annotate ppc/fdct_altivec.c @ 2144:e33371c1e2b4 libavcodec

better context model (0.2-1% lower bitrate) a tiny bit cleanup
author michael
date Tue, 27 Jul 2004 01:30:44 +0000
parents 6a4cfc5f9f96
children 210cca8f5764
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 *
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
5 * This library is free software; you can redistribute it and/or
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
6 * 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
7 * 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
8 * 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
9 *
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
10 * This library is distributed in the hope that it will be useful,
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
11 * 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
12 * 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
13 * Lesser General Public License for more details.
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
14 *
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
15 * You should have received a copy of the GNU Lesser General Public
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
16 * License along with this library; if not, write to the Free Software
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
18 */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
19
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 #include "../common.h"
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
22 #include "../dsputil.h"
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
23 #include "dsputil_altivec.h"
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
24 #include "gcc_fixes.h"
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
25
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
26
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
27 #define vs16(v) ((vector signed short)(v))
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
28 #define vs32(v) ((vector signed int)(v))
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
29 #define vu8(v) ((vector unsigned char)(v))
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
30 #define vu16(v) ((vector unsigned short)(v))
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
31 #define vu32(v) ((vector unsigned int)(v))
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
32
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
33
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
34 #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
35 #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
36 #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
37 #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
38 #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
39 #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
40 #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
41 #define SQRT_2 1.41421353816986083984375000 /* sqrt(2) */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
42
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
43
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
44 #define W0 -(2 * C2)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
45 #define W1 (2 * C6)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
46 #define W2 (SQRT_2 * C6)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
47 #define W3 (SQRT_2 * C3)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
48 #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
49 #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
50 #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
51 #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
52 #define W8 (SQRT_2 * ( C7 - C3))
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
53 #define W9 (SQRT_2 * (-C1 - C3))
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
54 #define WA (SQRT_2 * (-C3 - C5))
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
55 #define WB (SQRT_2 * ( C5 - C3))
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
56
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
57
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
58 static vector float fdctconsts[3] = {
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
59 (vector float)AVV( W0, W1, W2, W3 ),
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
60 (vector float)AVV( W4, W5, W6, W7 ),
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
61 (vector float)AVV( W8, W9, WA, WB )
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
62 };
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
63
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
64 #define LD_W0 vec_splat(cnsts0, 0)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
65 #define LD_W1 vec_splat(cnsts0, 1)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
66 #define LD_W2 vec_splat(cnsts0, 2)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
67 #define LD_W3 vec_splat(cnsts0, 3)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
68 #define LD_W4 vec_splat(cnsts1, 0)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
69 #define LD_W5 vec_splat(cnsts1, 1)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
70 #define LD_W6 vec_splat(cnsts1, 2)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
71 #define LD_W7 vec_splat(cnsts1, 3)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
72 #define LD_W8 vec_splat(cnsts2, 0)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
73 #define LD_W9 vec_splat(cnsts2, 1)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
74 #define LD_WA vec_splat(cnsts2, 2)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
75 #define LD_WB vec_splat(cnsts2, 3)
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
76
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
77
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
78 #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
79 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
80 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
81 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
82 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
83 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
84 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
85 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
86 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
87 \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
88 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
89 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
90 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
91 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
92 \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
93 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
94 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
95 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
96 cnst = LD_W2; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
97 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
98 cnst = LD_W1; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
99 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
100 cnst = LD_W0; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
101 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
102 \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
103 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
104 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
105 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
106 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
107 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
108 cnst = LD_W3; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
109 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
110 \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
111 cnst = LD_W8; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
112 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
113 cnst = LD_W9; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
114 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
115 cnst = LD_WA; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
116 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
117 cnst = LD_WB; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
118 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
119 \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
120 cnst = LD_W4; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
121 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
122 cnst = LD_W5; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
123 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
124 cnst = LD_W6; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
125 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
126 cnst = LD_W7; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
127 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
128 \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
129 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
130 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
131 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
132 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
133 /* }}} */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
134
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
135 #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
136 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
137 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
138 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
139 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
140 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
141 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
142 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
143 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
144 \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
145 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
146 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
147 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
148 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
149 \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
150 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
151 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
152 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
153 cnst = LD_W2; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
154 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
155 cnst = LD_W1; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
156 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
157 cnst = LD_W0; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
158 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
159 \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
160 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
161 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
162 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
163 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
164 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
165 cnst = LD_W3; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
166 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
167 \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
168 cnst = LD_W8; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
169 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
170 cnst = LD_W9; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
171 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
172 cnst = LD_WA; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
173 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
174 cnst = LD_WB; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
175 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
176 \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
177 cnst = LD_W4; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
178 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
179 cnst = LD_W5; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
180 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
181 cnst = LD_W6; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
182 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
183 cnst = LD_W7; \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
184 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
185 \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
186 b7 = vec_add(b7, x2); /* b7 += x2; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
187 b5 = vec_add(b5, x3); /* b5 += x3; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
188 b3 = vec_add(b3, x2); /* b3 += x2; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
189 b1 = vec_add(b1, x3); /* b1 += x3; */ \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
190 /* }}} */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
191
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 /* two dimensional discrete cosine transform */
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 void fdct_altivec(int16_t *block)
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 POWERPC_PERF_DECLARE(altivec_fdct, 1);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
199 #ifdef ALTIVEC_USE_REFERENCE_C_CODE
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
200 POWERPC_PERF_START_COUNT(altivec_fdct, 1);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
201 void ff_jpeg_fdct_islow(int16_t *block);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
202 ff_jpeg_fdct_islow(block);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
203 POWERPC_PERF_STOP_COUNT(altivec_fdct, 1);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
204 #else /* ALTIVEC_USE_REFERENCE_C_CODE */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
205 vector signed short *bp;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
206 vector float *cp;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
207 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
208 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
209 vector float mzero, cnst, cnsts0, cnsts1, cnsts2;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
210 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
211
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
212 POWERPC_PERF_START_COUNT(altivec_fdct, 1);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
213
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
214
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
215 /* setup constants {{{ */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
216 /* mzero = -0.0 */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
217 vu32(mzero) = vec_splat_u32(-1);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
218 vu32(mzero) = vec_sl(vu32(mzero), vu32(mzero));
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
219 cp = fdctconsts;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
220 cnsts0 = vec_ld(0, cp); cp++;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
221 cnsts1 = vec_ld(0, cp); cp++;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
222 cnsts2 = vec_ld(0, cp);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
223 /* }}} */
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
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
226 /* 8x8 matrix transpose (vector short[8]) {{{ */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
227 #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
228
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
229 bp = (vector signed short*)block;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
230 vs16(b00) = vec_ld(0, bp);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
231 vs16(b40) = vec_ld(16*4, bp);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
232 vs16(b01) = MERGE_S16(h, b00, b40);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
233 vs16(b11) = MERGE_S16(l, b00, b40);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
234 bp++;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
235 vs16(b10) = vec_ld(0, bp);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
236 vs16(b50) = vec_ld(16*4, bp);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
237 vs16(b21) = MERGE_S16(h, b10, b50);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
238 vs16(b31) = MERGE_S16(l, b10, b50);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
239 bp++;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
240 vs16(b20) = vec_ld(0, bp);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
241 vs16(b60) = vec_ld(16*4, bp);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
242 vs16(b41) = MERGE_S16(h, b20, b60);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
243 vs16(b51) = MERGE_S16(l, b20, b60);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
244 bp++;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
245 vs16(b30) = vec_ld(0, bp);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
246 vs16(b70) = vec_ld(16*4, bp);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
247 vs16(b61) = MERGE_S16(h, b30, b70);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
248 vs16(b71) = MERGE_S16(l, b30, b70);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
249
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
250 vs16(x0) = MERGE_S16(h, b01, b41);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
251 vs16(x1) = MERGE_S16(l, b01, b41);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
252 vs16(x2) = MERGE_S16(h, b11, b51);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
253 vs16(x3) = MERGE_S16(l, b11, b51);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
254 vs16(x4) = MERGE_S16(h, b21, b61);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
255 vs16(x5) = MERGE_S16(l, b21, b61);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
256 vs16(x6) = MERGE_S16(h, b31, b71);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
257 vs16(x7) = MERGE_S16(l, b31, b71);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
258
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
259 vs16(b00) = MERGE_S16(h, x0, x4);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
260 vs16(b10) = MERGE_S16(l, x0, x4);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
261 vs16(b20) = MERGE_S16(h, x1, x5);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
262 vs16(b30) = MERGE_S16(l, x1, x5);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
263 vs16(b40) = MERGE_S16(h, x2, x6);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
264 vs16(b50) = MERGE_S16(l, x2, x6);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
265 vs16(b60) = MERGE_S16(h, x3, x7);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
266 vs16(b70) = MERGE_S16(l, x3, x7);
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 #undef MERGE_S16
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
269 /* }}} */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
270
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 /* 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
273 * 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
274 * of this.
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
275 */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
276 #if 1
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
277 /* fdct rows {{{ */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
278 vs16(x0) = vec_add(vs16(b00), vs16(b70));
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
279 vs16(x7) = vec_sub(vs16(b00), vs16(b70));
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
280 vs16(x1) = vec_add(vs16(b10), vs16(b60));
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
281 vs16(x6) = vec_sub(vs16(b10), vs16(b60));
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
282 vs16(x2) = vec_add(vs16(b20), vs16(b50));
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
283 vs16(x5) = vec_sub(vs16(b20), vs16(b50));
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
284 vs16(x3) = vec_add(vs16(b30), vs16(b40));
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
285 vs16(x4) = vec_sub(vs16(b30), vs16(b40));
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
286
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
287 vs16(b70) = vec_add(vs16(x0), vs16(x3));
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
288 vs16(b10) = vec_add(vs16(x1), vs16(x2));
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
289
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
290 vs16(b00) = vec_add(vs16(b70), vs16(b10));
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
291 vs16(b40) = vec_sub(vs16(b70), vs16(b10));
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
292
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
293 #define CTF0(n) \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
294 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
295 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
296 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
297 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
298
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
299 CTF0(0);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
300 CTF0(4);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
301
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
302 vs16(b20) = vec_sub(vs16(x0), vs16(x3));
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
303 vs16(b60) = vec_sub(vs16(x1), vs16(x2));
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
304
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
305 CTF0(2);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
306 CTF0(6);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
307
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
308 #undef CTF0
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
309
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
310 x0 = vec_add(b60, b20);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
311 x1 = vec_add(b61, b21);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
312
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
313 cnst = LD_W2;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
314 x0 = vec_madd(cnst, x0, mzero);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
315 x1 = vec_madd(cnst, x1, mzero);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
316 cnst = LD_W1;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
317 b20 = vec_madd(cnst, b20, x0);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
318 b21 = vec_madd(cnst, b21, x1);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
319 cnst = LD_W0;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
320 b60 = vec_madd(cnst, b60, x0);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
321 b61 = vec_madd(cnst, b61, x1);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
322
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
323 #define CTFX(x,b) \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
324 vs32(b##0) = vec_unpackh(vs16(x)); \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
325 vs32(b##1) = vec_unpackl(vs16(x)); \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
326 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
327 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
328
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
329 CTFX(x4, b7);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
330 CTFX(x5, b5);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
331 CTFX(x6, b3);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
332 CTFX(x7, b1);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
333
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
334 #undef CTFX
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
335
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
336
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
337 x0 = vec_add(b70, b10);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
338 x1 = vec_add(b50, b30);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
339 x2 = vec_add(b70, b30);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
340 x3 = vec_add(b50, b10);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
341 x8 = vec_add(x2, x3);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
342 cnst = LD_W3;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
343 x8 = vec_madd(cnst, x8, mzero);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
344
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
345 cnst = LD_W8;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
346 x0 = vec_madd(cnst, x0, mzero);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
347 cnst = LD_W9;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
348 x1 = vec_madd(cnst, x1, mzero);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
349 cnst = LD_WA;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
350 x2 = vec_madd(cnst, x2, x8);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
351 cnst = LD_WB;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
352 x3 = vec_madd(cnst, x3, x8);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
353
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
354 cnst = LD_W4;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
355 b70 = vec_madd(cnst, b70, x0);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
356 cnst = LD_W5;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
357 b50 = vec_madd(cnst, b50, x1);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
358 cnst = LD_W6;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
359 b30 = vec_madd(cnst, b30, x1);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
360 cnst = LD_W7;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
361 b10 = vec_madd(cnst, b10, x0);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
362
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
363 b70 = vec_add(b70, x2);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
364 b50 = vec_add(b50, x3);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
365 b30 = vec_add(b30, x2);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
366 b10 = vec_add(b10, x3);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
367
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
368
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
369 x0 = vec_add(b71, b11);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
370 x1 = vec_add(b51, b31);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
371 x2 = vec_add(b71, b31);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
372 x3 = vec_add(b51, b11);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
373 x8 = vec_add(x2, x3);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
374 cnst = LD_W3;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
375 x8 = vec_madd(cnst, x8, mzero);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
376
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
377 cnst = LD_W8;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
378 x0 = vec_madd(cnst, x0, mzero);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
379 cnst = LD_W9;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
380 x1 = vec_madd(cnst, x1, mzero);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
381 cnst = LD_WA;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
382 x2 = vec_madd(cnst, x2, x8);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
383 cnst = LD_WB;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
384 x3 = vec_madd(cnst, x3, x8);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
385
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
386 cnst = LD_W4;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
387 b71 = vec_madd(cnst, b71, x0);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
388 cnst = LD_W5;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
389 b51 = vec_madd(cnst, b51, x1);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
390 cnst = LD_W6;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
391 b31 = vec_madd(cnst, b31, x1);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
392 cnst = LD_W7;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
393 b11 = vec_madd(cnst, b11, x0);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
394
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
395 b71 = vec_add(b71, x2);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
396 b51 = vec_add(b51, x3);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
397 b31 = vec_add(b31, x2);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
398 b11 = vec_add(b11, x3);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
399 /* }}} */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
400 #else
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
401 /* convert to float {{{ */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
402 #define CTF(n) \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
403 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
404 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
405 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
406 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
407
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
408 CTF(0);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
409 CTF(1);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
410 CTF(2);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
411 CTF(3);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
412 CTF(4);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
413 CTF(5);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
414 CTF(6);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
415 CTF(7);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
416
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
417 #undef CTF
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
418 /* }}} */
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 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
421 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
422 #endif
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
423
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
424
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
425 /* 8x8 matrix transpose (vector float[8][2]) {{{ */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
426 x0 = vec_mergel(b00, b20);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
427 x1 = vec_mergeh(b00, b20);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
428 x2 = vec_mergel(b10, b30);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
429 x3 = vec_mergeh(b10, b30);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
430
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
431 b00 = vec_mergeh(x1, x3);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
432 b10 = vec_mergel(x1, x3);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
433 b20 = vec_mergeh(x0, x2);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
434 b30 = vec_mergel(x0, x2);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
435
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
436 x4 = vec_mergel(b41, b61);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
437 x5 = vec_mergeh(b41, b61);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
438 x6 = vec_mergel(b51, b71);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
439 x7 = vec_mergeh(b51, b71);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
440
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
441 b41 = vec_mergeh(x5, x7);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
442 b51 = vec_mergel(x5, x7);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
443 b61 = vec_mergeh(x4, x6);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
444 b71 = vec_mergel(x4, x6);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
445
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
446 x0 = vec_mergel(b01, b21);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
447 x1 = vec_mergeh(b01, b21);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
448 x2 = vec_mergel(b11, b31);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
449 x3 = vec_mergeh(b11, b31);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
450
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
451 x4 = vec_mergel(b40, b60);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
452 x5 = vec_mergeh(b40, b60);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
453 x6 = vec_mergel(b50, b70);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
454 x7 = vec_mergeh(b50, b70);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
455
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
456 b40 = vec_mergeh(x1, x3);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
457 b50 = vec_mergel(x1, x3);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
458 b60 = vec_mergeh(x0, x2);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
459 b70 = vec_mergel(x0, x2);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
460
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
461 b01 = vec_mergeh(x5, x7);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
462 b11 = vec_mergel(x5, x7);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
463 b21 = vec_mergeh(x4, x6);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
464 b31 = vec_mergel(x4, x6);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
465 /* }}} */
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 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
469 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
470
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
471
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
472 /* round, convert back to short {{{ */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
473 #define CTS(n) \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
474 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
475 b##n##1 = vec_round(b##n##1); \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
476 vs32(b##n##0) = vec_cts(b##n##0, 0); \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
477 vs32(b##n##1) = vec_cts(b##n##1, 0); \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
478 vs16(b##n##0) = vec_pack(vs32(b##n##0), vs32(b##n##1)); \
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
479 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
480
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
481 bp = (vector signed short*)block;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
482 CTS(0); bp++;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
483 CTS(1); bp++;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
484 CTS(2); bp++;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
485 CTS(3); bp++;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
486 CTS(4); bp++;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
487 CTS(5); bp++;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
488 CTS(6); bp++;
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
489 CTS(7);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
490
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
491 #undef CTS
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
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
494 POWERPC_PERF_STOP_COUNT(altivec_fdct, 1);
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
495 #endif /* ALTIVEC_USE_REFERENCE_C_CODE */
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
496 }
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
497
6a4cfc5f9f96 AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
michael
parents:
diff changeset
498 /* vim:set foldmethod=marker foldlevel=0: */