1 /******************************************************************************
2 * *
3 * Copyright (C) 2018 The Android Open Source Project
4 *
5 * Licensed under the Apache License, Version 2.0 (the "License");
6 * you may not use this file except in compliance with the License.
7 * You may obtain a copy of the License at:
8 *
9 * http://www.apache.org/licenses/LICENSE-2.0
10 *
11 * Unless required by applicable law or agreed to in writing, software
12 * distributed under the License is distributed on an "AS IS" BASIS,
13 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14 * See the License for the specific language governing permissions and
15 * limitations under the License.
16 *
17 *****************************************************************************
18 * Originally developed and contributed by Ittiam Systems Pvt. Ltd, Bangalore
19 */
20 #include <string.h>
21
22 #include "ixheaacd_sbr_common.h"
23 #include <ixheaacd_type_def.h>
24
25 #include "ixheaacd_constants.h"
26 #include "ixheaacd_basic_ops32.h"
27 #include "ixheaacd_basic_ops16.h"
28 #include "ixheaacd_basic_ops40.h"
29 #include "ixheaacd_basic_ops.h"
30
31 #include "ixheaacd_intrinsics.h"
32 #include "ixheaacd_common_rom.h"
33 #include "ixheaacd_bitbuffer.h"
34 #include "ixheaacd_sbrdecsettings.h"
35 #include "ixheaacd_sbr_scale.h"
36 #include "ixheaacd_lpp_tran.h"
37 #include "ixheaacd_env_extr_part.h"
38 #include "ixheaacd_sbr_rom.h"
39 #include "ixheaacd_hybrid.h"
40 #include "ixheaacd_ps_dec.h"
41 #include "ixheaacd_env_extr.h"
42 #include "ixheaacd_qmf_dec.h"
43
44 #include <ixheaacd_basic_op.h>
45 #include "ixheaacd_env_calc.h"
46
47 #include "ixheaacd_interface.h"
48
49 #include "ixheaacd_function_selector.h"
50 #include "ixheaacd_audioobjtypes.h"
51 #if !__ARM_NEON__
52
53 #define DCT3_LEN (32)
54 #define DCT2_LEN (64)
55
56 #define LP_SHIFT_VAL 7
57 #define HQ_SHIFT_64 4
58 #define RADIXSHIFT 1
59 #define HQ_SHIFT_VAL 4
60
ixheaacd_dct3_32(WORD32 * input,WORD32 * output,const WORD16 * main_twidle_fwd,const WORD16 * post_tbl,const WORD16 * w_16,const WORD32 * p_table)61 VOID ixheaacd_dct3_32(WORD32 *input, WORD32 *output,
62 const WORD16 *main_twidle_fwd, const WORD16 *post_tbl,
63 const WORD16 *w_16, const WORD32 *p_table) {
64 WORD32 n, k;
65
66 WORD32 temp1[6];
67 WORD32 temp2[4];
68 WORD16 twid_re, twid_im;
69 WORD32 *ptr_reverse, *ptr_forward, *p_out, *ptr_out1;
70 const WORD16 *twidle_fwd, *twidle_rev;
71
72 ptr_forward = &input[49];
73 ptr_reverse = &input[47];
74
75 p_out = output;
76 twidle_fwd = main_twidle_fwd;
77 twidle_fwd += 4;
78
79 *p_out++ = input[48] >> LP_SHIFT_VAL;
80 *p_out++ = 0;
81
82 for (n = 1; n < DCT3_LEN / 2; n++) {
83 temp1[0] = *ptr_forward++;
84 temp1[1] = *ptr_reverse--;
85 temp1[0] = ixheaacd_add32_sat(ixheaacd_shr32(temp1[0], LP_SHIFT_VAL),
86 ixheaacd_shr32(temp1[1], LP_SHIFT_VAL));
87
88 temp1[2] = *(ptr_forward - 33);
89 temp1[3] = *(ptr_reverse - 31);
90 temp1[1] = ixheaacd_sub32_sat(ixheaacd_shr32(temp1[2], LP_SHIFT_VAL),
91 ixheaacd_shr32(temp1[3], LP_SHIFT_VAL));
92 twid_re = *twidle_fwd++;
93
94 twid_im = *twidle_fwd;
95 twidle_fwd += 3;
96
97 *p_out++ = ixheaacd_mult32x16in32(temp1[0], twid_re) +
98 ixheaacd_mult32x16in32(temp1[1], twid_im);
99 *p_out++ = -ixheaacd_mult32x16in32(temp1[1], twid_re) +
100 ixheaacd_mult32x16in32(temp1[0], twid_im);
101 }
102
103 twid_re = *twidle_fwd++;
104
105 twid_im = *twidle_fwd;
106 twidle_fwd += 3;
107
108 temp1[1] = *ptr_reverse--;
109 temp1[0] = *(ptr_reverse - 31);
110 temp1[1] = ixheaacd_sub32_sat(ixheaacd_shr32(temp1[1], LP_SHIFT_VAL),
111 ixheaacd_shr32(temp1[0], LP_SHIFT_VAL));
112
113 temp1[0] = temp1[1];
114
115 temp2[2] = ixheaacd_mult32x16in32(temp1[0], twid_re) +
116 ixheaacd_mult32x16in32(temp1[1], twid_im);
117 temp2[3] = -ixheaacd_mult32x16in32(temp1[1], twid_re) +
118 ixheaacd_mult32x16in32(temp1[0], twid_im);
119
120 ptr_forward = output;
121 ptr_reverse = &output[DCT3_LEN - 1];
122
123 temp2[0] = *ptr_forward++;
124 temp2[1] = *ptr_forward--;
125
126 temp1[0] = -temp2[1] - temp2[3];
127 temp1[1] = temp2[0] - temp2[2];
128 temp2[0] = (temp2[0] + temp2[2] + temp1[0]);
129 temp2[1] = (temp2[1] - temp2[3] + temp1[1]);
130
131 temp2[0] >>= 1;
132 temp2[1] >>= 1;
133
134 *ptr_forward++ = temp2[0];
135 *ptr_forward++ = temp2[1];
136
137 twidle_fwd = post_tbl + 2;
138 twidle_rev = post_tbl + 14;
139
140 for (n = 1; n < DCT3_LEN / 4; n++) {
141 temp2[0] = *ptr_forward++;
142 temp2[1] = *ptr_forward--;
143 temp2[3] = *ptr_reverse--;
144 temp2[2] = *ptr_reverse++;
145
146 twid_re = *twidle_rev;
147 twidle_rev -= 2;
148 twid_im = *twidle_fwd;
149 twidle_fwd += 2;
150
151 temp1[0] = temp2[0] - temp2[2];
152 temp1[1] = (temp2[0] + temp2[2]);
153
154 temp1[2] = temp2[1] + temp2[3];
155 temp1[3] = (temp2[1] - temp2[3]);
156
157 temp1[4] = ixheaacd_mult32x16in32(temp1[0], twid_re) +
158 ixheaacd_mult32x16in32(temp1[2], twid_im);
159 temp1[5] = -ixheaacd_mult32x16in32(temp1[2], twid_re) +
160 ixheaacd_mult32x16in32(temp1[0], twid_im);
161
162 temp1[1] >>= 1;
163 temp1[3] >>= 1;
164
165 *ptr_forward++ = temp1[1] - temp1[4];
166 *ptr_forward++ = temp1[3] + temp1[5];
167
168 *ptr_reverse-- = -temp1[3] + temp1[5];
169 *ptr_reverse-- = temp1[1] + temp1[4];
170 }
171 temp2[0] = *ptr_forward++;
172 temp2[1] = *ptr_forward--;
173 temp2[3] = *ptr_reverse--;
174 temp2[2] = *ptr_reverse++;
175
176 twid_re = -*twidle_rev;
177 twidle_rev -= 2;
178 twid_im = *twidle_fwd;
179 twidle_fwd += 2;
180
181 temp1[0] = temp2[0] - temp2[2];
182 temp1[1] = (temp2[0] + temp2[2]);
183
184 temp1[2] = temp2[1] + temp2[3];
185 temp1[3] = (temp2[1] - temp2[3]);
186
187 temp1[4] = ixheaacd_mult32x16in32(temp1[0], twid_re) -
188 ixheaacd_mult32x16in32(temp1[2], twid_im);
189 temp1[5] = ixheaacd_mult32x16in32(temp1[2], twid_re) +
190 ixheaacd_mult32x16in32(temp1[0], twid_im);
191
192 temp1[1] >>= 1;
193 temp1[3] >>= 1;
194 *ptr_forward++ = temp1[1] + temp1[4];
195 *ptr_forward++ = temp1[3] + temp1[5];
196
197 ixheaacd_radix4bfly(w_16, output, 1, 4);
198 ixheaacd_postradixcompute4(input, output, p_table, 16);
199
200 output[0] = input[0];
201 output[2] = input[1];
202
203 p_out = input + 2;
204 ptr_forward = output + 1;
205 ptr_reverse = output + 30;
206 ptr_out1 = input + 18;
207
208 for (k = (DCT3_LEN / 4) - 1; k != 0; k--) {
209 WORD32 tempre, tempim;
210
211 tempre = *p_out++;
212 tempim = *p_out++;
213 *ptr_forward = (tempim);
214 ptr_forward += 2;
215 *ptr_forward = (tempre);
216 ptr_forward += 2;
217
218 tempre = *ptr_out1++;
219 tempim = *ptr_out1++;
220 *ptr_reverse = (tempim);
221 ptr_reverse -= 2;
222 *ptr_reverse = (tempre);
223 ptr_reverse -= 2;
224 }
225
226 {
227 WORD32 tempre, tempim;
228 tempre = *p_out++;
229 tempim = *p_out++;
230 *ptr_forward = (tempim);
231 ptr_forward += 2;
232 *ptr_forward = (tempre);
233 ptr_forward += 2;
234 }
235
236 return;
237 }
238
ixheaacd_dct2_64(WORD32 * x,WORD32 * X,ia_qmf_dec_tables_struct * qmf_dec_tables_ptr,WORD16 * filter_states)239 VOID ixheaacd_dct2_64(WORD32 *x, WORD32 *X,
240 ia_qmf_dec_tables_struct *qmf_dec_tables_ptr,
241 WORD16 *filter_states) {
242 ixheaacd_pretwdct2(x, X);
243
244 ixheaacd_radix4bfly(qmf_dec_tables_ptr->w_32, X, 1, 8);
245
246 ixheaacd_radix4bfly(qmf_dec_tables_ptr->w_32 + 48, X, 4, 2);
247
248 ixheaacd_postradixcompute2(x, X, qmf_dec_tables_ptr->dig_rev_table2_32, 32);
249
250 ixheaacd_fftposttw(x, qmf_dec_tables_ptr);
251
252 ixheaacd_posttwdct2(x, filter_states, qmf_dec_tables_ptr);
253
254 return;
255 }
256
ixheaacd_cos_sin_mod(WORD32 * subband,ia_sbr_qmf_filter_bank_struct * qmf_bank,WORD16 * p_twiddle,WORD32 * p_dig_rev_tbl)257 VOID ixheaacd_cos_sin_mod(WORD32 *subband,
258 ia_sbr_qmf_filter_bank_struct *qmf_bank,
259 WORD16 *p_twiddle, WORD32 *p_dig_rev_tbl) {
260 WORD32 re2, re3;
261 WORD16 wim, wre;
262
263 WORD32 i, M_2;
264 WORD32 M = ixheaacd_shr32(qmf_bank->no_channels, 1);
265
266 const WORD16 *p_sin;
267 const WORD16 *p_sin_cos = &qmf_bank->cos_twiddle[0];
268 WORD32 subband_tmp[128];
269 WORD32 re;
270 WORD32 im;
271 WORD32 *psubband, *psubband1;
272 WORD32 *psubband_t, *psubband1_t;
273 WORD32 *psubband2, *psubband12;
274 WORD32 *psubband_t2, *psubband1_t2;
275
276 M_2 = ixheaacd_shr32(M, 1);
277
278 psubband = &subband[0];
279 psubband1 = &subband[2 * M - 1];
280 psubband_t = subband_tmp;
281 psubband1_t = &subband_tmp[2 * M - 1];
282
283 psubband2 = &subband[64];
284 psubband12 = &subband[2 * M - 1 + 64];
285 psubband_t2 = &subband_tmp[64];
286 psubband1_t2 = &subband_tmp[2 * M - 1 + 64];
287
288 for (i = (M_2 >> 1) - 1; i >= 0; i--) {
289 re = *psubband++;
290 im = *psubband1--;
291
292 wim = *p_sin_cos++;
293 wre = *p_sin_cos++;
294
295 *psubband_t++ = ixheaacd_add32_sat(ixheaacd_mult32x16in32(re, wre),
296 ixheaacd_mult32x16in32(im, wim));
297 *psubband_t++ = ixheaacd_sub32_sat(ixheaacd_mult32x16in32(im, wre),
298 ixheaacd_mult32x16in32(re, wim));
299
300 re = *psubband2++;
301 im = *psubband12--;
302
303 *psubband_t2++ = ixheaacd_sub32_sat(ixheaacd_mult32x16in32(im, wim),
304 ixheaacd_mult32x16in32(re, wre));
305 *psubband_t2++ = ixheaacd_add32_sat(ixheaacd_mult32x16in32(re, wim),
306 ixheaacd_mult32x16in32(im, wre));
307
308 re = *psubband1--;
309 im = *psubband++;
310
311 wim = *p_sin_cos++;
312 wre = *p_sin_cos++;
313
314 *psubband1_t-- = ixheaacd_sub32_sat(ixheaacd_mult32x16in32(im, wre),
315 ixheaacd_mult32x16in32(re, wim));
316 *psubband1_t-- = ixheaacd_add32_sat(ixheaacd_mult32x16in32(re, wre),
317 ixheaacd_mult32x16in32(im, wim));
318
319 re = *psubband12--;
320 im = *psubband2++;
321
322 *psubband1_t2-- = ixheaacd_add32_sat(ixheaacd_mult32x16in32(re, wim),
323 ixheaacd_mult32x16in32(im, wre));
324 *psubband1_t2-- = ixheaacd_sub32_sat(ixheaacd_mult32x16in32(im, wim),
325 ixheaacd_mult32x16in32(re, wre));
326
327 re = *psubband++;
328 im = *psubband1--;
329
330 wim = *p_sin_cos++;
331 wre = *p_sin_cos++;
332
333 *psubband_t++ = ixheaacd_add32_sat(ixheaacd_mult32x16in32(re, wre),
334 ixheaacd_mult32x16in32(im, wim));
335 *psubband_t++ = ixheaacd_sub32_sat(ixheaacd_mult32x16in32(im, wre),
336 ixheaacd_mult32x16in32(re, wim));
337
338 re = *psubband2++;
339 im = *psubband12--;
340
341 *psubband_t2++ = ixheaacd_sub32_sat(ixheaacd_mult32x16in32(im, wim),
342 ixheaacd_mult32x16in32(re, wre));
343 *psubband_t2++ = ixheaacd_add32_sat(ixheaacd_mult32x16in32(re, wim),
344 ixheaacd_mult32x16in32(im, wre));
345
346 re = *psubband1--;
347 im = *psubband++;
348
349 wim = *p_sin_cos++;
350 wre = *p_sin_cos++;
351
352 *psubband1_t-- = ixheaacd_sub32_sat(ixheaacd_mult32x16in32(im, wre),
353 ixheaacd_mult32x16in32(re, wim));
354 *psubband1_t-- = ixheaacd_add32_sat(ixheaacd_mult32x16in32(re, wre),
355 ixheaacd_mult32x16in32(im, wim));
356
357 re = *psubband12--;
358 im = *psubband2++;
359 ;
360
361 *psubband1_t2-- = ixheaacd_add32_sat(ixheaacd_mult32x16in32(re, wim),
362 ixheaacd_mult32x16in32(im, wre));
363 *psubband1_t2-- = ixheaacd_sub32_sat(ixheaacd_mult32x16in32(im, wim),
364 ixheaacd_mult32x16in32(re, wre));
365 }
366
367 if (M == 32) {
368 ixheaacd_radix4bfly(p_twiddle, subband_tmp, 1, 8);
369 ixheaacd_radix4bfly(p_twiddle + 48, subband_tmp, 4, 2);
370 ixheaacd_postradixcompute2(subband, subband_tmp, p_dig_rev_tbl, 32);
371
372 ixheaacd_radix4bfly(p_twiddle, &subband_tmp[64], 1, 8);
373 ixheaacd_radix4bfly(p_twiddle + 48, &subband_tmp[64], 4, 2);
374 ixheaacd_postradixcompute2(&subband[64], &subband_tmp[64], p_dig_rev_tbl,
375 32);
376
377 } else {
378 ixheaacd_radix4bfly(p_twiddle, subband_tmp, 1, 4);
379 ixheaacd_postradixcompute4(subband, subband_tmp, p_dig_rev_tbl, 16);
380
381 ixheaacd_radix4bfly(p_twiddle, &subband_tmp[64], 1, 4);
382 ixheaacd_postradixcompute4(&subband[64], &subband_tmp[64], p_dig_rev_tbl,
383 16);
384 }
385
386 psubband = &subband[0];
387 psubband1 = &subband[2 * M - 1];
388
389 re = *psubband1;
390
391 *psubband = *psubband >> 1;
392 psubband++;
393 *psubband1 = ixheaacd_negate32(*psubband >> 1);
394 psubband1--;
395
396 p_sin = &qmf_bank->alt_sin_twiddle[0];
397 wim = *p_sin++;
398 wre = *p_sin++;
399
400 im = *psubband1;
401
402 *psubband1-- = ixheaacd_add32_sat(ixheaacd_mult32x16in32(re, wre),
403 ixheaacd_mult32x16in32(im, wim));
404 *psubband++ = ixheaacd_sub32_sat(ixheaacd_mult32x16in32(im, wre),
405 ixheaacd_mult32x16in32(re, wim));
406
407 psubband2 = &subband[64];
408 psubband12 = &subband[2 * M - 1 + 64];
409
410 re = *psubband12;
411
412 *psubband12-- = ixheaacd_negate32_sat(*psubband2 >> 1);
413
414 *psubband2 = psubband2[1] >> 1;
415
416 psubband2++;
417
418 im = *psubband12;
419
420 *psubband2++ = ixheaacd_negate32_sat(ixheaacd_add32_sat(
421 ixheaacd_mult32x16in32(re, wre), ixheaacd_mult32x16in32(im, wim)));
422 *psubband12-- = ixheaacd_sub32_sat(ixheaacd_mult32x16in32(re, wim),
423 ixheaacd_mult32x16in32(im, wre));
424
425 for (i = (M_2 - 2); i >= 0; i--) {
426 im = psubband[0];
427
428 re = psubband[1];
429
430 re2 = *psubband1;
431
432 *psubband++ = ixheaacd_add32_sat(ixheaacd_mult32x16in32(re, wim),
433 ixheaacd_mult32x16in32(im, wre));
434 *psubband1-- = ixheaacd_sub32_sat(ixheaacd_mult32x16in32(im, wim),
435 ixheaacd_mult32x16in32(re, wre));
436
437 im = psubband2[0];
438
439 re = psubband2[1];
440
441 re3 = *psubband12;
442
443 *psubband12-- = ixheaacd_negate32_sat(ixheaacd_add32_sat(
444 ixheaacd_mult32x16in32(re, wim), ixheaacd_mult32x16in32(im, wre)));
445 *psubband2++ = ixheaacd_sub32_sat(ixheaacd_mult32x16in32(re, wre),
446 ixheaacd_mult32x16in32(im, wim));
447
448 wim = *p_sin++;
449 wre = *p_sin++;
450 im = psubband1[0];
451
452 *psubband1-- = ixheaacd_add32_sat(ixheaacd_mult32x16in32(re2, wre),
453 ixheaacd_mult32x16in32(im, wim));
454 *psubband++ = ixheaacd_sub32_sat(ixheaacd_mult32x16in32(im, wre),
455 ixheaacd_mult32x16in32(re2, wim));
456
457 im = psubband12[0];
458
459 *psubband2++ = ixheaacd_negate32_sat(ixheaacd_add32_sat(
460 ixheaacd_mult32x16in32(re3, wre), ixheaacd_mult32x16in32(im, wim)));
461 *psubband12-- = ixheaacd_sub32_sat(ixheaacd_mult32x16in32(re3, wim),
462 ixheaacd_mult32x16in32(im, wre));
463 }
464 }
465
ixheaacd_fwd_modulation(const WORD32 * p_time_in1,WORD32 * real_subband,WORD32 * imag_subband,ia_sbr_qmf_filter_bank_struct * qmf_bank,ia_qmf_dec_tables_struct * qmf_dec_tables_ptr)466 VOID ixheaacd_fwd_modulation(const WORD32 *p_time_in1, WORD32 *real_subband,
467 WORD32 *imag_subband,
468 ia_sbr_qmf_filter_bank_struct *qmf_bank,
469 ia_qmf_dec_tables_struct *qmf_dec_tables_ptr) {
470 WORD32 i;
471 const WORD32 *p_time_in2 = &p_time_in1[2 * NO_ANALYSIS_CHANNELS - 1];
472 WORD32 temp1, temp2;
473 WORD32 *t_real_subband = real_subband;
474 WORD32 *t_imag_subband = imag_subband;
475 const WORD16 *tcos;
476
477 for (i = NO_ANALYSIS_CHANNELS - 1; i >= 0; i--) {
478 temp1 = ixheaacd_shr32(*p_time_in1++, HQ_SHIFT_VAL);
479 temp2 = ixheaacd_shr32(*p_time_in2--, HQ_SHIFT_VAL);
480
481 *t_real_subband++ = ixheaacd_sub32_sat(temp1, temp2);
482
483 *t_imag_subband++ = ixheaacd_add32_sat(temp1, temp2);
484 }
485
486 ixheaacd_cos_sin_mod(real_subband, qmf_bank, qmf_dec_tables_ptr->w_16,
487 qmf_dec_tables_ptr->dig_rev_table4_16);
488
489 tcos = qmf_bank->t_cos;
490
491 for (i = (qmf_bank->usb - qmf_bank->lsb - 1); i >= 0; i--) {
492 WORD16 cosh, sinh;
493 WORD32 re, im;
494
495 re = *real_subband;
496 im = *imag_subband;
497 cosh = *tcos++;
498 sinh = *tcos++;
499 *real_subband++ = ixheaacd_add32_sat(ixheaacd_mult32x16in32_shl(re, cosh),
500 ixheaacd_mult32x16in32_shl(im, sinh));
501 *imag_subband++ = ixheaacd_sub32_sat(ixheaacd_mult32x16in32_shl(im, cosh),
502 ixheaacd_mult32x16in32_shl(re, sinh));
503 }
504 }
505
ixheaacd_sbr_qmfanal32_winadd(WORD16 * inp1,WORD16 * inp2,WORD16 * p_qmf1,WORD16 * p_qmf2,WORD32 * p_out)506 VOID ixheaacd_sbr_qmfanal32_winadd(WORD16 *inp1, WORD16 *inp2, WORD16 *p_qmf1,
507 WORD16 *p_qmf2, WORD32 *p_out) {
508 WORD32 n;
509
510 for (n = 0; n < 32; n += 2) {
511 WORD32 accu;
512
513 accu = ixheaacd_mult16x16in32(inp1[n + 0], p_qmf1[2 * (n + 0)]);
514 accu = ixheaacd_add32_sat(
515 accu, ixheaacd_mult16x16in32(inp1[n + 64], p_qmf1[2 * (n + 64)]));
516 accu = ixheaacd_add32_sat(
517 accu, ixheaacd_mult16x16in32(inp1[n + 128], p_qmf1[2 * (n + 128)]));
518 accu = ixheaacd_add32_sat(
519 accu, ixheaacd_mult16x16in32(inp1[n + 192], p_qmf1[2 * (n + 192)]));
520 accu = ixheaacd_add32_sat(
521 accu, ixheaacd_mult16x16in32(inp1[n + 256], p_qmf1[2 * (n + 256)]));
522 p_out[n] = accu;
523
524 accu = ixheaacd_mult16x16in32(inp1[n + 1 + 0], p_qmf1[2 * (n + 1 + 0)]);
525 accu = ixheaacd_add32_sat(
526 accu,
527 ixheaacd_mult16x16in32(inp1[n + 1 + 64], p_qmf1[2 * (n + 1 + 64)]));
528 accu = ixheaacd_add32_sat(
529 accu,
530 ixheaacd_mult16x16in32(inp1[n + 1 + 128], p_qmf1[2 * (n + 1 + 128)]));
531 accu = ixheaacd_add32_sat(
532 accu,
533 ixheaacd_mult16x16in32(inp1[n + 1 + 192], p_qmf1[2 * (n + 1 + 192)]));
534 accu = ixheaacd_add32_sat(
535 accu,
536 ixheaacd_mult16x16in32(inp1[n + 1 + 256], p_qmf1[2 * (n + 1 + 256)]));
537 p_out[n + 1] = accu;
538
539 accu = ixheaacd_mult16x16in32(inp2[n + 0], p_qmf2[2 * (n + 0)]);
540 accu = ixheaacd_add32_sat(
541 accu, ixheaacd_mult16x16in32(inp2[n + 64], p_qmf2[2 * (n + 64)]));
542 accu = ixheaacd_add32_sat(
543 accu, ixheaacd_mult16x16in32(inp2[n + 128], p_qmf2[2 * (n + 128)]));
544 accu = ixheaacd_add32_sat(
545 accu, ixheaacd_mult16x16in32(inp2[n + 192], p_qmf2[2 * (n + 192)]));
546 accu = ixheaacd_add32_sat(
547 accu, ixheaacd_mult16x16in32(inp2[n + 256], p_qmf2[2 * (n + 256)]));
548 p_out[n + 32] = accu;
549
550 accu = ixheaacd_mult16x16in32(inp2[n + 1 + 0], p_qmf2[2 * (n + 1 + 0)]);
551 accu = ixheaacd_add32_sat(
552 accu,
553 ixheaacd_mult16x16in32(inp2[n + 1 + 64], p_qmf2[2 * (n + 1 + 64)]));
554 accu = ixheaacd_add32_sat(
555 accu,
556 ixheaacd_mult16x16in32(inp2[n + 1 + 128], p_qmf2[2 * (n + 1 + 128)]));
557 accu = ixheaacd_add32_sat(
558 accu,
559 ixheaacd_mult16x16in32(inp2[n + 1 + 192], p_qmf2[2 * (n + 1 + 192)]));
560 accu = ixheaacd_add32_sat(
561 accu,
562 ixheaacd_mult16x16in32(inp2[n + 1 + 256], p_qmf2[2 * (n + 1 + 256)]));
563 p_out[n + 1 + 32] = accu;
564 }
565 }
566
ixheaacd_cplx_anal_qmffilt(const WORD16 * time_sample_buf,ia_sbr_scale_fact_struct * sbr_scale_factor,WORD32 ** qmf_real,WORD32 ** qmf_imag,ia_sbr_qmf_filter_bank_struct * qmf_bank,ia_qmf_dec_tables_struct * qmf_dec_tables_ptr,WORD32 ch_fac,WORD32 low_pow_flag,WORD audio_object_type)567 VOID ixheaacd_cplx_anal_qmffilt(const WORD16 *time_sample_buf,
568 ia_sbr_scale_fact_struct *sbr_scale_factor,
569 WORD32 **qmf_real, WORD32 **qmf_imag,
570 ia_sbr_qmf_filter_bank_struct *qmf_bank,
571 ia_qmf_dec_tables_struct *qmf_dec_tables_ptr,
572 WORD32 ch_fac, WORD32 low_pow_flag,
573 WORD audio_object_type) {
574 WORD32 i, k;
575 WORD32 num_time_slots = qmf_bank->num_time_slots;
576
577 WORD32 analysis_buffer[4 * NO_ANALYSIS_CHANNELS];
578 WORD16 *filter_states = qmf_bank->core_samples_buffer;
579
580 WORD16 *fp1, *fp2, *tmp;
581
582 WORD16 *filter_1;
583 WORD16 *filter_2;
584 WORD16 *filt_ptr;
585 if (audio_object_type != AOT_ER_AAC_ELD &&
586 audio_object_type != AOT_ER_AAC_LD) {
587 qmf_bank->filter_pos +=
588 (qmf_dec_tables_ptr->qmf_c - qmf_bank->analy_win_coeff);
589 qmf_bank->analy_win_coeff = qmf_dec_tables_ptr->qmf_c;
590 } else {
591 qmf_bank->filter_pos +=
592 (qmf_dec_tables_ptr->qmf_c_eld3 - qmf_bank->analy_win_coeff);
593 qmf_bank->analy_win_coeff = qmf_dec_tables_ptr->qmf_c_eld3;
594 }
595
596 filter_1 = qmf_bank->filter_pos;
597
598 if (audio_object_type != AOT_ER_AAC_ELD &&
599 audio_object_type != AOT_ER_AAC_LD) {
600 filter_2 = filter_1 + 64;
601 } else {
602 filter_2 = filter_1 + 32;
603 }
604
605 sbr_scale_factor->st_lb_scale = 0;
606 sbr_scale_factor->lb_scale = -10;
607 if (!low_pow_flag) {
608 if (audio_object_type != AOT_ER_AAC_ELD &&
609 audio_object_type != AOT_ER_AAC_LD) {
610 sbr_scale_factor->lb_scale = -8;
611 } else {
612 sbr_scale_factor->lb_scale = -9;
613 }
614 qmf_bank->cos_twiddle =
615 (WORD16 *)qmf_dec_tables_ptr->sbr_sin_cos_twiddle_l32;
616 qmf_bank->alt_sin_twiddle =
617 (WORD16 *)qmf_dec_tables_ptr->sbr_alt_sin_twiddle_l32;
618 if (audio_object_type != AOT_ER_AAC_ELD &&
619 audio_object_type != AOT_ER_AAC_LD) {
620 qmf_bank->t_cos = (WORD16 *)qmf_dec_tables_ptr->sbr_t_cos_sin_l32;
621 } else {
622 qmf_bank->t_cos =
623 (WORD16 *)qmf_dec_tables_ptr->ixheaacd_sbr_t_cos_sin_l32_eld;
624 }
625 }
626
627 fp1 = qmf_bank->anal_filter_states;
628 fp2 = qmf_bank->anal_filter_states + NO_ANALYSIS_CHANNELS;
629
630 if (audio_object_type == AOT_ER_AAC_ELD ||
631 audio_object_type == AOT_ER_AAC_LD) {
632 filter_2 = qmf_bank->filter_2;
633 fp1 = qmf_bank->fp1_anal;
634 fp2 = qmf_bank->fp2_anal;
635 }
636
637 for (i = 0; i < num_time_slots; i++) {
638 for (k = 0; k < NO_ANALYSIS_CHANNELS; k++)
639 filter_states[NO_ANALYSIS_CHANNELS - 1 - k] = time_sample_buf[ch_fac * k];
640
641 if (audio_object_type != AOT_ER_AAC_ELD &&
642 audio_object_type != AOT_ER_AAC_LD) {
643 ixheaacd_sbr_qmfanal32_winadd(fp1, fp2, filter_1, filter_2,
644 analysis_buffer);
645 } else {
646 ixheaacd_sbr_qmfanal32_winadd_eld(fp1, fp2, filter_1, filter_2,
647 analysis_buffer);
648 }
649
650 time_sample_buf += NO_ANALYSIS_CHANNELS * ch_fac;
651
652 filter_states -= NO_ANALYSIS_CHANNELS;
653 if (filter_states < qmf_bank->anal_filter_states) {
654 filter_states = qmf_bank->anal_filter_states + 288;
655 }
656
657 tmp = fp1;
658 fp1 = fp2;
659 fp2 = tmp;
660 if (audio_object_type != AOT_ER_AAC_ELD &&
661 audio_object_type != AOT_ER_AAC_LD) {
662 filter_1 += 64;
663 filter_2 += 64;
664 } else {
665 filter_1 += 32;
666 filter_2 += 32;
667 }
668
669 filt_ptr = filter_1;
670 filter_1 = filter_2;
671 filter_2 = filt_ptr;
672 if (audio_object_type != AOT_ER_AAC_ELD &&
673 audio_object_type != AOT_ER_AAC_LD) {
674 if (filter_2 > (qmf_bank->analy_win_coeff + 640)) {
675 filter_1 = (WORD16 *)qmf_bank->analy_win_coeff;
676 filter_2 = (WORD16 *)qmf_bank->analy_win_coeff + 64;
677 }
678 } else {
679 if (filter_2 > (qmf_bank->analy_win_coeff + 320)) {
680 filter_1 = (WORD16 *)qmf_bank->analy_win_coeff;
681 filter_2 = (WORD16 *)qmf_bank->analy_win_coeff + 32;
682 }
683 }
684
685 if (!low_pow_flag) {
686 ixheaacd_fwd_modulation(analysis_buffer, qmf_real[i], qmf_imag[i],
687 qmf_bank, qmf_dec_tables_ptr);
688 } else {
689 ixheaacd_dct3_32(
690 (WORD32 *)analysis_buffer, qmf_real[i], qmf_dec_tables_ptr->dct23_tw,
691 qmf_dec_tables_ptr->post_fft_tbl, qmf_dec_tables_ptr->w_16,
692 qmf_dec_tables_ptr->dig_rev_table4_16);
693 }
694 }
695
696 qmf_bank->filter_pos = filter_1;
697 qmf_bank->core_samples_buffer = filter_states;
698
699 if (audio_object_type == AOT_ER_AAC_ELD || audio_object_type == AOT_ER_AAC_LD)
700
701 {
702 qmf_bank->fp1_anal = fp1;
703 qmf_bank->fp2_anal = fp2;
704 qmf_bank->filter_2 = filter_2;
705 }
706 }
707
ixheaacd_inv_modulation_lp(WORD32 * qmf_real,WORD16 * filter_states,ia_sbr_qmf_filter_bank_struct * syn_qmf,ia_qmf_dec_tables_struct * qmf_dec_tables_ptr)708 VOID ixheaacd_inv_modulation_lp(WORD32 *qmf_real, WORD16 *filter_states,
709 ia_sbr_qmf_filter_bank_struct *syn_qmf,
710 ia_qmf_dec_tables_struct *qmf_dec_tables_ptr) {
711 WORD32 L = syn_qmf->no_channels;
712 const WORD32 M = (L >> 1);
713 WORD32 *dct_in = qmf_real;
714 WORD32 time_out[2 * NO_SYNTHESIS_CHANNELS];
715
716 WORD32 *ptime_out = &time_out[0];
717
718 if (L == 64)
719 ixheaacd_dct2_64(dct_in, ptime_out, qmf_dec_tables_ptr, filter_states + M);
720 else
721 ixheaacd_dct2_32(dct_in, time_out, qmf_dec_tables_ptr, filter_states);
722
723 filter_states[3 * M] = 0;
724 }
725
ixheaacd_inv_emodulation(WORD32 * qmf_real,ia_sbr_qmf_filter_bank_struct * syn_qmf,ia_qmf_dec_tables_struct * qmf_dec_tables_ptr)726 VOID ixheaacd_inv_emodulation(WORD32 *qmf_real,
727 ia_sbr_qmf_filter_bank_struct *syn_qmf,
728 ia_qmf_dec_tables_struct *qmf_dec_tables_ptr) {
729 if (syn_qmf->no_channels == 64)
730 ixheaacd_cos_sin_mod(qmf_real, syn_qmf, qmf_dec_tables_ptr->w_32,
731 qmf_dec_tables_ptr->dig_rev_table2_32);
732 else
733 ixheaacd_cos_sin_mod(qmf_real, syn_qmf, qmf_dec_tables_ptr->w_16,
734 qmf_dec_tables_ptr->dig_rev_table4_16);
735 }
736
ixheaacd_esbr_radix4bfly(const WORD32 * w,WORD32 * x,WORD32 index1,WORD32 index)737 VOID ixheaacd_esbr_radix4bfly(const WORD32 *w, WORD32 *x, WORD32 index1,
738 WORD32 index) {
739 int i;
740 WORD32 l1, l2, h2, fft_jmp;
741 WORD32 xt0_0, yt0_0, xt1_0, yt1_0, xt2_0, yt2_0;
742 WORD32 xh0_0, xh1_0, xh20_0, xh21_0, xl0_0, xl1_0, xl20_0, xl21_0;
743 WORD32 x_0, x_1, x_l1_0, x_l1_1, x_l2_0, x_l2_1;
744 WORD32 x_h2_0, x_h2_1;
745 WORD32 si10, si20, si30, co10, co20, co30;
746
747 WORD64 mul_1, mul_2, mul_3, mul_4, mul_5, mul_6;
748 WORD64 mul_7, mul_8, mul_9, mul_10, mul_11, mul_12;
749 WORD32 *x_l1;
750 WORD32 *x_l2;
751 WORD32 *x_h2;
752 const WORD32 *w_ptr = w;
753 WORD32 i1;
754
755 h2 = index << 1;
756 l1 = index << 2;
757 l2 = (index << 2) + (index << 1);
758
759 x_l1 = &(x[l1]);
760 x_l2 = &(x[l2]);
761 x_h2 = &(x[h2]);
762
763 fft_jmp = 6 * (index);
764
765 for (i1 = 0; i1 < index1; i1++) {
766 for (i = 0; i < index; i++) {
767 si10 = (*w_ptr++);
768 co10 = (*w_ptr++);
769 si20 = (*w_ptr++);
770 co20 = (*w_ptr++);
771 si30 = (*w_ptr++);
772 co30 = (*w_ptr++);
773
774 x_0 = x[0];
775 x_h2_0 = x[h2];
776 x_l1_0 = x[l1];
777 x_l2_0 = x[l2];
778
779 xh0_0 = ixheaacd_add32_sat(x_0, x_l1_0);
780 xl0_0 = ixheaacd_sub32_sat(x_0, x_l1_0);
781
782 xh20_0 = ixheaacd_add32_sat(x_h2_0, x_l2_0);
783 xl20_0 = ixheaacd_sub32_sat(x_h2_0, x_l2_0);
784
785 x[0] = ixheaacd_add32_sat(xh0_0, xh20_0);
786 xt0_0 = ixheaacd_sub32_sat(xh0_0, xh20_0);
787
788 x_1 = x[1];
789 x_h2_1 = x[h2 + 1];
790 x_l1_1 = x[l1 + 1];
791 x_l2_1 = x[l2 + 1];
792
793 xh1_0 = ixheaacd_add32_sat(x_1, x_l1_1);
794 xl1_0 = ixheaacd_sub32_sat(x_1, x_l1_1);
795
796 xh21_0 = ixheaacd_add32_sat(x_h2_1, x_l2_1);
797 xl21_0 = ixheaacd_sub32_sat(x_h2_1, x_l2_1);
798
799 x[1] = ixheaacd_add32_sat(xh1_0, xh21_0);
800 yt0_0 = ixheaacd_sub32_sat(xh1_0, xh21_0);
801
802 xt1_0 = ixheaacd_add32_sat(xl0_0, xl21_0);
803 xt2_0 = ixheaacd_sub32_sat(xl0_0, xl21_0);
804
805 yt2_0 = ixheaacd_add32_sat(xl1_0, xl20_0);
806 yt1_0 = ixheaacd_sub32_sat(xl1_0, xl20_0);
807
808 mul_11 = ixheaacd_mult64(xt2_0, co30);
809 mul_3 = ixheaacd_mult64(yt2_0, si30);
810 x[l2] = (WORD32)((mul_3 + mul_11) >> 32) << RADIXSHIFT;
811
812 mul_5 = ixheaacd_mult64(xt2_0, si30);
813 mul_9 = ixheaacd_mult64(yt2_0, co30);
814 x[l2 + 1] = (WORD32)((mul_9 - mul_5) >> 32) << RADIXSHIFT;
815
816 mul_12 = ixheaacd_mult64(xt0_0, co20);
817 mul_2 = ixheaacd_mult64(yt0_0, si20);
818 x[l1] = (WORD32)((mul_2 + mul_12) >> 32) << RADIXSHIFT;
819
820 mul_6 = ixheaacd_mult64(xt0_0, si20);
821 mul_8 = ixheaacd_mult64(yt0_0, co20);
822 x[l1 + 1] = (WORD32)((mul_8 - mul_6) >> 32) << RADIXSHIFT;
823
824 mul_4 = ixheaacd_mult64(xt1_0, co10);
825 mul_1 = ixheaacd_mult64(yt1_0, si10);
826 x[h2] = (WORD32)((mul_1 + mul_4) >> 32) << RADIXSHIFT;
827
828 mul_10 = ixheaacd_mult64(xt1_0, si10);
829 mul_7 = ixheaacd_mult64(yt1_0, co10);
830 x[h2 + 1] = (WORD32)((mul_7 - mul_10) >> 32) << RADIXSHIFT;
831
832 x += 2;
833 }
834 x += fft_jmp;
835 w_ptr = w_ptr - fft_jmp;
836 }
837 }
838
ixheaacd_esbr_postradixcompute2(WORD32 * ptr_y,WORD32 * ptr_x,const WORD32 * pdig_rev_tbl,WORD32 npoints)839 VOID ixheaacd_esbr_postradixcompute2(WORD32 *ptr_y, WORD32 *ptr_x,
840 const WORD32 *pdig_rev_tbl,
841 WORD32 npoints) {
842 WORD32 i, k;
843 WORD32 h2;
844 WORD32 x_0, x_1, x_2, x_3;
845 WORD32 x_4, x_5, x_6, x_7;
846 WORD32 x_8, x_9, x_a, x_b, x_c, x_d, x_e, x_f;
847 WORD32 n00, n10, n20, n30, n01, n11, n21, n31;
848 WORD32 n02, n12, n22, n32, n03, n13, n23, n33;
849 WORD32 n0, j0;
850 WORD32 *x2, *x0;
851 WORD32 *y0, *y1, *y2, *y3;
852
853 y0 = ptr_y;
854 y2 = ptr_y + (WORD32)npoints;
855 x0 = ptr_x;
856 x2 = ptr_x + (WORD32)(npoints >> 1);
857
858 y1 = y0 + (WORD32)(npoints >> 2);
859 y3 = y2 + (WORD32)(npoints >> 2);
860 j0 = 8;
861 n0 = npoints >> 1;
862
863 for (k = 0; k < 2; k++) {
864 for (i = 0; i<npoints>> 1; i += 8) {
865 h2 = *pdig_rev_tbl++ >> 2;
866
867 x_0 = *x0++;
868 x_1 = *x0++;
869 x_2 = *x0++;
870 x_3 = *x0++;
871 x_4 = *x0++;
872 x_5 = *x0++;
873 x_6 = *x0++;
874 x_7 = *x0++;
875
876 n00 = ixheaacd_add32_sat(x_0, x_2);
877 n01 = ixheaacd_add32_sat(x_1, x_3);
878 n20 = ixheaacd_sub32_sat(x_0, x_2);
879 n21 = ixheaacd_sub32_sat(x_1, x_3);
880 n10 = ixheaacd_add32_sat(x_4, x_6);
881 n11 = ixheaacd_add32_sat(x_5, x_7);
882 n30 = ixheaacd_sub32_sat(x_4, x_6);
883 n31 = ixheaacd_sub32_sat(x_5, x_7);
884
885 y0[h2] = n00;
886 y0[h2 + 1] = n01;
887 y1[h2] = n10;
888 y1[h2 + 1] = n11;
889 y2[h2] = n20;
890 y2[h2 + 1] = n21;
891 y3[h2] = n30;
892 y3[h2 + 1] = n31;
893
894 x_8 = *x2++;
895 x_9 = *x2++;
896 x_a = *x2++;
897 x_b = *x2++;
898 x_c = *x2++;
899 x_d = *x2++;
900 x_e = *x2++;
901 x_f = *x2++;
902
903 n02 = ixheaacd_add32_sat(x_8, x_a);
904 n03 = ixheaacd_add32_sat(x_9, x_b);
905 n22 = ixheaacd_sub32_sat(x_8, x_a);
906 n23 = ixheaacd_sub32_sat(x_9, x_b);
907 n12 = ixheaacd_add32_sat(x_c, x_e);
908 n13 = ixheaacd_add32_sat(x_d, x_f);
909 n32 = ixheaacd_sub32_sat(x_c, x_e);
910 n33 = ixheaacd_sub32_sat(x_d, x_f);
911
912 y0[h2 + 2] = n02;
913 y0[h2 + 3] = n03;
914 y1[h2 + 2] = n12;
915 y1[h2 + 3] = n13;
916 y2[h2 + 2] = n22;
917 y2[h2 + 3] = n23;
918 y3[h2 + 2] = n32;
919 y3[h2 + 3] = n33;
920 }
921 x0 += (WORD32)npoints >> 1;
922 x2 += (WORD32)npoints >> 1;
923 }
924 }
925
ixheaacd_esbr_postradixcompute4(WORD32 * ptr_y,WORD32 * ptr_x,const WORD32 * p_dig_rev_tbl,WORD32 npoints)926 VOID ixheaacd_esbr_postradixcompute4(WORD32 *ptr_y, WORD32 *ptr_x,
927 const WORD32 *p_dig_rev_tbl,
928 WORD32 npoints) {
929 WORD32 i, k;
930 WORD32 h2;
931 WORD32 xh0_0, xh1_0, xl0_0, xl1_0;
932 WORD32 xh0_1, xh1_1, xl0_1, xl1_1;
933 WORD32 x_0, x_1, x_2, x_3;
934 WORD32 xh0_2, xh1_2, xl0_2, xl1_2, xh0_3, xh1_3, xl0_3, xl1_3;
935 WORD32 x_4, x_5, x_6, x_7;
936 WORD32 x_8, x_9, x_a, x_b, x_c, x_d, x_e, x_f;
937 WORD32 n00, n10, n20, n30, n01, n11, n21, n31;
938 WORD32 n02, n12, n22, n32, n03, n13, n23, n33;
939 WORD32 n0, j0;
940 WORD32 *x2, *x0;
941 WORD32 *y0, *y1, *y2, *y3;
942
943 y0 = ptr_y;
944 y2 = ptr_y + (WORD32)npoints;
945 x0 = ptr_x;
946 x2 = ptr_x + (WORD32)(npoints >> 1);
947
948 y1 = y0 + (WORD32)(npoints >> 1);
949 y3 = y2 + (WORD32)(npoints >> 1);
950
951 j0 = 4;
952 n0 = npoints >> 2;
953
954 for (k = 0; k < 2; k++) {
955 for (i = 0; i<npoints>> 1; i += 8) {
956 h2 = *p_dig_rev_tbl++ >> 2;
957 x_0 = *x0++;
958 x_1 = *x0++;
959 x_2 = *x0++;
960 x_3 = *x0++;
961 x_4 = *x0++;
962 x_5 = *x0++;
963 x_6 = *x0++;
964 x_7 = *x0++;
965
966 xh0_0 = ixheaacd_add32_sat(x_0, x_4);
967 xh1_0 = ixheaacd_add32_sat(x_1, x_5);
968 xl0_0 = ixheaacd_sub32_sat(x_0, x_4);
969 xl1_0 = ixheaacd_sub32_sat(x_1, x_5);
970 xh0_1 = ixheaacd_add32_sat(x_2, x_6);
971 xh1_1 = ixheaacd_add32_sat(x_3, x_7);
972 xl0_1 = ixheaacd_sub32_sat(x_2, x_6);
973 xl1_1 = ixheaacd_sub32_sat(x_3, x_7);
974
975 n00 = ixheaacd_add32_sat(xh0_0, xh0_1);
976 n01 = ixheaacd_add32_sat(xh1_0, xh1_1);
977 n10 = ixheaacd_add32_sat(xl0_0, xl1_1);
978 n11 = ixheaacd_sub32_sat(xl1_0, xl0_1);
979 n20 = ixheaacd_sub32_sat(xh0_0, xh0_1);
980 n21 = ixheaacd_sub32_sat(xh1_0, xh1_1);
981 n30 = ixheaacd_sub32_sat(xl0_0, xl1_1);
982 n31 = ixheaacd_add32_sat(xl1_0, xl0_1);
983
984 y0[h2] = n00;
985 y0[h2 + 1] = n01;
986 y1[h2] = n10;
987 y1[h2 + 1] = n11;
988 y2[h2] = n20;
989 y2[h2 + 1] = n21;
990 y3[h2] = n30;
991 y3[h2 + 1] = n31;
992
993 x_8 = *x2++;
994 x_9 = *x2++;
995 x_a = *x2++;
996 x_b = *x2++;
997 x_c = *x2++;
998 x_d = *x2++;
999 x_e = *x2++;
1000 x_f = *x2++;
1001
1002 xh0_2 = ixheaacd_add32_sat(x_8, x_c);
1003 xh1_2 = ixheaacd_add32_sat(x_9, x_d);
1004 xl0_2 = ixheaacd_sub32_sat(x_8, x_c);
1005 xl1_2 = ixheaacd_sub32_sat(x_9, x_d);
1006 xh0_3 = ixheaacd_add32_sat(x_a, x_e);
1007 xh1_3 = ixheaacd_add32_sat(x_b, x_f);
1008 xl0_3 = ixheaacd_sub32_sat(x_a, x_e);
1009 xl1_3 = ixheaacd_sub32_sat(x_b, x_f);
1010
1011 n02 = ixheaacd_add32_sat(xh0_2, xh0_3);
1012 n03 = ixheaacd_add32_sat(xh1_2, xh1_3);
1013 n12 = ixheaacd_add32_sat(xl0_2, xl1_3);
1014 n13 = ixheaacd_sub32_sat(xl1_2, xl0_3);
1015 n22 = ixheaacd_sub32_sat(xh0_2, xh0_3);
1016 n23 = ixheaacd_sub32_sat(xh1_2, xh1_3);
1017 n32 = ixheaacd_sub32_sat(xl0_2, xl1_3);
1018 n33 = ixheaacd_add32_sat(xl1_2, xl0_3);
1019
1020 y0[h2 + 2] = n02;
1021 y0[h2 + 3] = n03;
1022 y1[h2 + 2] = n12;
1023 y1[h2 + 3] = n13;
1024 y2[h2 + 2] = n22;
1025 y2[h2 + 3] = n23;
1026 y3[h2 + 2] = n32;
1027 y3[h2 + 3] = n33;
1028 }
1029 x0 += (WORD32)npoints >> 1;
1030 x2 += (WORD32)npoints >> 1;
1031 }
1032 }
1033
ixheaacd_esbr_cos_sin_mod(WORD32 * subband,ia_sbr_qmf_filter_bank_struct * qmf_bank,WORD32 * p_twiddle,WORD32 * p_dig_rev_tbl)1034 VOID ixheaacd_esbr_cos_sin_mod(WORD32 *subband,
1035 ia_sbr_qmf_filter_bank_struct *qmf_bank,
1036 WORD32 *p_twiddle, WORD32 *p_dig_rev_tbl) {
1037 WORD32 z;
1038 WORD32 temp[128];
1039 WORD32 scaleshift = 0;
1040
1041 WORD32 re2, re3;
1042 WORD32 wim, wre;
1043
1044 WORD32 i, M_2;
1045 WORD32 M = ixheaacd_shr32(qmf_bank->no_channels, 1);
1046
1047 const WORD32 *p_sin;
1048 const WORD32 *p_sin_cos;
1049
1050 WORD32 subband_tmp[128];
1051 WORD32 re;
1052 WORD32 im;
1053 WORD32 *psubband, *psubband1;
1054 WORD32 *psubband_t, *psubband1_t;
1055 WORD32 *psubband2, *psubband12;
1056 WORD32 *psubband_t2, *psubband1_t2;
1057
1058 M_2 = ixheaacd_shr32(M, 1);
1059
1060 p_sin_cos = qmf_bank->esbr_cos_twiddle;
1061
1062 psubband = &subband[0];
1063 psubband1 = &subband[2 * M - 1];
1064 psubband_t = subband_tmp;
1065 psubband1_t = &subband_tmp[2 * M - 1];
1066
1067 psubband2 = &subband[64];
1068 psubband12 = &subband[2 * M - 1 + 64];
1069 psubband_t2 = &subband_tmp[64];
1070 psubband1_t2 = &subband_tmp[2 * M - 1 + 64];
1071
1072 for (i = (M_2 >> 1) - 1; i >= 0; i--) {
1073 re = *psubband++;
1074 im = *psubband1--;
1075
1076 wim = *p_sin_cos++;
1077 wre = *p_sin_cos++;
1078
1079 *psubband_t++ = (WORD32)(
1080 (ixheaacd_add64(ixheaacd_mult64(re, wre), ixheaacd_mult64(im, wim))) >>
1081 32);
1082 *psubband_t++ = (WORD32)((ixheaacd_sub64_sat(ixheaacd_mult64(im, wre),
1083 ixheaacd_mult64(re, wim))) >>
1084 32);
1085
1086 re = *psubband2++;
1087 im = *psubband12--;
1088
1089 *psubband_t2++ = (WORD32)((ixheaacd_sub64_sat(ixheaacd_mult64(im, wim),
1090 ixheaacd_mult64(re, wre))) >>
1091 32);
1092 *psubband_t2++ = (WORD32)(
1093 (ixheaacd_add64(ixheaacd_mult64(re, wim), ixheaacd_mult64(im, wre))) >>
1094 32);
1095
1096 re = *psubband1--;
1097 im = *psubband++;
1098
1099 wim = *p_sin_cos++;
1100 wre = *p_sin_cos++;
1101
1102 *psubband1_t-- = (WORD32)((ixheaacd_sub64_sat(ixheaacd_mult64(im, wre),
1103 ixheaacd_mult64(re, wim))) >>
1104 32);
1105 *psubband1_t-- = (WORD32)(
1106 (ixheaacd_add64(ixheaacd_mult64(re, wre), ixheaacd_mult64(im, wim))) >>
1107 32);
1108
1109 re = *psubband12--;
1110 im = *psubband2++;
1111
1112 *psubband1_t2-- = (WORD32)(
1113 (ixheaacd_add64(ixheaacd_mult64(re, wim), ixheaacd_mult64(im, wre))) >>
1114 32);
1115 *psubband1_t2-- = (WORD32)((ixheaacd_sub64_sat(ixheaacd_mult64(im, wim),
1116 ixheaacd_mult64(re, wre))) >>
1117 32);
1118
1119 re = *psubband++;
1120 im = *psubband1--;
1121
1122 wim = *p_sin_cos++;
1123 wre = *p_sin_cos++;
1124
1125 *psubband_t++ = (WORD32)(
1126 (ixheaacd_add64(ixheaacd_mult64(re, wre), ixheaacd_mult64(im, wim))) >>
1127 32);
1128 *psubband_t++ = (WORD32)((ixheaacd_sub64_sat(ixheaacd_mult64(im, wre),
1129 ixheaacd_mult64(re, wim))) >>
1130 32);
1131
1132 re = *psubband2++;
1133 im = *psubband12--;
1134
1135 *psubband_t2++ = (WORD32)((ixheaacd_sub64_sat(ixheaacd_mult64(im, wim),
1136 ixheaacd_mult64(re, wre))) >>
1137 32);
1138 *psubband_t2++ = (WORD32)(
1139 (ixheaacd_add64(ixheaacd_mult64(re, wim), ixheaacd_mult64(im, wre))) >>
1140 32);
1141
1142 re = *psubband1--;
1143 im = *psubband++;
1144
1145 wim = *p_sin_cos++;
1146 wre = *p_sin_cos++;
1147
1148 *psubband1_t-- = (WORD32)((ixheaacd_sub64_sat(ixheaacd_mult64(im, wre),
1149 ixheaacd_mult64(re, wim))) >>
1150 32);
1151 *psubband1_t-- = (WORD32)(
1152 (ixheaacd_add64(ixheaacd_mult64(re, wre), ixheaacd_mult64(im, wim))) >>
1153 32);
1154
1155 re = *psubband12--;
1156 im = *psubband2++;
1157
1158 *psubband1_t2-- = (WORD32)(
1159 (ixheaacd_add64(ixheaacd_mult64(re, wim), ixheaacd_mult64(im, wre))) >>
1160 32);
1161 *psubband1_t2-- = (WORD32)((ixheaacd_sub64_sat(ixheaacd_mult64(im, wim),
1162 ixheaacd_mult64(re, wre))) >>
1163 32);
1164 }
1165
1166 if (M == 32) {
1167 ixheaacd_esbr_radix4bfly(p_twiddle, subband_tmp, 1, 8);
1168 ixheaacd_esbr_radix4bfly(p_twiddle + 48, subband_tmp, 4, 2);
1169 ixheaacd_esbr_postradixcompute2(subband, subband_tmp, p_dig_rev_tbl, 32);
1170
1171 ixheaacd_esbr_radix4bfly(p_twiddle, &subband_tmp[64], 1, 8);
1172 ixheaacd_esbr_radix4bfly(p_twiddle + 48, &subband_tmp[64], 4, 2);
1173 ixheaacd_esbr_postradixcompute2(&subband[64], &subband_tmp[64],
1174 p_dig_rev_tbl, 32);
1175
1176 }
1177
1178 else if (M == 16) {
1179 ixheaacd_esbr_radix4bfly(p_twiddle, subband_tmp, 1, 4);
1180 ixheaacd_esbr_postradixcompute4(subband, subband_tmp, p_dig_rev_tbl, 16);
1181
1182 ixheaacd_esbr_radix4bfly(p_twiddle, &subband_tmp[64], 1, 4);
1183 ixheaacd_esbr_postradixcompute4(&subband[64], &subband_tmp[64],
1184 p_dig_rev_tbl, 16);
1185
1186 }
1187
1188 else if (M == 12) {
1189 for (z = 0; z < (qmf_bank->no_channels >> 1); z++) {
1190 temp[z] = subband_tmp[2 * z];
1191 temp[12 + z] = subband_tmp[2 * z + 1];
1192 }
1193
1194 ixheaacd_complex_fft_p3(temp, &temp[12], 12, -1, &scaleshift);
1195
1196 for (z = 0; z < (qmf_bank->no_channels >> 1); z++) {
1197 subband[2 * z] = temp[z];
1198 subband[2 * z + 1] = temp[z + 12];
1199 }
1200 scaleshift = 0;
1201 for (z = 0; z < (qmf_bank->no_channels >> 1); z++) {
1202 temp[z] = subband_tmp[64 + 2 * z];
1203 temp[12 + z] = subband_tmp[64 + 2 * z + 1];
1204 }
1205
1206 ixheaacd_complex_fft_p3(temp, &temp[12], 12, -1, &scaleshift);
1207
1208 for (z = 0; z < (qmf_bank->no_channels >> 1); z++) {
1209 subband[64 + 2 * z] = temp[z];
1210 subband[64 + 2 * z + 1] = temp[z + 12];
1211 }
1212
1213 }
1214
1215 else {
1216 for (z = 0; z < (qmf_bank->no_channels >> 1); z++) {
1217 temp[z] = subband_tmp[2 * z];
1218 temp[8 + z] = subband_tmp[2 * z + 1];
1219 }
1220
1221 (*ixheaacd_complex_fft_p2)(temp, &temp[8], 8, -1, &scaleshift);
1222
1223 for (z = 0; z < (qmf_bank->no_channels >> 1); z++) {
1224 subband[2 * z] = temp[z] << scaleshift;
1225 subband[2 * z + 1] = temp[z + 8] << scaleshift;
1226 }
1227 scaleshift = 0;
1228 for (z = 0; z < (qmf_bank->no_channels >> 1); z++) {
1229 temp[z] = subband_tmp[64 + 2 * z];
1230 temp[8 + z] = subband_tmp[64 + 2 * z + 1];
1231 }
1232
1233 (*ixheaacd_complex_fft_p2)(temp, &temp[8], 8, -1, &scaleshift);
1234
1235 for (z = 0; z < (qmf_bank->no_channels >> 1); z++) {
1236 subband[64 + 2 * z] = temp[z] << scaleshift;
1237 subband[64 + 2 * z + 1] = temp[8 + z] << scaleshift;
1238 }
1239 }
1240
1241 psubband = &subband[0];
1242 psubband1 = &subband[2 * M - 1];
1243
1244 re = *psubband1;
1245
1246 *psubband = *psubband >> 1;
1247 psubband++;
1248 *psubband1 = ixheaacd_negate32(*psubband >> 1);
1249 psubband1--;
1250
1251 p_sin = qmf_bank->esbr_alt_sin_twiddle;
1252
1253 wim = *p_sin++;
1254 wre = *p_sin++;
1255
1256 im = *psubband1;
1257
1258 *psubband1-- = (WORD32)(
1259 (ixheaacd_add64(ixheaacd_mult64(re, wre), ixheaacd_mult64(im, wim))) >>
1260 32);
1261 *psubband++ = (WORD32)((ixheaacd_sub64_sat(ixheaacd_mult64(im, wre),
1262 ixheaacd_mult64(re, wim))) >>
1263 32);
1264
1265 psubband2 = &subband[64];
1266 psubband12 = &subband[2 * M - 1 + 64];
1267
1268 re = *psubband12;
1269
1270 *psubband12-- = ixheaacd_negate32_sat(*psubband2 >> 1);
1271
1272 *psubband2 = psubband2[1] >> 1;
1273
1274 psubband2++;
1275
1276 im = *psubband12;
1277
1278 *psubband2++ = ixheaacd_negate32_sat((WORD32)(
1279 (ixheaacd_add64(ixheaacd_mult64(re, wre), ixheaacd_mult64(im, wim))) >>
1280 32));
1281 *psubband12-- = (WORD32)((ixheaacd_sub64_sat(ixheaacd_mult64(re, wim),
1282 ixheaacd_mult64(im, wre))) >>
1283 32);
1284
1285 for (i = (M_2 - 2); i >= 0; i--) {
1286 im = psubband[0];
1287
1288 re = psubband[1];
1289
1290 re2 = *psubband1;
1291
1292 *psubband++ = (WORD32)(
1293 (ixheaacd_add64(ixheaacd_mult64(re, wim), ixheaacd_mult64(im, wre))) >>
1294 32);
1295 *psubband1-- = (WORD32)((ixheaacd_sub64_sat(ixheaacd_mult64(im, wim),
1296 ixheaacd_mult64(re, wre))) >>
1297 32);
1298
1299 im = psubband2[0];
1300
1301 re = psubband2[1];
1302
1303 re3 = *psubband12;
1304
1305 *psubband12-- = ixheaacd_negate32_sat((WORD32)(
1306 (ixheaacd_add64(ixheaacd_mult64(re, wim), ixheaacd_mult64(im, wre))) >>
1307 32));
1308 *psubband2++ = (WORD32)((ixheaacd_sub64_sat(ixheaacd_mult64(re, wre),
1309 ixheaacd_mult64(im, wim))) >>
1310 32);
1311
1312 wim = *p_sin++;
1313 wre = *p_sin++;
1314 im = psubband1[0];
1315
1316 *psubband1-- = (WORD32)(
1317 (ixheaacd_add64(ixheaacd_mult64(re2, wre), ixheaacd_mult64(im, wim))) >>
1318 32);
1319 *psubband++ = (WORD32)((ixheaacd_sub64_sat(ixheaacd_mult64(im, wre),
1320 ixheaacd_mult64(re2, wim))) >>
1321 32);
1322
1323 im = psubband12[0];
1324
1325 *psubband2++ = ixheaacd_negate32_sat((WORD32)(
1326 (ixheaacd_add64(ixheaacd_mult64(re3, wre), ixheaacd_mult64(im, wim))) >>
1327 32));
1328 *psubband12-- = (WORD32)((ixheaacd_sub64_sat(ixheaacd_mult64(re3, wim),
1329 ixheaacd_mult64(im, wre))) >>
1330 32);
1331 }
1332 }
1333
ixheaacd_esbr_fwd_modulation(const WORD32 * time_sample_buf,WORD32 * real_subband,WORD32 * imag_subband,ia_sbr_qmf_filter_bank_struct * qmf_bank,ia_qmf_dec_tables_struct * qmf_dec_tables_ptr)1334 VOID ixheaacd_esbr_fwd_modulation(
1335 const WORD32 *time_sample_buf, WORD32 *real_subband, WORD32 *imag_subband,
1336 ia_sbr_qmf_filter_bank_struct *qmf_bank,
1337 ia_qmf_dec_tables_struct *qmf_dec_tables_ptr) {
1338 WORD32 i;
1339 const WORD32 *time_sample_buf1 =
1340 &time_sample_buf[2 * qmf_bank->no_channels - 1];
1341 WORD32 temp1, temp2;
1342 WORD32 *t_real_subband = real_subband;
1343 WORD32 *t_imag_subband = imag_subband;
1344 const WORD32 *tcos;
1345
1346 for (i = qmf_bank->no_channels - 1; i >= 0; i--) {
1347 temp1 = ixheaacd_shr32(*time_sample_buf++, HQ_SHIFT_64);
1348 temp2 = ixheaacd_shr32(*time_sample_buf1--, HQ_SHIFT_64);
1349
1350 *t_real_subband++ = ixheaacd_sub32_sat(temp1, temp2);
1351
1352 *t_imag_subband++ = ixheaacd_add32_sat(temp1, temp2);
1353 }
1354
1355 ixheaacd_esbr_cos_sin_mod(real_subband, qmf_bank,
1356 qmf_dec_tables_ptr->esbr_w_16,
1357 qmf_dec_tables_ptr->dig_rev_table4_16);
1358
1359 tcos = qmf_bank->esbr_t_cos;
1360
1361 for (i = (qmf_bank->usb - qmf_bank->lsb - 1); i >= 0; i--) {
1362 WORD32 cosh, sinh;
1363 WORD32 re, im;
1364
1365 re = *real_subband;
1366 im = *imag_subband;
1367 cosh = *tcos++;
1368 sinh = *tcos++;
1369 *real_subband++ = (WORD32)((ixheaacd_add64(ixheaacd_mult64(re, cosh),
1370 ixheaacd_mult64(im, sinh))) >>
1371 31);
1372 *imag_subband++ =
1373 (WORD32)((ixheaacd_sub64_sat(ixheaacd_mult64(im, cosh),
1374 ixheaacd_mult64(re, sinh))) >>
1375 31);
1376 }
1377 }
1378
ixheaacd_sbr_qmfsyn64_winadd(WORD16 * tmp1,WORD16 * tmp2,WORD16 * inp1,WORD16 * sample_buffer,FLAG shift,WORD32 ch_fac)1379 VOID ixheaacd_sbr_qmfsyn64_winadd(WORD16 *tmp1, WORD16 *tmp2, WORD16 *inp1,
1380 WORD16 *sample_buffer, FLAG shift,
1381 WORD32 ch_fac) {
1382 WORD32 k;
1383 WORD32 rounding_fac = 0x8000;
1384 rounding_fac = rounding_fac >> shift;
1385
1386 for (k = 0; k < 64; k++) {
1387 WORD32 syn_out = rounding_fac;
1388
1389 syn_out = ixheaacd_add32_sat(syn_out,
1390 ixheaacd_mult16x16in32(tmp1[0 + k], inp1[k + 0]));
1391 syn_out = ixheaacd_add32_sat(
1392 syn_out, ixheaacd_mult16x16in32(tmp1[256 + k], inp1[k + 128]));
1393 syn_out = ixheaacd_add32_sat(
1394 syn_out, ixheaacd_mult16x16in32(tmp1[512 + k], inp1[k + 256]));
1395 syn_out = ixheaacd_add32_sat(
1396 syn_out, ixheaacd_mult16x16in32(tmp1[768 + k], inp1[k + 384]));
1397 syn_out = ixheaacd_add32_sat(
1398 syn_out, ixheaacd_mult16x16in32(tmp1[1024 + k], inp1[k + 512]));
1399
1400 syn_out = ixheaacd_add32_sat(
1401 syn_out, ixheaacd_mult16x16in32(tmp2[128 + k], inp1[k + 64]));
1402 syn_out = ixheaacd_add32_sat(
1403 syn_out, ixheaacd_mult16x16in32(tmp2[384 + k], inp1[k + 192]));
1404 syn_out = ixheaacd_add32_sat(
1405 syn_out, ixheaacd_mult16x16in32(tmp2[640 + k], inp1[k + 320]));
1406 syn_out = ixheaacd_add32_sat(
1407 syn_out, ixheaacd_mult16x16in32(tmp2[896 + k], inp1[k + 448]));
1408 syn_out = ixheaacd_add32_sat(
1409 syn_out, ixheaacd_mult16x16in32(tmp2[1152 + k], inp1[k + 576]));
1410
1411 sample_buffer[ch_fac * k] = (ixheaacd_shl32_sat(syn_out, shift) >> 16);
1412 }
1413 }
1414
ixheaacd_esbr_qmfsyn64_winadd(WORD32 * tmp1,WORD32 * tmp2,WORD32 * inp1,WORD32 * sample_buffer,WORD32 ch_fac)1415 VOID ixheaacd_esbr_qmfsyn64_winadd(WORD32 *tmp1, WORD32 *tmp2, WORD32 *inp1,
1416 WORD32 *sample_buffer, WORD32 ch_fac) {
1417 WORD32 k;
1418
1419 for (k = 0; k < 64; k++) {
1420 WORD64 syn_out = 0;
1421
1422 syn_out =
1423 ixheaacd_add64(syn_out, ixheaacd_mult64(tmp1[0 + k], inp1[k + 0]));
1424 syn_out =
1425 ixheaacd_add64(syn_out, ixheaacd_mult64(tmp1[256 + k], inp1[k + 128]));
1426 syn_out =
1427 ixheaacd_add64(syn_out, ixheaacd_mult64(tmp1[512 + k], inp1[k + 256]));
1428 syn_out =
1429 ixheaacd_add64(syn_out, ixheaacd_mult64(tmp1[768 + k], inp1[k + 384]));
1430 syn_out =
1431 ixheaacd_add64(syn_out, ixheaacd_mult64(tmp1[1024 + k], inp1[k + 512]));
1432
1433 syn_out =
1434 ixheaacd_add64(syn_out, ixheaacd_mult64(tmp2[128 + k], inp1[k + 64]));
1435 syn_out =
1436 ixheaacd_add64(syn_out, ixheaacd_mult64(tmp2[384 + k], inp1[k + 192]));
1437 syn_out =
1438 ixheaacd_add64(syn_out, ixheaacd_mult64(tmp2[640 + k], inp1[k + 320]));
1439 syn_out =
1440 ixheaacd_add64(syn_out, ixheaacd_mult64(tmp2[896 + k], inp1[k + 448]));
1441 syn_out =
1442 ixheaacd_add64(syn_out, ixheaacd_mult64(tmp2[1152 + k], inp1[k + 576]));
1443
1444 sample_buffer[ch_fac * k] = (WORD32)(syn_out >> 31);
1445 }
1446 }
1447
ixheaacd_shiftrountine(WORD32 * qmf_real,WORD32 * qmf_imag,WORD32 len,WORD32 common_shift)1448 VOID ixheaacd_shiftrountine(WORD32 *qmf_real, WORD32 *qmf_imag, WORD32 len,
1449 WORD32 common_shift) {
1450 WORD32 treal, timag;
1451 WORD32 j;
1452
1453 if (common_shift < 0) {
1454 WORD32 cshift = -common_shift;
1455 cshift = ixheaacd_min32(cshift, 31);
1456 for (j = len - 1; j >= 0; j--) {
1457 treal = *qmf_real;
1458 timag = *qmf_imag;
1459
1460 treal = (ixheaacd_shr32(treal, cshift));
1461 timag = (ixheaacd_shr32(timag, cshift));
1462
1463 *qmf_real++ = treal;
1464 *qmf_imag++ = timag;
1465 }
1466 } else {
1467 for (j = len - 1; j >= 0; j--) {
1468 treal = (ixheaacd_shl32_sat(*qmf_real, common_shift));
1469 timag = (ixheaacd_shl32_sat(*qmf_imag, common_shift));
1470 *qmf_real++ = treal;
1471 *qmf_imag++ = timag;
1472 }
1473 }
1474 }
1475
ixheaacd_shiftrountine_with_rnd(WORD32 * qmf_real,WORD32 * qmf_imag,WORD16 * filter_states,WORD32 len,WORD32 shift)1476 VOID ixheaacd_shiftrountine_with_rnd(WORD32 *qmf_real, WORD32 *qmf_imag,
1477 WORD16 *filter_states, WORD32 len,
1478 WORD32 shift) {
1479 WORD16 *filter_states_rev = filter_states + len;
1480 WORD32 treal, timag;
1481 WORD32 j;
1482
1483 for (j = (len - 1); j >= 0; j -= 2) {
1484 WORD32 r1, r2, i1, i2;
1485 i2 = qmf_imag[j];
1486 r2 = qmf_real[j];
1487
1488 r1 = *qmf_real++;
1489 i1 = *qmf_imag++;
1490
1491 timag = ixheaacd_add32_sat(i1, r1);
1492 timag = (ixheaacd_shl32_sat(timag, shift));
1493 filter_states_rev[j] = ixheaacd_round16(timag);
1494
1495 treal = ixheaacd_sub32_sat(i2, r2);
1496 treal = (ixheaacd_shl32_sat(treal, shift));
1497 filter_states[j] = ixheaacd_round16(treal);
1498
1499 treal = ixheaacd_sub32_sat(i1, r1);
1500 treal = (ixheaacd_shl32_sat(treal, shift));
1501 *filter_states++ = ixheaacd_round16(treal);
1502
1503 timag = ixheaacd_add32_sat(i2, r2);
1504 ;
1505 timag = (ixheaacd_shl32_sat(timag, shift));
1506 *filter_states_rev++ = ixheaacd_round16(timag);
1507 }
1508 }
1509
ixheaacd_shiftrountine_with_rnd_eld(WORD32 * qmf_real,WORD32 * qmf_imag,WORD16 * filter_states,WORD32 len,WORD32 shift)1510 VOID ixheaacd_shiftrountine_with_rnd_eld(WORD32 *qmf_real, WORD32 *qmf_imag,
1511 WORD16 *filter_states, WORD32 len,
1512 WORD32 shift) {
1513 WORD16 *filter_states_rev = filter_states + len;
1514 WORD32 treal, timag;
1515 WORD32 j;
1516
1517 for (j = (len - 1); j >= 0; j -= 2) {
1518 WORD32 r1, r2, i1, i2;
1519 i2 = qmf_imag[j];
1520 r2 = qmf_real[j];
1521 r1 = *qmf_real++;
1522 i1 = *qmf_imag++;
1523
1524 timag = ixheaacd_negate32(ixheaacd_add32_sat(i1, r1));
1525 timag = (ixheaacd_shl32_sat(timag, shift));
1526 filter_states_rev[j] = ixheaacd_round16(timag);
1527
1528 treal = ixheaacd_sub32_sat(r2, i2);
1529 treal = (ixheaacd_shl32_sat(treal, shift));
1530 filter_states[j] = ixheaacd_round16(treal);
1531
1532 treal = ixheaacd_sub32_sat(r1, i1);
1533 treal = (ixheaacd_shl32_sat(treal, shift));
1534 *filter_states++ = ixheaacd_round16(treal);
1535
1536 timag = ixheaacd_negate32(ixheaacd_add32_sat(i2, r2));
1537 timag = (ixheaacd_shl32_sat(timag, shift));
1538 *filter_states_rev++ = ixheaacd_round16(timag);
1539 }
1540 }
1541
ixheaacd_shiftrountine_with_rnd_hq(WORD32 * qmf_real,WORD32 * qmf_imag,WORD32 * filter_states,WORD32 len,WORD32 shift)1542 VOID ixheaacd_shiftrountine_with_rnd_hq(WORD32 *qmf_real, WORD32 *qmf_imag,
1543 WORD32 *filter_states, WORD32 len,
1544 WORD32 shift) {
1545 WORD32 *filter_states_rev = filter_states + len;
1546 WORD32 treal, timag;
1547 WORD32 j;
1548
1549 for (j = (len - 1); j >= 0; j -= 2) {
1550 WORD32 r1, r2, i1, i2;
1551 i2 = qmf_imag[j];
1552 r2 = qmf_real[j];
1553 r1 = *qmf_real++;
1554 i1 = *qmf_imag++;
1555
1556 timag = ixheaacd_add32_sat(i1, r1);
1557 timag = (ixheaacd_shl32_sat(timag, shift));
1558 filter_states_rev[j] = timag;
1559
1560 treal = ixheaacd_sub32_sat(i2, r2);
1561 treal = (ixheaacd_shl32_sat(treal, shift));
1562 filter_states[j] = treal;
1563
1564 treal = ixheaacd_sub32_sat(i1, r1);
1565 treal = (ixheaacd_shl32_sat(treal, shift));
1566 *filter_states++ = treal;
1567
1568 timag = ixheaacd_add32_sat(i2, r2);
1569 timag = (ixheaacd_shl32_sat(timag, shift));
1570 *filter_states_rev++ = timag;
1571 }
1572 }
1573
ixheaacd_radix4bfly(const WORD16 * w,WORD32 * x,WORD32 index1,WORD32 index)1574 VOID ixheaacd_radix4bfly(const WORD16 *w, WORD32 *x, WORD32 index1,
1575 WORD32 index) {
1576 int i;
1577 WORD32 l1, l2, h2, fft_jmp;
1578 WORD32 xt0_0, yt0_0, xt1_0, yt1_0, xt2_0, yt2_0;
1579 WORD32 xh0_0, xh1_0, xh20_0, xh21_0, xl0_0, xl1_0, xl20_0, xl21_0;
1580 WORD32 x_0, x_1, x_l1_0, x_l1_1, x_l2_0, x_l2_1;
1581 WORD32 x_h2_0, x_h2_1;
1582 WORD16 si10, si20, si30, co10, co20, co30;
1583
1584 WORD32 mul_1, mul_2, mul_3, mul_4, mul_5, mul_6;
1585 WORD32 mul_7, mul_8, mul_9, mul_10, mul_11, mul_12;
1586 WORD32 *x_l1;
1587 WORD32 *x_l2;
1588 WORD32 *x_h2;
1589 const WORD16 *w_ptr = w;
1590 WORD32 i1;
1591
1592 h2 = index << 1;
1593 l1 = index << 2;
1594 l2 = (index << 2) + (index << 1);
1595
1596 x_l1 = &(x[l1]);
1597 x_l2 = &(x[l2]);
1598 x_h2 = &(x[h2]);
1599
1600 fft_jmp = 6 * (index);
1601
1602 for (i1 = 0; i1 < index1; i1++) {
1603 for (i = 0; i < index; i++) {
1604 si10 = (*w_ptr++);
1605 co10 = (*w_ptr++);
1606 si20 = (*w_ptr++);
1607 co20 = (*w_ptr++);
1608 si30 = (*w_ptr++);
1609 co30 = (*w_ptr++);
1610
1611 x_0 = x[0];
1612 x_h2_0 = x[h2];
1613 x_l1_0 = x[l1];
1614 x_l2_0 = x[l2];
1615
1616 xh0_0 = ixheaacd_add32_sat(x_0, x_l1_0);
1617 xl0_0 = ixheaacd_sub32_sat(x_0, x_l1_0);
1618
1619 xh20_0 = ixheaacd_add32_sat(x_h2_0, x_l2_0);
1620 xl20_0 = ixheaacd_sub32_sat(x_h2_0, x_l2_0);
1621
1622 x[0] = ixheaacd_add32_sat(xh0_0, xh20_0);
1623 xt0_0 = ixheaacd_sub32_sat(xh0_0, xh20_0);
1624
1625 x_1 = x[1];
1626 x_h2_1 = x[h2 + 1];
1627 x_l1_1 = x[l1 + 1];
1628 x_l2_1 = x[l2 + 1];
1629
1630 xh1_0 = ixheaacd_add32_sat(x_1, x_l1_1);
1631 xl1_0 = ixheaacd_sub32_sat(x_1, x_l1_1);
1632
1633 xh21_0 = ixheaacd_add32_sat(x_h2_1, x_l2_1);
1634 xl21_0 = ixheaacd_sub32_sat(x_h2_1, x_l2_1);
1635
1636 x[1] = ixheaacd_add32_sat(xh1_0, xh21_0);
1637 yt0_0 = ixheaacd_sub32_sat(xh1_0, xh21_0);
1638
1639 xt1_0 = ixheaacd_add32_sat(xl0_0, xl21_0);
1640 xt2_0 = ixheaacd_sub32_sat(xl0_0, xl21_0);
1641
1642 yt2_0 = ixheaacd_add32_sat(xl1_0, xl20_0);
1643 yt1_0 = ixheaacd_sub32_sat(xl1_0, xl20_0);
1644
1645 mul_11 = ixheaacd_mult32x16in32(xt2_0, co30);
1646 mul_3 = ixheaacd_mult32x16in32(yt2_0, si30);
1647 x[l2] = (mul_3 + mul_11) << RADIXSHIFT;
1648
1649 mul_5 = ixheaacd_mult32x16in32(xt2_0, si30);
1650 mul_9 = ixheaacd_mult32x16in32(yt2_0, co30);
1651 x[l2 + 1] = (mul_9 - mul_5) << RADIXSHIFT;
1652
1653 mul_12 = ixheaacd_mult32x16in32(xt0_0, co20);
1654 mul_2 = ixheaacd_mult32x16in32(yt0_0, si20);
1655 x[l1] = (mul_2 + mul_12) << RADIXSHIFT;
1656
1657 mul_6 = ixheaacd_mult32x16in32(xt0_0, si20);
1658 mul_8 = ixheaacd_mult32x16in32(yt0_0, co20);
1659 x[l1 + 1] = (mul_8 - mul_6) << RADIXSHIFT;
1660
1661 mul_4 = ixheaacd_mult32x16in32(xt1_0, co10);
1662 mul_1 = ixheaacd_mult32x16in32(yt1_0, si10);
1663 x[h2] = (mul_1 + mul_4) << RADIXSHIFT;
1664
1665 mul_10 = ixheaacd_mult32x16in32(xt1_0, si10);
1666 mul_7 = ixheaacd_mult32x16in32(yt1_0, co10);
1667 x[h2 + 1] = (mul_7 - mul_10) << RADIXSHIFT;
1668
1669 x += 2;
1670 }
1671 x += fft_jmp;
1672 w_ptr = w_ptr - fft_jmp;
1673 }
1674 }
1675
ixheaacd_postradixcompute4(WORD32 * ptr_y,WORD32 * ptr_x,const WORD32 * p_dig_rev_tbl,WORD32 npoints)1676 VOID ixheaacd_postradixcompute4(WORD32 *ptr_y, WORD32 *ptr_x,
1677 const WORD32 *p_dig_rev_tbl, WORD32 npoints) {
1678 WORD32 i, k;
1679 WORD32 h2;
1680 WORD32 xh0_0, xh1_0, xl0_0, xl1_0;
1681 WORD32 xh0_1, xh1_1, xl0_1, xl1_1;
1682 WORD32 x_0, x_1, x_2, x_3;
1683 WORD32 xh0_2, xh1_2, xl0_2, xl1_2, xh0_3, xh1_3, xl0_3, xl1_3;
1684 WORD32 x_4, x_5, x_6, x_7;
1685 WORD32 x_8, x_9, x_a, x_b, x_c, x_d, x_e, x_f;
1686 WORD32 n00, n10, n20, n30, n01, n11, n21, n31;
1687 WORD32 n02, n12, n22, n32, n03, n13, n23, n33;
1688 WORD32 n0, j0;
1689 WORD32 *x2, *x0;
1690 WORD32 *y0, *y1, *y2, *y3;
1691
1692 y0 = ptr_y;
1693 y2 = ptr_y + (WORD32)npoints;
1694 x0 = ptr_x;
1695 x2 = ptr_x + (WORD32)(npoints >> 1);
1696
1697 y1 = y0 + (WORD32)(npoints >> 1);
1698 y3 = y2 + (WORD32)(npoints >> 1);
1699
1700 j0 = 4;
1701 n0 = npoints >> 2;
1702
1703 for (k = 0; k < 2; k++) {
1704 for (i = 0; i<npoints>> 1; i += 8) {
1705 h2 = *p_dig_rev_tbl++ >> 2;
1706 x_0 = *x0++;
1707 x_1 = *x0++;
1708 x_2 = *x0++;
1709 x_3 = *x0++;
1710 x_4 = *x0++;
1711 x_5 = *x0++;
1712 x_6 = *x0++;
1713 x_7 = *x0++;
1714
1715 xh0_0 = ixheaacd_add32_sat(x_0, x_4);
1716 xh1_0 = ixheaacd_add32_sat(x_1, x_5);
1717 xl0_0 = ixheaacd_sub32_sat(x_0, x_4);
1718 xl1_0 = ixheaacd_sub32_sat(x_1, x_5);
1719 xh0_1 = ixheaacd_add32_sat(x_2, x_6);
1720 xh1_1 = ixheaacd_add32_sat(x_3, x_7);
1721 xl0_1 = ixheaacd_sub32_sat(x_2, x_6);
1722 xl1_1 = ixheaacd_sub32_sat(x_3, x_7);
1723
1724 n00 = ixheaacd_add32_sat(xh0_0, xh0_1);
1725 n01 = ixheaacd_add32_sat(xh1_0, xh1_1);
1726 n10 = ixheaacd_add32_sat(xl0_0, xl1_1);
1727 n11 = ixheaacd_sub32_sat(xl1_0, xl0_1);
1728 n20 = ixheaacd_sub32_sat(xh0_0, xh0_1);
1729 n21 = ixheaacd_sub32_sat(xh1_0, xh1_1);
1730 n30 = ixheaacd_sub32_sat(xl0_0, xl1_1);
1731 n31 = ixheaacd_add32_sat(xl1_0, xl0_1);
1732
1733 y0[h2] = n00;
1734 y0[h2 + 1] = n01;
1735 y1[h2] = n10;
1736 y1[h2 + 1] = n11;
1737 y2[h2] = n20;
1738 y2[h2 + 1] = n21;
1739 y3[h2] = n30;
1740 y3[h2 + 1] = n31;
1741
1742 x_8 = *x2++;
1743 x_9 = *x2++;
1744 x_a = *x2++;
1745 x_b = *x2++;
1746 x_c = *x2++;
1747 x_d = *x2++;
1748 x_e = *x2++;
1749 x_f = *x2++;
1750
1751 xh0_2 = ixheaacd_add32_sat(x_8, x_c);
1752 xh1_2 = ixheaacd_add32_sat(x_9, x_d);
1753 xl0_2 = ixheaacd_sub32_sat(x_8, x_c);
1754 xl1_2 = ixheaacd_sub32_sat(x_9, x_d);
1755 xh0_3 = ixheaacd_add32_sat(x_a, x_e);
1756 xh1_3 = ixheaacd_add32_sat(x_b, x_f);
1757 xl0_3 = ixheaacd_sub32_sat(x_a, x_e);
1758 xl1_3 = ixheaacd_sub32_sat(x_b, x_f);
1759
1760 n02 = ixheaacd_add32_sat(xh0_2, xh0_3);
1761 n03 = ixheaacd_add32_sat(xh1_2, xh1_3);
1762 n12 = ixheaacd_add32_sat(xl0_2, xl1_3);
1763 n13 = ixheaacd_sub32_sat(xl1_2, xl0_3);
1764 n22 = ixheaacd_sub32_sat(xh0_2, xh0_3);
1765 n23 = ixheaacd_sub32_sat(xh1_2, xh1_3);
1766 n32 = ixheaacd_sub32_sat(xl0_2, xl1_3);
1767 n33 = ixheaacd_add32_sat(xl1_2, xl0_3);
1768
1769 y0[h2 + 2] = n02;
1770 y0[h2 + 3] = n03;
1771 y1[h2 + 2] = n12;
1772 y1[h2 + 3] = n13;
1773 y2[h2 + 2] = n22;
1774 y2[h2 + 3] = n23;
1775 y3[h2 + 2] = n32;
1776 y3[h2 + 3] = n33;
1777 }
1778 x0 += (WORD32)npoints >> 1;
1779 x2 += (WORD32)npoints >> 1;
1780 }
1781 }
1782
ixheaacd_postradixcompute2(WORD32 * ptr_y,WORD32 * ptr_x,const WORD32 * pdig_rev_tbl,WORD32 npoints)1783 VOID ixheaacd_postradixcompute2(WORD32 *ptr_y, WORD32 *ptr_x,
1784 const WORD32 *pdig_rev_tbl, WORD32 npoints) {
1785 WORD32 i, k;
1786 WORD32 h2;
1787 WORD32 x_0, x_1, x_2, x_3;
1788 WORD32 x_4, x_5, x_6, x_7;
1789 WORD32 x_8, x_9, x_a, x_b, x_c, x_d, x_e, x_f;
1790 WORD32 n00, n10, n20, n30, n01, n11, n21, n31;
1791 WORD32 n02, n12, n22, n32, n03, n13, n23, n33;
1792 WORD32 n0, j0;
1793 WORD32 *x2, *x0;
1794 WORD32 *y0, *y1, *y2, *y3;
1795
1796 y0 = ptr_y;
1797 y2 = ptr_y + (WORD32)npoints;
1798 x0 = ptr_x;
1799 x2 = ptr_x + (WORD32)(npoints >> 1);
1800
1801 y1 = y0 + (WORD32)(npoints >> 2);
1802 y3 = y2 + (WORD32)(npoints >> 2);
1803 j0 = 8;
1804 n0 = npoints >> 1;
1805
1806 for (k = 0; k < 2; k++) {
1807 for (i = 0; i<npoints>> 1; i += 8) {
1808 h2 = *pdig_rev_tbl++ >> 2;
1809
1810 x_0 = *x0++;
1811 x_1 = *x0++;
1812 x_2 = *x0++;
1813 x_3 = *x0++;
1814 x_4 = *x0++;
1815 x_5 = *x0++;
1816 x_6 = *x0++;
1817 x_7 = *x0++;
1818
1819 n00 = ixheaacd_add32_sat(x_0, x_2);
1820 n01 = ixheaacd_add32_sat(x_1, x_3);
1821 n20 = ixheaacd_sub32_sat(x_0, x_2);
1822 n21 = ixheaacd_sub32_sat(x_1, x_3);
1823 n10 = ixheaacd_add32_sat(x_4, x_6);
1824 n11 = ixheaacd_add32_sat(x_5, x_7);
1825 n30 = ixheaacd_sub32_sat(x_4, x_6);
1826 n31 = ixheaacd_sub32_sat(x_5, x_7);
1827
1828 y0[h2] = n00;
1829 y0[h2 + 1] = n01;
1830 y1[h2] = n10;
1831 y1[h2 + 1] = n11;
1832 y2[h2] = n20;
1833 y2[h2 + 1] = n21;
1834 y3[h2] = n30;
1835 y3[h2 + 1] = n31;
1836
1837 x_8 = *x2++;
1838 x_9 = *x2++;
1839 x_a = *x2++;
1840 x_b = *x2++;
1841 x_c = *x2++;
1842 x_d = *x2++;
1843 x_e = *x2++;
1844 x_f = *x2++;
1845
1846 n02 = ixheaacd_add32_sat(x_8, x_a);
1847 n03 = ixheaacd_add32_sat(x_9, x_b);
1848 n22 = ixheaacd_sub32_sat(x_8, x_a);
1849 n23 = ixheaacd_sub32_sat(x_9, x_b);
1850 n12 = ixheaacd_add32_sat(x_c, x_e);
1851 n13 = ixheaacd_add32_sat(x_d, x_f);
1852 n32 = ixheaacd_sub32_sat(x_c, x_e);
1853 n33 = ixheaacd_sub32_sat(x_d, x_f);
1854
1855 y0[h2 + 2] = n02;
1856 y0[h2 + 3] = n03;
1857 y1[h2 + 2] = n12;
1858 y1[h2 + 3] = n13;
1859 y2[h2 + 2] = n22;
1860 y2[h2 + 3] = n23;
1861 y3[h2 + 2] = n32;
1862 y3[h2 + 3] = n33;
1863 }
1864 x0 += (WORD32)npoints >> 1;
1865 x2 += (WORD32)npoints >> 1;
1866 }
1867 }
1868 #endif
1869