annotate liba52/imdct.c @ 20567:a45bc176a16d

paragraph about groff and UTF-8
author diego
date Wed, 01 Nov 2006 17:31:11 +0000
parents 6334c14b38eb
children a7b716b53e9f
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
1 /*
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
2 * imdct.c
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
3 * Copyright (C) 2000-2002 Michel Lespinasse <walken@zoy.org>
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
4 * Copyright (C) 1999-2000 Aaron Holtzman <aholtzma@ess.engr.uvic.ca>
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
5 *
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
6 * The ifft algorithms in this file have been largely inspired by Dan
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
7 * Bernstein's work, djbfft, available at http://cr.yp.to/djbfft.html
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
8 *
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
9 * This file is part of a52dec, a free ATSC A-52 stream decoder.
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
10 * See http://liba52.sourceforge.net/ for updates.
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
11 *
14991
07f1e7669772 Mark modified files as such to comply more closely with GPL ¡ø2a.
diego
parents: 12303
diff changeset
12 * Modified for use with MPlayer, changes contained in liba52_changes.diff.
18783
0783dd397f74 CVS --> Subversion in copyright notices
diego
parents: 18721
diff changeset
13 * detailed changelog at http://svn.mplayerhq.hu/mplayer/trunk/
14991
07f1e7669772 Mark modified files as such to comply more closely with GPL ¡ø2a.
diego
parents: 12303
diff changeset
14 * $Id$
07f1e7669772 Mark modified files as such to comply more closely with GPL ¡ø2a.
diego
parents: 12303
diff changeset
15 *
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
16 * a52dec is free software; you can redistribute it and/or modify
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
17 * it under the terms of the GNU General Public License as published by
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
18 * the Free Software Foundation; either version 2 of the License, or
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
19 * (at your option) any later version.
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
20 *
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
21 * a52dec is distributed in the hope that it will be useful,
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
22 * but WITHOUT ANY WARRANTY; without even the implied warranty of
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
23 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
24 * GNU General Public License for more details.
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
25 *
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
26 * You should have received a copy of the GNU General Public License
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
27 * along with this program; if not, write to the Free Software
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
28 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
3579
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
29 *
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
30 * SSE optimizations from Michael Niedermayer (michaelni@gmx.at)
3884
0410677eda4a ported 3dnow(ex) optimizations from libac3
michael
parents: 3720
diff changeset
31 * 3DNOW optimizations from Nick Kurshev <nickols_k@mail.ru>
0410677eda4a ported 3dnow(ex) optimizations from libac3
michael
parents: 3720
diff changeset
32 * michael did port them from libac3 (untested, perhaps totally broken)
9001
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
33 * AltiVec optimizations from Romain Dolbeau (romain@dolbeau.org)
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
34 */
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
35
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
36 #include "config.h"
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
37
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
38 #include <math.h>
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
39 #include <stdio.h>
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
40 #ifdef LIBA52_DJBFFT
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
41 #include <fftc4.h>
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
42 #endif
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
43 #ifndef M_PI
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
44 #define M_PI 3.1415926535897932384626433832795029
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
45 #endif
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
46 #include <inttypes.h>
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
47
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
48 #include "a52.h"
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
49 #include "a52_internal.h"
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
50 #include "mm_accel.h"
4247
2dbd637ffe05 mangle for win32 in liba52 (includes dummy mangle.h pointing to the one in main)
atmos4
parents: 3908
diff changeset
51 #include "mangle.h"
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
52
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
53 void (*a52_imdct_512) (sample_t * data, sample_t * delay, sample_t bias);
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
54
3884
0410677eda4a ported 3dnow(ex) optimizations from libac3
michael
parents: 3720
diff changeset
55 #ifdef RUNTIME_CPUDETECT
0410677eda4a ported 3dnow(ex) optimizations from libac3
michael
parents: 3720
diff changeset
56 #undef HAVE_3DNOWEX
0410677eda4a ported 3dnow(ex) optimizations from libac3
michael
parents: 3720
diff changeset
57 #endif
0410677eda4a ported 3dnow(ex) optimizations from libac3
michael
parents: 3720
diff changeset
58
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
59 typedef struct complex_s {
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
60 sample_t real;
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
61 sample_t imag;
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
62 } complex_t;
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
63
12303
f881c918739b attribute_used patch by (VMiklos <mamajom at axelero dot hu>)
michael
parents: 9122
diff changeset
64 static const int pm128[128] attribute_used __attribute__((aligned(16))) =
3884
0410677eda4a ported 3dnow(ex) optimizations from libac3
michael
parents: 3720
diff changeset
65 {
0410677eda4a ported 3dnow(ex) optimizations from libac3
michael
parents: 3720
diff changeset
66 0, 16, 32, 48, 64, 80, 96, 112, 8, 40, 72, 104, 24, 56, 88, 120,
0410677eda4a ported 3dnow(ex) optimizations from libac3
michael
parents: 3720
diff changeset
67 4, 20, 36, 52, 68, 84, 100, 116, 12, 28, 44, 60, 76, 92, 108, 124,
0410677eda4a ported 3dnow(ex) optimizations from libac3
michael
parents: 3720
diff changeset
68 2, 18, 34, 50, 66, 82, 98, 114, 10, 42, 74, 106, 26, 58, 90, 122,
0410677eda4a ported 3dnow(ex) optimizations from libac3
michael
parents: 3720
diff changeset
69 6, 22, 38, 54, 70, 86, 102, 118, 14, 46, 78, 110, 30, 62, 94, 126,
0410677eda4a ported 3dnow(ex) optimizations from libac3
michael
parents: 3720
diff changeset
70 1, 17, 33, 49, 65, 81, 97, 113, 9, 41, 73, 105, 25, 57, 89, 121,
0410677eda4a ported 3dnow(ex) optimizations from libac3
michael
parents: 3720
diff changeset
71 5, 21, 37, 53, 69, 85, 101, 117, 13, 29, 45, 61, 77, 93, 109, 125,
0410677eda4a ported 3dnow(ex) optimizations from libac3
michael
parents: 3720
diff changeset
72 3, 19, 35, 51, 67, 83, 99, 115, 11, 43, 75, 107, 27, 59, 91, 123,
0410677eda4a ported 3dnow(ex) optimizations from libac3
michael
parents: 3720
diff changeset
73 7, 23, 39, 55, 71, 87, 103, 119, 15, 31, 47, 63, 79, 95, 111, 127
0410677eda4a ported 3dnow(ex) optimizations from libac3
michael
parents: 3720
diff changeset
74 };
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
75
12303
f881c918739b attribute_used patch by (VMiklos <mamajom at axelero dot hu>)
michael
parents: 9122
diff changeset
76 static uint8_t attribute_used bit_reverse_512[] = {
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
77 0x00, 0x40, 0x20, 0x60, 0x10, 0x50, 0x30, 0x70,
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
78 0x08, 0x48, 0x28, 0x68, 0x18, 0x58, 0x38, 0x78,
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
79 0x04, 0x44, 0x24, 0x64, 0x14, 0x54, 0x34, 0x74,
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
80 0x0c, 0x4c, 0x2c, 0x6c, 0x1c, 0x5c, 0x3c, 0x7c,
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
81 0x02, 0x42, 0x22, 0x62, 0x12, 0x52, 0x32, 0x72,
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
82 0x0a, 0x4a, 0x2a, 0x6a, 0x1a, 0x5a, 0x3a, 0x7a,
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
83 0x06, 0x46, 0x26, 0x66, 0x16, 0x56, 0x36, 0x76,
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
84 0x0e, 0x4e, 0x2e, 0x6e, 0x1e, 0x5e, 0x3e, 0x7e,
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
85 0x01, 0x41, 0x21, 0x61, 0x11, 0x51, 0x31, 0x71,
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
86 0x09, 0x49, 0x29, 0x69, 0x19, 0x59, 0x39, 0x79,
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
87 0x05, 0x45, 0x25, 0x65, 0x15, 0x55, 0x35, 0x75,
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
88 0x0d, 0x4d, 0x2d, 0x6d, 0x1d, 0x5d, 0x3d, 0x7d,
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
89 0x03, 0x43, 0x23, 0x63, 0x13, 0x53, 0x33, 0x73,
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
90 0x0b, 0x4b, 0x2b, 0x6b, 0x1b, 0x5b, 0x3b, 0x7b,
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
91 0x07, 0x47, 0x27, 0x67, 0x17, 0x57, 0x37, 0x77,
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
92 0x0f, 0x4f, 0x2f, 0x6f, 0x1f, 0x5f, 0x3f, 0x7f};
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
93
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
94 static uint8_t fftorder[] = {
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
95 0,128, 64,192, 32,160,224, 96, 16,144, 80,208,240,112, 48,176,
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
96 8,136, 72,200, 40,168,232,104,248,120, 56,184, 24,152,216, 88,
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
97 4,132, 68,196, 36,164,228,100, 20,148, 84,212,244,116, 52,180,
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
98 252,124, 60,188, 28,156,220, 92, 12,140, 76,204,236,108, 44,172,
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
99 2,130, 66,194, 34,162,226, 98, 18,146, 82,210,242,114, 50,178,
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
100 10,138, 74,202, 42,170,234,106,250,122, 58,186, 26,154,218, 90,
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
101 254,126, 62,190, 30,158,222, 94, 14,142, 78,206,238,110, 46,174,
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
102 6,134, 70,198, 38,166,230,102,246,118, 54,182, 22,150,214, 86
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
103 };
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
104
3508
b5220cf63fc3 some SSE optimizations
michael
parents: 3394
diff changeset
105 static complex_t __attribute__((aligned(16))) buf[128];
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
106
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
107 /* Twiddle factor LUT */
9001
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
108 static complex_t __attribute__((aligned(16))) w_1[1];
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
109 static complex_t __attribute__((aligned(16))) w_2[2];
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
110 static complex_t __attribute__((aligned(16))) w_4[4];
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
111 static complex_t __attribute__((aligned(16))) w_8[8];
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
112 static complex_t __attribute__((aligned(16))) w_16[16];
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
113 static complex_t __attribute__((aligned(16))) w_32[32];
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
114 static complex_t __attribute__((aligned(16))) w_64[64];
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
115 static complex_t __attribute__((aligned(16))) * w[7] = {w_1, w_2, w_4, w_8, w_16, w_32, w_64};
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
116
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
117 /* Twiddle factors for IMDCT */
9001
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
118 static sample_t __attribute__((aligned(16))) xcos1[128];
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
119 static sample_t __attribute__((aligned(16))) xsin1[128];
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
120
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
121 #if defined(ARCH_X86) || defined(ARCH_X86_64)
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
122 // NOTE: SSE needs 16byte alignment or it will segfault
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
123 //
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
124 static float __attribute__((aligned(16))) sseSinCos1c[256];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
125 static float __attribute__((aligned(16))) sseSinCos1d[256];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
126 static float attribute_used __attribute__((aligned(16))) ps111_1[4]={1,1,1,-1};
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
127 //static float __attribute__((aligned(16))) sseW0[4];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
128 static float __attribute__((aligned(16))) sseW1[8];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
129 static float __attribute__((aligned(16))) sseW2[16];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
130 static float __attribute__((aligned(16))) sseW3[32];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
131 static float __attribute__((aligned(16))) sseW4[64];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
132 static float __attribute__((aligned(16))) sseW5[128];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
133 static float __attribute__((aligned(16))) sseW6[256];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
134 static float __attribute__((aligned(16))) *sseW[7]=
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
135 {NULL /*sseW0*/,sseW1,sseW2,sseW3,sseW4,sseW5,sseW6};
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
136 static float __attribute__((aligned(16))) sseWindow[512];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
137 #endif
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
138
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
139 /* Root values for IFFT */
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
140 static sample_t roots16[3];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
141 static sample_t roots32[7];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
142 static sample_t roots64[15];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
143 static sample_t roots128[31];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
144
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
145 /* Twiddle factors for IMDCT */
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
146 static complex_t pre1[128];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
147 static complex_t post1[64];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
148 static complex_t pre2[64];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
149 static complex_t post2[32];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
150
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
151 static sample_t a52_imdct_window[256];
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
152
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
153 static void (* ifft128) (complex_t * buf);
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
154 static void (* ifft64) (complex_t * buf);
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
155
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
156 static inline void ifft2 (complex_t * buf)
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
157 {
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
158 double r, i;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
159
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
160 r = buf[0].real;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
161 i = buf[0].imag;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
162 buf[0].real += buf[1].real;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
163 buf[0].imag += buf[1].imag;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
164 buf[1].real = r - buf[1].real;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
165 buf[1].imag = i - buf[1].imag;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
166 }
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
167
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
168 static inline void ifft4 (complex_t * buf)
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
169 {
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
170 double tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp8;
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
171
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
172 tmp1 = buf[0].real + buf[1].real;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
173 tmp2 = buf[3].real + buf[2].real;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
174 tmp3 = buf[0].imag + buf[1].imag;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
175 tmp4 = buf[2].imag + buf[3].imag;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
176 tmp5 = buf[0].real - buf[1].real;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
177 tmp6 = buf[0].imag - buf[1].imag;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
178 tmp7 = buf[2].imag - buf[3].imag;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
179 tmp8 = buf[3].real - buf[2].real;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
180
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
181 buf[0].real = tmp1 + tmp2;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
182 buf[0].imag = tmp3 + tmp4;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
183 buf[2].real = tmp1 - tmp2;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
184 buf[2].imag = tmp3 - tmp4;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
185 buf[1].real = tmp5 + tmp7;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
186 buf[1].imag = tmp6 + tmp8;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
187 buf[3].real = tmp5 - tmp7;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
188 buf[3].imag = tmp6 - tmp8;
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
189 }
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
190
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
191 /* the basic split-radix ifft butterfly */
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
192
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
193 #define BUTTERFLY(a0,a1,a2,a3,wr,wi) do { \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
194 tmp5 = a2.real * wr + a2.imag * wi; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
195 tmp6 = a2.imag * wr - a2.real * wi; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
196 tmp7 = a3.real * wr - a3.imag * wi; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
197 tmp8 = a3.imag * wr + a3.real * wi; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
198 tmp1 = tmp5 + tmp7; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
199 tmp2 = tmp6 + tmp8; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
200 tmp3 = tmp6 - tmp8; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
201 tmp4 = tmp7 - tmp5; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
202 a2.real = a0.real - tmp1; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
203 a2.imag = a0.imag - tmp2; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
204 a3.real = a1.real - tmp3; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
205 a3.imag = a1.imag - tmp4; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
206 a0.real += tmp1; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
207 a0.imag += tmp2; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
208 a1.real += tmp3; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
209 a1.imag += tmp4; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
210 } while (0)
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
211
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
212 /* split-radix ifft butterfly, specialized for wr=1 wi=0 */
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
213
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
214 #define BUTTERFLY_ZERO(a0,a1,a2,a3) do { \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
215 tmp1 = a2.real + a3.real; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
216 tmp2 = a2.imag + a3.imag; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
217 tmp3 = a2.imag - a3.imag; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
218 tmp4 = a3.real - a2.real; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
219 a2.real = a0.real - tmp1; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
220 a2.imag = a0.imag - tmp2; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
221 a3.real = a1.real - tmp3; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
222 a3.imag = a1.imag - tmp4; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
223 a0.real += tmp1; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
224 a0.imag += tmp2; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
225 a1.real += tmp3; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
226 a1.imag += tmp4; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
227 } while (0)
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
228
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
229 /* split-radix ifft butterfly, specialized for wr=wi */
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
230
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
231 #define BUTTERFLY_HALF(a0,a1,a2,a3,w) do { \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
232 tmp5 = (a2.real + a2.imag) * w; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
233 tmp6 = (a2.imag - a2.real) * w; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
234 tmp7 = (a3.real - a3.imag) * w; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
235 tmp8 = (a3.imag + a3.real) * w; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
236 tmp1 = tmp5 + tmp7; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
237 tmp2 = tmp6 + tmp8; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
238 tmp3 = tmp6 - tmp8; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
239 tmp4 = tmp7 - tmp5; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
240 a2.real = a0.real - tmp1; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
241 a2.imag = a0.imag - tmp2; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
242 a3.real = a1.real - tmp3; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
243 a3.imag = a1.imag - tmp4; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
244 a0.real += tmp1; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
245 a0.imag += tmp2; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
246 a1.real += tmp3; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
247 a1.imag += tmp4; \
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
248 } while (0)
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
249
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
250 static inline void ifft8 (complex_t * buf)
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
251 {
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
252 double tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp8;
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
253
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
254 ifft4 (buf);
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
255 ifft2 (buf + 4);
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
256 ifft2 (buf + 6);
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
257 BUTTERFLY_ZERO (buf[0], buf[2], buf[4], buf[6]);
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
258 BUTTERFLY_HALF (buf[1], buf[3], buf[5], buf[7], roots16[1]);
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
259 }
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
260
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
261 static void ifft_pass (complex_t * buf, sample_t * weight, int n)
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
262 {
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
263 complex_t * buf1;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
264 complex_t * buf2;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
265 complex_t * buf3;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
266 double tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp8;
8254
772d6d27fd66 warning patch by (Dominik Mierzejewski <dominik at rangers dot eu dot org>)
michael
parents: 4497
diff changeset
267 int i;
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
268
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
269 buf++;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
270 buf1 = buf + n;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
271 buf2 = buf + 2 * n;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
272 buf3 = buf + 3 * n;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
273
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
274 BUTTERFLY_ZERO (buf[-1], buf1[-1], buf2[-1], buf3[-1]);
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
275
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
276 i = n - 1;
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
277
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
278 do {
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
279 BUTTERFLY (buf[0], buf1[0], buf2[0], buf3[0], weight[n], weight[2*i]);
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
280 buf++;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
281 buf1++;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
282 buf2++;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
283 buf3++;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
284 weight++;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
285 } while (--i);
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
286 }
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
287
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
288 static void ifft16 (complex_t * buf)
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
289 {
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
290 ifft8 (buf);
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
291 ifft4 (buf + 8);
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
292 ifft4 (buf + 12);
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
293 ifft_pass (buf, roots16 - 4, 4);
18721
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
294 }
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
295
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
296 static void ifft32 (complex_t * buf)
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
297 {
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
298 ifft16 (buf);
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
299 ifft8 (buf + 16);
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
300 ifft8 (buf + 24);
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
301 ifft_pass (buf, roots32 - 8, 8);
18721
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
302 }
3579
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
303
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
304 static void ifft64_c (complex_t * buf)
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
305 {
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
306 ifft32 (buf);
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
307 ifft16 (buf + 32);
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
308 ifft16 (buf + 48);
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
309 ifft_pass (buf, roots64 - 16, 16);
18721
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
310 }
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
311
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
312 static void ifft128_c (complex_t * buf)
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
313 {
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
314 ifft32 (buf);
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
315 ifft16 (buf + 32);
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
316 ifft16 (buf + 48);
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
317 ifft_pass (buf, roots64 - 16, 16);
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
318
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
319 ifft32 (buf + 64);
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
320 ifft32 (buf + 96);
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
321 ifft_pass (buf, roots128 - 32, 32);
18721
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
322 }
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
323
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
324 void imdct_do_512 (sample_t * data, sample_t * delay, sample_t bias)
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
325 {
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
326 int i, k;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
327 sample_t t_r, t_i, a_r, a_i, b_r, b_i, w_1, w_2;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
328 const sample_t * window = a52_imdct_window;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
329 complex_t buf[128];
18721
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
330
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
331 for (i = 0; i < 128; i++) {
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
332 k = fftorder[i];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
333 t_r = pre1[i].real;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
334 t_i = pre1[i].imag;
18721
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
335
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
336 buf[i].real = t_i * data[255-k] + t_r * data[k];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
337 buf[i].imag = t_r * data[255-k] - t_i * data[k];
3579
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
338 }
18721
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
339
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
340 ifft128 (buf);
3579
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
341
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
342 /* Post IFFT complex multiply plus IFFT complex conjugate*/
3579
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
343 /* Window and convert to real valued signal */
18721
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
344 for (i = 0; i < 64; i++) {
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
345 /* y[n] = z[n] * (xcos1[n] + j * xsin1[n]) ; */
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
346 t_r = post1[i].real;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
347 t_i = post1[i].imag;
18721
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
348
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
349 a_r = t_r * buf[i].real + t_i * buf[i].imag;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
350 a_i = t_i * buf[i].real - t_r * buf[i].imag;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
351 b_r = t_i * buf[127-i].real + t_r * buf[127-i].imag;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
352 b_i = t_r * buf[127-i].real - t_i * buf[127-i].imag;
3579
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
353
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
354 w_1 = window[2*i];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
355 w_2 = window[255-2*i];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
356 data[2*i] = delay[2*i] * w_2 - a_r * w_1 + bias;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
357 data[255-2*i] = delay[2*i] * w_1 + a_r * w_2 + bias;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
358 delay[2*i] = a_i;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
359
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
360 w_1 = window[2*i+1];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
361 w_2 = window[254-2*i];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
362 data[2*i+1] = delay[2*i+1] * w_2 + b_r * w_1 + bias;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
363 data[254-2*i] = delay[2*i+1] * w_1 - b_r * w_2 + bias;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
364 delay[2*i+1] = b_i;
3579
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
365 }
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
366 }
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
367
9001
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
368 #ifdef HAVE_ALTIVEC
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
369
9122
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9001
diff changeset
370 #ifndef SYS_DARWIN
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9001
diff changeset
371 #include <altivec.h>
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9001
diff changeset
372 #endif
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9001
diff changeset
373
9001
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
374 // used to build registers permutation vectors (vcprm)
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
375 // the 's' are for words in the _s_econd vector
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
376 #define WORD_0 0x00,0x01,0x02,0x03
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
377 #define WORD_1 0x04,0x05,0x06,0x07
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
378 #define WORD_2 0x08,0x09,0x0a,0x0b
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
379 #define WORD_3 0x0c,0x0d,0x0e,0x0f
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
380 #define WORD_s0 0x10,0x11,0x12,0x13
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
381 #define WORD_s1 0x14,0x15,0x16,0x17
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
382 #define WORD_s2 0x18,0x19,0x1a,0x1b
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
383 #define WORD_s3 0x1c,0x1d,0x1e,0x1f
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
384
9122
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9001
diff changeset
385 #ifdef SYS_DARWIN
9001
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
386 #define vcprm(a,b,c,d) (const vector unsigned char)(WORD_ ## a, WORD_ ## b, WORD_ ## c, WORD_ ## d)
9122
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9001
diff changeset
387 #else
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9001
diff changeset
388 #define vcprm(a,b,c,d) (const vector unsigned char){WORD_ ## a, WORD_ ## b, WORD_ ## c, WORD_ ## d}
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9001
diff changeset
389 #endif
9001
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
390
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
391 // vcprmle is used to keep the same index as in the SSE version.
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
392 // it's the same as vcprm, with the index inversed
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
393 // ('le' is Little Endian)
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
394 #define vcprmle(a,b,c,d) vcprm(d,c,b,a)
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
395
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
396 // used to build inverse/identity vectors (vcii)
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
397 // n is _n_egative, p is _p_ositive
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
398 #define FLOAT_n -1.
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
399 #define FLOAT_p 1.
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
400
9122
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9001
diff changeset
401 #ifdef SYS_DARWIN
9001
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
402 #define vcii(a,b,c,d) (const vector float)(FLOAT_ ## a, FLOAT_ ## b, FLOAT_ ## c, FLOAT_ ## d)
9122
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9001
diff changeset
403 #else
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9001
diff changeset
404 #define vcii(a,b,c,d) (const vector float){FLOAT_ ## a, FLOAT_ ## b, FLOAT_ ## c, FLOAT_ ## d}
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9001
diff changeset
405 #endif
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9001
diff changeset
406
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9001
diff changeset
407 #ifdef SYS_DARWIN
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9001
diff changeset
408 #define FOUROF(a) (a)
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9001
diff changeset
409 #else
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9001
diff changeset
410 #define FOUROF(a) {a,a,a,a}
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9001
diff changeset
411 #endif
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9001
diff changeset
412
9001
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
413
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
414 void
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
415 imdct_do_512_altivec(sample_t data[],sample_t delay[], sample_t bias)
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
416 {
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
417 int i;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
418 int k;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
419 int p,q;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
420 int m;
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
421 long two_m;
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
422 long two_m_plus_one;
9001
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
423
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
424 sample_t tmp_b_i;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
425 sample_t tmp_b_r;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
426 sample_t tmp_a_i;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
427 sample_t tmp_a_r;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
428
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
429 sample_t *data_ptr;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
430 sample_t *delay_ptr;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
431 sample_t *window_ptr;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
432
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
433 /* 512 IMDCT with source and dest data in 'data' */
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
434
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
435 /* Pre IFFT complex multiply plus IFFT cmplx conjugate & reordering*/
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
436 for( i=0; i < 128; i++) {
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
437 /* z[i] = (X[256-2*i-1] + j * X[2*i]) * (xcos1[i] + j * xsin1[i]) ; */
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
438 int j= bit_reverse_512[i];
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
439 buf[i].real = (data[256-2*j-1] * xcos1[j]) - (data[2*j] * xsin1[j]);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
440 buf[i].imag = -1.0 * ((data[2*j] * xcos1[j]) + (data[256-2*j-1] * xsin1[j]));
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
441 }
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
442
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
443 /* 1. iteration */
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
444 for(i = 0; i < 128; i += 2) {
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
445 #if 0
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
446 tmp_a_r = buf[i].real;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
447 tmp_a_i = buf[i].imag;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
448 tmp_b_r = buf[i+1].real;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
449 tmp_b_i = buf[i+1].imag;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
450 buf[i].real = tmp_a_r + tmp_b_r;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
451 buf[i].imag = tmp_a_i + tmp_b_i;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
452 buf[i+1].real = tmp_a_r - tmp_b_r;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
453 buf[i+1].imag = tmp_a_i - tmp_b_i;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
454 #else
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
455 vector float temp, bufv;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
456
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
457 bufv = vec_ld(i << 3, (float*)buf);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
458 temp = vec_perm(bufv, bufv, vcprm(2,3,0,1));
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
459 bufv = vec_madd(bufv, vcii(p,p,n,n), temp);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
460 vec_st(bufv, i << 3, (float*)buf);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
461 #endif
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
462 }
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
463
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
464 /* 2. iteration */
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
465 // Note w[1]={{1,0}, {0,-1}}
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
466 for(i = 0; i < 128; i += 4) {
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
467 #if 0
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
468 tmp_a_r = buf[i].real;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
469 tmp_a_i = buf[i].imag;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
470 tmp_b_r = buf[i+2].real;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
471 tmp_b_i = buf[i+2].imag;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
472 buf[i].real = tmp_a_r + tmp_b_r;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
473 buf[i].imag = tmp_a_i + tmp_b_i;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
474 buf[i+2].real = tmp_a_r - tmp_b_r;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
475 buf[i+2].imag = tmp_a_i - tmp_b_i;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
476 tmp_a_r = buf[i+1].real;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
477 tmp_a_i = buf[i+1].imag;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
478 /* WARNING: im <-> re here ! */
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
479 tmp_b_r = buf[i+3].imag;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
480 tmp_b_i = buf[i+3].real;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
481 buf[i+1].real = tmp_a_r + tmp_b_r;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
482 buf[i+1].imag = tmp_a_i - tmp_b_i;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
483 buf[i+3].real = tmp_a_r - tmp_b_r;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
484 buf[i+3].imag = tmp_a_i + tmp_b_i;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
485 #else
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
486 vector float buf01, buf23, temp1, temp2;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
487
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
488 buf01 = vec_ld((i + 0) << 3, (float*)buf);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
489 buf23 = vec_ld((i + 2) << 3, (float*)buf);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
490 buf23 = vec_perm(buf23,buf23,vcprm(0,1,3,2));
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
491
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
492 temp1 = vec_madd(buf23, vcii(p,p,p,n), buf01);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
493 temp2 = vec_madd(buf23, vcii(n,n,n,p), buf01);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
494
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
495 vec_st(temp1, (i + 0) << 3, (float*)buf);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
496 vec_st(temp2, (i + 2) << 3, (float*)buf);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
497 #endif
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
498 }
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
499
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
500 /* 3. iteration */
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
501 for(i = 0; i < 128; i += 8) {
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
502 #if 0
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
503 tmp_a_r = buf[i].real;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
504 tmp_a_i = buf[i].imag;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
505 tmp_b_r = buf[i+4].real;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
506 tmp_b_i = buf[i+4].imag;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
507 buf[i].real = tmp_a_r + tmp_b_r;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
508 buf[i].imag = tmp_a_i + tmp_b_i;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
509 buf[i+4].real = tmp_a_r - tmp_b_r;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
510 buf[i+4].imag = tmp_a_i - tmp_b_i;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
511 tmp_a_r = buf[1+i].real;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
512 tmp_a_i = buf[1+i].imag;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
513 tmp_b_r = (buf[i+5].real + buf[i+5].imag) * w[2][1].real;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
514 tmp_b_i = (buf[i+5].imag - buf[i+5].real) * w[2][1].real;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
515 buf[1+i].real = tmp_a_r + tmp_b_r;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
516 buf[1+i].imag = tmp_a_i + tmp_b_i;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
517 buf[i+5].real = tmp_a_r - tmp_b_r;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
518 buf[i+5].imag = tmp_a_i - tmp_b_i;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
519 tmp_a_r = buf[i+2].real;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
520 tmp_a_i = buf[i+2].imag;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
521 /* WARNING re <-> im & sign */
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
522 tmp_b_r = buf[i+6].imag;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
523 tmp_b_i = - buf[i+6].real;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
524 buf[i+2].real = tmp_a_r + tmp_b_r;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
525 buf[i+2].imag = tmp_a_i + tmp_b_i;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
526 buf[i+6].real = tmp_a_r - tmp_b_r;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
527 buf[i+6].imag = tmp_a_i - tmp_b_i;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
528 tmp_a_r = buf[i+3].real;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
529 tmp_a_i = buf[i+3].imag;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
530 tmp_b_r = (buf[i+7].real - buf[i+7].imag) * w[2][3].imag;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
531 tmp_b_i = (buf[i+7].imag + buf[i+7].real) * w[2][3].imag;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
532 buf[i+3].real = tmp_a_r + tmp_b_r;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
533 buf[i+3].imag = tmp_a_i + tmp_b_i;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
534 buf[i+7].real = tmp_a_r - tmp_b_r;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
535 buf[i+7].imag = tmp_a_i - tmp_b_i;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
536 #else
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
537 vector float buf01, buf23, buf45, buf67;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
538
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
539 buf01 = vec_ld((i + 0) << 3, (float*)buf);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
540 buf23 = vec_ld((i + 2) << 3, (float*)buf);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
541
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
542 tmp_b_r = (buf[i+5].real + buf[i+5].imag) * w[2][1].real;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
543 tmp_b_i = (buf[i+5].imag - buf[i+5].real) * w[2][1].real;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
544 buf[i+5].real = tmp_b_r;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
545 buf[i+5].imag = tmp_b_i;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
546 tmp_b_r = (buf[i+7].real - buf[i+7].imag) * w[2][3].imag;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
547 tmp_b_i = (buf[i+7].imag + buf[i+7].real) * w[2][3].imag;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
548 buf[i+7].real = tmp_b_r;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
549 buf[i+7].imag = tmp_b_i;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
550
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
551 buf23 = vec_ld((i + 2) << 3, (float*)buf);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
552 buf45 = vec_ld((i + 4) << 3, (float*)buf);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
553 buf67 = vec_ld((i + 6) << 3, (float*)buf);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
554 buf67 = vec_perm(buf67, buf67, vcprm(1,0,2,3));
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
555
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
556 vec_st(vec_add(buf01, buf45), (i + 0) << 3, (float*)buf);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
557 vec_st(vec_madd(buf67, vcii(p,n,p,p), buf23), (i + 2) << 3, (float*)buf);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
558 vec_st(vec_sub(buf01, buf45), (i + 4) << 3, (float*)buf);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
559 vec_st(vec_nmsub(buf67, vcii(p,n,p,p), buf23), (i + 6) << 3, (float*)buf);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
560 #endif
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
561 }
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
562
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
563 /* 4-7. iterations */
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
564 for (m=3; m < 7; m++) {
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
565 two_m = (1 << m);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
566
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
567 two_m_plus_one = two_m<<1;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
568
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
569 for(i = 0; i < 128; i += two_m_plus_one) {
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
570 for(k = 0; k < two_m; k+=2) {
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
571 #if 0
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
572 int p = k + i;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
573 int q = p + two_m;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
574 tmp_a_r = buf[p].real;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
575 tmp_a_i = buf[p].imag;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
576 tmp_b_r =
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
577 buf[q].real * w[m][k].real -
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
578 buf[q].imag * w[m][k].imag;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
579 tmp_b_i =
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
580 buf[q].imag * w[m][k].real +
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
581 buf[q].real * w[m][k].imag;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
582 buf[p].real = tmp_a_r + tmp_b_r;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
583 buf[p].imag = tmp_a_i + tmp_b_i;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
584 buf[q].real = tmp_a_r - tmp_b_r;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
585 buf[q].imag = tmp_a_i - tmp_b_i;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
586
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
587 tmp_a_r = buf[(p + 1)].real;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
588 tmp_a_i = buf[(p + 1)].imag;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
589 tmp_b_r =
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
590 buf[(q + 1)].real * w[m][(k + 1)].real -
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
591 buf[(q + 1)].imag * w[m][(k + 1)].imag;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
592 tmp_b_i =
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
593 buf[(q + 1)].imag * w[m][(k + 1)].real +
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
594 buf[(q + 1)].real * w[m][(k + 1)].imag;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
595 buf[(p + 1)].real = tmp_a_r + tmp_b_r;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
596 buf[(p + 1)].imag = tmp_a_i + tmp_b_i;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
597 buf[(q + 1)].real = tmp_a_r - tmp_b_r;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
598 buf[(q + 1)].imag = tmp_a_i - tmp_b_i;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
599 #else
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
600 int p = k + i;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
601 int q = p + two_m;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
602 vector float vecp, vecq, vecw, temp1, temp2, temp3, temp4;
9122
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9001
diff changeset
603 const vector float vczero = (const vector float)FOUROF(0.);
9001
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
604 // first compute buf[q] and buf[q+1]
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
605 vecq = vec_ld(q << 3, (float*)buf);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
606 vecw = vec_ld(0, (float*)&(w[m][k]));
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
607 temp1 = vec_madd(vecq, vecw, vczero);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
608 temp2 = vec_perm(vecq, vecq, vcprm(1,0,3,2));
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
609 temp2 = vec_madd(temp2, vecw, vczero);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
610 temp3 = vec_perm(temp1, temp2, vcprm(0,s0,2,s2));
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
611 temp4 = vec_perm(temp1, temp2, vcprm(1,s1,3,s3));
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
612 vecq = vec_madd(temp4, vcii(n,p,n,p), temp3);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
613 // then butterfly with buf[p] and buf[p+1]
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
614 vecp = vec_ld(p << 3, (float*)buf);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
615
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
616 temp1 = vec_add(vecp, vecq);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
617 temp2 = vec_sub(vecp, vecq);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
618
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
619 vec_st(temp1, p << 3, (float*)buf);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
620 vec_st(temp2, q << 3, (float*)buf);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
621 #endif
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
622 }
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
623 }
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
624 }
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
625
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
626 /* Post IFFT complex multiply plus IFFT complex conjugate*/
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
627 for( i=0; i < 128; i+=4) {
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
628 /* y[n] = z[n] * (xcos1[n] + j * xsin1[n]) ; */
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
629 #if 0
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
630 tmp_a_r = buf[(i + 0)].real;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
631 tmp_a_i = -1.0 * buf[(i + 0)].imag;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
632 buf[(i + 0)].real =
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
633 (tmp_a_r * xcos1[(i + 0)]) - (tmp_a_i * xsin1[(i + 0)]);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
634 buf[(i + 0)].imag =
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
635 (tmp_a_r * xsin1[(i + 0)]) + (tmp_a_i * xcos1[(i + 0)]);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
636
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
637 tmp_a_r = buf[(i + 1)].real;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
638 tmp_a_i = -1.0 * buf[(i + 1)].imag;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
639 buf[(i + 1)].real =
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
640 (tmp_a_r * xcos1[(i + 1)]) - (tmp_a_i * xsin1[(i + 1)]);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
641 buf[(i + 1)].imag =
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
642 (tmp_a_r * xsin1[(i + 1)]) + (tmp_a_i * xcos1[(i + 1)]);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
643
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
644 tmp_a_r = buf[(i + 2)].real;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
645 tmp_a_i = -1.0 * buf[(i + 2)].imag;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
646 buf[(i + 2)].real =
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
647 (tmp_a_r * xcos1[(i + 2)]) - (tmp_a_i * xsin1[(i + 2)]);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
648 buf[(i + 2)].imag =
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
649 (tmp_a_r * xsin1[(i + 2)]) + (tmp_a_i * xcos1[(i + 2)]);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
650
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
651 tmp_a_r = buf[(i + 3)].real;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
652 tmp_a_i = -1.0 * buf[(i + 3)].imag;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
653 buf[(i + 3)].real =
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
654 (tmp_a_r * xcos1[(i + 3)]) - (tmp_a_i * xsin1[(i + 3)]);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
655 buf[(i + 3)].imag =
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
656 (tmp_a_r * xsin1[(i + 3)]) + (tmp_a_i * xcos1[(i + 3)]);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
657 #else
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
658 vector float bufv_0, bufv_2, cosv, sinv, temp1, temp2;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
659 vector float temp0022, temp1133, tempCS01;
9122
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9001
diff changeset
660 const vector float vczero = (const vector float)FOUROF(0.);
9001
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
661
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
662 bufv_0 = vec_ld((i + 0) << 3, (float*)buf);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
663 bufv_2 = vec_ld((i + 2) << 3, (float*)buf);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
664
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
665 cosv = vec_ld(i << 2, xcos1);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
666 sinv = vec_ld(i << 2, xsin1);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
667
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
668 temp0022 = vec_perm(bufv_0, bufv_0, vcprm(0,0,2,2));
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
669 temp1133 = vec_perm(bufv_0, bufv_0, vcprm(1,1,3,3));
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
670 tempCS01 = vec_perm(cosv, sinv, vcprm(0,s0,1,s1));
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
671 temp1 = vec_madd(temp0022, tempCS01, vczero);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
672 tempCS01 = vec_perm(cosv, sinv, vcprm(s0,0,s1,1));
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
673 temp2 = vec_madd(temp1133, tempCS01, vczero);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
674 bufv_0 = vec_madd(temp2, vcii(p,n,p,n), temp1);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
675
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
676 vec_st(bufv_0, (i + 0) << 3, (float*)buf);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
677
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
678 /* idem with bufv_2 and high-order cosv/sinv */
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
679
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
680 temp0022 = vec_perm(bufv_2, bufv_2, vcprm(0,0,2,2));
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
681 temp1133 = vec_perm(bufv_2, bufv_2, vcprm(1,1,3,3));
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
682 tempCS01 = vec_perm(cosv, sinv, vcprm(2,s2,3,s3));
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
683 temp1 = vec_madd(temp0022, tempCS01, vczero);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
684 tempCS01 = vec_perm(cosv, sinv, vcprm(s2,2,s3,3));
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
685 temp2 = vec_madd(temp1133, tempCS01, vczero);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
686 bufv_2 = vec_madd(temp2, vcii(p,n,p,n), temp1);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
687
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
688 vec_st(bufv_2, (i + 2) << 3, (float*)buf);
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
689
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
690 #endif
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
691 }
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
692
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
693 data_ptr = data;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
694 delay_ptr = delay;
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
695 window_ptr = a52_imdct_window;
9001
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
696
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
697 /* Window and convert to real valued signal */
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
698 for(i=0; i< 64; i++) {
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
699 *data_ptr++ = -buf[64+i].imag * *window_ptr++ + *delay_ptr++ + bias;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
700 *data_ptr++ = buf[64-i-1].real * *window_ptr++ + *delay_ptr++ + bias;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
701 }
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
702
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
703 for(i=0; i< 64; i++) {
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
704 *data_ptr++ = -buf[i].real * *window_ptr++ + *delay_ptr++ + bias;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
705 *data_ptr++ = buf[128-i-1].imag * *window_ptr++ + *delay_ptr++ + bias;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
706 }
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
707
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
708 /* The trailing edge of the window goes into the delay line */
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
709 delay_ptr = delay;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
710
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
711 for(i=0; i< 64; i++) {
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
712 *delay_ptr++ = -buf[64+i].real * *--window_ptr;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
713 *delay_ptr++ = buf[64-i-1].imag * *--window_ptr;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
714 }
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
715
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
716 for(i=0; i<64; i++) {
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
717 *delay_ptr++ = buf[i].imag * *--window_ptr;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
718 *delay_ptr++ = -buf[128-i-1].real * *--window_ptr;
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
719 }
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
720 }
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
721 #endif
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
722
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
723
4497
d3aedd7db02c Restore K7 support
nick
parents: 4247
diff changeset
724 // Stuff below this line is borrowed from libac3
d3aedd7db02c Restore K7 support
nick
parents: 4247
diff changeset
725 #include "srfftp.h"
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
726 #if defined(ARCH_X86) || defined(ARCH_X86_64)
4497
d3aedd7db02c Restore K7 support
nick
parents: 4247
diff changeset
727 #ifndef HAVE_3DNOW
d3aedd7db02c Restore K7 support
nick
parents: 4247
diff changeset
728 #define HAVE_3DNOW 1
d3aedd7db02c Restore K7 support
nick
parents: 4247
diff changeset
729 #endif
3884
0410677eda4a ported 3dnow(ex) optimizations from libac3
michael
parents: 3720
diff changeset
730 #include "srfftp_3dnow.h"
0410677eda4a ported 3dnow(ex) optimizations from libac3
michael
parents: 3720
diff changeset
731
8451
fb88ccbc5ccc compiler warning fixes
arpi
parents: 8254
diff changeset
732 const i_cmplx_t x_plus_minus_3dnow __attribute__ ((aligned (8))) = {{ 0x00000000UL, 0x80000000UL }};
fb88ccbc5ccc compiler warning fixes
arpi
parents: 8254
diff changeset
733 const i_cmplx_t x_minus_plus_3dnow __attribute__ ((aligned (8))) = {{ 0x80000000UL, 0x00000000UL }};
3884
0410677eda4a ported 3dnow(ex) optimizations from libac3
michael
parents: 3720
diff changeset
734 const complex_t HSQRT2_3DNOW __attribute__ ((aligned (8))) = { 0.707106781188, 0.707106781188 };
0410677eda4a ported 3dnow(ex) optimizations from libac3
michael
parents: 3720
diff changeset
735
4497
d3aedd7db02c Restore K7 support
nick
parents: 4247
diff changeset
736 #undef HAVE_3DNOWEX
d3aedd7db02c Restore K7 support
nick
parents: 4247
diff changeset
737 #include "imdct_3dnow.h"
d3aedd7db02c Restore K7 support
nick
parents: 4247
diff changeset
738 #define HAVE_3DNOWEX
d3aedd7db02c Restore K7 support
nick
parents: 4247
diff changeset
739 #include "imdct_3dnow.h"
3884
0410677eda4a ported 3dnow(ex) optimizations from libac3
michael
parents: 3720
diff changeset
740
3579
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
741 void
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
742 imdct_do_512_sse(sample_t data[],sample_t delay[], sample_t bias)
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
743 {
8254
772d6d27fd66 warning patch by (Dominik Mierzejewski <dominik at rangers dot eu dot org>)
michael
parents: 4497
diff changeset
744 /* int i,k;
772d6d27fd66 warning patch by (Dominik Mierzejewski <dominik at rangers dot eu dot org>)
michael
parents: 4497
diff changeset
745 int p,q;*/
3579
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
746 int m;
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
747 long two_m;
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
748 long two_m_plus_one;
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
749 long two_m_plus_one_shl3;
15617
130dd060f723 one bugfix and a few gcc4 bug workaorunds by (Gianluigi Tiesi: mplayer, netfarm it)
michael
parents: 14991
diff changeset
750 complex_t *buf_offset;
3579
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
751
8254
772d6d27fd66 warning patch by (Dominik Mierzejewski <dominik at rangers dot eu dot org>)
michael
parents: 4497
diff changeset
752 /* sample_t tmp_a_i;
3579
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
753 sample_t tmp_a_r;
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
754 sample_t tmp_b_i;
8254
772d6d27fd66 warning patch by (Dominik Mierzejewski <dominik at rangers dot eu dot org>)
michael
parents: 4497
diff changeset
755 sample_t tmp_b_r;*/
3579
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
756
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
757 sample_t *data_ptr;
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
758 sample_t *delay_ptr;
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
759 sample_t *window_ptr;
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
760
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
761 /* 512 IMDCT with source and dest data in 'data' */
3623
3f1c2c06d0d8 adding some comments
michael
parents: 3584
diff changeset
762 /* see the c version (dct_do_512()), its allmost identical, just in C */
3f1c2c06d0d8 adding some comments
michael
parents: 3584
diff changeset
763
3579
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
764 /* Pre IFFT complex multiply plus IFFT cmplx conjugate */
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
765 /* Bit reversed shuffling */
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
766 asm volatile(
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
767 "xor %%"REG_S", %%"REG_S" \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
768 "lea "MANGLE(bit_reverse_512)", %%"REG_a"\n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
769 "mov $1008, %%"REG_D" \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
770 "push %%"REG_BP" \n\t" //use ebp without telling gcc
19372
6334c14b38eb Replace asmalign.h hack by ASMALIGN cpp macros from config.h.
diego
parents: 18783
diff changeset
771 ASMALIGN(4)
3579
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
772 "1: \n\t"
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
773 "movlps (%0, %%"REG_S"), %%xmm0 \n\t" // XXXI
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
774 "movhps 8(%0, %%"REG_D"), %%xmm0 \n\t" // RXXI
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
775 "movlps 8(%0, %%"REG_S"), %%xmm1 \n\t" // XXXi
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
776 "movhps (%0, %%"REG_D"), %%xmm1 \n\t" // rXXi
3584
7c4046c04be3 removing unnecessary sse sin/cos LUT
michael
parents: 3581
diff changeset
777 "shufps $0x33, %%xmm1, %%xmm0 \n\t" // irIR
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
778 "movaps "MANGLE(sseSinCos1c)"(%%"REG_S"), %%xmm2\n\t"
3584
7c4046c04be3 removing unnecessary sse sin/cos LUT
michael
parents: 3581
diff changeset
779 "mulps %%xmm0, %%xmm2 \n\t"
7c4046c04be3 removing unnecessary sse sin/cos LUT
michael
parents: 3581
diff changeset
780 "shufps $0xB1, %%xmm0, %%xmm0 \n\t" // riRI
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
781 "mulps "MANGLE(sseSinCos1d)"(%%"REG_S"), %%xmm0\n\t"
3584
7c4046c04be3 removing unnecessary sse sin/cos LUT
michael
parents: 3581
diff changeset
782 "subps %%xmm0, %%xmm2 \n\t"
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
783 "movzb (%%"REG_a"), %%"REG_d" \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
784 "movzb 1(%%"REG_a"), %%"REG_BP" \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
785 "movlps %%xmm2, (%1, %%"REG_d", 8) \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
786 "movhps %%xmm2, (%1, %%"REG_BP", 8) \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
787 "add $16, %%"REG_S" \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
788 "add $2, %%"REG_a" \n\t" // avoid complex addressing for P4 crap
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
789 "sub $16, %%"REG_D" \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
790 "jnc 1b \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
791 "pop %%"REG_BP" \n\t"//no we didnt touch ebp *g*
16189
72764c0dad8a Fixes segfault on IA-32 machines caused by the ASM patch for AMD-64 for a52.
gpoirier
parents: 16173
diff changeset
792 :: "b" (data), "c" (buf)
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
793 : "%"REG_S, "%"REG_D, "%"REG_a, "%"REG_d
3579
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
794 );
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
795
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
796
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
797 /* FFT Merge */
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
798 /* unoptimized variant
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
799 for (m=1; m < 7; m++) {
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
800 if(m)
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
801 two_m = (1 << m);
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
802 else
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
803 two_m = 1;
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
804
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
805 two_m_plus_one = (1 << (m+1));
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
806
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
807 for(i = 0; i < 128; i += two_m_plus_one) {
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
808 for(k = 0; k < two_m; k++) {
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
809 p = k + i;
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
810 q = p + two_m;
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
811 tmp_a_r = buf[p].real;
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
812 tmp_a_i = buf[p].imag;
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
813 tmp_b_r = buf[q].real * w[m][k].real - buf[q].imag * w[m][k].imag;
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
814 tmp_b_i = buf[q].imag * w[m][k].real + buf[q].real * w[m][k].imag;
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
815 buf[p].real = tmp_a_r + tmp_b_r;
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
816 buf[p].imag = tmp_a_i + tmp_b_i;
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
817 buf[q].real = tmp_a_r - tmp_b_r;
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
818 buf[q].imag = tmp_a_i - tmp_b_i;
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
819 }
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
820 }
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
821 }
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
822 */
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
823
3623
3f1c2c06d0d8 adding some comments
michael
parents: 3584
diff changeset
824 /* 1. iteration */
3549
2e21accd86a8 cleanup
michael
parents: 3546
diff changeset
825 // Note w[0][0]={1,0}
3508
b5220cf63fc3 some SSE optimizations
michael
parents: 3394
diff changeset
826 asm volatile(
b5220cf63fc3 some SSE optimizations
michael
parents: 3394
diff changeset
827 "xorps %%xmm1, %%xmm1 \n\t"
b5220cf63fc3 some SSE optimizations
michael
parents: 3394
diff changeset
828 "xorps %%xmm2, %%xmm2 \n\t"
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
829 "mov %0, %%"REG_S" \n\t"
19372
6334c14b38eb Replace asmalign.h hack by ASMALIGN cpp macros from config.h.
diego
parents: 18783
diff changeset
830 ASMALIGN(4)
3508
b5220cf63fc3 some SSE optimizations
michael
parents: 3394
diff changeset
831 "1: \n\t"
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
832 "movlps (%%"REG_S"), %%xmm0\n\t" //buf[p]
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
833 "movlps 8(%%"REG_S"), %%xmm1\n\t" //buf[q]
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
834 "movhps (%%"REG_S"), %%xmm0\n\t" //buf[p]
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
835 "movhps 8(%%"REG_S"), %%xmm2\n\t" //buf[q]
3508
b5220cf63fc3 some SSE optimizations
michael
parents: 3394
diff changeset
836 "addps %%xmm1, %%xmm0 \n\t"
b5220cf63fc3 some SSE optimizations
michael
parents: 3394
diff changeset
837 "subps %%xmm2, %%xmm0 \n\t"
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
838 "movaps %%xmm0, (%%"REG_S")\n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
839 "add $16, %%"REG_S" \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
840 "cmp %1, %%"REG_S" \n\t"
3508
b5220cf63fc3 some SSE optimizations
michael
parents: 3394
diff changeset
841 " jb 1b \n\t"
b5220cf63fc3 some SSE optimizations
michael
parents: 3394
diff changeset
842 :: "g" (buf), "r" (buf + 128)
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
843 : "%"REG_S
3508
b5220cf63fc3 some SSE optimizations
michael
parents: 3394
diff changeset
844 );
3549
2e21accd86a8 cleanup
michael
parents: 3546
diff changeset
845
3623
3f1c2c06d0d8 adding some comments
michael
parents: 3584
diff changeset
846 /* 2. iteration */
3512
1f166e420b15 a bit more SSE optimizations
michael
parents: 3508
diff changeset
847 // Note w[1]={{1,0}, {0,-1}}
1f166e420b15 a bit more SSE optimizations
michael
parents: 3508
diff changeset
848 asm volatile(
4247
2dbd637ffe05 mangle for win32 in liba52 (includes dummy mangle.h pointing to the one in main)
atmos4
parents: 3908
diff changeset
849 "movaps "MANGLE(ps111_1)", %%xmm7\n\t" // 1,1,1,-1
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
850 "mov %0, %%"REG_S" \n\t"
19372
6334c14b38eb Replace asmalign.h hack by ASMALIGN cpp macros from config.h.
diego
parents: 18783
diff changeset
851 ASMALIGN(4)
3512
1f166e420b15 a bit more SSE optimizations
michael
parents: 3508
diff changeset
852 "1: \n\t"
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
853 "movaps 16(%%"REG_S"), %%xmm2 \n\t" //r2,i2,r3,i3
3512
1f166e420b15 a bit more SSE optimizations
michael
parents: 3508
diff changeset
854 "shufps $0xB4, %%xmm2, %%xmm2 \n\t" //r2,i2,i3,r3
1f166e420b15 a bit more SSE optimizations
michael
parents: 3508
diff changeset
855 "mulps %%xmm7, %%xmm2 \n\t" //r2,i2,i3,-r3
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
856 "movaps (%%"REG_S"), %%xmm0 \n\t" //r0,i0,r1,i1
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
857 "movaps (%%"REG_S"), %%xmm1 \n\t" //r0,i0,r1,i1
3512
1f166e420b15 a bit more SSE optimizations
michael
parents: 3508
diff changeset
858 "addps %%xmm2, %%xmm0 \n\t"
1f166e420b15 a bit more SSE optimizations
michael
parents: 3508
diff changeset
859 "subps %%xmm2, %%xmm1 \n\t"
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
860 "movaps %%xmm0, (%%"REG_S") \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
861 "movaps %%xmm1, 16(%%"REG_S") \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
862 "add $32, %%"REG_S" \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
863 "cmp %1, %%"REG_S" \n\t"
3512
1f166e420b15 a bit more SSE optimizations
michael
parents: 3508
diff changeset
864 " jb 1b \n\t"
1f166e420b15 a bit more SSE optimizations
michael
parents: 3508
diff changeset
865 :: "g" (buf), "r" (buf + 128)
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
866 : "%"REG_S
3512
1f166e420b15 a bit more SSE optimizations
michael
parents: 3508
diff changeset
867 );
3549
2e21accd86a8 cleanup
michael
parents: 3546
diff changeset
868
3623
3f1c2c06d0d8 adding some comments
michael
parents: 3584
diff changeset
869 /* 3. iteration */
3534
3483390a902b sse opt
michael
parents: 3529
diff changeset
870 /*
3483390a902b sse opt
michael
parents: 3529
diff changeset
871 Note sseW2+0={1,1,sqrt(2),sqrt(2))
3483390a902b sse opt
michael
parents: 3529
diff changeset
872 Note sseW2+16={0,0,sqrt(2),-sqrt(2))
3483390a902b sse opt
michael
parents: 3529
diff changeset
873 Note sseW2+32={0,0,-sqrt(2),-sqrt(2))
3483390a902b sse opt
michael
parents: 3529
diff changeset
874 Note sseW2+48={1,-1,sqrt(2),-sqrt(2))
3483390a902b sse opt
michael
parents: 3529
diff changeset
875 */
3483390a902b sse opt
michael
parents: 3529
diff changeset
876 asm volatile(
4247
2dbd637ffe05 mangle for win32 in liba52 (includes dummy mangle.h pointing to the one in main)
atmos4
parents: 3908
diff changeset
877 "movaps 48+"MANGLE(sseW2)", %%xmm6\n\t"
2dbd637ffe05 mangle for win32 in liba52 (includes dummy mangle.h pointing to the one in main)
atmos4
parents: 3908
diff changeset
878 "movaps 16+"MANGLE(sseW2)", %%xmm7\n\t"
3534
3483390a902b sse opt
michael
parents: 3529
diff changeset
879 "xorps %%xmm5, %%xmm5 \n\t"
3483390a902b sse opt
michael
parents: 3529
diff changeset
880 "xorps %%xmm2, %%xmm2 \n\t"
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
881 "mov %0, %%"REG_S" \n\t"
19372
6334c14b38eb Replace asmalign.h hack by ASMALIGN cpp macros from config.h.
diego
parents: 18783
diff changeset
882 ASMALIGN(4)
3534
3483390a902b sse opt
michael
parents: 3529
diff changeset
883 "1: \n\t"
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
884 "movaps 32(%%"REG_S"), %%xmm2 \n\t" //r4,i4,r5,i5
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
885 "movaps 48(%%"REG_S"), %%xmm3 \n\t" //r6,i6,r7,i7
4247
2dbd637ffe05 mangle for win32 in liba52 (includes dummy mangle.h pointing to the one in main)
atmos4
parents: 3908
diff changeset
886 "movaps "MANGLE(sseW2)", %%xmm4 \n\t" //r4,i4,r5,i5
2dbd637ffe05 mangle for win32 in liba52 (includes dummy mangle.h pointing to the one in main)
atmos4
parents: 3908
diff changeset
887 "movaps 32+"MANGLE(sseW2)", %%xmm5\n\t" //r6,i6,r7,i7
3537
d7e5a32643c9 C optimizations
michael
parents: 3534
diff changeset
888 "mulps %%xmm2, %%xmm4 \n\t"
d7e5a32643c9 C optimizations
michael
parents: 3534
diff changeset
889 "mulps %%xmm3, %%xmm5 \n\t"
3534
3483390a902b sse opt
michael
parents: 3529
diff changeset
890 "shufps $0xB1, %%xmm2, %%xmm2 \n\t" //i4,r4,i5,r5
3483390a902b sse opt
michael
parents: 3529
diff changeset
891 "shufps $0xB1, %%xmm3, %%xmm3 \n\t" //i6,r6,i7,r7
3537
d7e5a32643c9 C optimizations
michael
parents: 3534
diff changeset
892 "mulps %%xmm6, %%xmm3 \n\t"
3534
3483390a902b sse opt
michael
parents: 3529
diff changeset
893 "mulps %%xmm7, %%xmm2 \n\t"
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
894 "movaps (%%"REG_S"), %%xmm0 \n\t" //r0,i0,r1,i1
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
895 "movaps 16(%%"REG_S"), %%xmm1 \n\t" //r2,i2,r3,i3
3534
3483390a902b sse opt
michael
parents: 3529
diff changeset
896 "addps %%xmm4, %%xmm2 \n\t"
3483390a902b sse opt
michael
parents: 3529
diff changeset
897 "addps %%xmm5, %%xmm3 \n\t"
3483390a902b sse opt
michael
parents: 3529
diff changeset
898 "movaps %%xmm2, %%xmm4 \n\t"
3483390a902b sse opt
michael
parents: 3529
diff changeset
899 "movaps %%xmm3, %%xmm5 \n\t"
3483390a902b sse opt
michael
parents: 3529
diff changeset
900 "addps %%xmm0, %%xmm2 \n\t"
3483390a902b sse opt
michael
parents: 3529
diff changeset
901 "addps %%xmm1, %%xmm3 \n\t"
3483390a902b sse opt
michael
parents: 3529
diff changeset
902 "subps %%xmm4, %%xmm0 \n\t"
3483390a902b sse opt
michael
parents: 3529
diff changeset
903 "subps %%xmm5, %%xmm1 \n\t"
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
904 "movaps %%xmm2, (%%"REG_S") \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
905 "movaps %%xmm3, 16(%%"REG_S") \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
906 "movaps %%xmm0, 32(%%"REG_S") \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
907 "movaps %%xmm1, 48(%%"REG_S") \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
908 "add $64, %%"REG_S" \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
909 "cmp %1, %%"REG_S" \n\t"
3534
3483390a902b sse opt
michael
parents: 3529
diff changeset
910 " jb 1b \n\t"
3483390a902b sse opt
michael
parents: 3529
diff changeset
911 :: "g" (buf), "r" (buf + 128)
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
912 : "%"REG_S
3534
3483390a902b sse opt
michael
parents: 3529
diff changeset
913 );
3508
b5220cf63fc3 some SSE optimizations
michael
parents: 3394
diff changeset
914
3623
3f1c2c06d0d8 adding some comments
michael
parents: 3584
diff changeset
915 /* 4-7. iterations */
3546
4e772a3c6b62 sse opt
michael
parents: 3537
diff changeset
916 for (m=3; m < 7; m++) {
4e772a3c6b62 sse opt
michael
parents: 3537
diff changeset
917 two_m = (1 << m);
4e772a3c6b62 sse opt
michael
parents: 3537
diff changeset
918 two_m_plus_one = two_m<<1;
15617
130dd060f723 one bugfix and a few gcc4 bug workaorunds by (Gianluigi Tiesi: mplayer, netfarm it)
michael
parents: 14991
diff changeset
919 two_m_plus_one_shl3 = (two_m_plus_one<<3);
130dd060f723 one bugfix and a few gcc4 bug workaorunds by (Gianluigi Tiesi: mplayer, netfarm it)
michael
parents: 14991
diff changeset
920 buf_offset = buf+128;
3546
4e772a3c6b62 sse opt
michael
parents: 3537
diff changeset
921 asm volatile(
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
922 "mov %0, %%"REG_S" \n\t"
19372
6334c14b38eb Replace asmalign.h hack by ASMALIGN cpp macros from config.h.
diego
parents: 18783
diff changeset
923 ASMALIGN(4)
3546
4e772a3c6b62 sse opt
michael
parents: 3537
diff changeset
924 "1: \n\t"
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
925 "xor %%"REG_D", %%"REG_D" \n\t" // k
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
926 "lea (%%"REG_S", %3), %%"REG_d" \n\t"
3546
4e772a3c6b62 sse opt
michael
parents: 3537
diff changeset
927 "2: \n\t"
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
928 "movaps (%%"REG_d", %%"REG_D"), %%xmm1 \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
929 "movaps (%4, %%"REG_D", 2), %%xmm2 \n\t"
3546
4e772a3c6b62 sse opt
michael
parents: 3537
diff changeset
930 "mulps %%xmm1, %%xmm2 \n\t"
4e772a3c6b62 sse opt
michael
parents: 3537
diff changeset
931 "shufps $0xB1, %%xmm1, %%xmm1 \n\t"
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
932 "mulps 16(%4, %%"REG_D", 2), %%xmm1 \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
933 "movaps (%%"REG_S", %%"REG_D"), %%xmm0 \n\t"
3546
4e772a3c6b62 sse opt
michael
parents: 3537
diff changeset
934 "addps %%xmm2, %%xmm1 \n\t"
4e772a3c6b62 sse opt
michael
parents: 3537
diff changeset
935 "movaps %%xmm1, %%xmm2 \n\t"
4e772a3c6b62 sse opt
michael
parents: 3537
diff changeset
936 "addps %%xmm0, %%xmm1 \n\t"
4e772a3c6b62 sse opt
michael
parents: 3537
diff changeset
937 "subps %%xmm2, %%xmm0 \n\t"
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
938 "movaps %%xmm1, (%%"REG_S", %%"REG_D") \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
939 "movaps %%xmm0, (%%"REG_d", %%"REG_D") \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
940 "add $16, %%"REG_D" \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
941 "cmp %3, %%"REG_D" \n\t" //FIXME (opt) count against 0
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
942 "jb 2b \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
943 "add %2, %%"REG_S" \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
944 "cmp %1, %%"REG_S" \n\t"
3546
4e772a3c6b62 sse opt
michael
parents: 3537
diff changeset
945 " jb 1b \n\t"
15617
130dd060f723 one bugfix and a few gcc4 bug workaorunds by (Gianluigi Tiesi: mplayer, netfarm it)
michael
parents: 14991
diff changeset
946 :: "g" (buf), "m" (buf_offset), "m" (two_m_plus_one_shl3), "r" (two_m<<3),
3546
4e772a3c6b62 sse opt
michael
parents: 3537
diff changeset
947 "r" (sseW[m])
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
948 : "%"REG_S, "%"REG_D, "%"REG_d
3546
4e772a3c6b62 sse opt
michael
parents: 3537
diff changeset
949 );
4e772a3c6b62 sse opt
michael
parents: 3537
diff changeset
950 }
4e772a3c6b62 sse opt
michael
parents: 3537
diff changeset
951
3623
3f1c2c06d0d8 adding some comments
michael
parents: 3584
diff changeset
952 /* Post IFFT complex multiply plus IFFT complex conjugate*/
3581
8ddf654c4871 sse opt
michael
parents: 3579
diff changeset
953 asm volatile(
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
954 "mov $-1024, %%"REG_S" \n\t"
19372
6334c14b38eb Replace asmalign.h hack by ASMALIGN cpp macros from config.h.
diego
parents: 18783
diff changeset
955 ASMALIGN(4)
3581
8ddf654c4871 sse opt
michael
parents: 3579
diff changeset
956 "1: \n\t"
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
957 "movaps (%0, %%"REG_S"), %%xmm0 \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
958 "movaps (%0, %%"REG_S"), %%xmm1 \n\t"
3581
8ddf654c4871 sse opt
michael
parents: 3579
diff changeset
959 "shufps $0xB1, %%xmm0, %%xmm0 \n\t"
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
960 "mulps 1024+"MANGLE(sseSinCos1c)"(%%"REG_S"), %%xmm1\n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
961 "mulps 1024+"MANGLE(sseSinCos1d)"(%%"REG_S"), %%xmm0\n\t"
3581
8ddf654c4871 sse opt
michael
parents: 3579
diff changeset
962 "addps %%xmm1, %%xmm0 \n\t"
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
963 "movaps %%xmm0, (%0, %%"REG_S") \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
964 "add $16, %%"REG_S" \n\t"
3581
8ddf654c4871 sse opt
michael
parents: 3579
diff changeset
965 " jnz 1b \n\t"
8ddf654c4871 sse opt
michael
parents: 3579
diff changeset
966 :: "r" (buf+128)
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
967 : "%"REG_S
3581
8ddf654c4871 sse opt
michael
parents: 3579
diff changeset
968 );
8ddf654c4871 sse opt
michael
parents: 3579
diff changeset
969
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
970
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
971 data_ptr = data;
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
972 delay_ptr = delay;
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
973 window_ptr = a52_imdct_window;
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
974
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
975 /* Window and convert to real valued signal */
3552
9ff2e3801027 sse opt
michael
parents: 3549
diff changeset
976 asm volatile(
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
977 "xor %%"REG_D", %%"REG_D" \n\t" // 0
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
978 "xor %%"REG_S", %%"REG_S" \n\t" // 0
3552
9ff2e3801027 sse opt
michael
parents: 3549
diff changeset
979 "movss %3, %%xmm2 \n\t" // bias
9ff2e3801027 sse opt
michael
parents: 3549
diff changeset
980 "shufps $0x00, %%xmm2, %%xmm2 \n\t" // bias, bias, ...
19372
6334c14b38eb Replace asmalign.h hack by ASMALIGN cpp macros from config.h.
diego
parents: 18783
diff changeset
981 ASMALIGN(4)
3552
9ff2e3801027 sse opt
michael
parents: 3549
diff changeset
982 "1: \n\t"
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
983 "movlps (%0, %%"REG_S"), %%xmm0 \n\t" // ? ? A ?
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
984 "movlps 8(%0, %%"REG_S"), %%xmm1 \n\t" // ? ? C ?
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
985 "movhps -16(%0, %%"REG_D"), %%xmm1 \n\t" // ? D C ?
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
986 "movhps -8(%0, %%"REG_D"), %%xmm0 \n\t" // ? B A ?
3552
9ff2e3801027 sse opt
michael
parents: 3549
diff changeset
987 "shufps $0x99, %%xmm1, %%xmm0 \n\t" // D C B A
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
988 "mulps "MANGLE(sseWindow)"(%%"REG_S"), %%xmm0\n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
989 "addps (%2, %%"REG_S"), %%xmm0 \n\t"
3552
9ff2e3801027 sse opt
michael
parents: 3549
diff changeset
990 "addps %%xmm2, %%xmm0 \n\t"
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
991 "movaps %%xmm0, (%1, %%"REG_S") \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
992 "add $16, %%"REG_S" \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
993 "sub $16, %%"REG_D" \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
994 "cmp $512, %%"REG_S" \n\t"
3552
9ff2e3801027 sse opt
michael
parents: 3549
diff changeset
995 " jb 1b \n\t"
9ff2e3801027 sse opt
michael
parents: 3549
diff changeset
996 :: "r" (buf+64), "r" (data_ptr), "r" (delay_ptr), "m" (bias)
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
997 : "%"REG_S, "%"REG_D
3552
9ff2e3801027 sse opt
michael
parents: 3549
diff changeset
998 );
9ff2e3801027 sse opt
michael
parents: 3549
diff changeset
999 data_ptr+=128;
9ff2e3801027 sse opt
michael
parents: 3549
diff changeset
1000 delay_ptr+=128;
3553
a501627fc6db sse opt
michael
parents: 3552
diff changeset
1001 // window_ptr+=128;
3579
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
1002
3552
9ff2e3801027 sse opt
michael
parents: 3549
diff changeset
1003 asm volatile(
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1004 "mov $1024, %%"REG_D" \n\t" // 512
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1005 "xor %%"REG_S", %%"REG_S" \n\t" // 0
3552
9ff2e3801027 sse opt
michael
parents: 3549
diff changeset
1006 "movss %3, %%xmm2 \n\t" // bias
9ff2e3801027 sse opt
michael
parents: 3549
diff changeset
1007 "shufps $0x00, %%xmm2, %%xmm2 \n\t" // bias, bias, ...
19372
6334c14b38eb Replace asmalign.h hack by ASMALIGN cpp macros from config.h.
diego
parents: 18783
diff changeset
1008 ASMALIGN(4)
3552
9ff2e3801027 sse opt
michael
parents: 3549
diff changeset
1009 "1: \n\t"
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1010 "movlps (%0, %%"REG_S"), %%xmm0 \n\t" // ? ? ? A
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1011 "movlps 8(%0, %%"REG_S"), %%xmm1 \n\t" // ? ? ? C
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1012 "movhps -16(%0, %%"REG_D"), %%xmm1 \n\t" // D ? ? C
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1013 "movhps -8(%0, %%"REG_D"), %%xmm0 \n\t" // B ? ? A
3552
9ff2e3801027 sse opt
michael
parents: 3549
diff changeset
1014 "shufps $0xCC, %%xmm1, %%xmm0 \n\t" // D C B A
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1015 "mulps 512+"MANGLE(sseWindow)"(%%"REG_S"), %%xmm0\n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1016 "addps (%2, %%"REG_S"), %%xmm0 \n\t"
3552
9ff2e3801027 sse opt
michael
parents: 3549
diff changeset
1017 "addps %%xmm2, %%xmm0 \n\t"
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1018 "movaps %%xmm0, (%1, %%"REG_S") \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1019 "add $16, %%"REG_S" \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1020 "sub $16, %%"REG_D" \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1021 "cmp $512, %%"REG_S" \n\t"
3552
9ff2e3801027 sse opt
michael
parents: 3549
diff changeset
1022 " jb 1b \n\t"
9ff2e3801027 sse opt
michael
parents: 3549
diff changeset
1023 :: "r" (buf), "r" (data_ptr), "r" (delay_ptr), "m" (bias)
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1024 : "%"REG_S, "%"REG_D
3552
9ff2e3801027 sse opt
michael
parents: 3549
diff changeset
1025 );
9ff2e3801027 sse opt
michael
parents: 3549
diff changeset
1026 data_ptr+=128;
3553
a501627fc6db sse opt
michael
parents: 3552
diff changeset
1027 // window_ptr+=128;
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
1028
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
1029 /* The trailing edge of the window goes into the delay line */
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
1030 delay_ptr = delay;
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
1031
3553
a501627fc6db sse opt
michael
parents: 3552
diff changeset
1032 asm volatile(
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1033 "xor %%"REG_D", %%"REG_D" \n\t" // 0
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1034 "xor %%"REG_S", %%"REG_S" \n\t" // 0
19372
6334c14b38eb Replace asmalign.h hack by ASMALIGN cpp macros from config.h.
diego
parents: 18783
diff changeset
1035 ASMALIGN(4)
3553
a501627fc6db sse opt
michael
parents: 3552
diff changeset
1036 "1: \n\t"
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1037 "movlps (%0, %%"REG_S"), %%xmm0 \n\t" // ? ? ? A
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1038 "movlps 8(%0, %%"REG_S"), %%xmm1 \n\t" // ? ? ? C
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1039 "movhps -16(%0, %%"REG_D"), %%xmm1 \n\t" // D ? ? C
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1040 "movhps -8(%0, %%"REG_D"), %%xmm0 \n\t" // B ? ? A
3553
a501627fc6db sse opt
michael
parents: 3552
diff changeset
1041 "shufps $0xCC, %%xmm1, %%xmm0 \n\t" // D C B A
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1042 "mulps 1024+"MANGLE(sseWindow)"(%%"REG_S"), %%xmm0\n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1043 "movaps %%xmm0, (%1, %%"REG_S") \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1044 "add $16, %%"REG_S" \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1045 "sub $16, %%"REG_D" \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1046 "cmp $512, %%"REG_S" \n\t"
3553
a501627fc6db sse opt
michael
parents: 3552
diff changeset
1047 " jb 1b \n\t"
a501627fc6db sse opt
michael
parents: 3552
diff changeset
1048 :: "r" (buf+64), "r" (delay_ptr)
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1049 : "%"REG_S, "%"REG_D
3553
a501627fc6db sse opt
michael
parents: 3552
diff changeset
1050 );
a501627fc6db sse opt
michael
parents: 3552
diff changeset
1051 delay_ptr+=128;
a501627fc6db sse opt
michael
parents: 3552
diff changeset
1052 // window_ptr-=128;
3579
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
1053
3553
a501627fc6db sse opt
michael
parents: 3552
diff changeset
1054 asm volatile(
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1055 "mov $1024, %%"REG_D" \n\t" // 1024
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1056 "xor %%"REG_S", %%"REG_S" \n\t" // 0
19372
6334c14b38eb Replace asmalign.h hack by ASMALIGN cpp macros from config.h.
diego
parents: 18783
diff changeset
1057 ASMALIGN(4)
3553
a501627fc6db sse opt
michael
parents: 3552
diff changeset
1058 "1: \n\t"
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1059 "movlps (%0, %%"REG_S"), %%xmm0 \n\t" // ? ? A ?
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1060 "movlps 8(%0, %%"REG_S"), %%xmm1 \n\t" // ? ? C ?
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1061 "movhps -16(%0, %%"REG_D"), %%xmm1 \n\t" // ? D C ?
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1062 "movhps -8(%0, %%"REG_D"), %%xmm0 \n\t" // ? B A ?
3553
a501627fc6db sse opt
michael
parents: 3552
diff changeset
1063 "shufps $0x99, %%xmm1, %%xmm0 \n\t" // D C B A
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1064 "mulps 1536+"MANGLE(sseWindow)"(%%"REG_S"), %%xmm0\n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1065 "movaps %%xmm0, (%1, %%"REG_S") \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1066 "add $16, %%"REG_S" \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1067 "sub $16, %%"REG_D" \n\t"
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1068 "cmp $512, %%"REG_S" \n\t"
3553
a501627fc6db sse opt
michael
parents: 3552
diff changeset
1069 " jb 1b \n\t"
a501627fc6db sse opt
michael
parents: 3552
diff changeset
1070 :: "r" (buf), "r" (delay_ptr)
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1071 : "%"REG_S, "%"REG_D
3553
a501627fc6db sse opt
michael
parents: 3552
diff changeset
1072 );
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
1073 }
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1074 #endif // ARCH_X86 || ARCH_X86_64
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
1075
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1076 void a52_imdct_256(sample_t * data, sample_t * delay, sample_t bias)
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
1077 {
18721
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
1078 int i, k;
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1079 sample_t t_r, t_i, a_r, a_i, b_r, b_i, c_r, c_i, d_r, d_i, w_1, w_2;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1080 const sample_t * window = a52_imdct_window;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1081 complex_t buf1[64], buf2[64];
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
1082
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
1083 /* Pre IFFT complex multiply plus IFFT cmplx conjugate */
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1084 for (i = 0; i < 64; i++) {
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1085 k = fftorder[i];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1086 t_r = pre2[i].real;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1087 t_i = pre2[i].imag;
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
1088
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1089 buf1[i].real = t_i * data[254-k] + t_r * data[k];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1090 buf1[i].imag = t_r * data[254-k] - t_i * data[k];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1091
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1092 buf2[i].real = t_i * data[255-k] + t_r * data[k+1];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1093 buf2[i].imag = t_r * data[255-k] - t_i * data[k+1];
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
1094 }
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
1095
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1096 ifft64 (buf1);
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1097 ifft64 (buf2);
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1098
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1099 /* Post IFFT complex multiply */
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1100 /* Window and convert to real valued signal */
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1101 for (i = 0; i < 32; i++) {
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1102 /* y1[n] = z1[n] * (xcos2[n] + j * xs in2[n]) ; */
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1103 t_r = post2[i].real;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1104 t_i = post2[i].imag;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1105
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1106 a_r = t_r * buf1[i].real + t_i * buf1[i].imag;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1107 a_i = t_i * buf1[i].real - t_r * buf1[i].imag;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1108 b_r = t_i * buf1[63-i].real + t_r * buf1[63-i].imag;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1109 b_i = t_r * buf1[63-i].real - t_i * buf1[63-i].imag;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1110
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1111 c_r = t_r * buf2[i].real + t_i * buf2[i].imag;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1112 c_i = t_i * buf2[i].real - t_r * buf2[i].imag;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1113 d_r = t_i * buf2[63-i].real + t_r * buf2[63-i].imag;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1114 d_i = t_r * buf2[63-i].real - t_i * buf2[63-i].imag;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1115
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1116 w_1 = window[2*i];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1117 w_2 = window[255-2*i];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1118 data[2*i] = delay[2*i] * w_2 - a_r * w_1 + bias;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1119 data[255-2*i] = delay[2*i] * w_1 + a_r * w_2 + bias;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1120 delay[2*i] = c_i;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1121
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1122 w_1 = window[128+2*i];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1123 w_2 = window[127-2*i];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1124 data[128+2*i] = delay[127-2*i] * w_2 + a_i * w_1 + bias;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1125 data[127-2*i] = delay[127-2*i] * w_1 - a_i * w_2 + bias;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1126 delay[127-2*i] = c_r;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1127
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1128 w_1 = window[2*i+1];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1129 w_2 = window[254-2*i];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1130 data[2*i+1] = delay[2*i+1] * w_2 - b_i * w_1 + bias;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1131 data[254-2*i] = delay[2*i+1] * w_1 + b_i * w_2 + bias;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1132 delay[2*i+1] = d_r;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1133
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1134 w_1 = window[129+2*i];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1135 w_2 = window[126-2*i];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1136 data[129+2*i] = delay[126-2*i] * w_2 + b_r * w_1 + bias;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1137 data[126-2*i] = delay[126-2*i] * w_1 - b_r * w_2 + bias;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1138 delay[126-2*i] = d_i;
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
1139 }
18721
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
1140 }
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
1141
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1142 static double besselI0 (double x)
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1143 {
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1144 double bessel = 1;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1145 int i = 100;
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
1146
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1147 do
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1148 bessel = bessel * x / (i * i) + 1;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1149 while (--i);
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1150 return bessel;
18721
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
1151 }
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
1152
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1153 void a52_imdct_init (uint32_t mm_accel)
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1154 {
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1155 int i, j, k;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1156 double sum;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1157
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1158 /* compute imdct window - kaiser-bessel derived window, alpha = 5.0 */
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1159 sum = 0;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1160 for (i = 0; i < 256; i++) {
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1161 sum += besselI0 (i * (256 - i) * (5 * M_PI / 256) * (5 * M_PI / 256));
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1162 a52_imdct_window[i] = sum;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1163 }
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1164 sum++;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1165 for (i = 0; i < 256; i++)
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1166 a52_imdct_window[i] = sqrt (a52_imdct_window[i] / sum);
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1167
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1168 for (i = 0; i < 3; i++)
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1169 roots16[i] = cos ((M_PI / 8) * (i + 1));
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
1170
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1171 for (i = 0; i < 7; i++)
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1172 roots32[i] = cos ((M_PI / 16) * (i + 1));
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1173
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1174 for (i = 0; i < 15; i++)
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1175 roots64[i] = cos ((M_PI / 32) * (i + 1));
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1176
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1177 for (i = 0; i < 31; i++)
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1178 roots128[i] = cos ((M_PI / 64) * (i + 1));
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1179
18721
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
1180 for (i = 0; i < 64; i++) {
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1181 k = fftorder[i] / 2 + 64;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1182 pre1[i].real = cos ((M_PI / 256) * (k - 0.25));
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1183 pre1[i].imag = sin ((M_PI / 256) * (k - 0.25));
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1184 }
18721
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
1185
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1186 for (i = 64; i < 128; i++) {
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1187 k = fftorder[i] / 2 + 64;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1188 pre1[i].real = -cos ((M_PI / 256) * (k - 0.25));
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1189 pre1[i].imag = -sin ((M_PI / 256) * (k - 0.25));
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
1190 }
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
1191
18721
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
1192 for (i = 0; i < 64; i++) {
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1193 post1[i].real = cos ((M_PI / 256) * (i + 0.5));
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1194 post1[i].imag = sin ((M_PI / 256) * (i + 0.5));
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
1195 }
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
1196
18721
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
1197 for (i = 0; i < 64; i++) {
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1198 k = fftorder[i] / 4;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1199 pre2[i].real = cos ((M_PI / 128) * (k - 0.25));
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1200 pre2[i].imag = sin ((M_PI / 128) * (k - 0.25));
18721
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
1201 }
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
1202
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1203 for (i = 0; i < 32; i++) {
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1204 post2[i].real = cos ((M_PI / 128) * (i + 0.5));
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1205 post2[i].imag = sin ((M_PI / 128) * (i + 0.5));
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1206 }
18721
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
1207 for (i = 0; i < 128; i++) {
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
1208 xcos1[i] = -cos ((M_PI / 2048) * (8 * i + 1));
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
1209 xsin1[i] = -sin ((M_PI / 2048) * (8 * i + 1));
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
1210 }
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
1211 for (i = 0; i < 7; i++) {
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
1212 j = 1 << i;
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
1213 for (k = 0; k < j; k++) {
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
1214 w[i][k].real = cos (-M_PI * k / j);
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
1215 w[i][k].imag = sin (-M_PI * k / j);
3394
35b18ed357c2 imported from liba52 CVS
arpi
parents:
diff changeset
1216 }
18721
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
1217 }
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1218 #if defined(ARCH_X86) || defined(ARCH_X86_64)
3527
5a88b21cfe8a sse opt
michael
parents: 3512
diff changeset
1219 for (i = 0; i < 128; i++) {
3581
8ddf654c4871 sse opt
michael
parents: 3579
diff changeset
1220 sseSinCos1c[2*i+0]= xcos1[i];
8ddf654c4871 sse opt
michael
parents: 3579
diff changeset
1221 sseSinCos1c[2*i+1]= -xcos1[i];
8ddf654c4871 sse opt
michael
parents: 3579
diff changeset
1222 sseSinCos1d[2*i+0]= xsin1[i];
8ddf654c4871 sse opt
michael
parents: 3579
diff changeset
1223 sseSinCos1d[2*i+1]= xsin1[i];
3527
5a88b21cfe8a sse opt
michael
parents: 3512
diff changeset
1224 }
3534
3483390a902b sse opt
michael
parents: 3529
diff changeset
1225 for (i = 1; i < 7; i++) {
3483390a902b sse opt
michael
parents: 3529
diff changeset
1226 j = 1 << i;
3483390a902b sse opt
michael
parents: 3529
diff changeset
1227 for (k = 0; k < j; k+=2) {
3483390a902b sse opt
michael
parents: 3529
diff changeset
1228
3483390a902b sse opt
michael
parents: 3529
diff changeset
1229 sseW[i][4*k + 0] = w[i][k+0].real;
3483390a902b sse opt
michael
parents: 3529
diff changeset
1230 sseW[i][4*k + 1] = w[i][k+0].real;
3483390a902b sse opt
michael
parents: 3529
diff changeset
1231 sseW[i][4*k + 2] = w[i][k+1].real;
3483390a902b sse opt
michael
parents: 3529
diff changeset
1232 sseW[i][4*k + 3] = w[i][k+1].real;
3483390a902b sse opt
michael
parents: 3529
diff changeset
1233
3483390a902b sse opt
michael
parents: 3529
diff changeset
1234 sseW[i][4*k + 4] = -w[i][k+0].imag;
3483390a902b sse opt
michael
parents: 3529
diff changeset
1235 sseW[i][4*k + 5] = w[i][k+0].imag;
3483390a902b sse opt
michael
parents: 3529
diff changeset
1236 sseW[i][4*k + 6] = -w[i][k+1].imag;
3483390a902b sse opt
michael
parents: 3529
diff changeset
1237 sseW[i][4*k + 7] = w[i][k+1].imag;
3483390a902b sse opt
michael
parents: 3529
diff changeset
1238
3483390a902b sse opt
michael
parents: 3529
diff changeset
1239 //we multiply more or less uninitalized numbers so we need to use exactly 0.0
3483390a902b sse opt
michael
parents: 3529
diff changeset
1240 if(k==0)
3483390a902b sse opt
michael
parents: 3529
diff changeset
1241 {
3483390a902b sse opt
michael
parents: 3529
diff changeset
1242 // sseW[i][4*k + 0]= sseW[i][4*k + 1]= 1.0;
3483390a902b sse opt
michael
parents: 3529
diff changeset
1243 sseW[i][4*k + 4]= sseW[i][4*k + 5]= 0.0;
3483390a902b sse opt
michael
parents: 3529
diff changeset
1244 }
3483390a902b sse opt
michael
parents: 3529
diff changeset
1245
3483390a902b sse opt
michael
parents: 3529
diff changeset
1246 if(2*k == j)
3483390a902b sse opt
michael
parents: 3529
diff changeset
1247 {
3483390a902b sse opt
michael
parents: 3529
diff changeset
1248 sseW[i][4*k + 0]= sseW[i][4*k + 1]= 0.0;
3483390a902b sse opt
michael
parents: 3529
diff changeset
1249 // sseW[i][4*k + 4]= -(sseW[i][4*k + 5]= -1.0);
3483390a902b sse opt
michael
parents: 3529
diff changeset
1250 }
3483390a902b sse opt
michael
parents: 3529
diff changeset
1251 }
3483390a902b sse opt
michael
parents: 3529
diff changeset
1252 }
3552
9ff2e3801027 sse opt
michael
parents: 3549
diff changeset
1253
9ff2e3801027 sse opt
michael
parents: 3549
diff changeset
1254 for(i=0; i<128; i++)
9ff2e3801027 sse opt
michael
parents: 3549
diff changeset
1255 {
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1256 sseWindow[2*i+0]= -a52_imdct_window[2*i+0];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1257 sseWindow[2*i+1]= a52_imdct_window[2*i+1];
3552
9ff2e3801027 sse opt
michael
parents: 3549
diff changeset
1258 }
3553
a501627fc6db sse opt
michael
parents: 3552
diff changeset
1259
a501627fc6db sse opt
michael
parents: 3552
diff changeset
1260 for(i=0; i<64; i++)
a501627fc6db sse opt
michael
parents: 3552
diff changeset
1261 {
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1262 sseWindow[256 + 2*i+0]= -a52_imdct_window[254 - 2*i+1];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1263 sseWindow[256 + 2*i+1]= a52_imdct_window[254 - 2*i+0];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1264 sseWindow[384 + 2*i+0]= a52_imdct_window[126 - 2*i+1];
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1265 sseWindow[384 + 2*i+1]= -a52_imdct_window[126 - 2*i+0];
3553
a501627fc6db sse opt
michael
parents: 3552
diff changeset
1266 }
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1267 #endif
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1268 a52_imdct_512 = imdct_do_512;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1269 ifft128 = ifft128_c;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1270 ifft64 = ifft64_c;
3579
831860fada69 runtime cpu detection for the idct
michael
parents: 3553
diff changeset
1271
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1272 #if defined(ARCH_X86) || defined(ARCH_X86_64)
4497
d3aedd7db02c Restore K7 support
nick
parents: 4247
diff changeset
1273 if(mm_accel & MM_ACCEL_X86_SSE)
d3aedd7db02c Restore K7 support
nick
parents: 4247
diff changeset
1274 {
d3aedd7db02c Restore K7 support
nick
parents: 4247
diff changeset
1275 fprintf (stderr, "Using SSE optimized IMDCT transform\n");
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1276 a52_imdct_512 = imdct_do_512_sse;
18721
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
1277 }
4497
d3aedd7db02c Restore K7 support
nick
parents: 4247
diff changeset
1278 else
d3aedd7db02c Restore K7 support
nick
parents: 4247
diff changeset
1279 if(mm_accel & MM_ACCEL_X86_3DNOWEXT)
d3aedd7db02c Restore K7 support
nick
parents: 4247
diff changeset
1280 {
d3aedd7db02c Restore K7 support
nick
parents: 4247
diff changeset
1281 fprintf (stderr, "Using 3DNowEx optimized IMDCT transform\n");
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1282 a52_imdct_512 = imdct_do_512_3dnowex;
4497
d3aedd7db02c Restore K7 support
nick
parents: 4247
diff changeset
1283 }
d3aedd7db02c Restore K7 support
nick
parents: 4247
diff changeset
1284 else
d3aedd7db02c Restore K7 support
nick
parents: 4247
diff changeset
1285 if(mm_accel & MM_ACCEL_X86_3DNOW)
d3aedd7db02c Restore K7 support
nick
parents: 4247
diff changeset
1286 {
d3aedd7db02c Restore K7 support
nick
parents: 4247
diff changeset
1287 fprintf (stderr, "Using 3DNow optimized IMDCT transform\n");
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1288 a52_imdct_512 = imdct_do_512_3dnow;
4497
d3aedd7db02c Restore K7 support
nick
parents: 4247
diff changeset
1289 }
d3aedd7db02c Restore K7 support
nick
parents: 4247
diff changeset
1290 else
16173
d6219ce521e9 liba52 asm optimizations ported to amd64
aurel
parents: 15617
diff changeset
1291 #endif // ARCH_X86 || ARCH_X86_64
9001
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
1292 #ifdef HAVE_ALTIVEC
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
1293 if (mm_accel & MM_ACCEL_PPC_ALTIVEC)
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
1294 {
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
1295 fprintf(stderr, "Using AltiVec optimized IMDCT transform\n");
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1296 a52_imdct_512 = imdct_do_512_altivec;
9001
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
1297 }
18721
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
1298 else
9001
01a9cf43074c An AltiVec-enhanced IMDCT for liba52 (liba52/imdct.c)
arpi
parents: 8451
diff changeset
1299 #endif
3884
0410677eda4a ported 3dnow(ex) optimizations from libac3
michael
parents: 3720
diff changeset
1300
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1301 #ifdef LIBA52_DJBFFT
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1302 if (mm_accel & MM_ACCEL_DJBFFT) {
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1303 fprintf (stderr, "Using djbfft for IMDCT transform\n");
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1304 ifft128 = (void (*) (complex_t *)) fftc4_un128;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1305 ifft64 = (void (*) (complex_t *)) fftc4_un64;
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1306 } else
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1307 #endif
18721
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
1308 {
18720
4bad7f00556e sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18104
diff changeset
1309 fprintf (stderr, "No accelerated IMDCT transform found\n");
18721
722ac20fac5f sync with liba52 0.7.4, patch by Emanuele Giaquinta >emanuele.giaquinta ! gmail * com<
rathann
parents: 18720
diff changeset
1310 }
3884
0410677eda4a ported 3dnow(ex) optimizations from libac3
michael
parents: 3720
diff changeset
1311 }