annotate ppc/mpegvideo_altivec.c @ 12512:58bd8cb86f5d libavcodec

Remove unused variable.
author rbultje
date Fri, 24 Sep 2010 15:31:46 +0000
parents 9fef0a8ddd63
children
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
1 /*
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
2 * Copyright (c) 2002 Dieter Shirley
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
3 *
1942
9241d99f7443 forgotten copyright & dct_quantize_altivec disabled patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
4 * dct_unquantize_h263_altivec:
9241d99f7443 forgotten copyright & dct_quantize_altivec disabled patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
5 * Copyright (c) 2003 Romain Dolbeau <romain@dolbeau.org>
9241d99f7443 forgotten copyright & dct_quantize_altivec disabled patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 1839
diff changeset
6 *
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3036
diff changeset
7 * This file is part of FFmpeg.
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3036
diff changeset
8 *
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3036
diff changeset
9 * 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:
diff changeset
10 * 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:
diff changeset
11 * 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: 3036
diff changeset
12 * 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:
diff changeset
13 *
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3036
diff changeset
14 * 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:
diff changeset
15 * 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:
diff changeset
16 * 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:
diff changeset
17 * Lesser General Public License for more details.
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
18 *
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
19 * 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: 3036
diff changeset
20 * 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
21 * 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:
diff changeset
22 */
1277
f3152eb76f1a altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents: 1107
diff changeset
23
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
24 #include <stdlib.h>
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
25 #include <stdio.h>
12475
9fef0a8ddd63 Move mm_support() from libavcodec to libavutil, make it a public
stefano
parents: 12473
diff changeset
26 #include "libavutil/cpu.h"
6763
f7cbb7733146 Use full path for #includes from another directory.
diego
parents: 6486
diff changeset
27 #include "libavcodec/dsputil.h"
f7cbb7733146 Use full path for #includes from another directory.
diego
parents: 6486
diff changeset
28 #include "libavcodec/mpegvideo.h"
1277
f3152eb76f1a altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents: 1107
diff changeset
29
5750
09f99af1db40 Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents: 5746
diff changeset
30 #include "util_altivec.h"
10079
71ead14665e3 PPC: simplify loading some values into altivec registers
mru
parents: 9421
diff changeset
31 #include "types_altivec.h"
11382
50415a8f1451 PPC: move prototypes to headers and make some functions static
mru
parents: 11369
diff changeset
32 #include "dsputil_altivec.h"
10079
71ead14665e3 PPC: simplify loading some values into altivec registers
mru
parents: 9421
diff changeset
33
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
34 // Swaps two variables (used for altivec registers)
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
35 #define SWAP(a,b) \
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
36 do { \
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
37 __typeof__(a) swap_temp=a; \
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
38 a=b; \
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
39 b=swap_temp; \
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
40 } while (0)
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
41
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
42 // transposes a matrix consisting of four vectors with four elements each
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
43 #define TRANSPOSE4(a,b,c,d) \
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
44 do { \
7333
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
45 __typeof__(a) _trans_ach = vec_mergeh(a, c); \
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
46 __typeof__(a) _trans_acl = vec_mergel(a, c); \
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
47 __typeof__(a) _trans_bdh = vec_mergeh(b, d); \
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
48 __typeof__(a) _trans_bdl = vec_mergel(b, d); \
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
49 \
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
50 a = vec_mergeh(_trans_ach, _trans_bdh); \
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
51 b = vec_mergel(_trans_ach, _trans_bdh); \
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
52 c = vec_mergeh(_trans_acl, _trans_bdl); \
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
53 d = vec_mergel(_trans_acl, _trans_bdl); \
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
54 } while (0)
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
55
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
56
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
57 // Loads a four-byte value (int or float) from the target address
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
58 // into every element in the target vector. Only works if the
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
59 // target address is four-byte aligned (which should be always).
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
60 #define LOAD4(vec, address) \
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
61 { \
7333
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
62 __typeof__(vec)* _load_addr = (__typeof__(vec)*)(address); \
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
63 vector unsigned char _perm_vec = vec_lvsl(0,(address)); \
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
64 vec = vec_ld(0, _load_addr); \
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
65 vec = vec_perm(vec, vec, _perm_vec); \
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
66 vec = vec_splat(vec, 0); \
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
67 }
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
68
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1015
diff changeset
69
7373
266d4949aa15 Remove AltiVec vector declaration compiler compatibility macros.
diego
parents: 7333
diff changeset
70 #define FOUROF(a) {a,a,a,a}
6367
f02434b252ce Refactor the FOUROF macro using the AVV macro.
diego
parents: 6366
diff changeset
71
11382
50415a8f1451 PPC: move prototypes to headers and make some functions static
mru
parents: 11369
diff changeset
72 static int dct_quantize_altivec(MpegEncContext* s,
7333
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
73 DCTELEM* data, int n,
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
74 int qscale, int* overflow)
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
75 {
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
76 int lastNonZero;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
77 vector float row0, row1, row2, row3, row4, row5, row6, row7;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
78 vector float alt0, alt1, alt2, alt3, alt4, alt5, alt6, alt7;
5746
55ed6dc5d476 Remove const vector macro indirection that is useless and obfuscating
diego
parents: 5579
diff changeset
79 const vector float zero = (const vector float)FOUROF(0.);
5963
80103098c797 spelling
vitor
parents: 5761
diff changeset
80 // used after quantize step
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1352
diff changeset
81 int oldBaseValue = 0;
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
82
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
83 // Load the data into the row/alt vectors
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
84 {
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
85 vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
86
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
87 data0 = vec_ld(0, data);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
88 data1 = vec_ld(16, data);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
89 data2 = vec_ld(32, data);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
90 data3 = vec_ld(48, data);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
91 data4 = vec_ld(64, data);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
92 data5 = vec_ld(80, data);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
93 data6 = vec_ld(96, data);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
94 data7 = vec_ld(112, data);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
95
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
96 // Transpose the data before we start
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
97 TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
98
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
99 // load the data into floating point vectors. We load
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
100 // the high half of each row into the main row vectors
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
101 // and the low half into the alt vectors.
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
102 row0 = vec_ctf(vec_unpackh(data0), 0);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
103 alt0 = vec_ctf(vec_unpackl(data0), 0);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
104 row1 = vec_ctf(vec_unpackh(data1), 0);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
105 alt1 = vec_ctf(vec_unpackl(data1), 0);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
106 row2 = vec_ctf(vec_unpackh(data2), 0);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
107 alt2 = vec_ctf(vec_unpackl(data2), 0);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
108 row3 = vec_ctf(vec_unpackh(data3), 0);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
109 alt3 = vec_ctf(vec_unpackl(data3), 0);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
110 row4 = vec_ctf(vec_unpackh(data4), 0);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
111 alt4 = vec_ctf(vec_unpackl(data4), 0);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
112 row5 = vec_ctf(vec_unpackh(data5), 0);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
113 alt5 = vec_ctf(vec_unpackl(data5), 0);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
114 row6 = vec_ctf(vec_unpackh(data6), 0);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
115 alt6 = vec_ctf(vec_unpackl(data6), 0);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
116 row7 = vec_ctf(vec_unpackh(data7), 0);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
117 alt7 = vec_ctf(vec_unpackl(data7), 0);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
118 }
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
119
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
120 // The following block could exist as a separate an altivec dct
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
121 // function. However, if we put it inline, the DCT data can remain
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
122 // in the vector local variables, as floats, which we'll use during the
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
123 // quantize step...
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
124 {
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1015
diff changeset
125 const vector float vec_0_298631336 = (vector float)FOUROF(0.298631336f);
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1015
diff changeset
126 const vector float vec_0_390180644 = (vector float)FOUROF(-0.390180644f);
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1015
diff changeset
127 const vector float vec_0_541196100 = (vector float)FOUROF(0.541196100f);
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1015
diff changeset
128 const vector float vec_0_765366865 = (vector float)FOUROF(0.765366865f);
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1015
diff changeset
129 const vector float vec_0_899976223 = (vector float)FOUROF(-0.899976223f);
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1015
diff changeset
130 const vector float vec_1_175875602 = (vector float)FOUROF(1.175875602f);
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1015
diff changeset
131 const vector float vec_1_501321110 = (vector float)FOUROF(1.501321110f);
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1015
diff changeset
132 const vector float vec_1_847759065 = (vector float)FOUROF(-1.847759065f);
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1015
diff changeset
133 const vector float vec_1_961570560 = (vector float)FOUROF(-1.961570560f);
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1015
diff changeset
134 const vector float vec_2_053119869 = (vector float)FOUROF(2.053119869f);
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1015
diff changeset
135 const vector float vec_2_562915447 = (vector float)FOUROF(-2.562915447f);
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1015
diff changeset
136 const vector float vec_3_072711026 = (vector float)FOUROF(3.072711026f);
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
137
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
138
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
139 int whichPass, whichHalf;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
140
7333
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
141 for(whichPass = 1; whichPass<=2; whichPass++) {
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
142 for(whichHalf = 1; whichHalf<=2; whichHalf++) {
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
143 vector float tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
144 vector float tmp10, tmp11, tmp12, tmp13;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
145 vector float z1, z2, z3, z4, z5;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
146
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
147 tmp0 = vec_add(row0, row7); // tmp0 = dataptr[0] + dataptr[7];
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
148 tmp7 = vec_sub(row0, row7); // tmp7 = dataptr[0] - dataptr[7];
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
149 tmp3 = vec_add(row3, row4); // tmp3 = dataptr[3] + dataptr[4];
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
150 tmp4 = vec_sub(row3, row4); // tmp4 = dataptr[3] - dataptr[4];
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
151 tmp1 = vec_add(row1, row6); // tmp1 = dataptr[1] + dataptr[6];
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
152 tmp6 = vec_sub(row1, row6); // tmp6 = dataptr[1] - dataptr[6];
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
153 tmp2 = vec_add(row2, row5); // tmp2 = dataptr[2] + dataptr[5];
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
154 tmp5 = vec_sub(row2, row5); // tmp5 = dataptr[2] - dataptr[5];
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
155
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
156 tmp10 = vec_add(tmp0, tmp3); // tmp10 = tmp0 + tmp3;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
157 tmp13 = vec_sub(tmp0, tmp3); // tmp13 = tmp0 - tmp3;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
158 tmp11 = vec_add(tmp1, tmp2); // tmp11 = tmp1 + tmp2;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
159 tmp12 = vec_sub(tmp1, tmp2); // tmp12 = tmp1 - tmp2;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
160
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
161
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
162 // dataptr[0] = (DCTELEM) ((tmp10 + tmp11) << PASS1_BITS);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
163 row0 = vec_add(tmp10, tmp11);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
164
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
165 // dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
166 row4 = vec_sub(tmp10, tmp11);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
167
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
168
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
169 // z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
170 z1 = vec_madd(vec_add(tmp12, tmp13), vec_0_541196100, (vector float)zero);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
171
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
172 // dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
173 // CONST_BITS-PASS1_BITS);
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
174 row2 = vec_madd(tmp13, vec_0_765366865, z1);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
175
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
176 // dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
177 // CONST_BITS-PASS1_BITS);
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
178 row6 = vec_madd(tmp12, vec_1_847759065, z1);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
179
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
180 z1 = vec_add(tmp4, tmp7); // z1 = tmp4 + tmp7;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
181 z2 = vec_add(tmp5, tmp6); // z2 = tmp5 + tmp6;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
182 z3 = vec_add(tmp4, tmp6); // z3 = tmp4 + tmp6;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
183 z4 = vec_add(tmp5, tmp7); // z4 = tmp5 + tmp7;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
184
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
185 // z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
186 z5 = vec_madd(vec_add(z3, z4), vec_1_175875602, (vector float)zero);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
187
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
188 // z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
189 z3 = vec_madd(z3, vec_1_961570560, z5);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
190
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
191 // z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
192 z4 = vec_madd(z4, vec_0_390180644, z5);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
193
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
194 // The following adds are rolled into the multiplies above
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
195 // z3 = vec_add(z3, z5); // z3 += z5;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
196 // z4 = vec_add(z4, z5); // z4 += z5;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
197
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
198 // z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
6486
3f96dc62c4b9 misc spelling fixes
diego
parents: 6371
diff changeset
199 // Wow! It's actually more efficient to roll this multiply
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
200 // into the adds below, even thought the multiply gets done twice!
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
201 // z2 = vec_madd(z2, vec_2_562915447, (vector float)zero);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
202
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
203 // z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
204 // Same with this one...
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
205 // z1 = vec_madd(z1, vec_0_899976223, (vector float)zero);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
206
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
207 // tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
208 // dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
209 row7 = vec_madd(tmp4, vec_0_298631336, vec_madd(z1, vec_0_899976223, z3));
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
210
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
211 // tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
212 // dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
213 row5 = vec_madd(tmp5, vec_2_053119869, vec_madd(z2, vec_2_562915447, z4));
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
214
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
215 // tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
216 // dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
217 row3 = vec_madd(tmp6, vec_3_072711026, vec_madd(z2, vec_2_562915447, z3));
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
218
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
219 // tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
220 // dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
221 row1 = vec_madd(z1, vec_0_899976223, vec_madd(tmp7, vec_1_501321110, z4));
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
222
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
223 // Swap the row values with the alts. If this is the first half,
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
224 // this sets up the low values to be acted on in the second half.
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
225 // If this is the second half, it puts the high values back in
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
226 // the row values where they are expected to be when we're done.
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
227 SWAP(row0, alt0);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
228 SWAP(row1, alt1);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
229 SWAP(row2, alt2);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
230 SWAP(row3, alt3);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
231 SWAP(row4, alt4);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
232 SWAP(row5, alt5);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
233 SWAP(row6, alt6);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
234 SWAP(row7, alt7);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
235 }
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
236
7333
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
237 if (whichPass == 1) {
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
238 // transpose the data for the second pass
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 1942
diff changeset
239
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
240 // First, block transpose the upper right with lower left.
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
241 SWAP(row4, alt0);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
242 SWAP(row5, alt1);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
243 SWAP(row6, alt2);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
244 SWAP(row7, alt3);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
245
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
246 // Now, transpose each block of four
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
247 TRANSPOSE4(row0, row1, row2, row3);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
248 TRANSPOSE4(row4, row5, row6, row7);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
249 TRANSPOSE4(alt0, alt1, alt2, alt3);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
250 TRANSPOSE4(alt4, alt5, alt6, alt7);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
251 }
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
252 }
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
253 }
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
254
5963
80103098c797 spelling
vitor
parents: 5761
diff changeset
255 // perform the quantize step, using the floating point data
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
256 // still in the row/alt registers
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
257 {
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
258 const int* biasAddr;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
259 const vector signed int* qmat;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
260 vector float bias, negBias;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
261
7333
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
262 if (s->mb_intra) {
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
263 vector signed int baseVector;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
264
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
265 // We must cache element 0 in the intra case
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
266 // (it needs special handling).
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
267 baseVector = vec_cts(vec_splat(row0, 0), 0);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
268 vec_ste(baseVector, 0, &oldBaseValue);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
269
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
270 qmat = (vector signed int*)s->q_intra_matrix[qscale];
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
271 biasAddr = &(s->intra_quant_bias);
7333
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
272 } else {
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
273 qmat = (vector signed int*)s->q_inter_matrix[qscale];
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
274 biasAddr = &(s->inter_quant_bias);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
275 }
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
276
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
277 // Load the bias vector (We add 0.5 to the bias so that we're
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
278 // rounding when we convert to int, instead of flooring.)
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
279 {
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
280 vector signed int biasInt;
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1015
diff changeset
281 const vector float negOneFloat = (vector float)FOUROF(-1.0f);
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
282 LOAD4(biasInt, biasAddr);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
283 bias = vec_ctf(biasInt, QUANT_BIAS_SHIFT);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
284 negBias = vec_madd(bias, negOneFloat, zero);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
285 }
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
286
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
287 {
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
288 vector float q0, q1, q2, q3, q4, q5, q6, q7;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
289
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
290 q0 = vec_ctf(qmat[0], QMAT_SHIFT);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
291 q1 = vec_ctf(qmat[2], QMAT_SHIFT);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
292 q2 = vec_ctf(qmat[4], QMAT_SHIFT);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
293 q3 = vec_ctf(qmat[6], QMAT_SHIFT);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
294 q4 = vec_ctf(qmat[8], QMAT_SHIFT);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
295 q5 = vec_ctf(qmat[10], QMAT_SHIFT);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
296 q6 = vec_ctf(qmat[12], QMAT_SHIFT);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
297 q7 = vec_ctf(qmat[14], QMAT_SHIFT);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
298
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
299 row0 = vec_sel(vec_madd(row0, q0, negBias), vec_madd(row0, q0, bias),
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
300 vec_cmpgt(row0, zero));
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
301 row1 = vec_sel(vec_madd(row1, q1, negBias), vec_madd(row1, q1, bias),
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
302 vec_cmpgt(row1, zero));
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
303 row2 = vec_sel(vec_madd(row2, q2, negBias), vec_madd(row2, q2, bias),
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
304 vec_cmpgt(row2, zero));
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
305 row3 = vec_sel(vec_madd(row3, q3, negBias), vec_madd(row3, q3, bias),
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
306 vec_cmpgt(row3, zero));
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
307 row4 = vec_sel(vec_madd(row4, q4, negBias), vec_madd(row4, q4, bias),
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
308 vec_cmpgt(row4, zero));
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
309 row5 = vec_sel(vec_madd(row5, q5, negBias), vec_madd(row5, q5, bias),
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
310 vec_cmpgt(row5, zero));
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
311 row6 = vec_sel(vec_madd(row6, q6, negBias), vec_madd(row6, q6, bias),
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
312 vec_cmpgt(row6, zero));
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
313 row7 = vec_sel(vec_madd(row7, q7, negBias), vec_madd(row7, q7, bias),
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
314 vec_cmpgt(row7, zero));
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
315
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
316 q0 = vec_ctf(qmat[1], QMAT_SHIFT);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
317 q1 = vec_ctf(qmat[3], QMAT_SHIFT);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
318 q2 = vec_ctf(qmat[5], QMAT_SHIFT);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
319 q3 = vec_ctf(qmat[7], QMAT_SHIFT);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
320 q4 = vec_ctf(qmat[9], QMAT_SHIFT);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
321 q5 = vec_ctf(qmat[11], QMAT_SHIFT);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
322 q6 = vec_ctf(qmat[13], QMAT_SHIFT);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
323 q7 = vec_ctf(qmat[15], QMAT_SHIFT);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
324
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
325 alt0 = vec_sel(vec_madd(alt0, q0, negBias), vec_madd(alt0, q0, bias),
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
326 vec_cmpgt(alt0, zero));
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
327 alt1 = vec_sel(vec_madd(alt1, q1, negBias), vec_madd(alt1, q1, bias),
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
328 vec_cmpgt(alt1, zero));
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
329 alt2 = vec_sel(vec_madd(alt2, q2, negBias), vec_madd(alt2, q2, bias),
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
330 vec_cmpgt(alt2, zero));
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
331 alt3 = vec_sel(vec_madd(alt3, q3, negBias), vec_madd(alt3, q3, bias),
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
332 vec_cmpgt(alt3, zero));
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
333 alt4 = vec_sel(vec_madd(alt4, q4, negBias), vec_madd(alt4, q4, bias),
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
334 vec_cmpgt(alt4, zero));
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
335 alt5 = vec_sel(vec_madd(alt5, q5, negBias), vec_madd(alt5, q5, bias),
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
336 vec_cmpgt(alt5, zero));
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
337 alt6 = vec_sel(vec_madd(alt6, q6, negBias), vec_madd(alt6, q6, bias),
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
338 vec_cmpgt(alt6, zero));
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
339 alt7 = vec_sel(vec_madd(alt7, q7, negBias), vec_madd(alt7, q7, bias),
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
340 vec_cmpgt(alt7, zero));
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
341 }
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
342
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 1942
diff changeset
343
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
344 }
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
345
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
346 // Store the data back into the original block
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
347 {
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
348 vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
349
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
350 data0 = vec_pack(vec_cts(row0, 0), vec_cts(alt0, 0));
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
351 data1 = vec_pack(vec_cts(row1, 0), vec_cts(alt1, 0));
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
352 data2 = vec_pack(vec_cts(row2, 0), vec_cts(alt2, 0));
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
353 data3 = vec_pack(vec_cts(row3, 0), vec_cts(alt3, 0));
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
354 data4 = vec_pack(vec_cts(row4, 0), vec_cts(alt4, 0));
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
355 data5 = vec_pack(vec_cts(row5, 0), vec_cts(alt5, 0));
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
356 data6 = vec_pack(vec_cts(row6, 0), vec_cts(alt6, 0));
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
357 data7 = vec_pack(vec_cts(row7, 0), vec_cts(alt7, 0));
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
358
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
359 {
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
360 // Clamp for overflow
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
361 vector signed int max_q_int, min_q_int;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
362 vector signed short max_q, min_q;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
363
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
364 LOAD4(max_q_int, &(s->max_qcoeff));
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
365 LOAD4(min_q_int, &(s->min_qcoeff));
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
366
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
367 max_q = vec_pack(max_q_int, max_q_int);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
368 min_q = vec_pack(min_q_int, min_q_int);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
369
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
370 data0 = vec_max(vec_min(data0, max_q), min_q);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
371 data1 = vec_max(vec_min(data1, max_q), min_q);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
372 data2 = vec_max(vec_min(data2, max_q), min_q);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
373 data4 = vec_max(vec_min(data4, max_q), min_q);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
374 data5 = vec_max(vec_min(data5, max_q), min_q);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
375 data6 = vec_max(vec_min(data6, max_q), min_q);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
376 data7 = vec_max(vec_min(data7, max_q), min_q);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
377 }
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
378
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1352
diff changeset
379 {
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
380 vector bool char zero_01, zero_23, zero_45, zero_67;
6905
d163c4f3f4ab consistency cosmetics: indices --> indexes in variable names
diego
parents: 6763
diff changeset
381 vector signed char scanIndexes_01, scanIndexes_23, scanIndexes_45, scanIndexes_67;
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
382 vector signed char negOne = vec_splat_s8(-1);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
383 vector signed char* scanPtr =
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
384 (vector signed char*)(s->intra_scantable.inverse);
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1352
diff changeset
385 signed char lastNonZeroChar;
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
386
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
387 // Determine the largest non-zero index.
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1352
diff changeset
388 zero_01 = vec_pack(vec_cmpeq(data0, (vector signed short)zero),
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1352
diff changeset
389 vec_cmpeq(data1, (vector signed short)zero));
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1352
diff changeset
390 zero_23 = vec_pack(vec_cmpeq(data2, (vector signed short)zero),
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1352
diff changeset
391 vec_cmpeq(data3, (vector signed short)zero));
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1352
diff changeset
392 zero_45 = vec_pack(vec_cmpeq(data4, (vector signed short)zero),
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1352
diff changeset
393 vec_cmpeq(data5, (vector signed short)zero));
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1352
diff changeset
394 zero_67 = vec_pack(vec_cmpeq(data6, (vector signed short)zero),
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1352
diff changeset
395 vec_cmpeq(data7, (vector signed short)zero));
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
396
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
397 // 64 biggest values
6905
d163c4f3f4ab consistency cosmetics: indices --> indexes in variable names
diego
parents: 6763
diff changeset
398 scanIndexes_01 = vec_sel(scanPtr[0], negOne, zero_01);
d163c4f3f4ab consistency cosmetics: indices --> indexes in variable names
diego
parents: 6763
diff changeset
399 scanIndexes_23 = vec_sel(scanPtr[1], negOne, zero_23);
d163c4f3f4ab consistency cosmetics: indices --> indexes in variable names
diego
parents: 6763
diff changeset
400 scanIndexes_45 = vec_sel(scanPtr[2], negOne, zero_45);
d163c4f3f4ab consistency cosmetics: indices --> indexes in variable names
diego
parents: 6763
diff changeset
401 scanIndexes_67 = vec_sel(scanPtr[3], negOne, zero_67);
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
402
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
403 // 32 largest values
6905
d163c4f3f4ab consistency cosmetics: indices --> indexes in variable names
diego
parents: 6763
diff changeset
404 scanIndexes_01 = vec_max(scanIndexes_01, scanIndexes_23);
d163c4f3f4ab consistency cosmetics: indices --> indexes in variable names
diego
parents: 6763
diff changeset
405 scanIndexes_45 = vec_max(scanIndexes_45, scanIndexes_67);
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
406
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
407 // 16 largest values
6905
d163c4f3f4ab consistency cosmetics: indices --> indexes in variable names
diego
parents: 6763
diff changeset
408 scanIndexes_01 = vec_max(scanIndexes_01, scanIndexes_45);
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
409
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
410 // 8 largest values
6905
d163c4f3f4ab consistency cosmetics: indices --> indexes in variable names
diego
parents: 6763
diff changeset
411 scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne),
d163c4f3f4ab consistency cosmetics: indices --> indexes in variable names
diego
parents: 6763
diff changeset
412 vec_mergel(scanIndexes_01, negOne));
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
413
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
414 // 4 largest values
6905
d163c4f3f4ab consistency cosmetics: indices --> indexes in variable names
diego
parents: 6763
diff changeset
415 scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne),
d163c4f3f4ab consistency cosmetics: indices --> indexes in variable names
diego
parents: 6763
diff changeset
416 vec_mergel(scanIndexes_01, negOne));
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
417
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
418 // 2 largest values
6905
d163c4f3f4ab consistency cosmetics: indices --> indexes in variable names
diego
parents: 6763
diff changeset
419 scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne),
d163c4f3f4ab consistency cosmetics: indices --> indexes in variable names
diego
parents: 6763
diff changeset
420 vec_mergel(scanIndexes_01, negOne));
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
421
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
422 // largest value
6905
d163c4f3f4ab consistency cosmetics: indices --> indexes in variable names
diego
parents: 6763
diff changeset
423 scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne),
d163c4f3f4ab consistency cosmetics: indices --> indexes in variable names
diego
parents: 6763
diff changeset
424 vec_mergel(scanIndexes_01, negOne));
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
425
6905
d163c4f3f4ab consistency cosmetics: indices --> indexes in variable names
diego
parents: 6763
diff changeset
426 scanIndexes_01 = vec_splat(scanIndexes_01, 0);
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
427
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
428
6905
d163c4f3f4ab consistency cosmetics: indices --> indexes in variable names
diego
parents: 6763
diff changeset
429 vec_ste(scanIndexes_01, 0, &lastNonZeroChar);
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
430
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
431 lastNonZero = lastNonZeroChar;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 1942
diff changeset
432
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
433 // While the data is still in vectors we check for the transpose IDCT permute
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
434 // and handle it using the vector unit if we can. This is the permute used
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
435 // by the altivec idct, so it is common when using the altivec dct.
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
436
7333
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
437 if ((lastNonZero > 0) && (s->dsp.idct_permutation_type == FF_TRANSPOSE_IDCT_PERM)) {
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
438 TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
439 }
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
440
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
441 vec_st(data0, 0, data);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
442 vec_st(data1, 16, data);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
443 vec_st(data2, 32, data);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
444 vec_st(data3, 48, data);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
445 vec_st(data4, 64, data);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
446 vec_st(data5, 80, data);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
447 vec_st(data6, 96, data);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
448 vec_st(data7, 112, data);
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1352
diff changeset
449 }
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
450 }
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
451
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
452 // special handling of block[0]
7333
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
453 if (s->mb_intra) {
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
454 if (!s->h263_aic) {
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
455 if (n < 4)
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
456 oldBaseValue /= s->y_dc_scale;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
457 else
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
458 oldBaseValue /= s->c_dc_scale;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
459 }
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
460
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
461 // Divide by 8, rounding the result
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
462 data[0] = (oldBaseValue + 4) >> 3;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
463 }
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
464
5963
80103098c797 spelling
vitor
parents: 5761
diff changeset
465 // We handled the transpose permutation above and we don't
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
466 // need to permute the "no" permutation case.
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
467 if ((lastNonZero > 0) &&
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1033
diff changeset
468 (s->dsp.idct_permutation_type != FF_TRANSPOSE_IDCT_PERM) &&
7333
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
469 (s->dsp.idct_permutation_type != FF_NO_IDCT_PERM)) {
1107
28f0e53706a7 altivec compilation fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents: 1092
diff changeset
470 ff_block_permute(data, s->dsp.idct_permutation,
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
471 s->intra_scantable.scantable, lastNonZero);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
472 }
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
473
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
474 return lastNonZero;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
475 }
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
476
7333
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
477 /* AltiVec version of dct_unquantize_h263
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
478 this code assumes `block' is 16 bytes-aligned */
11382
50415a8f1451 PPC: move prototypes to headers and make some functions static
mru
parents: 11369
diff changeset
479 static void dct_unquantize_h263_altivec(MpegEncContext *s,
1001
95cbffdc98a9 dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 995
diff changeset
480 DCTELEM *block, int n, int qscale)
95cbffdc98a9 dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 995
diff changeset
481 {
95cbffdc98a9 dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 995
diff changeset
482 int i, level, qmul, qadd;
95cbffdc98a9 dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 995
diff changeset
483 int nCoeffs;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 1942
diff changeset
484
1001
95cbffdc98a9 dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 995
diff changeset
485 assert(s->block_last_index[n]>=0);
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1001
diff changeset
486
1001
95cbffdc98a9 dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 995
diff changeset
487 qadd = (qscale - 1) | 1;
95cbffdc98a9 dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 995
diff changeset
488 qmul = qscale << 1;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 1942
diff changeset
489
1001
95cbffdc98a9 dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 995
diff changeset
490 if (s->mb_intra) {
95cbffdc98a9 dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 995
diff changeset
491 if (!s->h263_aic) {
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 1942
diff changeset
492 if (n < 4)
1001
95cbffdc98a9 dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 995
diff changeset
493 block[0] = block[0] * s->y_dc_scale;
95cbffdc98a9 dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 995
diff changeset
494 else
95cbffdc98a9 dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 995
diff changeset
495 block[0] = block[0] * s->c_dc_scale;
95cbffdc98a9 dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 995
diff changeset
496 }else
95cbffdc98a9 dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 995
diff changeset
497 qadd = 0;
95cbffdc98a9 dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 995
diff changeset
498 i = 1;
4795
522e52c630bd typos/grammar
diego
parents: 3983
diff changeset
499 nCoeffs= 63; //does not always use zigzag table
1001
95cbffdc98a9 dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 995
diff changeset
500 } else {
95cbffdc98a9 dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 995
diff changeset
501 i = 0;
95cbffdc98a9 dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 995
diff changeset
502 nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
95cbffdc98a9 dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 995
diff changeset
503 }
95cbffdc98a9 dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 995
diff changeset
504
95cbffdc98a9 dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 995
diff changeset
505 {
7333
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
506 register const vector signed short vczero = (const vector signed short)vec_splat_s16(0);
11369
98970e51365a Remove DECLARE_ALIGNED_{8,16} macros
mru
parents: 10079
diff changeset
507 DECLARE_ALIGNED(16, short, qmul8) = qmul;
98970e51365a Remove DECLARE_ALIGNED_{8,16} macros
mru
parents: 10079
diff changeset
508 DECLARE_ALIGNED(16, short, qadd8) = qadd;
7333
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
509 register vector signed short blockv, qmulv, qaddv, nqaddv, temp1;
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
510 register vector bool short blockv_null, blockv_neg;
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
511 register short backup_0 = block[0];
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
512 register int j = 0;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 1942
diff changeset
513
10079
71ead14665e3 PPC: simplify loading some values into altivec registers
mru
parents: 9421
diff changeset
514 qmulv = vec_splat((vec_s16)vec_lde(0, &qmul8), 0);
71ead14665e3 PPC: simplify loading some values into altivec registers
mru
parents: 9421
diff changeset
515 qaddv = vec_splat((vec_s16)vec_lde(0, &qadd8), 0);
71ead14665e3 PPC: simplify loading some values into altivec registers
mru
parents: 9421
diff changeset
516 nqaddv = vec_sub(vczero, qaddv);
1001
95cbffdc98a9 dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 995
diff changeset
517
7333
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
518 #if 0 // block *is* 16 bytes-aligned, it seems.
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
519 // first make sure block[j] is 16 bytes-aligned
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
520 for(j = 0; (j <= nCoeffs) && ((((unsigned long)block) + (j << 1)) & 0x0000000F) ; j++) {
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
521 level = block[j];
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
522 if (level) {
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
523 if (level < 0) {
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
524 level = level * qmul - qadd;
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
525 } else {
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
526 level = level * qmul + qadd;
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
527 }
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
528 block[j] = level;
1001
95cbffdc98a9 dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 995
diff changeset
529 }
95cbffdc98a9 dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 995
diff changeset
530 }
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 1001
diff changeset
531 #endif
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 1942
diff changeset
532
7333
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
533 // vectorize all the 16 bytes-aligned blocks
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
534 // of 8 elements
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
535 for(; (j + 7) <= nCoeffs ; j+=8) {
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
536 blockv = vec_ld(j << 1, block);
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
537 blockv_neg = vec_cmplt(blockv, vczero);
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
538 blockv_null = vec_cmpeq(blockv, vczero);
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
539 // choose between +qadd or -qadd as the third operand
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
540 temp1 = vec_sel(qaddv, nqaddv, blockv_neg);
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
541 // multiply & add (block{i,i+7} * qmul [+-] qadd)
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
542 temp1 = vec_mladd(blockv, qmulv, temp1);
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
543 // put 0 where block[{i,i+7} used to have 0
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
544 blockv = vec_sel(temp1, blockv, blockv_null);
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
545 vec_st(blockv, j << 1, block);
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
546 }
1001
95cbffdc98a9 dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 995
diff changeset
547
7333
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
548 // if nCoeffs isn't a multiple of 8, finish the job
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
549 // using good old scalar units.
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
550 // (we could do it using a truncated vector,
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
551 // but I'm not sure it's worth the hassle)
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
552 for(; j <= nCoeffs ; j++) {
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
553 level = block[j];
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
554 if (level) {
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
555 if (level < 0) {
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
556 level = level * qmul - qadd;
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
557 } else {
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
558 level = level * qmul + qadd;
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
559 }
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
560 block[j] = level;
1001
95cbffdc98a9 dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 995
diff changeset
561 }
95cbffdc98a9 dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 995
diff changeset
562 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 1942
diff changeset
563
7333
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
564 if (i == 1) {
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
565 // cheat. this avoid special-casing the first iteration
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
566 block[0] = backup_0;
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
567 }
1001
95cbffdc98a9 dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 995
diff changeset
568 }
95cbffdc98a9 dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 995
diff changeset
569 }
5751
3260fa3c89eb Merge mpegvideo AltiVec code into mpegvideo_altivec.c where it belongs.
diego
parents: 5750
diff changeset
570
3260fa3c89eb Merge mpegvideo AltiVec code into mpegvideo_altivec.c where it belongs.
diego
parents: 5750
diff changeset
571
5761
72b59f7613af Rename MPV_common_init_ppc to MPV_common_init_altivec, the function is
diego
parents: 5751
diff changeset
572 void MPV_common_init_altivec(MpegEncContext *s)
5751
3260fa3c89eb Merge mpegvideo AltiVec code into mpegvideo_altivec.c where it belongs.
diego
parents: 5750
diff changeset
573 {
12475
9fef0a8ddd63 Move mm_support() from libavcodec to libavutil, make it a public
stefano
parents: 12473
diff changeset
574 if (!(av_get_cpu_flags() & AV_CPU_FLAG_ALTIVEC)) return;
6365
80c7e9c015c6 fix crash on non-AltiVec powered machines: MPV_common_init_altivec doesn't check mm_flags
gpoirier
parents: 5963
diff changeset
575
7333
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
576 if (s->avctx->lowres==0) {
5751
3260fa3c89eb Merge mpegvideo AltiVec code into mpegvideo_altivec.c where it belongs.
diego
parents: 5750
diff changeset
577 if ((s->avctx->idct_algo == FF_IDCT_AUTO) ||
7333
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
578 (s->avctx->idct_algo == FF_IDCT_ALTIVEC)) {
5751
3260fa3c89eb Merge mpegvideo AltiVec code into mpegvideo_altivec.c where it belongs.
diego
parents: 5750
diff changeset
579 s->dsp.idct_put = idct_put_altivec;
3260fa3c89eb Merge mpegvideo AltiVec code into mpegvideo_altivec.c where it belongs.
diego
parents: 5750
diff changeset
580 s->dsp.idct_add = idct_add_altivec;
3260fa3c89eb Merge mpegvideo AltiVec code into mpegvideo_altivec.c where it belongs.
diego
parents: 5750
diff changeset
581 s->dsp.idct_permutation_type = FF_TRANSPOSE_IDCT_PERM;
3260fa3c89eb Merge mpegvideo AltiVec code into mpegvideo_altivec.c where it belongs.
diego
parents: 5750
diff changeset
582 }
3260fa3c89eb Merge mpegvideo AltiVec code into mpegvideo_altivec.c where it belongs.
diego
parents: 5750
diff changeset
583 }
3260fa3c89eb Merge mpegvideo AltiVec code into mpegvideo_altivec.c where it belongs.
diego
parents: 5750
diff changeset
584
3260fa3c89eb Merge mpegvideo AltiVec code into mpegvideo_altivec.c where it belongs.
diego
parents: 5750
diff changeset
585 // Test to make sure that the dct required alignments are met.
3260fa3c89eb Merge mpegvideo AltiVec code into mpegvideo_altivec.c where it belongs.
diego
parents: 5750
diff changeset
586 if ((((long)(s->q_intra_matrix) & 0x0f) != 0) ||
7333
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
587 (((long)(s->q_inter_matrix) & 0x0f) != 0)) {
5751
3260fa3c89eb Merge mpegvideo AltiVec code into mpegvideo_altivec.c where it belongs.
diego
parents: 5750
diff changeset
588 av_log(s->avctx, AV_LOG_INFO, "Internal Error: q-matrix blocks must be 16-byte aligned "
3260fa3c89eb Merge mpegvideo AltiVec code into mpegvideo_altivec.c where it belongs.
diego
parents: 5750
diff changeset
589 "to use AltiVec DCT. Reverting to non-AltiVec version.\n");
3260fa3c89eb Merge mpegvideo AltiVec code into mpegvideo_altivec.c where it belongs.
diego
parents: 5750
diff changeset
590 return;
3260fa3c89eb Merge mpegvideo AltiVec code into mpegvideo_altivec.c where it belongs.
diego
parents: 5750
diff changeset
591 }
3260fa3c89eb Merge mpegvideo AltiVec code into mpegvideo_altivec.c where it belongs.
diego
parents: 5750
diff changeset
592
7333
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
593 if (((long)(s->intra_scantable.inverse) & 0x0f) != 0) {
5751
3260fa3c89eb Merge mpegvideo AltiVec code into mpegvideo_altivec.c where it belongs.
diego
parents: 5750
diff changeset
594 av_log(s->avctx, AV_LOG_INFO, "Internal Error: scan table blocks must be 16-byte aligned "
3260fa3c89eb Merge mpegvideo AltiVec code into mpegvideo_altivec.c where it belongs.
diego
parents: 5750
diff changeset
595 "to use AltiVec DCT. Reverting to non-AltiVec version.\n");
3260fa3c89eb Merge mpegvideo AltiVec code into mpegvideo_altivec.c where it belongs.
diego
parents: 5750
diff changeset
596 return;
3260fa3c89eb Merge mpegvideo AltiVec code into mpegvideo_altivec.c where it belongs.
diego
parents: 5750
diff changeset
597 }
3260fa3c89eb Merge mpegvideo AltiVec code into mpegvideo_altivec.c where it belongs.
diego
parents: 5750
diff changeset
598
3260fa3c89eb Merge mpegvideo AltiVec code into mpegvideo_altivec.c where it belongs.
diego
parents: 5750
diff changeset
599
3260fa3c89eb Merge mpegvideo AltiVec code into mpegvideo_altivec.c where it belongs.
diego
parents: 5750
diff changeset
600 if ((s->avctx->dct_algo == FF_DCT_AUTO) ||
7333
a8a79f5385f6 cosmetics: Reformat PPC code in libavcodec according to style guidelines.
diego
parents: 6905
diff changeset
601 (s->avctx->dct_algo == FF_DCT_ALTIVEC)) {
5751
3260fa3c89eb Merge mpegvideo AltiVec code into mpegvideo_altivec.c where it belongs.
diego
parents: 5750
diff changeset
602 #if 0 /* seems to cause trouble under some circumstances */
3260fa3c89eb Merge mpegvideo AltiVec code into mpegvideo_altivec.c where it belongs.
diego
parents: 5750
diff changeset
603 s->dct_quantize = dct_quantize_altivec;
3260fa3c89eb Merge mpegvideo AltiVec code into mpegvideo_altivec.c where it belongs.
diego
parents: 5750
diff changeset
604 #endif
3260fa3c89eb Merge mpegvideo AltiVec code into mpegvideo_altivec.c where it belongs.
diego
parents: 5750
diff changeset
605 s->dct_unquantize_h263_intra = dct_unquantize_h263_altivec;
3260fa3c89eb Merge mpegvideo AltiVec code into mpegvideo_altivec.c where it belongs.
diego
parents: 5750
diff changeset
606 s->dct_unquantize_h263_inter = dct_unquantize_h263_altivec;
3260fa3c89eb Merge mpegvideo AltiVec code into mpegvideo_altivec.c where it belongs.
diego
parents: 5750
diff changeset
607 }
3260fa3c89eb Merge mpegvideo AltiVec code into mpegvideo_altivec.c where it belongs.
diego
parents: 5750
diff changeset
608 }