2, uint64_t l3, uint64_t l4)
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
77 {
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
78 uint64_t r1 = ((l1 & ~BYTE_VEC(0x03)) >> 2)
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
79 + ((l2 & ~BYTE_VEC(0x03)) >> 2)
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
80 + ((l3 & ~BYTE_VEC(0x03)) >> 2)
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
81 + ((l4 & ~BYTE_VEC(0x03)) >> 2);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
82 uint64_t r2 = (( (l1 & BYTE_VEC(0x03))
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
83 + (l2 & BYTE_VEC(0x03))
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
84 + (l3 & BYTE_VEC(0x03))
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
85 + (l4 & BYTE_VEC(0x03))
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
86 + BYTE_VEC(0x02)) >> 2) & BYTE_VEC(0x03);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
87 return r1 + r2;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
88 }
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
89
1708
dea5b2946999 interlaced motion estimation
michael
parents: 705
diff changeset
90 int pix_abs8x8_mvi(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
586
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
91 {
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
92 int result = 0;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
93
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
94 if ((size_t) pix2 & 0x7) {
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
95 /* works only when pix2 is actually unaligned */
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
96 do { /* do 8 pixel a time */
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
97 uint64_t p1, p2;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
98
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
99 p1 = ldq(pix1);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
100 p2 = uldq(pix2);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
101 result += perr(p1, p2);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
102
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
103 pix1 += line_size;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
104 pix2 += line_size;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
105 } while (--h);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
106 } else {
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
107 do {
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
108 uint64_t p1, p2;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
109
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
110 p1 = ldq(pix1);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
111 p2 = ldq(pix2);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
112 result += perr(p1, p2);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
113
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
114 pix1 += line_size;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
115 pix2 += line_size;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
116 } while (--h);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
117 }
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
118
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
119 return result;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
120 }
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
121
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
122 #if 0 /* now done in assembly */
586
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
123 int pix_abs16x16_mvi(uint8_t *pix1, uint8_t *pix2, int line_size)
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
124 {
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
125 int result = 0;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
126 int h = 16;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
127
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
128 if ((size_t) pix2 & 0x7) {
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
129 /* works only when pix2 is actually unaligned */
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
130 do { /* do 16 pixel a time */
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
131 uint64_t p1_l, p1_r, p2_l, p2_r;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
132 uint64_t t;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
133
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
134 p1_l = ldq(pix1);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
135 p1_r = ldq(pix1 + 8);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
136 t = ldq_u(pix2 + 8);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
137 p2_l = extql(ldq_u(pix2), pix2) | extqh(t, pix2);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
138 p2_r = extql(t, pix2) | extqh(ldq_u(pix2 + 16), pix2);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
139 pix1 += line_size;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
140 pix2 += line_size;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
141
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
142 result += perr(p1_l, p2_l)
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
143 + perr(p1_r, p2_r);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
144 } while (--h);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
145 } else {
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
146 do {
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
147 uint64_t p1_l, p1_r, p2_l, p2_r;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
148
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
149 p1_l = ldq(pix1);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
150 p1_r = ldq(pix1 + 8);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
151 p2_l = ldq(pix2);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
152 p2_r = ldq(pix2 + 8);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
153 pix1 += line_size;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
154 pix2 += line_size;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
155
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
156 result += perr(p1_l, p2_l)
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
157 + perr(p1_r, p2_r);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
158 } while (--h);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
159 }
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
160
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
161 return result;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
162 }
705
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents: 586
diff changeset
163 #endif
586
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
164
1708
dea5b2946999 interlaced motion estimation
michael
parents: 705
diff changeset
165 int pix_abs16x16_x2_mvi(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
586
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
166 {
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
167 int result = 0;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
168 uint64_t disalign = (size_t) pix2 & 0x7;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
169
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
170 switch (disalign) {
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
171 case 0:
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
172 do {
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
173 uint64_t p1_l, p1_r, p2_l, p2_r;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
174 uint64_t l, r;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
175
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
176 p1_l = ldq(pix1);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
177 p1_r = ldq(pix1 + 8);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
178 l = ldq(pix2);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
179 r = ldq(pix2 + 8);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
180 p2_l = avg2(l, (l >> 8) | ((uint64_t) r << 56));
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
181 p2_r = avg2(r, (r >> 8) | ((uint64_t) pix2[16] << 56));
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
182 pix1 += line_size;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
183 pix2 += line_size;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
184
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
185 result += perr(p1_l, p2_l)
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
186 + perr(p1_r, p2_r);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
187 } while (--h);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
188 break;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
189 case 7:
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
190 /* |.......l|lllllllr|rrrrrrr*|
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
191 This case is special because disalign1 would be 8, which
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
192 gets treated as 0 by extqh. At least it is a bit faster
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 1708
diff changeset
193 that way :) */
586
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
194 do {
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
195 uint64_t p1_l, p1_r, p2_l, p2_r;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
196 uint64_t l, m, r;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
197
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
198 p1_l = ldq(pix1);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
199 p1_r = ldq(pix1 + 8);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
200 l = ldq_u(pix2);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
201 m = ldq_u(pix2 + 8);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
202 r = ldq_u(pix2 + 16);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
203 p2_l = avg2(extql(l, disalign) | extqh(m, disalign), m);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
204 p2_r = avg2(extql(m, disalign) | extqh(r, disalign), r);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
205 pix1 += line_size;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
206 pix2 += line_size;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 1708
diff changeset
207
586
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
208 result += perr(p1_l, p2_l)
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
209 + perr(p1_r, p2_r);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
210 } while (--h);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
211 break;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
212 default:
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
213 do {
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
214 uint64_t disalign1 = disalign + 1;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
215 uint64_t p1_l, p1_r, p2_l, p2_r;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
216 uint64_t l, m, r;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
217
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
218 p1_l = ldq(pix1);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
219 p1_r = ldq(pix1 + 8);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
220 l = ldq_u(pix2);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
221 m = ldq_u(pix2 + 8);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
222 r = ldq_u(pix2 + 16);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
223 p2_l = avg2(extql(l, disalign) | extqh(m, disalign),
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
224 extql(l, disalign1) | extqh(m, disalign1));
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
225 p2_r = avg2(extql(m, disalign) | extqh(r, disalign),
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
226 extql(m, disalign1) | extqh(r, disalign1));
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
227 pix1 += line_size;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
228 pix2 += line_size;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
229
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
230 result += perr(p1_l, p2_l)
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
231 + perr(p1_r, p2_r);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
232 } while (--h);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
233 break;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
234 }
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
235 return result;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
236 }
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
237
1708
dea5b2946999 interlaced motion estimation
michael
parents: 705
diff changeset
238 int pix_abs16x16_y2_mvi(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
586
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
239 {
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
240 int result = 0;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
241
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
242 if ((size_t) pix2 & 0x7) {
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
243 uint64_t t, p2_l, p2_r;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
244 t = ldq_u(pix2 + 8);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
245 p2_l = extql(ldq_u(pix2), pix2) | extqh(t, pix2);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
246 p2_r = extql(t, pix2) | extqh(ldq_u(pix2 + 16), pix2);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
247
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
248 do {
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
249 uint64_t p1_l, p1_r, np2_l, np2_r;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
250 uint64_t t;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
251
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
252 p1_l = ldq(pix1);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
253 p1_r = ldq(pix1 + 8);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
254 pix2 += line_size;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
255 t = ldq_u(pix2 + 8);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
256 np2_l = extql(ldq_u(pix2), pix2) | extqh(t, pix2);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
257 np2_r = extql(t, pix2) | extqh(ldq_u(pix2 + 16), pix2);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
258
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
259 result += perr(p1_l, avg2(p2_l, np2_l))
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
260 + perr(p1_r, avg2(p2_r, np2_r));
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
261
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
262 pix1 += line_size;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
263 p2_l = np2_l;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
264 p2_r = np2_r;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
265
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
266 } while (--h);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
267 } else {
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
268 uint64_t p2_l, p2_r;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
269 p2_l = ldq(pix2);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
270 p2_r = ldq(pix2 + 8);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
271 do {
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
272 uint64_t p1_l, p1_r, np2_l, np2_r;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
273
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
274 p1_l = ldq(pix1);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
275 p1_r = ldq(pix1 + 8);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
276 pix2 += line_size;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
277 np2_l = ldq(pix2);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
278 np2_r = ldq(pix2 + 8);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
279
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
280 result += perr(p1_l, avg2(p2_l, np2_l))
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
281 + perr(p1_r, avg2(p2_r, np2_r));
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
282
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
283 pix1 += line_size;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
284 p2_l = np2_l;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
285 p2_r = np2_r;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
286 } while (--h);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
287 }
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
288 return result;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
289 }
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
290
1708
dea5b2946999 interlaced motion estimation
michael
parents: 705
diff changeset
291 int pix_abs16x16_xy2_mvi(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
586
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
292 {
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
293 int result = 0;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 1708
diff changeset
294
586
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
295 uint64_t p1_l, p1_r;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
296 uint64_t p2_l, p2_r, p2_x;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
297
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
298 p1_l = ldq(pix1);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
299 p1_r = ldq(pix1 + 8);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
300
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
301 if ((size_t) pix2 & 0x7) { /* could be optimized a lot */
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
302 p2_l = uldq(pix2);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
303 p2_r = uldq(pix2 + 8);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
304 p2_x = (uint64_t) pix2[16] << 56;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
305 } else {
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
306 p2_l = ldq(pix2);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
307 p2_r = ldq(pix2 + 8);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
308 p2_x = ldq(pix2 + 16) << 56;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
309 }
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
310
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
311 do {
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
312 uint64_t np1_l, np1_r;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
313 uint64_t np2_l, np2_r, np2_x;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
314
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
315 pix1 += line_size;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
316 pix2 += line_size;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
317
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
318 np1_l = ldq(pix1);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
319 np1_r = ldq(pix1 + 8);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
320
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
321 if ((size_t) pix2 & 0x7) { /* could be optimized a lot */
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
322 np2_l = uldq(pix2);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
323 np2_r = uldq(pix2 + 8);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
324 np2_x = (uint64_t) pix2[16] << 56;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
325 } else {
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
326 np2_l = ldq(pix2);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
327 np2_r = ldq(pix2 + 8);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
328 np2_x = ldq(pix2 + 16) << 56;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
329 }
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
330
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
331 result += perr(p1_l,
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
332 avg4( p2_l, ( p2_l >> 8) | ((uint64_t) p2_r << 56),
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
333 np2_l, (np2_l >> 8) | ((uint64_t) np2_r << 56)))
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
334 + perr(p1_r,
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
335 avg4( p2_r, ( p2_r >> 8) | ((uint64_t) p2_x),
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
336 np2_r, (np2_r >> 8) | ((uint64_t) np2_x)));
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
337
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
338 p1_l = np1_l;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
339 p1_r = np1_r;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
340 p2_l = np2_l;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
341 p2_r = np2_r;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
342 p2_x = np2_x;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
343 } while (--h);
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
344
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
345 return result;
54b1c94977d5 MVI optimizations for motion estimation.
mellum
parents:
diff changeset
346 }