annotate mp3lib/dct64_altivec.c @ 14059:384ae6f896cf

Explain how (not) to handle execute permissions and binary files.
author diego
date Mon, 29 Nov 2004 11:28:02 +0000
parents d155623271e3
children 2029204dd999
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
9002
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
1
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
2 /*
11980
72256bd9557b copyright update by the author
diego
parents: 9122
diff changeset
3 * Discrete Cosine Tansform (DCT) for Altivec
72256bd9557b copyright update by the author
diego
parents: 9122
diff changeset
4 * Copyright (c) 2004 Romain Dolbeau <romain@dolbeau.org>
72256bd9557b copyright update by the author
diego
parents: 9122
diff changeset
5 * based upon code from "mp3lib/dct64.c"
9002
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
6 */
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
7
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
8 #define real float
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
9
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
10 #include "mpg123.h"
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
11
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
12 #ifdef HAVE_ALTIVEC
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
13
9122
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9002
diff changeset
14 #ifndef SYS_DARWIN
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9002
diff changeset
15 #include <altivec.h>
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9002
diff changeset
16 #endif
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9002
diff changeset
17
9002
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
18 // used to build registers permutation vectors (vcprm)
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
19 // the 's' are for words in the _s_econd vector
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
20 #define WORD_0 0x00,0x01,0x02,0x03
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
21 #define WORD_1 0x04,0x05,0x06,0x07
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
22 #define WORD_2 0x08,0x09,0x0a,0x0b
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
23 #define WORD_3 0x0c,0x0d,0x0e,0x0f
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
24 #define WORD_s0 0x10,0x11,0x12,0x13
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
25 #define WORD_s1 0x14,0x15,0x16,0x17
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
26 #define WORD_s2 0x18,0x19,0x1a,0x1b
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
27 #define WORD_s3 0x1c,0x1d,0x1e,0x1f
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
28
9122
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9002
diff changeset
29 #ifdef SYS_DARWIN
9002
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
30 #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: 9002
diff changeset
31 #else
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9002
diff changeset
32 #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: 9002
diff changeset
33 #endif
9002
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
34
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
35 // vcprmle is used to keep the same index as in the SSE version.
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
36 // it's the same as vcprm, with the index inversed
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
37 // ('le' is Little Endian)
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
38 #define vcprmle(a,b,c,d) vcprm(d,c,b,a)
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
39
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
40 // used to build inverse/identity vectors (vcii)
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
41 // n is _n_egative, p is _p_ositive
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
42 #define FLOAT_n -1.
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
43 #define FLOAT_p 1.
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
44
9122
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9002
diff changeset
45 #ifdef SYS_DARWIN
9002
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
46 #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: 9002
diff changeset
47 #else
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9002
diff changeset
48 #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: 9002
diff changeset
49 #endif
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9002
diff changeset
50
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9002
diff changeset
51 #ifdef SYS_DARWIN
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9002
diff changeset
52 #define FOUROF(a) (a)
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9002
diff changeset
53 #else
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9002
diff changeset
54 #define FOUROF(a) {a,a,a,a}
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9002
diff changeset
55 #endif
9002
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
56
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
57 void dct64_altivec(real *a,real *b,real *c)
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
58 {
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
59 real __attribute__ ((aligned(16))) b1[0x20];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
60 real __attribute__ ((aligned(16))) b2[0x20];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
61
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
62 real *out0 = a;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
63 real *out1 = b;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
64 real *samples = c;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
65
9122
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9002
diff changeset
66 const vector float vczero = (const vector float)FOUROF(0.);
9002
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
67 const vector unsigned char reverse = (const vector unsigned char)vcprm(3,2,1,0);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
68
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
69
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
70 if (((unsigned long)b1 & 0x0000000F) ||
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
71 ((unsigned long)b2 & 0x0000000F))
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
72
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
73 {
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
74 printf("MISALIGNED:\t%p\t%p\t%p\t%p\t%p\n",
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
75 b1, b2, a, b, samples);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
76 }
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
77
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
78
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
79 #ifdef ALTIVEC_USE_REFERENCE_C_CODE
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
80
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
81 {
12131
d155623271e3 fix symbol clashes when linking with libmp3lame including mp3 decoder, man, mp3lib is so much bloated
alex
parents: 11980
diff changeset
82 register real *costab = mp3lib_pnts[0];
9002
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
83
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
84 b1[0x00] = samples[0x00] + samples[0x1F];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
85 b1[0x01] = samples[0x01] + samples[0x1E];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
86 b1[0x02] = samples[0x02] + samples[0x1D];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
87 b1[0x03] = samples[0x03] + samples[0x1C];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
88 b1[0x04] = samples[0x04] + samples[0x1B];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
89 b1[0x05] = samples[0x05] + samples[0x1A];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
90 b1[0x06] = samples[0x06] + samples[0x19];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
91 b1[0x07] = samples[0x07] + samples[0x18];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
92 b1[0x08] = samples[0x08] + samples[0x17];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
93 b1[0x09] = samples[0x09] + samples[0x16];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
94 b1[0x0A] = samples[0x0A] + samples[0x15];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
95 b1[0x0B] = samples[0x0B] + samples[0x14];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
96 b1[0x0C] = samples[0x0C] + samples[0x13];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
97 b1[0x0D] = samples[0x0D] + samples[0x12];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
98 b1[0x0E] = samples[0x0E] + samples[0x11];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
99 b1[0x0F] = samples[0x0F] + samples[0x10];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
100 b1[0x10] = (samples[0x0F] - samples[0x10]) * costab[0xF];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
101 b1[0x11] = (samples[0x0E] - samples[0x11]) * costab[0xE];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
102 b1[0x12] = (samples[0x0D] - samples[0x12]) * costab[0xD];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
103 b1[0x13] = (samples[0x0C] - samples[0x13]) * costab[0xC];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
104 b1[0x14] = (samples[0x0B] - samples[0x14]) * costab[0xB];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
105 b1[0x15] = (samples[0x0A] - samples[0x15]) * costab[0xA];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
106 b1[0x16] = (samples[0x09] - samples[0x16]) * costab[0x9];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
107 b1[0x17] = (samples[0x08] - samples[0x17]) * costab[0x8];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
108 b1[0x18] = (samples[0x07] - samples[0x18]) * costab[0x7];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
109 b1[0x19] = (samples[0x06] - samples[0x19]) * costab[0x6];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
110 b1[0x1A] = (samples[0x05] - samples[0x1A]) * costab[0x5];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
111 b1[0x1B] = (samples[0x04] - samples[0x1B]) * costab[0x4];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
112 b1[0x1C] = (samples[0x03] - samples[0x1C]) * costab[0x3];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
113 b1[0x1D] = (samples[0x02] - samples[0x1D]) * costab[0x2];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
114 b1[0x1E] = (samples[0x01] - samples[0x1E]) * costab[0x1];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
115 b1[0x1F] = (samples[0x00] - samples[0x1F]) * costab[0x0];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
116
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
117 }
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
118 {
12131
d155623271e3 fix symbol clashes when linking with libmp3lame including mp3 decoder, man, mp3lib is so much bloated
alex
parents: 11980
diff changeset
119 register real *costab = mp3lib_pnts[1];
9002
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
120
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
121 b2[0x00] = b1[0x00] + b1[0x0F];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
122 b2[0x01] = b1[0x01] + b1[0x0E];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
123 b2[0x02] = b1[0x02] + b1[0x0D];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
124 b2[0x03] = b1[0x03] + b1[0x0C];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
125 b2[0x04] = b1[0x04] + b1[0x0B];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
126 b2[0x05] = b1[0x05] + b1[0x0A];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
127 b2[0x06] = b1[0x06] + b1[0x09];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
128 b2[0x07] = b1[0x07] + b1[0x08];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
129 b2[0x08] = (b1[0x07] - b1[0x08]) * costab[7];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
130 b2[0x09] = (b1[0x06] - b1[0x09]) * costab[6];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
131 b2[0x0A] = (b1[0x05] - b1[0x0A]) * costab[5];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
132 b2[0x0B] = (b1[0x04] - b1[0x0B]) * costab[4];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
133 b2[0x0C] = (b1[0x03] - b1[0x0C]) * costab[3];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
134 b2[0x0D] = (b1[0x02] - b1[0x0D]) * costab[2];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
135 b2[0x0E] = (b1[0x01] - b1[0x0E]) * costab[1];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
136 b2[0x0F] = (b1[0x00] - b1[0x0F]) * costab[0];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
137 b2[0x10] = b1[0x10] + b1[0x1F];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
138 b2[0x11] = b1[0x11] + b1[0x1E];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
139 b2[0x12] = b1[0x12] + b1[0x1D];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
140 b2[0x13] = b1[0x13] + b1[0x1C];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
141 b2[0x14] = b1[0x14] + b1[0x1B];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
142 b2[0x15] = b1[0x15] + b1[0x1A];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
143 b2[0x16] = b1[0x16] + b1[0x19];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
144 b2[0x17] = b1[0x17] + b1[0x18];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
145 b2[0x18] = (b1[0x18] - b1[0x17]) * costab[7];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
146 b2[0x19] = (b1[0x19] - b1[0x16]) * costab[6];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
147 b2[0x1A] = (b1[0x1A] - b1[0x15]) * costab[5];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
148 b2[0x1B] = (b1[0x1B] - b1[0x14]) * costab[4];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
149 b2[0x1C] = (b1[0x1C] - b1[0x13]) * costab[3];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
150 b2[0x1D] = (b1[0x1D] - b1[0x12]) * costab[2];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
151 b2[0x1E] = (b1[0x1E] - b1[0x11]) * costab[1];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
152 b2[0x1F] = (b1[0x1F] - b1[0x10]) * costab[0];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
153
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
154 }
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
155
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
156 {
12131
d155623271e3 fix symbol clashes when linking with libmp3lame including mp3 decoder, man, mp3lib is so much bloated
alex
parents: 11980
diff changeset
157 register real *costab = mp3lib_pnts[2];
9002
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
158
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
159 b1[0x00] = b2[0x00] + b2[0x07];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
160 b1[0x01] = b2[0x01] + b2[0x06];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
161 b1[0x02] = b2[0x02] + b2[0x05];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
162 b1[0x03] = b2[0x03] + b2[0x04];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
163 b1[0x04] = (b2[0x03] - b2[0x04]) * costab[3];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
164 b1[0x05] = (b2[0x02] - b2[0x05]) * costab[2];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
165 b1[0x06] = (b2[0x01] - b2[0x06]) * costab[1];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
166 b1[0x07] = (b2[0x00] - b2[0x07]) * costab[0];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
167 b1[0x08] = b2[0x08] + b2[0x0F];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
168 b1[0x09] = b2[0x09] + b2[0x0E];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
169 b1[0x0A] = b2[0x0A] + b2[0x0D];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
170 b1[0x0B] = b2[0x0B] + b2[0x0C];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
171 b1[0x0C] = (b2[0x0C] - b2[0x0B]) * costab[3];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
172 b1[0x0D] = (b2[0x0D] - b2[0x0A]) * costab[2];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
173 b1[0x0E] = (b2[0x0E] - b2[0x09]) * costab[1];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
174 b1[0x0F] = (b2[0x0F] - b2[0x08]) * costab[0];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
175 b1[0x10] = b2[0x10] + b2[0x17];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
176 b1[0x11] = b2[0x11] + b2[0x16];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
177 b1[0x12] = b2[0x12] + b2[0x15];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
178 b1[0x13] = b2[0x13] + b2[0x14];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
179 b1[0x14] = (b2[0x13] - b2[0x14]) * costab[3];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
180 b1[0x15] = (b2[0x12] - b2[0x15]) * costab[2];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
181 b1[0x16] = (b2[0x11] - b2[0x16]) * costab[1];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
182 b1[0x17] = (b2[0x10] - b2[0x17]) * costab[0];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
183 b1[0x18] = b2[0x18] + b2[0x1F];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
184 b1[0x19] = b2[0x19] + b2[0x1E];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
185 b1[0x1A] = b2[0x1A] + b2[0x1D];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
186 b1[0x1B] = b2[0x1B] + b2[0x1C];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
187 b1[0x1C] = (b2[0x1C] - b2[0x1B]) * costab[3];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
188 b1[0x1D] = (b2[0x1D] - b2[0x1A]) * costab[2];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
189 b1[0x1E] = (b2[0x1E] - b2[0x19]) * costab[1];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
190 b1[0x1F] = (b2[0x1F] - b2[0x18]) * costab[0];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
191 }
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
192
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
193 #else /* ALTIVEC_USE_REFERENCE_C_CODE */
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
194
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
195 // How does it work ?
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
196 // the first three passes are reproducted in the three block below
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
197 // all computations are done on a 4 elements vector
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
198 // 'reverse' is a special perumtation vector used to reverse
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
199 // the order of the elements inside a vector.
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
200 // note that all loads/stores to b1 (b2) between passes 1 and 2 (2 and 3)
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
201 // have been removed, all elements are stored inside b1vX (b2vX)
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
202 {
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
203 register vector float
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
204 b1v0, b1v1, b1v2, b1v3,
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
205 b1v4, b1v5, b1v6, b1v7;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
206 register vector float
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
207 temp1, temp2;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
208
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
209 {
12131
d155623271e3 fix symbol clashes when linking with libmp3lame including mp3 decoder, man, mp3lib is so much bloated
alex
parents: 11980
diff changeset
210 register real *costab = mp3lib_pnts[0];
9002
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
211
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
212 register vector float
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
213 samplesv1, samplesv2, samplesv3, samplesv4,
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
214 samplesv5, samplesv6, samplesv7, samplesv8,
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
215 samplesv9;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
216 register vector unsigned char samples_perm = vec_lvsl(0, samples);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
217 register vector float costabv1, costabv2, costabv3, costabv4, costabv5;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
218 register vector unsigned char costab_perm = vec_lvsl(0, costab);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
219
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
220 samplesv1 = vec_ld(0, samples);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
221 samplesv2 = vec_ld(16, samples);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
222 samplesv1 = vec_perm(samplesv1, samplesv2, samples_perm);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
223 samplesv3 = vec_ld(32, samples);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
224 samplesv2 = vec_perm(samplesv2, samplesv3, samples_perm);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
225 samplesv4 = vec_ld(48, samples);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
226 samplesv3 = vec_perm(samplesv3, samplesv4, samples_perm);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
227 samplesv5 = vec_ld(64, samples);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
228 samplesv4 = vec_perm(samplesv4, samplesv5, samples_perm);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
229 samplesv6 = vec_ld(80, samples);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
230 samplesv5 = vec_perm(samplesv5, samplesv6, samples_perm);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
231 samplesv7 = vec_ld(96, samples);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
232 samplesv6 = vec_perm(samplesv6, samplesv7, samples_perm);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
233 samplesv8 = vec_ld(112, samples);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
234 samplesv7 = vec_perm(samplesv7, samplesv8, samples_perm);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
235 samplesv9 = vec_ld(128, samples);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
236 samplesv8 = vec_perm(samplesv8, samplesv9, samples_perm);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
237
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
238 temp1 = vec_add(samplesv1,
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
239 vec_perm(samplesv8, samplesv8, reverse));
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
240 //vec_st(temp1, 0, b1);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
241 b1v0 = temp1;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
242 temp1 = vec_add(samplesv2,
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
243 vec_perm(samplesv7, samplesv7, reverse));
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
244 //vec_st(temp1, 16, b1);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
245 b1v1 = temp1;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
246 temp1 = vec_add(samplesv3,
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
247 vec_perm(samplesv6, samplesv6, reverse));
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
248 //vec_st(temp1, 32, b1);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
249 b1v2 = temp1;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
250 temp1 = vec_add(samplesv4,
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
251 vec_perm(samplesv5, samplesv5, reverse));
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
252 //vec_st(temp1, 48, b1);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
253 b1v3 = temp1;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
254
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
255 costabv1 = vec_ld(0, costab);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
256 costabv2 = vec_ld(16, costab);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
257 costabv1 = vec_perm(costabv1, costabv2, costab_perm);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
258 costabv3 = vec_ld(32, costab);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
259 costabv2 = vec_perm(costabv2, costabv3, costab_perm);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
260 costabv4 = vec_ld(48, costab);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
261 costabv3 = vec_perm(costabv3, costabv4, costab_perm);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
262 costabv5 = vec_ld(64, costab);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
263 costabv4 = vec_perm(costabv4, costabv5, costab_perm);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
264
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
265 temp1 = vec_sub(vec_perm(samplesv4, samplesv4, reverse),
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
266 samplesv5);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
267 temp2 = vec_madd(temp1,
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
268 vec_perm(costabv4, costabv4, reverse),
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
269 vczero);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
270 //vec_st(temp2, 64, b1);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
271 b1v4 = temp2;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
272
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
273 temp1 = vec_sub(vec_perm(samplesv3, samplesv3, reverse),
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
274 samplesv6);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
275 temp2 = vec_madd(temp1,
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
276 vec_perm(costabv3, costabv3, reverse),
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
277 vczero);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
278 //vec_st(temp2, 80, b1);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
279 b1v5 = temp2;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
280 temp1 = vec_sub(vec_perm(samplesv2, samplesv2, reverse),
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
281 samplesv7);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
282 temp2 = vec_madd(temp1,
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
283 vec_perm(costabv2, costabv2, reverse),
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
284 vczero);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
285 //vec_st(temp2, 96, b1);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
286 b1v6 = temp2;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
287
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
288 temp1 = vec_sub(vec_perm(samplesv1, samplesv1, reverse),
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
289 samplesv8);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
290 temp2 = vec_madd(temp1,
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
291 vec_perm(costabv1, costabv1, reverse),
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
292 vczero);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
293 //vec_st(temp2, 112, b1);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
294 b1v7 = temp2;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
295
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
296 }
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
297
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
298 {
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
299 register vector float
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
300 b2v0, b2v1, b2v2, b2v3,
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
301 b2v4, b2v5, b2v6, b2v7;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
302 {
12131
d155623271e3 fix symbol clashes when linking with libmp3lame including mp3 decoder, man, mp3lib is so much bloated
alex
parents: 11980
diff changeset
303 register real *costab = mp3lib_pnts[1];
9002
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
304 register vector float costabv1r, costabv2r, costabv1, costabv2, costabv3;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
305 register vector unsigned char costab_perm = vec_lvsl(0, costab);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
306
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
307 costabv1 = vec_ld(0, costab);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
308 costabv2 = vec_ld(16, costab);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
309 costabv1 = vec_perm(costabv1, costabv2, costab_perm);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
310 costabv3 = vec_ld(32, costab);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
311 costabv2 = vec_perm(costabv2, costabv3 , costab_perm);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
312 costabv1r = vec_perm(costabv1, costabv1, reverse);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
313 costabv2r = vec_perm(costabv2, costabv2, reverse);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
314
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
315 temp1 = vec_add(b1v0, vec_perm(b1v3, b1v3, reverse));
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
316 //vec_st(temp1, 0, b2);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
317 b2v0 = temp1;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
318 temp1 = vec_add(b1v1, vec_perm(b1v2, b1v2, reverse));
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
319 //vec_st(temp1, 16, b2);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
320 b2v1 = temp1;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
321 temp2 = vec_sub(vec_perm(b1v1, b1v1, reverse), b1v2);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
322 temp1 = vec_madd(temp2, costabv2r, vczero);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
323 //vec_st(temp1, 32, b2);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
324 b2v2 = temp1;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
325 temp2 = vec_sub(vec_perm(b1v0, b1v0, reverse), b1v3);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
326 temp1 = vec_madd(temp2, costabv1r, vczero);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
327 //vec_st(temp1, 48, b2);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
328 b2v3 = temp1;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
329 temp1 = vec_add(b1v4, vec_perm(b1v7, b1v7, reverse));
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
330 //vec_st(temp1, 64, b2);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
331 b2v4 = temp1;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
332 temp1 = vec_add(b1v5, vec_perm(b1v6, b1v6, reverse));
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
333 //vec_st(temp1, 80, b2);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
334 b2v5 = temp1;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
335 temp2 = vec_sub(b1v6, vec_perm(b1v5, b1v5, reverse));
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
336 temp1 = vec_madd(temp2, costabv2r, vczero);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
337 //vec_st(temp1, 96, b2);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
338 b2v6 = temp1;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
339 temp2 = vec_sub(b1v7, vec_perm(b1v4, b1v4, reverse));
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
340 temp1 = vec_madd(temp2, costabv1r, vczero);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
341 //vec_st(temp1, 112, b2);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
342 b2v7 = temp1;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
343 }
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
344
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
345 {
12131
d155623271e3 fix symbol clashes when linking with libmp3lame including mp3 decoder, man, mp3lib is so much bloated
alex
parents: 11980
diff changeset
346 register real *costab = mp3lib_pnts[2];
9002
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
347
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
348
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
349 vector float costabv1r, costabv1, costabv2;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
350 vector unsigned char costab_perm = vec_lvsl(0, costab);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
351
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
352 costabv1 = vec_ld(0, costab);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
353 costabv2 = vec_ld(16, costab);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
354 costabv1 = vec_perm(costabv1, costabv2, costab_perm);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
355 costabv1r = vec_perm(costabv1, costabv1, reverse);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
356
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
357 temp1 = vec_add(b2v0, vec_perm(b2v1, b2v1, reverse));
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
358 vec_st(temp1, 0, b1);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
359 temp2 = vec_sub(vec_perm(b2v0, b2v0, reverse), b2v1);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
360 temp1 = vec_madd(temp2, costabv1r, vczero);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
361 vec_st(temp1, 16, b1);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
362
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
363 temp1 = vec_add(b2v2, vec_perm(b2v3, b2v3, reverse));
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
364 vec_st(temp1, 32, b1);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
365 temp2 = vec_sub(b2v3, vec_perm(b2v2, b2v2, reverse));
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
366 temp1 = vec_madd(temp2, costabv1r, vczero);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
367 vec_st(temp1, 48, b1);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
368
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
369 temp1 = vec_add(b2v4, vec_perm(b2v5, b2v5, reverse));
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
370 vec_st(temp1, 64, b1);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
371 temp2 = vec_sub(vec_perm(b2v4, b2v4, reverse), b2v5);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
372 temp1 = vec_madd(temp2, costabv1r, vczero);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
373 vec_st(temp1, 80, b1);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
374
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
375 temp1 = vec_add(b2v6, vec_perm(b2v7, b2v7, reverse));
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
376 vec_st(temp1, 96, b1);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
377 temp2 = vec_sub(b2v7, vec_perm(b2v6, b2v6, reverse));
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
378 temp1 = vec_madd(temp2, costabv1r, vczero);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
379 vec_st(temp1, 112, b1);
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
380
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
381 }
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
382 }
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
383 }
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
384
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
385 #endif /* ALTIVEC_USE_REFERENCE_C_CODE */
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
386
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
387 {
12131
d155623271e3 fix symbol clashes when linking with libmp3lame including mp3 decoder, man, mp3lib is so much bloated
alex
parents: 11980
diff changeset
388 register real const cos0 = mp3lib_pnts[3][0];
d155623271e3 fix symbol clashes when linking with libmp3lame including mp3 decoder, man, mp3lib is so much bloated
alex
parents: 11980
diff changeset
389 register real const cos1 = mp3lib_pnts[3][1];
9002
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
390
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
391 b2[0x00] = b1[0x00] + b1[0x03];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
392 b2[0x01] = b1[0x01] + b1[0x02];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
393 b2[0x02] = (b1[0x01] - b1[0x02]) * cos1;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
394 b2[0x03] = (b1[0x00] - b1[0x03]) * cos0;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
395 b2[0x04] = b1[0x04] + b1[0x07];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
396 b2[0x05] = b1[0x05] + b1[0x06];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
397 b2[0x06] = (b1[0x06] - b1[0x05]) * cos1;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
398 b2[0x07] = (b1[0x07] - b1[0x04]) * cos0;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
399 b2[0x08] = b1[0x08] + b1[0x0B];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
400 b2[0x09] = b1[0x09] + b1[0x0A];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
401 b2[0x0A] = (b1[0x09] - b1[0x0A]) * cos1;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
402 b2[0x0B] = (b1[0x08] - b1[0x0B]) * cos0;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
403 b2[0x0C] = b1[0x0C] + b1[0x0F];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
404 b2[0x0D] = b1[0x0D] + b1[0x0E];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
405 b2[0x0E] = (b1[0x0E] - b1[0x0D]) * cos1;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
406 b2[0x0F] = (b1[0x0F] - b1[0x0C]) * cos0;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
407 b2[0x10] = b1[0x10] + b1[0x13];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
408 b2[0x11] = b1[0x11] + b1[0x12];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
409 b2[0x12] = (b1[0x11] - b1[0x12]) * cos1;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
410 b2[0x13] = (b1[0x10] - b1[0x13]) * cos0;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
411 b2[0x14] = b1[0x14] + b1[0x17];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
412 b2[0x15] = b1[0x15] + b1[0x16];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
413 b2[0x16] = (b1[0x16] - b1[0x15]) * cos1;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
414 b2[0x17] = (b1[0x17] - b1[0x14]) * cos0;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
415 b2[0x18] = b1[0x18] + b1[0x1B];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
416 b2[0x19] = b1[0x19] + b1[0x1A];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
417 b2[0x1A] = (b1[0x19] - b1[0x1A]) * cos1;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
418 b2[0x1B] = (b1[0x18] - b1[0x1B]) * cos0;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
419 b2[0x1C] = b1[0x1C] + b1[0x1F];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
420 b2[0x1D] = b1[0x1D] + b1[0x1E];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
421 b2[0x1E] = (b1[0x1E] - b1[0x1D]) * cos1;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
422 b2[0x1F] = (b1[0x1F] - b1[0x1C]) * cos0;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
423 }
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
424
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
425 {
12131
d155623271e3 fix symbol clashes when linking with libmp3lame including mp3 decoder, man, mp3lib is so much bloated
alex
parents: 11980
diff changeset
426 register real const cos0 = mp3lib_pnts[4][0];
9002
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
427
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
428 b1[0x00] = b2[0x00] + b2[0x01];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
429 b1[0x01] = (b2[0x00] - b2[0x01]) * cos0;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
430 b1[0x02] = b2[0x02] + b2[0x03];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
431 b1[0x03] = (b2[0x03] - b2[0x02]) * cos0;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
432 b1[0x02] += b1[0x03];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
433
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
434 b1[0x04] = b2[0x04] + b2[0x05];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
435 b1[0x05] = (b2[0x04] - b2[0x05]) * cos0;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
436 b1[0x06] = b2[0x06] + b2[0x07];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
437 b1[0x07] = (b2[0x07] - b2[0x06]) * cos0;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
438 b1[0x06] += b1[0x07];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
439 b1[0x04] += b1[0x06];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
440 b1[0x06] += b1[0x05];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
441 b1[0x05] += b1[0x07];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
442
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
443 b1[0x08] = b2[0x08] + b2[0x09];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
444 b1[0x09] = (b2[0x08] - b2[0x09]) * cos0;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
445 b1[0x0A] = b2[0x0A] + b2[0x0B];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
446 b1[0x0B] = (b2[0x0B] - b2[0x0A]) * cos0;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
447 b1[0x0A] += b1[0x0B];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
448
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
449 b1[0x0C] = b2[0x0C] + b2[0x0D];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
450 b1[0x0D] = (b2[0x0C] - b2[0x0D]) * cos0;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
451 b1[0x0E] = b2[0x0E] + b2[0x0F];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
452 b1[0x0F] = (b2[0x0F] - b2[0x0E]) * cos0;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
453 b1[0x0E] += b1[0x0F];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
454 b1[0x0C] += b1[0x0E];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
455 b1[0x0E] += b1[0x0D];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
456 b1[0x0D] += b1[0x0F];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
457
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
458 b1[0x10] = b2[0x10] + b2[0x11];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
459 b1[0x11] = (b2[0x10] - b2[0x11]) * cos0;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
460 b1[0x12] = b2[0x12] + b2[0x13];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
461 b1[0x13] = (b2[0x13] - b2[0x12]) * cos0;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
462 b1[0x12] += b1[0x13];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
463
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
464 b1[0x14] = b2[0x14] + b2[0x15];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
465 b1[0x15] = (b2[0x14] - b2[0x15]) * cos0;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
466 b1[0x16] = b2[0x16] + b2[0x17];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
467 b1[0x17] = (b2[0x17] - b2[0x16]) * cos0;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
468 b1[0x16] += b1[0x17];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
469 b1[0x14] += b1[0x16];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
470 b1[0x16] += b1[0x15];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
471 b1[0x15] += b1[0x17];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
472
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
473 b1[0x18] = b2[0x18] + b2[0x19];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
474 b1[0x19] = (b2[0x18] - b2[0x19]) * cos0;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
475 b1[0x1A] = b2[0x1A] + b2[0x1B];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
476 b1[0x1B] = (b2[0x1B] - b2[0x1A]) * cos0;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
477 b1[0x1A] += b1[0x1B];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
478
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
479 b1[0x1C] = b2[0x1C] + b2[0x1D];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
480 b1[0x1D] = (b2[0x1C] - b2[0x1D]) * cos0;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
481 b1[0x1E] = b2[0x1E] + b2[0x1F];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
482 b1[0x1F] = (b2[0x1F] - b2[0x1E]) * cos0;
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
483 b1[0x1E] += b1[0x1F];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
484 b1[0x1C] += b1[0x1E];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
485 b1[0x1E] += b1[0x1D];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
486 b1[0x1D] += b1[0x1F];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
487 }
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
488
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
489 out0[0x10*16] = b1[0x00];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
490 out0[0x10*12] = b1[0x04];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
491 out0[0x10* 8] = b1[0x02];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
492 out0[0x10* 4] = b1[0x06];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
493 out0[0x10* 0] = b1[0x01];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
494 out1[0x10* 0] = b1[0x01];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
495 out1[0x10* 4] = b1[0x05];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
496 out1[0x10* 8] = b1[0x03];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
497 out1[0x10*12] = b1[0x07];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
498
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
499 b1[0x08] += b1[0x0C];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
500 out0[0x10*14] = b1[0x08];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
501 b1[0x0C] += b1[0x0a];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
502 out0[0x10*10] = b1[0x0C];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
503 b1[0x0A] += b1[0x0E];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
504 out0[0x10* 6] = b1[0x0A];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
505 b1[0x0E] += b1[0x09];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
506 out0[0x10* 2] = b1[0x0E];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
507 b1[0x09] += b1[0x0D];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
508 out1[0x10* 2] = b1[0x09];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
509 b1[0x0D] += b1[0x0B];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
510 out1[0x10* 6] = b1[0x0D];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
511 b1[0x0B] += b1[0x0F];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
512 out1[0x10*10] = b1[0x0B];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
513 out1[0x10*14] = b1[0x0F];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
514
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
515 b1[0x18] += b1[0x1C];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
516 out0[0x10*15] = b1[0x10] + b1[0x18];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
517 out0[0x10*13] = b1[0x18] + b1[0x14];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
518 b1[0x1C] += b1[0x1a];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
519 out0[0x10*11] = b1[0x14] + b1[0x1C];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
520 out0[0x10* 9] = b1[0x1C] + b1[0x12];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
521 b1[0x1A] += b1[0x1E];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
522 out0[0x10* 7] = b1[0x12] + b1[0x1A];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
523 out0[0x10* 5] = b1[0x1A] + b1[0x16];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
524 b1[0x1E] += b1[0x19];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
525 out0[0x10* 3] = b1[0x16] + b1[0x1E];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
526 out0[0x10* 1] = b1[0x1E] + b1[0x11];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
527 b1[0x19] += b1[0x1D];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
528 out1[0x10* 1] = b1[0x11] + b1[0x19];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
529 out1[0x10* 3] = b1[0x19] + b1[0x15];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
530 b1[0x1D] += b1[0x1B];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
531 out1[0x10* 5] = b1[0x15] + b1[0x1D];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
532 out1[0x10* 7] = b1[0x1D] + b1[0x13];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
533 b1[0x1B] += b1[0x1F];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
534 out1[0x10* 9] = b1[0x13] + b1[0x1B];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
535 out1[0x10*11] = b1[0x1B] + b1[0x17];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
536 out1[0x10*13] = b1[0x17] + b1[0x1F];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
537 out1[0x10*15] = b1[0x1F];
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
538 }
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
539
9122
5ba896a38d75 The two attached patches *should* allow for proper
arpi
parents: 9002
diff changeset
540 #endif /* HAVE_ALTIVEC */
9002
60d144a16088 An altivec-optimized DCT64 for mp3lib
arpi
parents:
diff changeset
541