annotate ppc/dsputil_altivec.c @ 4379:291db895a809 libavcodec

remove useless #ifdef CONFIG_ZLIB
author mru
date Sun, 21 Jan 2007 21:55:10 +0000
parents b28edd190fc0
children c867ae28d4de
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
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
23 #include "../dsputil.h"
1277
f3152eb76f1a altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents: 1064
diff changeset
24
f3152eb76f1a altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents: 1064
diff changeset
25 #include "gcc_fixes.h"
f3152eb76f1a altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents: 1064
diff changeset
26
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
27 #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
28
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
29 #ifdef CONFIG_DARWIN
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
30 #include <sys/sysctl.h>
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
31 #else /* CONFIG_DARWIN */
2286
0f937b191cc3 Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents: 2068
diff changeset
32 #ifdef __AMIGAOS4__
0f937b191cc3 Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents: 2068
diff changeset
33 #include <exec/exec.h>
0f937b191cc3 Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents: 2068
diff changeset
34 #include <interfaces/exec.h>
0f937b191cc3 Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents: 2068
diff changeset
35 #include <proto/exec.h>
0f937b191cc3 Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents: 2068
diff changeset
36 #else /* __AMIGAOS4__ */
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
37 #include <signal.h>
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
38 #include <setjmp.h>
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
39
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
40 static sigjmp_buf jmpbuf;
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
41 static volatile sig_atomic_t canjump = 0;
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
42
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
43 static void sigill_handler (int sig)
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
44 {
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
45 if (!canjump) {
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
46 signal (sig, SIG_DFL);
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
47 raise (sig);
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
48 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
49
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
50 canjump = 0;
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
51 siglongjmp (jmpbuf, 1);
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
52 }
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
53 #endif /* CONFIG_DARWIN */
2286
0f937b191cc3 Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents: 2068
diff changeset
54 #endif /* __AMIGAOS4__ */
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
55
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
56 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
57 {
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
58 int i;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
59 int s __attribute__((aligned(16)));
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
60 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
61 vector unsigned char *tv;
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
62 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
63 vector unsigned int sad;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
64 vector signed int sumdiffs;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
65
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
66 s = 0;
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
67 sad = (vector unsigned int)vec_splat_u32(0);
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
68 for(i=0;i<h;i++) {
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
69 /*
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
70 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
71 pix1v: pix1[0]-pix1[15]
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
72 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
73 */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
74 tv = (vector unsigned char *) pix1;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
75 pix1v = vec_perm(tv[0], tv[1], vec_lvsl(0, pix1));
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
76
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
77 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
78 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
79
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
80 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
81 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
82
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
83 /* Calculate the average vector */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
84 avgv = vec_avg(pix2v, pix2iv);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
85
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
86 /* 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
87 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
88
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
89 /* 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
90 sad = vec_sum4s(t5, sad);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
91
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
92 pix1 += line_size;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
93 pix2 += line_size;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
94 }
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
95 /* 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
96 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
97 sumdiffs = vec_splat(sumdiffs, 3);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
98 vec_ste(sumdiffs, 0, &s);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
99
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
100 return s;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
101 }
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
102
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
103 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
104 {
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
105 int i;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
106 int s __attribute__((aligned(16)));
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
107 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
108 vector unsigned char *tv;
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
109 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
110 vector unsigned int sad;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
111 vector signed int sumdiffs;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
112 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
113
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
114 s = 0;
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
115 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
116
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
117 /*
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
118 Due to the fact that pix3 = pix2 + line_size, the pix3 of one
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
119 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
120 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
121 time around the loop.
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
122 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
123 pix2v: pix2[0]-pix2[15]
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
124 Split the pixel vectors into shorts
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
125 */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
126 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
127 pix2v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[0]));
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
128
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
129 for(i=0;i<h;i++) {
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
130 /*
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
131 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
132 pix1v: pix1[0]-pix1[15]
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
133 pix3v: pix3[0]-pix3[15]
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
134 */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
135 tv = (vector unsigned char *) pix1;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
136 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
137
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
138 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
139 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
140
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
141 /* Calculate the average vector */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
142 avgv = vec_avg(pix2v, pix3v);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
143
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
144 /* 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
145 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
146
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
147 /* 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
148 sad = vec_sum4s(t5, sad);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
149
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
150 pix1 += line_size;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
151 pix2v = pix3v;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
152 pix3 += line_size;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
153
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
154 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
155
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
156 /* 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
157 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
158 sumdiffs = vec_splat(sumdiffs, 3);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
159 vec_ste(sumdiffs, 0, &s);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
160 return s;
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
161 }
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
162
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
163 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
164 {
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
165 int i;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
166 int s __attribute__((aligned(16)));
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
167 uint8_t *pix3 = pix2 + line_size;
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
168 const_vector unsigned char zero = (const_vector unsigned char)vec_splat_u8(0);
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
169 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
170 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
171 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
172 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
173 vector unsigned short pix3lv, pix3hv, pix3ilv, pix3ihv;
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
174 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
175 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
176 vector unsigned int sad;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
177 vector signed int sumdiffs;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
178
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
179 sad = (vector unsigned int)vec_splat_u32(0);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
180
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
181 s = 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 /*
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
184 Due to the fact that pix3 = pix2 + line_size, the pix3 of one
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
185 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
186 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
187 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
188 Read unaligned pixels into our vectors. The vectors are as follows:
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
189 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
190 Split the pixel vectors into shorts
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 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
193 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
194
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
195 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
196 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
197
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
198 pix2hv = (vector unsigned short) vec_mergeh(zero, pix2v);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
199 pix2lv = (vector unsigned short) vec_mergel(zero, pix2v);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
200 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
201 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
202 t1 = vec_add(pix2hv, pix2ihv);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
203 t2 = vec_add(pix2lv, pix2ilv);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
204
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
205 for(i=0;i<h;i++) {
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
206 /*
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
207 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
208 pix1v: pix1[0]-pix1[15]
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
209 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
210 */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
211 tv = (vector unsigned char *) pix1;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
212 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
213
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
214 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
215 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
216
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
217 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
218 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
219
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
220 /*
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
221 Note that Altivec does have vec_avg, but this works on vector pairs
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
222 and rounds up. We could do avg(avg(a,b),avg(c,d)), but the rounding
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
223 would mean that, for example, avg(3,0,0,1) = 2, when it should be 1.
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
224 Instead, we have to split the pixel vectors into vectors of shorts,
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
225 and do the averaging by hand.
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
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
228 /* Split the pixel vectors into shorts */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
229 pix3hv = (vector unsigned short) vec_mergeh(zero, pix3v);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
230 pix3lv = (vector unsigned short) vec_mergel(zero, pix3v);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
231 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
232 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
233
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
234 /* Do the averaging on them */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
235 t3 = vec_add(pix3hv, pix3ihv);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
236 t4 = vec_add(pix3lv, pix3ilv);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
237
884
2cef5c4c0ca6 * altivec and pix_norm patch by Brian Foley
kabi
parents: 878
diff changeset
238 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
239 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
240
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
241 /* 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
242 avgv = vec_pack(avghv, avglv);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
243
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
244 /* 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
245 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
246
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
247 /* 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
248 sad = vec_sum4s(t5, sad);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
249
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
250 pix1 += line_size;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
251 pix3 += line_size;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
252 /* 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
253 t1 = t3;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
254 t2 = t4;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
255 }
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
256 /* 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
257 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
258 sumdiffs = vec_splat(sumdiffs, 3);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
259 vec_ste(sumdiffs, 0, &s);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
260
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
261 return s;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
262 }
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
263
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
264 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
265 {
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
266 int i;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
267 int s __attribute__((aligned(16)));
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
268 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
269 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
270 vector unsigned char t1, t2, t3,t4, t5;
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
271 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
272 vector signed int sumdiffs;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
273
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
274 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
275
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
276
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
277 for(i=0;i<h;i++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
278 /* 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
279 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
280 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
281 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
282 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
283 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
284 t2 = vec_perm(pix2v[0], pix2v[1], perm2);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
285
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
286 /* 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
287 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
288 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
289 t5 = vec_sub(t3, t4);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
290
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
291 /* 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
292 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
293
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
294 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
295 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
296 }
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
297
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
298 /* 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
299 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
300 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
301 vec_ste(sumdiffs, 0, &s);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
302
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
303 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
304 }
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
305
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
306 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
307 {
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
308 int i;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
309 int s __attribute__((aligned(16)));
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
310 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
311 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
312 vector unsigned char t1, t2, t3,t4, t5;
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
313 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
314 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
315
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
316 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
317
f3152eb76f1a altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents: 1064
diff changeset
318 permclear = (vector unsigned char)AVV(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
319
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
320 for(i=0;i<h;i++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
321 /* Read potentially unaligned pixels into t1 and t2
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
322 Since we're reading 16 pixels, and actually only want 8,
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
323 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
324 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
325 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
326 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
327 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
328 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
329 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
330
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
331 /* 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
332 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
333 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
334 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
335
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
336 /* 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
337 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
338
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
339 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
340 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
341 }
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
342
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
343 /* 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
344 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
345 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
346 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
347
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
348 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
349 }
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
350
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
351 int pix_norm1_altivec(uint8_t *pix, int line_size)
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
352 {
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
353 int i;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
354 int s __attribute__((aligned(16)));
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
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 *tv;
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
357 vector unsigned char pixv;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
358 vector unsigned int sv;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
359 vector signed int sum;
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 sv = (vector unsigned int)vec_splat_u32(0);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
362
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
363 s = 0;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
364 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
365 /* 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
366 tv = (vector unsigned char *) pix;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
367 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
368
884
2cef5c4c0ca6 * altivec and pix_norm patch by Brian Foley
kabi
parents: 878
diff changeset
369 /* Square the values, and add them to our sum */
2cef5c4c0ca6 * altivec and pix_norm patch by Brian Foley
kabi
parents: 878
diff changeset
370 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
371
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
372 pix += line_size;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
373 }
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
374 /* 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
375 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
376 sum = vec_splat(sum, 3);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
377 vec_ste(sum, 0, &s);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
378
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
379 return s;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
380 }
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
381
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
382 /**
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
383 * Sum of Squared Errors for a 8x8 block.
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
384 * AltiVec-enhanced.
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
385 * It's the sad8_altivec code above w/ squaring added.
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
386 */
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
387 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
388 {
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
389 int i;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
390 int s __attribute__((aligned(16)));
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
391 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
392 vector unsigned char perm1, perm2, permclear, *pix1v, *pix2v;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
393 vector unsigned char t1, t2, t3,t4, t5;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
394 vector unsigned int sum;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
395 vector signed int sumsqr;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
396
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
397 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
398
f3152eb76f1a altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents: 1064
diff changeset
399 permclear = (vector unsigned char)AVV(255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0);
f3152eb76f1a altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents: 1064
diff changeset
400
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
401
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
402 for(i=0;i<h;i++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
403 /* Read potentially unaligned pixels into t1 and t2
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
404 Since we're reading 16 pixels, and actually only want 8,
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
405 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
406 perm1 = vec_lvsl(0, pix1);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
407 pix1v = (vector unsigned char *) pix1;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
408 perm2 = vec_lvsl(0, pix2);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
409 pix2v = (vector unsigned char *) pix2;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
410 t1 = vec_and(vec_perm(pix1v[0], pix1v[1], perm1), permclear);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
411 t2 = vec_and(vec_perm(pix2v[0], pix2v[1], perm2), permclear);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
412
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
413 /*
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
414 Since we want to use unsigned chars, we can take advantage
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
415 of the fact that abs(a-b)^2 = (a-b)^2.
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
416 */
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
417
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
418 /* Calculate abs differences vector */
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
419 t3 = vec_max(t1, t2);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
420 t4 = vec_min(t1, t2);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
421 t5 = vec_sub(t3, t4);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
422
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
423 /* Square the values and add them to our sum */
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
424 sum = vec_msum(t5, t5, sum);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
425
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
426 pix1 += line_size;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
427 pix2 += line_size;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
428 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
429
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
430 /* Sum up the four partial sums, and put the result into s */
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
431 sumsqr = vec_sums((vector signed int) sum, (vector signed int) zero);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
432 sumsqr = vec_splat(sumsqr, 3);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
433 vec_ste(sumsqr, 0, &s);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
434
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
435 return s;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
436 }
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
437
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
438 /**
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
439 * Sum of Squared Errors for a 16x16 block.
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
440 * AltiVec-enhanced.
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
441 * It's the sad16_altivec code above w/ squaring added.
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
442 */
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
443 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
444 {
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
445 int i;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
446 int s __attribute__((aligned(16)));
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
447 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
448 vector unsigned char perm1, perm2, *pix1v, *pix2v;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
449 vector unsigned char t1, t2, t3,t4, t5;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
450 vector unsigned int sum;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
451 vector signed int sumsqr;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
452
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
453 sum = (vector unsigned int)vec_splat_u32(0);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
454
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
455 for(i=0;i<h;i++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
456 /* Read potentially unaligned pixels into t1 and t2 */
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
457 perm1 = vec_lvsl(0, pix1);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
458 pix1v = (vector unsigned char *) pix1;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
459 perm2 = vec_lvsl(0, pix2);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
460 pix2v = (vector unsigned char *) pix2;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
461 t1 = vec_perm(pix1v[0], pix1v[1], perm1);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
462 t2 = vec_perm(pix2v[0], pix2v[1], perm2);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
463
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
464 /*
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
465 Since we want to use unsigned chars, we can take advantage
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
466 of the fact that abs(a-b)^2 = (a-b)^2.
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
467 */
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
468
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
469 /* Calculate abs differences vector */
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
470 t3 = vec_max(t1, t2);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
471 t4 = vec_min(t1, t2);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
472 t5 = vec_sub(t3, t4);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
473
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
474 /* Square the values and add them to our sum */
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
475 sum = vec_msum(t5, t5, sum);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
476
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
477 pix1 += line_size;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
478 pix2 += line_size;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
479 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
480
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
481 /* Sum up the four partial sums, and put the result into s */
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
482 sumsqr = vec_sums((vector signed int) sum, (vector signed int) zero);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
483 sumsqr = vec_splat(sumsqr, 3);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
484 vec_ste(sumsqr, 0, &s);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
485
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
486 return s;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
487 }
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
488
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1033
diff changeset
489 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
490 {
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
491 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
492 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
493 vector unsigned char t1;
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
494 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
495 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
496
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
497 int i;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
498 int s __attribute__((aligned(16)));
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
499
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
500 sad = (vector unsigned int)vec_splat_u32(0);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
501
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
502 for (i = 0; i < 16; i++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
503 /* 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
504 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
505 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
506 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
507
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
508 /* 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
509 sad = vec_sum4s(t1, sad);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
510
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
511 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
512 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
513
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
514 /* 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
515 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
516 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
517 vec_ste(sumdiffs, 0, &s);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
518
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
519 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
520 }
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
521
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1033
diff changeset
522 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
523 {
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
524 int i;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
525 vector unsigned char perm, bytes, *pixv;
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
526 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
527 vector signed short shorts;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
528
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
529 for(i=0;i<8;i++)
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 // Read potentially unaligned pixels.
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
532 // 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
533 // but we simply ignore the extras.
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
534 perm = vec_lvsl(0, pixels);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
535 pixv = (vector unsigned char *) pixels;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
536 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
537
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
538 // convert the bytes into shorts
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
539 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
540
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
541 // 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
542 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
543
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
544 pixels += line_size;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
545 }
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
546 }
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
547
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1033
diff changeset
548 void diff_pixels_altivec(DCTELEM *restrict block, const uint8_t *s1,
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1033
diff changeset
549 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
550 {
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
551 int i;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
552 vector unsigned char perm, bytes, *pixv;
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
553 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
554 vector signed short shorts1, shorts2;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
555
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
556 for(i=0;i<4;i++)
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 // Read potentially unaligned pixels
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
559 // 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
560 // but we simply ignore the extras.
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
561 perm = vec_lvsl(0, s1);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
562 pixv = (vector unsigned char *) s1;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
563 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
564
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
565 // convert the bytes into shorts
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
566 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
567
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
568 // 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
569 perm = vec_lvsl(0, s2);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
570 pixv = (vector unsigned char *) s2;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
571 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
572
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
573 // convert the bytes into shorts
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
574 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
575
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
576 // Do the subtraction
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
577 shorts1 = vec_sub(shorts1, shorts2);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
578
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
579 // 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
580 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
581
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
582 s1 += stride;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
583 s2 += stride;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
584 block += 8;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
585
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
586
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
587 // 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
588 // unroll.
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
589
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
590 // Read potentially unaligned pixels
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
591 // 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
592 // but we simply ignore the extras.
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
593 perm = vec_lvsl(0, s1);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
594 pixv = (vector unsigned char *) s1;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
595 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
596
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
597 // convert the bytes into shorts
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
598 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
599
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
600 // 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
601 perm = vec_lvsl(0, s2);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
602 pixv = (vector unsigned char *) s2;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
603 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
604
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
605 // convert the bytes into shorts
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
606 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
607
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
608 // Do the subtraction
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
609 shorts1 = vec_sub(shorts1, shorts2);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
610
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
611 // 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
612 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
613
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
614 s1 += stride;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
615 s2 += stride;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
616 block += 8;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
617 }
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
618 }
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
619
995
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
620 void add_bytes_altivec(uint8_t *dst, uint8_t *src, int w) {
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
621 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
622 register vector unsigned char vdst, vsrc;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
623
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
624 /* dst and src are 16 bytes-aligned (guaranteed) */
3968
c86c7a54ba92 add_bytes passes tests
lu_zero
parents: 3947
diff changeset
625 for(i = 0 ; (i + 15) < w ; i+=16)
995
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
626 {
3968
c86c7a54ba92 add_bytes passes tests
lu_zero
parents: 3947
diff changeset
627 vdst = vec_ld(i, (unsigned char*)dst);
c86c7a54ba92 add_bytes passes tests
lu_zero
parents: 3947
diff changeset
628 vsrc = vec_ld(i, (unsigned char*)src);
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
629 vdst = vec_add(vsrc, vdst);
3968
c86c7a54ba92 add_bytes passes tests
lu_zero
parents: 3947
diff changeset
630 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
631 }
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
632 /* if w is not a multiple of 16 */
995
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
633 for (; (i < w) ; i++)
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
634 {
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
635 dst[i] = src[i];
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
636 }
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
637 }
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
638
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
639 /* 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
640 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
641 {
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
642 POWERPC_PERF_DECLARE(altivec_put_pixels16_num, 1);
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
643 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
644 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
645 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
646 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
647
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
648 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
649 int i;
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
650 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
651 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
652 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
653
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
654 POWERPC_PERF_START_COUNT(altivec_put_pixels16_num, 1);
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
655 // 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
656 // 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
657 // 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
658 // -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
659 // 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
660 #if 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
661 for(i=0; i<h; i++) {
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
662 pixelsv1 = vec_ld(0, (unsigned char*)pixels);
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
663 pixelsv2 = vec_ld(16, (unsigned char*)pixels);
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
664 vec_st(vec_perm(pixelsv1, pixelsv2, perm),
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
665 0, (unsigned char*)block);
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
666 pixels+=line_size;
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
667 block +=line_size;
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
668 }
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
669 #else
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
670 for(i=0; i<h; i+=4) {
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
671 pixelsv1 = vec_ld(0, (unsigned char*)pixels);
3533
0937cc91b574 avoid possible segfault situations
lu_zero
parents: 3346
diff changeset
672 pixelsv2 = vec_ld(15, (unsigned char*)pixels);
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
673 pixelsv1B = vec_ld(line_size, (unsigned char*)pixels);
3533
0937cc91b574 avoid possible segfault situations
lu_zero
parents: 3346
diff changeset
674 pixelsv2B = vec_ld(15 + line_size, (unsigned char*)pixels);
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
675 pixelsv1C = vec_ld(line_size_2, (unsigned char*)pixels);
3533
0937cc91b574 avoid possible segfault situations
lu_zero
parents: 3346
diff changeset
676 pixelsv2C = vec_ld(15 + line_size_2, (unsigned char*)pixels);
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
677 pixelsv1D = vec_ld(line_size_3, (unsigned char*)pixels);
3533
0937cc91b574 avoid possible segfault situations
lu_zero
parents: 3346
diff changeset
678 pixelsv2D = vec_ld(15 + line_size_3, (unsigned char*)pixels);
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
679 vec_st(vec_perm(pixelsv1, pixelsv2, perm),
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
680 0, (unsigned char*)block);
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
681 vec_st(vec_perm(pixelsv1B, pixelsv2B, perm),
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
682 line_size, (unsigned char*)block);
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
683 vec_st(vec_perm(pixelsv1C, pixelsv2C, perm),
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
684 line_size_2, (unsigned char*)block);
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
685 vec_st(vec_perm(pixelsv1D, pixelsv2D, perm),
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
686 line_size_3, (unsigned char*)block);
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
687 pixels+=line_size_4;
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
688 block +=line_size_4;
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
689 }
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
690 #endif
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
691 POWERPC_PERF_STOP_COUNT(altivec_put_pixels16_num, 1);
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
692 }
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
693
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
694 /* 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
695 #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
696 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
697 {
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
698 POWERPC_PERF_DECLARE(altivec_avg_pixels16_num, 1);
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
699 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
700 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
701 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
702
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
703 POWERPC_PERF_START_COUNT(altivec_avg_pixels16_num, 1);
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
704
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
705 for(i=0; i<h; i++) {
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
706 pixelsv1 = vec_ld(0, (unsigned char*)pixels);
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
707 pixelsv2 = vec_ld(16, (unsigned char*)pixels);
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
708 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
709 pixelsv = vec_perm(pixelsv1, pixelsv2, perm);
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
710 blockv = vec_avg(blockv,pixelsv);
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
711 vec_st(blockv, 0, (unsigned char*)block);
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
712 pixels+=line_size;
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
713 block +=line_size;
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
714 }
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
715
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
716 POWERPC_PERF_STOP_COUNT(altivec_avg_pixels16_num, 1);
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
717 }
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
718
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
719 /* next one assumes that ((line_size % 8) == 0) */
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
720 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
721 {
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
722 POWERPC_PERF_DECLARE(altivec_avg_pixels8_num, 1);
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
723 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
724 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
725
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
726 POWERPC_PERF_START_COUNT(altivec_avg_pixels8_num, 1);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
727
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
728 for (i = 0; i < h; 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
729 /*
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
730 block is 8 bytes-aligned, so we're either in the
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
731 left block (16 bytes-aligned) or in the right block (not)
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
732 */
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
733 int rightside = ((unsigned long)block & 0x0000000F);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
734
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
735 blockv = vec_ld(0, block);
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
736 pixelsv1 = vec_ld(0, (unsigned char*)pixels);
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
737 pixelsv2 = vec_ld(16, (unsigned char*)pixels);
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
738 pixelsv = vec_perm(pixelsv1, pixelsv2, vec_lvsl(0, pixels));
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
739
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
740 if (rightside)
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
741 {
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
742 pixelsv = vec_perm(blockv, pixelsv, vcprm(0,1,s0,s1));
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
743 }
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
744 else
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
745 {
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
746 pixelsv = vec_perm(blockv, pixelsv, vcprm(s0,s1,2,3));
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
747 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
748
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
749 blockv = vec_avg(blockv, pixelsv);
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
750
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
751 vec_st(blockv, 0, block);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
752
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
753 pixels += line_size;
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
754 block += line_size;
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
755 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
756
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
757 POWERPC_PERF_STOP_COUNT(altivec_avg_pixels8_num, 1);
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
758 }
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
759
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
760 /* next one assumes that ((line_size % 8) == 0) */
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
761 void put_pixels8_xy2_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
762 {
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
763 POWERPC_PERF_DECLARE(altivec_put_pixels8_xy2_num, 1);
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
764 register 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
765 register vector unsigned char
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
766 pixelsv1, pixelsv2,
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
767 pixelsavg;
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
768 register vector unsigned char
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
769 blockv, temp1, temp2;
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
770 register vector unsigned short
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
771 pixelssum1, pixelssum2, temp3;
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
772 register const_vector unsigned char vczero = (const_vector unsigned char)vec_splat_u8(0);
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
773 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
774
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
775 temp1 = vec_ld(0, pixels);
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
776 temp2 = vec_ld(16, pixels);
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
777 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
778 if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F)
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
779 {
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
780 pixelsv2 = temp2;
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
781 }
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
782 else
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
783 {
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
784 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
785 }
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
786 pixelsv1 = vec_mergeh(vczero, pixelsv1);
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
787 pixelsv2 = vec_mergeh(vczero, pixelsv2);
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
788 pixelssum1 = vec_add((vector unsigned short)pixelsv1,
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
789 (vector unsigned short)pixelsv2);
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
790 pixelssum1 = vec_add(pixelssum1, vctwo);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
791
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
792 POWERPC_PERF_START_COUNT(altivec_put_pixels8_xy2_num, 1);
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
793 for (i = 0; i < h ; 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
794 int rightside = ((unsigned long)block & 0x0000000F);
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
795 blockv = vec_ld(0, block);
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
796
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
797 temp1 = vec_ld(line_size, pixels);
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
798 temp2 = vec_ld(line_size + 16, pixels);
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
799 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
800 if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F)
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
801 {
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
802 pixelsv2 = temp2;
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
803 }
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
804 else
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
805 {
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
806 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
807 }
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
808
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
809 pixelsv1 = vec_mergeh(vczero, pixelsv1);
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
810 pixelsv2 = vec_mergeh(vczero, pixelsv2);
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
811 pixelssum2 = vec_add((vector unsigned short)pixelsv1,
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
812 (vector unsigned short)pixelsv2);
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
813 temp3 = vec_add(pixelssum1, pixelssum2);
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
814 temp3 = vec_sra(temp3, vctwo);
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
815 pixelssum1 = vec_add(pixelssum2, vctwo);
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
816 pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
817
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
818 if (rightside)
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
819 {
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
820 blockv = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1));
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
821 }
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
822 else
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
823 {
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
824 blockv = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3));
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
825 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
826
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
827 vec_st(blockv, 0, block);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
828
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
829 block += line_size;
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
830 pixels += line_size;
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
831 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
832
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
833 POWERPC_PERF_STOP_COUNT(altivec_put_pixels8_xy2_num, 1);
995
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
834 }
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
835
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
836 /* next one assumes that ((line_size % 8) == 0) */
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
837 void put_no_rnd_pixels8_xy2_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
838 {
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
839 POWERPC_PERF_DECLARE(altivec_put_no_rnd_pixels8_xy2_num, 1);
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
840 register int i;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
841 register vector unsigned char
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
842 pixelsv1, pixelsv2,
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
843 pixelsavg;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
844 register vector unsigned char
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
845 blockv, temp1, temp2;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
846 register vector unsigned short
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
847 pixelssum1, pixelssum2, temp3;
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
848 register const_vector unsigned char vczero = (const_vector unsigned char)vec_splat_u8(0);
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
849 register const_vector unsigned short vcone = (const_vector unsigned short)vec_splat_u16(1);
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
850 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
851
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
852 temp1 = vec_ld(0, pixels);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
853 temp2 = vec_ld(16, pixels);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
854 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
855 if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F)
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
856 {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
857 pixelsv2 = temp2;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
858 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
859 else
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
860 {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
861 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
862 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
863 pixelsv1 = vec_mergeh(vczero, pixelsv1);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
864 pixelsv2 = vec_mergeh(vczero, pixelsv2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
865 pixelssum1 = vec_add((vector unsigned short)pixelsv1,
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
866 (vector unsigned short)pixelsv2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
867 pixelssum1 = vec_add(pixelssum1, vcone);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
868
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
869 POWERPC_PERF_START_COUNT(altivec_put_no_rnd_pixels8_xy2_num, 1);
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
870 for (i = 0; i < h ; i++) {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
871 int rightside = ((unsigned long)block & 0x0000000F);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
872 blockv = vec_ld(0, block);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
873
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
874 temp1 = vec_ld(line_size, pixels);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
875 temp2 = vec_ld(line_size + 16, pixels);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
876 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
877 if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F)
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
878 {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
879 pixelsv2 = temp2;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
880 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
881 else
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
882 {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
883 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
884 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
885
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
886 pixelsv1 = vec_mergeh(vczero, pixelsv1);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
887 pixelsv2 = vec_mergeh(vczero, pixelsv2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
888 pixelssum2 = vec_add((vector unsigned short)pixelsv1,
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
889 (vector unsigned short)pixelsv2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
890 temp3 = vec_add(pixelssum1, pixelssum2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
891 temp3 = vec_sra(temp3, vctwo);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
892 pixelssum1 = vec_add(pixelssum2, vcone);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
893 pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
894
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
895 if (rightside)
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
896 {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
897 blockv = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1));
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 else
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
900 {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
901 blockv = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3));
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
902 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
903
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
904 vec_st(blockv, 0, block);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
905
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
906 block += line_size;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
907 pixels += line_size;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
908 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
909
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
910 POWERPC_PERF_STOP_COUNT(altivec_put_no_rnd_pixels8_xy2_num, 1);
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
911 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
912
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
913 /* next one assumes that ((line_size % 16) == 0) */
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
914 void put_pixels16_xy2_altivec(uint8_t * block, const uint8_t * pixels, int line_size, int h)
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
915 {
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
916 POWERPC_PERF_DECLARE(altivec_put_pixels16_xy2_num, 1);
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
917 register int i;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
918 register vector unsigned char
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
919 pixelsv1, pixelsv2, pixelsv3, pixelsv4;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
920 register vector unsigned char
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
921 blockv, temp1, temp2;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
922 register vector unsigned short
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
923 pixelssum1, pixelssum2, temp3,
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
924 pixelssum3, pixelssum4, temp4;
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
925 register const_vector unsigned char vczero = (const_vector unsigned char)vec_splat_u8(0);
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
926 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
927
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
928 POWERPC_PERF_START_COUNT(altivec_put_pixels16_xy2_num, 1);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
929
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
930 temp1 = vec_ld(0, pixels);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
931 temp2 = vec_ld(16, pixels);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
932 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
933 if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F)
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
934 {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
935 pixelsv2 = temp2;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
936 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
937 else
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
938 {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
939 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
940 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
941 pixelsv3 = vec_mergel(vczero, pixelsv1);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
942 pixelsv4 = vec_mergel(vczero, pixelsv2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
943 pixelsv1 = vec_mergeh(vczero, pixelsv1);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
944 pixelsv2 = vec_mergeh(vczero, pixelsv2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
945 pixelssum3 = vec_add((vector unsigned short)pixelsv3,
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
946 (vector unsigned short)pixelsv4);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
947 pixelssum3 = vec_add(pixelssum3, vctwo);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
948 pixelssum1 = vec_add((vector unsigned short)pixelsv1,
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
949 (vector unsigned short)pixelsv2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
950 pixelssum1 = vec_add(pixelssum1, vctwo);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
951
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
952 for (i = 0; i < h ; i++) {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
953 blockv = vec_ld(0, block);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
954
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
955 temp1 = vec_ld(line_size, pixels);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
956 temp2 = vec_ld(line_size + 16, pixels);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
957 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
958 if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F)
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
959 {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
960 pixelsv2 = temp2;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
961 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
962 else
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
963 {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
964 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
965 }
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 pixelsv3 = vec_mergel(vczero, pixelsv1);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
968 pixelsv4 = vec_mergel(vczero, pixelsv2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
969 pixelsv1 = vec_mergeh(vczero, pixelsv1);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
970 pixelsv2 = vec_mergeh(vczero, pixelsv2);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
971
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
972 pixelssum4 = vec_add((vector unsigned short)pixelsv3,
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
973 (vector unsigned short)pixelsv4);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
974 pixelssum2 = vec_add((vector unsigned short)pixelsv1,
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
975 (vector unsigned short)pixelsv2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
976 temp4 = vec_add(pixelssum3, pixelssum4);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
977 temp4 = vec_sra(temp4, vctwo);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
978 temp3 = vec_add(pixelssum1, pixelssum2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
979 temp3 = vec_sra(temp3, vctwo);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
980
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
981 pixelssum3 = vec_add(pixelssum4, vctwo);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
982 pixelssum1 = vec_add(pixelssum2, vctwo);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
983
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
984 blockv = vec_packsu(temp3, temp4);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
985
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
986 vec_st(blockv, 0, block);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
987
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
988 block += line_size;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
989 pixels += line_size;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
990 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
991
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
992 POWERPC_PERF_STOP_COUNT(altivec_put_pixels16_xy2_num, 1);
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
993 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
994
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
995 /* next one assumes that ((line_size % 16) == 0) */
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
996 void put_no_rnd_pixels16_xy2_altivec(uint8_t * block, const uint8_t * pixels, int line_size, int h)
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
997 {
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
998 POWERPC_PERF_DECLARE(altivec_put_no_rnd_pixels16_xy2_num, 1);
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
999 register int i;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1000 register vector unsigned char
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1001 pixelsv1, pixelsv2, pixelsv3, pixelsv4;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1002 register vector unsigned char
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1003 blockv, temp1, temp2;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1004 register vector unsigned short
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1005 pixelssum1, pixelssum2, temp3,
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1006 pixelssum3, pixelssum4, temp4;
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
1007 register const_vector unsigned char vczero = (const_vector unsigned char)vec_splat_u8(0);
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
1008 register const_vector unsigned short vcone = (const_vector unsigned short)vec_splat_u16(1);
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
1009 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
1010
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
1011 POWERPC_PERF_START_COUNT(altivec_put_no_rnd_pixels16_xy2_num, 1);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1012
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1013 temp1 = vec_ld(0, pixels);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1014 temp2 = vec_ld(16, pixels);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1015 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1016 if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F)
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1017 {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1018 pixelsv2 = temp2;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1019 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1020 else
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1021 {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1022 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1023 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1024 pixelsv3 = vec_mergel(vczero, pixelsv1);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1025 pixelsv4 = vec_mergel(vczero, pixelsv2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1026 pixelsv1 = vec_mergeh(vczero, pixelsv1);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1027 pixelsv2 = vec_mergeh(vczero, pixelsv2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1028 pixelssum3 = vec_add((vector unsigned short)pixelsv3,
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1029 (vector unsigned short)pixelsv4);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1030 pixelssum3 = vec_add(pixelssum3, vcone);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1031 pixelssum1 = vec_add((vector unsigned short)pixelsv1,
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1032 (vector unsigned short)pixelsv2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1033 pixelssum1 = vec_add(pixelssum1, vcone);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1034
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1035 for (i = 0; i < h ; i++) {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1036 blockv = vec_ld(0, block);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1037
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1038 temp1 = vec_ld(line_size, pixels);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1039 temp2 = vec_ld(line_size + 16, pixels);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1040 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1041 if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F)
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1042 {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1043 pixelsv2 = temp2;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1044 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1045 else
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1046 {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1047 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1048 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1049
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1050 pixelsv3 = vec_mergel(vczero, pixelsv1);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1051 pixelsv4 = vec_mergel(vczero, pixelsv2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1052 pixelsv1 = vec_mergeh(vczero, pixelsv1);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1053 pixelsv2 = vec_mergeh(vczero, pixelsv2);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1054
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1055 pixelssum4 = vec_add((vector unsigned short)pixelsv3,
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1056 (vector unsigned short)pixelsv4);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1057 pixelssum2 = vec_add((vector unsigned short)pixelsv1,
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1058 (vector unsigned short)pixelsv2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1059 temp4 = vec_add(pixelssum3, pixelssum4);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1060 temp4 = vec_sra(temp4, vctwo);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1061 temp3 = vec_add(pixelssum1, pixelssum2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1062 temp3 = vec_sra(temp3, vctwo);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1063
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1064 pixelssum3 = vec_add(pixelssum4, vcone);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1065 pixelssum1 = vec_add(pixelssum2, vcone);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1066
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1067 blockv = vec_packsu(temp3, temp4);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1068
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1069 vec_st(blockv, 0, block);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1070
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1071 block += line_size;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1072 pixels += line_size;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1073 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1074
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
1075 POWERPC_PERF_STOP_COUNT(altivec_put_no_rnd_pixels16_xy2_num, 1);
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1076 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1077
1949
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1078 int hadamard8_diff8x8_altivec(/*MpegEncContext*/ void *s, uint8_t *dst, uint8_t *src, int stride, int h){
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1079 POWERPC_PERF_DECLARE(altivec_hadamard8_diff8x8_num, 1);
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1080 int sum;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1081 register const_vector unsigned char vzero =
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1082 (const_vector unsigned char)vec_splat_u8(0);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1083 register vector signed short temp0, temp1, temp2, temp3, temp4,
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1084 temp5, temp6, temp7;
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1085 POWERPC_PERF_START_COUNT(altivec_hadamard8_diff8x8_num, 1);
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1086 {
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1087 register const_vector signed short vprod1 =(const_vector signed short)
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1088 AVV( 1,-1, 1,-1, 1,-1, 1,-1);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1089 register const_vector signed short vprod2 =(const_vector signed short)
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1090 AVV( 1, 1,-1,-1, 1, 1,-1,-1);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1091 register const_vector signed short vprod3 =(const_vector signed short)
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1092 AVV( 1, 1, 1, 1,-1,-1,-1,-1);
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1093 register const_vector unsigned char perm1 = (const_vector unsigned char)
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1094 AVV(0x02, 0x03, 0x00, 0x01, 0x06, 0x07, 0x04, 0x05,
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1095 0x0A, 0x0B, 0x08, 0x09, 0x0E, 0x0F, 0x0C, 0x0D);
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1096 register const_vector unsigned char perm2 = (const_vector unsigned char)
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1097 AVV(0x04, 0x05, 0x06, 0x07, 0x00, 0x01, 0x02, 0x03,
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1098 0x0C, 0x0D, 0x0E, 0x0F, 0x08, 0x09, 0x0A, 0x0B);
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1099 register const_vector unsigned char perm3 = (const_vector unsigned char)
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1100 AVV(0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F,
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1101 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
1102
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1103 #define ONEITERBUTTERFLY(i, res) \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1104 { \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1105 register vector unsigned char src1, src2, srcO; \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1106 register vector unsigned char dst1, dst2, dstO; \
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1107 register vector signed short srcV, dstV; \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1108 register vector signed short but0, but1, but2, op1, op2, op3; \
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1109 src1 = vec_ld(stride * i, src); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1110 if ((((stride * i) + (unsigned long)src) & 0x0000000F) > 8) \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1111 src2 = vec_ld((stride * i) + 16, src); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1112 srcO = vec_perm(src1, src2, vec_lvsl(stride * i, src)); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1113 dst1 = vec_ld(stride * i, dst); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1114 if ((((stride * i) + (unsigned long)dst) & 0x0000000F) > 8) \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1115 dst2 = vec_ld((stride * i) + 16, dst); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1116 dstO = vec_perm(dst1, dst2, vec_lvsl(stride * i, dst)); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1117 /* promote the unsigned chars to signed shorts */ \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1118 /* we're in the 8x8 function, we only care for the first 8 */ \
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1119 srcV = \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1120 (vector signed short)vec_mergeh((vector signed char)vzero, \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1121 (vector signed char)srcO); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1122 dstV = \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1123 (vector signed short)vec_mergeh((vector signed char)vzero, \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1124 (vector signed char)dstO); \
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1125 /* substractions inside the first butterfly */ \
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1126 but0 = vec_sub(srcV, dstV); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1127 op1 = vec_perm(but0, but0, perm1); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1128 but1 = vec_mladd(but0, vprod1, op1); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1129 op2 = vec_perm(but1, but1, perm2); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1130 but2 = vec_mladd(but1, vprod2, op2); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1131 op3 = vec_perm(but2, but2, perm3); \
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1132 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
1133 }
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1134 ONEITERBUTTERFLY(0, temp0);
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1135 ONEITERBUTTERFLY(1, temp1);
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1136 ONEITERBUTTERFLY(2, temp2);
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1137 ONEITERBUTTERFLY(3, temp3);
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1138 ONEITERBUTTERFLY(4, temp4);
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1139 ONEITERBUTTERFLY(5, temp5);
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1140 ONEITERBUTTERFLY(6, temp6);
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1141 ONEITERBUTTERFLY(7, temp7);
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1142 }
1949
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1143 #undef ONEITERBUTTERFLY
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1144 {
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1145 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
1146 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
1147 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
1148 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
1149 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
1150 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
1151 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
1152 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
1153 register vector signed short line7 = vec_sub(temp6, temp7);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1154
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1155 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
1156 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
1157 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
1158 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
1159 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
1160 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
1161 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
1162 register vector signed short line7B = vec_sub(line5, line7);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1163
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1164 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
1165 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
1166 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
1167 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
1168 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
1169 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
1170 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
1171 register vector signed short line7C = vec_sub(line3B, line7B);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1172
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1173 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
1174 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
1175 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
1176 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
1177 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
1178 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
1179 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
1180 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
1181 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
1182 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
1183 vec_ste(vsum, 0, &sum);
1949
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1184 }
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1185 POWERPC_PERF_STOP_COUNT(altivec_hadamard8_diff8x8_num, 1);
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1186 return sum;
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1187 }
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1188
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 16x8 works with 16 elements ; it allows to avoid replicating
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1191 loads, and give the compiler more rooms for scheduling.
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1192 It's only used from inside hadamard8_diff16_altivec.
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1193
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1194 Unfortunately, it seems gcc-3.3 is a bit dumb, and
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1195 the compiled code has a LOT of spill code, it seems
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1196 gcc (unlike xlc) cannot keep everything in registers
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1197 by itself. The following code include hand-made
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1198 registers allocation. It's not clean, but on
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1199 a 7450 the resulting code is much faster (best case
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1200 fall from 700+ cycles to 550).
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1201
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1202 xlc doesn't add spill code, but it doesn't know how to
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1203 schedule for the 7450, and its code isn't much faster than
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1204 gcc-3.3 on the 7450 (but uses 25% less instructions...)
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1205
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1206 On the 970, the hand-made RA is still a win (arount 690
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1207 vs. around 780), but xlc goes to around 660 on the
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1208 regular C code...
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1209 */
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1210
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1211 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
1212 int sum;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1213 register vector signed short
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1214 temp0 REG_v(v0),
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1215 temp1 REG_v(v1),
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1216 temp2 REG_v(v2),
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1217 temp3 REG_v(v3),
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1218 temp4 REG_v(v4),
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1219 temp5 REG_v(v5),
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1220 temp6 REG_v(v6),
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1221 temp7 REG_v(v7);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1222 register vector signed short
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1223 temp0S REG_v(v8),
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1224 temp1S REG_v(v9),
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1225 temp2S REG_v(v10),
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1226 temp3S REG_v(v11),
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1227 temp4S REG_v(v12),
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1228 temp5S REG_v(v13),
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1229 temp6S REG_v(v14),
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1230 temp7S REG_v(v15);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1231 register const_vector unsigned char vzero REG_v(v31)=
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1232 (const_vector unsigned char)vec_splat_u8(0);
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1233 {
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1234 register const_vector signed short vprod1 REG_v(v16)=
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1235 (const_vector signed short)AVV( 1,-1, 1,-1, 1,-1, 1,-1);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1236 register const_vector signed short vprod2 REG_v(v17)=
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1237 (const_vector signed short)AVV( 1, 1,-1,-1, 1, 1,-1,-1);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1238 register const_vector signed short vprod3 REG_v(v18)=
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1239 (const_vector signed short)AVV( 1, 1, 1, 1,-1,-1,-1,-1);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1240 register const_vector unsigned char perm1 REG_v(v19)=
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1241 (const_vector unsigned char)
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1242 AVV(0x02, 0x03, 0x00, 0x01, 0x06, 0x07, 0x04, 0x05,
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1243 0x0A, 0x0B, 0x08, 0x09, 0x0E, 0x0F, 0x0C, 0x0D);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1244 register const_vector unsigned char perm2 REG_v(v20)=
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1245 (const_vector unsigned char)
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1246 AVV(0x04, 0x05, 0x06, 0x07, 0x00, 0x01, 0x02, 0x03,
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1247 0x0C, 0x0D, 0x0E, 0x0F, 0x08, 0x09, 0x0A, 0x0B);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1248 register const_vector unsigned char perm3 REG_v(v21)=
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1249 (const_vector unsigned char)
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1250 AVV(0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F,
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1251 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
1252
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1253 #define ONEITERBUTTERFLY(i, res1, res2) \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1254 { \
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1255 register vector unsigned char src1 REG_v(v22), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1256 src2 REG_v(v23), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1257 dst1 REG_v(v24), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1258 dst2 REG_v(v25), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1259 srcO REG_v(v22), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1260 dstO REG_v(v23); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1261 \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1262 register vector signed short srcV REG_v(v24), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1263 dstV REG_v(v25), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1264 srcW REG_v(v26), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1265 dstW REG_v(v27), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1266 but0 REG_v(v28), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1267 but0S REG_v(v29), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1268 op1 REG_v(v30), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1269 but1 REG_v(v22), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1270 op1S REG_v(v23), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1271 but1S REG_v(v24), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1272 op2 REG_v(v25), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1273 but2 REG_v(v26), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1274 op2S REG_v(v27), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1275 but2S REG_v(v28), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1276 op3 REG_v(v29), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1277 op3S REG_v(v30); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1278 \
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1279 src1 = vec_ld(stride * i, src); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1280 src2 = vec_ld((stride * i) + 16, src); \
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1281 srcO = vec_perm(src1, src2, vec_lvsl(stride * i, src)); \
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1282 dst1 = vec_ld(stride * i, dst); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1283 dst2 = vec_ld((stride * i) + 16, dst); \
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1284 dstO = vec_perm(dst1, dst2, vec_lvsl(stride * i, dst)); \
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1285 /* promote the unsigned chars to signed shorts */ \
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1286 srcV = \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1287 (vector signed short)vec_mergeh((vector signed char)vzero, \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1288 (vector signed char)srcO); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1289 dstV = \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1290 (vector signed short)vec_mergeh((vector signed char)vzero, \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1291 (vector signed char)dstO); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1292 srcW = \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1293 (vector signed short)vec_mergel((vector signed char)vzero, \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1294 (vector signed char)srcO); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1295 dstW = \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1296 (vector signed short)vec_mergel((vector signed char)vzero, \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1297 (vector signed char)dstO); \
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1298 /* substractions inside the first butterfly */ \
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1299 but0 = vec_sub(srcV, dstV); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1300 but0S = vec_sub(srcW, dstW); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1301 op1 = vec_perm(but0, but0, perm1); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1302 but1 = vec_mladd(but0, vprod1, op1); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1303 op1S = vec_perm(but0S, but0S, perm1); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1304 but1S = vec_mladd(but0S, vprod1, op1S); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1305 op2 = vec_perm(but1, but1, perm2); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1306 but2 = vec_mladd(but1, vprod2, op2); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1307 op2S = vec_perm(but1S, but1S, perm2); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1308 but2S = vec_mladd(but1S, vprod2, op2S); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1309 op3 = vec_perm(but2, but2, perm3); \
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1310 res1 = vec_mladd(but2, vprod3, op3); \
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1311 op3S = vec_perm(but2S, but2S, perm3); \
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1312 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
1313 }
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1314 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
1315 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
1316 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
1317 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
1318 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
1319 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
1320 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
1321 ONEITERBUTTERFLY(7, temp7, temp7S);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1322 }
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1323 #undef ONEITERBUTTERFLY
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1324 {
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1325 register vector signed int vsum;
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1326 register vector signed short line0S, line1S, line2S, line3S, line4S,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1327 line5S, line6S, line7S, line0BS,line2BS,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1328 line1BS,line3BS,line4BS,line6BS,line5BS,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1329 line7BS,line0CS,line4CS,line1CS,line5CS,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1330 line2CS,line6CS,line3CS,line7CS;
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1331
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1332 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
1333 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
1334 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
1335 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
1336 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
1337 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
1338 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
1339 register vector signed short line7 = vec_sub(temp6, temp7);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1340
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1341 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
1342 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
1343 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
1344 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
1345 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
1346 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
1347 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
1348 register vector signed short line7B = vec_sub(line5, line7);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1349
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1350 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
1351 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
1352 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
1353 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
1354 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
1355 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
1356 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
1357 register vector signed short line7C = vec_sub(line3B, line7B);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1358
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1359 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
1360 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
1361 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
1362 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
1363 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
1364 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
1365 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
1366 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
1367
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1368 line0S = vec_add(temp0S, temp1S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1369 line1S = vec_sub(temp0S, temp1S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1370 line2S = vec_add(temp2S, temp3S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1371 line3S = vec_sub(temp2S, temp3S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1372 line4S = vec_add(temp4S, temp5S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1373 line5S = vec_sub(temp4S, temp5S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1374 line6S = vec_add(temp6S, temp7S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1375 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
1376
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1377 line0BS = vec_add(line0S, line2S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1378 line2BS = vec_sub(line0S, line2S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1379 line1BS = vec_add(line1S, line3S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1380 line3BS = vec_sub(line1S, line3S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1381 line4BS = vec_add(line4S, line6S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1382 line6BS = vec_sub(line4S, line6S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1383 line5BS = vec_add(line5S, line7S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1384 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
1385
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1386 line0CS = vec_add(line0BS, line4BS);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1387 line4CS = vec_sub(line0BS, line4BS);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1388 line1CS = vec_add(line1BS, line5BS);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1389 line5CS = vec_sub(line1BS, line5BS);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1390 line2CS = vec_add(line2BS, line6BS);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1391 line6CS = vec_sub(line2BS, line6BS);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1392 line3CS = vec_add(line3BS, line7BS);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1393 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
1394
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1395 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
1396 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
1397 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
1398 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
1399 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
1400 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
1401 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
1402 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
1403 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
1404 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
1405 vec_ste(vsum, 0, &sum);
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1406 }
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1407 return sum;
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1408 }
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1409
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1410 int hadamard8_diff16_altivec(/*MpegEncContext*/ void *s, uint8_t *dst, uint8_t *src, int stride, int h){
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1411 POWERPC_PERF_DECLARE(altivec_hadamard8_diff16_num, 1);
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1412 int score;
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1413 POWERPC_PERF_START_COUNT(altivec_hadamard8_diff16_num, 1);
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1414 score = hadamard8_diff16x8_altivec(s, dst, src, stride, 8);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1415 if (h==16) {
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1416 dst += 8*stride;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1417 src += 8*stride;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1418 score += hadamard8_diff16x8_altivec(s, dst, src, stride, 8);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1419 }
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1420 POWERPC_PERF_STOP_COUNT(altivec_hadamard8_diff16_num, 1);
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1421 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
1422 }
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1423
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
1424 int has_altivec(void)
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
1425 {
2286
0f937b191cc3 Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents: 2068
diff changeset
1426 #ifdef __AMIGAOS4__
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1427 ULONG result = 0;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1428 extern struct ExecIFace *IExec;
2286
0f937b191cc3 Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents: 2068
diff changeset
1429
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1430 IExec->GetCPUInfoTags(GCIT_VectorUnit, &result, TAG_DONE);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1431 if (result == VECTORTYPE_ALTIVEC) return 1;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1432 return 0;
2286
0f937b191cc3 Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents: 2068
diff changeset
1433 #else /* __AMIGAOS4__ */
0f937b191cc3 Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents: 2068
diff changeset
1434
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1435 #ifdef CONFIG_DARWIN
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
1436 int sels[2] = {CTL_HW, HW_VECTORUNIT};
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
1437 int has_vu = 0;
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
1438 size_t len = sizeof(has_vu);
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
1439 int err;
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
1440
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
1441 err = sysctl(sels, 2, &has_vu, &len, NULL, 0);
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
1442
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
1443 if (err == 0) return (has_vu != 0);
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1444 #else /* CONFIG_DARWIN */
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1445 /* no Darwin, do it the brute-force way */
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1446 /* this is borrowed from the libmpeg2 library */
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1447 {
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1448 signal (SIGILL, sigill_handler);
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1449 if (sigsetjmp (jmpbuf, 1)) {
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1450 signal (SIGILL, SIG_DFL);
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1451 } else {
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1452 canjump = 1;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1453
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1454 asm volatile ("mtspr 256, %0\n\t"
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1455 "vand %%v0, %%v0, %%v0"
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1456 :
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1457 : "r" (-1));
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1458
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1459 signal (SIGILL, SIG_DFL);
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1460 return 1;
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1461 }
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1462 }
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1463 #endif /* CONFIG_DARWIN */
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
1464 return 0;
2286
0f937b191cc3 Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents: 2068
diff changeset
1465 #endif /* __AMIGAOS4__ */
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
1466 }
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
1467
3543
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1468 static void vorbis_inverse_coupling_altivec(float *mag, float *ang,
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1469 int blocksize)
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1470 {
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1471 int i;
3546
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1472 vector float m, a;
3543
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1473 vector bool int t0, t1;
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1474 const vector unsigned int v_31 = //XXX
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1475 vec_add(vec_add(vec_splat_u32(15),vec_splat_u32(15)),vec_splat_u32(1));
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1476 for(i=0; i<blocksize; i+=4) {
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1477 m = vec_ld(0, mag+i);
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1478 a = vec_ld(0, ang+i);
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1479 t0 = vec_cmple(m, (vector float)vec_splat_u32(0));
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1480 t1 = vec_cmple(a, (vector float)vec_splat_u32(0));
3545
e2a589e55906 Minor fix
lu_zero
parents: 3543
diff changeset
1481 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
1482 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
1483 t1 = (vector bool int)vec_andc(a, t1);
3550
4f4c13574ad5 Yet another typo
lu_zero
parents: 3549
diff changeset
1484 a = vec_sub(m, (vector float)t1);
4f4c13574ad5 Yet another typo
lu_zero
parents: 3549
diff changeset
1485 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
1486 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
1487 vec_stl(m, 0, mag+i);
3543
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1488 }
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1489 }
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1490
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
1491 /* next one assumes that ((line_size % 8) == 0) */
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
1492 void avg_pixels8_xy2_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
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
1493 {
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
1494 POWERPC_PERF_DECLARE(altivec_avg_pixels8_xy2_num, 1);
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1495 register int i;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1496 register vector unsigned char pixelsv1, pixelsv2, pixelsavg;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1497 register vector unsigned char blockv, temp1, temp2, blocktemp;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1498 register vector unsigned short pixelssum1, pixelssum2, temp3;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1499
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1500 register const_vector unsigned char vczero = (const_vector unsigned char)
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1501 vec_splat_u8(0);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1502 register const_vector unsigned short vctwo = (const_vector unsigned short)
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1503 vec_splat_u16(2);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1504
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1505 temp1 = vec_ld(0, pixels);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1506 temp2 = vec_ld(16, pixels);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1507 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1508 if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F) {
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1509 pixelsv2 = temp2;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1510 } else {
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1511 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1512 }
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1513 pixelsv1 = vec_mergeh(vczero, pixelsv1);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1514 pixelsv2 = vec_mergeh(vczero, pixelsv2);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1515 pixelssum1 = vec_add((vector unsigned short)pixelsv1,
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1516 (vector unsigned short)pixelsv2);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1517 pixelssum1 = vec_add(pixelssum1, vctwo);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1518
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1519 POWERPC_PERF_START_COUNT(altivec_avg_pixels8_xy2_num, 1);
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1520 for (i = 0; i < h ; i++) {
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1521 int rightside = ((unsigned long)block & 0x0000000F);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1522 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
1523
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1524 temp1 = vec_ld(line_size, pixels);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1525 temp2 = vec_ld(line_size + 16, pixels);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1526 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1527 if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F)
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1528 {
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1529 pixelsv2 = temp2;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1530 } else {
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1531 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
1532 }
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
1533
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1534 pixelsv1 = vec_mergeh(vczero, pixelsv1);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1535 pixelsv2 = vec_mergeh(vczero, pixelsv2);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1536 pixelssum2 = vec_add((vector unsigned short)pixelsv1,
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1537 (vector unsigned short)pixelsv2);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1538 temp3 = vec_add(pixelssum1, pixelssum2);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1539 temp3 = vec_sra(temp3, vctwo);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1540 pixelssum1 = vec_add(pixelssum2, vctwo);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1541 pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1542
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1543 if (rightside) {
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1544 blocktemp = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1));
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1545 } else {
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1546 blocktemp = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3));
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1547 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1548
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1549 blockv = vec_avg(blocktemp, blockv);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1550 vec_st(blockv, 0, block);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1551
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1552 block += line_size;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1553 pixels += line_size;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1554 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1555
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
1556 POWERPC_PERF_STOP_COUNT(altivec_avg_pixels8_xy2_num, 1);
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
1557 }
3546
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1558
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1559 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
1560 {
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1561 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
1562 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
1563 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
1564 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
1565 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
1566 c->sad[0]= sad16_altivec;
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1567 c->sad[1]= sad8_altivec;
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1568 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
1569 c->sse[1]= sse8_altivec;
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1570 c->sse[0]= sse16_altivec;
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1571 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
1572 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
1573 c->get_pixels = get_pixels_altivec;
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1574 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
1575 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
1576 /* 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
1577 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
1578 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
1579 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
1580 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
1581 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
1582 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
1583 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
1584 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
1585
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1586 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
1587 c->hadamard8_diff[1] = hadamard8_diff8x8_altivec;
3572
d5f97ae4f24f keep in sync with dsputil, makes --disable-decoder=vorbis build
lu_zero
parents: 3554
diff changeset
1588 #ifdef CONFIG_VORBIS_DECODER
3546
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1589 c->vorbis_inverse_coupling = vorbis_inverse_coupling_altivec;
3572
d5f97ae4f24f keep in sync with dsputil, makes --disable-decoder=vorbis build
lu_zero
parents: 3554
diff changeset
1590 #endif
3546
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1591 }