3394
|
1 /*
|
|
2 * imdct.c
|
|
3 * Copyright (C) 2000-2001 Michel Lespinasse <walken@zoy.org>
|
|
4 * Copyright (C) 1999-2000 Aaron Holtzman <aholtzma@ess.engr.uvic.ca>
|
|
5 *
|
|
6 * This file is part of a52dec, a free ATSC A-52 stream decoder.
|
|
7 * See http://liba52.sourceforge.net/ for updates.
|
|
8 *
|
|
9 * a52dec is free software; you can redistribute it and/or modify
|
|
10 * it under the terms of the GNU General Public License as published by
|
|
11 * the Free Software Foundation; either version 2 of the License, or
|
|
12 * (at your option) any later version.
|
|
13 *
|
|
14 * a52dec is distributed in the hope that it will be useful,
|
|
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
|
17 * GNU General Public License for more details.
|
|
18 *
|
|
19 * You should have received a copy of the GNU General Public License
|
|
20 * along with this program; if not, write to the Free Software
|
|
21 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
|
|
22 */
|
|
23
|
|
24 #include "config.h"
|
|
25
|
|
26 #include <math.h>
|
|
27 #include <stdio.h>
|
|
28 #ifndef M_PI
|
|
29 #define M_PI 3.1415926535897932384626433832795029
|
|
30 #endif
|
|
31 #include <inttypes.h>
|
|
32
|
|
33 #include "a52.h"
|
|
34 #include "a52_internal.h"
|
|
35 #include "mm_accel.h"
|
|
36
|
|
37 void (* imdct_256) (sample_t data[], sample_t delay[], sample_t bias);
|
|
38 void (* imdct_512) (sample_t data[], sample_t delay[], sample_t bias);
|
|
39
|
|
40 typedef struct complex_s {
|
|
41 sample_t real;
|
|
42 sample_t imag;
|
|
43 } complex_t;
|
|
44
|
|
45
|
|
46 /* 128 point bit-reverse LUT */
|
|
47 static uint8_t bit_reverse_512[] = {
|
|
48 0x00, 0x40, 0x20, 0x60, 0x10, 0x50, 0x30, 0x70,
|
|
49 0x08, 0x48, 0x28, 0x68, 0x18, 0x58, 0x38, 0x78,
|
|
50 0x04, 0x44, 0x24, 0x64, 0x14, 0x54, 0x34, 0x74,
|
|
51 0x0c, 0x4c, 0x2c, 0x6c, 0x1c, 0x5c, 0x3c, 0x7c,
|
|
52 0x02, 0x42, 0x22, 0x62, 0x12, 0x52, 0x32, 0x72,
|
|
53 0x0a, 0x4a, 0x2a, 0x6a, 0x1a, 0x5a, 0x3a, 0x7a,
|
|
54 0x06, 0x46, 0x26, 0x66, 0x16, 0x56, 0x36, 0x76,
|
|
55 0x0e, 0x4e, 0x2e, 0x6e, 0x1e, 0x5e, 0x3e, 0x7e,
|
|
56 0x01, 0x41, 0x21, 0x61, 0x11, 0x51, 0x31, 0x71,
|
|
57 0x09, 0x49, 0x29, 0x69, 0x19, 0x59, 0x39, 0x79,
|
|
58 0x05, 0x45, 0x25, 0x65, 0x15, 0x55, 0x35, 0x75,
|
|
59 0x0d, 0x4d, 0x2d, 0x6d, 0x1d, 0x5d, 0x3d, 0x7d,
|
|
60 0x03, 0x43, 0x23, 0x63, 0x13, 0x53, 0x33, 0x73,
|
|
61 0x0b, 0x4b, 0x2b, 0x6b, 0x1b, 0x5b, 0x3b, 0x7b,
|
|
62 0x07, 0x47, 0x27, 0x67, 0x17, 0x57, 0x37, 0x77,
|
|
63 0x0f, 0x4f, 0x2f, 0x6f, 0x1f, 0x5f, 0x3f, 0x7f};
|
|
64
|
|
65 static uint8_t bit_reverse_256[] = {
|
|
66 0x00, 0x20, 0x10, 0x30, 0x08, 0x28, 0x18, 0x38,
|
|
67 0x04, 0x24, 0x14, 0x34, 0x0c, 0x2c, 0x1c, 0x3c,
|
|
68 0x02, 0x22, 0x12, 0x32, 0x0a, 0x2a, 0x1a, 0x3a,
|
|
69 0x06, 0x26, 0x16, 0x36, 0x0e, 0x2e, 0x1e, 0x3e,
|
|
70 0x01, 0x21, 0x11, 0x31, 0x09, 0x29, 0x19, 0x39,
|
|
71 0x05, 0x25, 0x15, 0x35, 0x0d, 0x2d, 0x1d, 0x3d,
|
|
72 0x03, 0x23, 0x13, 0x33, 0x0b, 0x2b, 0x1b, 0x3b,
|
|
73 0x07, 0x27, 0x17, 0x37, 0x0f, 0x2f, 0x1f, 0x3f};
|
|
74
|
3508
|
75 #ifdef HAVE_SSE
|
|
76 // NOTE: SSE needs 16byte alignment or it will segfault
|
|
77 static complex_t __attribute__((aligned(16))) buf[128];
|
3527
|
78 static float __attribute__((aligned(16))) sseSinCos1a[256];
|
|
79 static float __attribute__((aligned(16))) sseSinCos1b[256];
|
3512
|
80 static float __attribute__((aligned(16))) ps111_1[4]={1,1,1,-1};
|
3534
|
81 //static float __attribute__((aligned(16))) sseW0[4];
|
|
82 static float __attribute__((aligned(16))) sseW1[8];
|
|
83 static float __attribute__((aligned(16))) sseW2[16];
|
|
84 static float __attribute__((aligned(16))) sseW3[32];
|
|
85 static float __attribute__((aligned(16))) sseW4[64];
|
|
86 static float __attribute__((aligned(16))) sseW5[128];
|
|
87 static float __attribute__((aligned(16))) sseW6[256];
|
|
88 static float __attribute__((aligned(16))) *sseW[7]=
|
|
89 {NULL /*sseW0*/,sseW1,sseW2,sseW3,sseW4,sseW5,sseW6};
|
3508
|
90 #else
|
3394
|
91 static complex_t buf[128];
|
3508
|
92 #endif
|
3394
|
93
|
|
94 /* Twiddle factor LUT */
|
|
95 static complex_t w_1[1];
|
|
96 static complex_t w_2[2];
|
|
97 static complex_t w_4[4];
|
|
98 static complex_t w_8[8];
|
|
99 static complex_t w_16[16];
|
|
100 static complex_t w_32[32];
|
|
101 static complex_t w_64[64];
|
|
102 static complex_t * w[7] = {w_1, w_2, w_4, w_8, w_16, w_32, w_64};
|
|
103
|
|
104 /* Twiddle factors for IMDCT */
|
|
105 static sample_t xcos1[128];
|
|
106 static sample_t xsin1[128];
|
|
107 static sample_t xcos2[64];
|
|
108 static sample_t xsin2[64];
|
|
109
|
|
110 /* Windowing function for Modified DCT - Thank you acroread */
|
|
111 sample_t imdct_window[] = {
|
|
112 0.00014, 0.00024, 0.00037, 0.00051, 0.00067, 0.00086, 0.00107, 0.00130,
|
|
113 0.00157, 0.00187, 0.00220, 0.00256, 0.00297, 0.00341, 0.00390, 0.00443,
|
|
114 0.00501, 0.00564, 0.00632, 0.00706, 0.00785, 0.00871, 0.00962, 0.01061,
|
|
115 0.01166, 0.01279, 0.01399, 0.01526, 0.01662, 0.01806, 0.01959, 0.02121,
|
|
116 0.02292, 0.02472, 0.02662, 0.02863, 0.03073, 0.03294, 0.03527, 0.03770,
|
|
117 0.04025, 0.04292, 0.04571, 0.04862, 0.05165, 0.05481, 0.05810, 0.06153,
|
|
118 0.06508, 0.06878, 0.07261, 0.07658, 0.08069, 0.08495, 0.08935, 0.09389,
|
|
119 0.09859, 0.10343, 0.10842, 0.11356, 0.11885, 0.12429, 0.12988, 0.13563,
|
|
120 0.14152, 0.14757, 0.15376, 0.16011, 0.16661, 0.17325, 0.18005, 0.18699,
|
|
121 0.19407, 0.20130, 0.20867, 0.21618, 0.22382, 0.23161, 0.23952, 0.24757,
|
|
122 0.25574, 0.26404, 0.27246, 0.28100, 0.28965, 0.29841, 0.30729, 0.31626,
|
|
123 0.32533, 0.33450, 0.34376, 0.35311, 0.36253, 0.37204, 0.38161, 0.39126,
|
|
124 0.40096, 0.41072, 0.42054, 0.43040, 0.44030, 0.45023, 0.46020, 0.47019,
|
|
125 0.48020, 0.49022, 0.50025, 0.51028, 0.52031, 0.53033, 0.54033, 0.55031,
|
|
126 0.56026, 0.57019, 0.58007, 0.58991, 0.59970, 0.60944, 0.61912, 0.62873,
|
|
127 0.63827, 0.64774, 0.65713, 0.66643, 0.67564, 0.68476, 0.69377, 0.70269,
|
|
128 0.71150, 0.72019, 0.72877, 0.73723, 0.74557, 0.75378, 0.76186, 0.76981,
|
|
129 0.77762, 0.78530, 0.79283, 0.80022, 0.80747, 0.81457, 0.82151, 0.82831,
|
|
130 0.83496, 0.84145, 0.84779, 0.85398, 0.86001, 0.86588, 0.87160, 0.87716,
|
|
131 0.88257, 0.88782, 0.89291, 0.89785, 0.90264, 0.90728, 0.91176, 0.91610,
|
|
132 0.92028, 0.92432, 0.92822, 0.93197, 0.93558, 0.93906, 0.94240, 0.94560,
|
|
133 0.94867, 0.95162, 0.95444, 0.95713, 0.95971, 0.96217, 0.96451, 0.96674,
|
|
134 0.96887, 0.97089, 0.97281, 0.97463, 0.97635, 0.97799, 0.97953, 0.98099,
|
|
135 0.98236, 0.98366, 0.98488, 0.98602, 0.98710, 0.98811, 0.98905, 0.98994,
|
|
136 0.99076, 0.99153, 0.99225, 0.99291, 0.99353, 0.99411, 0.99464, 0.99513,
|
|
137 0.99558, 0.99600, 0.99639, 0.99674, 0.99706, 0.99736, 0.99763, 0.99788,
|
|
138 0.99811, 0.99831, 0.99850, 0.99867, 0.99882, 0.99895, 0.99908, 0.99919,
|
|
139 0.99929, 0.99938, 0.99946, 0.99953, 0.99959, 0.99965, 0.99969, 0.99974,
|
|
140 0.99978, 0.99981, 0.99984, 0.99986, 0.99988, 0.99990, 0.99992, 0.99993,
|
|
141 0.99994, 0.99995, 0.99996, 0.99997, 0.99998, 0.99998, 0.99998, 0.99999,
|
|
142 0.99999, 0.99999, 0.99999, 1.00000, 1.00000, 1.00000, 1.00000, 1.00000,
|
|
143 1.00000, 1.00000, 1.00000, 1.00000, 1.00000, 1.00000, 1.00000, 1.00000 };
|
|
144
|
|
145
|
|
146 static inline void swap_cmplx(complex_t *a, complex_t *b)
|
|
147 {
|
|
148 complex_t tmp;
|
|
149
|
|
150 tmp = *a;
|
|
151 *a = *b;
|
|
152 *b = tmp;
|
|
153 }
|
|
154
|
|
155
|
|
156
|
|
157 static inline complex_t cmplx_mult(complex_t a, complex_t b)
|
|
158 {
|
|
159 complex_t ret;
|
|
160
|
|
161 ret.real = a.real * b.real - a.imag * b.imag;
|
|
162 ret.imag = a.real * b.imag + a.imag * b.real;
|
|
163
|
|
164 return ret;
|
|
165 }
|
|
166
|
|
167 void
|
|
168 imdct_do_512(sample_t data[],sample_t delay[], sample_t bias)
|
|
169 {
|
|
170 int i,k;
|
|
171 int p,q;
|
|
172 int m;
|
|
173 int two_m;
|
|
174 int two_m_plus_one;
|
|
175
|
|
176 sample_t tmp_a_i;
|
|
177 sample_t tmp_a_r;
|
|
178 sample_t tmp_b_i;
|
|
179 sample_t tmp_b_r;
|
|
180
|
|
181 sample_t *data_ptr;
|
|
182 sample_t *delay_ptr;
|
|
183 sample_t *window_ptr;
|
|
184
|
|
185 /* 512 IMDCT with source and dest data in 'data' */
|
|
186
|
3529
|
187 #ifdef HAVE_SSE
|
3394
|
188 /* Pre IFFT complex multiply plus IFFT cmplx conjugate */
|
3529
|
189 /* Bit reversed shuffling */
|
3527
|
190 asm volatile(
|
|
191 "xorl %%esi, %%esi \n\t"
|
3529
|
192 "leal bit_reverse_512, %%eax \n\t"
|
3527
|
193 "movl $1008, %%edi \n\t"
|
3529
|
194 "pushl %%ebp \n\t" //use ebp without telling gcc
|
|
195 ".balign 16 \n\t"
|
3527
|
196 "1: \n\t"
|
|
197 "movaps (%0, %%esi), %%xmm0 \n\t"
|
|
198 "movaps (%0, %%edi), %%xmm1 \n\t"
|
|
199 "shufps $0xA0, %%xmm0, %%xmm0 \n\t"
|
|
200 "shufps $0x5F, %%xmm1, %%xmm1 \n\t"
|
|
201 "mulps sseSinCos1a(%%esi), %%xmm0 \n\t"
|
|
202 "mulps sseSinCos1b(%%esi), %%xmm1 \n\t"
|
|
203 "addps %%xmm1, %%xmm0 \n\t"
|
3529
|
204 "movzbl (%%eax), %%edx \n\t"
|
|
205 "movzbl 1(%%eax), %%ebp \n\t"
|
|
206 "movlps %%xmm0, (%1, %%edx,8) \n\t"
|
|
207 "movhps %%xmm0, (%1, %%ebp,8) \n\t"
|
3527
|
208 "addl $16, %%esi \n\t"
|
3529
|
209 "addl $2, %%eax \n\t" // avoid complex addressing for P4 crap
|
3527
|
210 "subl $16, %%edi \n\t"
|
|
211 " jnc 1b \n\t"
|
3529
|
212 "popl %%ebp \n\t"//no we didnt touch ebp *g*
|
|
213 :: "b" (data), "c" (buf)
|
|
214 : "%esi", "%edi", "%eax", "%edx"
|
3527
|
215 );
|
|
216 #else
|
3529
|
217 /* Pre IFFT complex multiply plus IFFT cmplx conjugate */
|
3394
|
218 for( i=0; i < 128; i++) {
|
|
219 /* z[i] = (X[256-2*i-1] + j * X[2*i]) * (xcos1[i] + j * xsin1[i]) ; */
|
|
220 buf[i].real = (data[256-2*i-1] * xcos1[i]) - (data[2*i] * xsin1[i]);
|
|
221 buf[i].imag = -1.0 * ((data[2*i] * xcos1[i]) + (data[256-2*i-1] * xsin1[i]));
|
|
222 }
|
|
223
|
|
224 /* Bit reversed shuffling */
|
|
225 for(i=0; i<128; i++) {
|
|
226 k = bit_reverse_512[i];
|
|
227 if (k < i)
|
|
228 swap_cmplx(&buf[i],&buf[k]);
|
|
229 }
|
3529
|
230 #endif
|
|
231
|
3394
|
232
|
|
233 /* FFT Merge */
|
3508
|
234 #ifdef HAVE_SSE
|
|
235 // Note w[0][0]={1,0}
|
|
236 // C Code for the following asm loop
|
|
237 /* for(i = 0; i < 128; i += 2) {
|
|
238 p = 0 + i;
|
|
239 q = p + 1;
|
|
240 tmp_a_r = buf[p].real;
|
|
241 tmp_a_i = buf[p].imag;
|
|
242 tmp_b_r = buf[q].real;
|
|
243 tmp_b_i = buf[q].imag;
|
|
244 buf[p].real = tmp_a_r + tmp_b_r;
|
|
245 buf[p].imag = tmp_a_i + tmp_b_i;
|
|
246 buf[q].real = tmp_a_r - tmp_b_r;
|
|
247 buf[q].imag = tmp_a_i - tmp_b_i;
|
|
248 }*/
|
|
249 asm volatile(
|
|
250 "xorps %%xmm1, %%xmm1 \n\t"
|
|
251 "xorps %%xmm2, %%xmm2 \n\t"
|
|
252 "movl %0, %%esi \n\t"
|
3529
|
253 ".balign 16 \n\t"
|
3508
|
254 "1: \n\t"
|
|
255 "movlps (%%esi), %%xmm0 \n\t" //buf[p]
|
|
256 "movlps 8(%%esi), %%xmm1\n\t" //buf[q]
|
|
257 "movhps (%%esi), %%xmm0 \n\t" //buf[p]
|
|
258 "movhps 8(%%esi), %%xmm2\n\t" //buf[q]
|
|
259 "addps %%xmm1, %%xmm0 \n\t"
|
|
260 "subps %%xmm2, %%xmm0 \n\t"
|
|
261 "movaps %%xmm0, (%%esi) \n\t"
|
|
262 "addl $16, %%esi \n\t"
|
|
263 "cmpl %1, %%esi \n\t"
|
|
264 " jb 1b \n\t"
|
|
265 :: "g" (buf), "r" (buf + 128)
|
|
266 : "%esi"
|
|
267 );
|
|
268
|
3512
|
269 // Note w[1]={{1,0}, {0,-1}}
|
|
270 // C Code for the following asm loop
|
|
271 /* for(i = 0; i < 128; i += 4) {
|
|
272 p = 0 + i;
|
|
273 q = p + 2;
|
|
274 tmp_a_r = buf[p].real;
|
|
275 tmp_a_i = buf[p].imag;
|
|
276 tmp_b_r = buf[q].real;
|
|
277 tmp_b_i = buf[q].imag;
|
|
278 buf[p].real = tmp_a_r + tmp_b_r;
|
|
279 buf[p].imag = tmp_a_i + tmp_b_i;
|
|
280 buf[q].real = tmp_a_r - tmp_b_r;
|
|
281 buf[q].imag = tmp_a_i - tmp_b_i;
|
|
282 tmp_a_r = buf[p+1].real;
|
|
283 tmp_a_i = buf[p+1].imag;
|
|
284 tmp_b_r = buf[q+1].imag;
|
|
285 tmp_b_i = buf[q+1].real;
|
|
286 buf[p+1].real = tmp_a_r + tmp_b_r;
|
|
287 buf[p+1].imag = tmp_a_i - tmp_b_i;
|
|
288 buf[q+1].real = tmp_a_r - tmp_b_r;
|
|
289 buf[q+1].imag = tmp_a_i + tmp_b_i;
|
|
290 }
|
|
291 */
|
|
292 asm volatile(
|
|
293 "movaps ps111_1, %%xmm7 \n\t" // 1,1,1,-1
|
|
294 "movl %0, %%esi \n\t"
|
3529
|
295 ".balign 16 \n\t"
|
3512
|
296 "1: \n\t"
|
|
297 "movaps 16(%%esi), %%xmm2 \n\t" //r2,i2,r3,i3
|
|
298 "shufps $0xB4, %%xmm2, %%xmm2 \n\t" //r2,i2,i3,r3
|
|
299 "mulps %%xmm7, %%xmm2 \n\t" //r2,i2,i3,-r3
|
|
300 "movaps (%%esi), %%xmm0 \n\t" //r0,i0,r1,i1
|
|
301 "movaps (%%esi), %%xmm1 \n\t" //r0,i0,r1,i1
|
|
302 "addps %%xmm2, %%xmm0 \n\t"
|
|
303 "subps %%xmm2, %%xmm1 \n\t"
|
|
304 "movaps %%xmm0, (%%esi) \n\t"
|
|
305 "movaps %%xmm1, 16(%%esi) \n\t"
|
|
306 "addl $32, %%esi \n\t"
|
|
307 "cmpl %1, %%esi \n\t"
|
|
308 " jb 1b \n\t"
|
|
309 :: "g" (buf), "r" (buf + 128)
|
|
310 : "%esi"
|
|
311 );
|
3534
|
312 /* C code for the next asm loop
|
|
313 for(k = 0; k < 4; k++) {
|
3512
|
314 for(i = 0; i < 128; i += 8) {
|
|
315 p = k + i;
|
3534
|
316 q = p + 4;
|
3512
|
317 tmp_a_r = buf[p].real;
|
|
318 tmp_a_i = buf[p].imag;
|
3534
|
319 tmp_b_r = buf[q].real * w[2][k].real - buf[q].imag * w[2][k].imag;
|
|
320 tmp_b_i = buf[q].imag * w[2][k].real + buf[q].real * w[2][k].imag;
|
3512
|
321 buf[p].real = tmp_a_r + tmp_b_r;
|
|
322 buf[p].imag = tmp_a_i + tmp_b_i;
|
|
323 buf[q].real = tmp_a_r - tmp_b_r;
|
|
324 buf[q].imag = tmp_a_i - tmp_b_i;
|
|
325 }
|
|
326 }
|
3534
|
327 */
|
|
328 /*
|
|
329 Note sseW2+0={1,1,sqrt(2),sqrt(2))
|
|
330 Note sseW2+16={0,0,sqrt(2),-sqrt(2))
|
|
331 Note sseW2+32={0,0,-sqrt(2),-sqrt(2))
|
|
332 Note sseW2+48={1,-1,sqrt(2),-sqrt(2))
|
|
333 */
|
|
334 asm volatile(
|
3537
|
335 "movaps 48+sseW2, %%xmm6 \n\t"
|
3534
|
336 "movaps 16+sseW2, %%xmm7 \n\t"
|
|
337 "xorps %%xmm5, %%xmm5 \n\t"
|
|
338 "xorps %%xmm2, %%xmm2 \n\t"
|
|
339 "movl %0, %%esi \n\t"
|
|
340 ".balign 16 \n\t"
|
|
341 "1: \n\t"
|
3537
|
342 "movaps 32(%%esi), %%xmm2 \n\t" //r4,i4,r5,i5
|
3534
|
343 "movaps 48(%%esi), %%xmm3 \n\t" //r6,i6,r7,i7
|
3537
|
344 "movaps sseW2, %%xmm4 \n\t" //r4,i4,r5,i5
|
|
345 "movaps 32+sseW2, %%xmm5 \n\t" //r6,i6,r7,i7
|
|
346 "mulps %%xmm2, %%xmm4 \n\t"
|
|
347 "mulps %%xmm3, %%xmm5 \n\t"
|
3534
|
348 "shufps $0xB1, %%xmm2, %%xmm2 \n\t" //i4,r4,i5,r5
|
|
349 "shufps $0xB1, %%xmm3, %%xmm3 \n\t" //i6,r6,i7,r7
|
3537
|
350 "mulps %%xmm6, %%xmm3 \n\t"
|
3534
|
351 "mulps %%xmm7, %%xmm2 \n\t"
|
|
352 "movaps (%%esi), %%xmm0 \n\t" //r0,i0,r1,i1
|
|
353 "movaps 16(%%esi), %%xmm1 \n\t" //r2,i2,r3,i3
|
|
354 "addps %%xmm4, %%xmm2 \n\t"
|
|
355 "addps %%xmm5, %%xmm3 \n\t"
|
|
356 "movaps %%xmm2, %%xmm4 \n\t"
|
|
357 "movaps %%xmm3, %%xmm5 \n\t"
|
|
358 "addps %%xmm0, %%xmm2 \n\t"
|
|
359 "addps %%xmm1, %%xmm3 \n\t"
|
|
360 "subps %%xmm4, %%xmm0 \n\t"
|
|
361 "subps %%xmm5, %%xmm1 \n\t"
|
|
362 "movaps %%xmm2, (%%esi) \n\t"
|
|
363 "movaps %%xmm3, 16(%%esi) \n\t"
|
|
364 "movaps %%xmm0, 32(%%esi) \n\t"
|
|
365 "movaps %%xmm1, 48(%%esi) \n\t"
|
|
366 "addl $64, %%esi \n\t"
|
|
367 "cmpl %1, %%esi \n\t"
|
|
368 " jb 1b \n\t"
|
|
369 :: "g" (buf), "r" (buf + 128)
|
|
370 : "%esi"
|
|
371 );
|
3512
|
372
|
3546
|
373 /* C version of the following asm loop
|
3512
|
374 for (m=3; m < 7; m++) {
|
|
375 two_m = (1 << m);
|
|
376
|
|
377 two_m_plus_one = two_m<<1;
|
3508
|
378
|
3537
|
379 for(i = 0; i < 128; i += two_m_plus_one) {
|
|
380 for(k = 0; k < two_m; k++) {
|
|
381 int p = k + i;
|
|
382 int q = p + two_m;
|
|
383 tmp_a_r = buf[p].real;
|
|
384 tmp_a_i = buf[p].imag;
|
|
385 tmp_b_r = buf[q].real * w[m][k].real - buf[q].imag * w[m][k].imag;
|
|
386 tmp_b_i = buf[q].imag * w[m][k].real + buf[q].real * w[m][k].imag;
|
|
387 buf[p].real = tmp_a_r + tmp_b_r;
|
|
388 buf[p].imag = tmp_a_i + tmp_b_i;
|
|
389 buf[q].real = tmp_a_r - tmp_b_r;
|
|
390 buf[q].imag = tmp_a_i - tmp_b_i;
|
|
391 }
|
|
392 }
|
|
393 }
|
3546
|
394 */
|
|
395 for (m=3; m < 7; m++) {
|
|
396 two_m = (1 << m);
|
|
397 two_m_plus_one = two_m<<1;
|
|
398 asm volatile(
|
|
399 "movl %0, %%esi \n\t"
|
|
400 ".balign 16 \n\t"
|
|
401 "1: \n\t"
|
|
402 "xorl %%edi, %%edi \n\t" // k
|
|
403 "leal (%%esi, %3), %%edx \n\t"
|
|
404 "2: \n\t"
|
|
405 "movaps (%%edx, %%edi), %%xmm1 \n\t"
|
|
406 "movaps (%4, %%edi, 2), %%xmm2 \n\t"
|
|
407 "mulps %%xmm1, %%xmm2 \n\t"
|
|
408 "shufps $0xB1, %%xmm1, %%xmm1 \n\t"
|
|
409 "mulps 16(%4, %%edi, 2), %%xmm1 \n\t"
|
|
410 "movaps (%%esi, %%edi), %%xmm0 \n\t"
|
|
411 "addps %%xmm2, %%xmm1 \n\t"
|
|
412 "movaps %%xmm1, %%xmm2 \n\t"
|
|
413 "addps %%xmm0, %%xmm1 \n\t"
|
|
414 "subps %%xmm2, %%xmm0 \n\t"
|
|
415 "movaps %%xmm1, (%%esi, %%edi) \n\t"
|
|
416 "movaps %%xmm0, (%%edx, %%edi) \n\t"
|
|
417 "addl $16, %%edi \n\t"
|
|
418 "cmpl %3, %%edi \n\t" //FIXME (opt) count against 0
|
|
419 " jb 2b \n\t"
|
|
420 "addl %2, %%esi \n\t"
|
|
421 "cmpl %1, %%esi \n\t"
|
|
422 " jb 1b \n\t"
|
|
423 :: "g" (buf), "m" (buf+128), "m" (two_m_plus_one<<3), "r" (two_m<<3),
|
|
424 "r" (sseW[m])
|
|
425 : "%esi", "%edi", "%edx"
|
|
426 );
|
|
427 }
|
|
428
|
3537
|
429 #else
|
|
430 /* unoptimized variant
|
|
431 for (m=1; m < 7; m++) {
|
|
432 if(m)
|
|
433 two_m = (1 << m);
|
|
434 else
|
|
435 two_m = 1;
|
|
436
|
|
437 two_m_plus_one = (1 << (m+1));
|
|
438
|
|
439 for(i = 0; i < 128; i += two_m_plus_one) {
|
|
440 for(k = 0; k < two_m; k++) {
|
3508
|
441 p = k + i;
|
|
442 q = p + two_m;
|
|
443 tmp_a_r = buf[p].real;
|
|
444 tmp_a_i = buf[p].imag;
|
|
445 tmp_b_r = buf[q].real * w[m][k].real - buf[q].imag * w[m][k].imag;
|
|
446 tmp_b_i = buf[q].imag * w[m][k].real + buf[q].real * w[m][k].imag;
|
|
447 buf[p].real = tmp_a_r + tmp_b_r;
|
|
448 buf[p].imag = tmp_a_i + tmp_b_i;
|
|
449 buf[q].real = tmp_a_r - tmp_b_r;
|
|
450 buf[q].imag = tmp_a_i - tmp_b_i;
|
|
451 }
|
|
452 }
|
|
453 }
|
3537
|
454 */
|
|
455
|
|
456 for(i = 0; i < 128; i += 2) {
|
|
457 tmp_a_r = buf[i].real;
|
|
458 tmp_a_i = buf[i].imag;
|
|
459 tmp_b_r = buf[i+1].real;
|
|
460 tmp_b_i = buf[i+1].imag;
|
|
461 buf[i].real = tmp_a_r + tmp_b_r;
|
|
462 buf[i].imag = tmp_a_i + tmp_b_i;
|
|
463 buf[i+1].real = tmp_a_r - tmp_b_r;
|
|
464 buf[i+1].imag = tmp_a_i - tmp_b_i;
|
|
465 }
|
|
466
|
|
467 for(i = 0; i < 128; i += 4) {
|
|
468 tmp_a_r = buf[i].real;
|
|
469 tmp_a_i = buf[i].imag;
|
|
470 tmp_b_r = buf[i+2].real;
|
|
471 tmp_b_i = buf[i+2].imag;
|
|
472 buf[i].real = tmp_a_r + tmp_b_r;
|
|
473 buf[i].imag = tmp_a_i + tmp_b_i;
|
|
474 buf[i+2].real = tmp_a_r - tmp_b_r;
|
|
475 buf[i+2].imag = tmp_a_i - tmp_b_i;
|
|
476 tmp_a_r = buf[i+1].real;
|
|
477 tmp_a_i = buf[i+1].imag;
|
|
478 tmp_b_r = buf[i+3].imag;
|
|
479 tmp_b_i = buf[i+3].real;
|
|
480 buf[i+1].real = tmp_a_r + tmp_b_r;
|
|
481 buf[i+1].imag = tmp_a_i - tmp_b_i;
|
|
482 buf[i+3].real = tmp_a_r - tmp_b_r;
|
|
483 buf[i+3].imag = tmp_a_i + tmp_b_i;
|
|
484 }
|
|
485
|
|
486 for(i = 0; i < 128; i += 8) {
|
|
487 tmp_a_r = buf[i].real;
|
|
488 tmp_a_i = buf[i].imag;
|
|
489 tmp_b_r = buf[i+4].real;
|
|
490 tmp_b_i = buf[i+4].imag;
|
|
491 buf[i].real = tmp_a_r + tmp_b_r;
|
|
492 buf[i].imag = tmp_a_i + tmp_b_i;
|
|
493 buf[i+4].real = tmp_a_r - tmp_b_r;
|
|
494 buf[i+4].imag = tmp_a_i - tmp_b_i;
|
|
495 tmp_a_r = buf[1+i].real;
|
|
496 tmp_a_i = buf[1+i].imag;
|
|
497 tmp_b_r = (buf[i+5].real + buf[i+5].imag) * w[2][1].real;
|
|
498 tmp_b_i = (buf[i+5].imag - buf[i+5].real) * w[2][1].real;
|
|
499 buf[1+i].real = tmp_a_r + tmp_b_r;
|
|
500 buf[1+i].imag = tmp_a_i + tmp_b_i;
|
|
501 buf[i+5].real = tmp_a_r - tmp_b_r;
|
|
502 buf[i+5].imag = tmp_a_i - tmp_b_i;
|
|
503 tmp_a_r = buf[i+2].real;
|
|
504 tmp_a_i = buf[i+2].imag;
|
|
505 tmp_b_r = buf[i+6].imag;
|
|
506 tmp_b_i = - buf[i+6].real;
|
|
507 buf[i+2].real = tmp_a_r + tmp_b_r;
|
|
508 buf[i+2].imag = tmp_a_i + tmp_b_i;
|
|
509 buf[i+6].real = tmp_a_r - tmp_b_r;
|
|
510 buf[i+6].imag = tmp_a_i - tmp_b_i;
|
|
511 tmp_a_r = buf[i+3].real;
|
|
512 tmp_a_i = buf[i+3].imag;
|
|
513 tmp_b_r = (buf[i+7].real - buf[i+7].imag) * w[2][3].imag;
|
|
514 tmp_b_i = (buf[i+7].imag + buf[i+7].real) * w[2][3].imag;
|
|
515 buf[i+3].real = tmp_a_r + tmp_b_r;
|
|
516 buf[i+3].imag = tmp_a_i + tmp_b_i;
|
|
517 buf[i+7].real = tmp_a_r - tmp_b_r;
|
|
518 buf[i+7].imag = tmp_a_i - tmp_b_i;
|
|
519 }
|
|
520
|
|
521 for (m=3; m < 7; m++) {
|
|
522 two_m = (1 << m);
|
3394
|
523
|
3537
|
524 two_m_plus_one = two_m<<1;
|
3394
|
525
|
3537
|
526 for(i = 0; i < 128; i += two_m_plus_one) {
|
|
527 for(k = 0; k < two_m; k++) {
|
|
528 int p = k + i;
|
|
529 int q = p + two_m;
|
3394
|
530 tmp_a_r = buf[p].real;
|
|
531 tmp_a_i = buf[p].imag;
|
|
532 tmp_b_r = buf[q].real * w[m][k].real - buf[q].imag * w[m][k].imag;
|
|
533 tmp_b_i = buf[q].imag * w[m][k].real + buf[q].real * w[m][k].imag;
|
|
534 buf[p].real = tmp_a_r + tmp_b_r;
|
|
535 buf[p].imag = tmp_a_i + tmp_b_i;
|
|
536 buf[q].real = tmp_a_r - tmp_b_r;
|
|
537 buf[q].imag = tmp_a_i - tmp_b_i;
|
|
538 }
|
|
539 }
|
|
540 }
|
3537
|
541
|
3508
|
542 #endif
|
|
543
|
3394
|
544 /* Post IFFT complex multiply plus IFFT complex conjugate*/
|
|
545 for( i=0; i < 128; i++) {
|
|
546 /* y[n] = z[n] * (xcos1[n] + j * xsin1[n]) ; */
|
|
547 tmp_a_r = buf[i].real;
|
|
548 tmp_a_i = -1.0 * buf[i].imag;
|
|
549 buf[i].real =(tmp_a_r * xcos1[i]) - (tmp_a_i * xsin1[i]);
|
|
550 buf[i].imag =(tmp_a_r * xsin1[i]) + (tmp_a_i * xcos1[i]);
|
|
551 }
|
|
552
|
|
553 data_ptr = data;
|
|
554 delay_ptr = delay;
|
|
555 window_ptr = imdct_window;
|
|
556
|
|
557 /* Window and convert to real valued signal */
|
|
558 for(i=0; i< 64; i++) {
|
|
559 *data_ptr++ = -buf[64+i].imag * *window_ptr++ + *delay_ptr++ + bias;
|
|
560 *data_ptr++ = buf[64-i-1].real * *window_ptr++ + *delay_ptr++ + bias;
|
|
561 }
|
|
562
|
|
563 for(i=0; i< 64; i++) {
|
|
564 *data_ptr++ = -buf[i].real * *window_ptr++ + *delay_ptr++ + bias;
|
|
565 *data_ptr++ = buf[128-i-1].imag * *window_ptr++ + *delay_ptr++ + bias;
|
|
566 }
|
|
567
|
|
568 /* The trailing edge of the window goes into the delay line */
|
|
569 delay_ptr = delay;
|
|
570
|
|
571 for(i=0; i< 64; i++) {
|
|
572 *delay_ptr++ = -buf[64+i].real * *--window_ptr;
|
|
573 *delay_ptr++ = buf[64-i-1].imag * *--window_ptr;
|
|
574 }
|
|
575
|
|
576 for(i=0; i<64; i++) {
|
|
577 *delay_ptr++ = buf[i].imag * *--window_ptr;
|
|
578 *delay_ptr++ = -buf[128-i-1].real * *--window_ptr;
|
|
579 }
|
|
580 }
|
|
581
|
|
582 void
|
|
583 imdct_do_256(sample_t data[],sample_t delay[],sample_t bias)
|
|
584 {
|
|
585 int i,k;
|
|
586 int p,q;
|
|
587 int m;
|
|
588 int two_m;
|
|
589 int two_m_plus_one;
|
|
590
|
|
591 sample_t tmp_a_i;
|
|
592 sample_t tmp_a_r;
|
|
593 sample_t tmp_b_i;
|
|
594 sample_t tmp_b_r;
|
|
595
|
|
596 sample_t *data_ptr;
|
|
597 sample_t *delay_ptr;
|
|
598 sample_t *window_ptr;
|
|
599
|
|
600 complex_t *buf_1, *buf_2;
|
|
601
|
|
602 buf_1 = &buf[0];
|
|
603 buf_2 = &buf[64];
|
|
604
|
|
605 /* Pre IFFT complex multiply plus IFFT cmplx conjugate */
|
|
606 for(k=0; k<64; k++) {
|
|
607 /* X1[k] = X[2*k] */
|
|
608 /* X2[k] = X[2*k+1] */
|
|
609
|
|
610 p = 2 * (128-2*k-1);
|
|
611 q = 2 * (2 * k);
|
|
612
|
|
613 /* Z1[k] = (X1[128-2*k-1] + j * X1[2*k]) * (xcos2[k] + j * xsin2[k]); */
|
|
614 buf_1[k].real = data[p] * xcos2[k] - data[q] * xsin2[k];
|
|
615 buf_1[k].imag = -1.0f * (data[q] * xcos2[k] + data[p] * xsin2[k]);
|
|
616 /* Z2[k] = (X2[128-2*k-1] + j * X2[2*k]) * (xcos2[k] + j * xsin2[k]); */
|
|
617 buf_2[k].real = data[p + 1] * xcos2[k] - data[q + 1] * xsin2[k];
|
|
618 buf_2[k].imag = -1.0f * ( data[q + 1] * xcos2[k] + data[p + 1] * xsin2[k]);
|
|
619 }
|
|
620
|
|
621 /* IFFT Bit reversed shuffling */
|
|
622 for(i=0; i<64; i++) {
|
|
623 k = bit_reverse_256[i];
|
|
624 if (k < i) {
|
|
625 swap_cmplx(&buf_1[i],&buf_1[k]);
|
|
626 swap_cmplx(&buf_2[i],&buf_2[k]);
|
|
627 }
|
|
628 }
|
|
629
|
|
630 /* FFT Merge */
|
|
631 for (m=0; m < 6; m++) {
|
|
632 two_m = (1 << m);
|
|
633 two_m_plus_one = (1 << (m+1));
|
|
634
|
|
635 /* FIXME */
|
|
636 if(m)
|
|
637 two_m = (1 << m);
|
|
638 else
|
|
639 two_m = 1;
|
|
640
|
|
641 for(k = 0; k < two_m; k++) {
|
|
642 for(i = 0; i < 64; i += two_m_plus_one) {
|
|
643 p = k + i;
|
|
644 q = p + two_m;
|
|
645 /* Do block 1 */
|
|
646 tmp_a_r = buf_1[p].real;
|
|
647 tmp_a_i = buf_1[p].imag;
|
|
648 tmp_b_r = buf_1[q].real * w[m][k].real - buf_1[q].imag * w[m][k].imag;
|
|
649 tmp_b_i = buf_1[q].imag * w[m][k].real + buf_1[q].real * w[m][k].imag;
|
|
650 buf_1[p].real = tmp_a_r + tmp_b_r;
|
|
651 buf_1[p].imag = tmp_a_i + tmp_b_i;
|
|
652 buf_1[q].real = tmp_a_r - tmp_b_r;
|
|
653 buf_1[q].imag = tmp_a_i - tmp_b_i;
|
|
654
|
|
655 /* Do block 2 */
|
|
656 tmp_a_r = buf_2[p].real;
|
|
657 tmp_a_i = buf_2[p].imag;
|
|
658 tmp_b_r = buf_2[q].real * w[m][k].real - buf_2[q].imag * w[m][k].imag;
|
|
659 tmp_b_i = buf_2[q].imag * w[m][k].real + buf_2[q].real * w[m][k].imag;
|
|
660 buf_2[p].real = tmp_a_r + tmp_b_r;
|
|
661 buf_2[p].imag = tmp_a_i + tmp_b_i;
|
|
662 buf_2[q].real = tmp_a_r - tmp_b_r;
|
|
663 buf_2[q].imag = tmp_a_i - tmp_b_i;
|
|
664 }
|
|
665 }
|
|
666 }
|
|
667
|
|
668 /* Post IFFT complex multiply */
|
|
669 for( i=0; i < 64; i++) {
|
|
670 /* y1[n] = z1[n] * (xcos2[n] + j * xs in2[n]) ; */
|
|
671 tmp_a_r = buf_1[i].real;
|
|
672 tmp_a_i = -buf_1[i].imag;
|
|
673 buf_1[i].real =(tmp_a_r * xcos2[i]) - (tmp_a_i * xsin2[i]);
|
|
674 buf_1[i].imag =(tmp_a_r * xsin2[i]) + (tmp_a_i * xcos2[i]);
|
|
675 /* y2[n] = z2[n] * (xcos2[n] + j * xsin2[n]) ; */
|
|
676 tmp_a_r = buf_2[i].real;
|
|
677 tmp_a_i = -buf_2[i].imag;
|
|
678 buf_2[i].real =(tmp_a_r * xcos2[i]) - (tmp_a_i * xsin2[i]);
|
|
679 buf_2[i].imag =(tmp_a_r * xsin2[i]) + (tmp_a_i * xcos2[i]);
|
|
680 }
|
|
681
|
|
682 data_ptr = data;
|
|
683 delay_ptr = delay;
|
|
684 window_ptr = imdct_window;
|
|
685
|
|
686 /* Window and convert to real valued signal */
|
|
687 for(i=0; i< 64; i++) {
|
|
688 *data_ptr++ = -buf_1[i].imag * *window_ptr++ + *delay_ptr++ + bias;
|
|
689 *data_ptr++ = buf_1[64-i-1].real * *window_ptr++ + *delay_ptr++ + bias;
|
|
690 }
|
|
691
|
|
692 for(i=0; i< 64; i++) {
|
|
693 *data_ptr++ = -buf_1[i].real * *window_ptr++ + *delay_ptr++ + bias;
|
|
694 *data_ptr++ = buf_1[64-i-1].imag * *window_ptr++ + *delay_ptr++ + bias;
|
|
695 }
|
|
696
|
|
697 delay_ptr = delay;
|
|
698
|
|
699 for(i=0; i< 64; i++) {
|
|
700 *delay_ptr++ = -buf_2[i].real * *--window_ptr;
|
|
701 *delay_ptr++ = buf_2[64-i-1].imag * *--window_ptr;
|
|
702 }
|
|
703
|
|
704 for(i=0; i< 64; i++) {
|
|
705 *delay_ptr++ = buf_2[i].imag * *--window_ptr;
|
|
706 *delay_ptr++ = -buf_2[64-i-1].real * *--window_ptr;
|
|
707 }
|
|
708 }
|
|
709
|
|
710 void imdct_init (uint32_t mm_accel)
|
|
711 {
|
|
712 #ifdef LIBA52_MLIB
|
|
713 if (mm_accel & MM_ACCEL_MLIB) {
|
|
714 fprintf (stderr, "Using mlib for IMDCT transform\n");
|
|
715 imdct_512 = imdct_do_512_mlib;
|
|
716 imdct_256 = imdct_do_256_mlib;
|
|
717 } else
|
|
718 #endif
|
|
719 {
|
|
720 int i, j, k;
|
|
721
|
|
722 fprintf (stderr, "No accelerated IMDCT transform found\n");
|
|
723
|
|
724 /* Twiddle factors to turn IFFT into IMDCT */
|
|
725 for (i = 0; i < 128; i++) {
|
|
726 xcos1[i] = -cos ((M_PI / 2048) * (8 * i + 1));
|
|
727 xsin1[i] = -sin ((M_PI / 2048) * (8 * i + 1));
|
|
728 }
|
3527
|
729 #ifdef HAVE_SSE
|
|
730 for (i = 0; i < 128; i++) {
|
|
731 sseSinCos1a[2*i+0]= -xsin1[i];
|
|
732 sseSinCos1a[2*i+1]= -xcos1[i];
|
|
733 sseSinCos1b[2*i+0]= xcos1[i];
|
|
734 sseSinCos1b[2*i+1]= -xsin1[i];
|
|
735 }
|
|
736 #endif
|
3394
|
737
|
|
738 /* More twiddle factors to turn IFFT into IMDCT */
|
|
739 for (i = 0; i < 64; i++) {
|
|
740 xcos2[i] = -cos ((M_PI / 1024) * (8 * i + 1));
|
|
741 xsin2[i] = -sin ((M_PI / 1024) * (8 * i + 1));
|
|
742 }
|
|
743
|
|
744 for (i = 0; i < 7; i++) {
|
|
745 j = 1 << i;
|
|
746 for (k = 0; k < j; k++) {
|
|
747 w[i][k].real = cos (-M_PI * k / j);
|
|
748 w[i][k].imag = sin (-M_PI * k / j);
|
|
749 }
|
|
750 }
|
3534
|
751 #ifdef HAVE_SSE
|
|
752 for (i = 1; i < 7; i++) {
|
|
753 j = 1 << i;
|
|
754 for (k = 0; k < j; k+=2) {
|
|
755
|
|
756 sseW[i][4*k + 0] = w[i][k+0].real;
|
|
757 sseW[i][4*k + 1] = w[i][k+0].real;
|
|
758 sseW[i][4*k + 2] = w[i][k+1].real;
|
|
759 sseW[i][4*k + 3] = w[i][k+1].real;
|
|
760
|
|
761 sseW[i][4*k + 4] = -w[i][k+0].imag;
|
|
762 sseW[i][4*k + 5] = w[i][k+0].imag;
|
|
763 sseW[i][4*k + 6] = -w[i][k+1].imag;
|
|
764 sseW[i][4*k + 7] = w[i][k+1].imag;
|
|
765
|
|
766 //we multiply more or less uninitalized numbers so we need to use exactly 0.0
|
|
767 if(k==0)
|
|
768 {
|
|
769 // sseW[i][4*k + 0]= sseW[i][4*k + 1]= 1.0;
|
|
770 sseW[i][4*k + 4]= sseW[i][4*k + 5]= 0.0;
|
|
771 }
|
|
772
|
|
773 if(2*k == j)
|
|
774 {
|
|
775 sseW[i][4*k + 0]= sseW[i][4*k + 1]= 0.0;
|
|
776 // sseW[i][4*k + 4]= -(sseW[i][4*k + 5]= -1.0);
|
|
777 }
|
|
778 }
|
|
779 }
|
|
780 #endif
|
|
781
|
3394
|
782 imdct_512 = imdct_do_512;
|
|
783 imdct_256 = imdct_do_256;
|
|
784 }
|
|
785 }
|