Mercurial > libavcodec.hg
annotate ppc/dsputil_altivec.c @ 845:32de034be20e libavcodec
hurry_up>=5 -> skip everything except header & set pict_type&key_frame
author | michaelni |
---|---|
date | Sun, 10 Nov 2002 10:54:07 +0000 |
parents | ace3ccd18dd2 |
children | 6ea69518e5f7 |
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 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
4 * |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
5 * 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
|
6 * 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
|
7 * 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
|
8 * 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
|
9 * |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
10 * 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
|
11 * 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
|
12 * 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
|
13 * 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
|
14 * |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
15 * 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
|
16 * License along with this library; if not, write to the Free Software |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
18 */ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
19 |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
20 #include "../dsputil.h" |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
21 #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
|
22 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
23 #if CONFIG_DARWIN |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
24 #include <sys/sysctl.h> |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
25 #endif |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
26 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
27 int pix_abs16x16_altivec(uint8_t *pix1, uint8_t *pix2, int 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
|
28 { |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
29 int i, s; |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
30 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
|
31 vector unsigned char t1, t2, t3,t4, t5; |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
32 vector unsigned int sad, zero; |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
33 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
|
34 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
35 zero = (vector unsigned int) (0); |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
36 sad = (vector unsigned int) (0); |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
37 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
38 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
39 for(i=0;i<16;i++) { |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
40 /* Read potentially unaligned pixels into t1 and t2 */ |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
41 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
|
42 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
|
43 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
|
44 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
|
45 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
|
46 t2 = vec_perm(pix2v[0], pix2v[1], perm2); |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
47 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
48 /* Calculate a sum of abs differences vector */ |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
49 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
|
50 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
|
51 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
|
52 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
53 /* Add each 4 pixel group together and put 4 results into sad */ |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
54 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
|
55 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
56 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
|
57 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
|
58 } |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
59 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
60 /* 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
|
61 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
|
62 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
|
63 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
|
64 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
65 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
|
66 } |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
67 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
68 int pix_abs8x8_altivec(uint8_t *pix1, uint8_t *pix2, int 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
|
69 { |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
70 int i, s; |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
71 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
|
72 vector unsigned char t1, t2, t3,t4, t5; |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
73 vector unsigned int sad, zero; |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
74 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
|
75 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
76 zero = (vector unsigned int) (0); |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
77 sad = (vector unsigned int) (0); |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
78 permclear = (vector unsigned char) (255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0); |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
79 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
80 for(i=0;i<8;i++) { |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
81 /* Read potentially unaligned pixels into t1 and t2 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
82 Since we're reading 16 pixels, and actually only want 8, |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
83 mask out the last 8 pixels. The 0s don't change the sum. */ |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
84 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
|
85 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
|
86 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
|
87 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
|
88 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
|
89 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
|
90 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
91 /* Calculate a sum of abs differences vector */ |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
92 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
|
93 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
|
94 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
|
95 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
96 /* Add each 4 pixel group together and put 4 results into sad */ |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
97 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
|
98 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
99 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
|
100 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
|
101 } |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
102 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
103 /* 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
|
104 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
|
105 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
|
106 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
|
107 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
108 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
|
109 } |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
110 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
111 int pix_sum_altivec(UINT8 * pix, int 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
|
112 { |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
113 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
114 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
|
115 vector unsigned char t1; |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
116 vector unsigned int sad, zero; |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
117 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
|
118 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
119 int s, i; |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
120 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
121 zero = (vector unsigned int) (0); |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
122 sad = (vector unsigned int) (0); |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
123 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
124 for (i = 0; i < 16; i++) { |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
125 /* Read the potentially unaligned 16 pixels into t1 */ |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
126 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
|
127 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
|
128 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
|
129 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
130 /* Add each 4 pixel group together and put 4 results into sad */ |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
131 sad = vec_sum4s(t1, sad); |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
132 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
133 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
|
134 } |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
135 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
136 /* 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
|
137 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
|
138 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
|
139 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
|
140 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
141 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
|
142 } |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
143 |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
144 void get_pixels_altivec(DCTELEM *restrict block, const UINT8 *pixels, int line_size) |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
145 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
146 int i; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
147 vector unsigned char perm, bytes, *pixv; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
148 vector unsigned char zero = (vector unsigned char) (0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
149 vector signed short shorts; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
150 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
151 for(i=0;i<8;i++) |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
152 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
153 // Read potentially unaligned pixels. |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
154 // 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
|
155 // but we simply ignore the extras. |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
156 perm = vec_lvsl(0, pixels); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
157 pixv = (vector unsigned char *) pixels; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
158 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
|
159 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
160 // convert the bytes into shorts |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
161 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
|
162 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
163 // 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
|
164 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
|
165 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
166 pixels += line_size; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
167 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
168 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
169 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
170 void diff_pixels_altivec(DCTELEM *restrict block, const UINT8 *s1, |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
171 const UINT8 *s2, int stride) |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
172 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
173 int i; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
174 vector unsigned char perm, bytes, *pixv; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
175 vector unsigned char zero = (vector unsigned char) (0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
176 vector signed short shorts1, shorts2; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
177 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
178 for(i=0;i<4;i++) |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
179 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
180 // Read potentially unaligned pixels |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
181 // 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
|
182 // but we simply ignore the extras. |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
183 perm = vec_lvsl(0, s1); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
184 pixv = (vector unsigned char *) s1; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
185 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
|
186 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
187 // convert the bytes into shorts |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
188 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
|
189 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
190 // 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
|
191 perm = vec_lvsl(0, s2); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
192 pixv = (vector unsigned char *) s2; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
193 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
|
194 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
195 // convert the bytes into shorts |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
196 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
|
197 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
198 // Do the subtraction |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
199 shorts1 = vec_sub(shorts1, shorts2); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
200 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
201 // 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
|
202 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
|
203 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
204 s1 += stride; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
205 s2 += stride; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
206 block += 8; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
207 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
208 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
209 // 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
|
210 // unroll. |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
211 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
212 // Read potentially unaligned pixels |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
213 // 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
|
214 // but we simply ignore the extras. |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
215 perm = vec_lvsl(0, s1); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
216 pixv = (vector unsigned char *) s1; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
217 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
|
218 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
219 // convert the bytes into shorts |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
220 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
|
221 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
222 // 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
|
223 perm = vec_lvsl(0, s2); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
224 pixv = (vector unsigned char *) s2; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
225 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
|
226 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
227 // convert the bytes into shorts |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
228 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
|
229 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
230 // Do the subtraction |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
231 shorts1 = vec_sub(shorts1, shorts2); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
232 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
233 // 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
|
234 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
|
235 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
236 s1 += stride; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
237 s2 += stride; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
238 block += 8; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
239 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
240 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
241 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
242 |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
243 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
|
244 { |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
245 #if CONFIG_DARWIN |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
246 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
|
247 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
|
248 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
|
249 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
|
250 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
251 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
|
252 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
253 if (err == 0) return (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
|
254 #endif |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
255 return 0; |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
256 } |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
257 |