Mercurial > libavcodec.hg
annotate ppc/dsputil_altivec.c @ 3963:075b3e6bdc05 libavcodec
Fixes:
rv10.c:494: warning: implicit declaration of function ¡Æff_mpeg4_init_direct_mv¡Ç
Approved by Michael.
author | rathann |
---|---|
date | Sun, 08 Oct 2006 21:26:19 +0000 |
parents | c8c591fe26f8 |
children | c86c7a54ba92 |
rev | line source |
---|---|
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
1 /* |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
2 * Copyright (c) 2002 Brian Foley |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
3 * Copyright (c) 2002 Dieter Shirley |
1949
66215baae7b9
hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1839
diff
changeset
|
4 * Copyright (c) 2003-2004 Romain Dolbeau <romain@dolbeau.org> |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
5 * |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3572
diff
changeset
|
6 * This file is part of FFmpeg. |
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3572
diff
changeset
|
7 * |
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3572
diff
changeset
|
8 * FFmpeg is free software; you can redistribute it and/or |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
9 * modify it under the terms of the GNU Lesser General Public |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
10 * License as published by the Free Software Foundation; either |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3572
diff
changeset
|
11 * version 2.1 of the License, or (at your option) any later version. |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
12 * |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3572
diff
changeset
|
13 * FFmpeg is distributed in the hope that it will be useful, |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
16 * Lesser General Public License for more details. |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
17 * |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
18 * You should have received a copy of the GNU Lesser General Public |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3572
diff
changeset
|
19 * License along with FFmpeg; if not, write to the Free Software |
3036
0b546eab515d
Update licensing information: The FSF changed postal address.
diego
parents:
2979
diff
changeset
|
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
21 */ |
2967 | 22 |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
23 #include "../dsputil.h" |
1277
f3152eb76f1a
altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents:
1064
diff
changeset
|
24 |
f3152eb76f1a
altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents:
1064
diff
changeset
|
25 #include "gcc_fixes.h" |
f3152eb76f1a
altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents:
1064
diff
changeset
|
26 |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
27 #include "dsputil_altivec.h" |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
28 |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
29 #ifdef CONFIG_DARWIN |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
30 #include <sys/sysctl.h> |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
31 #else /* CONFIG_DARWIN */ |
2286
0f937b191cc3
Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents:
2068
diff
changeset
|
32 #ifdef __AMIGAOS4__ |
0f937b191cc3
Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents:
2068
diff
changeset
|
33 #include <exec/exec.h> |
0f937b191cc3
Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents:
2068
diff
changeset
|
34 #include <interfaces/exec.h> |
0f937b191cc3
Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents:
2068
diff
changeset
|
35 #include <proto/exec.h> |
0f937b191cc3
Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents:
2068
diff
changeset
|
36 #else /* __AMIGAOS4__ */ |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
37 #include <signal.h> |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
38 #include <setjmp.h> |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
39 |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
40 static sigjmp_buf jmpbuf; |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
41 static volatile sig_atomic_t canjump = 0; |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
42 |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
43 static void sigill_handler (int sig) |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
44 { |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
45 if (!canjump) { |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
46 signal (sig, SIG_DFL); |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
47 raise (sig); |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
48 } |
2967 | 49 |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
50 canjump = 0; |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
51 siglongjmp (jmpbuf, 1); |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
52 } |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
53 #endif /* CONFIG_DARWIN */ |
2286
0f937b191cc3
Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents:
2068
diff
changeset
|
54 #endif /* __AMIGAOS4__ */ |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
55 |
1708 | 56 int sad16_x2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h) |
878
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
57 { |
981 | 58 int i; |
59 int s __attribute__((aligned(16))); | |
1839
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1708
diff
changeset
|
60 const_vector unsigned char zero = (const_vector unsigned char)vec_splat_u8(0); |
981 | 61 vector unsigned char *tv; |
878
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
62 vector unsigned char pix1v, pix2v, pix2iv, avgv, t5; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
63 vector unsigned int sad; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
64 vector signed int sumdiffs; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
65 |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
66 s = 0; |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
67 sad = (vector unsigned int)vec_splat_u32(0); |
1708 | 68 for(i=0;i<h;i++) { |
878
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
69 /* |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
70 Read unaligned pixels into our vectors. The vectors are as follows: |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
71 pix1v: pix1[0]-pix1[15] |
2979 | 72 pix2v: pix2[0]-pix2[15] pix2iv: pix2[1]-pix2[16] |
878
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
73 */ |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
74 tv = (vector unsigned char *) pix1; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
75 pix1v = vec_perm(tv[0], tv[1], vec_lvsl(0, pix1)); |
2967 | 76 |
878
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
77 tv = (vector unsigned char *) &pix2[0]; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
78 pix2v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[0])); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
79 |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
80 tv = (vector unsigned char *) &pix2[1]; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
81 pix2iv = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[1])); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
82 |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
83 /* Calculate the average vector */ |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
84 avgv = vec_avg(pix2v, pix2iv); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
85 |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
86 /* Calculate a sum of abs differences vector */ |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
87 t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv)); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
88 |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
89 /* Add each 4 pixel group together and put 4 results into sad */ |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
90 sad = vec_sum4s(t5, sad); |
2967 | 91 |
878
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
92 pix1 += line_size; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
93 pix2 += line_size; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
94 } |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
95 /* Sum up the four partial sums, and put the result into s */ |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
96 sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
97 sumdiffs = vec_splat(sumdiffs, 3); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
98 vec_ste(sumdiffs, 0, &s); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
99 |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
100 return s; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
101 } |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
102 |
1708 | 103 int sad16_y2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h) |
878
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
104 { |
981 | 105 int i; |
106 int s __attribute__((aligned(16))); | |
1839
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1708
diff
changeset
|
107 const_vector unsigned char zero = (const_vector unsigned char)vec_splat_u8(0); |
981 | 108 vector unsigned char *tv; |
878
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
109 vector unsigned char pix1v, pix2v, pix3v, avgv, t5; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
110 vector unsigned int sad; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
111 vector signed int sumdiffs; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
112 uint8_t *pix3 = pix2 + line_size; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
113 |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
114 s = 0; |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
115 sad = (vector unsigned int)vec_splat_u32(0); |
878
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
116 |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
117 /* |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
118 Due to the fact that pix3 = pix2 + line_size, the pix3 of one |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
119 iteration becomes pix2 in the next iteration. We can use this |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
120 fact to avoid a potentially expensive unaligned read, each |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
121 time around the loop. |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
122 Read unaligned pixels into our vectors. The vectors are as follows: |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
123 pix2v: pix2[0]-pix2[15] |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
124 Split the pixel vectors into shorts |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
125 */ |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
126 tv = (vector unsigned char *) &pix2[0]; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
127 pix2v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[0])); |
2967 | 128 |
1708 | 129 for(i=0;i<h;i++) { |
878
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
130 /* |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
131 Read unaligned pixels into our vectors. The vectors are as follows: |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
132 pix1v: pix1[0]-pix1[15] |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
133 pix3v: pix3[0]-pix3[15] |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
134 */ |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
135 tv = (vector unsigned char *) pix1; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
136 pix1v = vec_perm(tv[0], tv[1], vec_lvsl(0, pix1)); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
137 |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
138 tv = (vector unsigned char *) &pix3[0]; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
139 pix3v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix3[0])); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
140 |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
141 /* Calculate the average vector */ |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
142 avgv = vec_avg(pix2v, pix3v); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
143 |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
144 /* Calculate a sum of abs differences vector */ |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
145 t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv)); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
146 |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
147 /* Add each 4 pixel group together and put 4 results into sad */ |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
148 sad = vec_sum4s(t5, sad); |
2967 | 149 |
878
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
150 pix1 += line_size; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
151 pix2v = pix3v; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
152 pix3 += line_size; |
2967 | 153 |
878
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
154 } |
2967 | 155 |
878
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
156 /* Sum up the four partial sums, and put the result into s */ |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
157 sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
158 sumdiffs = vec_splat(sumdiffs, 3); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
159 vec_ste(sumdiffs, 0, &s); |
2967 | 160 return s; |
878
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
161 } |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
162 |
1708 | 163 int sad16_xy2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h) |
878
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
164 { |
981 | 165 int i; |
166 int s __attribute__((aligned(16))); | |
878
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
167 uint8_t *pix3 = pix2 + line_size; |
1839
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1708
diff
changeset
|
168 const_vector unsigned char zero = (const_vector unsigned char)vec_splat_u8(0); |
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1708
diff
changeset
|
169 const_vector unsigned short two = (const_vector unsigned short)vec_splat_u16(2); |
981 | 170 vector unsigned char *tv, avgv, t5; |
878
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
171 vector unsigned char pix1v, pix2v, pix3v, pix2iv, pix3iv; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
172 vector unsigned short pix2lv, pix2hv, pix2ilv, pix2ihv; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
173 vector unsigned short pix3lv, pix3hv, pix3ilv, pix3ihv; |
981 | 174 vector unsigned short avghv, avglv; |
878
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
175 vector unsigned short t1, t2, t3, t4; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
176 vector unsigned int sad; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
177 vector signed int sumdiffs; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
178 |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
179 sad = (vector unsigned int)vec_splat_u32(0); |
2967 | 180 |
878
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
181 s = 0; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
182 |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
183 /* |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
184 Due to the fact that pix3 = pix2 + line_size, the pix3 of one |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
185 iteration becomes pix2 in the next iteration. We can use this |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
186 fact to avoid a potentially expensive unaligned read, as well |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
187 as some splitting, and vector addition each time around the loop. |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
188 Read unaligned pixels into our vectors. The vectors are as follows: |
2979 | 189 pix2v: pix2[0]-pix2[15] pix2iv: pix2[1]-pix2[16] |
878
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
190 Split the pixel vectors into shorts |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
191 */ |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
192 tv = (vector unsigned char *) &pix2[0]; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
193 pix2v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[0])); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
194 |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
195 tv = (vector unsigned char *) &pix2[1]; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
196 pix2iv = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[1])); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
197 |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
198 pix2hv = (vector unsigned short) vec_mergeh(zero, pix2v); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
199 pix2lv = (vector unsigned short) vec_mergel(zero, pix2v); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
200 pix2ihv = (vector unsigned short) vec_mergeh(zero, pix2iv); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
201 pix2ilv = (vector unsigned short) vec_mergel(zero, pix2iv); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
202 t1 = vec_add(pix2hv, pix2ihv); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
203 t2 = vec_add(pix2lv, pix2ilv); |
2967 | 204 |
1708 | 205 for(i=0;i<h;i++) { |
878
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
206 /* |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
207 Read unaligned pixels into our vectors. The vectors are as follows: |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
208 pix1v: pix1[0]-pix1[15] |
2979 | 209 pix3v: pix3[0]-pix3[15] pix3iv: pix3[1]-pix3[16] |
878
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
210 */ |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
211 tv = (vector unsigned char *) pix1; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
212 pix1v = vec_perm(tv[0], tv[1], vec_lvsl(0, pix1)); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
213 |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
214 tv = (vector unsigned char *) &pix3[0]; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
215 pix3v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix3[0])); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
216 |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
217 tv = (vector unsigned char *) &pix3[1]; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
218 pix3iv = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix3[1])); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
219 |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
220 /* |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
221 Note that Altivec does have vec_avg, but this works on vector pairs |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
222 and rounds up. We could do avg(avg(a,b),avg(c,d)), but the rounding |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
223 would mean that, for example, avg(3,0,0,1) = 2, when it should be 1. |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
224 Instead, we have to split the pixel vectors into vectors of shorts, |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
225 and do the averaging by hand. |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
226 */ |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
227 |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
228 /* Split the pixel vectors into shorts */ |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
229 pix3hv = (vector unsigned short) vec_mergeh(zero, pix3v); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
230 pix3lv = (vector unsigned short) vec_mergel(zero, pix3v); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
231 pix3ihv = (vector unsigned short) vec_mergeh(zero, pix3iv); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
232 pix3ilv = (vector unsigned short) vec_mergel(zero, pix3iv); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
233 |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
234 /* Do the averaging on them */ |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
235 t3 = vec_add(pix3hv, pix3ihv); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
236 t4 = vec_add(pix3lv, pix3ilv); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
237 |
884 | 238 avghv = vec_sr(vec_add(vec_add(t1, t3), two), two); |
239 avglv = vec_sr(vec_add(vec_add(t2, t4), two), two); | |
878
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
240 |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
241 /* Pack the shorts back into a result */ |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
242 avgv = vec_pack(avghv, avglv); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
243 |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
244 /* Calculate a sum of abs differences vector */ |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
245 t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv)); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
246 |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
247 /* Add each 4 pixel group together and put 4 results into sad */ |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
248 sad = vec_sum4s(t5, sad); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
249 |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
250 pix1 += line_size; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
251 pix3 += line_size; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
252 /* Transfer the calculated values for pix3 into pix2 */ |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
253 t1 = t3; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
254 t2 = t4; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
255 } |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
256 /* Sum up the four partial sums, and put the result into s */ |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
257 sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
258 sumdiffs = vec_splat(sumdiffs, 3); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
259 vec_ste(sumdiffs, 0, &s); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
260 |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
261 return s; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
262 } |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
263 |
1708 | 264 int sad16_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h) |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
265 { |
981 | 266 int i; |
267 int s __attribute__((aligned(16))); | |
1839
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1708
diff
changeset
|
268 const_vector unsigned int zero = (const_vector unsigned int)vec_splat_u32(0); |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
269 vector unsigned char perm1, perm2, *pix1v, *pix2v; |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
270 vector unsigned char t1, t2, t3,t4, t5; |
981 | 271 vector unsigned int sad; |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
272 vector signed int sumdiffs; |
2967 | 273 |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
274 sad = (vector unsigned int)vec_splat_u32(0); |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
275 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
276 |
1708 | 277 for(i=0;i<h;i++) { |
2979 | 278 /* Read potentially unaligned pixels into t1 and t2 */ |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
279 perm1 = vec_lvsl(0, pix1); |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
280 pix1v = (vector unsigned char *) pix1; |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
281 perm2 = vec_lvsl(0, pix2); |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
282 pix2v = (vector unsigned char *) pix2; |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
283 t1 = vec_perm(pix1v[0], pix1v[1], perm1); |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
284 t2 = vec_perm(pix2v[0], pix2v[1], perm2); |
2967 | 285 |
2979 | 286 /* Calculate a sum of abs differences vector */ |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
287 t3 = vec_max(t1, t2); |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
288 t4 = vec_min(t1, t2); |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
289 t5 = vec_sub(t3, t4); |
2967 | 290 |
2979 | 291 /* Add each 4 pixel group together and put 4 results into sad */ |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
292 sad = vec_sum4s(t5, sad); |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
293 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
294 pix1 += line_size; |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
295 pix2 += line_size; |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
296 } |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
297 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
298 /* Sum up the four partial sums, and put the result into s */ |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
299 sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero); |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
300 sumdiffs = vec_splat(sumdiffs, 3); |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
301 vec_ste(sumdiffs, 0, &s); |
2967 | 302 |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
303 return s; |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
304 } |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
305 |
1708 | 306 int sad8_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h) |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
307 { |
981 | 308 int i; |
309 int s __attribute__((aligned(16))); | |
1839
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1708
diff
changeset
|
310 const_vector unsigned int zero = (const_vector unsigned int)vec_splat_u32(0); |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
311 vector unsigned char perm1, perm2, permclear, *pix1v, *pix2v; |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
312 vector unsigned char t1, t2, t3,t4, t5; |
981 | 313 vector unsigned int sad; |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
314 vector signed int sumdiffs; |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
315 |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
316 sad = (vector unsigned int)vec_splat_u32(0); |
1277
f3152eb76f1a
altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents:
1064
diff
changeset
|
317 |
f3152eb76f1a
altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents:
1064
diff
changeset
|
318 permclear = (vector unsigned char)AVV(255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0); |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
319 |
1708 | 320 for(i=0;i<h;i++) { |
2979 | 321 /* Read potentially unaligned pixels into t1 and t2 |
322 Since we're reading 16 pixels, and actually only want 8, | |
323 mask out the last 8 pixels. The 0s don't change the sum. */ | |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
324 perm1 = vec_lvsl(0, pix1); |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
325 pix1v = (vector unsigned char *) pix1; |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
326 perm2 = vec_lvsl(0, pix2); |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
327 pix2v = (vector unsigned char *) pix2; |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
328 t1 = vec_and(vec_perm(pix1v[0], pix1v[1], perm1), permclear); |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
329 t2 = vec_and(vec_perm(pix2v[0], pix2v[1], perm2), permclear); |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
330 |
2979 | 331 /* Calculate a sum of abs differences vector */ |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
332 t3 = vec_max(t1, t2); |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
333 t4 = vec_min(t1, t2); |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
334 t5 = vec_sub(t3, t4); |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
335 |
2979 | 336 /* Add each 4 pixel group together and put 4 results into sad */ |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
337 sad = vec_sum4s(t5, sad); |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
338 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
339 pix1 += line_size; |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
340 pix2 += line_size; |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
341 } |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
342 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
343 /* Sum up the four partial sums, and put the result into s */ |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
344 sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero); |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
345 sumdiffs = vec_splat(sumdiffs, 3); |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
346 vec_ste(sumdiffs, 0, &s); |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
347 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
348 return s; |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
349 } |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
350 |
878
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
351 int pix_norm1_altivec(uint8_t *pix, int line_size) |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
352 { |
981 | 353 int i; |
354 int s __attribute__((aligned(16))); | |
1839
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1708
diff
changeset
|
355 const_vector unsigned int zero = (const_vector unsigned int)vec_splat_u32(0); |
981 | 356 vector unsigned char *tv; |
878
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
357 vector unsigned char pixv; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
358 vector unsigned int sv; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
359 vector signed int sum; |
2967 | 360 |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
361 sv = (vector unsigned int)vec_splat_u32(0); |
2967 | 362 |
878
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
363 s = 0; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
364 for (i = 0; i < 16; i++) { |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
365 /* Read in the potentially unaligned pixels */ |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
366 tv = (vector unsigned char *) pix; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
367 pixv = vec_perm(tv[0], tv[1], vec_lvsl(0, pix)); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
368 |
884 | 369 /* Square the values, and add them to our sum */ |
370 sv = vec_msum(pixv, pixv, sv); | |
878
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
371 |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
372 pix += line_size; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
373 } |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
374 /* Sum up the four partial sums, and put the result into s */ |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
375 sum = vec_sums((vector signed int) sv, (vector signed int) zero); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
376 sum = vec_splat(sum, 3); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
377 vec_ste(sum, 0, &s); |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
378 |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
379 return s; |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
380 } |
6ea69518e5f7
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
828
diff
changeset
|
381 |
981 | 382 /** |
383 * Sum of Squared Errors for a 8x8 block. | |
384 * AltiVec-enhanced. | |
1708 | 385 * It's the sad8_altivec code above w/ squaring added. |
981 | 386 */ |
1708 | 387 int sse8_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h) |
981 | 388 { |
389 int i; | |
390 int s __attribute__((aligned(16))); | |
1839
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1708
diff
changeset
|
391 const_vector unsigned int zero = (const_vector unsigned int)vec_splat_u32(0); |
981 | 392 vector unsigned char perm1, perm2, permclear, *pix1v, *pix2v; |
393 vector unsigned char t1, t2, t3,t4, t5; | |
394 vector unsigned int sum; | |
395 vector signed int sumsqr; | |
2967 | 396 |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
397 sum = (vector unsigned int)vec_splat_u32(0); |
1277
f3152eb76f1a
altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents:
1064
diff
changeset
|
398 |
f3152eb76f1a
altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents:
1064
diff
changeset
|
399 permclear = (vector unsigned char)AVV(255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0); |
f3152eb76f1a
altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents:
1064
diff
changeset
|
400 |
2967 | 401 |
1708 | 402 for(i=0;i<h;i++) { |
2979 | 403 /* Read potentially unaligned pixels into t1 and t2 |
404 Since we're reading 16 pixels, and actually only want 8, | |
405 mask out the last 8 pixels. The 0s don't change the sum. */ | |
981 | 406 perm1 = vec_lvsl(0, pix1); |
407 pix1v = (vector unsigned char *) pix1; | |
408 perm2 = vec_lvsl(0, pix2); | |
409 pix2v = (vector unsigned char *) pix2; | |
410 t1 = vec_and(vec_perm(pix1v[0], pix1v[1], perm1), permclear); | |
411 t2 = vec_and(vec_perm(pix2v[0], pix2v[1], perm2), permclear); | |
412 | |
413 /* | |
414 Since we want to use unsigned chars, we can take advantage | |
415 of the fact that abs(a-b)^2 = (a-b)^2. | |
416 */ | |
2967 | 417 |
2979 | 418 /* Calculate abs differences vector */ |
981 | 419 t3 = vec_max(t1, t2); |
420 t4 = vec_min(t1, t2); | |
421 t5 = vec_sub(t3, t4); | |
2967 | 422 |
981 | 423 /* Square the values and add them to our sum */ |
424 sum = vec_msum(t5, t5, sum); | |
2967 | 425 |
981 | 426 pix1 += line_size; |
427 pix2 += line_size; | |
428 } | |
2967 | 429 |
981 | 430 /* Sum up the four partial sums, and put the result into s */ |
431 sumsqr = vec_sums((vector signed int) sum, (vector signed int) zero); | |
432 sumsqr = vec_splat(sumsqr, 3); | |
433 vec_ste(sumsqr, 0, &s); | |
2967 | 434 |
981 | 435 return s; |
436 } | |
437 | |
438 /** | |
439 * Sum of Squared Errors for a 16x16 block. | |
440 * AltiVec-enhanced. | |
1708 | 441 * It's the sad16_altivec code above w/ squaring added. |
981 | 442 */ |
1708 | 443 int sse16_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h) |
981 | 444 { |
445 int i; | |
446 int s __attribute__((aligned(16))); | |
1839
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1708
diff
changeset
|
447 const_vector unsigned int zero = (const_vector unsigned int)vec_splat_u32(0); |
981 | 448 vector unsigned char perm1, perm2, *pix1v, *pix2v; |
449 vector unsigned char t1, t2, t3,t4, t5; | |
450 vector unsigned int sum; | |
451 vector signed int sumsqr; | |
2967 | 452 |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
453 sum = (vector unsigned int)vec_splat_u32(0); |
2967 | 454 |
1708 | 455 for(i=0;i<h;i++) { |
2979 | 456 /* Read potentially unaligned pixels into t1 and t2 */ |
981 | 457 perm1 = vec_lvsl(0, pix1); |
458 pix1v = (vector unsigned char *) pix1; | |
459 perm2 = vec_lvsl(0, pix2); | |
460 pix2v = (vector unsigned char *) pix2; | |
461 t1 = vec_perm(pix1v[0], pix1v[1], perm1); | |
462 t2 = vec_perm(pix2v[0], pix2v[1], perm2); | |
463 | |
464 /* | |
465 Since we want to use unsigned chars, we can take advantage | |
466 of the fact that abs(a-b)^2 = (a-b)^2. | |
467 */ | |
2967 | 468 |
2979 | 469 /* Calculate abs differences vector */ |
981 | 470 t3 = vec_max(t1, t2); |
471 t4 = vec_min(t1, t2); | |
472 t5 = vec_sub(t3, t4); | |
2967 | 473 |
981 | 474 /* Square the values and add them to our sum */ |
475 sum = vec_msum(t5, t5, sum); | |
2967 | 476 |
981 | 477 pix1 += line_size; |
478 pix2 += line_size; | |
479 } | |
2967 | 480 |
981 | 481 /* Sum up the four partial sums, and put the result into s */ |
482 sumsqr = vec_sums((vector signed int) sum, (vector signed int) zero); | |
483 sumsqr = vec_splat(sumsqr, 3); | |
484 vec_ste(sumsqr, 0, &s); | |
2967 | 485 |
981 | 486 return s; |
487 } | |
488 | |
1064 | 489 int pix_sum_altivec(uint8_t * pix, int line_size) |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
490 { |
1839
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1708
diff
changeset
|
491 const_vector unsigned int zero = (const_vector unsigned int)vec_splat_u32(0); |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
492 vector unsigned char perm, *pixv; |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
493 vector unsigned char t1; |
981 | 494 vector unsigned int sad; |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
495 vector signed int sumdiffs; |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
496 |
981 | 497 int i; |
498 int s __attribute__((aligned(16))); | |
2967 | 499 |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
500 sad = (vector unsigned int)vec_splat_u32(0); |
2967 | 501 |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
502 for (i = 0; i < 16; i++) { |
2979 | 503 /* Read the potentially unaligned 16 pixels into t1 */ |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
504 perm = vec_lvsl(0, pix); |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
505 pixv = (vector unsigned char *) pix; |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
506 t1 = vec_perm(pixv[0], pixv[1], perm); |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
507 |
2979 | 508 /* Add each 4 pixel group together and put 4 results into sad */ |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
509 sad = vec_sum4s(t1, sad); |
2967 | 510 |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
511 pix += line_size; |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
512 } |
2967 | 513 |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
514 /* Sum up the four partial sums, and put the result into s */ |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
515 sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero); |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
516 sumdiffs = vec_splat(sumdiffs, 3); |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
517 vec_ste(sumdiffs, 0, &s); |
2967 | 518 |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
519 return s; |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
520 } |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
521 |
1064 | 522 void get_pixels_altivec(DCTELEM *restrict block, const uint8_t *pixels, int line_size) |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
523 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
524 int i; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
525 vector unsigned char perm, bytes, *pixv; |
1839
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1708
diff
changeset
|
526 const_vector unsigned char zero = (const_vector unsigned char)vec_splat_u8(0); |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
527 vector signed short shorts; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
528 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
529 for(i=0;i<8;i++) |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
530 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
531 // Read potentially unaligned pixels. |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
532 // We're reading 16 pixels, and actually only want 8, |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
533 // but we simply ignore the extras. |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
534 perm = vec_lvsl(0, pixels); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
535 pixv = (vector unsigned char *) pixels; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
536 bytes = vec_perm(pixv[0], pixv[1], perm); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
537 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
538 // convert the bytes into shorts |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
539 shorts = (vector signed short)vec_mergeh(zero, bytes); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
540 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
541 // save the data to the block, we assume the block is 16-byte aligned |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
542 vec_st(shorts, i*16, (vector signed short*)block); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
543 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
544 pixels += line_size; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
545 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
546 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
547 |
1064 | 548 void diff_pixels_altivec(DCTELEM *restrict block, const uint8_t *s1, |
549 const uint8_t *s2, int stride) | |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
550 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
551 int i; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
552 vector unsigned char perm, bytes, *pixv; |
1839
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1708
diff
changeset
|
553 const_vector unsigned char zero = (const_vector unsigned char)vec_splat_u8(0); |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
554 vector signed short shorts1, shorts2; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
555 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
556 for(i=0;i<4;i++) |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
557 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
558 // Read potentially unaligned pixels |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
559 // We're reading 16 pixels, and actually only want 8, |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
560 // but we simply ignore the extras. |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
561 perm = vec_lvsl(0, s1); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
562 pixv = (vector unsigned char *) s1; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
563 bytes = vec_perm(pixv[0], pixv[1], perm); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
564 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
565 // convert the bytes into shorts |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
566 shorts1 = (vector signed short)vec_mergeh(zero, bytes); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
567 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
568 // Do the same for the second block of pixels |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
569 perm = vec_lvsl(0, s2); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
570 pixv = (vector unsigned char *) s2; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
571 bytes = vec_perm(pixv[0], pixv[1], perm); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
572 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
573 // convert the bytes into shorts |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
574 shorts2 = (vector signed short)vec_mergeh(zero, bytes); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
575 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
576 // Do the subtraction |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
577 shorts1 = vec_sub(shorts1, shorts2); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
578 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
579 // save the data to the block, we assume the block is 16-byte aligned |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
580 vec_st(shorts1, 0, (vector signed short*)block); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
581 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
582 s1 += stride; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
583 s2 += stride; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
584 block += 8; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
585 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
586 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
587 // The code below is a copy of the code above... This is a manual |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
588 // unroll. |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
589 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
590 // Read potentially unaligned pixels |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
591 // We're reading 16 pixels, and actually only want 8, |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
592 // but we simply ignore the extras. |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
593 perm = vec_lvsl(0, s1); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
594 pixv = (vector unsigned char *) s1; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
595 bytes = vec_perm(pixv[0], pixv[1], perm); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
596 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
597 // convert the bytes into shorts |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
598 shorts1 = (vector signed short)vec_mergeh(zero, bytes); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
599 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
600 // Do the same for the second block of pixels |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
601 perm = vec_lvsl(0, s2); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
602 pixv = (vector unsigned char *) s2; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
603 bytes = vec_perm(pixv[0], pixv[1], perm); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
604 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
605 // convert the bytes into shorts |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
606 shorts2 = (vector signed short)vec_mergeh(zero, bytes); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
607 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
608 // Do the subtraction |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
609 shorts1 = vec_sub(shorts1, shorts2); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
610 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
611 // save the data to the block, we assume the block is 16-byte aligned |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
612 vec_st(shorts1, 0, (vector signed short*)block); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
613 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
614 s1 += stride; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
615 s2 += stride; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
616 block += 8; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
617 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
618 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
619 |
995
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
981
diff
changeset
|
620 void add_bytes_altivec(uint8_t *dst, uint8_t *src, int w) { |
1009
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
621 #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
|
622 int i; |
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
981
diff
changeset
|
623 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
|
624 dst[i+0] += src[i+0]; |
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
981
diff
changeset
|
625 dst[i+1] += src[i+1]; |
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
981
diff
changeset
|
626 dst[i+2] += src[i+2]; |
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
981
diff
changeset
|
627 dst[i+3] += src[i+3]; |
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
981
diff
changeset
|
628 dst[i+4] += src[i+4]; |
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
981
diff
changeset
|
629 dst[i+5] += src[i+5]; |
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
981
diff
changeset
|
630 dst[i+6] += src[i+6]; |
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
981
diff
changeset
|
631 dst[i+7] += src[i+7]; |
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
981
diff
changeset
|
632 } |
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
981
diff
changeset
|
633 for(; i<w; i++) |
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
981
diff
changeset
|
634 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
|
635 #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
|
636 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
|
637 register vector unsigned char vdst, vsrc; |
2967 | 638 |
1009
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
639 /* 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
|
640 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
|
641 { |
1009
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_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
|
643 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
|
644 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
|
645 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
|
646 } |
1009
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
647 /* 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
|
648 for (; (i < w) ; i++) |
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
981
diff
changeset
|
649 { |
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
981
diff
changeset
|
650 dst[i] = src[i]; |
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
981
diff
changeset
|
651 } |
1009
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
652 #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
|
653 } |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
654 |
1024
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
655 /* 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
|
656 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
|
657 { |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
658 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
|
659 #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
|
660 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
|
661 |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
662 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
|
663 |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
664 for(i=0; i<h; i++) { |
1839
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1708
diff
changeset
|
665 *((uint32_t*)(block)) = LD32(pixels); |
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1708
diff
changeset
|
666 *((uint32_t*)(block+4)) = LD32(pixels+4); |
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1708
diff
changeset
|
667 *((uint32_t*)(block+8)) = LD32(pixels+8); |
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1708
diff
changeset
|
668 *((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
|
669 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
|
670 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
|
671 } |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
672 |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
673 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
|
674 |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
675 #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
|
676 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
|
677 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
|
678 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
|
679 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
|
680 |
1024
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
681 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
|
682 int i; |
1352
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_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
|
684 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
|
685 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
|
686 |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
687 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
|
688 // 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
|
689 // 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
|
690 // 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
|
691 // -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
|
692 // 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
|
693 #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
|
694 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
|
695 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
|
696 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
|
697 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
|
698 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
|
699 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
|
700 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
|
701 } |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
702 #else |
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
703 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
|
704 pixelsv1 = vec_ld(0, (unsigned char*)pixels); |
3533 | 705 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
|
706 pixelsv1B = vec_ld(line_size, (unsigned char*)pixels); |
3533 | 707 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
|
708 pixelsv1C = vec_ld(line_size_2, (unsigned char*)pixels); |
3533 | 709 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
|
710 pixelsv1D = vec_ld(line_size_3, (unsigned char*)pixels); |
3533 | 711 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
|
712 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
|
713 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
|
714 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
|
715 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
|
716 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
|
717 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
|
718 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
|
719 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
|
720 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
|
721 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
|
722 } |
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
723 #endif |
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
724 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
|
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 #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
|
727 } |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
728 |
1024
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
729 /* 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
|
730 #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
|
731 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
|
732 { |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
733 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
|
734 #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
|
735 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
|
736 |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
737 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
|
738 |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
739 for(i=0; i<h; i++) { |
1839
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1708
diff
changeset
|
740 op_avg(*((uint32_t*)(block)),LD32(pixels)); |
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1708
diff
changeset
|
741 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
|
742 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
|
743 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
|
744 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
|
745 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
|
746 } |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
747 |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
748 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
|
749 |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
750 #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
|
751 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
|
752 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
|
753 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
|
754 |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
755 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
|
756 |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
757 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
|
758 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
|
759 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
|
760 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
|
761 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
|
762 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
|
763 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
|
764 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
|
765 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
|
766 } |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
767 |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
768 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
|
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 #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
|
771 } |
35cf2f4a0f8c
PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1009
diff
changeset
|
772 |
1024
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
773 /* 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
|
774 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
|
775 { |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
776 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
|
777 #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
|
778 int i; |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
779 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
|
780 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
|
781 *((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
|
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))->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
|
784 ((((*((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
|
785 ((((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
|
786 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
|
787 *((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
|
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 + 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
|
790 ((((*((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
|
791 ((((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
|
792 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
|
793 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
|
794 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
|
795 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
|
796 } |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
797 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
|
798 |
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 #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
|
800 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
|
801 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
|
802 |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
803 POWERPC_PERF_START_COUNT(altivec_avg_pixels8_num, 1); |
2967 | 804 |
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
|
805 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
|
806 /* |
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 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
|
808 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
|
809 */ |
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 int rightside = ((unsigned long)block & 0x0000000F); |
2967 | 811 |
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
|
812 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
|
813 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
|
814 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
|
815 pixelsv = vec_perm(pixelsv1, pixelsv2, vec_lvsl(0, pixels)); |
2967 | 816 |
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
|
817 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
|
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 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
|
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 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
|
822 { |
35cf2f4a0f8c
PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1009
diff
changeset
|
823 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
|
824 } |
2967 | 825 |
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
|
826 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
|
827 |
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 vec_st(blockv, 0, block); |
2967 | 829 |
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
|
830 pixels += line_size; |
35cf2f4a0f8c
PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1009
diff
changeset
|
831 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
|
832 } |
2967 | 833 |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
834 POWERPC_PERF_STOP_COUNT(altivec_avg_pixels8_num, 1); |
2967 | 835 |
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
|
836 #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
|
837 } |
1009
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
838 |
1024
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
839 /* 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
|
840 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
|
841 { |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
842 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
|
843 #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
|
844 int j; |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
845 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
|
846 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
|
847 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
|
848 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
|
849 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
|
850 (((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
|
851 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
|
852 (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
|
853 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
|
854 ((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
|
855 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
|
856 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
|
857 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
|
858 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
|
859 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
|
860 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
|
861 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
|
862 *((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
|
863 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
|
864 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
|
865 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
|
866 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
|
867 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
|
868 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
|
869 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
|
870 *((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
|
871 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
|
872 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
|
873 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
|
874 } 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
|
875 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
|
876 } |
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 |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
878 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
|
879 |
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 #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
|
881 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
|
882 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
|
883 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
|
884 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
|
885 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
|
886 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
|
887 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
|
888 pixelssum1, pixelssum2, temp3; |
1839
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1708
diff
changeset
|
889 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
|
890 register const_vector unsigned short vctwo = (const_vector unsigned short)vec_splat_u16(2); |
2967 | 891 |
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
|
892 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
|
893 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
|
894 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
|
895 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
|
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 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
|
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 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
|
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 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
|
902 } |
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 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
|
904 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
|
905 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
|
906 (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
|
907 pixelssum1 = vec_add(pixelssum1, vctwo); |
2967 | 908 |
909 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
|
910 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
|
911 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
|
912 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
|
913 |
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 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
|
915 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
|
916 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
|
917 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
|
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 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
|
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 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
|
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 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
|
924 } |
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 |
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 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
|
927 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
|
928 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
|
929 (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
|
930 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
|
931 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
|
932 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
|
933 pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero); |
2967 | 934 |
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
|
935 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
|
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 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
|
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 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
|
940 { |
35cf2f4a0f8c
PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1009
diff
changeset
|
941 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
|
942 } |
2967 | 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 vec_st(blockv, 0, block); |
2967 | 945 |
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
|
946 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
|
947 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
|
948 } |
2967 | 949 |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
950 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
|
951 #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
|
952 } |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
638
diff
changeset
|
953 |
1024
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
954 /* 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
|
955 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
|
956 { |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
957 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
|
958 #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
|
959 int j; |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
960 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
|
961 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
|
962 int i; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
963 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
|
964 const uint32_t b = |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
965 (((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
|
966 uint32_t l0 = |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
967 (a & 0x03030303UL) + (b & 0x03030303UL) + 0x01010101UL; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
968 uint32_t h0 = |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
969 ((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
|
970 uint32_t l1, h1; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
971 pixels += line_size; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
972 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
|
973 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
|
974 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
|
975 l1 = (a & 0x03030303UL) + (b & 0x03030303UL); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
976 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
|
977 *((uint32_t *) block) = |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
978 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
|
979 pixels += line_size; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
980 block += line_size; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
981 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
|
982 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
|
983 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
|
984 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
|
985 *((uint32_t *) block) = |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
986 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
|
987 pixels += line_size; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
988 block += line_size; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
989 } pixels += 4 - line_size * (h + 1); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
990 block += 4 - line_size * h; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
991 } |
2967 | 992 |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
993 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
|
994 |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
995 #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
|
996 register int i; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
997 register vector unsigned char |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
998 pixelsv1, pixelsv2, |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
999 pixelsavg; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1000 register vector unsigned char |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1001 blockv, temp1, temp2; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1002 register vector unsigned short |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1003 pixelssum1, pixelssum2, temp3; |
1839
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1708
diff
changeset
|
1004 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
|
1005 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
|
1006 register const_vector unsigned short vctwo = (const_vector unsigned short)vec_splat_u16(2); |
2967 | 1007 |
1024
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1008 temp1 = vec_ld(0, pixels); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1009 temp2 = vec_ld(16, pixels); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1010 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
|
1011 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
|
1012 { |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1013 pixelsv2 = temp2; |
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 else |
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 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
|
1018 } |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1019 pixelsv1 = vec_mergeh(vczero, pixelsv1); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1020 pixelsv2 = vec_mergeh(vczero, pixelsv2); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1021 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
|
1022 (vector unsigned short)pixelsv2); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1023 pixelssum1 = vec_add(pixelssum1, vcone); |
2967 | 1024 |
1025 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
|
1026 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
|
1027 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
|
1028 blockv = vec_ld(0, block); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1029 |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1030 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
|
1031 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
|
1032 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
|
1033 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
|
1034 { |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1035 pixelsv2 = temp2; |
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 else |
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 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
|
1040 } |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1041 |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1042 pixelsv1 = vec_mergeh(vczero, pixelsv1); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1043 pixelsv2 = vec_mergeh(vczero, pixelsv2); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1044 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
|
1045 (vector unsigned short)pixelsv2); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1046 temp3 = vec_add(pixelssum1, pixelssum2); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1047 temp3 = vec_sra(temp3, vctwo); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1048 pixelssum1 = vec_add(pixelssum2, vcone); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1049 pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero); |
2967 | 1050 |
1024
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1051 if (rightside) |
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 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
|
1054 } |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1055 else |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1056 { |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1057 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
|
1058 } |
2967 | 1059 |
1024
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1060 vec_st(blockv, 0, block); |
2967 | 1061 |
1024
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1062 block += line_size; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1063 pixels += line_size; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1064 } |
2967 | 1065 |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
1066 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
|
1067 #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
|
1068 } |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1069 |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1070 /* 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
|
1071 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
|
1072 { |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
1073 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
|
1074 #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
|
1075 int j; |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
1076 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
|
1077 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
|
1078 int i; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1079 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
|
1080 const uint32_t b = |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1081 (((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
|
1082 uint32_t l0 = |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1083 (a & 0x03030303UL) + (b & 0x03030303UL) + 0x02020202UL; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1084 uint32_t h0 = |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1085 ((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
|
1086 uint32_t l1, h1; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1087 pixels += line_size; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1088 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
|
1089 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
|
1090 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
|
1091 l1 = (a & 0x03030303UL) + (b & 0x03030303UL); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1092 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
|
1093 *((uint32_t *) block) = |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1094 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
|
1095 pixels += line_size; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1096 block += line_size; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1097 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
|
1098 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
|
1099 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
|
1100 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
|
1101 *((uint32_t *) block) = |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1102 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
|
1103 pixels += line_size; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1104 block += line_size; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1105 } 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
|
1106 block += 4 - line_size * h; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1107 } |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1108 |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
1109 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
|
1110 |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1111 #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
|
1112 register int i; |
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 pixelsv1, pixelsv2, pixelsv3, pixelsv4; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1115 register vector unsigned char |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1116 blockv, temp1, temp2; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1117 register vector unsigned short |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1118 pixelssum1, pixelssum2, temp3, |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1119 pixelssum3, pixelssum4, temp4; |
1839
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1708
diff
changeset
|
1120 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
|
1121 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
|
1122 |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
1123 POWERPC_PERF_START_COUNT(altivec_put_pixels16_xy2_num, 1); |
2967 | 1124 |
1024
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1125 temp1 = vec_ld(0, pixels); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1126 temp2 = vec_ld(16, pixels); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1127 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
|
1128 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
|
1129 { |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1130 pixelsv2 = temp2; |
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 else |
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 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
|
1135 } |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1136 pixelsv3 = vec_mergel(vczero, pixelsv1); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1137 pixelsv4 = vec_mergel(vczero, pixelsv2); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1138 pixelsv1 = vec_mergeh(vczero, pixelsv1); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1139 pixelsv2 = vec_mergeh(vczero, pixelsv2); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1140 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
|
1141 (vector unsigned short)pixelsv4); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1142 pixelssum3 = vec_add(pixelssum3, vctwo); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1143 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
|
1144 (vector unsigned short)pixelsv2); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1145 pixelssum1 = vec_add(pixelssum1, vctwo); |
2967 | 1146 |
1024
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1147 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
|
1148 blockv = vec_ld(0, block); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1149 |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1150 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
|
1151 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
|
1152 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
|
1153 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
|
1154 { |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1155 pixelsv2 = temp2; |
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 else |
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 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
|
1160 } |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1161 |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1162 pixelsv3 = vec_mergel(vczero, pixelsv1); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1163 pixelsv4 = vec_mergel(vczero, pixelsv2); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1164 pixelsv1 = vec_mergeh(vczero, pixelsv1); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1165 pixelsv2 = vec_mergeh(vczero, pixelsv2); |
2967 | 1166 |
1024
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1167 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
|
1168 (vector unsigned short)pixelsv4); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1169 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
|
1170 (vector unsigned short)pixelsv2); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1171 temp4 = vec_add(pixelssum3, pixelssum4); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1172 temp4 = vec_sra(temp4, vctwo); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1173 temp3 = vec_add(pixelssum1, pixelssum2); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1174 temp3 = vec_sra(temp3, vctwo); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1175 |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1176 pixelssum3 = vec_add(pixelssum4, vctwo); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1177 pixelssum1 = vec_add(pixelssum2, vctwo); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1178 |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1179 blockv = vec_packsu(temp3, temp4); |
2967 | 1180 |
1024
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1181 vec_st(blockv, 0, block); |
2967 | 1182 |
1024
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1183 block += line_size; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1184 pixels += line_size; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1185 } |
2967 | 1186 |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
1187 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
|
1188 #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
|
1189 } |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1190 |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1191 /* 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
|
1192 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
|
1193 { |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
1194 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
|
1195 #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
|
1196 int j; |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
1197 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
|
1198 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
|
1199 int i; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1200 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
|
1201 const uint32_t b = |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1202 (((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
|
1203 uint32_t l0 = |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1204 (a & 0x03030303UL) + (b & 0x03030303UL) + 0x01010101UL; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1205 uint32_t h0 = |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1206 ((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
|
1207 uint32_t l1, h1; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1208 pixels += line_size; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1209 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
|
1210 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
|
1211 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
|
1212 l1 = (a & 0x03030303UL) + (b & 0x03030303UL); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1213 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
|
1214 *((uint32_t *) block) = |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1215 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
|
1216 pixels += line_size; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1217 block += line_size; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1218 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
|
1219 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
|
1220 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
|
1221 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
|
1222 *((uint32_t *) block) = |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1223 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
|
1224 pixels += line_size; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1225 block += line_size; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1226 } 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
|
1227 block += 4 - line_size * h; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1228 } |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1229 |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
1230 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
|
1231 |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1232 #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
|
1233 register int i; |
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 pixelsv1, pixelsv2, pixelsv3, pixelsv4; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1236 register vector unsigned char |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1237 blockv, temp1, temp2; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1238 register vector unsigned short |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1239 pixelssum1, pixelssum2, temp3, |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1240 pixelssum3, pixelssum4, temp4; |
1839
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1708
diff
changeset
|
1241 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
|
1242 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
|
1243 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
|
1244 |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
1245 POWERPC_PERF_START_COUNT(altivec_put_no_rnd_pixels16_xy2_num, 1); |
2967 | 1246 |
1024
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1247 temp1 = vec_ld(0, pixels); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1248 temp2 = vec_ld(16, pixels); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1249 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
|
1250 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
|
1251 { |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1252 pixelsv2 = temp2; |
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 else |
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 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
|
1257 } |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1258 pixelsv3 = vec_mergel(vczero, pixelsv1); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1259 pixelsv4 = vec_mergel(vczero, pixelsv2); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1260 pixelsv1 = vec_mergeh(vczero, pixelsv1); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1261 pixelsv2 = vec_mergeh(vczero, pixelsv2); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1262 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
|
1263 (vector unsigned short)pixelsv4); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1264 pixelssum3 = vec_add(pixelssum3, vcone); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1265 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
|
1266 (vector unsigned short)pixelsv2); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1267 pixelssum1 = vec_add(pixelssum1, vcone); |
2967 | 1268 |
1024
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1269 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
|
1270 blockv = vec_ld(0, block); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1271 |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1272 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
|
1273 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
|
1274 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
|
1275 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
|
1276 { |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1277 pixelsv2 = temp2; |
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 else |
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 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
|
1282 } |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1283 |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1284 pixelsv3 = vec_mergel(vczero, pixelsv1); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1285 pixelsv4 = vec_mergel(vczero, pixelsv2); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1286 pixelsv1 = vec_mergeh(vczero, pixelsv1); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1287 pixelsv2 = vec_mergeh(vczero, pixelsv2); |
2967 | 1288 |
1024
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1289 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
|
1290 (vector unsigned short)pixelsv4); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1291 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
|
1292 (vector unsigned short)pixelsv2); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1293 temp4 = vec_add(pixelssum3, pixelssum4); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1294 temp4 = vec_sra(temp4, vctwo); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1295 temp3 = vec_add(pixelssum1, pixelssum2); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1296 temp3 = vec_sra(temp3, vctwo); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1297 |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1298 pixelssum3 = vec_add(pixelssum4, vcone); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1299 pixelssum1 = vec_add(pixelssum2, vcone); |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1300 |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1301 blockv = vec_packsu(temp3, temp4); |
2967 | 1302 |
1024
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1303 vec_st(blockv, 0, block); |
2967 | 1304 |
1024
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1305 block += line_size; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1306 pixels += line_size; |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1307 } |
2967 | 1308 |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1340
diff
changeset
|
1309 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
|
1310 #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
|
1311 } |
9cc1031e1864
More AltiVec MC functions patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1015
diff
changeset
|
1312 |
1949
66215baae7b9
hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1839
diff
changeset
|
1313 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
|
1314 POWERPC_PERF_DECLARE(altivec_hadamard8_diff8x8_num, 1); |
3554 | 1315 int sum; |
1316 register const_vector unsigned char vzero = | |
1317 (const_vector unsigned char)vec_splat_u8(0); | |
1318 register vector signed short temp0, temp1, temp2, temp3, temp4, | |
1319 temp5, temp6, temp7; | |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1320 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
|
1321 { |
3554 | 1322 register const_vector signed short vprod1 =(const_vector signed short) |
1323 AVV( 1,-1, 1,-1, 1,-1, 1,-1); | |
1324 register const_vector signed short vprod2 =(const_vector signed short) | |
1325 AVV( 1, 1,-1,-1, 1, 1,-1,-1); | |
1326 register const_vector signed short vprod3 =(const_vector signed short) | |
1327 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
|
1328 register const_vector unsigned char perm1 = (const_vector unsigned char) |
3554 | 1329 AVV(0x02, 0x03, 0x00, 0x01, 0x06, 0x07, 0x04, 0x05, |
1330 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
|
1331 register const_vector unsigned char perm2 = (const_vector unsigned char) |
3554 | 1332 AVV(0x04, 0x05, 0x06, 0x07, 0x00, 0x01, 0x02, 0x03, |
1333 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
|
1334 register const_vector unsigned char perm3 = (const_vector unsigned char) |
3554 | 1335 AVV(0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F, |
1336 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
|
1337 |
2979 | 1338 #define ONEITERBUTTERFLY(i, res) \ |
1339 { \ | |
1340 register vector unsigned char src1, src2, srcO; \ | |
1341 register vector unsigned char dst1, dst2, dstO; \ | |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1342 register vector signed short srcV, dstV; \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1343 register vector signed short but0, but1, but2, op1, op2, op3; \ |
2979 | 1344 src1 = vec_ld(stride * i, src); \ |
1345 if ((((stride * i) + (unsigned long)src) & 0x0000000F) > 8) \ | |
1346 src2 = vec_ld((stride * i) + 16, src); \ | |
1347 srcO = vec_perm(src1, src2, vec_lvsl(stride * i, src)); \ | |
1348 dst1 = vec_ld(stride * i, dst); \ | |
1349 if ((((stride * i) + (unsigned long)dst) & 0x0000000F) > 8) \ | |
1350 dst2 = vec_ld((stride * i) + 16, dst); \ | |
1351 dstO = vec_perm(dst1, dst2, vec_lvsl(stride * i, dst)); \ | |
1352 /* promote the unsigned chars to signed shorts */ \ | |
1353 /* 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
|
1354 srcV = \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1355 (vector signed short)vec_mergeh((vector signed char)vzero, \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1356 (vector signed char)srcO); \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1357 dstV = \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1358 (vector signed short)vec_mergeh((vector signed char)vzero, \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1359 (vector signed char)dstO); \ |
2979 | 1360 /* substractions inside the first butterfly */ \ |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1361 but0 = vec_sub(srcV, dstV); \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1362 op1 = vec_perm(but0, but0, perm1); \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1363 but1 = vec_mladd(but0, vprod1, op1); \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1364 op2 = vec_perm(but1, but1, perm2); \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1365 but2 = vec_mladd(but1, vprod2, op2); \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1366 op3 = vec_perm(but2, but2, perm3); \ |
2979 | 1367 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
|
1368 } |
66215baae7b9
hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1839
diff
changeset
|
1369 ONEITERBUTTERFLY(0, temp0); |
66215baae7b9
hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1839
diff
changeset
|
1370 ONEITERBUTTERFLY(1, temp1); |
66215baae7b9
hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1839
diff
changeset
|
1371 ONEITERBUTTERFLY(2, temp2); |
66215baae7b9
hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1839
diff
changeset
|
1372 ONEITERBUTTERFLY(3, temp3); |
66215baae7b9
hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1839
diff
changeset
|
1373 ONEITERBUTTERFLY(4, temp4); |
66215baae7b9
hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1839
diff
changeset
|
1374 ONEITERBUTTERFLY(5, temp5); |
66215baae7b9
hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1839
diff
changeset
|
1375 ONEITERBUTTERFLY(6, temp6); |
66215baae7b9
hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1839
diff
changeset
|
1376 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
|
1377 } |
1949
66215baae7b9
hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1839
diff
changeset
|
1378 #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
|
1379 { |
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 int vsum; |
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 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
|
1382 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
|
1383 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
|
1384 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
|
1385 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
|
1386 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
|
1387 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
|
1388 register vector signed short line7 = vec_sub(temp6, temp7); |
2967 | 1389 |
1951
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 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
|
1391 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
|
1392 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
|
1393 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
|
1394 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
|
1395 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
|
1396 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
|
1397 register vector signed short line7B = vec_sub(line5, line7); |
2967 | 1398 |
1951
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 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
|
1400 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
|
1401 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
|
1402 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
|
1403 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
|
1404 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
|
1405 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
|
1406 register vector signed short line7C = vec_sub(line3B, line7B); |
2967 | 1407 |
1951
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(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
|
1409 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
|
1410 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
|
1411 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
|
1412 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
|
1413 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
|
1414 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
|
1415 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
|
1416 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
|
1417 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
|
1418 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
|
1419 } |
66215baae7b9
hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1839
diff
changeset
|
1420 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
|
1421 return sum; |
66215baae7b9
hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1839
diff
changeset
|
1422 } |
66215baae7b9
hadamard8_diff8x8 in AltiVec, the 16bits edition by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1839
diff
changeset
|
1423 |
1951
2599b8444831
better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1949
diff
changeset
|
1424 /* |
2599b8444831
better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1949
diff
changeset
|
1425 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
|
1426 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
|
1427 It's only used from inside hadamard8_diff16_altivec. |
2967 | 1428 |
1951
2599b8444831
better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1949
diff
changeset
|
1429 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
|
1430 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
|
1431 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
|
1432 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
|
1433 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
|
1434 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
|
1435 fall from 700+ cycles to 550). |
2967 | 1436 |
1951
2599b8444831
better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1949
diff
changeset
|
1437 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
|
1438 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
|
1439 gcc-3.3 on the 7450 (but uses 25% less instructions...) |
2967 | 1440 |
1951
2599b8444831
better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1949
diff
changeset
|
1441 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
|
1442 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
|
1443 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
|
1444 */ |
2599b8444831
better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1949
diff
changeset
|
1445 |
2599b8444831
better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1949
diff
changeset
|
1446 static int hadamard8_diff16x8_altivec(/*MpegEncContext*/ void *s, uint8_t *dst, uint8_t *src, int stride, int h) { |
3554 | 1447 int sum; |
1448 register vector signed short | |
1449 temp0 REG_v(v0), | |
1450 temp1 REG_v(v1), | |
1451 temp2 REG_v(v2), | |
1452 temp3 REG_v(v3), | |
1453 temp4 REG_v(v4), | |
1454 temp5 REG_v(v5), | |
1455 temp6 REG_v(v6), | |
1456 temp7 REG_v(v7); | |
1457 register vector signed short | |
1458 temp0S REG_v(v8), | |
1459 temp1S REG_v(v9), | |
1460 temp2S REG_v(v10), | |
1461 temp3S REG_v(v11), | |
1462 temp4S REG_v(v12), | |
1463 temp5S REG_v(v13), | |
1464 temp6S REG_v(v14), | |
1465 temp7S REG_v(v15); | |
1466 register const_vector unsigned char vzero REG_v(v31)= | |
1467 (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
|
1468 { |
3554 | 1469 register const_vector signed short vprod1 REG_v(v16)= |
1470 (const_vector signed short)AVV( 1,-1, 1,-1, 1,-1, 1,-1); | |
1471 register const_vector signed short vprod2 REG_v(v17)= | |
1472 (const_vector signed short)AVV( 1, 1,-1,-1, 1, 1,-1,-1); | |
1473 register const_vector signed short vprod3 REG_v(v18)= | |
1474 (const_vector signed short)AVV( 1, 1, 1, 1,-1,-1,-1,-1); | |
1475 register const_vector unsigned char perm1 REG_v(v19)= | |
1476 (const_vector unsigned char) | |
1477 AVV(0x02, 0x03, 0x00, 0x01, 0x06, 0x07, 0x04, 0x05, | |
1478 0x0A, 0x0B, 0x08, 0x09, 0x0E, 0x0F, 0x0C, 0x0D); | |
1479 register const_vector unsigned char perm2 REG_v(v20)= | |
1480 (const_vector unsigned char) | |
1481 AVV(0x04, 0x05, 0x06, 0x07, 0x00, 0x01, 0x02, 0x03, | |
1482 0x0C, 0x0D, 0x0E, 0x0F, 0x08, 0x09, 0x0A, 0x0B); | |
1483 register const_vector unsigned char perm3 REG_v(v21)= | |
1484 (const_vector unsigned char) | |
1485 AVV(0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F, | |
1486 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07); | |
1980 | 1487 |
2979 | 1488 #define ONEITERBUTTERFLY(i, res1, res2) \ |
1489 { \ | |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1490 register vector unsigned char src1 REG_v(v22), \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1491 src2 REG_v(v23), \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1492 dst1 REG_v(v24), \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1493 dst2 REG_v(v25), \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1494 srcO REG_v(v22), \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1495 dstO REG_v(v23); \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1496 \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1497 register vector signed short srcV REG_v(v24), \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1498 dstV REG_v(v25), \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1499 srcW REG_v(v26), \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1500 dstW REG_v(v27), \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1501 but0 REG_v(v28), \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1502 but0S REG_v(v29), \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1503 op1 REG_v(v30), \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1504 but1 REG_v(v22), \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1505 op1S REG_v(v23), \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1506 but1S REG_v(v24), \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1507 op2 REG_v(v25), \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1508 but2 REG_v(v26), \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1509 op2S REG_v(v27), \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1510 but2S REG_v(v28), \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1511 op3 REG_v(v29), \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1512 op3S REG_v(v30); \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1513 \ |
2979 | 1514 src1 = vec_ld(stride * i, src); \ |
1515 src2 = vec_ld((stride * i) + 16, src); \ | |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1516 srcO = vec_perm(src1, src2, vec_lvsl(stride * i, src)); \ |
2979 | 1517 dst1 = vec_ld(stride * i, dst); \ |
1518 dst2 = vec_ld((stride * i) + 16, dst); \ | |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1519 dstO = vec_perm(dst1, dst2, vec_lvsl(stride * i, dst)); \ |
2979 | 1520 /* promote the unsigned chars to signed shorts */ \ |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1521 srcV = \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1522 (vector signed short)vec_mergeh((vector signed char)vzero, \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1523 (vector signed char)srcO); \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1524 dstV = \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1525 (vector signed short)vec_mergeh((vector signed char)vzero, \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1526 (vector signed char)dstO); \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1527 srcW = \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1528 (vector signed short)vec_mergel((vector signed char)vzero, \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1529 (vector signed char)srcO); \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1530 dstW = \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1531 (vector signed short)vec_mergel((vector signed char)vzero, \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1532 (vector signed char)dstO); \ |
2979 | 1533 /* substractions inside the first butterfly */ \ |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1534 but0 = vec_sub(srcV, dstV); \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1535 but0S = vec_sub(srcW, dstW); \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1536 op1 = vec_perm(but0, but0, perm1); \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1537 but1 = vec_mladd(but0, vprod1, op1); \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1538 op1S = vec_perm(but0S, but0S, perm1); \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1539 but1S = vec_mladd(but0S, vprod1, op1S); \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1540 op2 = vec_perm(but1, but1, perm2); \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1541 but2 = vec_mladd(but1, vprod2, op2); \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1542 op2S = vec_perm(but1S, but1S, perm2); \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1543 but2S = vec_mladd(but1S, vprod2, op2S); \ |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1544 op3 = vec_perm(but2, but2, perm3); \ |
2979 | 1545 res1 = vec_mladd(but2, vprod3, op3); \ |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1546 op3S = vec_perm(but2S, but2S, perm3); \ |
2979 | 1547 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
|
1548 } |
2599b8444831
better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1949
diff
changeset
|
1549 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
|
1550 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
|
1551 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
|
1552 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
|
1553 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
|
1554 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
|
1555 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
|
1556 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
|
1557 } |
2599b8444831
better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1949
diff
changeset
|
1558 #undef ONEITERBUTTERFLY |
2599b8444831
better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1949
diff
changeset
|
1559 { |
2599b8444831
better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1949
diff
changeset
|
1560 register vector signed int vsum; |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1561 register vector signed short line0S, line1S, line2S, line3S, line4S, |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1562 line5S, line6S, line7S, line0BS,line2BS, |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1563 line1BS,line3BS,line4BS,line6BS,line5BS, |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1564 line7BS,line0CS,line4CS,line1CS,line5CS, |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1565 line2CS,line6CS,line3CS,line7CS; |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1566 |
1951
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 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
|
1568 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
|
1569 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
|
1570 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
|
1571 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
|
1572 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
|
1573 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
|
1574 register vector signed short line7 = vec_sub(temp6, temp7); |
2967 | 1575 |
1951
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 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
|
1577 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
|
1578 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
|
1579 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
|
1580 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
|
1581 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
|
1582 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
|
1583 register vector signed short line7B = vec_sub(line5, line7); |
2967 | 1584 |
1951
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 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
|
1586 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
|
1587 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
|
1588 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
|
1589 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
|
1590 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
|
1591 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
|
1592 register vector signed short line7C = vec_sub(line3B, line7B); |
2967 | 1593 |
1951
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(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
|
1595 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
|
1596 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
|
1597 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
|
1598 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
|
1599 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
|
1600 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
|
1601 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
|
1602 |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1603 line0S = vec_add(temp0S, temp1S); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1604 line1S = vec_sub(temp0S, temp1S); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1605 line2S = vec_add(temp2S, temp3S); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1606 line3S = vec_sub(temp2S, temp3S); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1607 line4S = vec_add(temp4S, temp5S); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1608 line5S = vec_sub(temp4S, temp5S); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1609 line6S = vec_add(temp6S, temp7S); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1610 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
|
1611 |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1612 line0BS = vec_add(line0S, line2S); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1613 line2BS = vec_sub(line0S, line2S); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1614 line1BS = vec_add(line1S, line3S); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1615 line3BS = vec_sub(line1S, line3S); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1616 line4BS = vec_add(line4S, line6S); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1617 line6BS = vec_sub(line4S, line6S); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1618 line5BS = vec_add(line5S, line7S); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1619 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
|
1620 |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1621 line0CS = vec_add(line0BS, line4BS); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1622 line4CS = vec_sub(line0BS, line4BS); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1623 line1CS = vec_add(line1BS, line5BS); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1624 line5CS = vec_sub(line1BS, line5BS); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1625 line2CS = vec_add(line2BS, line6BS); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1626 line6CS = vec_sub(line2BS, line6BS); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1627 line3CS = vec_add(line3BS, line7BS); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3252
diff
changeset
|
1628 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
|
1629 |
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(line0CS), 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(line1CS), 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(line2CS), 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(line3CS), 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(line4CS), 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(line5CS), 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_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
|
1637 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
|
1638 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
|
1639 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
|
1640 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
|
1641 } |
2599b8444831
better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1949
diff
changeset
|
1642 return sum; |
2599b8444831
better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1949
diff
changeset
|
1643 } |
2599b8444831
better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1949
diff
changeset
|
1644 |
2599b8444831
better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1949
diff
changeset
|
1645 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
|
1646 POWERPC_PERF_DECLARE(altivec_hadamard8_diff16_num, 1); |
3554 | 1647 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
|
1648 POWERPC_PERF_START_COUNT(altivec_hadamard8_diff16_num, 1); |
3554 | 1649 score = hadamard8_diff16x8_altivec(s, dst, src, stride, 8); |
1650 if (h==16) { | |
1651 dst += 8*stride; | |
1652 src += 8*stride; | |
1653 score += hadamard8_diff16x8_altivec(s, dst, src, stride, 8); | |
1654 } | |
1951
2599b8444831
better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1949
diff
changeset
|
1655 POWERPC_PERF_STOP_COUNT(altivec_hadamard8_diff16_num, 1); |
3554 | 1656 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
|
1657 } |
2599b8444831
better hadamard8_diff16 in AltiVec, and more patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1949
diff
changeset
|
1658 |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
1659 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
|
1660 { |
2286
0f937b191cc3
Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents:
2068
diff
changeset
|
1661 #ifdef __AMIGAOS4__ |
3554 | 1662 ULONG result = 0; |
1663 extern struct ExecIFace *IExec; | |
2286
0f937b191cc3
Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents:
2068
diff
changeset
|
1664 |
3554 | 1665 IExec->GetCPUInfoTags(GCIT_VectorUnit, &result, TAG_DONE); |
1666 if (result == VECTORTYPE_ALTIVEC) return 1; | |
1667 return 0; | |
2286
0f937b191cc3
Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents:
2068
diff
changeset
|
1668 #else /* __AMIGAOS4__ */ |
0f937b191cc3
Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents:
2068
diff
changeset
|
1669 |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
1670 #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
|
1671 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
|
1672 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
|
1673 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
|
1674 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
|
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 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
|
1677 |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
diff
changeset
|
1678 if (err == 0) return (has_vu != 0); |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
1679 #else /* CONFIG_DARWIN */ |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
1680 /* no Darwin, do it the brute-force way */ |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
1681 /* this is borrowed from the libmpeg2 library */ |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
1682 { |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
1683 signal (SIGILL, sigill_handler); |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
1684 if (sigsetjmp (jmpbuf, 1)) { |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
1685 signal (SIGILL, SIG_DFL); |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
1686 } else { |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
1687 canjump = 1; |
2967 | 1688 |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
1689 asm volatile ("mtspr 256, %0\n\t" |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
1690 "vand %%v0, %%v0, %%v0" |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
1691 : |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
1692 : "r" (-1)); |
2967 | 1693 |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
1694 signal (SIGILL, SIG_DFL); |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
1695 return 1; |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
1696 } |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
1697 } |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1024
diff
changeset
|
1698 #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
|
1699 return 0; |
2286
0f937b191cc3
Altivec test on AmigaOS4 patch by (Chip <szarlada at freemail dot hu>)
michael
parents:
2068
diff
changeset
|
1700 #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
|
1701 } |
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
|
1702 |
3543 | 1703 static void vorbis_inverse_coupling_altivec(float *mag, float *ang, |
1704 int blocksize) | |
1705 { | |
1706 int i; | |
3546
5f97ba9a4eaa
Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents:
3545
diff
changeset
|
1707 vector float m, a; |
3543 | 1708 vector bool int t0, t1; |
1709 const vector unsigned int v_31 = //XXX | |
1710 vec_add(vec_add(vec_splat_u32(15),vec_splat_u32(15)),vec_splat_u32(1)); | |
1711 for(i=0; i<blocksize; i+=4) { | |
1712 m = vec_ld(0, mag+i); | |
1713 a = vec_ld(0, ang+i); | |
1714 t0 = vec_cmple(m, (vector float)vec_splat_u32(0)); | |
1715 t1 = vec_cmple(a, (vector float)vec_splat_u32(0)); | |
3545 | 1716 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
|
1717 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
|
1718 t1 = (vector bool int)vec_andc(a, t1); |
3550 | 1719 a = vec_sub(m, (vector float)t1); |
1720 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
|
1721 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
|
1722 vec_stl(m, 0, mag+i); |
3543 | 1723 } |
1724 } | |
1725 | |
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
|
1726 /* 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
|
1727 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
|
1728 { |
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 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
|
1730 #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
|
1731 |
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 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
|
1733 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
|
1734 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
|
1735 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
|
1736 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
|
1737 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
|
1738 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
|
1739 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
|
1740 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
|
1741 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
|
1742 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
|
1743 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
|
1744 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
|
1745 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
|
1746 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
|
1747 *((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
|
1748 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
|
1749 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
|
1750 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
|
1751 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
|
1752 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
|
1753 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
|
1754 *((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
|
1755 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
|
1756 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
|
1757 } 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
|
1758 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
|
1759 } |
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
|
1760 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
|
1761 #else /* ALTIVEC_USE_REFERENCE_C_CODE */ |
3554 | 1762 register int i; |
1763 register vector unsigned char pixelsv1, pixelsv2, pixelsavg; | |
1764 register vector unsigned char blockv, temp1, temp2, blocktemp; | |
1765 register vector unsigned short pixelssum1, pixelssum2, temp3; | |
1766 | |
1767 register const_vector unsigned char vczero = (const_vector unsigned char) | |
1768 vec_splat_u8(0); | |
1769 register const_vector unsigned short vctwo = (const_vector unsigned short) | |
1770 vec_splat_u16(2); | |
2967 | 1771 |
3554 | 1772 temp1 = vec_ld(0, pixels); |
1773 temp2 = vec_ld(16, pixels); | |
1774 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels)); | |
1775 if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F) { | |
1776 pixelsv2 = temp2; | |
1777 } else { | |
1778 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels)); | |
1779 } | |
1780 pixelsv1 = vec_mergeh(vczero, pixelsv1); | |
1781 pixelsv2 = vec_mergeh(vczero, pixelsv2); | |
1782 pixelssum1 = vec_add((vector unsigned short)pixelsv1, | |
1783 (vector unsigned short)pixelsv2); | |
1784 pixelssum1 = vec_add(pixelssum1, vctwo); | |
2967 | 1785 |
1786 POWERPC_PERF_START_COUNT(altivec_avg_pixels8_xy2_num, 1); | |
3554 | 1787 for (i = 0; i < h ; i++) { |
1788 int rightside = ((unsigned long)block & 0x0000000F); | |
1789 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
|
1790 |
3554 | 1791 temp1 = vec_ld(line_size, pixels); |
1792 temp2 = vec_ld(line_size + 16, pixels); | |
1793 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels)); | |
1794 if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F) | |
1795 { | |
1796 pixelsv2 = temp2; | |
1797 } else { | |
1798 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels)); | |
1799 } | |
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
|
1800 |
3554 | 1801 pixelsv1 = vec_mergeh(vczero, pixelsv1); |
1802 pixelsv2 = vec_mergeh(vczero, pixelsv2); | |
1803 pixelssum2 = vec_add((vector unsigned short)pixelsv1, | |
1804 (vector unsigned short)pixelsv2); | |
1805 temp3 = vec_add(pixelssum1, pixelssum2); | |
1806 temp3 = vec_sra(temp3, vctwo); | |
1807 pixelssum1 = vec_add(pixelssum2, vctwo); | |
1808 pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero); | |
2967 | 1809 |
3554 | 1810 if (rightside) { |
1811 blocktemp = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1)); | |
1812 } else { | |
1813 blocktemp = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3)); | |
1814 } | |
2967 | 1815 |
3554 | 1816 blockv = vec_avg(blocktemp, blockv); |
1817 vec_st(blockv, 0, block); | |
2967 | 1818 |
3554 | 1819 block += line_size; |
1820 pixels += line_size; | |
1821 } | |
2967 | 1822 |
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
|
1823 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
|
1824 #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
|
1825 } |
3546
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 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
|
1828 { |
5f97ba9a4eaa
Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents:
3545
diff
changeset
|
1829 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
|
1830 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
|
1831 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
|
1832 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
|
1833 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
|
1834 c->sad[0]= sad16_altivec; |
5f97ba9a4eaa
Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents:
3545
diff
changeset
|
1835 c->sad[1]= sad8_altivec; |
5f97ba9a4eaa
Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents:
3545
diff
changeset
|
1836 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
|
1837 c->sse[1]= sse8_altivec; |
5f97ba9a4eaa
Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents:
3545
diff
changeset
|
1838 c->sse[0]= sse16_altivec; |
5f97ba9a4eaa
Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents:
3545
diff
changeset
|
1839 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
|
1840 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
|
1841 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
|
1842 // 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
|
1843 #if 0 |
5f97ba9a4eaa
Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents:
3545
diff
changeset
|
1844 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
|
1845 #endif /* 0 */ |
5f97ba9a4eaa
Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents:
3545
diff
changeset
|
1846 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
|
1847 /* 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
|
1848 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
|
1849 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
|
1850 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
|
1851 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
|
1852 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
|
1853 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
|
1854 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
|
1855 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
|
1856 |
5f97ba9a4eaa
Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents:
3545
diff
changeset
|
1857 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
|
1858 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
|
1859 #ifdef CONFIG_VORBIS_DECODER |
3546
5f97ba9a4eaa
Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents:
3545
diff
changeset
|
1860 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
|
1861 #endif |
3546
5f97ba9a4eaa
Almost cosmetic changes in dsputil_init_ppc and vorbis_inverse_coupling_altivec:
lu_zero
parents:
3545
diff
changeset
|
1862 } |