?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Speed up AAC-HE SBR by 2% on S5L8701. Use MEM_ALIGN on critical arrays and avoid...
[kugel-rb.git] / apps / codecs / libfaad / sbr_qmf.c
blobbb6e176a3fb53b80e0deec27e26bdb428ef0c1bc
1 /*
2 ** FAAD2 - Freeware Advanced Audio (AAC) Decoder including SBR decoding
3 ** Copyright (C) 2003-2004 M. Bakker, Ahead Software AG, http://www.nero.com
4 **
5 ** This program is free software; you can redistribute it and/or modify
6 ** it under the terms of the GNU General Public License as published by
7 ** the Free Software Foundation; either version 2 of the License, or
8 ** (at your option) any later version.
9 **
10 ** This program is distributed in the hope that it will be useful,
11 ** but WITHOUT ANY WARRANTY; without even the implied warranty of
12 ** MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 ** GNU General Public License for more details.
14 **
15 ** You should have received a copy of the GNU General Public License
16 ** along with this program; if not, write to the Free Software
17 ** Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
18 **
19 ** Any non-GPL usage of this software or parts of this software is strictly
20 ** forbidden.
21 **
22 ** Commercial non-GPL licensing of this software is possible.
23 ** For more info contact Ahead Software through Mpeg4AAClicense@nero.com.
24 **
25 ** $Id$
26 **/
27
28 #include "common.h"
29 #include "structs.h"
30
31 #ifdef SBR_DEC
32
33
34 #include <stdlib.h>
35 #include <string.h>
36 #include "sbr_dct.h"
37 #include "sbr_qmf.h"
38 #include "sbr_qmf_c.h"
39 #include "sbr_syntax.h"
40
41 #ifdef FIXED_POINT
42 #define FAAD_SYNTHESIS_SCALE(X) ((X)>>1)
43 #define FAAD_ANALYSIS_SCALE1(X) ((X)>>4)
44 #define FAAD_ANALYSIS_SCALE2(X) ((X))
45 #define FAAD_ANALYSIS_SCALE3(X) ((X))
46 #else
47 #define FAAD_SYNTHESIS_SCALE(X) ((X)/64.0f)
48 #define FAAD_ANALYSIS_SCALE1(X) ((X))
49 #define FAAD_ANALYSIS_SCALE2(X) (2.0f*(X))
50 #define FAAD_ANALYSIS_SCALE3(X) ((X)/32.0f)
51 #endif
52
53 qmfa_info *qmfa_init(uint8_t channels)
54 {
55 qmfa_info *qmfa = (qmfa_info*)faad_malloc(sizeof(qmfa_info));
56
57 /* x is implemented as double ringbuffer */
58 qmfa->x = (real_t*)faad_malloc(2 * channels * 10 * sizeof(real_t));
59 memset(qmfa->x, 0, 2 * channels * 10 * sizeof(real_t));
60
61 /* ringbuffer index */
62 qmfa->x_index = 0;
63
64 qmfa->channels = channels;
65
66 return qmfa;
67 }
68
69 void qmfa_end(qmfa_info *qmfa)
70 {
71 if (qmfa)
72 {
73 if (qmfa->x) faad_free(qmfa->x);
74 faad_free(qmfa);
75 }
76 }
77
78 void sbr_qmf_analysis_32(sbr_info *sbr, qmfa_info *qmfa, const real_t *input,
79 qmf_t X[MAX_NTSRHFG][64], uint8_t offset, uint8_t kx)
80 {
81 real_t u[64] MEM_ALIGN_ATTR;
82 #ifndef SBR_LOW_POWER
83 real_t real[32] MEM_ALIGN_ATTR;
84 real_t imag[32] MEM_ALIGN_ATTR;
85 #else
86 real_t y[32] MEM_ALIGN_ATTR;
87 #endif
88 qmf_t *pX;
89 uint32_t in = 0;
90 uint32_t l, idx0, idx1;
91
92 /* qmf subsample l */
93 for (l = 0; l < sbr->numTimeSlotsRate; l++)
94 {
95 int32_t n;
96
97 /* shift input buffer x */
98 /* input buffer is not shifted anymore, x is implemented as double ringbuffer */
99 //memmove(qmfa->x + 32, qmfa->x, (320-32)*sizeof(real_t));
100
101 /* add new samples to input buffer x */
102 idx0 = qmfa->x_index + 31; idx1 = idx0 + 320;
103 for (n = 0; n < 32; n+=4)
104 {
105 qmfa->x[idx0--] = qmfa->x[idx1--] = (input[in++]);
106 qmfa->x[idx0--] = qmfa->x[idx1--] = (input[in++]);
107 qmfa->x[idx0--] = qmfa->x[idx1--] = (input[in++]);
108 qmfa->x[idx0--] = qmfa->x[idx1--] = (input[in++]);
109 }
110
111 /* window and summation to create array u */
112 for (n = 0; n < 32; n++)
113 {
114 idx0 = qmfa->x_index + n; idx1 = n * 20;
115 u[n] = FAAD_ANALYSIS_SCALE1(
116 MUL_F(qmfa->x[idx0 ], qmf_c[idx1 ]) +
117 MUL_F(qmfa->x[idx0 + 64], qmf_c[idx1 + 2]) +
118 MUL_F(qmfa->x[idx0 + 128], qmf_c[idx1 + 4]) +
119 MUL_F(qmfa->x[idx0 + 192], qmf_c[idx1 + 6]) +
120 MUL_F(qmfa->x[idx0 + 256], qmf_c[idx1 + 8]));
121 }
122 for (n = 32; n < 64; n++)
123 {
124 idx0 = qmfa->x_index + n; idx1 = n * 20 - 639;
125 u[n] = FAAD_ANALYSIS_SCALE1(
126 MUL_F(qmfa->x[idx0 ], qmf_c[idx1 ]) +
127 MUL_F(qmfa->x[idx0 + 64], qmf_c[idx1 + 2]) +
128 MUL_F(qmfa->x[idx0 + 128], qmf_c[idx1 + 4]) +
129 MUL_F(qmfa->x[idx0 + 192], qmf_c[idx1 + 6]) +
130 MUL_F(qmfa->x[idx0 + 256], qmf_c[idx1 + 8]));
131 }
132
133 /* update ringbuffer index */
134 qmfa->x_index -= 32;
135 if (qmfa->x_index < 0)
136 qmfa->x_index = (320-32);
137
138 /* calculate 32 subband samples by introducing X */
139 #ifdef SBR_LOW_POWER
140 y[0] = u[48];
141 for (n = 1; n < 16; n++)
142 y[n] = u[n+48] + u[48-n];
143 for (n = 16; n < 32; n++)
144 y[n] = -u[n-16] + u[48-n];
145
146 DCT3_32_unscaled(u, y);
147
148 for (n = 0; n < 32; n++)
149 {
150 if (n < kx)
151 {
152 QMF_RE(X[l + offset][n]) = FAAD_ANALYSIS_SCALE2(u[n]);
153 } else {
154 QMF_RE(X[l + offset][n]) = 0;
155 }
156 }
157 #else /* #ifdef SBR_LOW_POWER */
158
159 // Reordering of data moved from DCT_IV to here
160 idx0 = 30; idx1 = 63;
161 imag[31] = u[ 1]; real[ 0] = u[ 0];
162 for (n = 1; n < 31; n+=3)
163 {
164 imag[idx0--] = u[n+1]; real[n ] = -u[idx1--];
165 imag[idx0--] = u[n+2]; real[n+1] = -u[idx1--];
166 imag[idx0--] = u[n+3]; real[n+2] = -u[idx1--];
167 }
168 imag[ 0] = u[32]; real[31] = -u[33];
169
170 // dct4_kernel is DCT_IV without reordering which is done before and after FFT
171 dct4_kernel(real, imag);
172
173 // Reordering of data moved from DCT_IV to here
174 /* Step 1: Calculate all non-zero pairs */
175 pX = X[l + offset];
176 for (n = 0; n < kx/2; n++) {
177 idx0 = 2*n; idx1 = idx0 + 1;
178 QMF_RE(pX[idx0]) = FAAD_ANALYSIS_SCALE2( real[n ]);
179 QMF_IM(pX[idx0]) = FAAD_ANALYSIS_SCALE2( imag[n ]);
180 QMF_RE(pX[idx1]) = FAAD_ANALYSIS_SCALE2(-imag[31-n]);
181 QMF_IM(pX[idx1]) = FAAD_ANALYSIS_SCALE2(-real[31-n]);
182 }
183 /* Step 2: Calculate a single pair with half zero'ed */
184 if (kx&1) {
185 idx0 = 2*n; idx1 = idx0 + 1;
186 QMF_RE(pX[idx0]) = FAAD_ANALYSIS_SCALE2( real[n]);
187 QMF_IM(pX[idx0]) = FAAD_ANALYSIS_SCALE2( imag[n]);
188 QMF_RE(pX[idx1]) = QMF_IM(pX[idx1]) = 0;
189 n++;
190 }
191 /* Step 3: All other are zero'ed */
192 for (; n < 16; n++) {
193 idx0 = 2*n; idx1 = idx0 + 1;
194 QMF_RE(pX[idx0]) = QMF_IM(pX[idx0]) = 0;
195 QMF_RE(pX[idx1]) = QMF_IM(pX[idx1]) = 0;
196 }
197 #endif /* #ifdef SBR_LOW_POWER */
198 }
199 }
200
201 qmfs_info *qmfs_init(uint8_t channels)
202 {
203 qmfs_info *qmfs = (qmfs_info*)faad_malloc(sizeof(qmfs_info));
204
205 /* v is a double ringbuffer */
206 qmfs->v = (real_t*)faad_malloc(2 * channels * 20 * sizeof(real_t));
207 memset(qmfs->v, 0, 2 * channels * 20 * sizeof(real_t));
208
209 qmfs->v_index = 0;
210
211 qmfs->channels = channels;
212
213 return qmfs;
214 }
215
216 void qmfs_end(qmfs_info *qmfs)
217 {
218 if (qmfs)
219 {
220 if (qmfs->v) faad_free(qmfs->v);
221 faad_free(qmfs);
222 }
223 }
224
225 #ifdef SBR_LOW_POWER
226
227 void sbr_qmf_synthesis_32(sbr_info *sbr, qmfs_info *qmfs, qmf_t X[MAX_NTSRHFG][64],
228 real_t *output)
229 {
230 real_t x[16] MEM_ALIGN_ATTR;
231 real_t y[16] MEM_ALIGN_ATTR;
232 int16_t n, k, out = 0;
233 uint8_t l;
234
235 /* qmf subsample l */
236 for (l = 0; l < sbr->numTimeSlotsRate; l++)
237 {
238 /* shift buffers */
239 /* we are not shifting v, it is a double ringbuffer */
240 //memmove(qmfs->v + 64, qmfs->v, (640-64)*sizeof(real_t));
241
242 /* calculate 64 samples */
243 for (k = 0; k < 16; k++)
244 {
245 y[k] = FAAD_ANALYSIS_SCALE3((QMF_RE(X[l][k]) - QMF_RE(X[l][31-k])));
246 x[k] = FAAD_ANALYSIS_SCALE3((QMF_RE(X[l][k]) + QMF_RE(X[l][31-k])));
247 }
248
249 /* even n samples */
250 DCT2_16_unscaled(x, x);
251 /* odd n samples */
252 DCT4_16(y, y);
253
254 for (n = 8; n < 24; n++)
255 {
256 qmfs->v[qmfs->v_index + n*2 ] = qmfs->v[qmfs->v_index + 640 + n*2 ] = x[n-8];
257 qmfs->v[qmfs->v_index + n*2+1] = qmfs->v[qmfs->v_index + 640 + n*2+1] = y[n-8];
258 }
259 for (n = 0; n < 16; n++)
260 {
261 qmfs->v[qmfs->v_index + n] = qmfs->v[qmfs->v_index + 640 + n] = qmfs->v[qmfs->v_index + 32-n];
262 }
263 qmfs->v[qmfs->v_index + 48] = qmfs->v[qmfs->v_index + 640 + 48] = 0;
264 for (n = 1; n < 16; n++)
265 {
266 qmfs->v[qmfs->v_index + 48+n] = qmfs->v[qmfs->v_index + 640 + 48+n] = -qmfs->v[qmfs->v_index + 48-n];
267 }
268
269 /* calculate 32 output samples and window */
270 for (k = 0; k < 32; k++)
271 {
272 output[out++] = MUL_F(qmfs->v[qmfs->v_index + k], qmf_c[ 2*k*10]) +
273 MUL_F(qmfs->v[qmfs->v_index + 96 + k], qmf_c[1 + 2*k*10]) +
274 MUL_F(qmfs->v[qmfs->v_index + 128 + k], qmf_c[2 + 2*k*10]) +
275 MUL_F(qmfs->v[qmfs->v_index + 224 + k], qmf_c[3 + 2*k*10]) +
276 MUL_F(qmfs->v[qmfs->v_index + 256 + k], qmf_c[4 + 2*k*10]) +
277 MUL_F(qmfs->v[qmfs->v_index + 352 + k], qmf_c[5 + 2*k*10]) +
278 MUL_F(qmfs->v[qmfs->v_index + 384 + k], qmf_c[6 + 2*k*10]) +
279 MUL_F(qmfs->v[qmfs->v_index + 480 + k], qmf_c[7 + 2*k*10]) +
280 MUL_F(qmfs->v[qmfs->v_index + 512 + k], qmf_c[8 + 2*k*10]) +
281 MUL_F(qmfs->v[qmfs->v_index + 608 + k], qmf_c[9 + 2*k*10]);
282 }
283
284 /* update the ringbuffer index */
285 qmfs->v_index -= 64;
286 if (qmfs->v_index < 0)
287 qmfs->v_index = (640-64);
288 }
289 }
290
291 void sbr_qmf_synthesis_64(sbr_info *sbr, qmfs_info *qmfs, qmf_t X[MAX_NTSRHFG][64],
292 real_t *output)
293 {
294 real_t x[64] MEM_ALIGN_ATTR;
295 real_t y[64] MEM_ALIGN_ATTR;
296 int16_t n, k, out = 0;
297 uint8_t l;
298
299
300 /* qmf subsample l */
301 for (l = 0; l < sbr->numTimeSlotsRate; l++)
302 {
303 /* shift buffers */
304 /* we are not shifting v, it is a double ringbuffer */
305 //memmove(qmfs->v + 128, qmfs->v, (1280-128)*sizeof(real_t));
306
307 /* calculate 128 samples */
308 for (k = 0; k < 32; k++)
309 {
310 y[k] = FAAD_ANALYSIS_SCALE3((QMF_RE(X[l][k]) - QMF_RE(X[l][63-k])));
311 x[k] = FAAD_ANALYSIS_SCALE3((QMF_RE(X[l][k]) + QMF_RE(X[l][63-k])));
312 }
313
314 /* even n samples */
315 DCT2_32_unscaled(x, x);
316 /* odd n samples */
317 DCT4_32(y, y);
318
319 for (n = 16; n < 48; n++)
320 {
321 qmfs->v[qmfs->v_index + n*2] = qmfs->v[qmfs->v_index + 1280 + n*2 ] = x[n-16];
322 qmfs->v[qmfs->v_index + n*2+1] = qmfs->v[qmfs->v_index + 1280 + n*2+1] = y[n-16];
323 }
324 for (n = 0; n < 32; n++)
325 {
326 qmfs->v[qmfs->v_index + n] = qmfs->v[qmfs->v_index + 1280 + n] = qmfs->v[qmfs->v_index + 64-n];
327 }
328 qmfs->v[qmfs->v_index + 96] = qmfs->v[qmfs->v_index + 1280 + 96] = 0;
329 for (n = 1; n < 32; n++)
330 {
331 qmfs->v[qmfs->v_index + 96+n] = qmfs->v[qmfs->v_index + 1280 + 96+n] = -qmfs->v[qmfs->v_index + 96-n];
332 }
333
334 /* calculate 64 output samples and window */
335 for (k = 0; k < 64; k++)
336 {
337 output[out++] = MUL_F(qmfs->v[qmfs->v_index + k], qmf_c[ k*10]) +
338 MUL_F(qmfs->v[qmfs->v_index + 192 + k], qmf_c[1 + k*10]) +
339 MUL_F(qmfs->v[qmfs->v_index + 256 + k], qmf_c[2 + k*10]) +
340 MUL_F(qmfs->v[qmfs->v_index + 256 + 192 + k], qmf_c[3 + k*10]) +
341 MUL_F(qmfs->v[qmfs->v_index + 512 + k], qmf_c[4 + k*10]) +
342 MUL_F(qmfs->v[qmfs->v_index + 512 + 192 + k], qmf_c[5 + k*10]) +
343 MUL_F(qmfs->v[qmfs->v_index + 768 + k], qmf_c[6 + k*10]) +
344 MUL_F(qmfs->v[qmfs->v_index + 768 + 192 + k], qmf_c[7 + k*10]) +
345 MUL_F(qmfs->v[qmfs->v_index + 1024 + k], qmf_c[8 + k*10]) +
346 MUL_F(qmfs->v[qmfs->v_index + 1024 + 192 + k], qmf_c[9 + k*10]);
347 }
348
349 /* update the ringbuffer index */
350 qmfs->v_index -= 128;
351 if (qmfs->v_index < 0)
352 qmfs->v_index = (1280-128);
353 }
354 }
355 #else /* #ifdef SBR_LOW_POWER */
356
357 static const complex_t qmf32_pre_twiddle[] =
358 {
359 { FRAC_CONST(0.999924701839145), FRAC_CONST(-0.012271538285720) },
360 { FRAC_CONST(0.999322384588350), FRAC_CONST(-0.036807222941359) },
361 { FRAC_CONST(0.998118112900149), FRAC_CONST(-0.061320736302209) },
362 { FRAC_CONST(0.996312612182778), FRAC_CONST(-0.085797312344440) },
363 { FRAC_CONST(0.993906970002356), FRAC_CONST(-0.110222207293883) },
364 { FRAC_CONST(0.990902635427780), FRAC_CONST(-0.134580708507126) },
365 { FRAC_CONST(0.987301418157858), FRAC_CONST(-0.158858143333861) },
366 { FRAC_CONST(0.983105487431216), FRAC_CONST(-0.183039887955141) },
367 { FRAC_CONST(0.978317370719628), FRAC_CONST(-0.207111376192219) },
368 { FRAC_CONST(0.972939952205560), FRAC_CONST(-0.231058108280671) },
369 { FRAC_CONST(0.966976471044852), FRAC_CONST(-0.254865659604515) },
370 { FRAC_CONST(0.960430519415566), FRAC_CONST(-0.278519689385053) },
371 { FRAC_CONST(0.953306040354194), FRAC_CONST(-0.302005949319228) },
372 { FRAC_CONST(0.945607325380521), FRAC_CONST(-0.325310292162263) },
373 { FRAC_CONST(0.937339011912575), FRAC_CONST(-0.348418680249435) },
374 { FRAC_CONST(0.928506080473216), FRAC_CONST(-0.371317193951838) },
375 { FRAC_CONST(0.919113851690058), FRAC_CONST(-0.393992040061048) },
376 { FRAC_CONST(0.909167983090522), FRAC_CONST(-0.416429560097637) },
377 { FRAC_CONST(0.898674465693954), FRAC_CONST(-0.438616238538528) },
378 { FRAC_CONST(0.887639620402854), FRAC_CONST(-0.460538710958240) },
379 { FRAC_CONST(0.876070094195407), FRAC_CONST(-0.482183772079123) },
380 { FRAC_CONST(0.863972856121587), FRAC_CONST(-0.503538383725718) },
381 { FRAC_CONST(0.851355193105265), FRAC_CONST(-0.524589682678469) },
382 { FRAC_CONST(0.838224705554838), FRAC_CONST(-0.545324988422046) },
383 { FRAC_CONST(0.824589302785025), FRAC_CONST(-0.565731810783613) },
384 { FRAC_CONST(0.810457198252595), FRAC_CONST(-0.585797857456439) },
385 { FRAC_CONST(0.795836904608884), FRAC_CONST(-0.605511041404326) },
386 { FRAC_CONST(0.780737228572094), FRAC_CONST(-0.624859488142386) },
387 { FRAC_CONST(0.765167265622459), FRAC_CONST(-0.643831542889791) },
388 { FRAC_CONST(0.749136394523459), FRAC_CONST(-0.662415777590172) },
389 { FRAC_CONST(0.732654271672413), FRAC_CONST(-0.680600997795453) },
390 { FRAC_CONST(0.715730825283819), FRAC_CONST(-0.698376249408973) }
391 };
392
393 #define FAAD_CMPLX_PRETWIDDLE_SUB(k) \
394 (MUL_F(QMF_RE(X[l][k]), RE(qmf32_pre_twiddle[k])) - \
395 MUL_F(QMF_IM(X[l][k]), IM(qmf32_pre_twiddle[k])))
396
397 #define FAAD_CMPLX_PRETWIDDLE_ADD(k) \
398 (MUL_F(QMF_IM(X[l][k]), RE(qmf32_pre_twiddle[k])) + \
399 MUL_F(QMF_RE(X[l][k]), IM(qmf32_pre_twiddle[k])))
400
401 void sbr_qmf_synthesis_32(sbr_info *sbr, qmfs_info *qmfs, qmf_t X[MAX_NTSRHFG][64],
402 real_t *output)
403 {
404 real_t x1[32] MEM_ALIGN_ATTR;
405 real_t x2[32] MEM_ALIGN_ATTR;
406 int32_t n, k, idx0, idx1, out = 0;
407 uint32_t l;
408
409 /* qmf subsample l */
410 for (l = 0; l < sbr->numTimeSlotsRate; l++)
411 {
412 /* shift buffer v */
413 /* buffer is not shifted, we are using a ringbuffer */
414 //memmove(qmfs->v + 64, qmfs->v, (640-64)*sizeof(real_t));
415
416 /* calculate 64 samples */
417 /* complex pre-twiddle */
418 for (k = 0; k < 32;)
419 {
420 x1[k] = FAAD_CMPLX_PRETWIDDLE_SUB(k); x2[k] = FAAD_CMPLX_PRETWIDDLE_ADD(k); k++;
421 x1[k] = FAAD_CMPLX_PRETWIDDLE_SUB(k); x2[k] = FAAD_CMPLX_PRETWIDDLE_ADD(k); k++;
422 x1[k] = FAAD_CMPLX_PRETWIDDLE_SUB(k); x2[k] = FAAD_CMPLX_PRETWIDDLE_ADD(k); k++;
423 x1[k] = FAAD_CMPLX_PRETWIDDLE_SUB(k); x2[k] = FAAD_CMPLX_PRETWIDDLE_ADD(k); k++;
424 }
425
426 /* transform */
427 DCT4_32(x1, x1);
428 DST4_32(x2, x2);
429
430 idx0 = qmfs->v_index;
431 idx1 = qmfs->v_index + 63;
432 for (n = 0; n < 32; n+=2)
433 {
434 qmfs->v[idx0] = qmfs->v[idx0 + 640] = -x1[n ] + x2[n ]; idx0++;
435 qmfs->v[idx1] = qmfs->v[idx1 + 640] = x1[n ] + x2[n ]; idx1--;
436 qmfs->v[idx0] = qmfs->v[idx0 + 640] = -x1[n+1] + x2[n+1]; idx0++;
437 qmfs->v[idx1] = qmfs->v[idx1 + 640] = x1[n+1] + x2[n+1]; idx1--;
438 }
439
440 /* calculate 32 output samples and window */
441 for (k = 0; k < 32; k++)
442 {
443 idx0 = qmfs->v_index + k; idx1 = 2*k*10;
444 output[out++] = FAAD_SYNTHESIS_SCALE(
445 MUL_F(qmfs->v[idx0 ], qmf_c[idx1 ]) +
446 MUL_F(qmfs->v[idx0 + 96], qmf_c[idx1+1]) +
447 MUL_F(qmfs->v[idx0 + 128], qmf_c[idx1+2]) +
448 MUL_F(qmfs->v[idx0 + 224], qmf_c[idx1+3]) +
449 MUL_F(qmfs->v[idx0 + 256], qmf_c[idx1+4]) +
450 MUL_F(qmfs->v[idx0 + 352], qmf_c[idx1+5]) +
451 MUL_F(qmfs->v[idx0 + 384], qmf_c[idx1+6]) +
452 MUL_F(qmfs->v[idx0 + 480], qmf_c[idx1+7]) +
453 MUL_F(qmfs->v[idx0 + 512], qmf_c[idx1+8]) +
454 MUL_F(qmfs->v[idx0 + 608], qmf_c[idx1+9]));
455 }
456
457 /* update ringbuffer index */
458 qmfs->v_index -= 64;
459 if (qmfs->v_index < 0)
460 qmfs->v_index = (640 - 64);
461 }
462 }
463
464 void sbr_qmf_synthesis_64(sbr_info *sbr, qmfs_info *qmfs, qmf_t X[MAX_NTSRHFG][64],
465 real_t *output)
466 {
467 real_t real1[32] MEM_ALIGN_ATTR;
468 real_t imag1[32] MEM_ALIGN_ATTR;
469 real_t real2[32] MEM_ALIGN_ATTR;
470 real_t imag2[32] MEM_ALIGN_ATTR;
471 qmf_t *pX;
472 real_t *p_buf_1, *p_buf_3;
473 int32_t n, k, idx0, idx1, out = 0;
474 uint32_t l;
475
476 /* qmf subsample l */
477 for (l = 0; l < sbr->numTimeSlotsRate; l++)
478 {
479 /* shift buffer v */
480 /* buffer is not shifted, we use double ringbuffer */
481 //memmove(qmfs->v + 128, qmfs->v, (1280-128)*sizeof(real_t));
482
483 /* calculate 128 samples */
484 pX = X[l];
485 for (k = 0; k < 32; k++)
486 {
487 idx0 = 2*k; idx1 = idx0+1;
488 real1[ k] = QMF_RE(pX[idx0]); imag2[ k] = QMF_IM(pX[idx0]);
489 imag1[31-k] = QMF_RE(pX[idx1]); real2[31-k] = QMF_IM(pX[idx1]);
490 }
491
492 // dct4_kernel is DCT_IV without reordering which is done before and after FFT
493 dct4_kernel(real1, imag1);
494 dct4_kernel(real2, imag2);
495
496 p_buf_1 = qmfs->v + qmfs->v_index;
497 p_buf_3 = p_buf_1 + 1280;
498
499 idx0 = 0; idx1 = 127;
500 for (n = 0; n < 32; n++)
501 {
502 p_buf_1[idx0] = p_buf_3[idx0] = real2[ n] - real1[ n]; idx0++;
503 p_buf_1[idx1] = p_buf_3[idx1] = real2[ n] + real1[ n]; idx1--;
504 p_buf_1[idx0] = p_buf_3[idx0] = imag2[31-n] + imag1[31-n]; idx0++;
505 p_buf_1[idx1] = p_buf_3[idx1] = imag2[31-n] - imag1[31-n]; idx1--;
506 }
507
508 p_buf_1 = qmfs->v + qmfs->v_index;
509
510 /* calculate 64 output samples and window */
511 #ifdef CPU_ARM
512 const real_t *qtab = qmf_c;
513 real_t *pbuf = p_buf_1;
514 for (k = 0; k < 64; k++, pbuf++)
515 {
516 real_t *pout = &output[out++];
517 asm volatile (
518 "ldmia %[qtab]!, { r0-r3 } \n\t"
519 "ldr r4, [%[pbuf]] \n\t"
520 "ldr r7, [%[pbuf], #192*4] \n\t"
521 "smull r5, r6, r4, r0 \n\t"
522 "ldr r4, [%[pbuf], #256*4] \n\t"
523 "smlal r5, r6, r7, r1 \n\t"
524 "ldr r7, [%[pbuf], #448*4] \n\t"
525 "smlal r5, r6, r4, r2 \n\t"
526 "ldr r4, [%[pbuf], #512*4] \n\t"
527 "smlal r5, r6, r7, r3 \n\t"
528
529 "ldmia %[qtab]!, { r0-r3 } \n\t"
530 "ldr r7, [%[pbuf], #704*4] \n\t"
531 "smlal r5, r6, r4, r0 \n\t"
532 "ldr r4, [%[pbuf], #768*4] \n\t"
533 "smlal r5, r6, r7, r1 \n\t"
534 "ldr r7, [%[pbuf], #960*4] \n\t"
535 "smlal r5, r6, r4, r2 \n\t"
536 "mov r2, #1024*4 \n\t"
537
538 "ldmia %[qtab]!, { r0-r1 } \n\t"
539 "ldr r4, [%[pbuf], r2] \n\t"
540 "smlal r5, r6, r7, r3 \n\t"
541 "mov r2, #1216*4 \n\t"
542 "ldr r7, [%[pbuf], r2] \n\t"
543 "smlal r5, r6, r4, r0 \n\t"
544 "smlal r5, r6, r7, r1 \n\t"
545
546 "str r6, [%[pout]] \n"
547 : [qtab] "+r" (qtab)
548 : [pbuf] "r" (pbuf), [pout] "r" (pout)
549 : "r0", "r1", "r2", "r3", "r4", "r5", "r6", "r7", "memory");
550 }
551 #elif defined CPU_COLDFIRE
552 const real_t *qtab = qmf_c;
553 real_t *pbuf = p_buf_1;
554 for (k = 0; k < 64; k++, pbuf++)
555 {
556 real_t *pout = &output[out++];
557 asm volatile (
558 "move.l (%[pbuf]), %%d5 \n"
559
560 "movem.l (%[qtab]), %%d0-%%d4 \n"
561 "mac.l %%d0, %%d5, (192*4, %[pbuf]), %%d5, %%acc0 \n"
562 "mac.l %%d1, %%d5, (256*4, %[pbuf]), %%d5, %%acc0 \n"
563 "mac.l %%d2, %%d5, (448*4, %[pbuf]), %%d5, %%acc0 \n"
564 "mac.l %%d3, %%d5, (512*4, %[pbuf]), %%d5, %%acc0 \n"
565 "mac.l %%d4, %%d5, (704*4, %[pbuf]), %%d5, %%acc0 \n"
566 "lea.l (20, %[qtab]), %[qtab] \n"
567
568 "movem.l (%[qtab]), %%d0-%%d4 \n"
569 "mac.l %%d0, %%d5, (768*4, %[pbuf]), %%d5, %%acc0 \n"
570 "mac.l %%d1, %%d5, (960*4, %[pbuf]), %%d5, %%acc0 \n"
571 "mac.l %%d2, %%d5, (1024*4, %[pbuf]), %%d5, %%acc0 \n"
572 "mac.l %%d3, %%d5, (1216*4, %[pbuf]), %%d5, %%acc0 \n"
573 "mac.l %%d4, %%d5, %%acc0 \n"
574 "lea.l (20, %[qtab]), %[qtab] \n"
575
576 "movclr.l %%acc0, %%d0 \n"
577 "move.l %%d0, (%[pout]) \n"
578 : [qtab] "+a" (qtab)
579 : [pbuf] "a" (pbuf),
580 [pout] "a" (pout)
581 : "d0", "d1", "d2", "d3", "d4", "d5", "memory");
582 }
583 #else
584 for (k = 0; k < 64; k++)
585 {
586 idx0 = k*10;
587 output[out++] = FAAD_SYNTHESIS_SCALE(
588 MUL_F(p_buf_1[k ], qmf_c[idx0 ]) +
589 MUL_F(p_buf_1[k+ 192 ], qmf_c[idx0+1]) +
590 MUL_F(p_buf_1[k+ 256 ], qmf_c[idx0+2]) +
591 MUL_F(p_buf_1[k+ 256+192], qmf_c[idx0+3]) +
592 MUL_F(p_buf_1[k+ 512 ], qmf_c[idx0+4]) +
593 MUL_F(p_buf_1[k+ 512+192], qmf_c[idx0+5]) +
594 MUL_F(p_buf_1[k+ 768 ], qmf_c[idx0+6]) +
595 MUL_F(p_buf_1[k+ 768+192], qmf_c[idx0+7]) +
596 MUL_F(p_buf_1[k+1024 ], qmf_c[idx0+8]) +
597 MUL_F(p_buf_1[k+1024+192], qmf_c[idx0+9]));
598 }
599 #endif
600
601 /* update ringbuffer index */
602 qmfs->v_index -= 128;
603 if (qmfs->v_index < 0)
604 qmfs->v_index = (1280 - 128);
605 }
606 }
607 #endif /* #ifdef SBR_LOW_POWER */
608
609 #endif /* #ifdef SBR_DEC */
My public Rockbox git repo