/src/libxaac/decoder/generic/ixheaacd_qmf_dec_generic.c
Line | Count | Source (jump to first uncovered line) |
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 "ixheaac_type_def.h" |
24 | | |
25 | | #include "ixheaac_constants.h" |
26 | | #include "ixheaac_basic_ops32.h" |
27 | | #include "ixheaac_basic_ops16.h" |
28 | | #include "ixheaac_basic_ops40.h" |
29 | | #include "ixheaac_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 "ixheaac_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 | 0 | #define DCT3_LEN (32) |
54 | | #define DCT2_LEN (64) |
55 | | |
56 | 0 | #define LP_SHIFT_VAL 7 |
57 | 713M | #define HQ_SHIFT_64 4 |
58 | 1.76G | #define RADIXSHIFT 1 |
59 | 66.7M | #define HQ_SHIFT_VAL 4 |
60 | | |
61 | | extern const WORD32 ixheaacd_ldmps_polyphase_filter_coeff_fix[1280]; |
62 | | |
63 | | VOID ixheaacd_dct3_32(WORD32 *input, WORD32 *output, |
64 | | const WORD16 *main_twidle_fwd, const WORD16 *post_tbl, |
65 | 0 | const WORD16 *w_16, const WORD32 *p_table) { |
66 | 0 | WORD32 n, k; |
67 | |
|
68 | 0 | WORD32 temp1[6]; |
69 | 0 | WORD32 temp2[4]; |
70 | 0 | WORD16 twid_re, twid_im; |
71 | 0 | WORD32 *ptr_reverse, *ptr_forward, *p_out, *ptr_out1; |
72 | 0 | const WORD16 *twidle_fwd, *twidle_rev; |
73 | |
|
74 | 0 | ptr_forward = &input[49]; |
75 | 0 | ptr_reverse = &input[47]; |
76 | |
|
77 | 0 | p_out = output; |
78 | 0 | twidle_fwd = main_twidle_fwd; |
79 | 0 | twidle_fwd += 4; |
80 | |
|
81 | 0 | *p_out++ = input[48] >> LP_SHIFT_VAL; |
82 | 0 | *p_out++ = 0; |
83 | |
|
84 | 0 | for (n = 1; n < DCT3_LEN / 2; n++) { |
85 | 0 | temp1[0] = *ptr_forward++; |
86 | 0 | temp1[1] = *ptr_reverse--; |
87 | 0 | temp1[0] = ixheaac_add32_sat(ixheaac_shr32(temp1[0], LP_SHIFT_VAL), |
88 | 0 | ixheaac_shr32(temp1[1], LP_SHIFT_VAL)); |
89 | |
|
90 | 0 | temp1[2] = *(ptr_forward - 33); |
91 | 0 | temp1[3] = *(ptr_reverse - 31); |
92 | 0 | temp1[1] = ixheaac_sub32_sat(ixheaac_shr32(temp1[2], LP_SHIFT_VAL), |
93 | 0 | ixheaac_shr32(temp1[3], LP_SHIFT_VAL)); |
94 | 0 | twid_re = *twidle_fwd++; |
95 | |
|
96 | 0 | twid_im = *twidle_fwd; |
97 | 0 | twidle_fwd += 3; |
98 | |
|
99 | 0 | *p_out++ = ixheaac_mult32x16in32(temp1[0], twid_re) + |
100 | 0 | ixheaac_mult32x16in32(temp1[1], twid_im); |
101 | 0 | *p_out++ = -ixheaac_mult32x16in32(temp1[1], twid_re) + |
102 | 0 | ixheaac_mult32x16in32(temp1[0], twid_im); |
103 | 0 | } |
104 | |
|
105 | 0 | twid_re = *twidle_fwd++; |
106 | |
|
107 | 0 | twid_im = *twidle_fwd; |
108 | 0 | twidle_fwd += 3; |
109 | |
|
110 | 0 | temp1[1] = *ptr_reverse--; |
111 | 0 | temp1[0] = *(ptr_reverse - 31); |
112 | 0 | temp1[1] = ixheaac_sub32_sat(ixheaac_shr32(temp1[1], LP_SHIFT_VAL), |
113 | 0 | ixheaac_shr32(temp1[0], LP_SHIFT_VAL)); |
114 | |
|
115 | 0 | temp1[0] = temp1[1]; |
116 | |
|
117 | 0 | temp2[2] = ixheaac_mult32x16in32(temp1[0], twid_re) + |
118 | 0 | ixheaac_mult32x16in32(temp1[1], twid_im); |
119 | 0 | temp2[3] = -ixheaac_mult32x16in32(temp1[1], twid_re) + |
120 | 0 | ixheaac_mult32x16in32(temp1[0], twid_im); |
121 | |
|
122 | 0 | ptr_forward = output; |
123 | 0 | ptr_reverse = &output[DCT3_LEN - 1]; |
124 | |
|
125 | 0 | temp2[0] = *ptr_forward++; |
126 | 0 | temp2[1] = *ptr_forward--; |
127 | |
|
128 | 0 | temp1[0] = -temp2[1] - temp2[3]; |
129 | 0 | temp1[1] = temp2[0] - temp2[2]; |
130 | 0 | temp2[0] = (temp2[0] + temp2[2] + temp1[0]); |
131 | 0 | temp2[1] = (temp2[1] - temp2[3] + temp1[1]); |
132 | |
|
133 | 0 | temp2[0] >>= 1; |
134 | 0 | temp2[1] >>= 1; |
135 | |
|
136 | 0 | *ptr_forward++ = temp2[0]; |
137 | 0 | *ptr_forward++ = temp2[1]; |
138 | |
|
139 | 0 | twidle_fwd = post_tbl + 2; |
140 | 0 | twidle_rev = post_tbl + 14; |
141 | |
|
142 | 0 | for (n = 1; n < DCT3_LEN / 4; n++) { |
143 | 0 | temp2[0] = *ptr_forward++; |
144 | 0 | temp2[1] = *ptr_forward--; |
145 | 0 | temp2[3] = *ptr_reverse--; |
146 | 0 | temp2[2] = *ptr_reverse++; |
147 | |
|
148 | 0 | twid_re = *twidle_rev; |
149 | 0 | twidle_rev -= 2; |
150 | 0 | twid_im = *twidle_fwd; |
151 | 0 | twidle_fwd += 2; |
152 | |
|
153 | 0 | temp1[0] = temp2[0] - temp2[2]; |
154 | 0 | temp1[1] = (temp2[0] + temp2[2]); |
155 | |
|
156 | 0 | temp1[2] = temp2[1] + temp2[3]; |
157 | 0 | temp1[3] = (temp2[1] - temp2[3]); |
158 | |
|
159 | 0 | temp1[4] = ixheaac_mult32x16in32(temp1[0], twid_re) + |
160 | 0 | ixheaac_mult32x16in32(temp1[2], twid_im); |
161 | 0 | temp1[5] = -ixheaac_mult32x16in32(temp1[2], twid_re) + |
162 | 0 | ixheaac_mult32x16in32(temp1[0], twid_im); |
163 | |
|
164 | 0 | temp1[1] >>= 1; |
165 | 0 | temp1[3] >>= 1; |
166 | |
|
167 | 0 | *ptr_forward++ = temp1[1] - temp1[4]; |
168 | 0 | *ptr_forward++ = temp1[3] + temp1[5]; |
169 | |
|
170 | 0 | *ptr_reverse-- = -temp1[3] + temp1[5]; |
171 | 0 | *ptr_reverse-- = temp1[1] + temp1[4]; |
172 | 0 | } |
173 | 0 | temp2[0] = *ptr_forward++; |
174 | 0 | temp2[1] = *ptr_forward--; |
175 | 0 | temp2[3] = *ptr_reverse--; |
176 | 0 | temp2[2] = *ptr_reverse++; |
177 | |
|
178 | 0 | twid_re = -*twidle_rev; |
179 | 0 | twidle_rev -= 2; |
180 | 0 | twid_im = *twidle_fwd; |
181 | 0 | twidle_fwd += 2; |
182 | |
|
183 | 0 | temp1[0] = temp2[0] - temp2[2]; |
184 | 0 | temp1[1] = (temp2[0] + temp2[2]); |
185 | |
|
186 | 0 | temp1[2] = temp2[1] + temp2[3]; |
187 | 0 | temp1[3] = (temp2[1] - temp2[3]); |
188 | |
|
189 | 0 | temp1[4] = ixheaac_mult32x16in32(temp1[0], twid_re) - |
190 | 0 | ixheaac_mult32x16in32(temp1[2], twid_im); |
191 | 0 | temp1[5] = ixheaac_mult32x16in32(temp1[2], twid_re) + |
192 | 0 | ixheaac_mult32x16in32(temp1[0], twid_im); |
193 | |
|
194 | 0 | temp1[1] >>= 1; |
195 | 0 | temp1[3] >>= 1; |
196 | 0 | *ptr_forward++ = temp1[1] + temp1[4]; |
197 | 0 | *ptr_forward++ = temp1[3] + temp1[5]; |
198 | |
|
199 | 0 | ixheaacd_radix4bfly(w_16, output, 1, 4); |
200 | 0 | ixheaacd_postradixcompute4(input, output, p_table, 16); |
201 | |
|
202 | 0 | output[0] = input[0]; |
203 | 0 | output[2] = input[1]; |
204 | |
|
205 | 0 | p_out = input + 2; |
206 | 0 | ptr_forward = output + 1; |
207 | 0 | ptr_reverse = output + 30; |
208 | 0 | ptr_out1 = input + 18; |
209 | |
|
210 | 0 | for (k = (DCT3_LEN / 4) - 1; k != 0; k--) { |
211 | 0 | WORD32 tempre, tempim; |
212 | |
|
213 | 0 | tempre = *p_out++; |
214 | 0 | tempim = *p_out++; |
215 | 0 | *ptr_forward = (tempim); |
216 | 0 | ptr_forward += 2; |
217 | 0 | *ptr_forward = (tempre); |
218 | 0 | ptr_forward += 2; |
219 | |
|
220 | 0 | tempre = *ptr_out1++; |
221 | 0 | tempim = *ptr_out1++; |
222 | 0 | *ptr_reverse = (tempim); |
223 | 0 | ptr_reverse -= 2; |
224 | 0 | *ptr_reverse = (tempre); |
225 | 0 | ptr_reverse -= 2; |
226 | 0 | } |
227 | |
|
228 | 0 | { |
229 | 0 | WORD32 tempre, tempim; |
230 | 0 | tempre = *p_out++; |
231 | 0 | tempim = *p_out++; |
232 | 0 | *ptr_forward = (tempim); |
233 | 0 | ptr_forward += 2; |
234 | 0 | *ptr_forward = (tempre); |
235 | 0 | ptr_forward += 2; |
236 | 0 | } |
237 | |
|
238 | 0 | return; |
239 | 0 | } |
240 | | |
241 | | VOID ixheaacd_dct2_64(WORD32 *x, WORD32 *X, |
242 | | ia_qmf_dec_tables_struct *qmf_dec_tables_ptr, |
243 | 0 | WORD16 *filter_states) { |
244 | 0 | ixheaacd_pretwdct2(x, X); |
245 | |
|
246 | 0 | ixheaacd_radix4bfly(qmf_dec_tables_ptr->w_32, X, 1, 8); |
247 | |
|
248 | 0 | ixheaacd_radix4bfly(qmf_dec_tables_ptr->w_32 + 48, X, 4, 2); |
249 | |
|
250 | 0 | ixheaacd_postradixcompute2(x, X, qmf_dec_tables_ptr->dig_rev_table2_32, 32); |
251 | |
|
252 | 0 | ixheaacd_fftposttw(x, qmf_dec_tables_ptr); |
253 | |
|
254 | 0 | ixheaacd_posttwdct2(x, filter_states, qmf_dec_tables_ptr); |
255 | |
|
256 | 0 | return; |
257 | 0 | } |
258 | | |
259 | | VOID ixheaacd_cos_sin_mod(WORD32 *subband, |
260 | | ia_sbr_qmf_filter_bank_struct *qmf_bank, |
261 | 1.75M | WORD16 *p_twiddle, WORD32 *p_dig_rev_tbl) { |
262 | 1.75M | WORD32 re2, re3; |
263 | 1.75M | WORD16 wim, wre; |
264 | | |
265 | 1.75M | WORD32 i, M_2; |
266 | 1.75M | WORD32 M = ixheaac_shr32(qmf_bank->no_channels, 1); |
267 | | |
268 | 1.75M | const WORD16 *p_sin; |
269 | 1.75M | const WORD16 *p_sin_cos = &qmf_bank->cos_twiddle[0]; |
270 | 1.75M | WORD32 subband_tmp[128]; |
271 | 1.75M | WORD32 re; |
272 | 1.75M | WORD32 im; |
273 | 1.75M | WORD32 *psubband, *psubband1; |
274 | 1.75M | WORD32 *psubband_t, *psubband1_t; |
275 | 1.75M | WORD32 *psubband2, *psubband12; |
276 | 1.75M | WORD32 *psubband_t2, *psubband1_t2; |
277 | | |
278 | 1.75M | M_2 = ixheaac_shr32(M, 1); |
279 | | |
280 | 1.75M | psubband = &subband[0]; |
281 | 1.75M | psubband1 = &subband[2 * M - 1]; |
282 | 1.75M | psubband_t = subband_tmp; |
283 | 1.75M | psubband1_t = &subband_tmp[2 * M - 1]; |
284 | | |
285 | 1.75M | psubband2 = &subband[64]; |
286 | 1.75M | psubband12 = &subband[2 * M - 1 + 64]; |
287 | 1.75M | psubband_t2 = &subband_tmp[64]; |
288 | 1.75M | psubband1_t2 = &subband_tmp[2 * M - 1 + 64]; |
289 | | |
290 | 10.5M | for (i = (M_2 >> 1) - 1; i >= 0; i--) { |
291 | 8.84M | re = *psubband++; |
292 | 8.84M | im = *psubband1--; |
293 | | |
294 | 8.84M | wim = *p_sin_cos++; |
295 | 8.84M | wre = *p_sin_cos++; |
296 | | |
297 | 8.84M | *psubband_t++ = ixheaac_add32_sat(ixheaac_mult32x16in32(re, wre), |
298 | 8.84M | ixheaac_mult32x16in32(im, wim)); |
299 | 8.84M | *psubband_t++ = ixheaac_sub32_sat(ixheaac_mult32x16in32(im, wre), |
300 | 8.84M | ixheaac_mult32x16in32(re, wim)); |
301 | | |
302 | 8.84M | re = *psubband2++; |
303 | 8.84M | im = *psubband12--; |
304 | | |
305 | 8.84M | *psubband_t2++ = ixheaac_sub32_sat(ixheaac_mult32x16in32(im, wim), |
306 | 8.84M | ixheaac_mult32x16in32(re, wre)); |
307 | 8.84M | *psubband_t2++ = ixheaac_add32_sat(ixheaac_mult32x16in32(re, wim), |
308 | 8.84M | ixheaac_mult32x16in32(im, wre)); |
309 | | |
310 | 8.84M | re = *psubband1--; |
311 | 8.84M | im = *psubband++; |
312 | | |
313 | 8.84M | wim = *p_sin_cos++; |
314 | 8.84M | wre = *p_sin_cos++; |
315 | | |
316 | 8.84M | *psubband1_t-- = ixheaac_sub32_sat(ixheaac_mult32x16in32(im, wre), |
317 | 8.84M | ixheaac_mult32x16in32(re, wim)); |
318 | 8.84M | *psubband1_t-- = ixheaac_add32_sat(ixheaac_mult32x16in32(re, wre), |
319 | 8.84M | ixheaac_mult32x16in32(im, wim)); |
320 | | |
321 | 8.84M | re = *psubband12--; |
322 | 8.84M | im = *psubband2++; |
323 | | |
324 | 8.84M | *psubband1_t2-- = ixheaac_add32_sat(ixheaac_mult32x16in32(re, wim), |
325 | 8.84M | ixheaac_mult32x16in32(im, wre)); |
326 | 8.84M | *psubband1_t2-- = ixheaac_sub32_sat(ixheaac_mult32x16in32(im, wim), |
327 | 8.84M | ixheaac_mult32x16in32(re, wre)); |
328 | | |
329 | 8.84M | re = *psubband++; |
330 | 8.84M | im = *psubband1--; |
331 | | |
332 | 8.84M | wim = *p_sin_cos++; |
333 | 8.84M | wre = *p_sin_cos++; |
334 | | |
335 | 8.84M | *psubband_t++ = ixheaac_add32_sat(ixheaac_mult32x16in32(re, wre), |
336 | 8.84M | ixheaac_mult32x16in32(im, wim)); |
337 | 8.84M | *psubband_t++ = ixheaac_sub32_sat(ixheaac_mult32x16in32(im, wre), |
338 | 8.84M | ixheaac_mult32x16in32(re, wim)); |
339 | | |
340 | 8.84M | re = *psubband2++; |
341 | 8.84M | im = *psubband12--; |
342 | | |
343 | 8.84M | *psubband_t2++ = ixheaac_sub32_sat(ixheaac_mult32x16in32(im, wim), |
344 | 8.84M | ixheaac_mult32x16in32(re, wre)); |
345 | 8.84M | *psubband_t2++ = ixheaac_add32_sat(ixheaac_mult32x16in32(re, wim), |
346 | 8.84M | ixheaac_mult32x16in32(im, wre)); |
347 | | |
348 | 8.84M | re = *psubband1--; |
349 | 8.84M | im = *psubband++; |
350 | | |
351 | 8.84M | wim = *p_sin_cos++; |
352 | 8.84M | wre = *p_sin_cos++; |
353 | | |
354 | 8.84M | *psubband1_t-- = ixheaac_sub32_sat(ixheaac_mult32x16in32(im, wre), |
355 | 8.84M | ixheaac_mult32x16in32(re, wim)); |
356 | 8.84M | *psubband1_t-- = ixheaac_add32_sat(ixheaac_mult32x16in32(re, wre), |
357 | 8.84M | ixheaac_mult32x16in32(im, wim)); |
358 | | |
359 | 8.84M | re = *psubband12--; |
360 | 8.84M | im = *psubband2++; |
361 | 8.84M | ; |
362 | | |
363 | 8.84M | *psubband1_t2-- = ixheaac_add32_sat(ixheaac_mult32x16in32(re, wim), |
364 | 8.84M | ixheaac_mult32x16in32(im, wre)); |
365 | 8.84M | *psubband1_t2-- = ixheaac_sub32_sat(ixheaac_mult32x16in32(im, wim), |
366 | 8.84M | ixheaac_mult32x16in32(re, wre)); |
367 | 8.84M | } |
368 | | |
369 | 1.75M | if (M == 32) { |
370 | 460k | ixheaacd_radix4bfly(p_twiddle, subband_tmp, 1, 8); |
371 | 460k | ixheaacd_radix4bfly(p_twiddle + 48, subband_tmp, 4, 2); |
372 | 460k | ixheaacd_postradixcompute2(subband, subband_tmp, p_dig_rev_tbl, 32); |
373 | | |
374 | 460k | ixheaacd_radix4bfly(p_twiddle, &subband_tmp[64], 1, 8); |
375 | 460k | ixheaacd_radix4bfly(p_twiddle + 48, &subband_tmp[64], 4, 2); |
376 | 460k | ixheaacd_postradixcompute2(&subband[64], &subband_tmp[64], p_dig_rev_tbl, |
377 | 460k | 32); |
378 | | |
379 | 1.28M | } else { |
380 | 1.28M | ixheaacd_radix4bfly(p_twiddle, subband_tmp, 1, 4); |
381 | 1.28M | ixheaacd_postradixcompute4(subband, subband_tmp, p_dig_rev_tbl, 16); |
382 | | |
383 | 1.28M | ixheaacd_radix4bfly(p_twiddle, &subband_tmp[64], 1, 4); |
384 | 1.28M | ixheaacd_postradixcompute4(&subband[64], &subband_tmp[64], p_dig_rev_tbl, |
385 | 1.28M | 16); |
386 | 1.28M | } |
387 | | |
388 | 1.75M | psubband = &subband[0]; |
389 | 1.75M | psubband1 = &subband[2 * M - 1]; |
390 | | |
391 | 1.75M | re = *psubband1; |
392 | | |
393 | 1.75M | *psubband = *psubband >> 1; |
394 | 1.75M | psubband++; |
395 | 1.75M | *psubband1 = ixheaac_negate32_sat(*psubband >> 1); |
396 | 1.75M | psubband1--; |
397 | | |
398 | 1.75M | p_sin = &qmf_bank->alt_sin_twiddle[0]; |
399 | 1.75M | wim = *p_sin++; |
400 | 1.75M | wre = *p_sin++; |
401 | | |
402 | 1.75M | im = *psubband1; |
403 | | |
404 | 1.75M | *psubband1-- = ixheaac_add32_sat(ixheaac_mult32x16in32(re, wre), |
405 | 1.75M | ixheaac_mult32x16in32(im, wim)); |
406 | 1.75M | *psubband++ = ixheaac_sub32_sat(ixheaac_mult32x16in32(im, wre), |
407 | 1.75M | ixheaac_mult32x16in32(re, wim)); |
408 | | |
409 | 1.75M | psubband2 = &subband[64]; |
410 | 1.75M | psubband12 = &subband[2 * M - 1 + 64]; |
411 | | |
412 | 1.75M | re = *psubband12; |
413 | | |
414 | 1.75M | *psubband12-- = ixheaac_negate32_sat(*psubband2 >> 1); |
415 | | |
416 | 1.75M | *psubband2 = psubband2[1] >> 1; |
417 | | |
418 | 1.75M | psubband2++; |
419 | | |
420 | 1.75M | im = *psubband12; |
421 | | |
422 | 1.75M | *psubband2++ = ixheaac_negate32_sat(ixheaac_add32_sat( |
423 | 1.75M | ixheaac_mult32x16in32(re, wre), ixheaac_mult32x16in32(im, wim))); |
424 | 1.75M | *psubband12-- = ixheaac_sub32_sat(ixheaac_mult32x16in32(re, wim), |
425 | 1.75M | ixheaac_mult32x16in32(im, wre)); |
426 | | |
427 | 17.6M | for (i = (M_2 - 2); i >= 0; i--) { |
428 | 15.9M | im = psubband[0]; |
429 | | |
430 | 15.9M | re = psubband[1]; |
431 | | |
432 | 15.9M | re2 = *psubband1; |
433 | | |
434 | 15.9M | *psubband++ = ixheaac_add32_sat(ixheaac_mult32x16in32(re, wim), |
435 | 15.9M | ixheaac_mult32x16in32(im, wre)); |
436 | 15.9M | *psubband1-- = ixheaac_sub32_sat(ixheaac_mult32x16in32(im, wim), |
437 | 15.9M | ixheaac_mult32x16in32(re, wre)); |
438 | | |
439 | 15.9M | im = psubband2[0]; |
440 | | |
441 | 15.9M | re = psubband2[1]; |
442 | | |
443 | 15.9M | re3 = *psubband12; |
444 | | |
445 | 15.9M | *psubband12-- = ixheaac_negate32_sat(ixheaac_add32_sat( |
446 | 15.9M | ixheaac_mult32x16in32(re, wim), ixheaac_mult32x16in32(im, wre))); |
447 | 15.9M | *psubband2++ = ixheaac_sub32_sat(ixheaac_mult32x16in32(re, wre), |
448 | 15.9M | ixheaac_mult32x16in32(im, wim)); |
449 | | |
450 | 15.9M | wim = *p_sin++; |
451 | 15.9M | wre = *p_sin++; |
452 | 15.9M | im = psubband1[0]; |
453 | | |
454 | 15.9M | *psubband1-- = ixheaac_add32_sat(ixheaac_mult32x16in32(re2, wre), |
455 | 15.9M | ixheaac_mult32x16in32(im, wim)); |
456 | 15.9M | *psubband++ = ixheaac_sub32_sat(ixheaac_mult32x16in32(im, wre), |
457 | 15.9M | ixheaac_mult32x16in32(re2, wim)); |
458 | | |
459 | 15.9M | im = psubband12[0]; |
460 | | |
461 | 15.9M | *psubband2++ = ixheaac_negate32_sat(ixheaac_add32_sat( |
462 | 15.9M | ixheaac_mult32x16in32(re3, wre), ixheaac_mult32x16in32(im, wim))); |
463 | 15.9M | *psubband12-- = ixheaac_sub32_sat(ixheaac_mult32x16in32(re3, wim), |
464 | 15.9M | ixheaac_mult32x16in32(im, wre)); |
465 | 15.9M | } |
466 | 1.75M | } |
467 | | |
468 | | VOID ixheaacd_fwd_modulation(const WORD32 *p_time_in1, WORD32 *real_subband, |
469 | | WORD32 *imag_subband, |
470 | | ia_sbr_qmf_filter_bank_struct *qmf_bank, |
471 | | ia_qmf_dec_tables_struct *qmf_dec_tables_ptr, |
472 | 875k | WORD32 ld_mps_flag) { |
473 | 875k | WORD32 i; |
474 | 875k | const WORD32 *p_time_in2 = &p_time_in1[2 * qmf_bank->no_channels - 1]; |
475 | 875k | WORD32 temp1, temp2; |
476 | 875k | WORD32 *t_real_subband = real_subband; |
477 | 875k | WORD32 *t_imag_subband = imag_subband; |
478 | 875k | const WORD16 *tcos; |
479 | | |
480 | 34.2M | for (i = qmf_bank->no_channels - 1; i >= 0; i--) { |
481 | 33.3M | temp1 = ixheaac_shr32(*p_time_in1++, HQ_SHIFT_VAL); |
482 | 33.3M | temp2 = ixheaac_shr32(*p_time_in2--, HQ_SHIFT_VAL); |
483 | | |
484 | 33.3M | *t_real_subband++ = ixheaac_sub32_sat(temp1, temp2); |
485 | | |
486 | 33.3M | *t_imag_subband++ = ixheaac_add32_sat(temp1, temp2); |
487 | 33.3M | } |
488 | | |
489 | 875k | if (qmf_bank->no_channels != 64) |
490 | 707k | ixheaacd_cos_sin_mod(real_subband, qmf_bank, qmf_dec_tables_ptr->w_16, |
491 | 707k | qmf_dec_tables_ptr->dig_rev_table4_16); |
492 | 168k | else |
493 | 168k | ixheaacd_cos_sin_mod(real_subband, qmf_bank, qmf_dec_tables_ptr->w_32, |
494 | 168k | qmf_dec_tables_ptr->dig_rev_table2_32); |
495 | | |
496 | 875k | if (ld_mps_flag == 0) { |
497 | 803k | tcos = qmf_bank->t_cos; |
498 | | |
499 | 14.6M | for (i = (qmf_bank->usb - qmf_bank->lsb - 1); i >= 0; i--) { |
500 | 13.8M | WORD16 cosh, sinh; |
501 | 13.8M | WORD32 re, im; |
502 | | |
503 | 13.8M | re = *real_subband; |
504 | 13.8M | im = *imag_subband; |
505 | 13.8M | cosh = *tcos++; |
506 | 13.8M | sinh = *tcos++; |
507 | 13.8M | *real_subband++ = |
508 | 13.8M | ixheaac_add32_sat(ixheaac_mult32x16in32_shl(re, cosh), |
509 | 13.8M | ixheaac_mult32x16in32_shl(im, sinh)); |
510 | 13.8M | *imag_subband++ = |
511 | 13.8M | ixheaac_sub32_sat(ixheaac_mult32x16in32_shl(im, cosh), |
512 | 13.8M | ixheaac_mult32x16in32_shl(re, sinh)); |
513 | 13.8M | } |
514 | 803k | } else { |
515 | 71.8k | WORD32 i_band; |
516 | 1.22M | for (i = 0; i < min(64, qmf_bank->no_channels); i += 2) { |
517 | 1.15M | i_band = real_subband[i]; |
518 | 1.15M | real_subband[i] = -imag_subband[i]; |
519 | 1.15M | imag_subband[i] = i_band; |
520 | | |
521 | 1.15M | i_band = -real_subband[i + 1]; |
522 | 1.15M | real_subband[i + 1] = imag_subband[i + 1]; |
523 | 1.15M | imag_subband[i + 1] = i_band; |
524 | 1.15M | } |
525 | 71.8k | } |
526 | 875k | } |
527 | | |
528 | | VOID ixheaacd_sbr_qmfanal32_winadd(WORD16 *inp1, WORD16 *inp2, |
529 | | const WORD16 *p_qmf1, const WORD16 *p_qmf2, |
530 | 0 | WORD32 *p_out) { |
531 | 0 | WORD32 n; |
532 | |
|
533 | 0 | for (n = 0; n < 32; n += 2) { |
534 | 0 | WORD32 accu; |
535 | |
|
536 | 0 | accu = ixheaac_mult16x16in32(inp1[n + 0], p_qmf1[2 * (n + 0)]); |
537 | 0 | accu = ixheaac_add32_sat( |
538 | 0 | accu, ixheaac_mult16x16in32(inp1[n + 64], p_qmf1[2 * (n + 64)])); |
539 | 0 | accu = ixheaac_add32_sat( |
540 | 0 | accu, ixheaac_mult16x16in32(inp1[n + 128], p_qmf1[2 * (n + 128)])); |
541 | 0 | accu = ixheaac_add32_sat( |
542 | 0 | accu, ixheaac_mult16x16in32(inp1[n + 192], p_qmf1[2 * (n + 192)])); |
543 | 0 | accu = ixheaac_add32_sat( |
544 | 0 | accu, ixheaac_mult16x16in32(inp1[n + 256], p_qmf1[2 * (n + 256)])); |
545 | 0 | p_out[n] = accu; |
546 | |
|
547 | 0 | accu = ixheaac_mult16x16in32(inp1[n + 1 + 0], p_qmf1[2 * (n + 1 + 0)]); |
548 | 0 | accu = ixheaac_add32_sat( |
549 | 0 | accu, |
550 | 0 | ixheaac_mult16x16in32(inp1[n + 1 + 64], p_qmf1[2 * (n + 1 + 64)])); |
551 | 0 | accu = ixheaac_add32_sat( |
552 | 0 | accu, |
553 | 0 | ixheaac_mult16x16in32(inp1[n + 1 + 128], p_qmf1[2 * (n + 1 + 128)])); |
554 | 0 | accu = ixheaac_add32_sat( |
555 | 0 | accu, |
556 | 0 | ixheaac_mult16x16in32(inp1[n + 1 + 192], p_qmf1[2 * (n + 1 + 192)])); |
557 | 0 | accu = ixheaac_add32_sat( |
558 | 0 | accu, |
559 | 0 | ixheaac_mult16x16in32(inp1[n + 1 + 256], p_qmf1[2 * (n + 1 + 256)])); |
560 | 0 | p_out[n + 1] = accu; |
561 | |
|
562 | 0 | accu = ixheaac_mult16x16in32(inp2[n + 0], p_qmf2[2 * (n + 0)]); |
563 | 0 | accu = ixheaac_add32_sat( |
564 | 0 | accu, ixheaac_mult16x16in32(inp2[n + 64], p_qmf2[2 * (n + 64)])); |
565 | 0 | accu = ixheaac_add32_sat( |
566 | 0 | accu, ixheaac_mult16x16in32(inp2[n + 128], p_qmf2[2 * (n + 128)])); |
567 | 0 | accu = ixheaac_add32_sat( |
568 | 0 | accu, ixheaac_mult16x16in32(inp2[n + 192], p_qmf2[2 * (n + 192)])); |
569 | 0 | accu = ixheaac_add32_sat( |
570 | 0 | accu, ixheaac_mult16x16in32(inp2[n + 256], p_qmf2[2 * (n + 256)])); |
571 | 0 | p_out[n + 32] = accu; |
572 | |
|
573 | 0 | accu = ixheaac_mult16x16in32(inp2[n + 1 + 0], p_qmf2[2 * (n + 1 + 0)]); |
574 | 0 | accu = ixheaac_add32_sat( |
575 | 0 | accu, |
576 | 0 | ixheaac_mult16x16in32(inp2[n + 1 + 64], p_qmf2[2 * (n + 1 + 64)])); |
577 | 0 | accu = ixheaac_add32_sat( |
578 | 0 | accu, |
579 | 0 | ixheaac_mult16x16in32(inp2[n + 1 + 128], p_qmf2[2 * (n + 1 + 128)])); |
580 | 0 | accu = ixheaac_add32_sat( |
581 | 0 | accu, |
582 | 0 | ixheaac_mult16x16in32(inp2[n + 1 + 192], p_qmf2[2 * (n + 1 + 192)])); |
583 | 0 | accu = ixheaac_add32_sat( |
584 | 0 | accu, |
585 | 0 | ixheaac_mult16x16in32(inp2[n + 1 + 256], p_qmf2[2 * (n + 1 + 256)])); |
586 | 0 | p_out[n + 1 + 32] = accu; |
587 | 0 | } |
588 | 0 | } |
589 | | |
590 | | VOID ixheaacd_cplx_anal_qmffilt(const WORD16 *time_sample_buf, |
591 | | ia_sbr_scale_fact_struct *sbr_scale_factor, |
592 | | WORD32 **qmf_real, WORD32 **qmf_imag, |
593 | | ia_sbr_qmf_filter_bank_struct *qmf_bank, |
594 | | ia_qmf_dec_tables_struct *qmf_dec_tables_ptr, |
595 | | WORD32 ch_fac, WORD32 low_pow_flag, |
596 | 51.4k | WORD audio_object_type) { |
597 | 51.4k | WORD32 i, k; |
598 | 51.4k | WORD32 num_time_slots = qmf_bank->num_time_slots; |
599 | | |
600 | 51.4k | WORD32 analysis_buffer[4 * NO_ANALYSIS_CHANNELS] = { 0 }; |
601 | 51.4k | WORD16 *filter_states = qmf_bank->core_samples_buffer; |
602 | | |
603 | 51.4k | WORD16 *fp1, *fp2, *tmp; |
604 | | |
605 | 51.4k | WORD16 *filter_1; |
606 | 51.4k | WORD16 *filter_2; |
607 | 51.4k | WORD16 *filt_ptr; |
608 | 51.4k | WORD32 start_slot = 0; |
609 | | |
610 | 51.4k | if (audio_object_type != AOT_ER_AAC_ELD && |
611 | 51.4k | audio_object_type != AOT_ER_AAC_LD) { |
612 | 0 | qmf_bank->filter_pos += |
613 | 0 | (qmf_dec_tables_ptr->qmf_c - qmf_bank->analy_win_coeff); |
614 | 0 | qmf_bank->analy_win_coeff = qmf_dec_tables_ptr->qmf_c; |
615 | 51.4k | } else { |
616 | 51.4k | qmf_bank->filter_pos += |
617 | 51.4k | (qmf_dec_tables_ptr->qmf_c_eld3 - qmf_bank->analy_win_coeff); |
618 | 51.4k | qmf_bank->analy_win_coeff = qmf_dec_tables_ptr->qmf_c_eld3; |
619 | 51.4k | } |
620 | | |
621 | 51.4k | filter_1 = qmf_bank->filter_pos; |
622 | | |
623 | 51.4k | if (audio_object_type != AOT_ER_AAC_ELD && |
624 | 51.4k | audio_object_type != AOT_ER_AAC_LD) { |
625 | 0 | filter_2 = filter_1 + 64; |
626 | 51.4k | } else { |
627 | 51.4k | filter_2 = filter_1 + qmf_bank->no_channels; |
628 | 51.4k | } |
629 | | |
630 | 51.4k | sbr_scale_factor->st_lb_scale = 0; |
631 | 51.4k | sbr_scale_factor->lb_scale = -10; |
632 | 51.4k | if (!low_pow_flag) { |
633 | 51.4k | if (audio_object_type != AOT_ER_AAC_ELD && |
634 | 51.4k | audio_object_type != AOT_ER_AAC_LD) { |
635 | 0 | sbr_scale_factor->lb_scale = -8; |
636 | 51.4k | } else { |
637 | 51.4k | sbr_scale_factor->lb_scale = -9; |
638 | 51.4k | } |
639 | 51.4k | if (qmf_bank->no_channels != 64) { |
640 | 40.6k | qmf_bank->cos_twiddle = |
641 | 40.6k | (WORD16 *)qmf_dec_tables_ptr->sbr_sin_cos_twiddle_l32; |
642 | 40.6k | qmf_bank->alt_sin_twiddle = |
643 | 40.6k | (WORD16 *)qmf_dec_tables_ptr->sbr_alt_sin_twiddle_l32; |
644 | 40.6k | } else { |
645 | 10.7k | qmf_bank->cos_twiddle = |
646 | 10.7k | (WORD16 *)qmf_dec_tables_ptr->sbr_sin_cos_twiddle_l64; |
647 | 10.7k | qmf_bank->alt_sin_twiddle = |
648 | 10.7k | (WORD16 *)qmf_dec_tables_ptr->sbr_alt_sin_twiddle_l64; |
649 | 10.7k | } |
650 | 51.4k | if (audio_object_type != AOT_ER_AAC_ELD && |
651 | 51.4k | audio_object_type != AOT_ER_AAC_LD) { |
652 | 0 | qmf_bank->t_cos = (WORD16 *)qmf_dec_tables_ptr->sbr_t_cos_sin_l32; |
653 | 51.4k | } else { |
654 | 51.4k | qmf_bank->t_cos = |
655 | 51.4k | (WORD16 *)qmf_dec_tables_ptr->ixheaacd_sbr_t_cos_sin_l32_eld; |
656 | 51.4k | } |
657 | 51.4k | } |
658 | | |
659 | 51.4k | fp1 = qmf_bank->anal_filter_states; |
660 | 51.4k | fp2 = qmf_bank->anal_filter_states + qmf_bank->no_channels; |
661 | | |
662 | 51.4k | if (audio_object_type == AOT_ER_AAC_ELD || |
663 | 51.4k | audio_object_type == AOT_ER_AAC_LD) { |
664 | 51.4k | filter_2 = qmf_bank->filter_2; |
665 | 51.4k | fp1 = qmf_bank->fp1_anal; |
666 | 51.4k | fp2 = qmf_bank->fp2_anal; |
667 | 51.4k | } |
668 | | |
669 | 854k | for (i = start_slot; i < num_time_slots + start_slot; i++) { |
670 | 31.8M | for (k = 0; k < qmf_bank->no_channels; k++) |
671 | 31.0M | filter_states[qmf_bank->no_channels - 1 - k] = |
672 | 31.0M | time_sample_buf[ch_fac * k]; |
673 | | |
674 | 803k | if (audio_object_type != AOT_ER_AAC_ELD && |
675 | 803k | audio_object_type != AOT_ER_AAC_LD) { |
676 | 0 | ixheaacd_sbr_qmfanal32_winadd(fp1, fp2, filter_1, filter_2, |
677 | 0 | analysis_buffer); |
678 | 803k | } else { |
679 | 803k | ixheaacd_sbr_qmfanal32_winadd_eld(fp1, fp2, filter_1, filter_2, |
680 | 803k | analysis_buffer); |
681 | 803k | } |
682 | | |
683 | 803k | time_sample_buf += qmf_bank->no_channels * ch_fac; |
684 | | |
685 | 803k | filter_states -= qmf_bank->no_channels; |
686 | 803k | if (filter_states < qmf_bank->anal_filter_states) { |
687 | 82.3k | filter_states = qmf_bank->anal_filter_states + |
688 | 82.3k | ((qmf_bank->no_channels * 10) - qmf_bank->no_channels); |
689 | 82.3k | } |
690 | | |
691 | 803k | tmp = fp1; |
692 | 803k | fp1 = fp2; |
693 | 803k | fp2 = tmp; |
694 | 803k | if (audio_object_type != AOT_ER_AAC_ELD && |
695 | 803k | audio_object_type != AOT_ER_AAC_LD) { |
696 | 0 | filter_1 += 64; |
697 | 0 | filter_2 += 64; |
698 | 803k | } else { |
699 | 803k | filter_1 += qmf_bank->no_channels; |
700 | 803k | filter_2 += qmf_bank->no_channels; |
701 | 803k | } |
702 | | |
703 | 803k | filt_ptr = filter_1; |
704 | 803k | filter_1 = filter_2; |
705 | 803k | filter_2 = filt_ptr; |
706 | 803k | if (audio_object_type != AOT_ER_AAC_ELD && |
707 | 803k | audio_object_type != AOT_ER_AAC_LD) { |
708 | 0 | if (filter_2 > (qmf_bank->analy_win_coeff + 640)) { |
709 | 0 | filter_1 = (WORD16 *)qmf_bank->analy_win_coeff; |
710 | 0 | filter_2 = (WORD16 *)qmf_bank->analy_win_coeff + 64; |
711 | 0 | } |
712 | 803k | } else { |
713 | 803k | if (filter_2 > |
714 | 803k | (qmf_bank->analy_win_coeff + (qmf_bank->no_channels * 10))) { |
715 | 78.0k | filter_1 = (WORD16 *)qmf_bank->analy_win_coeff; |
716 | 78.0k | filter_2 = (WORD16 *)qmf_bank->analy_win_coeff + qmf_bank->no_channels; |
717 | 78.0k | } |
718 | 803k | } |
719 | | |
720 | 803k | if (!low_pow_flag) { |
721 | 803k | ixheaacd_fwd_modulation(analysis_buffer, qmf_real[i], qmf_imag[i], |
722 | 803k | qmf_bank, qmf_dec_tables_ptr, 0); |
723 | 803k | } else { |
724 | 0 | ixheaacd_dct3_32( |
725 | 0 | (WORD32 *)analysis_buffer, qmf_real[i], qmf_dec_tables_ptr->dct23_tw, |
726 | 0 | qmf_dec_tables_ptr->post_fft_tbl, qmf_dec_tables_ptr->w_16, |
727 | 0 | qmf_dec_tables_ptr->dig_rev_table4_16); |
728 | 0 | } |
729 | 803k | } |
730 | | |
731 | 51.4k | qmf_bank->filter_pos = filter_1; |
732 | 51.4k | qmf_bank->core_samples_buffer = filter_states; |
733 | | |
734 | 51.4k | if (audio_object_type == AOT_ER_AAC_ELD || audio_object_type == AOT_ER_AAC_LD) |
735 | | |
736 | 51.4k | { |
737 | 51.4k | qmf_bank->fp1_anal = fp1; |
738 | 51.4k | qmf_bank->fp2_anal = fp2; |
739 | 51.4k | qmf_bank->filter_2 = filter_2; |
740 | 51.4k | } |
741 | 51.4k | } |
742 | | |
743 | | VOID ixheaacd_cplx_anal_qmffilt_32(const WORD32 *time_sample_buf, |
744 | | ia_sbr_scale_fact_struct *sbr_scale_factor, |
745 | | WORD32 **qmf_real, WORD32 **qmf_imag, |
746 | | ia_sbr_qmf_filter_bank_struct *qmf_bank, |
747 | | ia_qmf_dec_tables_struct *qmf_dec_tables_ptr, |
748 | 4.54k | WORD32 ch_fac, WORD32 ldsbr_present) { |
749 | 4.54k | WORD32 i, k; |
750 | 4.54k | WORD32 num_time_slots = qmf_bank->num_time_slots; |
751 | | |
752 | 4.54k | WORD32 analysis_buffer[4 * NO_ANALYSIS_CHANNELS]; |
753 | 4.54k | WORD32 *filter_states = qmf_bank->core_samples_buffer_32; |
754 | | |
755 | 4.54k | WORD32 *fp1, *fp2, *tmp; |
756 | | |
757 | 4.54k | WORD32 *filter_1; |
758 | 4.54k | WORD32 *filter_2; |
759 | 4.54k | WORD32 *filt_ptr; |
760 | 4.54k | WORD32 start_slot = 2; |
761 | | |
762 | 4.54k | if (ldsbr_present) { |
763 | 4.50k | qmf_bank->filter_pos_32 += |
764 | 4.50k | (qmf_dec_tables_ptr->qmf_c_ldsbr_mps - qmf_bank->analy_win_coeff_32); |
765 | 4.50k | qmf_bank->analy_win_coeff_32 = qmf_dec_tables_ptr->qmf_c_ldsbr_mps; |
766 | 4.50k | } else { |
767 | 43 | qmf_bank->filter_pos_32 += (ixheaacd_ldmps_polyphase_filter_coeff_fix - |
768 | 43 | qmf_bank->analy_win_coeff_32); |
769 | 43 | qmf_bank->analy_win_coeff_32 = |
770 | 43 | (WORD32 *)ixheaacd_ldmps_polyphase_filter_coeff_fix; |
771 | 43 | } |
772 | | |
773 | 4.54k | filter_1 = qmf_bank->filter_pos_32; |
774 | 4.54k | filter_2 = filter_1 + qmf_bank->no_channels; |
775 | | |
776 | 4.54k | sbr_scale_factor->st_lb_scale = 0; |
777 | 4.54k | sbr_scale_factor->lb_scale = -10; |
778 | | |
779 | 4.54k | sbr_scale_factor->lb_scale = -9; |
780 | 4.54k | if (qmf_bank->no_channels != 64) { |
781 | 4.53k | qmf_bank->cos_twiddle = |
782 | 4.53k | (WORD16 *)qmf_dec_tables_ptr->sbr_sin_cos_twiddle_l32; |
783 | 4.53k | qmf_bank->alt_sin_twiddle = |
784 | 4.53k | (WORD16 *)qmf_dec_tables_ptr->sbr_alt_sin_twiddle_l32; |
785 | 4.53k | } else { |
786 | 10 | qmf_bank->cos_twiddle = |
787 | 10 | (WORD16 *)qmf_dec_tables_ptr->sbr_sin_cos_twiddle_l64; |
788 | 10 | qmf_bank->alt_sin_twiddle = |
789 | 10 | (WORD16 *)qmf_dec_tables_ptr->sbr_alt_sin_twiddle_l64; |
790 | 10 | } |
791 | 4.54k | qmf_bank->t_cos = |
792 | 4.54k | (WORD16 *)qmf_dec_tables_ptr->ixheaacd_sbr_t_cos_sin_l32_eld; |
793 | | |
794 | 4.54k | fp1 = qmf_bank->anal_filter_states_32; |
795 | 4.54k | fp2 = qmf_bank->anal_filter_states_32 + qmf_bank->no_channels; |
796 | | |
797 | 4.54k | filter_2 = qmf_bank->filter_2_32; |
798 | 4.54k | fp1 = qmf_bank->fp1_anal_32; |
799 | 4.54k | fp2 = qmf_bank->fp2_anal_32; |
800 | | |
801 | 76.4k | for (i = start_slot; i < num_time_slots + start_slot; i++) { |
802 | 2.37M | for (k = 0; k < qmf_bank->no_channels; k++) |
803 | 2.30M | filter_states[qmf_bank->no_channels - 1 - k] = |
804 | 2.30M | time_sample_buf[ch_fac * k]; |
805 | | |
806 | 71.8k | if (ldsbr_present) { |
807 | 71.4k | ixheaacd_sbr_qmfanal32_winadd_eld_32(fp1, fp2, filter_1, filter_2, |
808 | 71.4k | analysis_buffer); |
809 | 71.4k | } else { |
810 | 400 | ixheaacd_sbr_qmfanal32_winadd_eld_mps(fp1, fp2, filter_1, filter_2, |
811 | 400 | analysis_buffer); |
812 | 400 | } |
813 | | |
814 | 71.8k | time_sample_buf += qmf_bank->no_channels * ch_fac; |
815 | | |
816 | 71.8k | filter_states -= qmf_bank->no_channels; |
817 | | |
818 | 71.8k | if (filter_states < qmf_bank->anal_filter_states_32) { |
819 | 7.63k | filter_states = qmf_bank->anal_filter_states_32 + |
820 | 7.63k | ((qmf_bank->no_channels * 10) - qmf_bank->no_channels); |
821 | 7.63k | } |
822 | | |
823 | 71.8k | tmp = fp1; |
824 | 71.8k | fp1 = fp2; |
825 | 71.8k | fp2 = tmp; |
826 | | |
827 | 71.8k | filter_1 += qmf_bank->no_channels; |
828 | 71.8k | filter_2 += qmf_bank->no_channels; |
829 | | |
830 | 71.8k | filt_ptr = filter_1; |
831 | 71.8k | filter_1 = filter_2; |
832 | 71.8k | filter_2 = filt_ptr; |
833 | | |
834 | 71.8k | if (filter_2 > |
835 | 71.8k | (qmf_bank->analy_win_coeff_32 + (qmf_bank->no_channels * 10))) { |
836 | 6.68k | filter_1 = (WORD32 *)qmf_bank->analy_win_coeff_32; |
837 | 6.68k | filter_2 = (WORD32 *)qmf_bank->analy_win_coeff_32 + qmf_bank->no_channels; |
838 | 6.68k | } |
839 | | |
840 | 71.8k | ixheaacd_fwd_modulation(analysis_buffer, qmf_real[i], qmf_imag[i], qmf_bank, |
841 | 71.8k | qmf_dec_tables_ptr, 1); |
842 | 71.8k | } |
843 | | |
844 | 4.54k | qmf_bank->filter_pos_32 = filter_1; |
845 | 4.54k | qmf_bank->core_samples_buffer_32 = filter_states; |
846 | | |
847 | 4.54k | qmf_bank->fp1_anal_32 = fp1; |
848 | 4.54k | qmf_bank->fp2_anal_32 = fp2; |
849 | 4.54k | qmf_bank->filter_2_32 = filter_2; |
850 | 4.54k | } |
851 | | VOID ixheaacd_inv_modulation_lp(WORD32 *qmf_real, WORD16 *filter_states, |
852 | | ia_sbr_qmf_filter_bank_struct *syn_qmf, |
853 | 0 | ia_qmf_dec_tables_struct *qmf_dec_tables_ptr) { |
854 | 0 | WORD32 L = syn_qmf->no_channels; |
855 | 0 | const WORD32 M = (L >> 1); |
856 | 0 | WORD32 *dct_in = qmf_real; |
857 | 0 | WORD32 time_out[2 * NO_SYNTHESIS_CHANNELS]; |
858 | |
|
859 | 0 | WORD32 *ptime_out = &time_out[0]; |
860 | |
|
861 | 0 | if (L == 64) |
862 | 0 | ixheaacd_dct2_64(dct_in, ptime_out, qmf_dec_tables_ptr, filter_states + M); |
863 | 0 | else |
864 | 0 | ixheaacd_dct2_32(dct_in, time_out, qmf_dec_tables_ptr, filter_states); |
865 | |
|
866 | 0 | filter_states[3 * M] = 0; |
867 | 0 | } |
868 | | |
869 | | VOID ixheaacd_inv_emodulation(WORD32 *qmf_real, |
870 | | ia_sbr_qmf_filter_bank_struct *syn_qmf, |
871 | 874k | ia_qmf_dec_tables_struct *qmf_dec_tables_ptr) { |
872 | 874k | if (syn_qmf->no_channels == 64) |
873 | 292k | ixheaacd_cos_sin_mod(qmf_real, syn_qmf, qmf_dec_tables_ptr->w_32, |
874 | 292k | qmf_dec_tables_ptr->dig_rev_table2_32); |
875 | 582k | else |
876 | 582k | ixheaacd_cos_sin_mod(qmf_real, syn_qmf, qmf_dec_tables_ptr->w_16, |
877 | 582k | qmf_dec_tables_ptr->dig_rev_table4_16); |
878 | 874k | } |
879 | | |
880 | | VOID ixheaacd_esbr_radix4bfly(const WORD32 *w, WORD32 *x, WORD32 index1, |
881 | 42.6M | WORD32 index) { |
882 | 42.6M | int i; |
883 | 42.6M | WORD32 l1, l2, h2, fft_jmp; |
884 | 42.6M | WORD32 xt0_0, yt0_0, xt1_0, yt1_0, xt2_0, yt2_0; |
885 | 42.6M | WORD32 xh0_0, xh1_0, xh20_0, xh21_0, xl0_0, xl1_0, xl20_0, xl21_0; |
886 | 42.6M | WORD32 x_0, x_1, x_l1_0, x_l1_1, x_l2_0, x_l2_1; |
887 | 42.6M | WORD32 x_h2_0, x_h2_1; |
888 | 42.6M | WORD32 si10, si20, si30, co10, co20, co30; |
889 | | |
890 | 42.6M | WORD64 mul_1, mul_2, mul_3, mul_4, mul_5, mul_6; |
891 | 42.6M | WORD64 mul_7, mul_8, mul_9, mul_10, mul_11, mul_12; |
892 | 42.6M | const WORD32 *w_ptr = w; |
893 | 42.6M | WORD32 i1; |
894 | | |
895 | 42.6M | h2 = index << 1; |
896 | 42.6M | l1 = index << 2; |
897 | 42.6M | l2 = (index << 2) + (index << 1); |
898 | | |
899 | 42.6M | fft_jmp = 6 * (index); |
900 | | |
901 | 122M | for (i1 = 0; i1 < index1; i1++) { |
902 | 347M | for (i = 0; i < index; i++) { |
903 | 268M | si10 = (*w_ptr++); |
904 | 268M | co10 = (*w_ptr++); |
905 | 268M | si20 = (*w_ptr++); |
906 | 268M | co20 = (*w_ptr++); |
907 | 268M | si30 = (*w_ptr++); |
908 | 268M | co30 = (*w_ptr++); |
909 | | |
910 | 268M | x_0 = x[0]; |
911 | 268M | x_h2_0 = x[h2]; |
912 | 268M | x_l1_0 = x[l1]; |
913 | 268M | x_l2_0 = x[l2]; |
914 | | |
915 | 268M | xh0_0 = ixheaac_add32_sat(x_0, x_l1_0); |
916 | 268M | xl0_0 = ixheaac_sub32_sat(x_0, x_l1_0); |
917 | | |
918 | 268M | xh20_0 = ixheaac_add32_sat(x_h2_0, x_l2_0); |
919 | 268M | xl20_0 = ixheaac_sub32_sat(x_h2_0, x_l2_0); |
920 | | |
921 | 268M | x[0] = ixheaac_add32_sat(xh0_0, xh20_0); |
922 | 268M | xt0_0 = ixheaac_sub32_sat(xh0_0, xh20_0); |
923 | | |
924 | 268M | x_1 = x[1]; |
925 | 268M | x_h2_1 = x[h2 + 1]; |
926 | 268M | x_l1_1 = x[l1 + 1]; |
927 | 268M | x_l2_1 = x[l2 + 1]; |
928 | | |
929 | 268M | xh1_0 = ixheaac_add32_sat(x_1, x_l1_1); |
930 | 268M | xl1_0 = ixheaac_sub32_sat(x_1, x_l1_1); |
931 | | |
932 | 268M | xh21_0 = ixheaac_add32_sat(x_h2_1, x_l2_1); |
933 | 268M | xl21_0 = ixheaac_sub32_sat(x_h2_1, x_l2_1); |
934 | | |
935 | 268M | x[1] = ixheaac_add32_sat(xh1_0, xh21_0); |
936 | 268M | yt0_0 = ixheaac_sub32_sat(xh1_0, xh21_0); |
937 | | |
938 | 268M | xt1_0 = ixheaac_add32_sat(xl0_0, xl21_0); |
939 | 268M | xt2_0 = ixheaac_sub32_sat(xl0_0, xl21_0); |
940 | | |
941 | 268M | yt2_0 = ixheaac_add32_sat(xl1_0, xl20_0); |
942 | 268M | yt1_0 = ixheaac_sub32_sat(xl1_0, xl20_0); |
943 | | |
944 | 268M | mul_11 = ixheaac_mult64(xt2_0, co30); |
945 | 268M | mul_3 = ixheaac_mult64(yt2_0, si30); |
946 | 268M | x[l2] = (WORD32)((mul_3 + mul_11) >> 32) << RADIXSHIFT; |
947 | | |
948 | 268M | mul_5 = ixheaac_mult64(xt2_0, si30); |
949 | 268M | mul_9 = ixheaac_mult64(yt2_0, co30); |
950 | 268M | x[l2 + 1] = (WORD32)((mul_9 - mul_5) >> 32) << RADIXSHIFT; |
951 | | |
952 | 268M | mul_12 = ixheaac_mult64(xt0_0, co20); |
953 | 268M | mul_2 = ixheaac_mult64(yt0_0, si20); |
954 | 268M | x[l1] = (WORD32)((mul_2 + mul_12) >> 32) << RADIXSHIFT; |
955 | | |
956 | 268M | mul_6 = ixheaac_mult64(xt0_0, si20); |
957 | 268M | mul_8 = ixheaac_mult64(yt0_0, co20); |
958 | 268M | x[l1 + 1] = (WORD32)((mul_8 - mul_6) >> 32) << RADIXSHIFT; |
959 | | |
960 | 268M | mul_4 = ixheaac_mult64(xt1_0, co10); |
961 | 268M | mul_1 = ixheaac_mult64(yt1_0, si10); |
962 | 268M | x[h2] = (WORD32)((mul_1 + mul_4) >> 32) << RADIXSHIFT; |
963 | | |
964 | 268M | mul_10 = ixheaac_mult64(xt1_0, si10); |
965 | 268M | mul_7 = ixheaac_mult64(yt1_0, co10); |
966 | 268M | x[h2 + 1] = (WORD32)((mul_7 - mul_10) >> 32) << RADIXSHIFT; |
967 | | |
968 | 268M | x += 2; |
969 | 268M | } |
970 | 79.3M | x += fft_jmp; |
971 | 79.3M | w_ptr = w_ptr - fft_jmp; |
972 | 79.3M | } |
973 | 42.6M | } |
974 | | |
975 | | VOID ixheaacd_esbr_postradixcompute2(WORD32 *ptr_y, WORD32 *ptr_x, |
976 | | const WORD32 *pdig_rev_tbl, |
977 | 12.2M | WORD32 npoints) { |
978 | 12.2M | WORD32 i, k; |
979 | 12.2M | WORD32 h2; |
980 | 12.2M | WORD32 x_0, x_1, x_2, x_3; |
981 | 12.2M | WORD32 x_4, x_5, x_6, x_7; |
982 | 12.2M | WORD32 x_8, x_9, x_a, x_b, x_c, x_d, x_e, x_f; |
983 | 12.2M | WORD32 n00, n10, n20, n30, n01, n11, n21, n31; |
984 | 12.2M | WORD32 n02, n12, n22, n32, n03, n13, n23, n33; |
985 | 12.2M | WORD32 *x2, *x0; |
986 | 12.2M | WORD32 *y0, *y1, *y2, *y3; |
987 | | |
988 | 12.2M | y0 = ptr_y; |
989 | 12.2M | y2 = ptr_y + (WORD32)npoints; |
990 | 12.2M | x0 = ptr_x; |
991 | 12.2M | x2 = ptr_x + (WORD32)(npoints >> 1); |
992 | | |
993 | 12.2M | y1 = y0 + (WORD32)(npoints >> 2); |
994 | 12.2M | y3 = y2 + (WORD32)(npoints >> 2); |
995 | | |
996 | 36.6M | for (k = 0; k < 2; k++) { |
997 | 73.2M | for (i = 0; i<npoints>> 1; i += 8) { |
998 | 48.8M | h2 = *pdig_rev_tbl++ >> 2; |
999 | | |
1000 | 48.8M | x_0 = *x0++; |
1001 | 48.8M | x_1 = *x0++; |
1002 | 48.8M | x_2 = *x0++; |
1003 | 48.8M | x_3 = *x0++; |
1004 | 48.8M | x_4 = *x0++; |
1005 | 48.8M | x_5 = *x0++; |
1006 | 48.8M | x_6 = *x0++; |
1007 | 48.8M | x_7 = *x0++; |
1008 | | |
1009 | 48.8M | n00 = ixheaac_add32_sat(x_0, x_2); |
1010 | 48.8M | n01 = ixheaac_add32_sat(x_1, x_3); |
1011 | 48.8M | n20 = ixheaac_sub32_sat(x_0, x_2); |
1012 | 48.8M | n21 = ixheaac_sub32_sat(x_1, x_3); |
1013 | 48.8M | n10 = ixheaac_add32_sat(x_4, x_6); |
1014 | 48.8M | n11 = ixheaac_add32_sat(x_5, x_7); |
1015 | 48.8M | n30 = ixheaac_sub32_sat(x_4, x_6); |
1016 | 48.8M | n31 = ixheaac_sub32_sat(x_5, x_7); |
1017 | | |
1018 | 48.8M | y0[h2] = n00; |
1019 | 48.8M | y0[h2 + 1] = n01; |
1020 | 48.8M | y1[h2] = n10; |
1021 | 48.8M | y1[h2 + 1] = n11; |
1022 | 48.8M | y2[h2] = n20; |
1023 | 48.8M | y2[h2 + 1] = n21; |
1024 | 48.8M | y3[h2] = n30; |
1025 | 48.8M | y3[h2 + 1] = n31; |
1026 | | |
1027 | 48.8M | x_8 = *x2++; |
1028 | 48.8M | x_9 = *x2++; |
1029 | 48.8M | x_a = *x2++; |
1030 | 48.8M | x_b = *x2++; |
1031 | 48.8M | x_c = *x2++; |
1032 | 48.8M | x_d = *x2++; |
1033 | 48.8M | x_e = *x2++; |
1034 | 48.8M | x_f = *x2++; |
1035 | | |
1036 | 48.8M | n02 = ixheaac_add32_sat(x_8, x_a); |
1037 | 48.8M | n03 = ixheaac_add32_sat(x_9, x_b); |
1038 | 48.8M | n22 = ixheaac_sub32_sat(x_8, x_a); |
1039 | 48.8M | n23 = ixheaac_sub32_sat(x_9, x_b); |
1040 | 48.8M | n12 = ixheaac_add32_sat(x_c, x_e); |
1041 | 48.8M | n13 = ixheaac_add32_sat(x_d, x_f); |
1042 | 48.8M | n32 = ixheaac_sub32_sat(x_c, x_e); |
1043 | 48.8M | n33 = ixheaac_sub32_sat(x_d, x_f); |
1044 | | |
1045 | 48.8M | y0[h2 + 2] = n02; |
1046 | 48.8M | y0[h2 + 3] = n03; |
1047 | 48.8M | y1[h2 + 2] = n12; |
1048 | 48.8M | y1[h2 + 3] = n13; |
1049 | 48.8M | y2[h2 + 2] = n22; |
1050 | 48.8M | y2[h2 + 3] = n23; |
1051 | 48.8M | y3[h2 + 2] = n32; |
1052 | 48.8M | y3[h2 + 3] = n33; |
1053 | 48.8M | } |
1054 | 24.4M | x0 += (WORD32)npoints >> 1; |
1055 | 24.4M | x2 += (WORD32)npoints >> 1; |
1056 | 24.4M | } |
1057 | 12.2M | } |
1058 | | |
1059 | | VOID ixheaacd_esbr_postradixcompute4(WORD32 *ptr_y, WORD32 *ptr_x, |
1060 | | const WORD32 *p_dig_rev_tbl, |
1061 | 18.2M | WORD32 npoints) { |
1062 | 18.2M | WORD32 i, k; |
1063 | 18.2M | WORD32 h2; |
1064 | 18.2M | WORD32 xh0_0, xh1_0, xl0_0, xl1_0; |
1065 | 18.2M | WORD32 xh0_1, xh1_1, xl0_1, xl1_1; |
1066 | 18.2M | WORD32 x_0, x_1, x_2, x_3; |
1067 | 18.2M | WORD32 xh0_2, xh1_2, xl0_2, xl1_2, xh0_3, xh1_3, xl0_3, xl1_3; |
1068 | 18.2M | WORD32 x_4, x_5, x_6, x_7; |
1069 | 18.2M | WORD32 x_8, x_9, x_a, x_b, x_c, x_d, x_e, x_f; |
1070 | 18.2M | WORD32 n00, n10, n20, n30, n01, n11, n21, n31; |
1071 | 18.2M | WORD32 n02, n12, n22, n32, n03, n13, n23, n33; |
1072 | 18.2M | WORD32 *x2, *x0; |
1073 | 18.2M | WORD32 *y0, *y1, *y2, *y3; |
1074 | | |
1075 | 18.2M | y0 = ptr_y; |
1076 | 18.2M | y2 = ptr_y + (WORD32)npoints; |
1077 | 18.2M | x0 = ptr_x; |
1078 | 18.2M | x2 = ptr_x + (WORD32)(npoints >> 1); |
1079 | | |
1080 | 18.2M | y1 = y0 + (WORD32)(npoints >> 1); |
1081 | 18.2M | y3 = y2 + (WORD32)(npoints >> 1); |
1082 | | |
1083 | 54.7M | for (k = 0; k < 2; k++) { |
1084 | 73.0M | for (i = 0; i<npoints>> 1; i += 8) { |
1085 | 36.5M | h2 = *p_dig_rev_tbl++ >> 2; |
1086 | 36.5M | x_0 = *x0++; |
1087 | 36.5M | x_1 = *x0++; |
1088 | 36.5M | x_2 = *x0++; |
1089 | 36.5M | x_3 = *x0++; |
1090 | 36.5M | x_4 = *x0++; |
1091 | 36.5M | x_5 = *x0++; |
1092 | 36.5M | x_6 = *x0++; |
1093 | 36.5M | x_7 = *x0++; |
1094 | | |
1095 | 36.5M | xh0_0 = ixheaac_add32_sat(x_0, x_4); |
1096 | 36.5M | xh1_0 = ixheaac_add32_sat(x_1, x_5); |
1097 | 36.5M | xl0_0 = ixheaac_sub32_sat(x_0, x_4); |
1098 | 36.5M | xl1_0 = ixheaac_sub32_sat(x_1, x_5); |
1099 | 36.5M | xh0_1 = ixheaac_add32_sat(x_2, x_6); |
1100 | 36.5M | xh1_1 = ixheaac_add32_sat(x_3, x_7); |
1101 | 36.5M | xl0_1 = ixheaac_sub32_sat(x_2, x_6); |
1102 | 36.5M | xl1_1 = ixheaac_sub32_sat(x_3, x_7); |
1103 | | |
1104 | 36.5M | n00 = ixheaac_add32_sat(xh0_0, xh0_1); |
1105 | 36.5M | n01 = ixheaac_add32_sat(xh1_0, xh1_1); |
1106 | 36.5M | n10 = ixheaac_add32_sat(xl0_0, xl1_1); |
1107 | 36.5M | n11 = ixheaac_sub32_sat(xl1_0, xl0_1); |
1108 | 36.5M | n20 = ixheaac_sub32_sat(xh0_0, xh0_1); |
1109 | 36.5M | n21 = ixheaac_sub32_sat(xh1_0, xh1_1); |
1110 | 36.5M | n30 = ixheaac_sub32_sat(xl0_0, xl1_1); |
1111 | 36.5M | n31 = ixheaac_add32_sat(xl1_0, xl0_1); |
1112 | | |
1113 | 36.5M | y0[h2] = n00; |
1114 | 36.5M | y0[h2 + 1] = n01; |
1115 | 36.5M | y1[h2] = n10; |
1116 | 36.5M | y1[h2 + 1] = n11; |
1117 | 36.5M | y2[h2] = n20; |
1118 | 36.5M | y2[h2 + 1] = n21; |
1119 | 36.5M | y3[h2] = n30; |
1120 | 36.5M | y3[h2 + 1] = n31; |
1121 | | |
1122 | 36.5M | x_8 = *x2++; |
1123 | 36.5M | x_9 = *x2++; |
1124 | 36.5M | x_a = *x2++; |
1125 | 36.5M | x_b = *x2++; |
1126 | 36.5M | x_c = *x2++; |
1127 | 36.5M | x_d = *x2++; |
1128 | 36.5M | x_e = *x2++; |
1129 | 36.5M | x_f = *x2++; |
1130 | | |
1131 | 36.5M | xh0_2 = ixheaac_add32_sat(x_8, x_c); |
1132 | 36.5M | xh1_2 = ixheaac_add32_sat(x_9, x_d); |
1133 | 36.5M | xl0_2 = ixheaac_sub32_sat(x_8, x_c); |
1134 | 36.5M | xl1_2 = ixheaac_sub32_sat(x_9, x_d); |
1135 | 36.5M | xh0_3 = ixheaac_add32_sat(x_a, x_e); |
1136 | 36.5M | xh1_3 = ixheaac_add32_sat(x_b, x_f); |
1137 | 36.5M | xl0_3 = ixheaac_sub32_sat(x_a, x_e); |
1138 | 36.5M | xl1_3 = ixheaac_sub32_sat(x_b, x_f); |
1139 | | |
1140 | 36.5M | n02 = ixheaac_add32_sat(xh0_2, xh0_3); |
1141 | 36.5M | n03 = ixheaac_add32_sat(xh1_2, xh1_3); |
1142 | 36.5M | n12 = ixheaac_add32_sat(xl0_2, xl1_3); |
1143 | 36.5M | n13 = ixheaac_sub32_sat(xl1_2, xl0_3); |
1144 | 36.5M | n22 = ixheaac_sub32_sat(xh0_2, xh0_3); |
1145 | 36.5M | n23 = ixheaac_sub32_sat(xh1_2, xh1_3); |
1146 | 36.5M | n32 = ixheaac_sub32_sat(xl0_2, xl1_3); |
1147 | 36.5M | n33 = ixheaac_add32_sat(xl1_2, xl0_3); |
1148 | | |
1149 | 36.5M | y0[h2 + 2] = n02; |
1150 | 36.5M | y0[h2 + 3] = n03; |
1151 | 36.5M | y1[h2 + 2] = n12; |
1152 | 36.5M | y1[h2 + 3] = n13; |
1153 | 36.5M | y2[h2 + 2] = n22; |
1154 | 36.5M | y2[h2 + 3] = n23; |
1155 | 36.5M | y3[h2 + 2] = n32; |
1156 | 36.5M | y3[h2 + 3] = n33; |
1157 | 36.5M | } |
1158 | 36.5M | x0 += (WORD32)npoints >> 1; |
1159 | 36.5M | x2 += (WORD32)npoints >> 1; |
1160 | 36.5M | } |
1161 | 18.2M | } |
1162 | | |
1163 | | VOID ixheaacd_esbr_cos_sin_mod(WORD32 *subband, |
1164 | | ia_sbr_qmf_filter_bank_struct *qmf_bank, |
1165 | 18.7M | WORD32 *p_twiddle, WORD32 *p_dig_rev_tbl) { |
1166 | 18.7M | WORD32 z; |
1167 | 18.7M | WORD32 temp[128]; |
1168 | 18.7M | WORD32 scaleshift = 0; |
1169 | | |
1170 | 18.7M | WORD32 re2, re3; |
1171 | 18.7M | WORD32 wim, wre; |
1172 | | |
1173 | 18.7M | WORD32 i, M_2; |
1174 | 18.7M | WORD32 M = ixheaac_shr32(qmf_bank->no_channels, 1); |
1175 | | |
1176 | 18.7M | const WORD32 *p_sin; |
1177 | 18.7M | const WORD32 *p_sin_cos; |
1178 | | |
1179 | 18.7M | WORD32 subband_tmp[128]; |
1180 | 18.7M | WORD32 re; |
1181 | 18.7M | WORD32 im; |
1182 | 18.7M | WORD32 *psubband, *psubband1; |
1183 | 18.7M | WORD32 *psubband_t, *psubband1_t; |
1184 | 18.7M | WORD32 *psubband2, *psubband12; |
1185 | 18.7M | WORD32 *psubband_t2, *psubband1_t2; |
1186 | | |
1187 | 18.7M | M_2 = ixheaac_shr32(M, 1); |
1188 | | |
1189 | 18.7M | p_sin_cos = qmf_bank->esbr_cos_twiddle; |
1190 | | |
1191 | 18.7M | psubband = &subband[0]; |
1192 | 18.7M | psubband1 = &subband[2 * M - 1]; |
1193 | 18.7M | psubband_t = subband_tmp; |
1194 | 18.7M | psubband1_t = &subband_tmp[2 * M - 1]; |
1195 | | |
1196 | 18.7M | psubband2 = &subband[64]; |
1197 | 18.7M | psubband12 = &subband[2 * M - 1 + 64]; |
1198 | 18.7M | psubband_t2 = &subband_tmp[64]; |
1199 | 18.7M | psubband1_t2 = &subband_tmp[2 * M - 1 + 64]; |
1200 | | |
1201 | 112M | for (i = (M_2 >> 1) - 1; i >= 0; i--) { |
1202 | 93.4M | re = *psubband++; |
1203 | 93.4M | im = *psubband1--; |
1204 | | |
1205 | 93.4M | wim = *p_sin_cos++; |
1206 | 93.4M | wre = *p_sin_cos++; |
1207 | | |
1208 | 93.4M | *psubband_t++ = (WORD32)( |
1209 | 93.4M | (ixheaac_add64(ixheaac_mult64(re, wre), ixheaac_mult64(im, wim))) >> |
1210 | 93.4M | 32); |
1211 | 93.4M | *psubband_t++ = (WORD32)((ixheaac_sub64_sat(ixheaac_mult64(im, wre), |
1212 | 93.4M | ixheaac_mult64(re, wim))) >> |
1213 | 93.4M | 32); |
1214 | | |
1215 | 93.4M | re = *psubband2++; |
1216 | 93.4M | im = *psubband12--; |
1217 | | |
1218 | 93.4M | *psubband_t2++ = (WORD32)((ixheaac_sub64_sat(ixheaac_mult64(im, wim), |
1219 | 93.4M | ixheaac_mult64(re, wre))) >> |
1220 | 93.4M | 32); |
1221 | 93.4M | *psubband_t2++ = (WORD32)( |
1222 | 93.4M | (ixheaac_add64(ixheaac_mult64(re, wim), ixheaac_mult64(im, wre))) >> |
1223 | 93.4M | 32); |
1224 | | |
1225 | 93.4M | re = *psubband1--; |
1226 | 93.4M | im = *psubband++; |
1227 | | |
1228 | 93.4M | wim = *p_sin_cos++; |
1229 | 93.4M | wre = *p_sin_cos++; |
1230 | | |
1231 | 93.4M | *psubband1_t-- = (WORD32)((ixheaac_sub64_sat(ixheaac_mult64(im, wre), |
1232 | 93.4M | ixheaac_mult64(re, wim))) >> |
1233 | 93.4M | 32); |
1234 | 93.4M | *psubband1_t-- = (WORD32)( |
1235 | 93.4M | (ixheaac_add64(ixheaac_mult64(re, wre), ixheaac_mult64(im, wim))) >> |
1236 | 93.4M | 32); |
1237 | | |
1238 | 93.4M | re = *psubband12--; |
1239 | 93.4M | im = *psubband2++; |
1240 | | |
1241 | 93.4M | *psubband1_t2-- = (WORD32)( |
1242 | 93.4M | (ixheaac_add64(ixheaac_mult64(re, wim), ixheaac_mult64(im, wre))) >> |
1243 | 93.4M | 32); |
1244 | 93.4M | *psubband1_t2-- = (WORD32)((ixheaac_sub64_sat(ixheaac_mult64(im, wim), |
1245 | 93.4M | ixheaac_mult64(re, wre))) >> |
1246 | 93.4M | 32); |
1247 | | |
1248 | 93.4M | re = *psubband++; |
1249 | 93.4M | im = *psubband1--; |
1250 | | |
1251 | 93.4M | wim = *p_sin_cos++; |
1252 | 93.4M | wre = *p_sin_cos++; |
1253 | | |
1254 | 93.4M | *psubband_t++ = (WORD32)( |
1255 | 93.4M | (ixheaac_add64(ixheaac_mult64(re, wre), ixheaac_mult64(im, wim))) >> |
1256 | 93.4M | 32); |
1257 | 93.4M | *psubband_t++ = (WORD32)((ixheaac_sub64_sat(ixheaac_mult64(im, wre), |
1258 | 93.4M | ixheaac_mult64(re, wim))) >> |
1259 | 93.4M | 32); |
1260 | | |
1261 | 93.4M | re = *psubband2++; |
1262 | 93.4M | im = *psubband12--; |
1263 | | |
1264 | 93.4M | *psubband_t2++ = (WORD32)((ixheaac_sub64_sat(ixheaac_mult64(im, wim), |
1265 | 93.4M | ixheaac_mult64(re, wre))) >> |
1266 | 93.4M | 32); |
1267 | 93.4M | *psubband_t2++ = (WORD32)( |
1268 | 93.4M | (ixheaac_add64(ixheaac_mult64(re, wim), ixheaac_mult64(im, wre))) >> |
1269 | 93.4M | 32); |
1270 | | |
1271 | 93.4M | re = *psubband1--; |
1272 | 93.4M | im = *psubband++; |
1273 | | |
1274 | 93.4M | wim = *p_sin_cos++; |
1275 | 93.4M | wre = *p_sin_cos++; |
1276 | | |
1277 | 93.4M | *psubband1_t-- = (WORD32)((ixheaac_sub64_sat(ixheaac_mult64(im, wre), |
1278 | 93.4M | ixheaac_mult64(re, wim))) >> |
1279 | 93.4M | 32); |
1280 | 93.4M | *psubband1_t-- = (WORD32)( |
1281 | 93.4M | (ixheaac_add64(ixheaac_mult64(re, wre), ixheaac_mult64(im, wim))) >> |
1282 | 93.4M | 32); |
1283 | | |
1284 | 93.4M | re = *psubband12--; |
1285 | 93.4M | im = *psubband2++; |
1286 | | |
1287 | 93.4M | *psubband1_t2-- = (WORD32)( |
1288 | 93.4M | (ixheaac_add64(ixheaac_mult64(re, wim), ixheaac_mult64(im, wre))) >> |
1289 | 93.4M | 32); |
1290 | 93.4M | *psubband1_t2-- = (WORD32)((ixheaac_sub64_sat(ixheaac_mult64(im, wim), |
1291 | 93.4M | ixheaac_mult64(re, wre))) >> |
1292 | 93.4M | 32); |
1293 | 93.4M | } |
1294 | | |
1295 | 18.7M | if (M == 32) { |
1296 | 6.10M | ixheaacd_esbr_radix4bfly(p_twiddle, subband_tmp, 1, 8); |
1297 | 6.10M | ixheaacd_esbr_radix4bfly(p_twiddle + 48, subband_tmp, 4, 2); |
1298 | 6.10M | ixheaacd_esbr_postradixcompute2(subband, subband_tmp, p_dig_rev_tbl, 32); |
1299 | | |
1300 | 6.10M | ixheaacd_esbr_radix4bfly(p_twiddle, &subband_tmp[64], 1, 8); |
1301 | 6.10M | ixheaacd_esbr_radix4bfly(p_twiddle + 48, &subband_tmp[64], 4, 2); |
1302 | 6.10M | ixheaacd_esbr_postradixcompute2(&subband[64], &subband_tmp[64], |
1303 | 6.10M | p_dig_rev_tbl, 32); |
1304 | | |
1305 | 6.10M | } |
1306 | | |
1307 | 12.6M | else if (M == 16) { |
1308 | 9.13M | ixheaacd_esbr_radix4bfly(p_twiddle, subband_tmp, 1, 4); |
1309 | 9.13M | ixheaacd_esbr_postradixcompute4(subband, subband_tmp, p_dig_rev_tbl, 16); |
1310 | | |
1311 | 9.13M | ixheaacd_esbr_radix4bfly(p_twiddle, &subband_tmp[64], 1, 4); |
1312 | 9.13M | ixheaacd_esbr_postradixcompute4(&subband[64], &subband_tmp[64], |
1313 | 9.13M | p_dig_rev_tbl, 16); |
1314 | | |
1315 | 9.13M | } |
1316 | | |
1317 | 3.52M | else if (M == 12) { |
1318 | 13.6M | for (z = 0; z < (qmf_bank->no_channels >> 1); z++) { |
1319 | 12.6M | temp[z] = subband_tmp[2 * z]; |
1320 | 12.6M | temp[12 + z] = subband_tmp[2 * z + 1]; |
1321 | 12.6M | } |
1322 | | |
1323 | 1.05M | ixheaacd_complex_fft_p3(temp, &temp[12], 12, -1, &scaleshift); |
1324 | | |
1325 | 13.6M | for (z = 0; z < (qmf_bank->no_channels >> 1); z++) { |
1326 | 12.6M | subband[2 * z] = temp[z]; |
1327 | 12.6M | subband[2 * z + 1] = temp[z + 12]; |
1328 | 12.6M | } |
1329 | 1.05M | scaleshift = 0; |
1330 | 13.6M | for (z = 0; z < (qmf_bank->no_channels >> 1); z++) { |
1331 | 12.6M | temp[z] = subband_tmp[64 + 2 * z]; |
1332 | 12.6M | temp[12 + z] = subband_tmp[64 + 2 * z + 1]; |
1333 | 12.6M | } |
1334 | | |
1335 | 1.05M | ixheaacd_complex_fft_p3(temp, &temp[12], 12, -1, &scaleshift); |
1336 | | |
1337 | 13.6M | for (z = 0; z < (qmf_bank->no_channels >> 1); z++) { |
1338 | 12.6M | subband[64 + 2 * z] = temp[z]; |
1339 | 12.6M | subband[64 + 2 * z + 1] = temp[z + 12]; |
1340 | 12.6M | } |
1341 | | |
1342 | 1.05M | } |
1343 | | |
1344 | 2.47M | else { |
1345 | 22.2M | for (z = 0; z < (qmf_bank->no_channels >> 1); z++) { |
1346 | 19.7M | temp[z] = subband_tmp[2 * z]; |
1347 | 19.7M | temp[8 + z] = subband_tmp[2 * z + 1]; |
1348 | 19.7M | } |
1349 | | |
1350 | 2.47M | (*ixheaacd_complex_fft_p2)(temp, &temp[8], 8, -1, &scaleshift); |
1351 | | |
1352 | 22.2M | for (z = 0; z < (qmf_bank->no_channels >> 1); z++) { |
1353 | 19.7M | subband[2 * z] = temp[z] << scaleshift; |
1354 | 19.7M | subband[2 * z + 1] = temp[z + 8] << scaleshift; |
1355 | 19.7M | } |
1356 | 2.47M | scaleshift = 0; |
1357 | 22.2M | for (z = 0; z < (qmf_bank->no_channels >> 1); z++) { |
1358 | 19.7M | temp[z] = subband_tmp[64 + 2 * z]; |
1359 | 19.7M | temp[8 + z] = subband_tmp[64 + 2 * z + 1]; |
1360 | 19.7M | } |
1361 | | |
1362 | 2.47M | (*ixheaacd_complex_fft_p2)(temp, &temp[8], 8, -1, &scaleshift); |
1363 | | |
1364 | 22.2M | for (z = 0; z < (qmf_bank->no_channels >> 1); z++) { |
1365 | 19.7M | subband[64 + 2 * z] = temp[z] << scaleshift; |
1366 | 19.7M | subband[64 + 2 * z + 1] = temp[8 + z] << scaleshift; |
1367 | 19.7M | } |
1368 | 2.47M | } |
1369 | | |
1370 | 18.7M | psubband = &subband[0]; |
1371 | 18.7M | psubband1 = &subband[2 * M - 1]; |
1372 | | |
1373 | 18.7M | re = *psubband1; |
1374 | | |
1375 | 18.7M | *psubband = *psubband >> 1; |
1376 | 18.7M | psubband++; |
1377 | 18.7M | *psubband1 = ixheaac_negate32_sat(*psubband >> 1); |
1378 | 18.7M | psubband1--; |
1379 | | |
1380 | 18.7M | p_sin = qmf_bank->esbr_alt_sin_twiddle; |
1381 | | |
1382 | 18.7M | wim = *p_sin++; |
1383 | 18.7M | wre = *p_sin++; |
1384 | | |
1385 | 18.7M | im = *psubband1; |
1386 | | |
1387 | 18.7M | *psubband1-- = (WORD32)( |
1388 | 18.7M | (ixheaac_add64(ixheaac_mult64(re, wre), ixheaac_mult64(im, wim))) >> |
1389 | 18.7M | 32); |
1390 | 18.7M | *psubband++ = (WORD32)((ixheaac_sub64_sat(ixheaac_mult64(im, wre), |
1391 | 18.7M | ixheaac_mult64(re, wim))) >> |
1392 | 18.7M | 32); |
1393 | | |
1394 | 18.7M | psubband2 = &subband[64]; |
1395 | 18.7M | psubband12 = &subband[2 * M - 1 + 64]; |
1396 | | |
1397 | 18.7M | re = *psubband12; |
1398 | | |
1399 | 18.7M | *psubband12-- = ixheaac_negate32_sat(*psubband2 >> 1); |
1400 | | |
1401 | 18.7M | *psubband2 = psubband2[1] >> 1; |
1402 | | |
1403 | 18.7M | psubband2++; |
1404 | | |
1405 | 18.7M | im = *psubband12; |
1406 | | |
1407 | 18.7M | *psubband2++ = ixheaac_negate32_sat((WORD32)( |
1408 | 18.7M | (ixheaac_add64(ixheaac_mult64(re, wre), ixheaac_mult64(im, wim))) >> |
1409 | 18.7M | 32)); |
1410 | 18.7M | *psubband12-- = (WORD32)((ixheaac_sub64_sat(ixheaac_mult64(re, wim), |
1411 | 18.7M | ixheaac_mult64(im, wre))) >> |
1412 | 18.7M | 32); |
1413 | | |
1414 | 186M | for (i = (M_2 - 2); i >= 0; i--) { |
1415 | 168M | im = psubband[0]; |
1416 | | |
1417 | 168M | re = psubband[1]; |
1418 | | |
1419 | 168M | re2 = *psubband1; |
1420 | | |
1421 | 168M | *psubband++ = (WORD32)( |
1422 | 168M | (ixheaac_add64(ixheaac_mult64(re, wim), ixheaac_mult64(im, wre))) >> |
1423 | 168M | 32); |
1424 | 168M | *psubband1-- = (WORD32)((ixheaac_sub64_sat(ixheaac_mult64(im, wim), |
1425 | 168M | ixheaac_mult64(re, wre))) >> |
1426 | 168M | 32); |
1427 | | |
1428 | 168M | im = psubband2[0]; |
1429 | | |
1430 | 168M | re = psubband2[1]; |
1431 | | |
1432 | 168M | re3 = *psubband12; |
1433 | | |
1434 | 168M | *psubband12-- = ixheaac_negate32_sat((WORD32)( |
1435 | 168M | (ixheaac_add64(ixheaac_mult64(re, wim), ixheaac_mult64(im, wre))) >> |
1436 | 168M | 32)); |
1437 | 168M | *psubband2++ = (WORD32)((ixheaac_sub64_sat(ixheaac_mult64(re, wre), |
1438 | 168M | ixheaac_mult64(im, wim))) >> |
1439 | 168M | 32); |
1440 | | |
1441 | 168M | wim = *p_sin++; |
1442 | 168M | wre = *p_sin++; |
1443 | 168M | im = psubband1[0]; |
1444 | | |
1445 | 168M | *psubband1-- = (WORD32)( |
1446 | 168M | (ixheaac_add64(ixheaac_mult64(re2, wre), ixheaac_mult64(im, wim))) >> |
1447 | 168M | 32); |
1448 | 168M | *psubband++ = (WORD32)((ixheaac_sub64_sat(ixheaac_mult64(im, wre), |
1449 | 168M | ixheaac_mult64(re2, wim))) >> |
1450 | 168M | 32); |
1451 | | |
1452 | 168M | im = psubband12[0]; |
1453 | | |
1454 | 168M | *psubband2++ = ixheaac_negate32_sat((WORD32)( |
1455 | 168M | (ixheaac_add64(ixheaac_mult64(re3, wre), ixheaac_mult64(im, wim))) >> |
1456 | 168M | 32)); |
1457 | 168M | *psubband12-- = (WORD32)((ixheaac_sub64_sat(ixheaac_mult64(re3, wim), |
1458 | 168M | ixheaac_mult64(im, wre))) >> |
1459 | 168M | 32); |
1460 | 168M | } |
1461 | 18.7M | } |
1462 | | |
1463 | | VOID ixheaacd_esbr_fwd_modulation( |
1464 | | const WORD32 *time_sample_buf, WORD32 *real_subband, WORD32 *imag_subband, |
1465 | | ia_sbr_qmf_filter_bank_struct *qmf_bank, |
1466 | 12.6M | ia_qmf_dec_tables_struct *qmf_dec_tables_ptr) { |
1467 | 12.6M | WORD32 i; |
1468 | 12.6M | const WORD32 *time_sample_buf1 = |
1469 | 12.6M | &time_sample_buf[2 * qmf_bank->no_channels - 1]; |
1470 | 12.6M | WORD32 temp1, temp2; |
1471 | 12.6M | WORD32 *t_real_subband = real_subband; |
1472 | 12.6M | WORD32 *t_imag_subband = imag_subband; |
1473 | 12.6M | const WORD32 *tcos; |
1474 | | |
1475 | 369M | for (i = qmf_bank->no_channels - 1; i >= 0; i--) { |
1476 | 356M | temp1 = ixheaac_shr32(*time_sample_buf++, HQ_SHIFT_64); |
1477 | 356M | temp2 = ixheaac_shr32(*time_sample_buf1--, HQ_SHIFT_64); |
1478 | | |
1479 | 356M | *t_real_subband++ = ixheaac_sub32_sat(temp1, temp2); |
1480 | | |
1481 | 356M | *t_imag_subband++ = ixheaac_add32_sat(temp1, temp2); |
1482 | 356M | } |
1483 | | |
1484 | 12.6M | ixheaacd_esbr_cos_sin_mod(real_subband, qmf_bank, |
1485 | 12.6M | qmf_dec_tables_ptr->esbr_w_16, |
1486 | 12.6M | qmf_dec_tables_ptr->dig_rev_table4_16); |
1487 | | |
1488 | 12.6M | tcos = qmf_bank->esbr_t_cos; |
1489 | | |
1490 | 369M | for (i = (qmf_bank->usb - qmf_bank->lsb - 1); i >= 0; i--) { |
1491 | 356M | WORD32 cosh, sinh; |
1492 | 356M | WORD32 re, im; |
1493 | | |
1494 | 356M | re = *real_subband; |
1495 | 356M | im = *imag_subband; |
1496 | 356M | cosh = *tcos++; |
1497 | 356M | sinh = *tcos++; |
1498 | 356M | *real_subband++ = (WORD32)((ixheaac_add64(ixheaac_mult64(re, cosh), |
1499 | 356M | ixheaac_mult64(im, sinh))) >> |
1500 | 356M | 31); |
1501 | 356M | *imag_subband++ = |
1502 | 356M | (WORD32)((ixheaac_sub64_sat(ixheaac_mult64(im, cosh), |
1503 | 356M | ixheaac_mult64(re, sinh))) >> |
1504 | 356M | 31); |
1505 | 356M | } |
1506 | 12.6M | } |
1507 | | |
1508 | | VOID ixheaacd_sbr_qmfsyn64_winadd(WORD16 *tmp1, WORD16 *tmp2, WORD16 *inp1, |
1509 | | WORD16 *sample_buffer, FLAG shift, |
1510 | 292k | WORD32 ch_fac) { |
1511 | 292k | WORD32 k; |
1512 | 292k | WORD32 rounding_fac = 0x8000; |
1513 | 292k | rounding_fac = rounding_fac >> shift; |
1514 | | |
1515 | 19.0M | for (k = 0; k < 64; k++) { |
1516 | 18.7M | WORD32 syn_out = rounding_fac; |
1517 | | |
1518 | 18.7M | syn_out = ixheaac_add32_sat( |
1519 | 18.7M | syn_out, ixheaac_mult16x16in32(tmp1[0 + k], inp1[k + 0])); |
1520 | 18.7M | syn_out = ixheaac_add32_sat( |
1521 | 18.7M | syn_out, ixheaac_mult16x16in32(tmp1[256 + k], inp1[k + 128])); |
1522 | 18.7M | syn_out = ixheaac_add32_sat( |
1523 | 18.7M | syn_out, ixheaac_mult16x16in32(tmp1[512 + k], inp1[k + 256])); |
1524 | 18.7M | syn_out = ixheaac_add32_sat( |
1525 | 18.7M | syn_out, ixheaac_mult16x16in32(tmp1[768 + k], inp1[k + 384])); |
1526 | 18.7M | syn_out = ixheaac_add32_sat( |
1527 | 18.7M | syn_out, ixheaac_mult16x16in32(tmp1[1024 + k], inp1[k + 512])); |
1528 | | |
1529 | 18.7M | syn_out = ixheaac_add32_sat( |
1530 | 18.7M | syn_out, ixheaac_mult16x16in32(tmp2[128 + k], inp1[k + 64])); |
1531 | 18.7M | syn_out = ixheaac_add32_sat( |
1532 | 18.7M | syn_out, ixheaac_mult16x16in32(tmp2[384 + k], inp1[k + 192])); |
1533 | 18.7M | syn_out = ixheaac_add32_sat( |
1534 | 18.7M | syn_out, ixheaac_mult16x16in32(tmp2[640 + k], inp1[k + 320])); |
1535 | 18.7M | syn_out = ixheaac_add32_sat( |
1536 | 18.7M | syn_out, ixheaac_mult16x16in32(tmp2[896 + k], inp1[k + 448])); |
1537 | 18.7M | syn_out = ixheaac_add32_sat( |
1538 | 18.7M | syn_out, ixheaac_mult16x16in32(tmp2[1152 + k], inp1[k + 576])); |
1539 | | |
1540 | 18.7M | sample_buffer[ch_fac * k] = (ixheaac_shl32_sat(syn_out, shift) >> 16); |
1541 | 18.7M | } |
1542 | 292k | } |
1543 | | |
1544 | | VOID ixheaacd_esbr_qmfsyn64_winadd(WORD32 *tmp1, WORD32 *tmp2, WORD32 *inp1, |
1545 | 6.10M | WORD32 *sample_buffer, WORD32 ch_fac) { |
1546 | 6.10M | WORD32 k; |
1547 | | |
1548 | 397M | for (k = 0; k < 64; k++) { |
1549 | 390M | WORD64 syn_out = 0; |
1550 | | |
1551 | 390M | syn_out = |
1552 | 390M | ixheaac_add64(syn_out, ixheaac_mult64(tmp1[0 + k], inp1[k + 0])); |
1553 | 390M | syn_out = |
1554 | 390M | ixheaac_add64(syn_out, ixheaac_mult64(tmp1[256 + k], inp1[k + 128])); |
1555 | 390M | syn_out = |
1556 | 390M | ixheaac_add64(syn_out, ixheaac_mult64(tmp1[512 + k], inp1[k + 256])); |
1557 | 390M | syn_out = |
1558 | 390M | ixheaac_add64(syn_out, ixheaac_mult64(tmp1[768 + k], inp1[k + 384])); |
1559 | 390M | syn_out = |
1560 | 390M | ixheaac_add64(syn_out, ixheaac_mult64(tmp1[1024 + k], inp1[k + 512])); |
1561 | | |
1562 | 390M | syn_out = |
1563 | 390M | ixheaac_add64(syn_out, ixheaac_mult64(tmp2[128 + k], inp1[k + 64])); |
1564 | 390M | syn_out = |
1565 | 390M | ixheaac_add64(syn_out, ixheaac_mult64(tmp2[384 + k], inp1[k + 192])); |
1566 | 390M | syn_out = |
1567 | 390M | ixheaac_add64(syn_out, ixheaac_mult64(tmp2[640 + k], inp1[k + 320])); |
1568 | 390M | syn_out = |
1569 | 390M | ixheaac_add64(syn_out, ixheaac_mult64(tmp2[896 + k], inp1[k + 448])); |
1570 | 390M | syn_out = |
1571 | 390M | ixheaac_add64(syn_out, ixheaac_mult64(tmp2[1152 + k], inp1[k + 576])); |
1572 | | |
1573 | 390M | sample_buffer[ch_fac * k] = (WORD32)(syn_out >> 31); |
1574 | 390M | } |
1575 | 6.10M | } |
1576 | | |
1577 | | VOID ixheaacd_esbr_qmfsyn32_winadd(WORD32 *tmp1, WORD32 *tmp2, WORD32 *inp1, |
1578 | 9.88k | WORD32 *sample_buffer, WORD32 ch_fac) { |
1579 | 9.88k | WORD32 k; |
1580 | | |
1581 | 326k | for (k = 0; k < 32; k++) { |
1582 | 316k | WORD64 syn_out = 0; |
1583 | | |
1584 | 316k | syn_out = |
1585 | 316k | ixheaac_add64(syn_out, ixheaac_mult64(tmp1[0 + k], inp1[2 * (k + 0)])); |
1586 | 316k | syn_out = |
1587 | 316k | ixheaac_add64(syn_out, ixheaac_mult64(tmp1[128 + k], inp1[2 * (k + 64)])); |
1588 | 316k | syn_out = |
1589 | 316k | ixheaac_add64(syn_out, ixheaac_mult64(tmp1[256 + k], inp1[2 * (k + 128)])); |
1590 | 316k | syn_out = |
1591 | 316k | ixheaac_add64(syn_out, ixheaac_mult64(tmp1[384 + k], inp1[2 * (k + 192)])); |
1592 | 316k | syn_out = |
1593 | 316k | ixheaac_add64(syn_out, ixheaac_mult64(tmp1[512 + k], inp1[2 * (k + 256)])); |
1594 | | |
1595 | 316k | syn_out = |
1596 | 316k | ixheaac_add64(syn_out, ixheaac_mult64(tmp2[64 + k], inp1[2 * (k + 32)])); |
1597 | 316k | syn_out = |
1598 | 316k | ixheaac_add64(syn_out, ixheaac_mult64(tmp2[192 + k], inp1[2 * (k + 96)])); |
1599 | 316k | syn_out = |
1600 | 316k | ixheaac_add64(syn_out, ixheaac_mult64(tmp2[320 + k], inp1[2 * (k + 160)])); |
1601 | 316k | syn_out = |
1602 | 316k | ixheaac_add64(syn_out, ixheaac_mult64(tmp2[448 + k], inp1[2 * (k + 224)])); |
1603 | 316k | syn_out = |
1604 | 316k | ixheaac_add64(syn_out, ixheaac_mult64(tmp2[576 + k], inp1[2 * (k + 288)])); |
1605 | | |
1606 | 316k | sample_buffer[ch_fac * k] = (WORD32)(syn_out >> 31); |
1607 | 316k | } |
1608 | 9.88k | } |
1609 | | |
1610 | | VOID ixheaacd_shiftrountine(WORD32 *qmf_real, WORD32 *qmf_imag, WORD32 len, |
1611 | 0 | WORD32 common_shift) { |
1612 | 0 | WORD32 treal, timag; |
1613 | 0 | WORD32 j; |
1614 | |
|
1615 | 0 | if (common_shift < 0) { |
1616 | 0 | WORD32 cshift = -common_shift; |
1617 | 0 | cshift = ixheaac_min32(cshift, 31); |
1618 | 0 | for (j = len - 1; j >= 0; j--) { |
1619 | 0 | treal = *qmf_real; |
1620 | 0 | timag = *qmf_imag; |
1621 | |
|
1622 | 0 | treal = (ixheaac_shr32(treal, cshift)); |
1623 | 0 | timag = (ixheaac_shr32(timag, cshift)); |
1624 | |
|
1625 | 0 | *qmf_real++ = treal; |
1626 | 0 | *qmf_imag++ = timag; |
1627 | 0 | } |
1628 | 0 | } else { |
1629 | 0 | for (j = len - 1; j >= 0; j--) { |
1630 | 0 | treal = (ixheaac_shl32_sat(*qmf_real, common_shift)); |
1631 | 0 | timag = (ixheaac_shl32_sat(*qmf_imag, common_shift)); |
1632 | 0 | *qmf_real++ = treal; |
1633 | 0 | *qmf_imag++ = timag; |
1634 | 0 | } |
1635 | 0 | } |
1636 | 0 | } |
1637 | | |
1638 | | VOID ixheaacd_shiftrountine_with_rnd(WORD32 *qmf_real, WORD32 *qmf_imag, |
1639 | | WORD16 *filter_states, WORD32 len, |
1640 | 0 | WORD32 shift) { |
1641 | 0 | WORD16 *filter_states_rev = filter_states + len; |
1642 | 0 | WORD32 treal, timag; |
1643 | 0 | WORD32 j; |
1644 | |
|
1645 | 0 | for (j = (len - 1); j >= 0; j -= 2) { |
1646 | 0 | WORD32 r1, r2, i1, i2; |
1647 | 0 | i2 = qmf_imag[j]; |
1648 | 0 | r2 = qmf_real[j]; |
1649 | |
|
1650 | 0 | r1 = *qmf_real++; |
1651 | 0 | i1 = *qmf_imag++; |
1652 | |
|
1653 | 0 | timag = ixheaac_add32_sat(i1, r1); |
1654 | 0 | timag = (ixheaac_shl32_sat(timag, shift)); |
1655 | 0 | filter_states_rev[j] = ixheaac_round16(timag); |
1656 | |
|
1657 | 0 | treal = ixheaac_sub32_sat(i2, r2); |
1658 | 0 | treal = (ixheaac_shl32_sat(treal, shift)); |
1659 | 0 | filter_states[j] = ixheaac_round16(treal); |
1660 | |
|
1661 | 0 | treal = ixheaac_sub32_sat(i1, r1); |
1662 | 0 | treal = (ixheaac_shl32_sat(treal, shift)); |
1663 | 0 | *filter_states++ = ixheaac_round16(treal); |
1664 | |
|
1665 | 0 | timag = ixheaac_add32_sat(i2, r2); |
1666 | 0 | ; |
1667 | 0 | timag = (ixheaac_shl32_sat(timag, shift)); |
1668 | 0 | *filter_states_rev++ = ixheaac_round16(timag); |
1669 | 0 | } |
1670 | 0 | } |
1671 | | |
1672 | | VOID ixheaacd_shiftrountine_with_rnd_eld(WORD32 *qmf_real, WORD32 *qmf_imag, |
1673 | | WORD16 *filter_states, WORD32 len, |
1674 | 874k | WORD32 shift) { |
1675 | 874k | WORD16 *filter_states_rev = filter_states + len; |
1676 | 874k | WORD32 treal, timag; |
1677 | 874k | WORD32 j; |
1678 | | |
1679 | 19.5M | for (j = (len - 1); j >= 0; j -= 2) { |
1680 | 18.6M | WORD32 r1, r2, i1, i2; |
1681 | 18.6M | i2 = qmf_imag[j]; |
1682 | 18.6M | r2 = qmf_real[j]; |
1683 | 18.6M | r1 = *qmf_real++; |
1684 | 18.6M | i1 = *qmf_imag++; |
1685 | | |
1686 | 18.6M | timag = ixheaac_negate32_sat(ixheaac_add32_sat(i1, r1)); |
1687 | 18.6M | timag = (ixheaac_shl32_sat(timag, shift)); |
1688 | 18.6M | filter_states_rev[j] = ixheaac_round16(timag); |
1689 | | |
1690 | 18.6M | treal = ixheaac_sub32_sat(r2, i2); |
1691 | 18.6M | treal = (ixheaac_shl32_sat(treal, shift)); |
1692 | 18.6M | filter_states[j] = ixheaac_round16(treal); |
1693 | | |
1694 | 18.6M | treal = ixheaac_sub32_sat(r1, i1); |
1695 | 18.6M | treal = (ixheaac_shl32_sat(treal, shift)); |
1696 | 18.6M | *filter_states++ = ixheaac_round16(treal); |
1697 | | |
1698 | 18.6M | timag = ixheaac_negate32_sat(ixheaac_add32_sat(i2, r2)); |
1699 | 18.6M | timag = (ixheaac_shl32_sat(timag, shift)); |
1700 | 18.6M | *filter_states_rev++ = ixheaac_round16(timag); |
1701 | 18.6M | } |
1702 | 874k | } |
1703 | | |
1704 | | VOID ixheaacd_shiftrountine_with_rnd_hq(WORD32 *qmf_real, WORD32 *qmf_imag, |
1705 | | WORD32 *filter_states, WORD32 len, |
1706 | 6.11M | WORD32 shift) { |
1707 | 6.11M | WORD32 *filter_states_rev = filter_states + len; |
1708 | 6.11M | WORD32 treal, timag; |
1709 | 6.11M | WORD32 j; |
1710 | | |
1711 | 201M | for (j = (len - 1); j >= 0; j -= 2) { |
1712 | 195M | WORD32 r1, r2, i1, i2; |
1713 | 195M | i2 = qmf_imag[j]; |
1714 | 195M | r2 = qmf_real[j]; |
1715 | 195M | r1 = *qmf_real++; |
1716 | 195M | i1 = *qmf_imag++; |
1717 | | |
1718 | 195M | timag = ixheaac_add32_sat(i1, r1); |
1719 | 195M | timag = (ixheaac_shl32_sat(timag, shift)); |
1720 | 195M | filter_states_rev[j] = timag; |
1721 | | |
1722 | 195M | treal = ixheaac_sub32_sat(i2, r2); |
1723 | 195M | treal = (ixheaac_shl32_sat(treal, shift)); |
1724 | 195M | filter_states[j] = treal; |
1725 | | |
1726 | 195M | treal = ixheaac_sub32_sat(i1, r1); |
1727 | 195M | treal = (ixheaac_shl32_sat(treal, shift)); |
1728 | 195M | *filter_states++ = treal; |
1729 | | |
1730 | 195M | timag = ixheaac_add32_sat(i2, r2); |
1731 | 195M | timag = (ixheaac_shl32_sat(timag, shift)); |
1732 | 195M | *filter_states_rev++ = timag; |
1733 | 195M | } |
1734 | 6.11M | } |
1735 | | |
1736 | | VOID ixheaacd_radix4bfly(const WORD16 *w, WORD32 *x, WORD32 index1, |
1737 | 4.42M | WORD32 index) { |
1738 | 4.42M | int i; |
1739 | 4.42M | WORD32 l1, l2, h2, fft_jmp; |
1740 | 4.42M | WORD32 xt0_0, yt0_0, xt1_0, yt1_0, xt2_0, yt2_0; |
1741 | 4.42M | WORD32 xh0_0, xh1_0, xh20_0, xh21_0, xl0_0, xl1_0, xl20_0, xl21_0; |
1742 | 4.42M | WORD32 x_0, x_1, x_l1_0, x_l1_1, x_l2_0, x_l2_1; |
1743 | 4.42M | WORD32 x_h2_0, x_h2_1; |
1744 | 4.42M | WORD16 si10, si20, si30, co10, co20, co30; |
1745 | | |
1746 | 4.42M | WORD32 mul_1, mul_2, mul_3, mul_4, mul_5, mul_6; |
1747 | 4.42M | WORD32 mul_7, mul_8, mul_9, mul_10, mul_11, mul_12; |
1748 | 4.42M | const WORD16 *w_ptr = w; |
1749 | 4.42M | WORD32 i1; |
1750 | | |
1751 | 4.42M | h2 = index << 1; |
1752 | 4.42M | l1 = index << 2; |
1753 | 4.42M | l2 = (index << 2) + (index << 1); |
1754 | | |
1755 | 4.42M | fft_jmp = 6 * (index); |
1756 | | |
1757 | 11.6M | for (i1 = 0; i1 < index1; i1++) { |
1758 | 32.2M | for (i = 0; i < index; i++) { |
1759 | 25.0M | si10 = (*w_ptr++); |
1760 | 25.0M | co10 = (*w_ptr++); |
1761 | 25.0M | si20 = (*w_ptr++); |
1762 | 25.0M | co20 = (*w_ptr++); |
1763 | 25.0M | si30 = (*w_ptr++); |
1764 | 25.0M | co30 = (*w_ptr++); |
1765 | | |
1766 | 25.0M | x_0 = x[0]; |
1767 | 25.0M | x_h2_0 = x[h2]; |
1768 | 25.0M | x_l1_0 = x[l1]; |
1769 | 25.0M | x_l2_0 = x[l2]; |
1770 | | |
1771 | 25.0M | xh0_0 = ixheaac_add32_sat(x_0, x_l1_0); |
1772 | 25.0M | xl0_0 = ixheaac_sub32_sat(x_0, x_l1_0); |
1773 | | |
1774 | 25.0M | xh20_0 = ixheaac_add32_sat(x_h2_0, x_l2_0); |
1775 | 25.0M | xl20_0 = ixheaac_sub32_sat(x_h2_0, x_l2_0); |
1776 | | |
1777 | 25.0M | x[0] = ixheaac_add32_sat(xh0_0, xh20_0); |
1778 | 25.0M | xt0_0 = ixheaac_sub32_sat(xh0_0, xh20_0); |
1779 | | |
1780 | 25.0M | x_1 = x[1]; |
1781 | 25.0M | x_h2_1 = x[h2 + 1]; |
1782 | 25.0M | x_l1_1 = x[l1 + 1]; |
1783 | 25.0M | x_l2_1 = x[l2 + 1]; |
1784 | | |
1785 | 25.0M | xh1_0 = ixheaac_add32_sat(x_1, x_l1_1); |
1786 | 25.0M | xl1_0 = ixheaac_sub32_sat(x_1, x_l1_1); |
1787 | | |
1788 | 25.0M | xh21_0 = ixheaac_add32_sat(x_h2_1, x_l2_1); |
1789 | 25.0M | xl21_0 = ixheaac_sub32_sat(x_h2_1, x_l2_1); |
1790 | | |
1791 | 25.0M | x[1] = ixheaac_add32_sat(xh1_0, xh21_0); |
1792 | 25.0M | yt0_0 = ixheaac_sub32_sat(xh1_0, xh21_0); |
1793 | | |
1794 | 25.0M | xt1_0 = ixheaac_add32_sat(xl0_0, xl21_0); |
1795 | 25.0M | xt2_0 = ixheaac_sub32_sat(xl0_0, xl21_0); |
1796 | | |
1797 | 25.0M | yt2_0 = ixheaac_add32_sat(xl1_0, xl20_0); |
1798 | 25.0M | yt1_0 = ixheaac_sub32_sat(xl1_0, xl20_0); |
1799 | | |
1800 | 25.0M | mul_11 = ixheaac_mult32x16in32(xt2_0, co30); |
1801 | 25.0M | mul_3 = ixheaac_mult32x16in32(yt2_0, si30); |
1802 | 25.0M | x[l2] = (mul_3 + mul_11) << RADIXSHIFT; |
1803 | | |
1804 | 25.0M | mul_5 = ixheaac_mult32x16in32(xt2_0, si30); |
1805 | 25.0M | mul_9 = ixheaac_mult32x16in32(yt2_0, co30); |
1806 | 25.0M | x[l2 + 1] = (mul_9 - mul_5) << RADIXSHIFT; |
1807 | | |
1808 | 25.0M | mul_12 = ixheaac_mult32x16in32(xt0_0, co20); |
1809 | 25.0M | mul_2 = ixheaac_mult32x16in32(yt0_0, si20); |
1810 | 25.0M | x[l1] = (mul_2 + mul_12) << RADIXSHIFT; |
1811 | | |
1812 | 25.0M | mul_6 = ixheaac_mult32x16in32(xt0_0, si20); |
1813 | 25.0M | mul_8 = ixheaac_mult32x16in32(yt0_0, co20); |
1814 | 25.0M | x[l1 + 1] = (mul_8 - mul_6) << RADIXSHIFT; |
1815 | | |
1816 | 25.0M | mul_4 = ixheaac_mult32x16in32(xt1_0, co10); |
1817 | 25.0M | mul_1 = ixheaac_mult32x16in32(yt1_0, si10); |
1818 | 25.0M | x[h2] = (mul_1 + mul_4) << RADIXSHIFT; |
1819 | | |
1820 | 25.0M | mul_10 = ixheaac_mult32x16in32(xt1_0, si10); |
1821 | 25.0M | mul_7 = ixheaac_mult32x16in32(yt1_0, co10); |
1822 | 25.0M | x[h2 + 1] = (mul_7 - mul_10) << RADIXSHIFT; |
1823 | | |
1824 | 25.0M | x += 2; |
1825 | 25.0M | } |
1826 | 7.18M | x += fft_jmp; |
1827 | 7.18M | w_ptr = w_ptr - fft_jmp; |
1828 | 7.18M | } |
1829 | 4.42M | } |
1830 | | |
1831 | | VOID ixheaacd_postradixcompute4(WORD32 *ptr_y, WORD32 *ptr_x, |
1832 | 2.57M | const WORD32 *p_dig_rev_tbl, WORD32 npoints) { |
1833 | 2.57M | WORD32 i, k; |
1834 | 2.57M | WORD32 h2; |
1835 | 2.57M | WORD32 xh0_0, xh1_0, xl0_0, xl1_0; |
1836 | 2.57M | WORD32 xh0_1, xh1_1, xl0_1, xl1_1; |
1837 | 2.57M | WORD32 x_0, x_1, x_2, x_3; |
1838 | 2.57M | WORD32 xh0_2, xh1_2, xl0_2, xl1_2, xh0_3, xh1_3, xl0_3, xl1_3; |
1839 | 2.57M | WORD32 x_4, x_5, x_6, x_7; |
1840 | 2.57M | WORD32 x_8, x_9, x_a, x_b, x_c, x_d, x_e, x_f; |
1841 | 2.57M | WORD32 n00, n10, n20, n30, n01, n11, n21, n31; |
1842 | 2.57M | WORD32 n02, n12, n22, n32, n03, n13, n23, n33; |
1843 | 2.57M | WORD32 *x2, *x0; |
1844 | 2.57M | WORD32 *y0, *y1, *y2, *y3; |
1845 | | |
1846 | 2.57M | y0 = ptr_y; |
1847 | 2.57M | y2 = ptr_y + (WORD32)npoints; |
1848 | 2.57M | x0 = ptr_x; |
1849 | 2.57M | x2 = ptr_x + (WORD32)(npoints >> 1); |
1850 | | |
1851 | 2.57M | y1 = y0 + (WORD32)(npoints >> 1); |
1852 | 2.57M | y3 = y2 + (WORD32)(npoints >> 1); |
1853 | | |
1854 | 7.73M | for (k = 0; k < 2; k++) { |
1855 | 10.3M | for (i = 0; i<npoints>> 1; i += 8) { |
1856 | 5.15M | h2 = *p_dig_rev_tbl++ >> 2; |
1857 | 5.15M | x_0 = *x0++; |
1858 | 5.15M | x_1 = *x0++; |
1859 | 5.15M | x_2 = *x0++; |
1860 | 5.15M | x_3 = *x0++; |
1861 | 5.15M | x_4 = *x0++; |
1862 | 5.15M | x_5 = *x0++; |
1863 | 5.15M | x_6 = *x0++; |
1864 | 5.15M | x_7 = *x0++; |
1865 | | |
1866 | 5.15M | xh0_0 = ixheaac_add32_sat(x_0, x_4); |
1867 | 5.15M | xh1_0 = ixheaac_add32_sat(x_1, x_5); |
1868 | 5.15M | xl0_0 = ixheaac_sub32_sat(x_0, x_4); |
1869 | 5.15M | xl1_0 = ixheaac_sub32_sat(x_1, x_5); |
1870 | 5.15M | xh0_1 = ixheaac_add32_sat(x_2, x_6); |
1871 | 5.15M | xh1_1 = ixheaac_add32_sat(x_3, x_7); |
1872 | 5.15M | xl0_1 = ixheaac_sub32_sat(x_2, x_6); |
1873 | 5.15M | xl1_1 = ixheaac_sub32_sat(x_3, x_7); |
1874 | | |
1875 | 5.15M | n00 = ixheaac_add32_sat(xh0_0, xh0_1); |
1876 | 5.15M | n01 = ixheaac_add32_sat(xh1_0, xh1_1); |
1877 | 5.15M | n10 = ixheaac_add32_sat(xl0_0, xl1_1); |
1878 | 5.15M | n11 = ixheaac_sub32_sat(xl1_0, xl0_1); |
1879 | 5.15M | n20 = ixheaac_sub32_sat(xh0_0, xh0_1); |
1880 | 5.15M | n21 = ixheaac_sub32_sat(xh1_0, xh1_1); |
1881 | 5.15M | n30 = ixheaac_sub32_sat(xl0_0, xl1_1); |
1882 | 5.15M | n31 = ixheaac_add32_sat(xl1_0, xl0_1); |
1883 | | |
1884 | 5.15M | y0[h2] = n00; |
1885 | 5.15M | y0[h2 + 1] = n01; |
1886 | 5.15M | y1[h2] = n10; |
1887 | 5.15M | y1[h2 + 1] = n11; |
1888 | 5.15M | y2[h2] = n20; |
1889 | 5.15M | y2[h2 + 1] = n21; |
1890 | 5.15M | y3[h2] = n30; |
1891 | 5.15M | y3[h2 + 1] = n31; |
1892 | | |
1893 | 5.15M | x_8 = *x2++; |
1894 | 5.15M | x_9 = *x2++; |
1895 | 5.15M | x_a = *x2++; |
1896 | 5.15M | x_b = *x2++; |
1897 | 5.15M | x_c = *x2++; |
1898 | 5.15M | x_d = *x2++; |
1899 | 5.15M | x_e = *x2++; |
1900 | 5.15M | x_f = *x2++; |
1901 | | |
1902 | 5.15M | xh0_2 = ixheaac_add32_sat(x_8, x_c); |
1903 | 5.15M | xh1_2 = ixheaac_add32_sat(x_9, x_d); |
1904 | 5.15M | xl0_2 = ixheaac_sub32_sat(x_8, x_c); |
1905 | 5.15M | xl1_2 = ixheaac_sub32_sat(x_9, x_d); |
1906 | 5.15M | xh0_3 = ixheaac_add32_sat(x_a, x_e); |
1907 | 5.15M | xh1_3 = ixheaac_add32_sat(x_b, x_f); |
1908 | 5.15M | xl0_3 = ixheaac_sub32_sat(x_a, x_e); |
1909 | 5.15M | xl1_3 = ixheaac_sub32_sat(x_b, x_f); |
1910 | | |
1911 | 5.15M | n02 = ixheaac_add32_sat(xh0_2, xh0_3); |
1912 | 5.15M | n03 = ixheaac_add32_sat(xh1_2, xh1_3); |
1913 | 5.15M | n12 = ixheaac_add32_sat(xl0_2, xl1_3); |
1914 | 5.15M | n13 = ixheaac_sub32_sat(xl1_2, xl0_3); |
1915 | 5.15M | n22 = ixheaac_sub32_sat(xh0_2, xh0_3); |
1916 | 5.15M | n23 = ixheaac_sub32_sat(xh1_2, xh1_3); |
1917 | 5.15M | n32 = ixheaac_sub32_sat(xl0_2, xl1_3); |
1918 | 5.15M | n33 = ixheaac_add32_sat(xl1_2, xl0_3); |
1919 | | |
1920 | 5.15M | y0[h2 + 2] = n02; |
1921 | 5.15M | y0[h2 + 3] = n03; |
1922 | 5.15M | y1[h2 + 2] = n12; |
1923 | 5.15M | y1[h2 + 3] = n13; |
1924 | 5.15M | y2[h2 + 2] = n22; |
1925 | 5.15M | y2[h2 + 3] = n23; |
1926 | 5.15M | y3[h2 + 2] = n32; |
1927 | 5.15M | y3[h2 + 3] = n33; |
1928 | 5.15M | } |
1929 | 5.15M | x0 += (WORD32)npoints >> 1; |
1930 | 5.15M | x2 += (WORD32)npoints >> 1; |
1931 | 5.15M | } |
1932 | 2.57M | } |
1933 | | |
1934 | | VOID ixheaacd_postradixcompute2(WORD32 *ptr_y, WORD32 *ptr_x, |
1935 | 921k | const WORD32 *pdig_rev_tbl, WORD32 npoints) { |
1936 | 921k | WORD32 i, k; |
1937 | 921k | WORD32 h2; |
1938 | 921k | WORD32 x_0, x_1, x_2, x_3; |
1939 | 921k | WORD32 x_4, x_5, x_6, x_7; |
1940 | 921k | WORD32 x_8, x_9, x_a, x_b, x_c, x_d, x_e, x_f; |
1941 | 921k | WORD32 n00, n10, n20, n30, n01, n11, n21, n31; |
1942 | 921k | WORD32 n02, n12, n22, n32, n03, n13, n23, n33; |
1943 | 921k | WORD32 *x2, *x0; |
1944 | 921k | WORD32 *y0, *y1, *y2, *y3; |
1945 | | |
1946 | 921k | y0 = ptr_y; |
1947 | 921k | y2 = ptr_y + (WORD32)npoints; |
1948 | 921k | x0 = ptr_x; |
1949 | 921k | x2 = ptr_x + (WORD32)(npoints >> 1); |
1950 | | |
1951 | 921k | y1 = y0 + (WORD32)(npoints >> 2); |
1952 | 921k | y3 = y2 + (WORD32)(npoints >> 2); |
1953 | | |
1954 | 2.76M | for (k = 0; k < 2; k++) { |
1955 | 5.53M | for (i = 0; i<npoints>> 1; i += 8) { |
1956 | 3.68M | h2 = *pdig_rev_tbl++ >> 2; |
1957 | | |
1958 | 3.68M | x_0 = *x0++; |
1959 | 3.68M | x_1 = *x0++; |
1960 | 3.68M | x_2 = *x0++; |
1961 | 3.68M | x_3 = *x0++; |
1962 | 3.68M | x_4 = *x0++; |
1963 | 3.68M | x_5 = *x0++; |
1964 | 3.68M | x_6 = *x0++; |
1965 | 3.68M | x_7 = *x0++; |
1966 | | |
1967 | 3.68M | n00 = ixheaac_add32_sat(x_0, x_2); |
1968 | 3.68M | n01 = ixheaac_add32_sat(x_1, x_3); |
1969 | 3.68M | n20 = ixheaac_sub32_sat(x_0, x_2); |
1970 | 3.68M | n21 = ixheaac_sub32_sat(x_1, x_3); |
1971 | 3.68M | n10 = ixheaac_add32_sat(x_4, x_6); |
1972 | 3.68M | n11 = ixheaac_add32_sat(x_5, x_7); |
1973 | 3.68M | n30 = ixheaac_sub32_sat(x_4, x_6); |
1974 | 3.68M | n31 = ixheaac_sub32_sat(x_5, x_7); |
1975 | | |
1976 | 3.68M | y0[h2] = n00; |
1977 | 3.68M | y0[h2 + 1] = n01; |
1978 | 3.68M | y1[h2] = n10; |
1979 | 3.68M | y1[h2 + 1] = n11; |
1980 | 3.68M | y2[h2] = n20; |
1981 | 3.68M | y2[h2 + 1] = n21; |
1982 | 3.68M | y3[h2] = n30; |
1983 | 3.68M | y3[h2 + 1] = n31; |
1984 | | |
1985 | 3.68M | x_8 = *x2++; |
1986 | 3.68M | x_9 = *x2++; |
1987 | 3.68M | x_a = *x2++; |
1988 | 3.68M | x_b = *x2++; |
1989 | 3.68M | x_c = *x2++; |
1990 | 3.68M | x_d = *x2++; |
1991 | 3.68M | x_e = *x2++; |
1992 | 3.68M | x_f = *x2++; |
1993 | | |
1994 | 3.68M | n02 = ixheaac_add32_sat(x_8, x_a); |
1995 | 3.68M | n03 = ixheaac_add32_sat(x_9, x_b); |
1996 | 3.68M | n22 = ixheaac_sub32_sat(x_8, x_a); |
1997 | 3.68M | n23 = ixheaac_sub32_sat(x_9, x_b); |
1998 | 3.68M | n12 = ixheaac_add32_sat(x_c, x_e); |
1999 | 3.68M | n13 = ixheaac_add32_sat(x_d, x_f); |
2000 | 3.68M | n32 = ixheaac_sub32_sat(x_c, x_e); |
2001 | 3.68M | n33 = ixheaac_sub32_sat(x_d, x_f); |
2002 | | |
2003 | 3.68M | y0[h2 + 2] = n02; |
2004 | 3.68M | y0[h2 + 3] = n03; |
2005 | 3.68M | y1[h2 + 2] = n12; |
2006 | 3.68M | y1[h2 + 3] = n13; |
2007 | 3.68M | y2[h2 + 2] = n22; |
2008 | 3.68M | y2[h2 + 3] = n23; |
2009 | 3.68M | y3[h2 + 2] = n32; |
2010 | 3.68M | y3[h2 + 3] = n33; |
2011 | 3.68M | } |
2012 | 1.84M | x0 += (WORD32)npoints >> 1; |
2013 | 1.84M | x2 += (WORD32)npoints >> 1; |
2014 | 1.84M | } |
2015 | 921k | } |
2016 | | #endif |