annotate ppc/dsputil_altivec.c @ 3663:8e180a3276fe libavcodec

New function for chroma MC (will be used in VC-1)
author kostya
date Sat, 02 Sep 2006 04:52:00 +0000
parents d5f97ae4f24f
children c8c591fe26f8
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 *
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
6 * This library is free software; you can redistribute it and/or
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
7 * 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
8 * License as published by the Free Software Foundation; either
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
9 * version 2 of the License, or (at your option) any later version.
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
10 *
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
11 * This library is distributed in the hope that it will be useful,
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
12 * 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
13 * 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
14 * 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
15 *
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
16 * You should have received a copy of the GNU Lesser General Public
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
17 * License along with this library; if not, write to the Free Software
3036
0b546eab515d Update licensing information: The FSF changed postal address.
diego
parents: 2979
diff changeset
18 * 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
19 */
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
20
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
21 #include "../dsputil.h"
1277
f3152eb76f1a altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents: 1064
diff changeset
22
f3152eb76f1a altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents: 1064
diff changeset
23 #include "gcc_fixes.h"
f3152eb76f1a altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents: 1064
diff changeset
24
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
25 #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
26
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
27 #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
28 #include <sys/sysctl.h>
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
29 #else /* CONFIG_DARWIN */
2286
0f937b191cc3 Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents: 2068
diff changeset
30 #ifdef __AMIGAOS4__
0f937b191cc3 Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents: 2068
diff changeset
31 #include <exec/exec.h>
0f937b191cc3 Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents: 2068
diff changeset
32 #include <interfaces/exec.h>
0f937b191cc3 Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents: 2068
diff changeset
33 #include <proto/exec.h>
0f937b191cc3 Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents: 2068
diff changeset
34 #else /* __AMIGAOS4__ */
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
35 #include <signal.h>
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
36 #include <setjmp.h>
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
37
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
38 static sigjmp_buf jmpbuf;
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
39 static volatile sig_atomic_t canjump = 0;
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
40
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
41 static void sigill_handler (int sig)
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 if (!canjump) {
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
44 signal (sig, SIG_DFL);
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
45 raise (sig);
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
46 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
47
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
48 canjump = 0;
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
49 siglongjmp (jmpbuf, 1);
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
50 }
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
51 #endif /* CONFIG_DARWIN */
2286
0f937b191cc3 Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents: 2068
diff changeset
52 #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
53
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
54 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
55 {
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
56 int i;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
57 int s __attribute__((aligned(16)));
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
58 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
59 vector unsigned char *tv;
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
60 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
61 vector unsigned int sad;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
62 vector signed int sumdiffs;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
63
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
64 s = 0;
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
65 sad = (vector unsigned int)vec_splat_u32(0);
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
66 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
67 /*
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
68 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
69 pix1v: pix1[0]-pix1[15]
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
70 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
71 */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
72 tv = (vector unsigned char *) pix1;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
73 pix1v = vec_perm(tv[0], tv[1], vec_lvsl(0, pix1));
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
74
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
75 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
76 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
77
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
78 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
79 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
80
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
81 /* Calculate the average vector */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
82 avgv = vec_avg(pix2v, pix2iv);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
83
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
84 /* 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
85 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
86
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
87 /* 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
88 sad = vec_sum4s(t5, sad);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
89
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
90 pix1 += line_size;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
91 pix2 += line_size;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
92 }
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
93 /* 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
94 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
95 sumdiffs = vec_splat(sumdiffs, 3);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
96 vec_ste(sumdiffs, 0, &s);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
97
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
98 return 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
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
101 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
102 {
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
103 int i;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
104 int s __attribute__((aligned(16)));
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
105 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
106 vector unsigned char *tv;
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
107 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
108 vector unsigned int sad;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
109 vector signed int sumdiffs;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
110 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
111
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
112 s = 0;
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
113 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
114
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
115 /*
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
116 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
117 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
118 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
119 time around the loop.
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
120 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
121 pix2v: pix2[0]-pix2[15]
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
122 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
123 */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
124 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
125 pix2v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[0]));
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
126
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
127 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
128 /*
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
129 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
130 pix1v: pix1[0]-pix1[15]
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
131 pix3v: pix3[0]-pix3[15]
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
132 */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
133 tv = (vector unsigned char *) pix1;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
134 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
135
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
136 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
137 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
138
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
139 /* Calculate the average vector */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
140 avgv = vec_avg(pix2v, pix3v);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
141
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
142 /* 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
143 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
144
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
145 /* 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
146 sad = vec_sum4s(t5, sad);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
147
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
148 pix1 += line_size;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
149 pix2v = pix3v;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
150 pix3 += line_size;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
151
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
152 }
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 /* 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
155 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
156 sumdiffs = vec_splat(sumdiffs, 3);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
157 vec_ste(sumdiffs, 0, &s);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
158 return s;
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
159 }
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
160
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
161 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
162 {
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
163 int i;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
164 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
165 uint8_t *pix3 = pix2 + line_size;
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
166 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
167 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
168 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
169 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
170 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
171 vector unsigned short pix3lv, pix3hv, pix3ilv, pix3ihv;
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
172 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
173 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
174 vector unsigned int sad;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
175 vector signed int sumdiffs;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
176
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
177 sad = (vector unsigned int)vec_splat_u32(0);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
178
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
179 s = 0;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
180
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
181 /*
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
182 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
183 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
184 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
185 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
186 Read unaligned pixels into our vectors. The vectors are as follows:
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
187 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
188 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
189 */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
190 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
191 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
192
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
193 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
194 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
195
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
196 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
197 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
198 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
199 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
200 t1 = vec_add(pix2hv, pix2ihv);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
201 t2 = vec_add(pix2lv, pix2ilv);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
202
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
203 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
204 /*
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
205 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
206 pix1v: pix1[0]-pix1[15]
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
207 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
208 */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
209 tv = (vector unsigned char *) pix1;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
210 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
211
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
212 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
213 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
214
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
215 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
216 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
217
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
218 /*
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
219 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
220 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
221 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
222 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
223 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
224 */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
225
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
226 /* 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
227 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
228 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
229 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
230 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
231
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
232 /* Do the averaging on them */
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
233 t3 = vec_add(pix3hv, pix3ihv);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
234 t4 = vec_add(pix3lv, pix3ilv);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
235
884
2cef5c4c0ca6 * altivec and pix_norm patch by Brian Foley
kabi
parents: 878
diff changeset
236 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
237 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
238
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
239 /* 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
240 avgv = vec_pack(avghv, avglv);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
241
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
242 /* 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
243 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
244
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
245 /* 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
246 sad = vec_sum4s(t5, sad);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
247
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
248 pix1 += line_size;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
249 pix3 += line_size;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
250 /* 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
251 t1 = t3;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
252 t2 = t4;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
253 }
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
254 /* 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
255 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
256 sumdiffs = vec_splat(sumdiffs, 3);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
257 vec_ste(sumdiffs, 0, &s);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
258
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
259 return 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
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
262 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
263 {
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
264 int i;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
265 int s __attribute__((aligned(16)));
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
266 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
267 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
268 vector unsigned char t1, t2, t3,t4, t5;
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
269 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
270 vector signed int sumdiffs;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
271
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
272 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
273
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
274
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
275 for(i=0;i<h;i++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
276 /* 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
277 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
278 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
279 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
280 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
281 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
282 t2 = vec_perm(pix2v[0], pix2v[1], perm2);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
283
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
284 /* 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
285 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
286 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
287 t5 = vec_sub(t3, t4);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
288
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
289 /* 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
290 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
291
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
292 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
293 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
294 }
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
295
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
296 /* 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
297 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
298 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
299 vec_ste(sumdiffs, 0, &s);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
300
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
301 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
302 }
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
303
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
304 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
305 {
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
306 int i;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
307 int s __attribute__((aligned(16)));
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
308 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
309 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
310 vector unsigned char t1, t2, t3,t4, t5;
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
311 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
312 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
313
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
314 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
315
f3152eb76f1a altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents: 1064
diff changeset
316 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
317
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
318 for(i=0;i<h;i++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
319 /* Read potentially unaligned pixels into t1 and t2
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
320 Since we're reading 16 pixels, and actually only want 8,
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
321 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
322 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
323 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
324 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
325 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
326 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
327 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
328
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
329 /* 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
330 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
331 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
332 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
333
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
334 /* 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
335 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
336
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
337 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
338 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
339 }
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
340
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
341 /* 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
342 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
343 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
344 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
345
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
346 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
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
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
349 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
350 {
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
351 int i;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
352 int s __attribute__((aligned(16)));
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
353 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
354 vector unsigned char *tv;
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
355 vector unsigned char pixv;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
356 vector unsigned int sv;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
357 vector signed int sum;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
358
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
359 sv = (vector unsigned int)vec_splat_u32(0);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
360
878
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
361 s = 0;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
362 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
363 /* 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
364 tv = (vector unsigned char *) pix;
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
365 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
366
884
2cef5c4c0ca6 * altivec and pix_norm patch by Brian Foley
kabi
parents: 878
diff changeset
367 /* Square the values, and add them to our sum */
2cef5c4c0ca6 * altivec and pix_norm patch by Brian Foley
kabi
parents: 878
diff changeset
368 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
369
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
370 pix += line_size;
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 /* 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
373 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
374 sum = vec_splat(sum, 3);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
375 vec_ste(sum, 0, &s);
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
376
6ea69518e5f7 altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 828
diff changeset
377 return 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
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
380 /**
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
381 * Sum of Squared Errors for a 8x8 block.
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
382 * AltiVec-enhanced.
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
383 * It's the sad8_altivec code above w/ squaring added.
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
384 */
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
385 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
386 {
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
387 int i;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
388 int s __attribute__((aligned(16)));
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
389 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
390 vector unsigned char perm1, perm2, permclear, *pix1v, *pix2v;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
391 vector unsigned char t1, t2, t3,t4, t5;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
392 vector unsigned int sum;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
393 vector signed int sumsqr;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
394
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
395 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
396
f3152eb76f1a altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents: 1064
diff changeset
397 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
398
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
399
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
400 for(i=0;i<h;i++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
401 /* Read potentially unaligned pixels into t1 and t2
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
402 Since we're reading 16 pixels, and actually only want 8,
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
403 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
404 perm1 = vec_lvsl(0, pix1);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
405 pix1v = (vector unsigned char *) pix1;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
406 perm2 = vec_lvsl(0, pix2);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
407 pix2v = (vector unsigned char *) pix2;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
408 t1 = vec_and(vec_perm(pix1v[0], pix1v[1], perm1), permclear);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
409 t2 = vec_and(vec_perm(pix2v[0], pix2v[1], perm2), permclear);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
410
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
411 /*
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
412 Since we want to use unsigned chars, we can take advantage
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
413 of the fact that abs(a-b)^2 = (a-b)^2.
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
414 */
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
415
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
416 /* Calculate abs differences vector */
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
417 t3 = vec_max(t1, t2);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
418 t4 = vec_min(t1, t2);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
419 t5 = vec_sub(t3, t4);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
420
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
421 /* Square the values and add them to our sum */
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
422 sum = vec_msum(t5, t5, sum);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
423
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
424 pix1 += line_size;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
425 pix2 += line_size;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
426 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
427
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
428 /* Sum up the four partial sums, and put the result into s */
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
429 sumsqr = vec_sums((vector signed int) sum, (vector signed int) zero);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
430 sumsqr = vec_splat(sumsqr, 3);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
431 vec_ste(sumsqr, 0, &s);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
432
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
433 return s;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
434 }
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
435
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
436 /**
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
437 * Sum of Squared Errors for a 16x16 block.
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
438 * AltiVec-enhanced.
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
439 * It's the sad16_altivec code above w/ squaring added.
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
440 */
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
441 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
442 {
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
443 int i;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
444 int s __attribute__((aligned(16)));
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
445 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
446 vector unsigned char perm1, perm2, *pix1v, *pix2v;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
447 vector unsigned char t1, t2, t3,t4, t5;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
448 vector unsigned int sum;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
449 vector signed int sumsqr;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
450
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
451 sum = (vector unsigned int)vec_splat_u32(0);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
452
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1352
diff changeset
453 for(i=0;i<h;i++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
454 /* Read potentially unaligned pixels into t1 and t2 */
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
455 perm1 = vec_lvsl(0, pix1);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
456 pix1v = (vector unsigned char *) pix1;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
457 perm2 = vec_lvsl(0, pix2);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
458 pix2v = (vector unsigned char *) pix2;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
459 t1 = vec_perm(pix1v[0], pix1v[1], perm1);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
460 t2 = vec_perm(pix2v[0], pix2v[1], perm2);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
461
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
462 /*
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
463 Since we want to use unsigned chars, we can take advantage
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
464 of the fact that abs(a-b)^2 = (a-b)^2.
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
465 */
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
466
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
467 /* Calculate abs differences vector */
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
468 t3 = vec_max(t1, t2);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
469 t4 = vec_min(t1, t2);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
470 t5 = vec_sub(t3, t4);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
471
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
472 /* Square the values and add them to our sum */
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
473 sum = vec_msum(t5, t5, sum);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
474
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
475 pix1 += line_size;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
476 pix2 += line_size;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
477 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
478
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
479 /* Sum up the four partial sums, and put the result into s */
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
480 sumsqr = vec_sums((vector signed int) sum, (vector signed int) zero);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
481 sumsqr = vec_splat(sumsqr, 3);
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
482 vec_ste(sumsqr, 0, &s);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
483
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
484 return s;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
485 }
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
486
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1033
diff changeset
487 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
488 {
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
489 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
490 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
491 vector unsigned char t1;
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
492 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
493 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
494
981
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
495 int i;
8bec850dc9c7 altivec patches by Romain Dolbeau
bellard
parents: 978
diff changeset
496 int s __attribute__((aligned(16)));
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
497
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
498 sad = (vector unsigned int)vec_splat_u32(0);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
499
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
500 for (i = 0; i < 16; i++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
501 /* 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
502 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
503 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
504 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
505
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
506 /* 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
507 sad = vec_sum4s(t1, sad);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
508
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 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
510 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
511
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
512 /* 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
513 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
514 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
515 vec_ste(sumdiffs, 0, &s);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
516
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
517 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
518 }
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
519
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1033
diff changeset
520 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
521 {
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
522 int i;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
523 vector unsigned char perm, bytes, *pixv;
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
524 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
525 vector signed short shorts;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
526
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
527 for(i=0;i<8;i++)
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 // Read potentially unaligned pixels.
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
530 // 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
531 // but we simply ignore the extras.
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
532 perm = vec_lvsl(0, pixels);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
533 pixv = (vector unsigned char *) pixels;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
534 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
535
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
536 // convert the bytes into shorts
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
537 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
538
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
539 // 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
540 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
541
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
542 pixels += line_size;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
543 }
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
544 }
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
545
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1033
diff changeset
546 void diff_pixels_altivec(DCTELEM *restrict block, const uint8_t *s1,
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1033
diff changeset
547 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
548 {
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
549 int i;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
550 vector unsigned char perm, bytes, *pixv;
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
551 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
552 vector signed short shorts1, shorts2;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
553
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
554 for(i=0;i<4;i++)
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 // Read potentially unaligned pixels
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
557 // 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
558 // but we simply ignore the extras.
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
559 perm = vec_lvsl(0, s1);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
560 pixv = (vector unsigned char *) s1;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
561 bytes = vec_perm(pixv[0], pixv[1], perm);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
562
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
563 // convert the bytes into shorts
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
564 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
565
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
566 // 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
567 perm = vec_lvsl(0, s2);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
568 pixv = (vector unsigned char *) s2;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
569 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
570
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
571 // convert the bytes into shorts
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
572 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
573
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
574 // Do the subtraction
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
575 shorts1 = vec_sub(shorts1, shorts2);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
576
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
577 // 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
578 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
579
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
580 s1 += stride;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
581 s2 += stride;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
582 block += 8;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
583
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
584
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
585 // 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
586 // unroll.
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
587
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
588 // Read potentially unaligned pixels
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
589 // 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
590 // but we simply ignore the extras.
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
591 perm = vec_lvsl(0, s1);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
592 pixv = (vector unsigned char *) s1;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
593 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
594
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
595 // convert the bytes into shorts
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
596 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
597
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
598 // 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
599 perm = vec_lvsl(0, s2);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
600 pixv = (vector unsigned char *) s2;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
601 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
602
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
603 // convert the bytes into shorts
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
604 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
605
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
606 // Do the subtraction
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
607 shorts1 = vec_sub(shorts1, shorts2);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
608
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
609 // 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
610 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
611
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
612 s1 += stride;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
613 s2 += stride;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
614 block += 8;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
615 }
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
616 }
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
617
995
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
618 void add_bytes_altivec(uint8_t *dst, uint8_t *src, int w) {
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
619 #ifdef ALTIVEC_USE_REFERENCE_C_CODE
995
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
620 int i;
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
621 for(i=0; i+7<w; i++){
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
622 dst[i+0] += src[i+0];
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
623 dst[i+1] += src[i+1];
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
624 dst[i+2] += src[i+2];
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
625 dst[i+3] += src[i+3];
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
626 dst[i+4] += src[i+4];
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
627 dst[i+5] += src[i+5];
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
628 dst[i+6] += src[i+6];
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
629 dst[i+7] += src[i+7];
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
630 }
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
631 for(; i<w; i++)
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
632 dst[i+0] += src[i+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
633 #else /* ALTIVEC_USE_REFERENCE_C_CODE */
995
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
634 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
635 register vector unsigned char vdst, vsrc;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
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 /* dst and src are 16 bytes-aligned (guaranteed) */
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
638 for(i = 0 ; (i + 15) < w ; i++)
995
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
639 {
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 vdst = vec_ld(i << 4, (unsigned char*)dst);
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
641 vsrc = vec_ld(i << 4, (unsigned char*)src);
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
642 vdst = vec_add(vsrc, vdst);
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
643 vec_st(vdst, i << 4, (unsigned char*)dst);
995
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
644 }
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
645 /* 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
646 for (; (i < w) ; i++)
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
647 {
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
648 dst[i] = src[i];
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
649 }
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
650 #endif /* ALTIVEC_USE_REFERENCE_C_CODE */
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
651 }
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
652
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
653 /* 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
654 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
655 {
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
656 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
657 #ifdef ALTIVEC_USE_REFERENCE_C_CODE
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
658 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
659
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
660 POWERPC_PERF_START_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
661
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
662 for(i=0; i<h; i++) {
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
663 *((uint32_t*)(block)) = LD32(pixels);
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
664 *((uint32_t*)(block+4)) = LD32(pixels+4);
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
665 *((uint32_t*)(block+8)) = LD32(pixels+8);
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
666 *((uint32_t*)(block+12)) = LD32(pixels+12);
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
667 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
668 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
669 }
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
670
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
671 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
672
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
673 #else /* ALTIVEC_USE_REFERENCE_C_CODE */
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
674 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
675 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
676 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
677 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
678
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
679 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
680 int i;
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
681 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
682 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
683 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
684
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
685 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
686 // 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
687 // 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
688 // 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
689 // -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
690 // 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
691 #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
692 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
693 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
694 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
695 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
696 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
697 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
698 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
699 }
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
700 #else
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
701 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
702 pixelsv1 = vec_ld(0, (unsigned char*)pixels);
3533
0937cc91b574 avoid possible segfault situations
lu_zero
parents: 3346
diff changeset
703 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
704 pixelsv1B = vec_ld(line_size, (unsigned char*)pixels);
3533
0937cc91b574 avoid possible segfault situations
lu_zero
parents: 3346
diff changeset
705 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
706 pixelsv1C = vec_ld(line_size_2, (unsigned char*)pixels);
3533
0937cc91b574 avoid possible segfault situations
lu_zero
parents: 3346
diff changeset
707 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
708 pixelsv1D = vec_ld(line_size_3, (unsigned char*)pixels);
3533
0937cc91b574 avoid possible segfault situations
lu_zero
parents: 3346
diff changeset
709 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
710 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
711 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
712 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
713 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
714 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
715 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
716 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
717 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
718 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
719 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
720 }
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
721 #endif
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
722 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
723
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
724 #endif /* ALTIVEC_USE_REFERENCE_C_CODE */
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
725 }
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
726
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
727 /* 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
728 #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
729 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
730 {
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
731 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
732 #ifdef ALTIVEC_USE_REFERENCE_C_CODE
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
733 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
734
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
735 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
736
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
737 for(i=0; i<h; i++) {
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
738 op_avg(*((uint32_t*)(block)),LD32(pixels));
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
739 op_avg(*((uint32_t*)(block+4)),LD32(pixels+4));
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
740 op_avg(*((uint32_t*)(block+8)),LD32(pixels+8));
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
741 op_avg(*((uint32_t*)(block+12)),LD32(pixels+12));
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
742 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
743 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
744 }
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
745
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
746 POWERPC_PERF_STOP_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
747
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
748 #else /* ALTIVEC_USE_REFERENCE_C_CODE */
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
749 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
750 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
751 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
752
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
753 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
754
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
755 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
756 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
757 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
758 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
759 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
760 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
761 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
762 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
763 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
764 }
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
765
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
766 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
767
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 #endif /* ALTIVEC_USE_REFERENCE_C_CODE */
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 }
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
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
771 /* next one assumes that ((line_size % 8) == 0) */
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
772 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
773 {
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
774 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
775 #ifdef ALTIVEC_USE_REFERENCE_C_CODE
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 int i;
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
777 POWERPC_PERF_START_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
778 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
779 *((uint32_t *) (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
780 (((*((uint32_t *) (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
781 ((((const struct unaligned_32 *) (pixels))->l))) -
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 ((((*((uint32_t *) (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
783 ((((const struct unaligned_32 *) (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
784 l))) & 0xFEFEFEFEUL) >> 1));
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 *((uint32_t *) (block + 4)) =
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 (((*((uint32_t *) (block + 4))) |
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 ((((const struct unaligned_32 *) (pixels + 4))->l))) -
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 ((((*((uint32_t *) (block + 4))) ^
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 ((((const struct unaligned_32 *) (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
790 4))->
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
791 l))) & 0xFEFEFEFEUL) >> 1));
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
792 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
793 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
794 }
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
795 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
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 #else /* ALTIVEC_USE_REFERENCE_C_CODE */
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 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
799 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
800
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
801 POWERPC_PERF_START_COUNT(altivec_avg_pixels8_num, 1);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
802
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
803 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
804 /*
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 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
806 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
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 int rightside = ((unsigned long)block & 0x0000000F);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
809
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
810 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
811 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
812 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
813 pixelsv = vec_perm(pixelsv1, pixelsv2, vec_lvsl(0, pixels));
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
814
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
815 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
816 {
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
817 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
818 }
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 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
820 {
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 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
822 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
823
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
824 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
825
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
826 vec_st(blockv, 0, block);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
827
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
828 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
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 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
831
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
832 POWERPC_PERF_STOP_COUNT(altivec_avg_pixels8_num, 1);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
833
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
834 #endif /* ALTIVEC_USE_REFERENCE_C_CODE */
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
835 }
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 995
diff changeset
836
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
837 /* 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
838 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
839 {
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
840 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
841 #ifdef ALTIVEC_USE_REFERENCE_C_CODE
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
842 int j;
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
843 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
844 for (j = 0; j < 2; j++) {
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
845 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
846 const uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
847 const uint32_t b =
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
848 (((const struct unaligned_32 *) (pixels + 1))->l);
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
849 uint32_t l0 =
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
850 (a & 0x03030303UL) + (b & 0x03030303UL) + 0x02020202UL;
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
851 uint32_t h0 =
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
852 ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
853 uint32_t l1, h1;
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
854 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
855 for (i = 0; i < h; i += 2) {
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
856 uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
857 uint32_t b = (((const struct unaligned_32 *) (pixels + 1))->l);
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
858 l1 = (a & 0x03030303UL) + (b & 0x03030303UL);
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
859 h1 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
860 *((uint32_t *) 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
861 h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL);
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
862 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
863 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
864 a = (((const struct unaligned_32 *) (pixels))->l);
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
865 b = (((const struct unaligned_32 *) (pixels + 1))->l);
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
866 l0 = (a & 0x03030303UL) + (b & 0x03030303UL) + 0x02020202UL;
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
867 h0 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
868 *((uint32_t *) 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
869 h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL);
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
870 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
871 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
872 } pixels += 4 - line_size * (h + 1);
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
873 block += 4 - line_size * 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
874 }
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
875
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
876 POWERPC_PERF_STOP_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
877
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
878 #else /* ALTIVEC_USE_REFERENCE_C_CODE */
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
879 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
880 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
881 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
882 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
883 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
884 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
885 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
886 pixelssum1, pixelssum2, temp3;
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
887 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
888 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
889
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
890 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
891 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
892 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
893 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
894 {
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
895 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
896 }
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
897 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
898 {
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
899 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
900 }
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
901 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
902 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
903 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
904 (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
905 pixelssum1 = vec_add(pixelssum1, vctwo);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
906
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
907 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
908 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
909 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
910 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
911
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
912 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
913 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
914 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
915 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
916 {
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
917 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
918 }
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
919 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
920 {
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
921 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
922 }
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
923
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
924 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
925 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
926 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
927 (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
928 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
929 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
930 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
931 pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
932
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
933 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
934 {
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
935 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
936 }
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
937 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
938 {
35cf2f4a0f8c PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1009
diff changeset
939 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
940 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
941
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
942 vec_st(blockv, 0, block);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
943
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
944 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
945 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
946 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
947
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
948 POWERPC_PERF_STOP_COUNT(altivec_put_pixels8_xy2_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
949 #endif /* ALTIVEC_USE_REFERENCE_C_CODE */
995
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 981
diff changeset
950 }
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents: 638
diff changeset
951
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
952 /* 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
953 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
954 {
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
955 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
956 #ifdef ALTIVEC_USE_REFERENCE_C_CODE
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
957 int j;
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
958 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
959 for (j = 0; j < 2; j++) {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
960 int i;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
961 const uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
962 const uint32_t b =
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
963 (((const struct unaligned_32 *) (pixels + 1))->l);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
964 uint32_t l0 =
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
965 (a & 0x03030303UL) + (b & 0x03030303UL) + 0x01010101UL;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
966 uint32_t h0 =
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
967 ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
968 uint32_t l1, h1;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
969 pixels += line_size;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
970 for (i = 0; i < h; i += 2) {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
971 uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
972 uint32_t b = (((const struct unaligned_32 *) (pixels + 1))->l);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
973 l1 = (a & 0x03030303UL) + (b & 0x03030303UL);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
974 h1 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
975 *((uint32_t *) block) =
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
976 h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
977 pixels += line_size;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
978 block += line_size;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
979 a = (((const struct unaligned_32 *) (pixels))->l);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
980 b = (((const struct unaligned_32 *) (pixels + 1))->l);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
981 l0 = (a & 0x03030303UL) + (b & 0x03030303UL) + 0x01010101UL;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
982 h0 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
983 *((uint32_t *) block) =
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
984 h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
985 pixels += line_size;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
986 block += line_size;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
987 } pixels += 4 - line_size * (h + 1);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
988 block += 4 - line_size * h;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
989 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
990
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
991 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
992
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
993 #else /* ALTIVEC_USE_REFERENCE_C_CODE */
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
994 register int i;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
995 register vector unsigned char
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
996 pixelsv1, pixelsv2,
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
997 pixelsavg;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
998 register vector unsigned char
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
999 blockv, temp1, temp2;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1000 register vector unsigned short
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1001 pixelssum1, pixelssum2, temp3;
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
1002 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
1003 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
1004 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
1005
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1006 temp1 = vec_ld(0, pixels);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1007 temp2 = vec_ld(16, pixels);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1008 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
1009 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
1010 {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1011 pixelsv2 = temp2;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1012 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1013 else
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1014 {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1015 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
1016 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1017 pixelsv1 = vec_mergeh(vczero, pixelsv1);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1018 pixelsv2 = vec_mergeh(vczero, pixelsv2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1019 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
1020 (vector unsigned short)pixelsv2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1021 pixelssum1 = vec_add(pixelssum1, vcone);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1022
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1023 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
1024 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
1025 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
1026 blockv = vec_ld(0, block);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1027
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1028 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
1029 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
1030 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
1031 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
1032 {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1033 pixelsv2 = temp2;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1034 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1035 else
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1036 {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1037 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
1038 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1039
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1040 pixelsv1 = vec_mergeh(vczero, pixelsv1);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1041 pixelsv2 = vec_mergeh(vczero, pixelsv2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1042 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
1043 (vector unsigned short)pixelsv2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1044 temp3 = vec_add(pixelssum1, pixelssum2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1045 temp3 = vec_sra(temp3, vctwo);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1046 pixelssum1 = vec_add(pixelssum2, vcone);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1047 pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1048
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1049 if (rightside)
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1050 {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1051 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
1052 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1053 else
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1054 {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1055 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
1056 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1057
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1058 vec_st(blockv, 0, block);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1059
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1060 block += line_size;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1061 pixels += line_size;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1062 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1063
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
1064 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
1065 #endif /* ALTIVEC_USE_REFERENCE_C_CODE */
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
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1068 /* 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
1069 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
1070 {
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
1071 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
1072 #ifdef ALTIVEC_USE_REFERENCE_C_CODE
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1073 int j;
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
1074 POWERPC_PERF_START_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
1075 for (j = 0; j < 4; j++) {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1076 int i;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1077 const uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1078 const uint32_t b =
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1079 (((const struct unaligned_32 *) (pixels + 1))->l);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1080 uint32_t l0 =
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1081 (a & 0x03030303UL) + (b & 0x03030303UL) + 0x02020202UL;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1082 uint32_t h0 =
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1083 ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1084 uint32_t l1, h1;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1085 pixels += line_size;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1086 for (i = 0; i < h; i += 2) {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1087 uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1088 uint32_t b = (((const struct unaligned_32 *) (pixels + 1))->l);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1089 l1 = (a & 0x03030303UL) + (b & 0x03030303UL);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1090 h1 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1091 *((uint32_t *) block) =
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1092 h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1093 pixels += line_size;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1094 block += line_size;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1095 a = (((const struct unaligned_32 *) (pixels))->l);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1096 b = (((const struct unaligned_32 *) (pixels + 1))->l);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1097 l0 = (a & 0x03030303UL) + (b & 0x03030303UL) + 0x02020202UL;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1098 h0 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1099 *((uint32_t *) block) =
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1100 h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1101 pixels += line_size;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1102 block += line_size;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1103 } pixels += 4 - line_size * (h + 1);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1104 block += 4 - line_size * h;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1105 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1106
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
1107 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
1108
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1109 #else /* ALTIVEC_USE_REFERENCE_C_CODE */
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1110 register int i;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1111 register vector unsigned char
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1112 pixelsv1, pixelsv2, pixelsv3, pixelsv4;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1113 register vector unsigned char
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1114 blockv, temp1, temp2;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1115 register vector unsigned short
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1116 pixelssum1, pixelssum2, temp3,
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1117 pixelssum3, pixelssum4, temp4;
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
1118 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
1119 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
1120
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
1121 POWERPC_PERF_START_COUNT(altivec_put_pixels16_xy2_num, 1);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1122
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1123 temp1 = vec_ld(0, pixels);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1124 temp2 = vec_ld(16, pixels);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1125 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
1126 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
1127 {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1128 pixelsv2 = temp2;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1129 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1130 else
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1131 {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1132 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
1133 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1134 pixelsv3 = vec_mergel(vczero, pixelsv1);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1135 pixelsv4 = vec_mergel(vczero, pixelsv2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1136 pixelsv1 = vec_mergeh(vczero, pixelsv1);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1137 pixelsv2 = vec_mergeh(vczero, pixelsv2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1138 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
1139 (vector unsigned short)pixelsv4);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1140 pixelssum3 = vec_add(pixelssum3, vctwo);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1141 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
1142 (vector unsigned short)pixelsv2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1143 pixelssum1 = vec_add(pixelssum1, vctwo);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1144
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1145 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
1146 blockv = vec_ld(0, block);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1147
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1148 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
1149 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
1150 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
1151 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
1152 {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1153 pixelsv2 = temp2;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1154 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1155 else
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1156 {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1157 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
1158 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1159
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1160 pixelsv3 = vec_mergel(vczero, pixelsv1);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1161 pixelsv4 = vec_mergel(vczero, pixelsv2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1162 pixelsv1 = vec_mergeh(vczero, pixelsv1);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1163 pixelsv2 = vec_mergeh(vczero, pixelsv2);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1164
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1165 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
1166 (vector unsigned short)pixelsv4);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1167 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
1168 (vector unsigned short)pixelsv2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1169 temp4 = vec_add(pixelssum3, pixelssum4);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1170 temp4 = vec_sra(temp4, vctwo);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1171 temp3 = vec_add(pixelssum1, pixelssum2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1172 temp3 = vec_sra(temp3, vctwo);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1173
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1174 pixelssum3 = vec_add(pixelssum4, vctwo);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1175 pixelssum1 = vec_add(pixelssum2, vctwo);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1176
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1177 blockv = vec_packsu(temp3, temp4);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1178
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1179 vec_st(blockv, 0, block);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1180
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1181 block += line_size;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1182 pixels += line_size;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1183 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1184
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
1185 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
1186 #endif /* ALTIVEC_USE_REFERENCE_C_CODE */
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1187 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1188
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1189 /* 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
1190 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
1191 {
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
1192 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
1193 #ifdef ALTIVEC_USE_REFERENCE_C_CODE
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1194 int j;
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
1195 POWERPC_PERF_START_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
1196 for (j = 0; j < 4; j++) {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1197 int i;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1198 const uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1199 const uint32_t b =
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1200 (((const struct unaligned_32 *) (pixels + 1))->l);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1201 uint32_t l0 =
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1202 (a & 0x03030303UL) + (b & 0x03030303UL) + 0x01010101UL;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1203 uint32_t h0 =
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1204 ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1205 uint32_t l1, h1;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1206 pixels += line_size;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1207 for (i = 0; i < h; i += 2) {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1208 uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1209 uint32_t b = (((const struct unaligned_32 *) (pixels + 1))->l);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1210 l1 = (a & 0x03030303UL) + (b & 0x03030303UL);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1211 h1 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1212 *((uint32_t *) block) =
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1213 h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1214 pixels += line_size;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1215 block += line_size;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1216 a = (((const struct unaligned_32 *) (pixels))->l);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1217 b = (((const struct unaligned_32 *) (pixels + 1))->l);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1218 l0 = (a & 0x03030303UL) + (b & 0x03030303UL) + 0x01010101UL;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1219 h0 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1220 *((uint32_t *) block) =
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1221 h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1222 pixels += line_size;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1223 block += line_size;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1224 } pixels += 4 - line_size * (h + 1);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1225 block += 4 - line_size * h;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1226 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1227
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
1228 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
1229
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1230 #else /* ALTIVEC_USE_REFERENCE_C_CODE */
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1231 register int i;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1232 register vector unsigned char
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1233 pixelsv1, pixelsv2, pixelsv3, pixelsv4;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1234 register vector unsigned char
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1235 blockv, temp1, temp2;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1236 register vector unsigned short
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1237 pixelssum1, pixelssum2, temp3,
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1238 pixelssum3, pixelssum4, temp4;
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1708
diff changeset
1239 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
1240 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
1241 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
1242
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
1243 POWERPC_PERF_START_COUNT(altivec_put_no_rnd_pixels16_xy2_num, 1);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1244
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1245 temp1 = vec_ld(0, pixels);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1246 temp2 = vec_ld(16, pixels);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1247 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
1248 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
1249 {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1250 pixelsv2 = temp2;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1251 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1252 else
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1253 {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1254 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
1255 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1256 pixelsv3 = vec_mergel(vczero, pixelsv1);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1257 pixelsv4 = vec_mergel(vczero, pixelsv2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1258 pixelsv1 = vec_mergeh(vczero, pixelsv1);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1259 pixelsv2 = vec_mergeh(vczero, pixelsv2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1260 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
1261 (vector unsigned short)pixelsv4);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1262 pixelssum3 = vec_add(pixelssum3, vcone);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1263 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
1264 (vector unsigned short)pixelsv2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1265 pixelssum1 = vec_add(pixelssum1, vcone);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1266
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1267 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
1268 blockv = vec_ld(0, block);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1269
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1270 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
1271 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
1272 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
1273 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
1274 {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1275 pixelsv2 = temp2;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1276 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1277 else
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1278 {
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1279 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
1280 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1281
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1282 pixelsv3 = vec_mergel(vczero, pixelsv1);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1283 pixelsv4 = vec_mergel(vczero, pixelsv2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1284 pixelsv1 = vec_mergeh(vczero, pixelsv1);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1285 pixelsv2 = vec_mergeh(vczero, pixelsv2);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1286
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1287 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
1288 (vector unsigned short)pixelsv4);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1289 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
1290 (vector unsigned short)pixelsv2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1291 temp4 = vec_add(pixelssum3, pixelssum4);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1292 temp4 = vec_sra(temp4, vctwo);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1293 temp3 = vec_add(pixelssum1, pixelssum2);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1294 temp3 = vec_sra(temp3, vctwo);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1295
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1296 pixelssum3 = vec_add(pixelssum4, vcone);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1297 pixelssum1 = vec_add(pixelssum2, vcone);
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1298
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1299 blockv = vec_packsu(temp3, temp4);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1300
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1301 vec_st(blockv, 0, block);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1302
1024
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1303 block += line_size;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1304 pixels += line_size;
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1305 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1306
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1340
diff changeset
1307 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
1308 #endif /* ALTIVEC_USE_REFERENCE_C_CODE */
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1309 }
9cc1031e1864 More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1015
diff changeset
1310
1949
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1311 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
1312 POWERPC_PERF_DECLARE(altivec_hadamard8_diff8x8_num, 1);
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1313 int sum;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1314 register const_vector unsigned char vzero =
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1315 (const_vector unsigned char)vec_splat_u8(0);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1316 register vector signed short temp0, temp1, temp2, temp3, temp4,
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1317 temp5, temp6, temp7;
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1318 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
1319 {
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1320 register const_vector signed short vprod1 =(const_vector signed short)
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1321 AVV( 1,-1, 1,-1, 1,-1, 1,-1);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1322 register const_vector signed short vprod2 =(const_vector signed short)
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1323 AVV( 1, 1,-1,-1, 1, 1,-1,-1);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1324 register const_vector signed short vprod3 =(const_vector signed short)
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1325 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
1326 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
1327 AVV(0x02, 0x03, 0x00, 0x01, 0x06, 0x07, 0x04, 0x05,
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1328 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
1329 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
1330 AVV(0x04, 0x05, 0x06, 0x07, 0x00, 0x01, 0x02, 0x03,
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1331 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
1332 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
1333 AVV(0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F,
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1334 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
1335
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1336 #define ONEITERBUTTERFLY(i, res) \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1337 { \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1338 register vector unsigned char src1, src2, srcO; \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1339 register vector unsigned char dst1, dst2, dstO; \
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1340 register vector signed short srcV, dstV; \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1341 register vector signed short but0, but1, but2, op1, op2, op3; \
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1342 src1 = vec_ld(stride * i, src); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1343 if ((((stride * i) + (unsigned long)src) & 0x0000000F) > 8) \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1344 src2 = vec_ld((stride * i) + 16, src); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1345 srcO = vec_perm(src1, src2, vec_lvsl(stride * i, src)); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1346 dst1 = vec_ld(stride * i, dst); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1347 if ((((stride * i) + (unsigned long)dst) & 0x0000000F) > 8) \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1348 dst2 = vec_ld((stride * i) + 16, dst); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1349 dstO = vec_perm(dst1, dst2, vec_lvsl(stride * i, dst)); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1350 /* promote the unsigned chars to signed shorts */ \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1351 /* 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
1352 srcV = \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1353 (vector signed short)vec_mergeh((vector signed char)vzero, \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1354 (vector signed char)srcO); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1355 dstV = \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1356 (vector signed short)vec_mergeh((vector signed char)vzero, \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1357 (vector signed char)dstO); \
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1358 /* substractions inside the first butterfly */ \
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1359 but0 = vec_sub(srcV, dstV); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1360 op1 = vec_perm(but0, but0, perm1); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1361 but1 = vec_mladd(but0, vprod1, op1); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1362 op2 = vec_perm(but1, but1, perm2); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1363 but2 = vec_mladd(but1, vprod2, op2); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1364 op3 = vec_perm(but2, but2, perm3); \
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1365 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
1366 }
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1367 ONEITERBUTTERFLY(0, temp0);
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1368 ONEITERBUTTERFLY(1, temp1);
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1369 ONEITERBUTTERFLY(2, temp2);
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1370 ONEITERBUTTERFLY(3, temp3);
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1371 ONEITERBUTTERFLY(4, temp4);
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1372 ONEITERBUTTERFLY(5, temp5);
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1373 ONEITERBUTTERFLY(6, temp6);
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1374 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
1375 }
1949
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1376 #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
1377 {
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1378 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
1379 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
1380 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
1381 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
1382 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
1383 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
1384 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
1385 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
1386 register vector signed short line7 = vec_sub(temp6, temp7);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1387
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1388 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
1389 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
1390 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
1391 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
1392 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
1393 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
1394 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
1395 register vector signed short line7B = vec_sub(line5, line7);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1396
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1397 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
1398 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
1399 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
1400 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
1401 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
1402 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
1403 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
1404 register vector signed short line7C = vec_sub(line3B, line7B);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1405
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1406 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
1407 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
1408 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
1409 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
1410 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
1411 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
1412 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
1413 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
1414 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
1415 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
1416 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
1417 }
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1418 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
1419 return sum;
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1420 }
66215baae7b9 hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
1421
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 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
1424 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
1425 It's only used from inside hadamard8_diff16_altivec.
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1426
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1427 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
1428 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
1429 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
1430 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
1431 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
1432 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
1433 fall from 700+ cycles to 550).
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1434
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1435 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
1436 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
1437 gcc-3.3 on the 7450 (but uses 25% less instructions...)
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1438
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1439 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
1440 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
1441 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
1442 */
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1443
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1444 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
1445 int sum;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1446 register vector signed short
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1447 temp0 REG_v(v0),
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1448 temp1 REG_v(v1),
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1449 temp2 REG_v(v2),
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1450 temp3 REG_v(v3),
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1451 temp4 REG_v(v4),
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1452 temp5 REG_v(v5),
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1453 temp6 REG_v(v6),
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1454 temp7 REG_v(v7);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1455 register vector signed short
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1456 temp0S REG_v(v8),
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1457 temp1S REG_v(v9),
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1458 temp2S REG_v(v10),
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1459 temp3S REG_v(v11),
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1460 temp4S REG_v(v12),
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1461 temp5S REG_v(v13),
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1462 temp6S REG_v(v14),
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1463 temp7S REG_v(v15);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1464 register const_vector unsigned char vzero REG_v(v31)=
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1465 (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
1466 {
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1467 register const_vector signed short vprod1 REG_v(v16)=
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1468 (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
1469 register const_vector signed short vprod2 REG_v(v17)=
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1470 (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
1471 register const_vector signed short vprod3 REG_v(v18)=
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1472 (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
1473 register const_vector unsigned char perm1 REG_v(v19)=
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1474 (const_vector unsigned char)
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1475 AVV(0x02, 0x03, 0x00, 0x01, 0x06, 0x07, 0x04, 0x05,
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1476 0x0A, 0x0B, 0x08, 0x09, 0x0E, 0x0F, 0x0C, 0x0D);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1477 register const_vector unsigned char perm2 REG_v(v20)=
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1478 (const_vector unsigned char)
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1479 AVV(0x04, 0x05, 0x06, 0x07, 0x00, 0x01, 0x02, 0x03,
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1480 0x0C, 0x0D, 0x0E, 0x0F, 0x08, 0x09, 0x0A, 0x0B);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1481 register const_vector unsigned char perm3 REG_v(v21)=
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1482 (const_vector unsigned char)
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1483 AVV(0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F,
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1484 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
1485
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1486 #define ONEITERBUTTERFLY(i, res1, res2) \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1487 { \
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1488 register vector unsigned char src1 REG_v(v22), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1489 src2 REG_v(v23), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1490 dst1 REG_v(v24), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1491 dst2 REG_v(v25), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1492 srcO REG_v(v22), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1493 dstO REG_v(v23); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1494 \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1495 register vector signed short srcV REG_v(v24), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1496 dstV REG_v(v25), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1497 srcW REG_v(v26), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1498 dstW REG_v(v27), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1499 but0 REG_v(v28), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1500 but0S REG_v(v29), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1501 op1 REG_v(v30), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1502 but1 REG_v(v22), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1503 op1S REG_v(v23), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1504 but1S REG_v(v24), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1505 op2 REG_v(v25), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1506 but2 REG_v(v26), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1507 op2S REG_v(v27), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1508 but2S REG_v(v28), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1509 op3 REG_v(v29), \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1510 op3S REG_v(v30); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1511 \
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1512 src1 = vec_ld(stride * i, src); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1513 src2 = vec_ld((stride * i) + 16, src); \
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1514 srcO = vec_perm(src1, src2, vec_lvsl(stride * i, src)); \
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1515 dst1 = vec_ld(stride * i, dst); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1516 dst2 = vec_ld((stride * i) + 16, dst); \
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1517 dstO = vec_perm(dst1, dst2, vec_lvsl(stride * i, dst)); \
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1518 /* promote the unsigned chars to signed shorts */ \
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1519 srcV = \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1520 (vector signed short)vec_mergeh((vector signed char)vzero, \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1521 (vector signed char)srcO); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1522 dstV = \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1523 (vector signed short)vec_mergeh((vector signed char)vzero, \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1524 (vector signed char)dstO); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1525 srcW = \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1526 (vector signed short)vec_mergel((vector signed char)vzero, \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1527 (vector signed char)srcO); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1528 dstW = \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1529 (vector signed short)vec_mergel((vector signed char)vzero, \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1530 (vector signed char)dstO); \
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1531 /* substractions inside the first butterfly */ \
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1532 but0 = vec_sub(srcV, dstV); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1533 but0S = vec_sub(srcW, dstW); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1534 op1 = vec_perm(but0, but0, perm1); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1535 but1 = vec_mladd(but0, vprod1, op1); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1536 op1S = vec_perm(but0S, but0S, perm1); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1537 but1S = vec_mladd(but0S, vprod1, op1S); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1538 op2 = vec_perm(but1, but1, perm2); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1539 but2 = vec_mladd(but1, vprod2, op2); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1540 op2S = vec_perm(but1S, but1S, perm2); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1541 but2S = vec_mladd(but1S, vprod2, op2S); \
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1542 op3 = vec_perm(but2, but2, perm3); \
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1543 res1 = vec_mladd(but2, vprod3, op3); \
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1544 op3S = vec_perm(but2S, but2S, perm3); \
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1545 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
1546 }
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1547 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
1548 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
1549 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
1550 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
1551 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
1552 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
1553 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
1554 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
1555 }
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1556 #undef ONEITERBUTTERFLY
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1557 {
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1558 register vector signed int vsum;
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1559 register vector signed short line0S, line1S, line2S, line3S, line4S,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1560 line5S, line6S, line7S, line0BS,line2BS,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1561 line1BS,line3BS,line4BS,line6BS,line5BS,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1562 line7BS,line0CS,line4CS,line1CS,line5CS,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1563 line2CS,line6CS,line3CS,line7CS;
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1564
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1565 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
1566 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
1567 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
1568 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
1569 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
1570 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
1571 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
1572 register vector signed short line7 = vec_sub(temp6, temp7);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1573
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1574 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
1575 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
1576 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
1577 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
1578 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
1579 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
1580 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
1581 register vector signed short line7B = vec_sub(line5, line7);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1582
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1583 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
1584 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
1585 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
1586 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
1587 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
1588 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
1589 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
1590 register vector signed short line7C = vec_sub(line3B, line7B);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1591
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1592 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
1593 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
1594 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
1595 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
1596 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
1597 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
1598 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
1599 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
1600
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1601 line0S = vec_add(temp0S, temp1S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1602 line1S = vec_sub(temp0S, temp1S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1603 line2S = vec_add(temp2S, temp3S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1604 line3S = vec_sub(temp2S, temp3S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1605 line4S = vec_add(temp4S, temp5S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1606 line5S = vec_sub(temp4S, temp5S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1607 line6S = vec_add(temp6S, temp7S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1608 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
1609
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1610 line0BS = vec_add(line0S, line2S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1611 line2BS = vec_sub(line0S, line2S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1612 line1BS = vec_add(line1S, line3S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1613 line3BS = vec_sub(line1S, line3S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1614 line4BS = vec_add(line4S, line6S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1615 line6BS = vec_sub(line4S, line6S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1616 line5BS = vec_add(line5S, line7S);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1617 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
1618
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1619 line0CS = vec_add(line0BS, line4BS);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1620 line4CS = vec_sub(line0BS, line4BS);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1621 line1CS = vec_add(line1BS, line5BS);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1622 line5CS = vec_sub(line1BS, line5BS);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1623 line2CS = vec_add(line2BS, line6BS);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1624 line6CS = vec_sub(line2BS, line6BS);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1625 line3CS = vec_add(line3BS, line7BS);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3252
diff changeset
1626 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
1627
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1628 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
1629 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
1630 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
1631 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
1632 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
1633 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
1634 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
1635 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
1636 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
1637 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
1638 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
1639 }
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1640 return sum;
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1641 }
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1642
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1643 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
1644 POWERPC_PERF_DECLARE(altivec_hadamard8_diff16_num, 1);
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1645 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
1646 POWERPC_PERF_START_COUNT(altivec_hadamard8_diff16_num, 1);
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1647 score = hadamard8_diff16x8_altivec(s, dst, src, stride, 8);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1648 if (h==16) {
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1649 dst += 8*stride;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1650 src += 8*stride;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1651 score += hadamard8_diff16x8_altivec(s, dst, src, stride, 8);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1652 }
1951
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1653 POWERPC_PERF_STOP_COUNT(altivec_hadamard8_diff16_num, 1);
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1654 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
1655 }
2599b8444831 better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1949
diff changeset
1656
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
1657 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
1658 {
2286
0f937b191cc3 Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents: 2068
diff changeset
1659 #ifdef __AMIGAOS4__
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1660 ULONG result = 0;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1661 extern struct ExecIFace *IExec;
2286
0f937b191cc3 Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents: 2068
diff changeset
1662
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1663 IExec->GetCPUInfoTags(GCIT_VectorUnit, &result, TAG_DONE);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1664 if (result == VECTORTYPE_ALTIVEC) return 1;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1665 return 0;
2286
0f937b191cc3 Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents: 2068
diff changeset
1666 #else /* __AMIGAOS4__ */
0f937b191cc3 Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents: 2068
diff changeset
1667
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1668 #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
1669 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
1670 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
1671 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
1672 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
1673
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
1674 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
1675
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff changeset
1676 if (err == 0) return (has_vu != 0);
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1677 #else /* CONFIG_DARWIN */
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1678 /* no Darwin, do it the brute-force way */
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1679 /* this is borrowed from the libmpeg2 library */
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1680 {
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1681 signal (SIGILL, sigill_handler);
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1682 if (sigsetjmp (jmpbuf, 1)) {
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1683 signal (SIGILL, SIG_DFL);
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1684 } else {
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1685 canjump = 1;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1686
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1687 asm volatile ("mtspr 256, %0\n\t"
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1688 "vand %%v0, %%v0, %%v0"
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1689 :
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1690 : "r" (-1));
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1691
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1692 signal (SIGILL, SIG_DFL);
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1693 return 1;
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1694 }
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1695 }
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1024
diff changeset
1696 #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
1697 return 0;
2286
0f937b191cc3 Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents: 2068
diff changeset
1698 #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
1699 }
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
1700
3543
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1701 static void vorbis_inverse_coupling_altivec(float *mag, float *ang,
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1702 int blocksize)
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1703 {
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1704 int i;
3546
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1705 vector float m, a;
3543
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1706 vector bool int t0, t1;
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1707 const vector unsigned int v_31 = //XXX
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1708 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
1709 for(i=0; i<blocksize; i+=4) {
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1710 m = vec_ld(0, mag+i);
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1711 a = vec_ld(0, ang+i);
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1712 t0 = vec_cmple(m, (vector float)vec_splat_u32(0));
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1713 t1 = vec_cmple(a, (vector float)vec_splat_u32(0));
3545
e2a589e55906 Minor fix
lu_zero
parents: 3543
diff changeset
1714 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
1715 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
1716 t1 = (vector bool int)vec_andc(a, t1);
3550
4f4c13574ad5 Yet another typo
lu_zero
parents: 3549
diff changeset
1717 a = vec_sub(m, (vector float)t1);
4f4c13574ad5 Yet another typo
lu_zero
parents: 3549
diff changeset
1718 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
1719 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
1720 vec_stl(m, 0, mag+i);
3543
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1721 }
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1722 }
6aabb2bec46c vorbis_inverse_coupling_altivec
lu_zero
parents: 3533
diff changeset
1723
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
1724 /* 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
1725 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
1726 {
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
1727 POWERPC_PERF_DECLARE(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
1728 #ifdef ALTIVEC_USE_REFERENCE_C_CODE
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
1729
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
1730 int j;
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
1731 POWERPC_PERF_START_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
1732 for (j = 0; j < 2; j++) {
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
1733 int i;
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
1734 const uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
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
1735 const uint32_t b = (((const struct unaligned_32 *) (pixels + 1))->l);
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
1736 uint32_t l0 = (a & 0x03030303UL) + (b & 0x03030303UL) + 0x02020202UL;
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
1737 uint32_t h0 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
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
1738 uint32_t l1, h1;
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
1739 pixels += line_size;
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
1740 for (i = 0; i < h; i += 2) {
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
1741 uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
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
1742 uint32_t b = (((const struct unaligned_32 *) (pixels + 1))->l);
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
1743 l1 = (a & 0x03030303UL) + (b & 0x03030303UL);
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
1744 h1 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
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
1745 *((uint32_t *) block) = rnd_avg32(*((uint32_t *) block), h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL));
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
1746 pixels += line_size;
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
1747 block += line_size;
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
1748 a = (((const struct unaligned_32 *) (pixels))->l);
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
1749 b = (((const struct unaligned_32 *) (pixels + 1))->l);
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
1750 l0 = (a & 0x03030303UL) + (b & 0x03030303UL) + 0x02020202UL;
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
1751 h0 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
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
1752 *((uint32_t *) block) = rnd_avg32(*((uint32_t *) block), h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL));
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
1753 pixels += line_size;
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
1754 block += line_size;
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
1755 } pixels += 4 - line_size * (h + 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
1756 block += 4 - line_size * 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
1757 }
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
1758 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
1759 #else /* ALTIVEC_USE_REFERENCE_C_CODE */
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1760 register int i;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1761 register vector unsigned char pixelsv1, pixelsv2, pixelsavg;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1762 register vector unsigned char blockv, temp1, temp2, blocktemp;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1763 register vector unsigned short pixelssum1, pixelssum2, temp3;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1764
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1765 register const_vector unsigned char vczero = (const_vector unsigned char)
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1766 vec_splat_u8(0);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1767 register const_vector unsigned short vctwo = (const_vector unsigned short)
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1768 vec_splat_u16(2);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1769
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1770 temp1 = vec_ld(0, pixels);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1771 temp2 = vec_ld(16, pixels);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1772 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1773 if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F) {
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1774 pixelsv2 = temp2;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1775 } else {
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1776 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1777 }
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1778 pixelsv1 = vec_mergeh(vczero, pixelsv1);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1779 pixelsv2 = vec_mergeh(vczero, pixelsv2);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1780 pixelssum1 = vec_add((vector unsigned short)pixelsv1,
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1781 (vector unsigned short)pixelsv2);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1782 pixelssum1 = vec_add(pixelssum1, vctwo);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1783
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1784 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
1785 for (i = 0; i < h ; i++) {
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1786 int rightside = ((unsigned long)block & 0x0000000F);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1787 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
1788
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1789 temp1 = vec_ld(line_size, pixels);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1790 temp2 = vec_ld(line_size + 16, pixels);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1791 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1792 if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F)
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1793 {
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1794 pixelsv2 = temp2;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1795 } else {
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1796 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
1797 }
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
1798
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1799 pixelsv1 = vec_mergeh(vczero, pixelsv1);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1800 pixelsv2 = vec_mergeh(vczero, pixelsv2);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1801 pixelssum2 = vec_add((vector unsigned short)pixelsv1,
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1802 (vector unsigned short)pixelsv2);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1803 temp3 = vec_add(pixelssum1, pixelssum2);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1804 temp3 = vec_sra(temp3, vctwo);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1805 pixelssum1 = vec_add(pixelssum2, vctwo);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1806 pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1807
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1808 if (rightside) {
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1809 blocktemp = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1));
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1810 } else {
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1811 blocktemp = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3));
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1812 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1813
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1814 blockv = vec_avg(blocktemp, blockv);
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1815 vec_st(blockv, 0, block);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1816
3554
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1817 block += line_size;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1818 pixels += line_size;
ce5554dd79ce Cosmetics: 2->4 spaces and some braces
lu_zero
parents: 3550
diff changeset
1819 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2286
diff changeset
1820
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
1821 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
1822 #endif /* ALTIVEC_USE_REFERENCE_C_CODE */
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
1823 }
3546
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1824
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1825 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
1826 {
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1827 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
1828 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
1829 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
1830 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
1831 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
1832 c->sad[0]= sad16_altivec;
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1833 c->sad[1]= sad8_altivec;
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1834 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
1835 c->sse[1]= sse8_altivec;
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1836 c->sse[0]= sse16_altivec;
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1837 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
1838 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
1839 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
1840 // next one disabled as it's untested.
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1841 #if 0
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1842 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
1843 #endif /* 0 */
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1844 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
1845 /* 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
1846 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
1847 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
1848 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
1849 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
1850 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
1851 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
1852 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
1853 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
1854
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1855 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
1856 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
1857 #ifdef CONFIG_VORBIS_DECODER
3546
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1858 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
1859 #endif
3546
5f97ba9a4eaa Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents: 3545
diff changeset
1860 }