annotate ppc/fdct_altivec.c @ 4631:cd3e32d16136 libavcodec

add a -ab with default of 64k change -b to only apply to video this returns -ab/-b to the old defautls and old behavior regarding ordering of them
author michael
date Wed, 07 Mar 2007 09:51:45 +0000
parents b28edd190fc0
children d5ba514e3f4a
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
1578
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
1 /* ffmpeg/libavcodec/ppc/fdct_altivec.c, this file is part of the
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
2 * AltiVec optimized library for the FFMPEG Multimedia System
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
3 * Copyright (C) 2003 James Klicman <james@klicman.org>
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
4 *
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3036
diff changeset
5 * This file is part of FFmpeg.
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3036
diff changeset
6 *
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3036
diff changeset
7 * FFmpeg is free software; you can redistribute it and/or
1578
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
8 * modify it under the terms of the GNU Lesser General Public
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
9 * License as published by the Free Software Foundation; either
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
10 * version 2.1 of the License, or (at your option) any later version.
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
11 *
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3036
diff changeset
12 * FFmpeg is distributed in the hope that it will be useful,
1578
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
15 * Lesser General Public License for more details.
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
16 *
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
17 * 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
18 * License along with FFmpeg; if not, write to the Free Software
3036
0b546eab515d Update licensing information: The FSF changed postal address.
diego
parents: 2819
diff changeset
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
1578
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
20 */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
21
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
22
2819
alex
parents: 2612
diff changeset
23 #include "common.h"
1578
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
24 #include "../dsputil.h"
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
25 #include "dsputil_altivec.h"
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
26 #include "gcc_fixes.h"
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
27
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
28
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
29 #define vs16(v) ((vector signed short)(v))
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
30 #define vs32(v) ((vector signed int)(v))
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
31 #define vu8(v) ((vector unsigned char)(v))
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
32 #define vu16(v) ((vector unsigned short)(v))
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
33 #define vu32(v) ((vector unsigned int)(v))
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
34
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
35
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
36 #define C1 0.98078525066375732421875000 /* cos(1*PI/16) */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
37 #define C2 0.92387950420379638671875000 /* cos(2*PI/16) */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
38 #define C3 0.83146959543228149414062500 /* cos(3*PI/16) */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
39 #define C4 0.70710676908493041992187500 /* cos(4*PI/16) */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
40 #define C5 0.55557024478912353515625000 /* cos(5*PI/16) */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
41 #define C6 0.38268342614173889160156250 /* cos(6*PI/16) */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
42 #define C7 0.19509032368659973144531250 /* cos(7*PI/16) */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
43 #define SQRT_2 1.41421353816986083984375000 /* sqrt(2) */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
44
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
45
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
46 #define W0 -(2 * C2)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
47 #define W1 (2 * C6)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
48 #define W2 (SQRT_2 * C6)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
49 #define W3 (SQRT_2 * C3)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
50 #define W4 (SQRT_2 * (-C1 + C3 + C5 - C7))
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
51 #define W5 (SQRT_2 * ( C1 + C3 - C5 + C7))
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
52 #define W6 (SQRT_2 * ( C1 + C3 + C5 - C7))
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
53 #define W7 (SQRT_2 * ( C1 + C3 - C5 - C7))
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
54 #define W8 (SQRT_2 * ( C7 - C3))
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
55 #define W9 (SQRT_2 * (-C1 - C3))
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
56 #define WA (SQRT_2 * (-C3 - C5))
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
57 #define WB (SQRT_2 * ( C5 - C3))
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
58
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
59
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
60 static vector float fdctconsts[3] = {
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
61 (vector float)AVV( W0, W1, W2, W3 ),
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
62 (vector float)AVV( W4, W5, W6, W7 ),
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
63 (vector float)AVV( W8, W9, WA, WB )
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
64 };
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
65
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
66 #define LD_W0 vec_splat(cnsts0, 0)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
67 #define LD_W1 vec_splat(cnsts0, 1)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
68 #define LD_W2 vec_splat(cnsts0, 2)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
69 #define LD_W3 vec_splat(cnsts0, 3)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
70 #define LD_W4 vec_splat(cnsts1, 0)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
71 #define LD_W5 vec_splat(cnsts1, 1)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
72 #define LD_W6 vec_splat(cnsts1, 2)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
73 #define LD_W7 vec_splat(cnsts1, 3)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
74 #define LD_W8 vec_splat(cnsts2, 0)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
75 #define LD_W9 vec_splat(cnsts2, 1)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
76 #define LD_WA vec_splat(cnsts2, 2)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
77 #define LD_WB vec_splat(cnsts2, 3)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
78
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
79
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
80 #define FDCTROW(b0,b1,b2,b3,b4,b5,b6,b7) /* {{{ */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
81 x0 = vec_add(b0, b7); /* x0 = b0 + b7; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
82 x7 = vec_sub(b0, b7); /* x7 = b0 - b7; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
83 x1 = vec_add(b1, b6); /* x1 = b1 + b6; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
84 x6 = vec_sub(b1, b6); /* x6 = b1 - b6; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
85 x2 = vec_add(b2, b5); /* x2 = b2 + b5; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
86 x5 = vec_sub(b2, b5); /* x5 = b2 - b5; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
87 x3 = vec_add(b3, b4); /* x3 = b3 + b4; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
88 x4 = vec_sub(b3, b4); /* x4 = b3 - b4; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
89 \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
90 b7 = vec_add(x0, x3); /* b7 = x0 + x3; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
91 b1 = vec_add(x1, x2); /* b1 = x1 + x2; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
92 b0 = vec_add(b7, b1); /* b0 = b7 + b1; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
93 b4 = vec_sub(b7, b1); /* b4 = b7 - b1; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
94 \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
95 b2 = vec_sub(x0, x3); /* b2 = x0 - x3; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
96 b6 = vec_sub(x1, x2); /* b6 = x1 - x2; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
97 b5 = vec_add(b6, b2); /* b5 = b6 + b2; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
98 cnst = LD_W2; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
99 b5 = vec_madd(cnst, b5, mzero); /* b5 = b5 * W2; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
100 cnst = LD_W1; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
101 b2 = vec_madd(cnst, b2, b5); /* b2 = b5 + b2 * W1; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
102 cnst = LD_W0; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
103 b6 = vec_madd(cnst, b6, b5); /* b6 = b5 + b6 * W0; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
104 \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
105 x0 = vec_add(x4, x7); /* x0 = x4 + x7; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
106 x1 = vec_add(x5, x6); /* x1 = x5 + x6; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
107 x2 = vec_add(x4, x6); /* x2 = x4 + x6; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
108 x3 = vec_add(x5, x7); /* x3 = x5 + x7; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
109 x8 = vec_add(x2, x3); /* x8 = x2 + x3; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
110 cnst = LD_W3; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
111 x8 = vec_madd(cnst, x8, mzero); /* x8 = x8 * W3; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
112 \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
113 cnst = LD_W8; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
114 x0 = vec_madd(cnst, x0, mzero); /* x0 *= W8; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
115 cnst = LD_W9; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
116 x1 = vec_madd(cnst, x1, mzero); /* x1 *= W9; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
117 cnst = LD_WA; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
118 x2 = vec_madd(cnst, x2, x8); /* x2 = x2 * WA + x8; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
119 cnst = LD_WB; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
120 x3 = vec_madd(cnst, x3, x8); /* x3 = x3 * WB + x8; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
121 \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
122 cnst = LD_W4; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
123 b7 = vec_madd(cnst, x4, x0); /* b7 = x4 * W4 + x0; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
124 cnst = LD_W5; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
125 b5 = vec_madd(cnst, x5, x1); /* b5 = x5 * W5 + x1; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
126 cnst = LD_W6; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
127 b3 = vec_madd(cnst, x6, x1); /* b3 = x6 * W6 + x1; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
128 cnst = LD_W7; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
129 b1 = vec_madd(cnst, x7, x0); /* b1 = x7 * W7 + x0; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
130 \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
131 b7 = vec_add(b7, x2); /* b7 = b7 + x2; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
132 b5 = vec_add(b5, x3); /* b5 = b5 + x3; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
133 b3 = vec_add(b3, x2); /* b3 = b3 + x2; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
134 b1 = vec_add(b1, x3); /* b1 = b1 + x3; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
135 /* }}} */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
136
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
137 #define FDCTCOL(b0,b1,b2,b3,b4,b5,b6,b7) /* {{{ */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
138 x0 = vec_add(b0, b7); /* x0 = b0 + b7; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
139 x7 = vec_sub(b0, b7); /* x7 = b0 - b7; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
140 x1 = vec_add(b1, b6); /* x1 = b1 + b6; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
141 x6 = vec_sub(b1, b6); /* x6 = b1 - b6; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
142 x2 = vec_add(b2, b5); /* x2 = b2 + b5; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
143 x5 = vec_sub(b2, b5); /* x5 = b2 - b5; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
144 x3 = vec_add(b3, b4); /* x3 = b3 + b4; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
145 x4 = vec_sub(b3, b4); /* x4 = b3 - b4; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
146 \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
147 b7 = vec_add(x0, x3); /* b7 = x0 + x3; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
148 b1 = vec_add(x1, x2); /* b1 = x1 + x2; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
149 b0 = vec_add(b7, b1); /* b0 = b7 + b1; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
150 b4 = vec_sub(b7, b1); /* b4 = b7 - b1; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
151 \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
152 b2 = vec_sub(x0, x3); /* b2 = x0 - x3; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
153 b6 = vec_sub(x1, x2); /* b6 = x1 - x2; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
154 b5 = vec_add(b6, b2); /* b5 = b6 + b2; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
155 cnst = LD_W2; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
156 b5 = vec_madd(cnst, b5, mzero); /* b5 = b5 * W2; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
157 cnst = LD_W1; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
158 b2 = vec_madd(cnst, b2, b5); /* b2 = b5 + b2 * W1; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
159 cnst = LD_W0; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
160 b6 = vec_madd(cnst, b6, b5); /* b6 = b5 + b6 * W0; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
161 \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
162 x0 = vec_add(x4, x7); /* x0 = x4 + x7; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
163 x1 = vec_add(x5, x6); /* x1 = x5 + x6; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
164 x2 = vec_add(x4, x6); /* x2 = x4 + x6; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
165 x3 = vec_add(x5, x7); /* x3 = x5 + x7; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
166 x8 = vec_add(x2, x3); /* x8 = x2 + x3; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
167 cnst = LD_W3; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
168 x8 = vec_madd(cnst, x8, mzero); /* x8 = x8 * W3; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
169 \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
170 cnst = LD_W8; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
171 x0 = vec_madd(cnst, x0, mzero); /* x0 *= W8; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
172 cnst = LD_W9; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
173 x1 = vec_madd(cnst, x1, mzero); /* x1 *= W9; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
174 cnst = LD_WA; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
175 x2 = vec_madd(cnst, x2, x8); /* x2 = x2 * WA + x8; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
176 cnst = LD_WB; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
177 x3 = vec_madd(cnst, x3, x8); /* x3 = x3 * WB + x8; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
178 \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
179 cnst = LD_W4; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
180 b7 = vec_madd(cnst, x4, x0); /* b7 = x4 * W4 + x0; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
181 cnst = LD_W5; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
182 b5 = vec_madd(cnst, x5, x1); /* b5 = x5 * W5 + x1; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
183 cnst = LD_W6; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
184 b3 = vec_madd(cnst, x6, x1); /* b3 = x6 * W6 + x1; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
185 cnst = LD_W7; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
186 b1 = vec_madd(cnst, x7, x0); /* b1 = x7 * W7 + x0; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
187 \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
188 b7 = vec_add(b7, x2); /* b7 += x2; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
189 b5 = vec_add(b5, x3); /* b5 += x3; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
190 b3 = vec_add(b3, x2); /* b3 += x2; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
191 b1 = vec_add(b1, x3); /* b1 += x3; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
192 /* }}} */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
193
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
194
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
195
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
196 /* two dimensional discrete cosine transform */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
197
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
198 void fdct_altivec(int16_t *block)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
199 {
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
200 POWERPC_PERF_DECLARE(altivec_fdct, 1);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
201 vector signed short *bp;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
202 vector float *cp;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
203 vector float b00, b10, b20, b30, b40, b50, b60, b70;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
204 vector float b01, b11, b21, b31, b41, b51, b61, b71;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
205 vector float mzero, cnst, cnsts0, cnsts1, cnsts2;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
206 vector float x0, x1, x2, x3, x4, x5, x6, x7, x8;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
207
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
208 POWERPC_PERF_START_COUNT(altivec_fdct, 1);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
209
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
210
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
211 /* setup constants {{{ */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
212 /* mzero = -0.0 */
2612
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
213 mzero = ((vector float)vec_splat_u32(-1));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
214 mzero = ((vector float)vec_sl(vu32(mzero), vu32(mzero)));
1578
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
215 cp = fdctconsts;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
216 cnsts0 = vec_ld(0, cp); cp++;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
217 cnsts1 = vec_ld(0, cp); cp++;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
218 cnsts2 = vec_ld(0, cp);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
219 /* }}} */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
220
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
221
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
222 /* 8x8 matrix transpose (vector short[8]) {{{ */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
223 #define MERGE_S16(hl,a,b) vec_merge##hl(vs16(a), vs16(b))
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
224
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
225 bp = (vector signed short*)block;
2612
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
226 b00 = ((vector float)vec_ld(0, bp));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
227 b40 = ((vector float)vec_ld(16*4, bp));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
228 b01 = ((vector float)MERGE_S16(h, b00, b40));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
229 b11 = ((vector float)MERGE_S16(l, b00, b40));
1578
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
230 bp++;
2612
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
231 b10 = ((vector float)vec_ld(0, bp));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
232 b50 = ((vector float)vec_ld(16*4, bp));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
233 b21 = ((vector float)MERGE_S16(h, b10, b50));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
234 b31 = ((vector float)MERGE_S16(l, b10, b50));
1578
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
235 bp++;
2612
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
236 b20 = ((vector float)vec_ld(0, bp));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
237 b60 = ((vector float)vec_ld(16*4, bp));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
238 b41 = ((vector float)MERGE_S16(h, b20, b60));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
239 b51 = ((vector float)MERGE_S16(l, b20, b60));
1578
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
240 bp++;
2612
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
241 b30 = ((vector float)vec_ld(0, bp));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
242 b70 = ((vector float)vec_ld(16*4, bp));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
243 b61 = ((vector float)MERGE_S16(h, b30, b70));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
244 b71 = ((vector float)MERGE_S16(l, b30, b70));
1578
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
245
2612
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
246 x0 = ((vector float)MERGE_S16(h, b01, b41));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
247 x1 = ((vector float)MERGE_S16(l, b01, b41));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
248 x2 = ((vector float)MERGE_S16(h, b11, b51));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
249 x3 = ((vector float)MERGE_S16(l, b11, b51));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
250 x4 = ((vector float)MERGE_S16(h, b21, b61));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
251 x5 = ((vector float)MERGE_S16(l, b21, b61));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
252 x6 = ((vector float)MERGE_S16(h, b31, b71));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
253 x7 = ((vector float)MERGE_S16(l, b31, b71));
1578
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
254
2612
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
255 b00 = ((vector float)MERGE_S16(h, x0, x4));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
256 b10 = ((vector float)MERGE_S16(l, x0, x4));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
257 b20 = ((vector float)MERGE_S16(h, x1, x5));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
258 b30 = ((vector float)MERGE_S16(l, x1, x5));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
259 b40 = ((vector float)MERGE_S16(h, x2, x6));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
260 b50 = ((vector float)MERGE_S16(l, x2, x6));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
261 b60 = ((vector float)MERGE_S16(h, x3, x7));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
262 b70 = ((vector float)MERGE_S16(l, x3, x7));
1578
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
263
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
264 #undef MERGE_S16
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
265 /* }}} */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
266
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
267
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
268 /* Some of the initial calculations can be done as vector short before
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
269 * conversion to vector float. The following code section takes advantage
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
270 * of this.
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
271 */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
272 #if 1
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
273 /* fdct rows {{{ */
2612
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
274 x0 = ((vector float)vec_add(vs16(b00), vs16(b70)));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
275 x7 = ((vector float)vec_sub(vs16(b00), vs16(b70)));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
276 x1 = ((vector float)vec_add(vs16(b10), vs16(b60)));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
277 x6 = ((vector float)vec_sub(vs16(b10), vs16(b60)));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
278 x2 = ((vector float)vec_add(vs16(b20), vs16(b50)));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
279 x5 = ((vector float)vec_sub(vs16(b20), vs16(b50)));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
280 x3 = ((vector float)vec_add(vs16(b30), vs16(b40)));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
281 x4 = ((vector float)vec_sub(vs16(b30), vs16(b40)));
1578
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
282
2612
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
283 b70 = ((vector float)vec_add(vs16(x0), vs16(x3)));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
284 b10 = ((vector float)vec_add(vs16(x1), vs16(x2)));
1578
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
285
2612
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
286 b00 = ((vector float)vec_add(vs16(b70), vs16(b10)));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
287 b40 = ((vector float)vec_sub(vs16(b70), vs16(b10)));
1578
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
288
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
289 #define CTF0(n) \
2612
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
290 b##n##1 = ((vector float)vec_unpackl(vs16(b##n##0))); \
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
291 b##n##0 = ((vector float)vec_unpackh(vs16(b##n##0))); \
1578
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
292 b##n##1 = vec_ctf(vs32(b##n##1), 0); \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
293 b##n##0 = vec_ctf(vs32(b##n##0), 0);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
294
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
295 CTF0(0);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
296 CTF0(4);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
297
2612
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
298 b20 = ((vector float)vec_sub(vs16(x0), vs16(x3)));
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
299 b60 = ((vector float)vec_sub(vs16(x1), vs16(x2)));
1578
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
300
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
301 CTF0(2);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
302 CTF0(6);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
303
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
304 #undef CTF0
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
305
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
306 x0 = vec_add(b60, b20);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
307 x1 = vec_add(b61, b21);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
308
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
309 cnst = LD_W2;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
310 x0 = vec_madd(cnst, x0, mzero);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
311 x1 = vec_madd(cnst, x1, mzero);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
312 cnst = LD_W1;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
313 b20 = vec_madd(cnst, b20, x0);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
314 b21 = vec_madd(cnst, b21, x1);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
315 cnst = LD_W0;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
316 b60 = vec_madd(cnst, b60, x0);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
317 b61 = vec_madd(cnst, b61, x1);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
318
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
319 #define CTFX(x,b) \
2612
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
320 b##0 = ((vector float)vec_unpackh(vs16(x))); \
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
321 b##1 = ((vector float)vec_unpackl(vs16(x))); \
1578
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
322 b##0 = vec_ctf(vs32(b##0), 0); \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
323 b##1 = vec_ctf(vs32(b##1), 0); \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
324
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
325 CTFX(x4, b7);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
326 CTFX(x5, b5);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
327 CTFX(x6, b3);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
328 CTFX(x7, b1);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
329
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
330 #undef CTFX
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
331
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
332
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
333 x0 = vec_add(b70, b10);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
334 x1 = vec_add(b50, b30);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
335 x2 = vec_add(b70, b30);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
336 x3 = vec_add(b50, b10);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
337 x8 = vec_add(x2, x3);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
338 cnst = LD_W3;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
339 x8 = vec_madd(cnst, x8, mzero);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
340
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
341 cnst = LD_W8;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
342 x0 = vec_madd(cnst, x0, mzero);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
343 cnst = LD_W9;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
344 x1 = vec_madd(cnst, x1, mzero);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
345 cnst = LD_WA;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
346 x2 = vec_madd(cnst, x2, x8);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
347 cnst = LD_WB;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
348 x3 = vec_madd(cnst, x3, x8);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
349
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
350 cnst = LD_W4;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
351 b70 = vec_madd(cnst, b70, x0);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
352 cnst = LD_W5;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
353 b50 = vec_madd(cnst, b50, x1);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
354 cnst = LD_W6;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
355 b30 = vec_madd(cnst, b30, x1);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
356 cnst = LD_W7;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
357 b10 = vec_madd(cnst, b10, x0);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
358
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
359 b70 = vec_add(b70, x2);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
360 b50 = vec_add(b50, x3);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
361 b30 = vec_add(b30, x2);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
362 b10 = vec_add(b10, x3);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
363
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
364
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
365 x0 = vec_add(b71, b11);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
366 x1 = vec_add(b51, b31);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
367 x2 = vec_add(b71, b31);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
368 x3 = vec_add(b51, b11);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
369 x8 = vec_add(x2, x3);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
370 cnst = LD_W3;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
371 x8 = vec_madd(cnst, x8, mzero);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
372
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
373 cnst = LD_W8;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
374 x0 = vec_madd(cnst, x0, mzero);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
375 cnst = LD_W9;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
376 x1 = vec_madd(cnst, x1, mzero);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
377 cnst = LD_WA;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
378 x2 = vec_madd(cnst, x2, x8);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
379 cnst = LD_WB;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
380 x3 = vec_madd(cnst, x3, x8);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
381
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
382 cnst = LD_W4;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
383 b71 = vec_madd(cnst, b71, x0);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
384 cnst = LD_W5;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
385 b51 = vec_madd(cnst, b51, x1);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
386 cnst = LD_W6;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
387 b31 = vec_madd(cnst, b31, x1);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
388 cnst = LD_W7;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
389 b11 = vec_madd(cnst, b11, x0);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
390
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
391 b71 = vec_add(b71, x2);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
392 b51 = vec_add(b51, x3);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
393 b31 = vec_add(b31, x2);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
394 b11 = vec_add(b11, x3);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
395 /* }}} */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
396 #else
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
397 /* convert to float {{{ */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
398 #define CTF(n) \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
399 vs32(b##n##1) = vec_unpackl(vs16(b##n##0)); \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
400 vs32(b##n##0) = vec_unpackh(vs16(b##n##0)); \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
401 b##n##1 = vec_ctf(vs32(b##n##1), 0); \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
402 b##n##0 = vec_ctf(vs32(b##n##0), 0); \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
403
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
404 CTF(0);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
405 CTF(1);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
406 CTF(2);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
407 CTF(3);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
408 CTF(4);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
409 CTF(5);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
410 CTF(6);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
411 CTF(7);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
412
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
413 #undef CTF
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
414 /* }}} */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
415
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
416 FDCTROW(b00, b10, b20, b30, b40, b50, b60, b70);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
417 FDCTROW(b01, b11, b21, b31, b41, b51, b61, b71);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
418 #endif
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
419
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
420
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
421 /* 8x8 matrix transpose (vector float[8][2]) {{{ */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
422 x0 = vec_mergel(b00, b20);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
423 x1 = vec_mergeh(b00, b20);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
424 x2 = vec_mergel(b10, b30);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
425 x3 = vec_mergeh(b10, b30);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
426
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
427 b00 = vec_mergeh(x1, x3);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
428 b10 = vec_mergel(x1, x3);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
429 b20 = vec_mergeh(x0, x2);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
430 b30 = vec_mergel(x0, x2);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
431
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
432 x4 = vec_mergel(b41, b61);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
433 x5 = vec_mergeh(b41, b61);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
434 x6 = vec_mergel(b51, b71);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
435 x7 = vec_mergeh(b51, b71);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
436
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
437 b41 = vec_mergeh(x5, x7);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
438 b51 = vec_mergel(x5, x7);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
439 b61 = vec_mergeh(x4, x6);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
440 b71 = vec_mergel(x4, x6);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
441
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
442 x0 = vec_mergel(b01, b21);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
443 x1 = vec_mergeh(b01, b21);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
444 x2 = vec_mergel(b11, b31);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
445 x3 = vec_mergeh(b11, b31);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
446
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
447 x4 = vec_mergel(b40, b60);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
448 x5 = vec_mergeh(b40, b60);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
449 x6 = vec_mergel(b50, b70);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
450 x7 = vec_mergeh(b50, b70);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
451
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
452 b40 = vec_mergeh(x1, x3);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
453 b50 = vec_mergel(x1, x3);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
454 b60 = vec_mergeh(x0, x2);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
455 b70 = vec_mergel(x0, x2);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
456
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
457 b01 = vec_mergeh(x5, x7);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
458 b11 = vec_mergel(x5, x7);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
459 b21 = vec_mergeh(x4, x6);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
460 b31 = vec_mergel(x4, x6);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
461 /* }}} */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
462
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
463
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
464 FDCTCOL(b00, b10, b20, b30, b40, b50, b60, b70);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
465 FDCTCOL(b01, b11, b21, b31, b41, b51, b61, b71);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
466
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
467
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
468 /* round, convert back to short {{{ */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
469 #define CTS(n) \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
470 b##n##0 = vec_round(b##n##0); \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
471 b##n##1 = vec_round(b##n##1); \
2612
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
472 b##n##0 = ((vector float)vec_cts(b##n##0, 0)); \
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
473 b##n##1 = ((vector float)vec_cts(b##n##1, 0)); \
210cca8f5764 gcc 4 compilation fix
diego
parents: 1578
diff changeset
474 b##n##0 = ((vector float)vec_pack(vs32(b##n##0), vs32(b##n##1))); \
1578
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
475 vec_st(vs16(b##n##0), 0, bp);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
476
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
477 bp = (vector signed short*)block;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
478 CTS(0); bp++;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
479 CTS(1); bp++;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
480 CTS(2); bp++;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
481 CTS(3); bp++;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
482 CTS(4); bp++;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
483 CTS(5); bp++;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
484 CTS(6); bp++;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
485 CTS(7);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
486
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
487 #undef CTS
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
488 /* }}} */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
489
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
490 POWERPC_PERF_STOP_COUNT(altivec_fdct, 1);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
491 }
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
492
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
493 /* vim:set foldmethod=marker foldlevel=0: */