annotate ppc/dsputil_altivec.c @ 12530:63edd10ad4bc libavcodec tip

Try to fix crashes introduced by r25218 r25218 made assumptions about the existence of past reference frames that weren't necessarily true.
author darkshikari
date Tue, 28 Sep 2010 09:06:22 +0000
parents 3cd4cd0509cd
children
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
1 /*
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
2 * Copyright (c) 2002 Brian Foley
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
3 * Copyright (c) 2002 Dieter Shirley
1949
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
4 * Copyright (c) 2003-2004 Romain Dolbeau <romain@dolbeau.org>
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
5 *
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3572
diff changeset
6 * This file is part of FFmpeg.
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3572
diff changeset
7 *
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3572
diff changeset
8 * FFmpeg is free software; you can redistribute it and/or
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
9 * modify it under the terms of the GNU Lesser General Public
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
10 * 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: 3572
diff changeset
11 * version 2.1 of the License, or (at your option) any later version.
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
12 *
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3572
diff changeset
13 * FFmpeg is distributed in the hope that it will be useful,
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
16 * Lesser General Public License for more details.
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
17 *
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
18 * 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: 3572
diff changeset
19 * 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
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
21 */
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
22
9421
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
23 #include "config.h"
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
24 #if HAVE_ALTIVEC_H
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
25 #include <altivec.h>
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
26 #endif
6763
f7cbb7733146 Use full path for #includes from another directory.
diego
parents: 6105
diff changeset
27 #include "libavcodec/dsputil.h"
5750
09f99af1db40 Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents: 5749
diff changeset
28 #include "util_altivec.h"
8307
bf5bc1f4cba0 add Altivec implementation of clear_block
gpoirier
parents: 7373
diff changeset
29 #include "types_altivec.h"
11382
50415a8f1451 PPC: move prototypes to headers and make some functions static
mru
parents: 10082
diff changeset
30 #include "dsputil_altivec.h"
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
31
11382
50415a8f1451 PPC: move prototypes to headers and make some functions static
mru
parents: 10082
diff changeset
32 static int sad16_x2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
33 {
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
34 int i;
10082
9f4b529bd5c0 PPC: remove unnecessary alignment on local variables
mru
parents: 9668
diff changeset
35 int s;
5746
55ed6dc5d476 Remove const vector macro indirection that is useless and obfuscating
diego
parents: 5609
diff changeset
36 const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
37 vector unsigned char *tv;
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
38 vector unsigned char pix1v, pix2v, pix2iv, avgv, t5;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
39 vector unsigned int sad;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
40 vector signed int sumdiffs;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
41
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
42 s = 0;
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
43 sad = (vector unsigned int)vec_splat_u32(0);
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
44 for (i = 0; i < h; i++) {
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
45 /* Read unaligned pixels into our vectors. The vectors are as follows:
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
46 pix1v: pix1[0]-pix1[15]
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
47 pix2v: pix2[0]-pix2[15] pix2iv: pix2[1]-pix2[16] */
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
48 tv = (vector unsigned char *) pix1;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
49 pix1v = vec_perm(tv[0], tv[1], vec_lvsl(0, pix1));
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
50
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
51 tv = (vector unsigned char *) &pix2[0];
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
52 pix2v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[0]));
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
53
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
54 tv = (vector unsigned char *) &pix2[1];
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
55 pix2iv = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[1]));
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
56
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
57 /* Calculate the average vector */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
58 avgv = vec_avg(pix2v, pix2iv);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
59
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
60 /* Calculate a sum of abs differences vector */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
61 t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv));
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
62
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
63 /* Add each 4 pixel group together and put 4 results into sad */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
64 sad = vec_sum4s(t5, sad);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
65
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
66 pix1 += line_size;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
67 pix2 += line_size;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
68 }
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
69 /* Sum up the four partial sums, and put the result into s */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
70 sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
71 sumdiffs = vec_splat(sumdiffs, 3);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
72 vec_ste(sumdiffs, 0, &s);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
73
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
74 return s;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
75 }
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
76
11382
50415a8f1451 PPC: move prototypes to headers and make some functions static
mru
parents: 10082
diff changeset
77 static int sad16_y2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
78 {
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
79 int i;
10082
9f4b529bd5c0 PPC: remove unnecessary alignment on local variables
mru
parents: 9668
diff changeset
80 int s;
5746
55ed6dc5d476 Remove const vector macro indirection that is useless and obfuscating
diego
parents: 5609
diff changeset
81 const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
82 vector unsigned char *tv;
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
83 vector unsigned char pix1v, pix2v, pix3v, avgv, t5;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
84 vector unsigned int sad;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
85 vector signed int sumdiffs;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
86 uint8_t *pix3 = pix2 + line_size;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
87
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
88 s = 0;
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
89 sad = (vector unsigned int)vec_splat_u32(0);
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
90
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
91 /* Due to the fact that pix3 = pix2 + line_size, the pix3 of one
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
92 iteration becomes pix2 in the next iteration. We can use this
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
93 fact to avoid a potentially expensive unaligned read, each
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
94 time around the loop.
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
95 Read unaligned pixels into our vectors. The vectors are as follows:
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
96 pix2v: pix2[0]-pix2[15]
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
97 Split the pixel vectors into shorts */
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
98 tv = (vector unsigned char *) &pix2[0];
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
99 pix2v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[0]));
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
100
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
101 for (i = 0; i < h; i++) {
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
102 /* Read unaligned pixels into our vectors. The vectors are as follows:
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
103 pix1v: pix1[0]-pix1[15]
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
104 pix3v: pix3[0]-pix3[15] */
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
105 tv = (vector unsigned char *) pix1;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
106 pix1v = vec_perm(tv[0], tv[1], vec_lvsl(0, pix1));
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
107
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
108 tv = (vector unsigned char *) &pix3[0];
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
109 pix3v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix3[0]));
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
110
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
111 /* Calculate the average vector */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
112 avgv = vec_avg(pix2v, pix3v);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
113
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
114 /* Calculate a sum of abs differences vector */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
115 t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv));
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
116
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
117 /* Add each 4 pixel group together and put 4 results into sad */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
118 sad = vec_sum4s(t5, sad);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
119
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
120 pix1 += line_size;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
121 pix2v = pix3v;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
122 pix3 += line_size;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
123
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
124 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
125
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
126 /* Sum up the four partial sums, and put the result into s */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
127 sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
128 sumdiffs = vec_splat(sumdiffs, 3);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
129 vec_ste(sumdiffs, 0, &s);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
130 return s;
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
131 }
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
132
11382
50415a8f1451 PPC: move prototypes to headers and make some functions static
mru
parents: 10082
diff changeset
133 static int sad16_xy2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
134 {
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
135 int i;
10082
9f4b529bd5c0 PPC: remove unnecessary alignment on local variables
mru
parents: 9668
diff changeset
136 int s;
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
137 uint8_t *pix3 = pix2 + line_size;
5746
55ed6dc5d476 Remove const vector macro indirection that is useless and obfuscating
diego
parents: 5609
diff changeset
138 const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);
55ed6dc5d476 Remove const vector macro indirection that is useless and obfuscating
diego
parents: 5609
diff changeset
139 const vector unsigned short two = (const vector unsigned short)vec_splat_u16(2);
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
140 vector unsigned char *tv, avgv, t5;
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
141 vector unsigned char pix1v, pix2v, pix3v, pix2iv, pix3iv;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
142 vector unsigned short pix2lv, pix2hv, pix2ilv, pix2ihv;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
143 vector unsigned short pix3lv, pix3hv, pix3ilv, pix3ihv;
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
144 vector unsigned short avghv, avglv;
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
145 vector unsigned short t1, t2, t3, t4;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
146 vector unsigned int sad;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
147 vector signed int sumdiffs;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
148
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
149 sad = (vector unsigned int)vec_splat_u32(0);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
150
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
151 s = 0;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
152
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
153 /* Due to the fact that pix3 = pix2 + line_size, the pix3 of one
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
154 iteration becomes pix2 in the next iteration. We can use this
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
155 fact to avoid a potentially expensive unaligned read, as well
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
156 as some splitting, and vector addition each time around the loop.
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
157 Read unaligned pixels into our vectors. The vectors are as follows:
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
158 pix2v: pix2[0]-pix2[15] pix2iv: pix2[1]-pix2[16]
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
159 Split the pixel vectors into shorts */
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
160 tv = (vector unsigned char *) &pix2[0];
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
161 pix2v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[0]));
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
162
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
163 tv = (vector unsigned char *) &pix2[1];
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
164 pix2iv = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[1]));
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
165
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
166 pix2hv = (vector unsigned short) vec_mergeh(zero, pix2v);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
167 pix2lv = (vector unsigned short) vec_mergel(zero, pix2v);
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
168 pix2ihv = (vector unsigned short) vec_mergeh(zero, pix2iv);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
169 pix2ilv = (vector unsigned short) vec_mergel(zero, pix2iv);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
170 t1 = vec_add(pix2hv, pix2ihv);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
171 t2 = vec_add(pix2lv, pix2ilv);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
172
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
173 for (i = 0; i < h; i++) {
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
174 /* Read unaligned pixels into our vectors. The vectors are as follows:
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
175 pix1v: pix1[0]-pix1[15]
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
176 pix3v: pix3[0]-pix3[15] pix3iv: pix3[1]-pix3[16] */
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
177 tv = (vector unsigned char *) pix1;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
178 pix1v = vec_perm(tv[0], tv[1], vec_lvsl(0, pix1));
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
179
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
180 tv = (vector unsigned char *) &pix3[0];
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
181 pix3v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix3[0]));
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
182
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
183 tv = (vector unsigned char *) &pix3[1];
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
184 pix3iv = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix3[1]));
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
185
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
186 /* Note that AltiVec does have vec_avg, but this works on vector pairs
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
187 and rounds up. We could do avg(avg(a,b),avg(c,d)), but the rounding
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
188 would mean that, for example, avg(3,0,0,1) = 2, when it should be 1.
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
189 Instead, we have to split the pixel vectors into vectors of shorts,
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
190 and do the averaging by hand. */
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
191
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
192 /* Split the pixel vectors into shorts */
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
193 pix3hv = (vector unsigned short) vec_mergeh(zero, pix3v);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
194 pix3lv = (vector unsigned short) vec_mergel(zero, pix3v);
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
195 pix3ihv = (vector unsigned short) vec_mergeh(zero, pix3iv);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
196 pix3ilv = (vector unsigned short) vec_mergel(zero, pix3iv);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
197
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
198 /* Do the averaging on them */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
199 t3 = vec_add(pix3hv, pix3ihv);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
200 t4 = vec_add(pix3lv, pix3ilv);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
201
884
2cef5c4c0ca6 * altivec and pix_norm patch by Brian Foley
kabi
parents: 878
diff changeset
202 avghv = vec_sr(vec_add(vec_add(t1, t3), two), two);
2cef5c4c0ca6 * altivec and pix_norm patch by Brian Foley
kabi
parents: 878
diff changeset
203 avglv = vec_sr(vec_add(vec_add(t2, t4), two), two);
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
204
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
205 /* Pack the shorts back into a result */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
206 avgv = vec_pack(avghv, avglv);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
207
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
208 /* Calculate a sum of abs differences vector */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
209 t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv));
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
210
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
211 /* Add each 4 pixel group together and put 4 results into sad */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
212 sad = vec_sum4s(t5, sad);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
213
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
214 pix1 += line_size;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
215 pix3 += line_size;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
216 /* Transfer the calculated values for pix3 into pix2 */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
217 t1 = t3;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
218 t2 = t4;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
219 }
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
220 /* Sum up the four partial sums, and put the result into s */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
221 sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
222 sumdiffs = vec_splat(sumdiffs, 3);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
223 vec_ste(sumdiffs, 0, &s);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
224
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
225 return s;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
226 }
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
227
11382
50415a8f1451 PPC: move prototypes to headers and make some functions static
mru
parents: 10082
diff changeset
228 static int sad16_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
229 {
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
230 int i;
10082
9f4b529bd5c0 PPC: remove unnecessary alignment on local variables
mru
parents: 9668
diff changeset
231 int s;
5746
55ed6dc5d476 Remove const vector macro indirection that is useless and obfuscating
diego
parents: 5609
diff changeset
232 const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
233 vector unsigned char perm1, perm2, *pix1v, *pix2v;
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
234 vector unsigned char t1, t2, t3,t4, t5;
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
235 vector unsigned int sad;
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
236 vector signed int sumdiffs;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
237
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
238 sad = (vector unsigned int)vec_splat_u32(0);
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
239
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
240
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
241 for (i = 0; i < h; i++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
242 /* Read potentially unaligned pixels into t1 and t2 */
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
243 perm1 = vec_lvsl(0, pix1);
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
244 pix1v = (vector unsigned char *) pix1;
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
245 perm2 = vec_lvsl(0, pix2);
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
246 pix2v = (vector unsigned char *) pix2;
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
247 t1 = vec_perm(pix1v[0], pix1v[1], perm1);
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
248 t2 = vec_perm(pix2v[0], pix2v[1], perm2);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
249
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
250 /* Calculate a sum of abs differences vector */
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
251 t3 = vec_max(t1, t2);
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
252 t4 = vec_min(t1, t2);
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
253 t5 = vec_sub(t3, t4);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
254
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
255 /* Add each 4 pixel group together and put 4 results into sad */
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
256 sad = vec_sum4s(t5, sad);
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
257
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
258 pix1 += line_size;
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
259 pix2 += line_size;
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
260 }
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
261
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
262 /* Sum up the four partial sums, and put the result into s */
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
263 sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
264 sumdiffs = vec_splat(sumdiffs, 3);
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
265 vec_ste(sumdiffs, 0, &s);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
266
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
267 return s;
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
268 }
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
269
11382
50415a8f1451 PPC: move prototypes to headers and make some functions static
mru
parents: 10082
diff changeset
270 static int sad8_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
271 {
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
272 int i;
10082
9f4b529bd5c0 PPC: remove unnecessary alignment on local variables
mru
parents: 9668
diff changeset
273 int s;
5746
55ed6dc5d476 Remove const vector macro indirection that is useless and obfuscating
diego
parents: 5609
diff changeset
274 const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
275 vector unsigned char perm1, perm2, permclear, *pix1v, *pix2v;
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
276 vector unsigned char t1, t2, t3,t4, t5;
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
277 vector unsigned int sad;
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
278 vector signed int sumdiffs;
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
279
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
280 sad = (vector unsigned int)vec_splat_u32(0);
1277
f3152eb76f1a altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents: 1064
diff changeset
281
7373
266d4949aa15 Remove AltiVec vector declaration compiler compatibility macros.
diego
parents: 7335
diff changeset
282 permclear = (vector unsigned char){255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0};
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
283
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
284 for (i = 0; i < h; i++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
285 /* Read potentially unaligned pixels into t1 and t2
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
286 Since we're reading 16 pixels, and actually only want 8,
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
287 mask out the last 8 pixels. The 0s don't change the sum. */
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
288 perm1 = vec_lvsl(0, pix1);
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
289 pix1v = (vector unsigned char *) pix1;
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
290 perm2 = vec_lvsl(0, pix2);
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
291 pix2v = (vector unsigned char *) pix2;
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
292 t1 = vec_and(vec_perm(pix1v[0], pix1v[1], perm1), permclear);
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
293 t2 = vec_and(vec_perm(pix2v[0], pix2v[1], perm2), permclear);
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
294
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
295 /* Calculate a sum of abs differences vector */
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
296 t3 = vec_max(t1, t2);
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
297 t4 = vec_min(t1, t2);
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
298 t5 = vec_sub(t3, t4);
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
299
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
300 /* Add each 4 pixel group together and put 4 results into sad */
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
301 sad = vec_sum4s(t5, sad);
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
302
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
303 pix1 += line_size;
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
304 pix2 += line_size;
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
305 }
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
306
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
307 /* Sum up the four partial sums, and put the result into s */
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
308 sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
309 sumdiffs = vec_splat(sumdiffs, 3);
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
310 vec_ste(sumdiffs, 0, &s);
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
311
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
312 return s;
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
313 }
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
314
11382
50415a8f1451 PPC: move prototypes to headers and make some functions static
mru
parents: 10082
diff changeset
315 static int pix_norm1_altivec(uint8_t *pix, int line_size)
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
316 {
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
317 int i;
10082
9f4b529bd5c0 PPC: remove unnecessary alignment on local variables
mru
parents: 9668
diff changeset
318 int s;
5746
55ed6dc5d476 Remove const vector macro indirection that is useless and obfuscating
diego
parents: 5609
diff changeset
319 const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
320 vector unsigned char *tv;
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
321 vector unsigned char pixv;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
322 vector unsigned int sv;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
323 vector signed int sum;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
324
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
325 sv = (vector unsigned int)vec_splat_u32(0);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
326
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
327 s = 0;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
328 for (i = 0; i < 16; i++) {
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
329 /* Read in the potentially unaligned pixels */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
330 tv = (vector unsigned char *) pix;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
331 pixv = vec_perm(tv[0], tv[1], vec_lvsl(0, pix));
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
332
884
2cef5c4c0ca6 * altivec and pix_norm patch by Brian Foley
kabi
parents: 878
diff changeset
333 /* Square the values, and add them to our sum */
2cef5c4c0ca6 * altivec and pix_norm patch by Brian Foley
kabi
parents: 878
diff changeset
334 sv = vec_msum(pixv, pixv, sv);
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
335
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
336 pix += line_size;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
337 }
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
338 /* Sum up the four partial sums, and put the result into s */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
339 sum = vec_sums((vector signed int) sv, (vector signed int) zero);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
340 sum = vec_splat(sum, 3);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
341 vec_ste(sum, 0, &s);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
342
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
343 return s;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
344 }
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
345
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
346 /**
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
347 * Sum of Squared Errors for a 8x8 block.
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
348 * AltiVec-enhanced.
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
349 * It's the sad8_altivec code above w/ squaring added.
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
350 */
11382
50415a8f1451 PPC: move prototypes to headers and make some functions static
mru
parents: 10082
diff changeset
351 static int sse8_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
352 {
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
353 int i;
10082
9f4b529bd5c0 PPC: remove unnecessary alignment on local variables
mru
parents: 9668
diff changeset
354 int s;
5746
55ed6dc5d476 Remove const vector macro indirection that is useless and obfuscating
diego
parents: 5609
diff changeset
355 const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
356 vector unsigned char perm1, perm2, permclear, *pix1v, *pix2v;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
357 vector unsigned char t1, t2, t3,t4, t5;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
358 vector unsigned int sum;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
359 vector signed int sumsqr;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
360
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
361 sum = (vector unsigned int)vec_splat_u32(0);
1277
f3152eb76f1a altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents: 1064
diff changeset
362
7373
266d4949aa15 Remove AltiVec vector declaration compiler compatibility macros.
diego
parents: 7335
diff changeset
363 permclear = (vector unsigned char){255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0};
1277
f3152eb76f1a altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents: 1064
diff changeset
364
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
365
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
366 for (i = 0; i < h; i++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
367 /* Read potentially unaligned pixels into t1 and t2
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
368 Since we're reading 16 pixels, and actually only want 8,
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
369 mask out the last 8 pixels. The 0s don't change the sum. */
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
370 perm1 = vec_lvsl(0, pix1);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
371 pix1v = (vector unsigned char *) pix1;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
372 perm2 = vec_lvsl(0, pix2);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
373 pix2v = (vector unsigned char *) pix2;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
374 t1 = vec_and(vec_perm(pix1v[0], pix1v[1], perm1), permclear);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
375 t2 = vec_and(vec_perm(pix2v[0], pix2v[1], perm2), permclear);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
376
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
377 /* Since we want to use unsigned chars, we can take advantage
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
378 of the fact that abs(a-b)^2 = (a-b)^2. */
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
379
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
380 /* Calculate abs differences vector */
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
381 t3 = vec_max(t1, t2);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
382 t4 = vec_min(t1, t2);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
383 t5 = vec_sub(t3, t4);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
384
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
385 /* Square the values and add them to our sum */
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
386 sum = vec_msum(t5, t5, sum);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
387
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
388 pix1 += line_size;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
389 pix2 += line_size;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
390 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
391
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
392 /* Sum up the four partial sums, and put the result into s */
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
393 sumsqr = vec_sums((vector signed int) sum, (vector signed int) zero);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
394 sumsqr = vec_splat(sumsqr, 3);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
395 vec_ste(sumsqr, 0, &s);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
396
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
397 return s;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
398 }
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
399
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
400 /**
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
401 * Sum of Squared Errors for a 16x16 block.
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
402 * AltiVec-enhanced.
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
403 * It's the sad16_altivec code above w/ squaring added.
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
404 */
11382
50415a8f1451 PPC: move prototypes to headers and make some functions static
mru
parents: 10082
diff changeset
405 static int sse16_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
406 {
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
407 int i;
10082
9f4b529bd5c0 PPC: remove unnecessary alignment on local variables
mru
parents: 9668
diff changeset
408 int s;
5746
55ed6dc5d476 Remove const vector macro indirection that is useless and obfuscating
diego
parents: 5609
diff changeset
409 const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
410 vector unsigned char perm1, perm2, *pix1v, *pix2v;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
411 vector unsigned char t1, t2, t3,t4, t5;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
412 vector unsigned int sum;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
413 vector signed int sumsqr;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
414
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
415 sum = (vector unsigned int)vec_splat_u32(0);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
416
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
417 for (i = 0; i < h; i++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
418 /* Read potentially unaligned pixels into t1 and t2 */
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
419 perm1 = vec_lvsl(0, pix1);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
420 pix1v = (vector unsigned char *) pix1;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
421 perm2 = vec_lvsl(0, pix2);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
422 pix2v = (vector unsigned char *) pix2;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
423 t1 = vec_perm(pix1v[0], pix1v[1], perm1);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
424 t2 = vec_perm(pix2v[0], pix2v[1], perm2);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
425
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
426 /* Since we want to use unsigned chars, we can take advantage
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
427 of the fact that abs(a-b)^2 = (a-b)^2. */
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
428
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
429 /* Calculate abs differences vector */
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
430 t3 = vec_max(t1, t2);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
431 t4 = vec_min(t1, t2);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
432 t5 = vec_sub(t3, t4);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
433
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
434 /* Square the values and add them to our sum */
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
435 sum = vec_msum(t5, t5, sum);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
436
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
437 pix1 += line_size;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
438 pix2 += line_size;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
439 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
440
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
441 /* Sum up the four partial sums, and put the result into s */
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
442 sumsqr = vec_sums((vector signed int) sum, (vector signed int) zero);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
443 sumsqr = vec_splat(sumsqr, 3);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
444 vec_ste(sumsqr, 0, &s);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
445
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
446 return s;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
447 }
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
448
11382
50415a8f1451 PPC: move prototypes to headers and make some functions static
mru
parents: 10082
diff changeset
449 static int pix_sum_altivec(uint8_t * pix, int line_size)
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
450 {
5746
55ed6dc5d476 Remove const vector macro indirection that is useless and obfuscating
diego
parents: 5609
diff changeset
451 const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
452 vector unsigned char perm, *pixv;
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
453 vector unsigned char t1;
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
454 vector unsigned int sad;
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
455 vector signed int sumdiffs;
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
456
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
457 int i;
10082
9f4b529bd5c0 PPC: remove unnecessary alignment on local variables
mru
parents: 9668
diff changeset
458 int s;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
459
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
460 sad = (vector unsigned int)vec_splat_u32(0);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
461
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
462 for (i = 0; i < 16; i++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
463 /* Read the potentially unaligned 16 pixels into t1 */
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
464 perm = vec_lvsl(0, pix);
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
465 pixv = (vector unsigned char *) pix;
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
466 t1 = vec_perm(pixv[0], pixv[1], perm);
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
467
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
468 /* Add each 4 pixel group together and put 4 results into sad */
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
469 sad = vec_sum4s(t1, sad);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
470
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
471 pix += line_size;
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
472 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
473
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
474 /* Sum up the four partial sums, and put the result into s */
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
475 sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
476 sumdiffs = vec_splat(sumdiffs, 3);
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
477 vec_ste(sumdiffs, 0, &s);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
478
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
479 return s;
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
480 }
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
481
11382
50415a8f1451 PPC: move prototypes to headers and make some functions static
mru
parents: 10082
diff changeset
482 static void get_pixels_altivec(DCTELEM *restrict block, const uint8_t *pixels, int line_size)
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
483 {
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
484 int i;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
485 vector unsigned char perm, bytes, *pixv;
5746
55ed6dc5d476 Remove const vector macro indirection that is useless and obfuscating
diego
parents: 5609
diff changeset
486 const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
487 vector signed short shorts;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
488
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
489 for (i = 0; i < 8; i++) {
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
490 // Read potentially unaligned pixels.
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
491 // We're reading 16 pixels, and actually only want 8,
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
492 // but we simply ignore the extras.
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
493 perm = vec_lvsl(0, pixels);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
494 pixv = (vector unsigned char *) pixels;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
495 bytes = vec_perm(pixv[0], pixv[1], perm);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
496
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
497 // convert the bytes into shorts
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
498 shorts = (vector signed short)vec_mergeh(zero, bytes);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
499
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
500 // save the data to the block, we assume the block is 16-byte aligned
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
501 vec_st(shorts, i*16, (vector signed short*)block);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
502
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
503 pixels += line_size;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
504 }
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
505 }
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
506
11382
50415a8f1451 PPC: move prototypes to headers and make some functions static
mru
parents: 10082
diff changeset
507 static void diff_pixels_altivec(DCTELEM *restrict block, const uint8_t *s1,
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1033
diff changeset
508 const uint8_t *s2, int stride)
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
509 {
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
510 int i;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
511 vector unsigned char perm, bytes, *pixv;
5746
55ed6dc5d476 Remove const vector macro indirection that is useless and obfuscating
diego
parents: 5609
diff changeset
512 const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
513 vector signed short shorts1, shorts2;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
514
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
515 for (i = 0; i < 4; i++) {
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
516 // Read potentially unaligned pixels
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
517 // We're reading 16 pixels, and actually only want 8,
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
518 // but we simply ignore the extras.
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
519 perm = vec_lvsl(0, s1);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
520 pixv = (vector unsigned char *) s1;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
521 bytes = vec_perm(pixv[0], pixv[1], perm);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
522
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
523 // convert the bytes into shorts
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
524 shorts1 = (vector signed short)vec_mergeh(zero, bytes);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
525
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
526 // Do the same for the second block of pixels
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
527 perm = vec_lvsl(0, s2);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
528 pixv = (vector unsigned char *) s2;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
529 bytes = vec_perm(pixv[0], pixv[1], perm);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
530
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
531 // convert the bytes into shorts
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
532 shorts2 = (vector signed short)vec_mergeh(zero, bytes);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
533
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
534 // Do the subtraction
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
535 shorts1 = vec_sub(shorts1, shorts2);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
536
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
537 // save the data to the block, we assume the block is 16-byte aligned
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
538 vec_st(shorts1, 0, (vector signed short*)block);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
539
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
540 s1 += stride;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
541 s2 += stride;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
542 block += 8;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
543
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
544
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
545 // The code below is a copy of the code above... This is a manual
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
546 // unroll.
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
547
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
548 // Read potentially unaligned pixels
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
549 // We're reading 16 pixels, and actually only want 8,
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
550 // but we simply ignore the extras.
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
551 perm = vec_lvsl(0, s1);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
552 pixv = (vector unsigned char *) s1;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
553 bytes = vec_perm(pixv[0], pixv[1], perm);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
554
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
555 // convert the bytes into shorts
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
556 shorts1 = (vector signed short)vec_mergeh(zero, bytes);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
557
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
558 // Do the same for the second block of pixels
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
559 perm = vec_lvsl(0, s2);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
560 pixv = (vector unsigned char *) s2;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
561 bytes = vec_perm(pixv[0], pixv[1], perm);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
562
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
563 // convert the bytes into shorts
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
564 shorts2 = (vector signed short)vec_mergeh(zero, bytes);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
565
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
566 // Do the subtraction
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
567 shorts1 = vec_sub(shorts1, shorts2);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
568
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
569 // save the data to the block, we assume the block is 16-byte aligned
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
570 vec_st(shorts1, 0, (vector signed short*)block);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
571
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
572 s1 += stride;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
573 s2 += stride;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
574 block += 8;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
575 }
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
576 }
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
577
8307
bf5bc1f4cba0 add Altivec implementation of clear_block
gpoirier
parents: 7373
diff changeset
578
bf5bc1f4cba0 add Altivec implementation of clear_block
gpoirier
parents: 7373
diff changeset
579 static void clear_block_altivec(DCTELEM *block) {
bf5bc1f4cba0 add Altivec implementation of clear_block
gpoirier
parents: 7373
diff changeset
580 LOAD_ZERO;
bf5bc1f4cba0 add Altivec implementation of clear_block
gpoirier
parents: 7373
diff changeset
581 vec_st(zero_s16v, 0, block);
bf5bc1f4cba0 add Altivec implementation of clear_block
gpoirier
parents: 7373
diff changeset
582 vec_st(zero_s16v, 16, block);
bf5bc1f4cba0 add Altivec implementation of clear_block
gpoirier
parents: 7373
diff changeset
583 vec_st(zero_s16v, 32, block);
bf5bc1f4cba0 add Altivec implementation of clear_block
gpoirier
parents: 7373
diff changeset
584 vec_st(zero_s16v, 48, block);
bf5bc1f4cba0 add Altivec implementation of clear_block
gpoirier
parents: 7373
diff changeset
585 vec_st(zero_s16v, 64, block);
bf5bc1f4cba0 add Altivec implementation of clear_block
gpoirier
parents: 7373
diff changeset
586 vec_st(zero_s16v, 80, block);
bf5bc1f4cba0 add Altivec implementation of clear_block
gpoirier
parents: 7373
diff changeset
587 vec_st(zero_s16v, 96, block);
bf5bc1f4cba0 add Altivec implementation of clear_block
gpoirier
parents: 7373
diff changeset
588 vec_st(zero_s16v, 112, block);
bf5bc1f4cba0 add Altivec implementation of clear_block
gpoirier
parents: 7373
diff changeset
589 }
bf5bc1f4cba0 add Altivec implementation of clear_block
gpoirier
parents: 7373
diff changeset
590
bf5bc1f4cba0 add Altivec implementation of clear_block
gpoirier
parents: 7373
diff changeset
591
11382
50415a8f1451 PPC: move prototypes to headers and make some functions static
mru
parents: 10082
diff changeset
592 static void add_bytes_altivec(uint8_t *dst, uint8_t *src, int w) {
995
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
593 register int i;
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
594 register vector unsigned char vdst, vsrc;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
595
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
596 /* dst and src are 16 bytes-aligned (guaranteed) */
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
597 for (i = 0 ; (i + 15) < w ; i+=16) {
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
598 vdst = vec_ld(i, (unsigned char*)dst);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
599 vsrc = vec_ld(i, (unsigned char*)src);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
600 vdst = vec_add(vsrc, vdst);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
601 vec_st(vdst, i, (unsigned char*)dst);
995
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
602 }
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
603 /* if w is not a multiple of 16 */
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
604 for (; (i < w) ; i++) {
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
605 dst[i] = src[i];
995
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
606 }
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
607 }
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
608
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
609 /* next one assumes that ((line_size % 16) == 0) */
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
610 void put_pixels16_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
611 {
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
612 register vector unsigned char pixelsv1, pixelsv2;
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
613 register vector unsigned char pixelsv1B, pixelsv2B;
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
614 register vector unsigned char pixelsv1C, pixelsv2C;
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
615 register vector unsigned char pixelsv1D, pixelsv2D;
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
616
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
617 register vector unsigned char perm = vec_lvsl(0, pixels);
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
618 int i;
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
619 register int line_size_2 = line_size << 1;
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
620 register int line_size_3 = line_size + line_size_2;
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
621 register int line_size_4 = line_size << 2;
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
622
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
623 // hand-unrolling the loop by 4 gains about 15%
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
624 // mininum execution time goes from 74 to 60 cycles
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
625 // it's faster than -funroll-loops, but using
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
626 // -funroll-loops w/ this is bad - 74 cycles again.
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
627 // all this is on a 7450, tuning for the 7450
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
628 #if 0
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
629 for (i = 0; i < h; i++) {
9668
5a730e38f1b3 Remove useless casts of vec_ld() parameters.
diego
parents: 9421
diff changeset
630 pixelsv1 = vec_ld(0, pixels);
5a730e38f1b3 Remove useless casts of vec_ld() parameters.
diego
parents: 9421
diff changeset
631 pixelsv2 = vec_ld(16, pixels);
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
632 vec_st(vec_perm(pixelsv1, pixelsv2, perm),
9668
5a730e38f1b3 Remove useless casts of vec_ld() parameters.
diego
parents: 9421
diff changeset
633 0, block);
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
634 pixels+=line_size;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
635 block +=line_size;
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
636 }
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
637 #else
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
638 for (i = 0; i < h; i += 4) {
9668
5a730e38f1b3 Remove useless casts of vec_ld() parameters.
diego
parents: 9421
diff changeset
639 pixelsv1 = vec_ld( 0, pixels);
5a730e38f1b3 Remove useless casts of vec_ld() parameters.
diego
parents: 9421
diff changeset
640 pixelsv2 = vec_ld(15, pixels);
5a730e38f1b3 Remove useless casts of vec_ld() parameters.
diego
parents: 9421
diff changeset
641 pixelsv1B = vec_ld(line_size, pixels);
5a730e38f1b3 Remove useless casts of vec_ld() parameters.
diego
parents: 9421
diff changeset
642 pixelsv2B = vec_ld(15 + line_size, pixels);
5a730e38f1b3 Remove useless casts of vec_ld() parameters.
diego
parents: 9421
diff changeset
643 pixelsv1C = vec_ld(line_size_2, pixels);
5a730e38f1b3 Remove useless casts of vec_ld() parameters.
diego
parents: 9421
diff changeset
644 pixelsv2C = vec_ld(15 + line_size_2, pixels);
5a730e38f1b3 Remove useless casts of vec_ld() parameters.
diego
parents: 9421
diff changeset
645 pixelsv1D = vec_ld(line_size_3, pixels);
5a730e38f1b3 Remove useless casts of vec_ld() parameters.
diego
parents: 9421
diff changeset
646 pixelsv2D = vec_ld(15 + line_size_3, pixels);
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
647 vec_st(vec_perm(pixelsv1, pixelsv2, perm),
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
648 0, (unsigned char*)block);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
649 vec_st(vec_perm(pixelsv1B, pixelsv2B, perm),
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
650 line_size, (unsigned char*)block);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
651 vec_st(vec_perm(pixelsv1C, pixelsv2C, perm),
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
652 line_size_2, (unsigned char*)block);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
653 vec_st(vec_perm(pixelsv1D, pixelsv2D, perm),
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
654 line_size_3, (unsigned char*)block);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
655 pixels+=line_size_4;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
656 block +=line_size_4;
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
657 }
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
658 #endif
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
659 }
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
660
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
661 /* next one assumes that ((line_size % 16) == 0) */
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
662 #define op_avg(a,b) a = ( ((a)|(b)) - ((((a)^(b))&0xFEFEFEFEUL)>>1) )
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
663 void avg_pixels16_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
664 {
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
665 register vector unsigned char pixelsv1, pixelsv2, pixelsv, blockv;
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
666 register vector unsigned char perm = vec_lvsl(0, pixels);
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
667 int i;
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
668
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
669 for (i = 0; i < h; i++) {
9668
5a730e38f1b3 Remove useless casts of vec_ld() parameters.
diego
parents: 9421
diff changeset
670 pixelsv1 = vec_ld( 0, pixels);
5a730e38f1b3 Remove useless casts of vec_ld() parameters.
diego
parents: 9421
diff changeset
671 pixelsv2 = vec_ld(16,pixels);
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
672 blockv = vec_ld(0, block);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
673 pixelsv = vec_perm(pixelsv1, pixelsv2, perm);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
674 blockv = vec_avg(blockv,pixelsv);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
675 vec_st(blockv, 0, (unsigned char*)block);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
676 pixels+=line_size;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
677 block +=line_size;
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
678 }
1015
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
679 }
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
680
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
681 /* next one assumes that ((line_size % 8) == 0) */
11382
50415a8f1451 PPC: move prototypes to headers and make some functions static
mru
parents: 10082
diff changeset
682 static void avg_pixels8_altivec(uint8_t * block, const uint8_t * pixels, int line_size, int h)
1015
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
683 {
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
684 register vector unsigned char pixelsv1, pixelsv2, pixelsv, blockv;
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
685 int i;
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
686
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
687 for (i = 0; i < h; i++) {
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
688 /* block is 8 bytes-aligned, so we're either in the
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
689 left block (16 bytes-aligned) or in the right block (not) */
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
690 int rightside = ((unsigned long)block & 0x0000000F);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
691
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
692 blockv = vec_ld(0, block);
9668
5a730e38f1b3 Remove useless casts of vec_ld() parameters.
diego
parents: 9421
diff changeset
693 pixelsv1 = vec_ld( 0, pixels);
5a730e38f1b3 Remove useless casts of vec_ld() parameters.
diego
parents: 9421
diff changeset
694 pixelsv2 = vec_ld(16, pixels);
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
695 pixelsv = vec_perm(pixelsv1, pixelsv2, vec_lvsl(0, pixels));
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
696
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
697 if (rightside) {
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
698 pixelsv = vec_perm(blockv, pixelsv, vcprm(0,1,s0,s1));
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
699 } else {
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
700 pixelsv = vec_perm(blockv, pixelsv, vcprm(s0,s1,2,3));
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
701 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
702
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
703 blockv = vec_avg(blockv, pixelsv);
1015
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
704
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
705 vec_st(blockv, 0, block);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
706
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
707 pixels += line_size;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
708 block += line_size;
1015
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
709 }
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
710 }
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
711
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
712 /* next one assumes that ((line_size % 8) == 0) */
11382
50415a8f1451 PPC: move prototypes to headers and make some functions static
mru
parents: 10082
diff changeset
713 static void put_pixels8_xy2_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
1015
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
714 {
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
715 register int i;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
716 register vector unsigned char pixelsv1, pixelsv2, pixelsavg;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
717 register vector unsigned char blockv, temp1, temp2;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
718 register vector unsigned short pixelssum1, pixelssum2, temp3;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
719 register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
720 register const vector unsigned short vctwo = (const vector unsigned short)vec_splat_u16(2);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
721
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
722 temp1 = vec_ld(0, pixels);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
723 temp2 = vec_ld(16, pixels);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
724 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
725 if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F) {
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
726 pixelsv2 = temp2;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
727 } else {
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
728 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
729 }
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
730 pixelsv1 = vec_mergeh(vczero, pixelsv1);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
731 pixelsv2 = vec_mergeh(vczero, pixelsv2);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
732 pixelssum1 = vec_add((vector unsigned short)pixelsv1,
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
733 (vector unsigned short)pixelsv2);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
734 pixelssum1 = vec_add(pixelssum1, vctwo);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
735
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
736 for (i = 0; i < h ; i++) {
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
737 int rightside = ((unsigned long)block & 0x0000000F);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
738 blockv = vec_ld(0, block);
1015
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
739
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
740 temp1 = vec_ld(line_size, pixels);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
741 temp2 = vec_ld(line_size + 16, pixels);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
742 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
743 if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F) {
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
744 pixelsv2 = temp2;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
745 } else {
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
746 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
747 }
1015
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
748
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
749 pixelsv1 = vec_mergeh(vczero, pixelsv1);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
750 pixelsv2 = vec_mergeh(vczero, pixelsv2);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
751 pixelssum2 = vec_add((vector unsigned short)pixelsv1,
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
752 (vector unsigned short)pixelsv2);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
753 temp3 = vec_add(pixelssum1, pixelssum2);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
754 temp3 = vec_sra(temp3, vctwo);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
755 pixelssum1 = vec_add(pixelssum2, vctwo);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
756 pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
757
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
758 if (rightside) {
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
759 blockv = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1));
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
760 } else {
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
761 blockv = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3));
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
762 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
763
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
764 vec_st(blockv, 0, block);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
765
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
766 block += line_size;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
767 pixels += line_size;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
768 }
995
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
769 }
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
770
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
771 /* next one assumes that ((line_size % 8) == 0) */
11382
50415a8f1451 PPC: move prototypes to headers and make some functions static
mru
parents: 10082
diff changeset
772 static void put_no_rnd_pixels8_xy2_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
773 {
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
774 register int i;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
775 register vector unsigned char pixelsv1, pixelsv2, pixelsavg;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
776 register vector unsigned char blockv, temp1, temp2;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
777 register vector unsigned short pixelssum1, pixelssum2, temp3;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
778 register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
779 register const vector unsigned short vcone = (const vector unsigned short)vec_splat_u16(1);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
780 register const vector unsigned short vctwo = (const vector unsigned short)vec_splat_u16(2);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
781
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
782 temp1 = vec_ld(0, pixels);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
783 temp2 = vec_ld(16, pixels);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
784 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
785 if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F) {
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
786 pixelsv2 = temp2;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
787 } else {
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
788 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
789 }
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
790 pixelsv1 = vec_mergeh(vczero, pixelsv1);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
791 pixelsv2 = vec_mergeh(vczero, pixelsv2);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
792 pixelssum1 = vec_add((vector unsigned short)pixelsv1,
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
793 (vector unsigned short)pixelsv2);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
794 pixelssum1 = vec_add(pixelssum1, vcone);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
795
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
796 for (i = 0; i < h ; i++) {
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
797 int rightside = ((unsigned long)block & 0x0000000F);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
798 blockv = vec_ld(0, block);
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
799
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
800 temp1 = vec_ld(line_size, pixels);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
801 temp2 = vec_ld(line_size + 16, pixels);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
802 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
803 if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F) {
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
804 pixelsv2 = temp2;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
805 } else {
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
806 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
807 }
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
808
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
809 pixelsv1 = vec_mergeh(vczero, pixelsv1);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
810 pixelsv2 = vec_mergeh(vczero, pixelsv2);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
811 pixelssum2 = vec_add((vector unsigned short)pixelsv1,
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
812 (vector unsigned short)pixelsv2);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
813 temp3 = vec_add(pixelssum1, pixelssum2);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
814 temp3 = vec_sra(temp3, vctwo);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
815 pixelssum1 = vec_add(pixelssum2, vcone);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
816 pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
817
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
818 if (rightside) {
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
819 blockv = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1));
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
820 } else {
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
821 blockv = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3));
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
822 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
823
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
824 vec_st(blockv, 0, block);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
825
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
826 block += line_size;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
827 pixels += line_size;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
828 }
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
829 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
830
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
831 /* next one assumes that ((line_size % 16) == 0) */
11382
50415a8f1451 PPC: move prototypes to headers and make some functions static
mru
parents: 10082
diff changeset
832 static void put_pixels16_xy2_altivec(uint8_t * block, const uint8_t * pixels, int line_size, int h)
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
833 {
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
834 register int i;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
835 register vector unsigned char pixelsv1, pixelsv2, pixelsv3, pixelsv4;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
836 register vector unsigned char blockv, temp1, temp2;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
837 register vector unsigned short temp3, temp4,
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
838 pixelssum1, pixelssum2, pixelssum3, pixelssum4;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
839 register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
840 register const vector unsigned short vctwo = (const vector unsigned short)vec_splat_u16(2);
1340
09b8fe0f0139 PPC fixes & clean-up patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1277
diff changeset
841
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
842 temp1 = vec_ld(0, pixels);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
843 temp2 = vec_ld(16, pixels);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
844 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
845 if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F) {
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
846 pixelsv2 = temp2;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
847 } else {
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
848 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
849 }
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
850 pixelsv3 = vec_mergel(vczero, pixelsv1);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
851 pixelsv4 = vec_mergel(vczero, pixelsv2);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
852 pixelsv1 = vec_mergeh(vczero, pixelsv1);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
853 pixelsv2 = vec_mergeh(vczero, pixelsv2);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
854 pixelssum3 = vec_add((vector unsigned short)pixelsv3,
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
855 (vector unsigned short)pixelsv4);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
856 pixelssum3 = vec_add(pixelssum3, vctwo);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
857 pixelssum1 = vec_add((vector unsigned short)pixelsv1,
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
858 (vector unsigned short)pixelsv2);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
859 pixelssum1 = vec_add(pixelssum1, vctwo);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
860
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
861 for (i = 0; i < h ; i++) {
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
862 blockv = vec_ld(0, block);
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
863
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
864 temp1 = vec_ld(line_size, pixels);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
865 temp2 = vec_ld(line_size + 16, pixels);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
866 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
867 if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F) {
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
868 pixelsv2 = temp2;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
869 } else {
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
870 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
871 }
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
872
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
873 pixelsv3 = vec_mergel(vczero, pixelsv1);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
874 pixelsv4 = vec_mergel(vczero, pixelsv2);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
875 pixelsv1 = vec_mergeh(vczero, pixelsv1);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
876 pixelsv2 = vec_mergeh(vczero, pixelsv2);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
877
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
878 pixelssum4 = vec_add((vector unsigned short)pixelsv3,
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
879 (vector unsigned short)pixelsv4);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
880 pixelssum2 = vec_add((vector unsigned short)pixelsv1,
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
881 (vector unsigned short)pixelsv2);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
882 temp4 = vec_add(pixelssum3, pixelssum4);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
883 temp4 = vec_sra(temp4, vctwo);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
884 temp3 = vec_add(pixelssum1, pixelssum2);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
885 temp3 = vec_sra(temp3, vctwo);
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
886
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
887 pixelssum3 = vec_add(pixelssum4, vctwo);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
888 pixelssum1 = vec_add(pixelssum2, vctwo);
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
889
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
890 blockv = vec_packsu(temp3, temp4);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
891
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
892 vec_st(blockv, 0, block);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
893
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
894 block += line_size;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
895 pixels += line_size;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
896 }
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
897 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
898
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
899 /* next one assumes that ((line_size % 16) == 0) */
11382
50415a8f1451 PPC: move prototypes to headers and make some functions static
mru
parents: 10082
diff changeset
900 static void put_no_rnd_pixels16_xy2_altivec(uint8_t * block, const uint8_t * pixels, int line_size, int h)
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
901 {
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
902 register int i;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
903 register vector unsigned char pixelsv1, pixelsv2, pixelsv3, pixelsv4;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
904 register vector unsigned char blockv, temp1, temp2;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
905 register vector unsigned short temp3, temp4,
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
906 pixelssum1, pixelssum2, pixelssum3, pixelssum4;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
907 register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
908 register const vector unsigned short vcone = (const vector unsigned short)vec_splat_u16(1);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
909 register const vector unsigned short vctwo = (const vector unsigned short)vec_splat_u16(2);
1340
09b8fe0f0139 PPC fixes & clean-up patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1277
diff changeset
910
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
911 temp1 = vec_ld(0, pixels);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
912 temp2 = vec_ld(16, pixels);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
913 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
914 if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F) {
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
915 pixelsv2 = temp2;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
916 } else {
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
917 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
918 }
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
919 pixelsv3 = vec_mergel(vczero, pixelsv1);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
920 pixelsv4 = vec_mergel(vczero, pixelsv2);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
921 pixelsv1 = vec_mergeh(vczero, pixelsv1);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
922 pixelsv2 = vec_mergeh(vczero, pixelsv2);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
923 pixelssum3 = vec_add((vector unsigned short)pixelsv3,
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
924 (vector unsigned short)pixelsv4);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
925 pixelssum3 = vec_add(pixelssum3, vcone);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
926 pixelssum1 = vec_add((vector unsigned short)pixelsv1,
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
927 (vector unsigned short)pixelsv2);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
928 pixelssum1 = vec_add(pixelssum1, vcone);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
929
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
930 for (i = 0; i < h ; i++) {
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
931 blockv = vec_ld(0, block);
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
932
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
933 temp1 = vec_ld(line_size, pixels);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
934 temp2 = vec_ld(line_size + 16, pixels);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
935 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
936 if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F) {
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
937 pixelsv2 = temp2;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
938 } else {
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
939 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
940 }
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
941
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
942 pixelsv3 = vec_mergel(vczero, pixelsv1);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
943 pixelsv4 = vec_mergel(vczero, pixelsv2);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
944 pixelsv1 = vec_mergeh(vczero, pixelsv1);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
945 pixelsv2 = vec_mergeh(vczero, pixelsv2);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
946
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
947 pixelssum4 = vec_add((vector unsigned short)pixelsv3,
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
948 (vector unsigned short)pixelsv4);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
949 pixelssum2 = vec_add((vector unsigned short)pixelsv1,
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
950 (vector unsigned short)pixelsv2);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
951 temp4 = vec_add(pixelssum3, pixelssum4);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
952 temp4 = vec_sra(temp4, vctwo);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
953 temp3 = vec_add(pixelssum1, pixelssum2);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
954 temp3 = vec_sra(temp3, vctwo);
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
955
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
956 pixelssum3 = vec_add(pixelssum4, vcone);
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
957 pixelssum1 = vec_add(pixelssum2, vcone);
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
958
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
959 blockv = vec_packsu(temp3, temp4);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
960
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
961 vec_st(blockv, 0, block);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
962
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
963 block += line_size;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
964 pixels += line_size;
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
965 }
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
966 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
967
11382
50415a8f1451 PPC: move prototypes to headers and make some functions static
mru
parents: 10082
diff changeset
968 static int hadamard8_diff8x8_altivec(/*MpegEncContext*/ void *s, uint8_t *dst, uint8_t *src, int stride, int h){
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
969 int sum;
5746
55ed6dc5d476 Remove const vector macro indirection that is useless and obfuscating
diego
parents: 5609
diff changeset
970 register const vector unsigned char vzero =
55ed6dc5d476 Remove const vector macro indirection that is useless and obfuscating
diego
parents: 5609
diff changeset
971 (const vector unsigned char)vec_splat_u8(0);
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
972 register vector signed short temp0, temp1, temp2, temp3, temp4,
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
973 temp5, temp6, temp7;
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
974 {
5746
55ed6dc5d476 Remove const vector macro indirection that is useless and obfuscating
diego
parents: 5609
diff changeset
975 register const vector signed short vprod1 =(const vector signed short)
7373
266d4949aa15 Remove AltiVec vector declaration compiler compatibility macros.
diego
parents: 7335
diff changeset
976 { 1,-1, 1,-1, 1,-1, 1,-1 };
5746
55ed6dc5d476 Remove const vector macro indirection that is useless and obfuscating
diego
parents: 5609
diff changeset
977 register const vector signed short vprod2 =(const vector signed short)
7373
266d4949aa15 Remove AltiVec vector declaration compiler compatibility macros.
diego
parents: 7335
diff changeset
978 { 1, 1,-1,-1, 1, 1,-1,-1 };
5746
55ed6dc5d476 Remove const vector macro indirection that is useless and obfuscating
diego
parents: 5609
diff changeset
979 register const vector signed short vprod3 =(const vector signed short)
7373
266d4949aa15 Remove AltiVec vector declaration compiler compatibility macros.
diego
parents: 7335
diff changeset
980 { 1, 1, 1, 1,-1,-1,-1,-1 };
5746
55ed6dc5d476 Remove const vector macro indirection that is useless and obfuscating
diego
parents: 5609
diff changeset
981 register const vector unsigned char perm1 = (const vector unsigned char)
7373
266d4949aa15 Remove AltiVec vector declaration compiler compatibility macros.
diego
parents: 7335
diff changeset
982 {0x02, 0x03, 0x00, 0x01, 0x06, 0x07, 0x04, 0x05,
266d4949aa15 Remove AltiVec vector declaration compiler compatibility macros.
diego
parents: 7335
diff changeset
983 0x0A, 0x0B, 0x08, 0x09, 0x0E, 0x0F, 0x0C, 0x0D};
5746
55ed6dc5d476 Remove const vector macro indirection that is useless and obfuscating
diego
parents: 5609
diff changeset
984 register const vector unsigned char perm2 = (const vector unsigned char)
7373
266d4949aa15 Remove AltiVec vector declaration compiler compatibility macros.
diego
parents: 7335
diff changeset
985 {0x04, 0x05, 0x06, 0x07, 0x00, 0x01, 0x02, 0x03,
266d4949aa15 Remove AltiVec vector declaration compiler compatibility macros.
diego
parents: 7335
diff changeset
986 0x0C, 0x0D, 0x0E, 0x0F, 0x08, 0x09, 0x0A, 0x0B};
5746
55ed6dc5d476 Remove const vector macro indirection that is useless and obfuscating
diego
parents: 5609
diff changeset
987 register const vector unsigned char perm3 = (const vector unsigned char)
7373
266d4949aa15 Remove AltiVec vector declaration compiler compatibility macros.
diego
parents: 7335
diff changeset
988 {0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F,
266d4949aa15 Remove AltiVec vector declaration compiler compatibility macros.
diego
parents: 7335
diff changeset
989 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07};
1949
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
990
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
991 #define ONEITERBUTTERFLY(i, res) \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
992 { \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
993 register vector unsigned char src1, src2, srcO; \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
994 register vector unsigned char dst1, dst2, dstO; \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
995 register vector signed short srcV, dstV; \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
996 register vector signed short but0, but1, but2, op1, op2, op3; \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
997 src1 = vec_ld(stride * i, src); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
998 src2 = vec_ld((stride * i) + 15, src); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
999 srcO = vec_perm(src1, src2, vec_lvsl(stride * i, src)); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1000 dst1 = vec_ld(stride * i, dst); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1001 dst2 = vec_ld((stride * i) + 15, dst); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1002 dstO = vec_perm(dst1, dst2, vec_lvsl(stride * i, dst)); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1003 /* promote the unsigned chars to signed shorts */ \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1004 /* we're in the 8x8 function, we only care for the first 8 */ \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1005 srcV = (vector signed short)vec_mergeh((vector signed char)vzero, \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1006 (vector signed char)srcO); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1007 dstV = (vector signed short)vec_mergeh((vector signed char)vzero, \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1008 (vector signed char)dstO); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1009 /* subtractions inside the first butterfly */ \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1010 but0 = vec_sub(srcV, dstV); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1011 op1 = vec_perm(but0, but0, perm1); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1012 but1 = vec_mladd(but0, vprod1, op1); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1013 op2 = vec_perm(but1, but1, perm2); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1014 but2 = vec_mladd(but1, vprod2, op2); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1015 op3 = vec_perm(but2, but2, perm3); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1016 res = vec_mladd(but2, vprod3, op3); \
1949
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1017 }
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1018 ONEITERBUTTERFLY(0, temp0);
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1019 ONEITERBUTTERFLY(1, temp1);
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1020 ONEITERBUTTERFLY(2, temp2);
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1021 ONEITERBUTTERFLY(3, temp3);
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1022 ONEITERBUTTERFLY(4, temp4);
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1023 ONEITERBUTTERFLY(5, temp5);
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1024 ONEITERBUTTERFLY(6, temp6);
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1025 ONEITERBUTTERFLY(7, temp7);
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1026 }
1949
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1027 #undef ONEITERBUTTERFLY
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1028 {
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1029 register vector signed int vsum;
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1030 register vector signed short line0 = vec_add(temp0, temp1);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1031 register vector signed short line1 = vec_sub(temp0, temp1);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1032 register vector signed short line2 = vec_add(temp2, temp3);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1033 register vector signed short line3 = vec_sub(temp2, temp3);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1034 register vector signed short line4 = vec_add(temp4, temp5);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1035 register vector signed short line5 = vec_sub(temp4, temp5);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1036 register vector signed short line6 = vec_add(temp6, temp7);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1037 register vector signed short line7 = vec_sub(temp6, temp7);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1038
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1039 register vector signed short line0B = vec_add(line0, line2);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1040 register vector signed short line2B = vec_sub(line0, line2);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1041 register vector signed short line1B = vec_add(line1, line3);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1042 register vector signed short line3B = vec_sub(line1, line3);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1043 register vector signed short line4B = vec_add(line4, line6);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1044 register vector signed short line6B = vec_sub(line4, line6);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1045 register vector signed short line5B = vec_add(line5, line7);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1046 register vector signed short line7B = vec_sub(line5, line7);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1047
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1048 register vector signed short line0C = vec_add(line0B, line4B);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1049 register vector signed short line4C = vec_sub(line0B, line4B);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1050 register vector signed short line1C = vec_add(line1B, line5B);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1051 register vector signed short line5C = vec_sub(line1B, line5B);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1052 register vector signed short line2C = vec_add(line2B, line6B);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1053 register vector signed short line6C = vec_sub(line2B, line6B);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1054 register vector signed short line3C = vec_add(line3B, line7B);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1055 register vector signed short line7C = vec_sub(line3B, line7B);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1056
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1057 vsum = vec_sum4s(vec_abs(line0C), vec_splat_s32(0));
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1058 vsum = vec_sum4s(vec_abs(line1C), vsum);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1059 vsum = vec_sum4s(vec_abs(line2C), vsum);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1060 vsum = vec_sum4s(vec_abs(line3C), vsum);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1061 vsum = vec_sum4s(vec_abs(line4C), vsum);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1062 vsum = vec_sum4s(vec_abs(line5C), vsum);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1063 vsum = vec_sum4s(vec_abs(line6C), vsum);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1064 vsum = vec_sum4s(vec_abs(line7C), vsum);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1065 vsum = vec_sums(vsum, (vector signed int)vzero);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1066 vsum = vec_splat(vsum, 3);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1067 vec_ste(vsum, 0, &sum);
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1068 }
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1069 return sum;
1949
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1070 }
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1071
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1072 /*
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1073 16x8 works with 16 elements; it allows to avoid replicating loads, and
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1074 give the compiler more rooms for scheduling. It's only used from
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1075 inside hadamard8_diff16_altivec.
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1076
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1077 Unfortunately, it seems gcc-3.3 is a bit dumb, and the compiled code has a LOT
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1078 of spill code, it seems gcc (unlike xlc) cannot keep everything in registers
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1079 by itself. The following code include hand-made registers allocation. It's not
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1080 clean, but on a 7450 the resulting code is much faster (best case fall from
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1081 700+ cycles to 550).
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1082
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1083 xlc doesn't add spill code, but it doesn't know how to schedule for the 7450,
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1084 and its code isn't much faster than gcc-3.3 on the 7450 (but uses 25% less
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1085 instructions...)
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1086
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1087 On the 970, the hand-made RA is still a win (around 690 vs. around 780), but
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1088 xlc goes to around 660 on the regular C code...
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1089 */
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1090
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1091 static int hadamard8_diff16x8_altivec(/*MpegEncContext*/ void *s, uint8_t *dst, uint8_t *src, int stride, int h) {
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1092 int sum;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1093 register vector signed short
9421
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1094 temp0 __asm__ ("v0"),
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1095 temp1 __asm__ ("v1"),
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1096 temp2 __asm__ ("v2"),
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1097 temp3 __asm__ ("v3"),
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1098 temp4 __asm__ ("v4"),
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1099 temp5 __asm__ ("v5"),
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1100 temp6 __asm__ ("v6"),
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1101 temp7 __asm__ ("v7");
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1102 register vector signed short
9421
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1103 temp0S __asm__ ("v8"),
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1104 temp1S __asm__ ("v9"),
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1105 temp2S __asm__ ("v10"),
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1106 temp3S __asm__ ("v11"),
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1107 temp4S __asm__ ("v12"),
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1108 temp5S __asm__ ("v13"),
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1109 temp6S __asm__ ("v14"),
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1110 temp7S __asm__ ("v15");
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1111 register const vector unsigned char vzero __asm__ ("v31") =
5746
55ed6dc5d476 Remove const vector macro indirection that is useless and obfuscating
diego
parents: 5609
diff changeset
1112 (const vector unsigned char)vec_splat_u8(0);
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1113 {
9421
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1114 register const vector signed short vprod1 __asm__ ("v16") =
7373
266d4949aa15 Remove AltiVec vector declaration compiler compatibility macros.
diego
parents: 7335
diff changeset
1115 (const vector signed short){ 1,-1, 1,-1, 1,-1, 1,-1 };
9421
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1116 register const vector signed short vprod2 __asm__ ("v17") =
7373
266d4949aa15 Remove AltiVec vector declaration compiler compatibility macros.
diego
parents: 7335
diff changeset
1117 (const vector signed short){ 1, 1,-1,-1, 1, 1,-1,-1 };
9421
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1118 register const vector signed short vprod3 __asm__ ("v18") =
7373
266d4949aa15 Remove AltiVec vector declaration compiler compatibility macros.
diego
parents: 7335
diff changeset
1119 (const vector signed short){ 1, 1, 1, 1,-1,-1,-1,-1 };
9421
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1120 register const vector unsigned char perm1 __asm__ ("v19") =
5746
55ed6dc5d476 Remove const vector macro indirection that is useless and obfuscating
diego
parents: 5609
diff changeset
1121 (const vector unsigned char)
7373
266d4949aa15 Remove AltiVec vector declaration compiler compatibility macros.
diego
parents: 7335
diff changeset
1122 {0x02, 0x03, 0x00, 0x01, 0x06, 0x07, 0x04, 0x05,
266d4949aa15 Remove AltiVec vector declaration compiler compatibility macros.
diego
parents: 7335
diff changeset
1123 0x0A, 0x0B, 0x08, 0x09, 0x0E, 0x0F, 0x0C, 0x0D};
9421
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1124 register const vector unsigned char perm2 __asm__ ("v20") =
5746
55ed6dc5d476 Remove const vector macro indirection that is useless and obfuscating
diego
parents: 5609
diff changeset
1125 (const vector unsigned char)
7373
266d4949aa15 Remove AltiVec vector declaration compiler compatibility macros.
diego
parents: 7335
diff changeset
1126 {0x04, 0x05, 0x06, 0x07, 0x00, 0x01, 0x02, 0x03,
266d4949aa15 Remove AltiVec vector declaration compiler compatibility macros.
diego
parents: 7335
diff changeset
1127 0x0C, 0x0D, 0x0E, 0x0F, 0x08, 0x09, 0x0A, 0x0B};
9421
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1128 register const vector unsigned char perm3 __asm__ ("v21") =
5746
55ed6dc5d476 Remove const vector macro indirection that is useless and obfuscating
diego
parents: 5609
diff changeset
1129 (const vector unsigned char)
7373
266d4949aa15 Remove AltiVec vector declaration compiler compatibility macros.
diego
parents: 7335
diff changeset
1130 {0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F,
266d4949aa15 Remove AltiVec vector declaration compiler compatibility macros.
diego
parents: 7335
diff changeset
1131 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07};
1980
a6972a4f90c8 use the AVV macro from gcc_fixes.h instead ifdefs
alex
parents: 1979
diff changeset
1132
9421
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1133 #define ONEITERBUTTERFLY(i, res1, res2) \
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1134 { \
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1135 register vector unsigned char src1 __asm__ ("v22"), \
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1136 src2 __asm__ ("v23"), \
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1137 dst1 __asm__ ("v24"), \
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1138 dst2 __asm__ ("v25"), \
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1139 srcO __asm__ ("v22"), \
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1140 dstO __asm__ ("v23"); \
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1141 \
9421
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1142 register vector signed short srcV __asm__ ("v24"), \
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1143 dstV __asm__ ("v25"), \
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1144 srcW __asm__ ("v26"), \
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1145 dstW __asm__ ("v27"), \
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1146 but0 __asm__ ("v28"), \
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1147 but0S __asm__ ("v29"), \
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1148 op1 __asm__ ("v30"), \
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1149 but1 __asm__ ("v22"), \
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1150 op1S __asm__ ("v23"), \
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1151 but1S __asm__ ("v24"), \
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1152 op2 __asm__ ("v25"), \
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1153 but2 __asm__ ("v26"), \
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1154 op2S __asm__ ("v27"), \
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1155 but2S __asm__ ("v28"), \
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1156 op3 __asm__ ("v29"), \
dd2b5e52336a Remove gcc_fixes.h. It only contains workarounds for unsupported gcc versions.
diego
parents: 8596
diff changeset
1157 op3S __asm__ ("v30"); \
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1158 \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1159 src1 = vec_ld(stride * i, src); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1160 src2 = vec_ld((stride * i) + 16, src); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1161 srcO = vec_perm(src1, src2, vec_lvsl(stride * i, src)); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1162 dst1 = vec_ld(stride * i, dst); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1163 dst2 = vec_ld((stride * i) + 16, dst); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1164 dstO = vec_perm(dst1, dst2, vec_lvsl(stride * i, dst)); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1165 /* promote the unsigned chars to signed shorts */ \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1166 srcV = (vector signed short)vec_mergeh((vector signed char)vzero, \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1167 (vector signed char)srcO); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1168 dstV = (vector signed short)vec_mergeh((vector signed char)vzero, \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1169 (vector signed char)dstO); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1170 srcW = (vector signed short)vec_mergel((vector signed char)vzero, \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1171 (vector signed char)srcO); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1172 dstW = (vector signed short)vec_mergel((vector signed char)vzero, \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1173 (vector signed char)dstO); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1174 /* subtractions inside the first butterfly */ \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1175 but0 = vec_sub(srcV, dstV); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1176 but0S = vec_sub(srcW, dstW); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1177 op1 = vec_perm(but0, but0, perm1); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1178 but1 = vec_mladd(but0, vprod1, op1); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1179 op1S = vec_perm(but0S, but0S, perm1); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1180 but1S = vec_mladd(but0S, vprod1, op1S); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1181 op2 = vec_perm(but1, but1, perm2); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1182 but2 = vec_mladd(but1, vprod2, op2); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1183 op2S = vec_perm(but1S, but1S, perm2); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1184 but2S = vec_mladd(but1S, vprod2, op2S); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1185 op3 = vec_perm(but2, but2, perm3); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1186 res1 = vec_mladd(but2, vprod3, op3); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1187 op3S = vec_perm(but2S, but2S, perm3); \
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1188 res2 = vec_mladd(but2S, vprod3, op3S); \
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1189 }
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1190 ONEITERBUTTERFLY(0, temp0, temp0S);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1191 ONEITERBUTTERFLY(1, temp1, temp1S);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1192 ONEITERBUTTERFLY(2, temp2, temp2S);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1193 ONEITERBUTTERFLY(3, temp3, temp3S);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1194 ONEITERBUTTERFLY(4, temp4, temp4S);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1195 ONEITERBUTTERFLY(5, temp5, temp5S);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1196 ONEITERBUTTERFLY(6, temp6, temp6S);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1197 ONEITERBUTTERFLY(7, temp7, temp7S);
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1198 }
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1199 #undef ONEITERBUTTERFLY
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1200 {
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1201 register vector signed int vsum;
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1202 register vector signed short line0S, line1S, line2S, line3S, line4S,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1203 line5S, line6S, line7S, line0BS,line2BS,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1204 line1BS,line3BS,line4BS,line6BS,line5BS,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1205 line7BS,line0CS,line4CS,line1CS,line5CS,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1206 line2CS,line6CS,line3CS,line7CS;
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1207
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1208 register vector signed short line0 = vec_add(temp0, temp1);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1209 register vector signed short line1 = vec_sub(temp0, temp1);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1210 register vector signed short line2 = vec_add(temp2, temp3);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1211 register vector signed short line3 = vec_sub(temp2, temp3);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1212 register vector signed short line4 = vec_add(temp4, temp5);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1213 register vector signed short line5 = vec_sub(temp4, temp5);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1214 register vector signed short line6 = vec_add(temp6, temp7);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1215 register vector signed short line7 = vec_sub(temp6, temp7);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1216
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1217 register vector signed short line0B = vec_add(line0, line2);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1218 register vector signed short line2B = vec_sub(line0, line2);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1219 register vector signed short line1B = vec_add(line1, line3);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1220 register vector signed short line3B = vec_sub(line1, line3);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1221 register vector signed short line4B = vec_add(line4, line6);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1222 register vector signed short line6B = vec_sub(line4, line6);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1223 register vector signed short line5B = vec_add(line5, line7);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1224 register vector signed short line7B = vec_sub(line5, line7);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1225
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1226 register vector signed short line0C = vec_add(line0B, line4B);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1227 register vector signed short line4C = vec_sub(line0B, line4B);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1228 register vector signed short line1C = vec_add(line1B, line5B);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1229 register vector signed short line5C = vec_sub(line1B, line5B);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1230 register vector signed short line2C = vec_add(line2B, line6B);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1231 register vector signed short line6C = vec_sub(line2B, line6B);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1232 register vector signed short line3C = vec_add(line3B, line7B);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1233 register vector signed short line7C = vec_sub(line3B, line7B);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1234
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1235 vsum = vec_sum4s(vec_abs(line0C), vec_splat_s32(0));
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1236 vsum = vec_sum4s(vec_abs(line1C), vsum);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1237 vsum = vec_sum4s(vec_abs(line2C), vsum);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1238 vsum = vec_sum4s(vec_abs(line3C), vsum);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1239 vsum = vec_sum4s(vec_abs(line4C), vsum);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1240 vsum = vec_sum4s(vec_abs(line5C), vsum);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1241 vsum = vec_sum4s(vec_abs(line6C), vsum);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1242 vsum = vec_sum4s(vec_abs(line7C), vsum);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1243
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1244 line0S = vec_add(temp0S, temp1S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1245 line1S = vec_sub(temp0S, temp1S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1246 line2S = vec_add(temp2S, temp3S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1247 line3S = vec_sub(temp2S, temp3S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1248 line4S = vec_add(temp4S, temp5S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1249 line5S = vec_sub(temp4S, temp5S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1250 line6S = vec_add(temp6S, temp7S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1251 line7S = vec_sub(temp6S, temp7S);
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1252
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1253 line0BS = vec_add(line0S, line2S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1254 line2BS = vec_sub(line0S, line2S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1255 line1BS = vec_add(line1S, line3S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1256 line3BS = vec_sub(line1S, line3S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1257 line4BS = vec_add(line4S, line6S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1258 line6BS = vec_sub(line4S, line6S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1259 line5BS = vec_add(line5S, line7S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1260 line7BS = vec_sub(line5S, line7S);
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1261
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1262 line0CS = vec_add(line0BS, line4BS);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1263 line4CS = vec_sub(line0BS, line4BS);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1264 line1CS = vec_add(line1BS, line5BS);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1265 line5CS = vec_sub(line1BS, line5BS);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1266 line2CS = vec_add(line2BS, line6BS);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1267 line6CS = vec_sub(line2BS, line6BS);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1268 line3CS = vec_add(line3BS, line7BS);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1269 line7CS = vec_sub(line3BS, line7BS);
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1270
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1271 vsum = vec_sum4s(vec_abs(line0CS), vsum);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1272 vsum = vec_sum4s(vec_abs(line1CS), vsum);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1273 vsum = vec_sum4s(vec_abs(line2CS), vsum);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1274 vsum = vec_sum4s(vec_abs(line3CS), vsum);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1275 vsum = vec_sum4s(vec_abs(line4CS), vsum);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1276 vsum = vec_sum4s(vec_abs(line5CS), vsum);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1277 vsum = vec_sum4s(vec_abs(line6CS), vsum);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1278 vsum = vec_sum4s(vec_abs(line7CS), vsum);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1279 vsum = vec_sums(vsum, (vector signed int)vzero);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1280 vsum = vec_splat(vsum, 3);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1281 vec_ste(vsum, 0, &sum);
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1282 }
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1283 return sum;
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1284 }
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1285
11382
50415a8f1451 PPC: move prototypes to headers and make some functions static
mru
parents: 10082
diff changeset
1286 static int hadamard8_diff16_altivec(/*MpegEncContext*/ void *s, uint8_t *dst, uint8_t *src, int stride, int h){
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1287 int score;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1288 score = hadamard8_diff16x8_altivec(s, dst, src, stride, 8);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1289 if (h==16) {
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1290 dst += 8*stride;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1291 src += 8*stride;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1292 score += hadamard8_diff16x8_altivec(s, dst, src, stride, 8);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1293 }
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1294 return score;
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1295 }
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1296
3543
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1297 static void vorbis_inverse_coupling_altivec(float *mag, float *ang,
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1298 int blocksize)
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1299 {
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1300 int i;
3546
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1301 vector float m, a;
3543
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1302 vector bool int t0, t1;
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1303 const vector unsigned int v_31 = //XXX
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1304 vec_add(vec_add(vec_splat_u32(15),vec_splat_u32(15)),vec_splat_u32(1));
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1305 for (i = 0; i < blocksize; i += 4) {
3543
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1306 m = vec_ld(0, mag+i);
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1307 a = vec_ld(0, ang+i);
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1308 t0 = vec_cmple(m, (vector float)vec_splat_u32(0));
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1309 t1 = vec_cmple(a, (vector float)vec_splat_u32(0));
3545
e2a589e55906 Minor fix
lu_zero
parents: 3543
diff changeset
1310 a = vec_xor(a, (vector float) vec_sl((vector unsigned int)t0, v_31));
3546
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1311 t0 = (vector bool int)vec_and(a, t1);
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1312 t1 = (vector bool int)vec_andc(a, t1);
3550
4f4c13574ad5 Yet another typo
lu_zero
parents: 3549
diff changeset
1313 a = vec_sub(m, (vector float)t1);
4f4c13574ad5 Yet another typo
lu_zero
parents: 3549
diff changeset
1314 m = vec_add(m, (vector float)t0);
3549
7b4e34f1ff1f Fix a stupid typo and another error, thanks to Emanuele Giaquinta <exg@gentoo.org> for pointing out the issue and the patch
lu_zero
parents: 3546
diff changeset
1315 vec_stl(a, 0, ang+i);
7b4e34f1ff1f Fix a stupid typo and another error, thanks to Emanuele Giaquinta <exg@gentoo.org> for pointing out the issue and the patch
lu_zero
parents: 3546
diff changeset
1316 vec_stl(m, 0, mag+i);
3543
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1317 }
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1318 }
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1319
2057
4c663228e020 avg_pixels8_xy2_altivec in AltiVec, enabling avg_pixels8_altivec, hadamard fix by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 2056
diff changeset
1320 /* next one assumes that ((line_size % 8) == 0) */
11382
50415a8f1451 PPC: move prototypes to headers and make some functions static
mru
parents: 10082
diff changeset
1321 static void avg_pixels8_xy2_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
2057
4c663228e020 avg_pixels8_xy2_altivec in AltiVec, enabling avg_pixels8_altivec, hadamard fix by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 2056
diff changeset
1322 {
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1323 register int i;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1324 register vector unsigned char pixelsv1, pixelsv2, pixelsavg;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1325 register vector unsigned char blockv, temp1, temp2, blocktemp;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1326 register vector unsigned short pixelssum1, pixelssum2, temp3;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1327
5746
55ed6dc5d476 Remove const vector macro indirection that is useless and obfuscating
diego
parents: 5609
diff changeset
1328 register const vector unsigned char vczero = (const vector unsigned char)
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1329 vec_splat_u8(0);
5746
55ed6dc5d476 Remove const vector macro indirection that is useless and obfuscating
diego
parents: 5609
diff changeset
1330 register const vector unsigned short vctwo = (const vector unsigned short)
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1331 vec_splat_u16(2);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1332
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1333 temp1 = vec_ld(0, pixels);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1334 temp2 = vec_ld(16, pixels);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1335 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1336 if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F) {
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1337 pixelsv2 = temp2;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1338 } else {
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1339 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1340 }
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1341 pixelsv1 = vec_mergeh(vczero, pixelsv1);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1342 pixelsv2 = vec_mergeh(vczero, pixelsv2);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1343 pixelssum1 = vec_add((vector unsigned short)pixelsv1,
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1344 (vector unsigned short)pixelsv2);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1345 pixelssum1 = vec_add(pixelssum1, vctwo);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1346
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1347 for (i = 0; i < h ; i++) {
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1348 int rightside = ((unsigned long)block & 0x0000000F);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1349 blockv = vec_ld(0, block);
2057
4c663228e020 avg_pixels8_xy2_altivec in AltiVec, enabling avg_pixels8_altivec, hadamard fix by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 2056
diff changeset
1350
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1351 temp1 = vec_ld(line_size, pixels);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1352 temp2 = vec_ld(line_size + 16, pixels);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1353 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
7335
d463d8ee7755 cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines.
diego
parents: 6763
diff changeset
1354 if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F) {
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1355 pixelsv2 = temp2;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1356 } else {
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1357 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1358 }
2057
4c663228e020 avg_pixels8_xy2_altivec in AltiVec, enabling avg_pixels8_altivec, hadamard fix by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 2056
diff changeset
1359
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1360 pixelsv1 = vec_mergeh(vczero, pixelsv1);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1361 pixelsv2 = vec_mergeh(vczero, pixelsv2);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1362 pixelssum2 = vec_add((vector unsigned short)pixelsv1,
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1363 (vector unsigned short)pixelsv2);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1364 temp3 = vec_add(pixelssum1, pixelssum2);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1365 temp3 = vec_sra(temp3, vctwo);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1366 pixelssum1 = vec_add(pixelssum2, vctwo);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1367 pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1368
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1369 if (rightside) {
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1370 blocktemp = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1));
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1371 } else {
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1372 blocktemp = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3));
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1373 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1374
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1375 blockv = vec_avg(blocktemp, blockv);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1376 vec_st(blockv, 0, block);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1377
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1378 block += line_size;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1379 pixels += line_size;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1380 }
2057
4c663228e020 avg_pixels8_xy2_altivec in AltiVec, enabling avg_pixels8_altivec, hadamard fix by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 2056
diff changeset
1381 }
3546
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1382
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1383 void dsputil_init_altivec(DSPContext* c, AVCodecContext *avctx)
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1384 {
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1385 c->pix_abs[0][1] = sad16_x2_altivec;
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1386 c->pix_abs[0][2] = sad16_y2_altivec;
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1387 c->pix_abs[0][3] = sad16_xy2_altivec;
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1388 c->pix_abs[0][0] = sad16_altivec;
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1389 c->pix_abs[1][0] = sad8_altivec;
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1390 c->sad[0]= sad16_altivec;
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1391 c->sad[1]= sad8_altivec;
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1392 c->pix_norm1 = pix_norm1_altivec;
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1393 c->sse[1]= sse8_altivec;
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1394 c->sse[0]= sse16_altivec;
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1395 c->pix_sum = pix_sum_altivec;
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1396 c->diff_pixels = diff_pixels_altivec;
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1397 c->get_pixels = get_pixels_altivec;
8307
bf5bc1f4cba0 add Altivec implementation of clear_block
gpoirier
parents: 7373
diff changeset
1398 c->clear_block = clear_block_altivec;
3546
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1399 c->add_bytes= add_bytes_altivec;
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1400 c->put_pixels_tab[0][0] = put_pixels16_altivec;
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1401 /* the two functions do the same thing, so use the same code */
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1402 c->put_no_rnd_pixels_tab[0][0] = put_pixels16_altivec;
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1403 c->avg_pixels_tab[0][0] = avg_pixels16_altivec;
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1404 c->avg_pixels_tab[1][0] = avg_pixels8_altivec;
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1405 c->avg_pixels_tab[1][3] = avg_pixels8_xy2_altivec;
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1406 c->put_pixels_tab[1][3] = put_pixels8_xy2_altivec;
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1407 c->put_no_rnd_pixels_tab[1][3] = put_no_rnd_pixels8_xy2_altivec;
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1408 c->put_pixels_tab[0][3] = put_pixels16_xy2_altivec;
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1409 c->put_no_rnd_pixels_tab[0][3] = put_no_rnd_pixels16_xy2_altivec;
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1410
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1411 c->hadamard8_diff[0] = hadamard8_diff16_altivec;
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1412 c->hadamard8_diff[1] = hadamard8_diff8x8_altivec;
8596
68e959302527 replace all occurrence of ENABLE_ by the corresponding CONFIG_, HAVE_ or ARCH_
aurel
parents: 8307
diff changeset
1413 if (CONFIG_VORBIS_DECODER)
5753
25429b089d96 cosmetics: Fix indentation after last commit.
diego
parents: 5752
diff changeset
1414 c->vorbis_inverse_coupling = vorbis_inverse_coupling_altivec;
3546
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1415 }