?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
typo fix
[mplayer/greg.git] / liba52 / imdct.c
bloba8a5815cb1eb773d216abe2201dcfc1d9951d0f8
1 /*
2 * imdct.c
3 * Copyright (C) 2000-2002 Michel Lespinasse <walken@zoy.org>
4 * Copyright (C) 1999-2000 Aaron Holtzman <aholtzma@ess.engr.uvic.ca>
5 *
6 * The ifft algorithms in this file have been largely inspired by Dan
7 * Bernstein's work, djbfft, available at http://cr.yp.to/djbfft.html
8 *
9 * This file is part of a52dec, a free ATSC A-52 stream decoder.
10 * See http://liba52.sourceforge.net/ for updates.
11 *
12 * Modified for use with MPlayer, changes contained in liba52_changes.diff.
13 * detailed changelog at http://svn.mplayerhq.hu/mplayer/trunk/
14 * $Id$
15 *
16 * a52dec is free software; you can redistribute it and/or modify
17 * it under the terms of the GNU General Public License as published by
18 * the Free Software Foundation; either version 2 of the License, or
19 * (at your option) any later version.
20 *
21 * a52dec is distributed in the hope that it will be useful,
22 * but WITHOUT ANY WARRANTY; without even the implied warranty of
23 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
24 * GNU General Public License for more details.
25 *
26 * You should have received a copy of the GNU General Public License
27 * along with this program; if not, write to the Free Software
28 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
29 *
30 * SSE optimizations from Michael Niedermayer (michaelni@gmx.at)
31 * 3DNOW optimizations from Nick Kurshev <nickols_k@mail.ru>
32 * michael did port them from libac3 (untested, perhaps totally broken)
33 * AltiVec optimizations from Romain Dolbeau (romain@dolbeau.org)
34 */
35
36 #include "config.h"
37
38 #include <math.h>
39 #include <stdio.h>
40 #ifdef LIBA52_DJBFFT
41 #include <fftc4.h>
42 #endif
43 #ifndef M_PI
44 #define M_PI 3.1415926535897932384626433832795029
45 #endif
46 #include <inttypes.h>
47
48 #include "a52.h"
49 #include "a52_internal.h"
50 #include "mm_accel.h"
51 #include "mangle.h"
52
53 void (*a52_imdct_512) (sample_t * data, sample_t * delay, sample_t bias);
54
55 #ifdef RUNTIME_CPUDETECT
56 #undef HAVE_3DNOWEX
57 #endif
58
59 typedef struct complex_s {
60 sample_t real;
61 sample_t imag;
62 } complex_t;
63
64 static const int pm128[128] attribute_used __attribute__((aligned(16))) =
65 {
66 0, 16, 32, 48, 64, 80, 96, 112, 8, 40, 72, 104, 24, 56, 88, 120,
67 4, 20, 36, 52, 68, 84, 100, 116, 12, 28, 44, 60, 76, 92, 108, 124,
68 2, 18, 34, 50, 66, 82, 98, 114, 10, 42, 74, 106, 26, 58, 90, 122,
69 6, 22, 38, 54, 70, 86, 102, 118, 14, 46, 78, 110, 30, 62, 94, 126,
70 1, 17, 33, 49, 65, 81, 97, 113, 9, 41, 73, 105, 25, 57, 89, 121,
71 5, 21, 37, 53, 69, 85, 101, 117, 13, 29, 45, 61, 77, 93, 109, 125,
72 3, 19, 35, 51, 67, 83, 99, 115, 11, 43, 75, 107, 27, 59, 91, 123,
73 7, 23, 39, 55, 71, 87, 103, 119, 15, 31, 47, 63, 79, 95, 111, 127
74 };
75
76 static uint8_t attribute_used bit_reverse_512[] = {
77 0x00, 0x40, 0x20, 0x60, 0x10, 0x50, 0x30, 0x70,
78 0x08, 0x48, 0x28, 0x68, 0x18, 0x58, 0x38, 0x78,
79 0x04, 0x44, 0x24, 0x64, 0x14, 0x54, 0x34, 0x74,
80 0x0c, 0x4c, 0x2c, 0x6c, 0x1c, 0x5c, 0x3c, 0x7c,
81 0x02, 0x42, 0x22, 0x62, 0x12, 0x52, 0x32, 0x72,
82 0x0a, 0x4a, 0x2a, 0x6a, 0x1a, 0x5a, 0x3a, 0x7a,
83 0x06, 0x46, 0x26, 0x66, 0x16, 0x56, 0x36, 0x76,
84 0x0e, 0x4e, 0x2e, 0x6e, 0x1e, 0x5e, 0x3e, 0x7e,
85 0x01, 0x41, 0x21, 0x61, 0x11, 0x51, 0x31, 0x71,
86 0x09, 0x49, 0x29, 0x69, 0x19, 0x59, 0x39, 0x79,
87 0x05, 0x45, 0x25, 0x65, 0x15, 0x55, 0x35, 0x75,
88 0x0d, 0x4d, 0x2d, 0x6d, 0x1d, 0x5d, 0x3d, 0x7d,
89 0x03, 0x43, 0x23, 0x63, 0x13, 0x53, 0x33, 0x73,
90 0x0b, 0x4b, 0x2b, 0x6b, 0x1b, 0x5b, 0x3b, 0x7b,
91 0x07, 0x47, 0x27, 0x67, 0x17, 0x57, 0x37, 0x77,
92 0x0f, 0x4f, 0x2f, 0x6f, 0x1f, 0x5f, 0x3f, 0x7f};
93
94 static uint8_t fftorder[] = {
95 0,128, 64,192, 32,160,224, 96, 16,144, 80,208,240,112, 48,176,
96 8,136, 72,200, 40,168,232,104,248,120, 56,184, 24,152,216, 88,
97 4,132, 68,196, 36,164,228,100, 20,148, 84,212,244,116, 52,180,
98 252,124, 60,188, 28,156,220, 92, 12,140, 76,204,236,108, 44,172,
99 2,130, 66,194, 34,162,226, 98, 18,146, 82,210,242,114, 50,178,
100 10,138, 74,202, 42,170,234,106,250,122, 58,186, 26,154,218, 90,
101 254,126, 62,190, 30,158,222, 94, 14,142, 78,206,238,110, 46,174,
102 6,134, 70,198, 38,166,230,102,246,118, 54,182, 22,150,214, 86
103 };
104
105 static complex_t __attribute__((aligned(16))) buf[128];
106
107 /* Twiddle factor LUT */
108 static complex_t __attribute__((aligned(16))) w_1[1];
109 static complex_t __attribute__((aligned(16))) w_2[2];
110 static complex_t __attribute__((aligned(16))) w_4[4];
111 static complex_t __attribute__((aligned(16))) w_8[8];
112 static complex_t __attribute__((aligned(16))) w_16[16];
113 static complex_t __attribute__((aligned(16))) w_32[32];
114 static complex_t __attribute__((aligned(16))) w_64[64];
115 static complex_t __attribute__((aligned(16))) * w[7] = {w_1, w_2, w_4, w_8, w_16, w_32, w_64};
116
117 /* Twiddle factors for IMDCT */
118 static sample_t __attribute__((aligned(16))) xcos1[128];
119 static sample_t __attribute__((aligned(16))) xsin1[128];
120
121 #if defined(ARCH_X86) || defined(ARCH_X86_64)
122 // NOTE: SSE needs 16byte alignment or it will segfault
123 //
124 static float __attribute__((aligned(16))) sseSinCos1c[256];
125 static float __attribute__((aligned(16))) sseSinCos1d[256];
126 static float attribute_used __attribute__((aligned(16))) ps111_1[4]={1,1,1,-1};
127 //static float __attribute__((aligned(16))) sseW0[4];
128 static float __attribute__((aligned(16))) sseW1[8];
129 static float __attribute__((aligned(16))) sseW2[16];
130 static float __attribute__((aligned(16))) sseW3[32];
131 static float __attribute__((aligned(16))) sseW4[64];
132 static float __attribute__((aligned(16))) sseW5[128];
133 static float __attribute__((aligned(16))) sseW6[256];
134 static float __attribute__((aligned(16))) *sseW[7]=
135 {NULL /*sseW0*/,sseW1,sseW2,sseW3,sseW4,sseW5,sseW6};
136 static float __attribute__((aligned(16))) sseWindow[512];
137 #endif
138
139 /* Root values for IFFT */
140 static sample_t roots16[3];
141 static sample_t roots32[7];
142 static sample_t roots64[15];
143 static sample_t roots128[31];
144
145 /* Twiddle factors for IMDCT */
146 static complex_t pre1[128];
147 static complex_t post1[64];
148 static complex_t pre2[64];
149 static complex_t post2[32];
150
151 static sample_t a52_imdct_window[256];
152
153 static void (* ifft128) (complex_t * buf);
154 static void (* ifft64) (complex_t * buf);
155
156 static inline void ifft2 (complex_t * buf)
157 {
158 double r, i;
159
160 r = buf[0].real;
161 i = buf[0].imag;
162 buf[0].real += buf[1].real;
163 buf[0].imag += buf[1].imag;
164 buf[1].real = r - buf[1].real;
165 buf[1].imag = i - buf[1].imag;
166 }
167
168 static inline void ifft4 (complex_t * buf)
169 {
170 double tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp8;
171
172 tmp1 = buf[0].real + buf[1].real;
173 tmp2 = buf[3].real + buf[2].real;
174 tmp3 = buf[0].imag + buf[1].imag;
175 tmp4 = buf[2].imag + buf[3].imag;
176 tmp5 = buf[0].real - buf[1].real;
177 tmp6 = buf[0].imag - buf[1].imag;
178 tmp7 = buf[2].imag - buf[3].imag;
179 tmp8 = buf[3].real - buf[2].real;
180
181 buf[0].real = tmp1 + tmp2;
182 buf[0].imag = tmp3 + tmp4;
183 buf[2].real = tmp1 - tmp2;
184 buf[2].imag = tmp3 - tmp4;
185 buf[1].real = tmp5 + tmp7;
186 buf[1].imag = tmp6 + tmp8;
187 buf[3].real = tmp5 - tmp7;
188 buf[3].imag = tmp6 - tmp8;
189 }
190
191 /* the basic split-radix ifft butterfly */
192
193 #define BUTTERFLY(a0,a1,a2,a3,wr,wi) do { \
194 tmp5 = a2.real * wr + a2.imag * wi; \
195 tmp6 = a2.imag * wr - a2.real * wi; \
196 tmp7 = a3.real * wr - a3.imag * wi; \
197 tmp8 = a3.imag * wr + a3.real * wi; \
198 tmp1 = tmp5 + tmp7; \
199 tmp2 = tmp6 + tmp8; \
200 tmp3 = tmp6 - tmp8; \
201 tmp4 = tmp7 - tmp5; \
202 a2.real = a0.real - tmp1; \
203 a2.imag = a0.imag - tmp2; \
204 a3.real = a1.real - tmp3; \
205 a3.imag = a1.imag - tmp4; \
206 a0.real += tmp1; \
207 a0.imag += tmp2; \
208 a1.real += tmp3; \
209 a1.imag += tmp4; \
210 } while (0)
211
212 /* split-radix ifft butterfly, specialized for wr=1 wi=0 */
213
214 #define BUTTERFLY_ZERO(a0,a1,a2,a3) do { \
215 tmp1 = a2.real + a3.real; \
216 tmp2 = a2.imag + a3.imag; \
217 tmp3 = a2.imag - a3.imag; \
218 tmp4 = a3.real - a2.real; \
219 a2.real = a0.real - tmp1; \
220 a2.imag = a0.imag - tmp2; \
221 a3.real = a1.real - tmp3; \
222 a3.imag = a1.imag - tmp4; \
223 a0.real += tmp1; \
224 a0.imag += tmp2; \
225 a1.real += tmp3; \
226 a1.imag += tmp4; \
227 } while (0)
228
229 /* split-radix ifft butterfly, specialized for wr=wi */
230
231 #define BUTTERFLY_HALF(a0,a1,a2,a3,w) do { \
232 tmp5 = (a2.real + a2.imag) * w; \
233 tmp6 = (a2.imag - a2.real) * w; \
234 tmp7 = (a3.real - a3.imag) * w; \
235 tmp8 = (a3.imag + a3.real) * w; \
236 tmp1 = tmp5 + tmp7; \
237 tmp2 = tmp6 + tmp8; \
238 tmp3 = tmp6 - tmp8; \
239 tmp4 = tmp7 - tmp5; \
240 a2.real = a0.real - tmp1; \
241 a2.imag = a0.imag - tmp2; \
242 a3.real = a1.real - tmp3; \
243 a3.imag = a1.imag - tmp4; \
244 a0.real += tmp1; \
245 a0.imag += tmp2; \
246 a1.real += tmp3; \
247 a1.imag += tmp4; \
248 } while (0)
249
250 static inline void ifft8 (complex_t * buf)
251 {
252 double tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp8;
253
254 ifft4 (buf);
255 ifft2 (buf + 4);
256 ifft2 (buf + 6);
257 BUTTERFLY_ZERO (buf[0], buf[2], buf[4], buf[6]);
258 BUTTERFLY_HALF (buf[1], buf[3], buf[5], buf[7], roots16[1]);
259 }
260
261 static void ifft_pass (complex_t * buf, sample_t * weight, int n)
262 {
263 complex_t * buf1;
264 complex_t * buf2;
265 complex_t * buf3;
266 double tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp8;
267 int i;
268
269 buf++;
270 buf1 = buf + n;
271 buf2 = buf + 2 * n;
272 buf3 = buf + 3 * n;
273
274 BUTTERFLY_ZERO (buf[-1], buf1[-1], buf2[-1], buf3[-1]);
275
276 i = n - 1;
277
278 do {
279 BUTTERFLY (buf[0], buf1[0], buf2[0], buf3[0], weight[n], weight[2*i]);
280 buf++;
281 buf1++;
282 buf2++;
283 buf3++;
284 weight++;
285 } while (--i);
286 }
287
288 static void ifft16 (complex_t * buf)
289 {
290 ifft8 (buf);
291 ifft4 (buf + 8);
292 ifft4 (buf + 12);
293 ifft_pass (buf, roots16 - 4, 4);
294 }
295
296 static void ifft32 (complex_t * buf)
297 {
298 ifft16 (buf);
299 ifft8 (buf + 16);
300 ifft8 (buf + 24);
301 ifft_pass (buf, roots32 - 8, 8);
302 }
303
304 static void ifft64_c (complex_t * buf)
305 {
306 ifft32 (buf);
307 ifft16 (buf + 32);
308 ifft16 (buf + 48);
309 ifft_pass (buf, roots64 - 16, 16);
310 }
311
312 static void ifft128_c (complex_t * buf)
313 {
314 ifft32 (buf);
315 ifft16 (buf + 32);
316 ifft16 (buf + 48);
317 ifft_pass (buf, roots64 - 16, 16);
318
319 ifft32 (buf + 64);
320 ifft32 (buf + 96);
321 ifft_pass (buf, roots128 - 32, 32);
322 }
323
324 void imdct_do_512 (sample_t * data, sample_t * delay, sample_t bias)
325 {
326 int i, k;
327 sample_t t_r, t_i, a_r, a_i, b_r, b_i, w_1, w_2;
328 const sample_t * window = a52_imdct_window;
329 complex_t buf[128];
330
331 for (i = 0; i < 128; i++) {
332 k = fftorder[i];
333 t_r = pre1[i].real;
334 t_i = pre1[i].imag;
335
336 buf[i].real = t_i * data[255-k] + t_r * data[k];
337 buf[i].imag = t_r * data[255-k] - t_i * data[k];
338 }
339
340 ifft128 (buf);
341
342 /* Post IFFT complex multiply plus IFFT complex conjugate*/
343 /* Window and convert to real valued signal */
344 for (i = 0; i < 64; i++) {
345 /* y[n] = z[n] * (xcos1[n] + j * xsin1[n]) ; */
346 t_r = post1[i].real;
347 t_i = post1[i].imag;
348
349 a_r = t_r * buf[i].real + t_i * buf[i].imag;
350 a_i = t_i * buf[i].real - t_r * buf[i].imag;
351 b_r = t_i * buf[127-i].real + t_r * buf[127-i].imag;
352 b_i = t_r * buf[127-i].real - t_i * buf[127-i].imag;
353
354 w_1 = window[2*i];
355 w_2 = window[255-2*i];
356 data[2*i] = delay[2*i] * w_2 - a_r * w_1 + bias;
357 data[255-2*i] = delay[2*i] * w_1 + a_r * w_2 + bias;
358 delay[2*i] = a_i;
359
360 w_1 = window[2*i+1];
361 w_2 = window[254-2*i];
362 data[2*i+1] = delay[2*i+1] * w_2 + b_r * w_1 + bias;
363 data[254-2*i] = delay[2*i+1] * w_1 - b_r * w_2 + bias;
364 delay[2*i+1] = b_i;
365 }
366 }
367
368 #ifdef HAVE_ALTIVEC
369
370 #ifdef HAVE_ALTIVEC_H
371 #include <altivec.h>
372 #endif
373
374 // used to build registers permutation vectors (vcprm)
375 // the 's' are for words in the _s_econd vector
376 #define WORD_0 0x00,0x01,0x02,0x03
377 #define WORD_1 0x04,0x05,0x06,0x07
378 #define WORD_2 0x08,0x09,0x0a,0x0b
379 #define WORD_3 0x0c,0x0d,0x0e,0x0f
380 #define WORD_s0 0x10,0x11,0x12,0x13
381 #define WORD_s1 0x14,0x15,0x16,0x17
382 #define WORD_s2 0x18,0x19,0x1a,0x1b
383 #define WORD_s3 0x1c,0x1d,0x1e,0x1f
384
385 #ifdef __APPLE_CC__
386 #define vcprm(a,b,c,d) (const vector unsigned char)(WORD_ ## a, WORD_ ## b, WORD_ ## c, WORD_ ## d)
387 #else
388 #define vcprm(a,b,c,d) (const vector unsigned char){WORD_ ## a, WORD_ ## b, WORD_ ## c, WORD_ ## d}
389 #endif
390
391 // vcprmle is used to keep the same index as in the SSE version.
392 // it's the same as vcprm, with the index inversed
393 // ('le' is Little Endian)
394 #define vcprmle(a,b,c,d) vcprm(d,c,b,a)
395
396 // used to build inverse/identity vectors (vcii)
397 // n is _n_egative, p is _p_ositive
398 #define FLOAT_n -1.
399 #define FLOAT_p 1.
400
401 #ifdef __APPLE_CC__
402 #define vcii(a,b,c,d) (const vector float)(FLOAT_ ## a, FLOAT_ ## b, FLOAT_ ## c, FLOAT_ ## d)
403 #else
404 #define vcii(a,b,c,d) (const vector float){FLOAT_ ## a, FLOAT_ ## b, FLOAT_ ## c, FLOAT_ ## d}
405 #endif
406
407 #ifdef __APPLE_CC__
408 #define FOUROF(a) (a)
409 #else
410 #define FOUROF(a) {a,a,a,a}
411 #endif
412
413
414 void
415 imdct_do_512_altivec(sample_t data[],sample_t delay[], sample_t bias)
416 {
417 int i;
418 int k;
419 int p,q;
420 int m;
421 long two_m;
422 long two_m_plus_one;
423
424 sample_t tmp_b_i;
425 sample_t tmp_b_r;
426 sample_t tmp_a_i;
427 sample_t tmp_a_r;
428
429 sample_t *data_ptr;
430 sample_t *delay_ptr;
431 sample_t *window_ptr;
432
433 /* 512 IMDCT with source and dest data in 'data' */
434
435 /* Pre IFFT complex multiply plus IFFT cmplx conjugate & reordering*/
436 for( i=0; i < 128; i++) {
437 /* z[i] = (X[256-2*i-1] + j * X[2*i]) * (xcos1[i] + j * xsin1[i]) ; */
438 int j= bit_reverse_512[i];
439 buf[i].real = (data[256-2*j-1] * xcos1[j]) - (data[2*j] * xsin1[j]);
440 buf[i].imag = -1.0 * ((data[2*j] * xcos1[j]) + (data[256-2*j-1] * xsin1[j]));
441 }
442
443 /* 1. iteration */
444 for(i = 0; i < 128; i += 2) {
445 #if 0
446 tmp_a_r = buf[i].real;
447 tmp_a_i = buf[i].imag;
448 tmp_b_r = buf[i+1].real;
449 tmp_b_i = buf[i+1].imag;
450 buf[i].real = tmp_a_r + tmp_b_r;
451 buf[i].imag = tmp_a_i + tmp_b_i;
452 buf[i+1].real = tmp_a_r - tmp_b_r;
453 buf[i+1].imag = tmp_a_i - tmp_b_i;
454 #else
455 vector float temp, bufv;
456
457 bufv = vec_ld(i << 3, (float*)buf);
458 temp = vec_perm(bufv, bufv, vcprm(2,3,0,1));
459 bufv = vec_madd(bufv, vcii(p,p,n,n), temp);
460 vec_st(bufv, i << 3, (float*)buf);
461 #endif
462 }
463
464 /* 2. iteration */
465 // Note w[1]={{1,0}, {0,-1}}
466 for(i = 0; i < 128; i += 4) {
467 #if 0
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 /* WARNING: im <-> re here ! */
479 tmp_b_r = buf[i+3].imag;
480 tmp_b_i = buf[i+3].real;
481 buf[i+1].real = tmp_a_r + tmp_b_r;
482 buf[i+1].imag = tmp_a_i - tmp_b_i;
483 buf[i+3].real = tmp_a_r - tmp_b_r;
484 buf[i+3].imag = tmp_a_i + tmp_b_i;
485 #else
486 vector float buf01, buf23, temp1, temp2;
487
488 buf01 = vec_ld((i + 0) << 3, (float*)buf);
489 buf23 = vec_ld((i + 2) << 3, (float*)buf);
490 buf23 = vec_perm(buf23,buf23,vcprm(0,1,3,2));
491
492 temp1 = vec_madd(buf23, vcii(p,p,p,n), buf01);
493 temp2 = vec_madd(buf23, vcii(n,n,n,p), buf01);
494
495 vec_st(temp1, (i + 0) << 3, (float*)buf);
496 vec_st(temp2, (i + 2) << 3, (float*)buf);
497 #endif
498 }
499
500 /* 3. iteration */
501 for(i = 0; i < 128; i += 8) {
502 #if 0
503 tmp_a_r = buf[i].real;
504 tmp_a_i = buf[i].imag;
505 tmp_b_r = buf[i+4].real;
506 tmp_b_i = buf[i+4].imag;
507 buf[i].real = tmp_a_r + tmp_b_r;
508 buf[i].imag = tmp_a_i + tmp_b_i;
509 buf[i+4].real = tmp_a_r - tmp_b_r;
510 buf[i+4].imag = tmp_a_i - tmp_b_i;
511 tmp_a_r = buf[1+i].real;
512 tmp_a_i = buf[1+i].imag;
513 tmp_b_r = (buf[i+5].real + buf[i+5].imag) * w[2][1].real;
514 tmp_b_i = (buf[i+5].imag - buf[i+5].real) * w[2][1].real;
515 buf[1+i].real = tmp_a_r + tmp_b_r;
516 buf[1+i].imag = tmp_a_i + tmp_b_i;
517 buf[i+5].real = tmp_a_r - tmp_b_r;
518 buf[i+5].imag = tmp_a_i - tmp_b_i;
519 tmp_a_r = buf[i+2].real;
520 tmp_a_i = buf[i+2].imag;
521 /* WARNING re <-> im & sign */
522 tmp_b_r = buf[i+6].imag;
523 tmp_b_i = - buf[i+6].real;
524 buf[i+2].real = tmp_a_r + tmp_b_r;
525 buf[i+2].imag = tmp_a_i + tmp_b_i;
526 buf[i+6].real = tmp_a_r - tmp_b_r;
527 buf[i+6].imag = tmp_a_i - tmp_b_i;
528 tmp_a_r = buf[i+3].real;
529 tmp_a_i = buf[i+3].imag;
530 tmp_b_r = (buf[i+7].real - buf[i+7].imag) * w[2][3].imag;
531 tmp_b_i = (buf[i+7].imag + buf[i+7].real) * w[2][3].imag;
532 buf[i+3].real = tmp_a_r + tmp_b_r;
533 buf[i+3].imag = tmp_a_i + tmp_b_i;
534 buf[i+7].real = tmp_a_r - tmp_b_r;
535 buf[i+7].imag = tmp_a_i - tmp_b_i;
536 #else
537 vector float buf01, buf23, buf45, buf67;
538
539 buf01 = vec_ld((i + 0) << 3, (float*)buf);
540 buf23 = vec_ld((i + 2) << 3, (float*)buf);
541
542 tmp_b_r = (buf[i+5].real + buf[i+5].imag) * w[2][1].real;
543 tmp_b_i = (buf[i+5].imag - buf[i+5].real) * w[2][1].real;
544 buf[i+5].real = tmp_b_r;
545 buf[i+5].imag = tmp_b_i;
546 tmp_b_r = (buf[i+7].real - buf[i+7].imag) * w[2][3].imag;
547 tmp_b_i = (buf[i+7].imag + buf[i+7].real) * w[2][3].imag;
548 buf[i+7].real = tmp_b_r;
549 buf[i+7].imag = tmp_b_i;
550
551 buf23 = vec_ld((i + 2) << 3, (float*)buf);
552 buf45 = vec_ld((i + 4) << 3, (float*)buf);
553 buf67 = vec_ld((i + 6) << 3, (float*)buf);
554 buf67 = vec_perm(buf67, buf67, vcprm(1,0,2,3));
555
556 vec_st(vec_add(buf01, buf45), (i + 0) << 3, (float*)buf);
557 vec_st(vec_madd(buf67, vcii(p,n,p,p), buf23), (i + 2) << 3, (float*)buf);
558 vec_st(vec_sub(buf01, buf45), (i + 4) << 3, (float*)buf);
559 vec_st(vec_nmsub(buf67, vcii(p,n,p,p), buf23), (i + 6) << 3, (float*)buf);
560 #endif
561 }
562
563 /* 4-7. iterations */
564 for (m=3; m < 7; m++) {
565 two_m = (1 << m);
566
567 two_m_plus_one = two_m<<1;
568
569 for(i = 0; i < 128; i += two_m_plus_one) {
570 for(k = 0; k < two_m; k+=2) {
571 #if 0
572 int p = k + i;
573 int q = p + two_m;
574 tmp_a_r = buf[p].real;
575 tmp_a_i = buf[p].imag;
576 tmp_b_r =
577 buf[q].real * w[m][k].real -
578 buf[q].imag * w[m][k].imag;
579 tmp_b_i =
580 buf[q].imag * w[m][k].real +
581 buf[q].real * w[m][k].imag;
582 buf[p].real = tmp_a_r + tmp_b_r;
583 buf[p].imag = tmp_a_i + tmp_b_i;
584 buf[q].real = tmp_a_r - tmp_b_r;
585 buf[q].imag = tmp_a_i - tmp_b_i;
586
587 tmp_a_r = buf[(p + 1)].real;
588 tmp_a_i = buf[(p + 1)].imag;
589 tmp_b_r =
590 buf[(q + 1)].real * w[m][(k + 1)].real -
591 buf[(q + 1)].imag * w[m][(k + 1)].imag;
592 tmp_b_i =
593 buf[(q + 1)].imag * w[m][(k + 1)].real +
594 buf[(q + 1)].real * w[m][(k + 1)].imag;
595 buf[(p + 1)].real = tmp_a_r + tmp_b_r;
596 buf[(p + 1)].imag = tmp_a_i + tmp_b_i;
597 buf[(q + 1)].real = tmp_a_r - tmp_b_r;
598 buf[(q + 1)].imag = tmp_a_i - tmp_b_i;
599 #else
600 int p = k + i;
601 int q = p + two_m;
602 vector float vecp, vecq, vecw, temp1, temp2, temp3, temp4;
603 const vector float vczero = (const vector float)FOUROF(0.);
604 // first compute buf[q] and buf[q+1]
605 vecq = vec_ld(q << 3, (float*)buf);
606 vecw = vec_ld(0, (float*)&(w[m][k]));
607 temp1 = vec_madd(vecq, vecw, vczero);
608 temp2 = vec_perm(vecq, vecq, vcprm(1,0,3,2));
609 temp2 = vec_madd(temp2, vecw, vczero);
610 temp3 = vec_perm(temp1, temp2, vcprm(0,s0,2,s2));
611 temp4 = vec_perm(temp1, temp2, vcprm(1,s1,3,s3));
612 vecq = vec_madd(temp4, vcii(n,p,n,p), temp3);
613 // then butterfly with buf[p] and buf[p+1]
614 vecp = vec_ld(p << 3, (float*)buf);
615
616 temp1 = vec_add(vecp, vecq);
617 temp2 = vec_sub(vecp, vecq);
618
619 vec_st(temp1, p << 3, (float*)buf);
620 vec_st(temp2, q << 3, (float*)buf);
621 #endif
622 }
623 }
624 }
625
626 /* Post IFFT complex multiply plus IFFT complex conjugate*/
627 for( i=0; i < 128; i+=4) {
628 /* y[n] = z[n] * (xcos1[n] + j * xsin1[n]) ; */
629 #if 0
630 tmp_a_r = buf[(i + 0)].real;
631 tmp_a_i = -1.0 * buf[(i + 0)].imag;
632 buf[(i + 0)].real =
633 (tmp_a_r * xcos1[(i + 0)]) - (tmp_a_i * xsin1[(i + 0)]);
634 buf[(i + 0)].imag =
635 (tmp_a_r * xsin1[(i + 0)]) + (tmp_a_i * xcos1[(i + 0)]);
636
637 tmp_a_r = buf[(i + 1)].real;
638 tmp_a_i = -1.0 * buf[(i + 1)].imag;
639 buf[(i + 1)].real =
640 (tmp_a_r * xcos1[(i + 1)]) - (tmp_a_i * xsin1[(i + 1)]);
641 buf[(i + 1)].imag =
642 (tmp_a_r * xsin1[(i + 1)]) + (tmp_a_i * xcos1[(i + 1)]);
643
644 tmp_a_r = buf[(i + 2)].real;
645 tmp_a_i = -1.0 * buf[(i + 2)].imag;
646 buf[(i + 2)].real =
647 (tmp_a_r * xcos1[(i + 2)]) - (tmp_a_i * xsin1[(i + 2)]);
648 buf[(i + 2)].imag =
649 (tmp_a_r * xsin1[(i + 2)]) + (tmp_a_i * xcos1[(i + 2)]);
650
651 tmp_a_r = buf[(i + 3)].real;
652 tmp_a_i = -1.0 * buf[(i + 3)].imag;
653 buf[(i + 3)].real =
654 (tmp_a_r * xcos1[(i + 3)]) - (tmp_a_i * xsin1[(i + 3)]);
655 buf[(i + 3)].imag =
656 (tmp_a_r * xsin1[(i + 3)]) + (tmp_a_i * xcos1[(i + 3)]);
657 #else
658 vector float bufv_0, bufv_2, cosv, sinv, temp1, temp2;
659 vector float temp0022, temp1133, tempCS01;
660 const vector float vczero = (const vector float)FOUROF(0.);
661
662 bufv_0 = vec_ld((i + 0) << 3, (float*)buf);
663 bufv_2 = vec_ld((i + 2) << 3, (float*)buf);
664
665 cosv = vec_ld(i << 2, xcos1);
666 sinv = vec_ld(i << 2, xsin1);
667
668 temp0022 = vec_perm(bufv_0, bufv_0, vcprm(0,0,2,2));
669 temp1133 = vec_perm(bufv_0, bufv_0, vcprm(1,1,3,3));
670 tempCS01 = vec_perm(cosv, sinv, vcprm(0,s0,1,s1));
671 temp1 = vec_madd(temp0022, tempCS01, vczero);
672 tempCS01 = vec_perm(cosv, sinv, vcprm(s0,0,s1,1));
673 temp2 = vec_madd(temp1133, tempCS01, vczero);
674 bufv_0 = vec_madd(temp2, vcii(p,n,p,n), temp1);
675
676 vec_st(bufv_0, (i + 0) << 3, (float*)buf);
677
678 /* idem with bufv_2 and high-order cosv/sinv */
679
680 temp0022 = vec_perm(bufv_2, bufv_2, vcprm(0,0,2,2));
681 temp1133 = vec_perm(bufv_2, bufv_2, vcprm(1,1,3,3));
682 tempCS01 = vec_perm(cosv, sinv, vcprm(2,s2,3,s3));
683 temp1 = vec_madd(temp0022, tempCS01, vczero);
684 tempCS01 = vec_perm(cosv, sinv, vcprm(s2,2,s3,3));
685 temp2 = vec_madd(temp1133, tempCS01, vczero);
686 bufv_2 = vec_madd(temp2, vcii(p,n,p,n), temp1);
687
688 vec_st(bufv_2, (i + 2) << 3, (float*)buf);
689
690 #endif
691 }
692
693 data_ptr = data;
694 delay_ptr = delay;
695 window_ptr = a52_imdct_window;
696
697 /* Window and convert to real valued signal */
698 for(i=0; i< 64; i++) {
699 *data_ptr++ = -buf[64+i].imag * *window_ptr++ + *delay_ptr++ + bias;
700 *data_ptr++ = buf[64-i-1].real * *window_ptr++ + *delay_ptr++ + bias;
701 }
702
703 for(i=0; i< 64; i++) {
704 *data_ptr++ = -buf[i].real * *window_ptr++ + *delay_ptr++ + bias;
705 *data_ptr++ = buf[128-i-1].imag * *window_ptr++ + *delay_ptr++ + bias;
706 }
707
708 /* The trailing edge of the window goes into the delay line */
709 delay_ptr = delay;
710
711 for(i=0; i< 64; i++) {
712 *delay_ptr++ = -buf[64+i].real * *--window_ptr;
713 *delay_ptr++ = buf[64-i-1].imag * *--window_ptr;
714 }
715
716 for(i=0; i<64; i++) {
717 *delay_ptr++ = buf[i].imag * *--window_ptr;
718 *delay_ptr++ = -buf[128-i-1].real * *--window_ptr;
719 }
720 }
721 #endif
722
723
724 // Stuff below this line is borrowed from libac3
725 #include "srfftp.h"
726 #if defined(ARCH_X86) || defined(ARCH_X86_64)
727 #ifndef HAVE_3DNOW
728 #define HAVE_3DNOW 1
729 #endif
730 #include "srfftp_3dnow.h"
731
732 const i_cmplx_t x_plus_minus_3dnow __attribute__ ((aligned (8))) = {{ 0x00000000UL, 0x80000000UL }};
733 const i_cmplx_t x_minus_plus_3dnow __attribute__ ((aligned (8))) = {{ 0x80000000UL, 0x00000000UL }};
734 const complex_t HSQRT2_3DNOW __attribute__ ((aligned (8))) = { 0.707106781188, 0.707106781188 };
735
736 #undef HAVE_3DNOWEX
737 #include "imdct_3dnow.h"
738 #define HAVE_3DNOWEX
739 #include "imdct_3dnow.h"
740
741 void
742 imdct_do_512_sse(sample_t data[],sample_t delay[], sample_t bias)
743 {
744 /* int i,k;
745 int p,q;*/
746 int m;
747 long two_m;
748 long two_m_plus_one;
749 long two_m_plus_one_shl3;
750 complex_t *buf_offset;
751
752 /* sample_t tmp_a_i;
753 sample_t tmp_a_r;
754 sample_t tmp_b_i;
755 sample_t tmp_b_r;*/
756
757 sample_t *data_ptr;
758 sample_t *delay_ptr;
759 sample_t *window_ptr;
760
761 /* 512 IMDCT with source and dest data in 'data' */
762 /* see the c version (dct_do_512()), its allmost identical, just in C */
763
764 /* Pre IFFT complex multiply plus IFFT cmplx conjugate */
765 /* Bit reversed shuffling */
766 asm volatile(
767 "xor %%"REG_S", %%"REG_S" \n\t"
768 "lea "MANGLE(bit_reverse_512)", %%"REG_a"\n\t"
769 "mov $1008, %%"REG_D" \n\t"
770 "push %%"REG_BP" \n\t" //use ebp without telling gcc
771 ASMALIGN(4)
772 "1: \n\t"
773 "movlps (%0, %%"REG_S"), %%xmm0 \n\t" // XXXI
774 "movhps 8(%0, %%"REG_D"), %%xmm0 \n\t" // RXXI
775 "movlps 8(%0, %%"REG_S"), %%xmm1 \n\t" // XXXi
776 "movhps (%0, %%"REG_D"), %%xmm1 \n\t" // rXXi
777 "shufps $0x33, %%xmm1, %%xmm0 \n\t" // irIR
778 "movaps "MANGLE(sseSinCos1c)"(%%"REG_S"), %%xmm2\n\t"
779 "mulps %%xmm0, %%xmm2 \n\t"
780 "shufps $0xB1, %%xmm0, %%xmm0 \n\t" // riRI
781 "mulps "MANGLE(sseSinCos1d)"(%%"REG_S"), %%xmm0\n\t"
782 "subps %%xmm0, %%xmm2 \n\t"
783 "movzb (%%"REG_a"), %%"REG_d" \n\t"
784 "movzb 1(%%"REG_a"), %%"REG_BP" \n\t"
785 "movlps %%xmm2, (%1, %%"REG_d", 8) \n\t"
786 "movhps %%xmm2, (%1, %%"REG_BP", 8) \n\t"
787 "add $16, %%"REG_S" \n\t"
788 "add $2, %%"REG_a" \n\t" // avoid complex addressing for P4 crap
789 "sub $16, %%"REG_D" \n\t"
790 "jnc 1b \n\t"
791 "pop %%"REG_BP" \n\t"//no we didnt touch ebp *g*
792 :: "b" (data), "c" (buf)
793 : "%"REG_S, "%"REG_D, "%"REG_a, "%"REG_d
794 );
795
796
797 /* FFT Merge */
798 /* unoptimized variant
799 for (m=1; m < 7; m++) {
800 if(m)
801 two_m = (1 << m);
802 else
803 two_m = 1;
804
805 two_m_plus_one = (1 << (m+1));
806
807 for(i = 0; i < 128; i += two_m_plus_one) {
808 for(k = 0; k < two_m; k++) {
809 p = k + i;
810 q = p + two_m;
811 tmp_a_r = buf[p].real;
812 tmp_a_i = buf[p].imag;
813 tmp_b_r = buf[q].real * w[m][k].real - buf[q].imag * w[m][k].imag;
814 tmp_b_i = buf[q].imag * w[m][k].real + buf[q].real * w[m][k].imag;
815 buf[p].real = tmp_a_r + tmp_b_r;
816 buf[p].imag = tmp_a_i + tmp_b_i;
817 buf[q].real = tmp_a_r - tmp_b_r;
818 buf[q].imag = tmp_a_i - tmp_b_i;
819 }
820 }
821 }
822 */
823
824 /* 1. iteration */
825 // Note w[0][0]={1,0}
826 asm volatile(
827 "xorps %%xmm1, %%xmm1 \n\t"
828 "xorps %%xmm2, %%xmm2 \n\t"
829 "mov %0, %%"REG_S" \n\t"
830 ASMALIGN(4)
831 "1: \n\t"
832 "movlps (%%"REG_S"), %%xmm0\n\t" //buf[p]
833 "movlps 8(%%"REG_S"), %%xmm1\n\t" //buf[q]
834 "movhps (%%"REG_S"), %%xmm0\n\t" //buf[p]
835 "movhps 8(%%"REG_S"), %%xmm2\n\t" //buf[q]
836 "addps %%xmm1, %%xmm0 \n\t"
837 "subps %%xmm2, %%xmm0 \n\t"
838 "movaps %%xmm0, (%%"REG_S")\n\t"
839 "add $16, %%"REG_S" \n\t"
840 "cmp %1, %%"REG_S" \n\t"
841 " jb 1b \n\t"
842 :: "g" (buf), "r" (buf + 128)
843 : "%"REG_S
844 );
845
846 /* 2. iteration */
847 // Note w[1]={{1,0}, {0,-1}}
848 asm volatile(
849 "movaps "MANGLE(ps111_1)", %%xmm7\n\t" // 1,1,1,-1
850 "mov %0, %%"REG_S" \n\t"
851 ASMALIGN(4)
852 "1: \n\t"
853 "movaps 16(%%"REG_S"), %%xmm2 \n\t" //r2,i2,r3,i3
854 "shufps $0xB4, %%xmm2, %%xmm2 \n\t" //r2,i2,i3,r3
855 "mulps %%xmm7, %%xmm2 \n\t" //r2,i2,i3,-r3
856 "movaps (%%"REG_S"), %%xmm0 \n\t" //r0,i0,r1,i1
857 "movaps (%%"REG_S"), %%xmm1 \n\t" //r0,i0,r1,i1
858 "addps %%xmm2, %%xmm0 \n\t"
859 "subps %%xmm2, %%xmm1 \n\t"
860 "movaps %%xmm0, (%%"REG_S") \n\t"
861 "movaps %%xmm1, 16(%%"REG_S") \n\t"
862 "add $32, %%"REG_S" \n\t"
863 "cmp %1, %%"REG_S" \n\t"
864 " jb 1b \n\t"
865 :: "g" (buf), "r" (buf + 128)
866 : "%"REG_S
867 );
868
869 /* 3. iteration */
870 /*
871 Note sseW2+0={1,1,sqrt(2),sqrt(2))
872 Note sseW2+16={0,0,sqrt(2),-sqrt(2))
873 Note sseW2+32={0,0,-sqrt(2),-sqrt(2))
874 Note sseW2+48={1,-1,sqrt(2),-sqrt(2))
875 */
876 asm volatile(
877 "movaps 48+"MANGLE(sseW2)", %%xmm6\n\t"
878 "movaps 16+"MANGLE(sseW2)", %%xmm7\n\t"
879 "xorps %%xmm5, %%xmm5 \n\t"
880 "xorps %%xmm2, %%xmm2 \n\t"
881 "mov %0, %%"REG_S" \n\t"
882 ASMALIGN(4)
883 "1: \n\t"
884 "movaps 32(%%"REG_S"), %%xmm2 \n\t" //r4,i4,r5,i5
885 "movaps 48(%%"REG_S"), %%xmm3 \n\t" //r6,i6,r7,i7
886 "movaps "MANGLE(sseW2)", %%xmm4 \n\t" //r4,i4,r5,i5
887 "movaps 32+"MANGLE(sseW2)", %%xmm5\n\t" //r6,i6,r7,i7
888 "mulps %%xmm2, %%xmm4 \n\t"
889 "mulps %%xmm3, %%xmm5 \n\t"
890 "shufps $0xB1, %%xmm2, %%xmm2 \n\t" //i4,r4,i5,r5
891 "shufps $0xB1, %%xmm3, %%xmm3 \n\t" //i6,r6,i7,r7
892 "mulps %%xmm6, %%xmm3 \n\t"
893 "mulps %%xmm7, %%xmm2 \n\t"
894 "movaps (%%"REG_S"), %%xmm0 \n\t" //r0,i0,r1,i1
895 "movaps 16(%%"REG_S"), %%xmm1 \n\t" //r2,i2,r3,i3
896 "addps %%xmm4, %%xmm2 \n\t"
897 "addps %%xmm5, %%xmm3 \n\t"
898 "movaps %%xmm2, %%xmm4 \n\t"
899 "movaps %%xmm3, %%xmm5 \n\t"
900 "addps %%xmm0, %%xmm2 \n\t"
901 "addps %%xmm1, %%xmm3 \n\t"
902 "subps %%xmm4, %%xmm0 \n\t"
903 "subps %%xmm5, %%xmm1 \n\t"
904 "movaps %%xmm2, (%%"REG_S") \n\t"
905 "movaps %%xmm3, 16(%%"REG_S") \n\t"
906 "movaps %%xmm0, 32(%%"REG_S") \n\t"
907 "movaps %%xmm1, 48(%%"REG_S") \n\t"
908 "add $64, %%"REG_S" \n\t"
909 "cmp %1, %%"REG_S" \n\t"
910 " jb 1b \n\t"
911 :: "g" (buf), "r" (buf + 128)
912 : "%"REG_S
913 );
914
915 /* 4-7. iterations */
916 for (m=3; m < 7; m++) {
917 two_m = (1 << m);
918 two_m_plus_one = two_m<<1;
919 two_m_plus_one_shl3 = (two_m_plus_one<<3);
920 buf_offset = buf+128;
921 asm volatile(
922 "mov %0, %%"REG_S" \n\t"
923 ASMALIGN(4)
924 "1: \n\t"
925 "xor %%"REG_D", %%"REG_D" \n\t" // k
926 "lea (%%"REG_S", %3), %%"REG_d" \n\t"
927 "2: \n\t"
928 "movaps (%%"REG_d", %%"REG_D"), %%xmm1 \n\t"
929 "movaps (%4, %%"REG_D", 2), %%xmm2 \n\t"
930 "mulps %%xmm1, %%xmm2 \n\t"
931 "shufps $0xB1, %%xmm1, %%xmm1 \n\t"
932 "mulps 16(%4, %%"REG_D", 2), %%xmm1 \n\t"
933 "movaps (%%"REG_S", %%"REG_D"), %%xmm0 \n\t"
934 "addps %%xmm2, %%xmm1 \n\t"
935 "movaps %%xmm1, %%xmm2 \n\t"
936 "addps %%xmm0, %%xmm1 \n\t"
937 "subps %%xmm2, %%xmm0 \n\t"
938 "movaps %%xmm1, (%%"REG_S", %%"REG_D") \n\t"
939 "movaps %%xmm0, (%%"REG_d", %%"REG_D") \n\t"
940 "add $16, %%"REG_D" \n\t"
941 "cmp %3, %%"REG_D" \n\t" //FIXME (opt) count against 0
942 "jb 2b \n\t"
943 "add %2, %%"REG_S" \n\t"
944 "cmp %1, %%"REG_S" \n\t"
945 " jb 1b \n\t"
946 :: "g" (buf), "m" (buf_offset), "m" (two_m_plus_one_shl3), "r" (two_m<<3),
947 "r" (sseW[m])
948 : "%"REG_S, "%"REG_D, "%"REG_d
949 );
950 }
951
952 /* Post IFFT complex multiply plus IFFT complex conjugate*/
953 asm volatile(
954 "mov $-1024, %%"REG_S" \n\t"
955 ASMALIGN(4)
956 "1: \n\t"
957 "movaps (%0, %%"REG_S"), %%xmm0 \n\t"
958 "movaps (%0, %%"REG_S"), %%xmm1 \n\t"
959 "shufps $0xB1, %%xmm0, %%xmm0 \n\t"
960 "mulps 1024+"MANGLE(sseSinCos1c)"(%%"REG_S"), %%xmm1\n\t"
961 "mulps 1024+"MANGLE(sseSinCos1d)"(%%"REG_S"), %%xmm0\n\t"
962 "addps %%xmm1, %%xmm0 \n\t"
963 "movaps %%xmm0, (%0, %%"REG_S") \n\t"
964 "add $16, %%"REG_S" \n\t"
965 " jnz 1b \n\t"
966 :: "r" (buf+128)
967 : "%"REG_S
968 );
969
970
971 data_ptr = data;
972 delay_ptr = delay;
973 window_ptr = a52_imdct_window;
974
975 /* Window and convert to real valued signal */
976 asm volatile(
977 "xor %%"REG_D", %%"REG_D" \n\t" // 0
978 "xor %%"REG_S", %%"REG_S" \n\t" // 0
979 "movss %3, %%xmm2 \n\t" // bias
980 "shufps $0x00, %%xmm2, %%xmm2 \n\t" // bias, bias, ...
981 ASMALIGN(4)
982 "1: \n\t"
983 "movlps (%0, %%"REG_S"), %%xmm0 \n\t" // ? ? A ?
984 "movlps 8(%0, %%"REG_S"), %%xmm1 \n\t" // ? ? C ?
985 "movhps -16(%0, %%"REG_D"), %%xmm1 \n\t" // ? D C ?
986 "movhps -8(%0, %%"REG_D"), %%xmm0 \n\t" // ? B A ?
987 "shufps $0x99, %%xmm1, %%xmm0 \n\t" // D C B A
988 "mulps "MANGLE(sseWindow)"(%%"REG_S"), %%xmm0\n\t"
989 "addps (%2, %%"REG_S"), %%xmm0 \n\t"
990 "addps %%xmm2, %%xmm0 \n\t"
991 "movaps %%xmm0, (%1, %%"REG_S") \n\t"
992 "add $16, %%"REG_S" \n\t"
993 "sub $16, %%"REG_D" \n\t"
994 "cmp $512, %%"REG_S" \n\t"
995 " jb 1b \n\t"
996 :: "r" (buf+64), "r" (data_ptr), "r" (delay_ptr), "m" (bias)
997 : "%"REG_S, "%"REG_D
998 );
999 data_ptr+=128;
1000 delay_ptr+=128;
1001 // window_ptr+=128;
1002
1003 asm volatile(
1004 "mov $1024, %%"REG_D" \n\t" // 512
1005 "xor %%"REG_S", %%"REG_S" \n\t" // 0
1006 "movss %3, %%xmm2 \n\t" // bias
1007 "shufps $0x00, %%xmm2, %%xmm2 \n\t" // bias, bias, ...
1008 ASMALIGN(4)
1009 "1: \n\t"
1010 "movlps (%0, %%"REG_S"), %%xmm0 \n\t" // ? ? ? A
1011 "movlps 8(%0, %%"REG_S"), %%xmm1 \n\t" // ? ? ? C
1012 "movhps -16(%0, %%"REG_D"), %%xmm1 \n\t" // D ? ? C
1013 "movhps -8(%0, %%"REG_D"), %%xmm0 \n\t" // B ? ? A
1014 "shufps $0xCC, %%xmm1, %%xmm0 \n\t" // D C B A
1015 "mulps 512+"MANGLE(sseWindow)"(%%"REG_S"), %%xmm0\n\t"
1016 "addps (%2, %%"REG_S"), %%xmm0 \n\t"
1017 "addps %%xmm2, %%xmm0 \n\t"
1018 "movaps %%xmm0, (%1, %%"REG_S") \n\t"
1019 "add $16, %%"REG_S" \n\t"
1020 "sub $16, %%"REG_D" \n\t"
1021 "cmp $512, %%"REG_S" \n\t"
1022 " jb 1b \n\t"
1023 :: "r" (buf), "r" (data_ptr), "r" (delay_ptr), "m" (bias)
1024 : "%"REG_S, "%"REG_D
1025 );
1026 data_ptr+=128;
1027 // window_ptr+=128;
1028
1029 /* The trailing edge of the window goes into the delay line */
1030 delay_ptr = delay;
1031
1032 asm volatile(
1033 "xor %%"REG_D", %%"REG_D" \n\t" // 0
1034 "xor %%"REG_S", %%"REG_S" \n\t" // 0
1035 ASMALIGN(4)
1036 "1: \n\t"
1037 "movlps (%0, %%"REG_S"), %%xmm0 \n\t" // ? ? ? A
1038 "movlps 8(%0, %%"REG_S"), %%xmm1 \n\t" // ? ? ? C
1039 "movhps -16(%0, %%"REG_D"), %%xmm1 \n\t" // D ? ? C
1040 "movhps -8(%0, %%"REG_D"), %%xmm0 \n\t" // B ? ? A
1041 "shufps $0xCC, %%xmm1, %%xmm0 \n\t" // D C B A
1042 "mulps 1024+"MANGLE(sseWindow)"(%%"REG_S"), %%xmm0\n\t"
1043 "movaps %%xmm0, (%1, %%"REG_S") \n\t"
1044 "add $16, %%"REG_S" \n\t"
1045 "sub $16, %%"REG_D" \n\t"
1046 "cmp $512, %%"REG_S" \n\t"
1047 " jb 1b \n\t"
1048 :: "r" (buf+64), "r" (delay_ptr)
1049 : "%"REG_S, "%"REG_D
1050 );
1051 delay_ptr+=128;
1052 // window_ptr-=128;
1053
1054 asm volatile(
1055 "mov $1024, %%"REG_D" \n\t" // 1024
1056 "xor %%"REG_S", %%"REG_S" \n\t" // 0
1057 ASMALIGN(4)
1058 "1: \n\t"
1059 "movlps (%0, %%"REG_S"), %%xmm0 \n\t" // ? ? A ?
1060 "movlps 8(%0, %%"REG_S"), %%xmm1 \n\t" // ? ? C ?
1061 "movhps -16(%0, %%"REG_D"), %%xmm1 \n\t" // ? D C ?
1062 "movhps -8(%0, %%"REG_D"), %%xmm0 \n\t" // ? B A ?
1063 "shufps $0x99, %%xmm1, %%xmm0 \n\t" // D C B A
1064 "mulps 1536+"MANGLE(sseWindow)"(%%"REG_S"), %%xmm0\n\t"
1065 "movaps %%xmm0, (%1, %%"REG_S") \n\t"
1066 "add $16, %%"REG_S" \n\t"
1067 "sub $16, %%"REG_D" \n\t"
1068 "cmp $512, %%"REG_S" \n\t"
1069 " jb 1b \n\t"
1070 :: "r" (buf), "r" (delay_ptr)
1071 : "%"REG_S, "%"REG_D
1072 );
1073 }
1074 #endif // ARCH_X86 || ARCH_X86_64
1075
1076 void a52_imdct_256(sample_t * data, sample_t * delay, sample_t bias)
1077 {
1078 int i, k;
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;
1080 const sample_t * window = a52_imdct_window;
1081 complex_t buf1[64], buf2[64];
1082
1083 /* Pre IFFT complex multiply plus IFFT cmplx conjugate */
1084 for (i = 0; i < 64; i++) {
1085 k = fftorder[i];
1086 t_r = pre2[i].real;
1087 t_i = pre2[i].imag;
1088
1089 buf1[i].real = t_i * data[254-k] + t_r * data[k];
1090 buf1[i].imag = t_r * data[254-k] - t_i * data[k];
1091
1092 buf2[i].real = t_i * data[255-k] + t_r * data[k+1];
1093 buf2[i].imag = t_r * data[255-k] - t_i * data[k+1];
1094 }
1095
1096 ifft64 (buf1);
1097 ifft64 (buf2);
1098
1099 /* Post IFFT complex multiply */
1100 /* Window and convert to real valued signal */
1101 for (i = 0; i < 32; i++) {
1102 /* y1[n] = z1[n] * (xcos2[n] + j * xs in2[n]) ; */
1103 t_r = post2[i].real;
1104 t_i = post2[i].imag;
1105
1106 a_r = t_r * buf1[i].real + t_i * buf1[i].imag;
1107 a_i = t_i * buf1[i].real - t_r * buf1[i].imag;
1108 b_r = t_i * buf1[63-i].real + t_r * buf1[63-i].imag;
1109 b_i = t_r * buf1[63-i].real - t_i * buf1[63-i].imag;
1110
1111 c_r = t_r * buf2[i].real + t_i * buf2[i].imag;
1112 c_i = t_i * buf2[i].real - t_r * buf2[i].imag;
1113 d_r = t_i * buf2[63-i].real + t_r * buf2[63-i].imag;
1114 d_i = t_r * buf2[63-i].real - t_i * buf2[63-i].imag;
1115
1116 w_1 = window[2*i];
1117 w_2 = window[255-2*i];
1118 data[2*i] = delay[2*i] * w_2 - a_r * w_1 + bias;
1119 data[255-2*i] = delay[2*i] * w_1 + a_r * w_2 + bias;
1120 delay[2*i] = c_i;
1121
1122 w_1 = window[128+2*i];
1123 w_2 = window[127-2*i];
1124 data[128+2*i] = delay[127-2*i] * w_2 + a_i * w_1 + bias;
1125 data[127-2*i] = delay[127-2*i] * w_1 - a_i * w_2 + bias;
1126 delay[127-2*i] = c_r;
1127
1128 w_1 = window[2*i+1];
1129 w_2 = window[254-2*i];
1130 data[2*i+1] = delay[2*i+1] * w_2 - b_i * w_1 + bias;
1131 data[254-2*i] = delay[2*i+1] * w_1 + b_i * w_2 + bias;
1132 delay[2*i+1] = d_r;
1133
1134 w_1 = window[129+2*i];
1135 w_2 = window[126-2*i];
1136 data[129+2*i] = delay[126-2*i] * w_2 + b_r * w_1 + bias;
1137 data[126-2*i] = delay[126-2*i] * w_1 - b_r * w_2 + bias;
1138 delay[126-2*i] = d_i;
1139 }
1140 }
1141
1142 static double besselI0 (double x)
1143 {
1144 double bessel = 1;
1145 int i = 100;
1146
1147 do
1148 bessel = bessel * x / (i * i) + 1;
1149 while (--i);
1150 return bessel;
1151 }
1152
1153 void a52_imdct_init (uint32_t mm_accel)
1154 {
1155 int i, j, k;
1156 double sum;
1157
1158 /* compute imdct window - kaiser-bessel derived window, alpha = 5.0 */
1159 sum = 0;
1160 for (i = 0; i < 256; i++) {
1161 sum += besselI0 (i * (256 - i) * (5 * M_PI / 256) * (5 * M_PI / 256));
1162 a52_imdct_window[i] = sum;
1163 }
1164 sum++;
1165 for (i = 0; i < 256; i++)
1166 a52_imdct_window[i] = sqrt (a52_imdct_window[i] / sum);
1167
1168 for (i = 0; i < 3; i++)
1169 roots16[i] = cos ((M_PI / 8) * (i + 1));
1170
1171 for (i = 0; i < 7; i++)
1172 roots32[i] = cos ((M_PI / 16) * (i + 1));
1173
1174 for (i = 0; i < 15; i++)
1175 roots64[i] = cos ((M_PI / 32) * (i + 1));
1176
1177 for (i = 0; i < 31; i++)
1178 roots128[i] = cos ((M_PI / 64) * (i + 1));
1179
1180 for (i = 0; i < 64; i++) {
1181 k = fftorder[i] / 2 + 64;
1182 pre1[i].real = cos ((M_PI / 256) * (k - 0.25));
1183 pre1[i].imag = sin ((M_PI / 256) * (k - 0.25));
1184 }
1185
1186 for (i = 64; i < 128; i++) {
1187 k = fftorder[i] / 2 + 64;
1188 pre1[i].real = -cos ((M_PI / 256) * (k - 0.25));
1189 pre1[i].imag = -sin ((M_PI / 256) * (k - 0.25));
1190 }
1191
1192 for (i = 0; i < 64; i++) {
1193 post1[i].real = cos ((M_PI / 256) * (i + 0.5));
1194 post1[i].imag = sin ((M_PI / 256) * (i + 0.5));
1195 }
1196
1197 for (i = 0; i < 64; i++) {
1198 k = fftorder[i] / 4;
1199 pre2[i].real = cos ((M_PI / 128) * (k - 0.25));
1200 pre2[i].imag = sin ((M_PI / 128) * (k - 0.25));
1201 }
1202
1203 for (i = 0; i < 32; i++) {
1204 post2[i].real = cos ((M_PI / 128) * (i + 0.5));
1205 post2[i].imag = sin ((M_PI / 128) * (i + 0.5));
1206 }
1207 for (i = 0; i < 128; i++) {
1208 xcos1[i] = -cos ((M_PI / 2048) * (8 * i + 1));
1209 xsin1[i] = -sin ((M_PI / 2048) * (8 * i + 1));
1210 }
1211 for (i = 0; i < 7; i++) {
1212 j = 1 << i;
1213 for (k = 0; k < j; k++) {
1214 w[i][k].real = cos (-M_PI * k / j);
1215 w[i][k].imag = sin (-M_PI * k / j);
1216 }
1217 }
1218 #if defined(ARCH_X86) || defined(ARCH_X86_64)
1219 for (i = 0; i < 128; i++) {
1220 sseSinCos1c[2*i+0]= xcos1[i];
1221 sseSinCos1c[2*i+1]= -xcos1[i];
1222 sseSinCos1d[2*i+0]= xsin1[i];
1223 sseSinCos1d[2*i+1]= xsin1[i];
1224 }
1225 for (i = 1; i < 7; i++) {
1226 j = 1 << i;
1227 for (k = 0; k < j; k+=2) {
1228
1229 sseW[i][4*k + 0] = w[i][k+0].real;
1230 sseW[i][4*k + 1] = w[i][k+0].real;
1231 sseW[i][4*k + 2] = w[i][k+1].real;
1232 sseW[i][4*k + 3] = w[i][k+1].real;
1233
1234 sseW[i][4*k + 4] = -w[i][k+0].imag;
1235 sseW[i][4*k + 5] = w[i][k+0].imag;
1236 sseW[i][4*k + 6] = -w[i][k+1].imag;
1237 sseW[i][4*k + 7] = w[i][k+1].imag;
1238
1239 //we multiply more or less uninitalized numbers so we need to use exactly 0.0
1240 if(k==0)
1241 {
1242 // sseW[i][4*k + 0]= sseW[i][4*k + 1]= 1.0;
1243 sseW[i][4*k + 4]= sseW[i][4*k + 5]= 0.0;
1244 }
1245
1246 if(2*k == j)
1247 {
1248 sseW[i][4*k + 0]= sseW[i][4*k + 1]= 0.0;
1249 // sseW[i][4*k + 4]= -(sseW[i][4*k + 5]= -1.0);
1250 }
1251 }
1252 }
1253
1254 for(i=0; i<128; i++)
1255 {
1256 sseWindow[2*i+0]= -a52_imdct_window[2*i+0];
1257 sseWindow[2*i+1]= a52_imdct_window[2*i+1];
1258 }
1259
1260 for(i=0; i<64; i++)
1261 {
1262 sseWindow[256 + 2*i+0]= -a52_imdct_window[254 - 2*i+1];
1263 sseWindow[256 + 2*i+1]= a52_imdct_window[254 - 2*i+0];
1264 sseWindow[384 + 2*i+0]= a52_imdct_window[126 - 2*i+1];
1265 sseWindow[384 + 2*i+1]= -a52_imdct_window[126 - 2*i+0];
1266 }
1267 #endif
1268 a52_imdct_512 = imdct_do_512;
1269 ifft128 = ifft128_c;
1270 ifft64 = ifft64_c;
1271
1272 #if defined(ARCH_X86) || defined(ARCH_X86_64)
1273 if(mm_accel & MM_ACCEL_X86_SSE)
1274 {
1275 fprintf (stderr, "Using SSE optimized IMDCT transform\n");
1276 a52_imdct_512 = imdct_do_512_sse;
1277 }
1278 else
1279 if(mm_accel & MM_ACCEL_X86_3DNOWEXT)
1280 {
1281 fprintf (stderr, "Using 3DNowEx optimized IMDCT transform\n");
1282 a52_imdct_512 = imdct_do_512_3dnowex;
1283 }
1284 else
1285 if(mm_accel & MM_ACCEL_X86_3DNOW)
1286 {
1287 fprintf (stderr, "Using 3DNow optimized IMDCT transform\n");
1288 a52_imdct_512 = imdct_do_512_3dnow;
1289 }
1290 else
1291 #endif // ARCH_X86 || ARCH_X86_64
1292 #ifdef HAVE_ALTIVEC
1293 if (mm_accel & MM_ACCEL_PPC_ALTIVEC)
1294 {
1295 fprintf(stderr, "Using AltiVec optimized IMDCT transform\n");
1296 a52_imdct_512 = imdct_do_512_altivec;
1297 }
1298 else
1299 #endif
1300
1301 #ifdef LIBA52_DJBFFT
1302 if (mm_accel & MM_ACCEL_DJBFFT) {
1303 fprintf (stderr, "Using djbfft for IMDCT transform\n");
1304 ifft128 = (void (*) (complex_t *)) fftc4_un128;
1305 ifft64 = (void (*) (complex_t *)) fftc4_un64;
1306 } else
1307 #endif
1308 {
1309 fprintf (stderr, "No accelerated IMDCT transform found\n");
1310 }
1311 }