/src/libxaac/decoder/ixheaacd_fft.c
Line | Count | Source |
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 <stdlib.h> |
21 | | #include <stdio.h> |
22 | | |
23 | | #include "ixheaac_type_def.h" |
24 | | #include "ixheaacd_interface.h" |
25 | | #include "ixheaac_constants.h" |
26 | | #include "ixheaac_basic_ops32.h" |
27 | | #include "ixheaac_basic_ops40.h" |
28 | | #include "ixheaacd_function_selector.h" |
29 | | |
30 | | extern const WORD32 ixheaacd_twiddle_table_fft_32x32[514]; |
31 | | extern const FLOAT32 ixheaacd_twiddle_table_fft[514]; |
32 | | extern const FLOAT32 ixheaacd_twiddle_table_fft_flt[16]; |
33 | | extern const WORD32 ixheaacd_twiddle_table_3pr[1155]; |
34 | | extern const WORD32 ixheaacd_twiddle_table_3pi[1155]; |
35 | | extern const WORD8 ixheaacd_mps_dig_rev[8]; |
36 | | |
37 | | #define PLATFORM_INLINE __inline |
38 | | |
39 | | #define DIG_REV(i, m, j) \ |
40 | 75.9M | do { \ |
41 | 75.9M | unsigned _ = (i); \ |
42 | 75.9M | _ = ((_ & 0x33333333) << 2) | ((_ & ~0x33333333) >> 2); \ |
43 | 75.9M | _ = ((_ & 0x0F0F0F0F) << 4) | ((_ & ~0x0F0F0F0F) >> 4); \ |
44 | 75.9M | _ = ((_ & 0x00FF00FF) << 8) | ((_ & ~0x00FF00FF) >> 8); \ |
45 | 75.9M | (j) = _ >> (m); \ |
46 | 75.9M | } while (0) |
47 | | |
48 | 2.09G | static PLATFORM_INLINE WORD32 ixheaacd_mult32_sat(WORD32 a, WORD32 b) { |
49 | 2.09G | WORD32 result; |
50 | 2.09G | WORD64 temp_result; |
51 | | |
52 | 2.09G | temp_result = (WORD64)a * (WORD64)b; |
53 | 2.09G | result = ixheaac_sat64_32(temp_result >> 31); |
54 | | |
55 | 2.09G | return (result); |
56 | 2.09G | } |
57 | | |
58 | 335M | static PLATFORM_INLINE WORD32 ixheaacd_mac32_sat(WORD32 a, WORD32 b, WORD32 c) { |
59 | 335M | WORD32 result; |
60 | | |
61 | 335M | result = ixheaac_add32_sat(a, ixheaacd_mult32_sat(b, c)); |
62 | | |
63 | 335M | return (result); |
64 | 335M | } |
65 | | |
66 | 1.78G | static PLATFORM_INLINE FLOAT32 ixheaacd_mult32X32float(FLOAT32 a, FLOAT32 b) { |
67 | 1.78G | FLOAT32 result; |
68 | | |
69 | 1.78G | result = a * b; |
70 | | |
71 | 1.78G | return result; |
72 | 1.78G | } |
73 | | |
74 | 331M | static PLATFORM_INLINE FLOAT32 ixheaacd_mac32X32float(FLOAT32 a, FLOAT32 b, FLOAT32 c) { |
75 | 331M | FLOAT32 result; |
76 | | |
77 | 331M | result = a + b * c; |
78 | | |
79 | 331M | return result; |
80 | 331M | } |
81 | | |
82 | | VOID ixheaacd_mps_synth_calc_fft(FLOAT32 *ptr_xr, FLOAT32 *ptr_xi, |
83 | 8.25M | WORD32 npoints) { |
84 | 8.25M | WORD32 i, j, k; |
85 | 8.25M | FLOAT32 y[64], z[64]; |
86 | 8.25M | FLOAT32 *ptr_y = y, *ptr_z = z; |
87 | 8.25M | const FLOAT32 *ptr_w = ixheaacd_twiddle_table_fft_flt; |
88 | | |
89 | 74.2M | for (i = 0; i < npoints; i += 4) { |
90 | 66.0M | FLOAT32 x0r, x0i, x1r, x1i, x2r, x2i, x3r, x3i; |
91 | 66.0M | FLOAT32 *inp = ptr_xr; |
92 | 66.0M | FLOAT32 tmk; |
93 | | |
94 | 66.0M | WORD32 h2 = ixheaacd_mps_dig_rev[i >> 2]; |
95 | | |
96 | 66.0M | inp += (h2); |
97 | | |
98 | 66.0M | x0r = *inp; |
99 | 66.0M | x0i = *(inp + 1); |
100 | 66.0M | inp += 16; |
101 | | |
102 | 66.0M | x1r = *inp; |
103 | 66.0M | x1i = *(inp + 1); |
104 | 66.0M | inp += 16; |
105 | | |
106 | 66.0M | x2r = *inp; |
107 | 66.0M | x2i = *(inp + 1); |
108 | 66.0M | inp += 16; |
109 | | |
110 | 66.0M | x3r = *inp; |
111 | 66.0M | x3i = *(inp + 1); |
112 | | |
113 | 66.0M | x0r = x0r + x2r; |
114 | 66.0M | x0i = x0i + x2i; |
115 | | |
116 | 66.0M | tmk = x0r - x2r; |
117 | 66.0M | x2r = tmk - x2r; |
118 | 66.0M | tmk = x0i - x2i; |
119 | 66.0M | x2i = tmk - x2i; |
120 | | |
121 | 66.0M | x1r = x1r + x3r; |
122 | 66.0M | x1i = x1i + x3i; |
123 | | |
124 | 66.0M | tmk = x1r - x3r; |
125 | 66.0M | x3r = tmk - x3r; |
126 | 66.0M | tmk = x1i - x3i; |
127 | 66.0M | x3i = tmk - x3i; |
128 | | |
129 | 66.0M | x0r = x0r + x1r; |
130 | 66.0M | x0i = x0i + x1i; |
131 | | |
132 | 66.0M | tmk = x0r - x1r; |
133 | 66.0M | x1r = tmk - x1r; |
134 | 66.0M | tmk = x0i - x1i; |
135 | 66.0M | x1i = tmk - x1i; |
136 | | |
137 | 66.0M | x2r = x2r + x3i; |
138 | 66.0M | x2i = x2i - x3r; |
139 | | |
140 | 66.0M | tmk = x2r - x3i; |
141 | 66.0M | x3i = tmk - x3i; |
142 | 66.0M | tmk = x2i + x3r; |
143 | 66.0M | x3r = tmk + x3r; |
144 | | |
145 | 66.0M | *ptr_y++ = x0r; |
146 | 66.0M | *ptr_y++ = x0i; |
147 | 66.0M | *ptr_y++ = x2r; |
148 | 66.0M | *ptr_y++ = x2i; |
149 | 66.0M | *ptr_y++ = x1r; |
150 | 66.0M | *ptr_y++ = x1i; |
151 | 66.0M | *ptr_y++ = x3i; |
152 | 66.0M | *ptr_y++ = x3r; |
153 | | |
154 | 66.0M | inp = ptr_xi; |
155 | | |
156 | 66.0M | inp += (h2); |
157 | | |
158 | 66.0M | x0r = *inp; |
159 | 66.0M | x0i = *(inp + 1); |
160 | 66.0M | inp += 16; |
161 | | |
162 | 66.0M | x1r = *inp; |
163 | 66.0M | x1i = *(inp + 1); |
164 | 66.0M | inp += 16; |
165 | | |
166 | 66.0M | x2r = *inp; |
167 | 66.0M | x2i = *(inp + 1); |
168 | 66.0M | inp += 16; |
169 | | |
170 | 66.0M | x3r = *inp; |
171 | 66.0M | x3i = *(inp + 1); |
172 | | |
173 | 66.0M | x0r = x0r + x2r; |
174 | 66.0M | x0i = x0i + x2i; |
175 | | |
176 | 66.0M | tmk = x0r - x2r; |
177 | 66.0M | x2r = tmk - x2r; |
178 | 66.0M | tmk = x0i - x2i; |
179 | 66.0M | x2i = tmk - x2i; |
180 | | |
181 | 66.0M | x1r = x1r + x3r; |
182 | 66.0M | x1i = x1i + x3i; |
183 | | |
184 | 66.0M | tmk = x1r - x3r; |
185 | 66.0M | x3r = tmk - x3r; |
186 | 66.0M | tmk = x1i - x3i; |
187 | 66.0M | x3i = tmk - x3i; |
188 | | |
189 | 66.0M | x0r = x0r + x1r; |
190 | 66.0M | x0i = x0i + x1i; |
191 | | |
192 | 66.0M | tmk = x0r - x1r; |
193 | 66.0M | x1r = tmk - x1r; |
194 | 66.0M | tmk = x0i - x1i; |
195 | 66.0M | x1i = tmk - x1i; |
196 | | |
197 | 66.0M | x2r = x2r + x3i; |
198 | 66.0M | x2i = x2i - x3r; |
199 | | |
200 | 66.0M | tmk = x2r - x3i; |
201 | 66.0M | x3i = tmk - x3i; |
202 | 66.0M | tmk = x2i + x3r; |
203 | 66.0M | x3r = tmk + x3r; |
204 | | |
205 | 66.0M | *ptr_z++ = x0r; |
206 | 66.0M | *ptr_z++ = x0i; |
207 | 66.0M | *ptr_z++ = x2r; |
208 | 66.0M | *ptr_z++ = x2i; |
209 | 66.0M | *ptr_z++ = x1r; |
210 | 66.0M | *ptr_z++ = x1i; |
211 | 66.0M | *ptr_z++ = x3i; |
212 | 66.0M | *ptr_z++ = x3r; |
213 | 66.0M | } |
214 | 8.25M | ptr_y -= 64; |
215 | 8.25M | ptr_z -= 64; |
216 | 8.25M | { |
217 | 8.25M | FLOAT32 *data_r = ptr_y; |
218 | 8.25M | FLOAT32 *data_i = ptr_z; |
219 | 24.7M | for (k = 2; k != 0; k--) { |
220 | 16.5M | FLOAT32 x0r, x0i, x1r, x1i, x2r, x2i, x3r, x3i; |
221 | | |
222 | 16.5M | x0r = (*data_r); |
223 | 16.5M | x0i = (*(data_r + 1)); |
224 | 16.5M | data_r += 8; |
225 | | |
226 | 16.5M | x1r = (*data_r); |
227 | 16.5M | x1i = (*(data_r + 1)); |
228 | 16.5M | data_r += 8; |
229 | | |
230 | 16.5M | x2r = (*data_r); |
231 | 16.5M | x2i = (*(data_r + 1)); |
232 | 16.5M | data_r += 8; |
233 | | |
234 | 16.5M | x3r = (*data_r); |
235 | 16.5M | x3i = (*(data_r + 1)); |
236 | 16.5M | data_r -= 24; |
237 | | |
238 | 16.5M | x0r = x0r + x2r; |
239 | 16.5M | x0i = x0i + x2i; |
240 | 16.5M | x2r = x0r - (x2r * 2); |
241 | 16.5M | x2i = x0i - (x2i * 2); |
242 | 16.5M | x1r = x1r + x3r; |
243 | 16.5M | x1i = x1i + x3i; |
244 | 16.5M | x3r = x1r - (x3r * 2); |
245 | 16.5M | x3i = x1i - (x3i * 2); |
246 | | |
247 | 16.5M | x0r = x0r + x1r; |
248 | 16.5M | x0i = x0i + x1i; |
249 | 16.5M | x1r = x0r - (x1r * 2); |
250 | 16.5M | x1i = x0i - (x1i * 2); |
251 | 16.5M | x2r = x2r + x3i; |
252 | 16.5M | x2i = x2i - x3r; |
253 | 16.5M | x3i = x2r - (x3i * 2); |
254 | 16.5M | x3r = x2i + (x3r * 2); |
255 | | |
256 | 16.5M | *data_r = x0r; |
257 | 16.5M | *(data_r + 1) = x0i; |
258 | 16.5M | data_r += 8; |
259 | | |
260 | 16.5M | *data_r = x2r; |
261 | 16.5M | *(data_r + 1) = x2i; |
262 | 16.5M | data_r += 8; |
263 | | |
264 | 16.5M | *data_r = x1r; |
265 | 16.5M | *(data_r + 1) = x1i; |
266 | 16.5M | data_r += 8; |
267 | | |
268 | 16.5M | *data_r = x3i; |
269 | 16.5M | *(data_r + 1) = x3r; |
270 | 16.5M | data_r += 8; |
271 | | |
272 | 16.5M | x0r = (*data_i); |
273 | 16.5M | x0i = (*(data_i + 1)); |
274 | 16.5M | data_i += 8; |
275 | | |
276 | 16.5M | x1r = (*data_i); |
277 | 16.5M | x1i = (*(data_i + 1)); |
278 | 16.5M | data_i += 8; |
279 | | |
280 | 16.5M | x2r = (*data_i); |
281 | 16.5M | x2i = (*(data_i + 1)); |
282 | 16.5M | data_i += 8; |
283 | | |
284 | 16.5M | x3r = (*data_i); |
285 | 16.5M | x3i = (*(data_i + 1)); |
286 | 16.5M | data_i -= 24; |
287 | | |
288 | 16.5M | x0r = x0r + x2r; |
289 | 16.5M | x0i = x0i + x2i; |
290 | 16.5M | x2r = x0r - (x2r * 2); |
291 | 16.5M | x2i = x0i - (x2i * 2); |
292 | 16.5M | x1r = x1r + x3r; |
293 | 16.5M | x1i = x1i + x3i; |
294 | 16.5M | x3r = x1r - (x3r * 2); |
295 | 16.5M | x3i = x1i - (x3i * 2); |
296 | | |
297 | 16.5M | x0r = x0r + x1r; |
298 | 16.5M | x0i = x0i + x1i; |
299 | 16.5M | x1r = x0r - (x1r * 2); |
300 | 16.5M | x1i = x0i - (x1i * 2); |
301 | 16.5M | x2r = x2r + x3i; |
302 | 16.5M | x2i = x2i - x3r; |
303 | 16.5M | x3i = x2r - (x3i * 2); |
304 | 16.5M | x3r = x2i + (x3r * 2); |
305 | | |
306 | 16.5M | *data_i = x0r; |
307 | 16.5M | *(data_i + 1) = x0i; |
308 | 16.5M | data_i += 8; |
309 | | |
310 | 16.5M | *data_i = x2r; |
311 | 16.5M | *(data_i + 1) = x2i; |
312 | 16.5M | data_i += 8; |
313 | | |
314 | 16.5M | *data_i = x1r; |
315 | 16.5M | *(data_i + 1) = x1i; |
316 | 16.5M | data_i += 8; |
317 | | |
318 | 16.5M | *data_i = x3i; |
319 | 16.5M | *(data_i + 1) = x3r; |
320 | 16.5M | data_i += 8; |
321 | 16.5M | } |
322 | 8.25M | data_r = ptr_y + 2; |
323 | 8.25M | data_i = ptr_z + 2; |
324 | | |
325 | 24.7M | for (k = 2; k != 0; k--) { |
326 | 16.5M | FLOAT32 tmp; |
327 | 16.5M | FLOAT32 x0r, x0i, x1r, x1i, x2r, x2i, x3r, x3i; |
328 | | |
329 | 16.5M | data_r += 8; |
330 | | |
331 | 16.5M | x1r = *data_r; |
332 | 16.5M | x1i = *(data_r + 1); |
333 | 16.5M | data_r += 8; |
334 | | |
335 | 16.5M | x2r = *data_r; |
336 | 16.5M | x2i = *(data_r + 1); |
337 | 16.5M | data_r += 8; |
338 | | |
339 | 16.5M | x3r = *data_r; |
340 | 16.5M | x3i = *(data_r + 1); |
341 | 16.5M | data_r -= 24; |
342 | | |
343 | 16.5M | tmp = (FLOAT32)(ixheaacd_mult32X32float((FLOAT32)x1r, 0.923880f) - |
344 | 16.5M | ixheaacd_mult32X32float((FLOAT32)x1i, -0.382683f)); |
345 | 16.5M | x1i = (FLOAT32)ixheaacd_mac32X32float(ixheaacd_mult32X32float((FLOAT32)x1r, -0.382683f), |
346 | 16.5M | (FLOAT32)x1i, 0.923880f); |
347 | 16.5M | x1r = tmp; |
348 | | |
349 | 16.5M | tmp = (FLOAT32)(ixheaacd_mult32X32float((FLOAT32)x2r, 0.707107f) - |
350 | 16.5M | ixheaacd_mult32X32float((FLOAT32)x2i, -0.707107f)); |
351 | 16.5M | x2i = (FLOAT32)ixheaacd_mac32X32float(ixheaacd_mult32X32float((FLOAT32)x2r, -0.707107f), |
352 | 16.5M | (FLOAT32)x2i, 0.707107f); |
353 | 16.5M | x2r = tmp; |
354 | | |
355 | 16.5M | tmp = (FLOAT32)(ixheaacd_mult32X32float((FLOAT32)x3r, 0.382683f) - |
356 | 16.5M | ixheaacd_mult32X32float((FLOAT32)x3i, -0.923880f)); |
357 | 16.5M | x3i = (FLOAT32)ixheaacd_mac32X32float(ixheaacd_mult32X32float((FLOAT32)x3r, -0.923880f), |
358 | 16.5M | (FLOAT32)x3i, 0.382683f); |
359 | 16.5M | x3r = tmp; |
360 | | |
361 | 16.5M | x0r = (*data_r); |
362 | 16.5M | x0i = (*(data_r + 1)); |
363 | | |
364 | 16.5M | x0r = x0r + (x2r); |
365 | 16.5M | x0i = x0i + (x2i); |
366 | 16.5M | x2r = x0r - (x2r * 2); |
367 | 16.5M | x2i = x0i - (x2i * 2); |
368 | 16.5M | x1r = x1r + x3r; |
369 | 16.5M | x1i = x1i + x3i; |
370 | 16.5M | x3r = x1r - (x3r * 2); |
371 | 16.5M | x3i = x1i - (x3i * 2); |
372 | | |
373 | 16.5M | x0r = x0r + (x1r); |
374 | 16.5M | x0i = x0i + (x1i); |
375 | 16.5M | x1r = x0r - (x1r * 2); |
376 | 16.5M | x1i = x0i - (x1i * 2); |
377 | 16.5M | x2r = x2r + (x3i); |
378 | 16.5M | x2i = x2i - (x3r); |
379 | 16.5M | x3i = x2r - (x3i * 2); |
380 | 16.5M | x3r = x2i + (x3r * 2); |
381 | | |
382 | 16.5M | *data_r = x0r; |
383 | 16.5M | *(data_r + 1) = x0i; |
384 | 16.5M | data_r += 8; |
385 | | |
386 | 16.5M | *data_r = x2r; |
387 | 16.5M | *(data_r + 1) = x2i; |
388 | 16.5M | data_r += 8; |
389 | | |
390 | 16.5M | *data_r = x1r; |
391 | 16.5M | *(data_r + 1) = x1i; |
392 | 16.5M | data_r += 8; |
393 | | |
394 | 16.5M | *data_r = x3i; |
395 | 16.5M | *(data_r + 1) = x3r; |
396 | 16.5M | data_r += 8; |
397 | 16.5M | data_i += 8; |
398 | | |
399 | 16.5M | x1r = *data_i; |
400 | 16.5M | x1i = *(data_i + 1); |
401 | 16.5M | data_i += 8; |
402 | | |
403 | 16.5M | x2r = *data_i; |
404 | 16.5M | x2i = *(data_i + 1); |
405 | 16.5M | data_i += 8; |
406 | | |
407 | 16.5M | x3r = *data_i; |
408 | 16.5M | x3i = *(data_i + 1); |
409 | 16.5M | data_i -= 24; |
410 | | |
411 | 16.5M | tmp = (FLOAT32)(ixheaacd_mult32X32float((FLOAT32)x1r, 0.923880f) - |
412 | 16.5M | ixheaacd_mult32X32float((FLOAT32)x1i, -0.382683f)); |
413 | 16.5M | x1i = (FLOAT32)ixheaacd_mac32X32float(ixheaacd_mult32X32float((FLOAT32)x1r, -0.382683f), |
414 | 16.5M | (FLOAT32)x1i, 0.923880f); |
415 | 16.5M | x1r = tmp; |
416 | | |
417 | 16.5M | tmp = (FLOAT32)(ixheaacd_mult32X32float((FLOAT32)x2r, 0.707107f) - |
418 | 16.5M | ixheaacd_mult32X32float((FLOAT32)x2i, -0.707107f)); |
419 | 16.5M | x2i = (FLOAT32)ixheaacd_mac32X32float(ixheaacd_mult32X32float((FLOAT32)x2r, -0.707107f), |
420 | 16.5M | (FLOAT32)x2i, 0.707107f); |
421 | 16.5M | x2r = tmp; |
422 | | |
423 | 16.5M | tmp = (FLOAT32)(ixheaacd_mult32X32float((FLOAT32)x3r, 0.382683f) - |
424 | 16.5M | ixheaacd_mult32X32float((FLOAT32)x3i, -0.923880f)); |
425 | 16.5M | x3i = (FLOAT32)ixheaacd_mac32X32float(ixheaacd_mult32X32float((FLOAT32)x3r, -0.923880f), |
426 | 16.5M | (FLOAT32)x3i, 0.382683f); |
427 | 16.5M | x3r = tmp; |
428 | | |
429 | 16.5M | x0r = (*data_i); |
430 | 16.5M | x0i = (*(data_i + 1)); |
431 | | |
432 | 16.5M | x0r = x0r + (x2r); |
433 | 16.5M | x0i = x0i + (x2i); |
434 | 16.5M | x2r = x0r - (x2r * 2); |
435 | 16.5M | x2i = x0i - (x2i * 2); |
436 | 16.5M | x1r = x1r + x3r; |
437 | 16.5M | x1i = x1i + x3i; |
438 | 16.5M | x3r = x1r - (x3r * 2); |
439 | 16.5M | x3i = x1i - (x3i * 2); |
440 | | |
441 | 16.5M | x0r = x0r + (x1r); |
442 | 16.5M | x0i = x0i + (x1i); |
443 | 16.5M | x1r = x0r - (x1r * 2); |
444 | 16.5M | x1i = x0i - (x1i * 2); |
445 | 16.5M | x2r = x2r + (x3i); |
446 | 16.5M | x2i = x2i - (x3r); |
447 | 16.5M | x3i = x2r - (x3i * 2); |
448 | 16.5M | x3r = x2i + (x3r * 2); |
449 | | |
450 | 16.5M | *data_i = x0r; |
451 | 16.5M | *(data_i + 1) = x0i; |
452 | 16.5M | data_i += 8; |
453 | | |
454 | 16.5M | *data_i = x2r; |
455 | 16.5M | *(data_i + 1) = x2i; |
456 | 16.5M | data_i += 8; |
457 | | |
458 | 16.5M | *data_i = x1r; |
459 | 16.5M | *(data_i + 1) = x1i; |
460 | 16.5M | data_i += 8; |
461 | | |
462 | 16.5M | *data_i = x3i; |
463 | 16.5M | *(data_i + 1) = x3r; |
464 | 16.5M | data_i += 8; |
465 | 16.5M | } |
466 | 8.25M | data_r -= 62; |
467 | 8.25M | data_i -= 62; |
468 | 24.7M | for (k = 2; k != 0; k--) { |
469 | 16.5M | FLOAT32 tmp; |
470 | 16.5M | FLOAT32 x0r, x0i, x1r, x1i, x2r, x2i, x3r, x3i; |
471 | | |
472 | 16.5M | data_r += 8; |
473 | | |
474 | 16.5M | x1r = *data_r; |
475 | 16.5M | x1i = *(data_r + 1); |
476 | 16.5M | data_r += 8; |
477 | | |
478 | 16.5M | x2r = *data_r; |
479 | 16.5M | x2i = *(data_r + 1); |
480 | 16.5M | data_r += 8; |
481 | | |
482 | 16.5M | x3r = *data_r; |
483 | 16.5M | x3i = *(data_r + 1); |
484 | 16.5M | data_r -= 24; |
485 | | |
486 | 16.5M | tmp = (FLOAT32)(ixheaacd_mult32X32float((FLOAT32)x1r, 0.707107f) - |
487 | 16.5M | ixheaacd_mult32X32float((FLOAT32)x1i, -0.707107f)); |
488 | 16.5M | x1i = (FLOAT32)ixheaacd_mac32X32float(ixheaacd_mult32X32float((FLOAT32)x1r, -0.707107f), |
489 | 16.5M | (FLOAT32)x1i, 0.707107f); |
490 | 16.5M | x1r = tmp; |
491 | | |
492 | 16.5M | tmp = x2i; |
493 | 16.5M | x2i = -x2r; |
494 | 16.5M | x2r = tmp; |
495 | | |
496 | 16.5M | tmp = (FLOAT32)(ixheaacd_mult32X32float((FLOAT32)x3r, -0.707107f) + |
497 | 16.5M | ixheaacd_mult32X32float((FLOAT32)x3i, 0.707107f)); |
498 | 16.5M | x3i = (FLOAT32)(-ixheaacd_mult32X32float((FLOAT32)x3r, 0.707107f) + |
499 | 16.5M | ixheaacd_mult32X32float((FLOAT32)x3i, -0.707107f)); |
500 | 16.5M | x3r = tmp; |
501 | | |
502 | 16.5M | x0r = (*data_r); |
503 | 16.5M | x0i = (*(data_r + 1)); |
504 | | |
505 | 16.5M | x0r = x0r + (x2r); |
506 | 16.5M | x0i = x0i + (x2i); |
507 | 16.5M | x2r = x0r - (x2r * 2); |
508 | 16.5M | x2i = x0i - (x2i * 2); |
509 | 16.5M | x1r = x1r + x3r; |
510 | 16.5M | x1i = x1i + x3i; |
511 | 16.5M | x3r = x1r - (x3r * 2); |
512 | 16.5M | x3i = x1i - (x3i * 2); |
513 | | |
514 | 16.5M | x0r = x0r + (x1r); |
515 | 16.5M | x0i = x0i + (x1i); |
516 | 16.5M | x1r = x0r - (x1r * 2); |
517 | 16.5M | x1i = x0i - (x1i * 2); |
518 | 16.5M | x2r = x2r + (x3i); |
519 | 16.5M | x2i = x2i - (x3r); |
520 | 16.5M | x3i = x2r - (x3i * 2); |
521 | 16.5M | x3r = x2i + (x3r * 2); |
522 | | |
523 | 16.5M | *data_r = x0r; |
524 | 16.5M | *(data_r + 1) = x0i; |
525 | 16.5M | data_r += 8; |
526 | | |
527 | 16.5M | *data_r = x2r; |
528 | 16.5M | *(data_r + 1) = x2i; |
529 | 16.5M | data_r += 8; |
530 | | |
531 | 16.5M | *data_r = x1r; |
532 | 16.5M | *(data_r + 1) = x1i; |
533 | 16.5M | data_r += 8; |
534 | | |
535 | 16.5M | *data_r = x3i; |
536 | 16.5M | *(data_r + 1) = x3r; |
537 | 16.5M | data_r += 8; |
538 | 16.5M | data_i += 8; |
539 | | |
540 | 16.5M | x1r = *data_i; |
541 | 16.5M | x1i = *(data_i + 1); |
542 | 16.5M | data_i += 8; |
543 | | |
544 | 16.5M | x2r = *data_i; |
545 | 16.5M | x2i = *(data_i + 1); |
546 | 16.5M | data_i += 8; |
547 | | |
548 | 16.5M | x3r = *data_i; |
549 | 16.5M | x3i = *(data_i + 1); |
550 | 16.5M | data_i -= 24; |
551 | | |
552 | 16.5M | tmp = (FLOAT32)(ixheaacd_mult32X32float((FLOAT32)x1r, 0.707107f) - |
553 | 16.5M | ixheaacd_mult32X32float((FLOAT32)x1i, -0.707107f)); |
554 | 16.5M | x1i = (FLOAT32)ixheaacd_mac32X32float(ixheaacd_mult32X32float((FLOAT32)x1r, -0.707107f), |
555 | 16.5M | (FLOAT32)x1i, 0.707107f); |
556 | 16.5M | x1r = tmp; |
557 | | |
558 | 16.5M | tmp = x2i; |
559 | 16.5M | x2i = -x2r; |
560 | 16.5M | x2r = tmp; |
561 | | |
562 | 16.5M | tmp = (FLOAT32)(ixheaacd_mult32X32float((FLOAT32)x3r, -0.707107f) + |
563 | 16.5M | ixheaacd_mult32X32float((FLOAT32)x3i, 0.707107f)); |
564 | 16.5M | x3i = (FLOAT32)(-ixheaacd_mult32X32float((FLOAT32)x3r, 0.707107f) + |
565 | 16.5M | ixheaacd_mult32X32float((FLOAT32)x3i, -0.707107f)); |
566 | 16.5M | x3r = tmp; |
567 | | |
568 | 16.5M | x0r = (*data_i); |
569 | 16.5M | x0i = (*(data_i + 1)); |
570 | | |
571 | 16.5M | x0r = x0r + (x2r); |
572 | 16.5M | x0i = x0i + (x2i); |
573 | 16.5M | x2r = x0r - (x2r * 2); |
574 | 16.5M | x2i = x0i - (x2i * 2); |
575 | 16.5M | x1r = x1r + x3r; |
576 | 16.5M | x1i = x1i + x3i; |
577 | 16.5M | x3r = x1r - (x3r * 2); |
578 | 16.5M | x3i = x1i - (x3i * 2); |
579 | | |
580 | 16.5M | x0r = x0r + (x1r); |
581 | 16.5M | x0i = x0i + (x1i); |
582 | 16.5M | x1r = x0r - (x1r * 2); |
583 | 16.5M | x1i = x0i - (x1i * 2); |
584 | 16.5M | x2r = x2r + (x3i); |
585 | 16.5M | x2i = x2i - (x3r); |
586 | 16.5M | x3i = x2r - (x3i * 2); |
587 | 16.5M | x3r = x2i + (x3r * 2); |
588 | | |
589 | 16.5M | *data_i = x0r; |
590 | 16.5M | *(data_i + 1) = x0i; |
591 | 16.5M | data_i += 8; |
592 | | |
593 | 16.5M | *data_i = x2r; |
594 | 16.5M | *(data_i + 1) = x2i; |
595 | 16.5M | data_i += 8; |
596 | | |
597 | 16.5M | *data_i = x1r; |
598 | 16.5M | *(data_i + 1) = x1i; |
599 | 16.5M | data_i += 8; |
600 | | |
601 | 16.5M | *data_i = x3i; |
602 | 16.5M | *(data_i + 1) = x3r; |
603 | 16.5M | data_i += 8; |
604 | 16.5M | } |
605 | 8.25M | data_r -= 62; |
606 | 8.25M | data_i -= 62; |
607 | 24.7M | for (k = 2; k != 0; k--) { |
608 | 16.5M | FLOAT32 tmp; |
609 | 16.5M | FLOAT32 x0r, x0i, x1r, x1i, x2r, x2i, x3r, x3i; |
610 | | |
611 | 16.5M | data_r += 8; |
612 | | |
613 | 16.5M | x1r = *data_r; |
614 | 16.5M | x1i = *(data_r + 1); |
615 | 16.5M | data_r += 8; |
616 | | |
617 | 16.5M | x2r = *data_r; |
618 | 16.5M | x2i = *(data_r + 1); |
619 | 16.5M | data_r += 8; |
620 | | |
621 | 16.5M | x3r = *data_r; |
622 | 16.5M | x3i = *(data_r + 1); |
623 | 16.5M | data_r -= 24; |
624 | | |
625 | 16.5M | tmp = (FLOAT32)(ixheaacd_mult32X32float((FLOAT32)x1r, 0.382683f) - |
626 | 16.5M | ixheaacd_mult32X32float((FLOAT32)x1i, -0.923880f)); |
627 | 16.5M | x1i = (FLOAT32)ixheaacd_mac32X32float(ixheaacd_mult32X32float((FLOAT32)x1r, -0.923880f), |
628 | 16.5M | (FLOAT32)x1i, 0.382683f); |
629 | 16.5M | x1r = tmp; |
630 | | |
631 | 16.5M | tmp = (FLOAT32)(ixheaacd_mult32X32float((FLOAT32)x2r, -0.707107f) + |
632 | 16.5M | ixheaacd_mult32X32float((FLOAT32)x2i, 0.707107f)); |
633 | 16.5M | x2i = (FLOAT32)(-ixheaacd_mult32X32float((FLOAT32)x2r, 0.707107f) + |
634 | 16.5M | ixheaacd_mult32X32float((FLOAT32)x2i, -0.707107f)); |
635 | 16.5M | x2r = tmp; |
636 | | |
637 | 16.5M | tmp = (FLOAT32)(-ixheaacd_mult32X32float((FLOAT32)x3r, 0.923880f) + |
638 | 16.5M | ixheaacd_mult32X32float((FLOAT32)x3i, -0.382683f)); |
639 | 16.5M | x3i = (FLOAT32)ixheaacd_mac32X32float(ixheaacd_mult32X32float((FLOAT32)x3r, -0.382683f), |
640 | 16.5M | (FLOAT32)x3i, 0.923880f); |
641 | 16.5M | x3r = tmp; |
642 | | |
643 | 16.5M | x0r = (*data_r); |
644 | 16.5M | x0i = (*(data_r + 1)); |
645 | | |
646 | 16.5M | x0r = x0r + (x2r); |
647 | 16.5M | x0i = x0i + (x2i); |
648 | 16.5M | x2r = x0r - (x2r * 2); |
649 | 16.5M | x2i = x0i - (x2i * 2); |
650 | 16.5M | x1r = x1r + x3r; |
651 | 16.5M | x1i = x1i - x3i; |
652 | 16.5M | x3r = x1r - (x3r * 2); |
653 | 16.5M | x3i = x1i + (x3i * 2); |
654 | | |
655 | 16.5M | x0r = x0r + (x1r); |
656 | 16.5M | x0i = x0i + (x1i); |
657 | 16.5M | x1r = x0r - (x1r * 2); |
658 | 16.5M | x1i = x0i - (x1i * 2); |
659 | 16.5M | x2r = x2r + (x3i); |
660 | 16.5M | x2i = x2i - (x3r); |
661 | 16.5M | x3i = x2r - (x3i * 2); |
662 | 16.5M | x3r = x2i + (x3r * 2); |
663 | | |
664 | 16.5M | *data_r = x0r; |
665 | 16.5M | *(data_r + 1) = x0i; |
666 | 16.5M | data_r += 8; |
667 | | |
668 | 16.5M | *data_r = x2r; |
669 | 16.5M | *(data_r + 1) = x2i; |
670 | 16.5M | data_r += 8; |
671 | | |
672 | 16.5M | *data_r = x1r; |
673 | 16.5M | *(data_r + 1) = x1i; |
674 | 16.5M | data_r += 8; |
675 | | |
676 | 16.5M | *data_r = x3i; |
677 | 16.5M | *(data_r + 1) = x3r; |
678 | 16.5M | data_r += 8; |
679 | 16.5M | data_i += 8; |
680 | | |
681 | 16.5M | x1r = *data_i; |
682 | 16.5M | x1i = *(data_i + 1); |
683 | 16.5M | data_i += 8; |
684 | | |
685 | 16.5M | x2r = *data_i; |
686 | 16.5M | x2i = *(data_i + 1); |
687 | 16.5M | data_i += 8; |
688 | | |
689 | 16.5M | x3r = *data_i; |
690 | 16.5M | x3i = *(data_i + 1); |
691 | 16.5M | data_i -= 24; |
692 | | |
693 | 16.5M | tmp = (FLOAT32)(ixheaacd_mult32X32float((FLOAT32)x1r, 0.382683f) - |
694 | 16.5M | ixheaacd_mult32X32float((FLOAT32)x1i, -0.923880f)); |
695 | 16.5M | x1i = (FLOAT32)ixheaacd_mac32X32float(ixheaacd_mult32X32float((FLOAT32)x1r, -0.923880f), |
696 | 16.5M | (FLOAT32)x1i, 0.382683f); |
697 | 16.5M | x1r = tmp; |
698 | | |
699 | 16.5M | tmp = (FLOAT32)(ixheaacd_mult32X32float((FLOAT32)x2r, -0.707107f) + |
700 | 16.5M | ixheaacd_mult32X32float((FLOAT32)x2i, 0.707107f)); |
701 | 16.5M | x2i = (FLOAT32)(-ixheaacd_mult32X32float((FLOAT32)x2r, 0.707107f) + |
702 | 16.5M | ixheaacd_mult32X32float((FLOAT32)x2i, -0.707107f)); |
703 | 16.5M | x2r = tmp; |
704 | | |
705 | 16.5M | tmp = (FLOAT32)(-ixheaacd_mult32X32float((FLOAT32)x3r, 0.923880f) + |
706 | 16.5M | ixheaacd_mult32X32float((FLOAT32)x3i, -0.382683f)); |
707 | 16.5M | x3i = (FLOAT32)ixheaacd_mac32X32float(ixheaacd_mult32X32float((FLOAT32)x3r, -0.382683f), |
708 | 16.5M | (FLOAT32)x3i, 0.923880f); |
709 | 16.5M | x3r = tmp; |
710 | | |
711 | 16.5M | x0r = (*data_i); |
712 | 16.5M | x0i = (*(data_i + 1)); |
713 | | |
714 | 16.5M | x0r = x0r + (x2r); |
715 | 16.5M | x0i = x0i + (x2i); |
716 | 16.5M | x2r = x0r - (x2r * 2); |
717 | 16.5M | x2i = x0i - (x2i * 2); |
718 | 16.5M | x1r = x1r + x3r; |
719 | 16.5M | x1i = x1i - x3i; |
720 | 16.5M | x3r = x1r - (x3r * 2); |
721 | 16.5M | x3i = x1i + (x3i * 2); |
722 | | |
723 | 16.5M | x0r = x0r + (x1r); |
724 | 16.5M | x0i = x0i + (x1i); |
725 | 16.5M | x1r = x0r - (x1r * 2); |
726 | 16.5M | x1i = x0i - (x1i * 2); |
727 | 16.5M | x2r = x2r + (x3i); |
728 | 16.5M | x2i = x2i - (x3r); |
729 | 16.5M | x3i = x2r - (x3i * 2); |
730 | 16.5M | x3r = x2i + (x3r * 2); |
731 | | |
732 | 16.5M | *data_i = x0r; |
733 | 16.5M | *(data_i + 1) = x0i; |
734 | 16.5M | data_i += 8; |
735 | | |
736 | 16.5M | *data_i = x2r; |
737 | 16.5M | *(data_i + 1) = x2i; |
738 | 16.5M | data_i += 8; |
739 | | |
740 | 16.5M | *data_i = x1r; |
741 | 16.5M | *(data_i + 1) = x1i; |
742 | 16.5M | data_i += 8; |
743 | | |
744 | 16.5M | *data_i = x3i; |
745 | 16.5M | *(data_i + 1) = x3r; |
746 | 16.5M | data_i += 8; |
747 | 16.5M | } |
748 | 8.25M | data_r -= 62; |
749 | 8.25M | data_i -= 62; |
750 | 8.25M | } |
751 | 8.25M | { |
752 | 8.25M | const FLOAT32 *twiddles = ptr_w; |
753 | 8.25M | FLOAT32 x0r, x0i, x1r, x1i; |
754 | 74.2M | for (j = 8; j != 0; j--) { |
755 | 66.0M | FLOAT32 W1 = *twiddles; |
756 | 66.0M | twiddles++; |
757 | 66.0M | FLOAT32 W4 = *twiddles; |
758 | 66.0M | twiddles++; |
759 | 66.0M | FLOAT32 tmp; |
760 | | |
761 | 66.0M | x0r = *ptr_y; |
762 | 66.0M | x0i = *(ptr_y + 1); |
763 | 66.0M | ptr_y += 32; |
764 | 66.0M | ptr_xr += 32; |
765 | | |
766 | 66.0M | x1r = *ptr_y; |
767 | 66.0M | x1i = *(ptr_y + 1); |
768 | | |
769 | 66.0M | tmp = (FLOAT32)(ixheaacd_mult32X32float((FLOAT32)x1r, W1) - |
770 | 66.0M | ixheaacd_mult32X32float((FLOAT32)x1i, W4)); |
771 | 66.0M | x1i = (FLOAT32)ixheaacd_mac32X32float(ixheaacd_mult32X32float((FLOAT32)x1r, W4), |
772 | 66.0M | (FLOAT32)x1i, W1); |
773 | 66.0M | x1r = tmp; |
774 | | |
775 | 66.0M | *ptr_xr = (x0r) - (x1r); |
776 | 66.0M | *(ptr_xr + 1) = (x0i) - (x1i); |
777 | 66.0M | ptr_y -= 32; |
778 | 66.0M | ptr_xr -= 32; |
779 | | |
780 | 66.0M | *ptr_xr = (x0r) + (x1r); |
781 | 66.0M | *(ptr_xr + 1) = (x0i) + (x1i); |
782 | 66.0M | ptr_y += 2; |
783 | 66.0M | ptr_xr += 2; |
784 | | |
785 | 66.0M | x0r = *ptr_z; |
786 | 66.0M | x0i = *(ptr_z + 1); |
787 | 66.0M | ptr_z += 32; |
788 | 66.0M | ptr_xi += 32; |
789 | | |
790 | 66.0M | x1r = *ptr_z; |
791 | 66.0M | x1i = *(ptr_z + 1); |
792 | | |
793 | 66.0M | tmp = (FLOAT32)(ixheaacd_mult32X32float((FLOAT32)x1r, W1) - |
794 | 66.0M | ixheaacd_mult32X32float((FLOAT32)x1i, W4)); |
795 | 66.0M | x1i = (FLOAT32)ixheaacd_mac32X32float(ixheaacd_mult32X32float((FLOAT32)x1r, W4), |
796 | 66.0M | (FLOAT32)x1i, W1); |
797 | 66.0M | x1r = tmp; |
798 | | |
799 | 66.0M | *ptr_xi = (x0r) - (x1r); |
800 | 66.0M | *(ptr_xi + 1) = (x0i) - (x1i); |
801 | 66.0M | ptr_z -= 32; |
802 | 66.0M | ptr_xi -= 32; |
803 | | |
804 | 66.0M | *ptr_xi = (x0r) + (x1r); |
805 | 66.0M | *(ptr_xi + 1) = (x0i) + (x1i); |
806 | 66.0M | ptr_z += 2; |
807 | 66.0M | ptr_xi += 2; |
808 | 66.0M | } |
809 | 8.25M | twiddles = ptr_w; |
810 | 74.2M | for (j = 8; j != 0; j--) { |
811 | 66.0M | FLOAT32 W1 = *twiddles; |
812 | 66.0M | twiddles++; |
813 | 66.0M | FLOAT32 W4 = *twiddles; |
814 | 66.0M | twiddles++; |
815 | 66.0M | FLOAT32 tmp; |
816 | | |
817 | 66.0M | x0r = *ptr_y; |
818 | 66.0M | x0i = *(ptr_y + 1); |
819 | 66.0M | ptr_y += 32; |
820 | 66.0M | ptr_xr += 32; |
821 | | |
822 | 66.0M | x1r = *ptr_y; |
823 | 66.0M | x1i = *(ptr_y + 1); |
824 | | |
825 | 66.0M | tmp = (FLOAT32)(ixheaacd_mult32X32float((FLOAT32)x1r, W4) + |
826 | 66.0M | ixheaacd_mult32X32float((FLOAT32)x1i, W1)); |
827 | 66.0M | x1i = (FLOAT32)(-ixheaacd_mult32X32float((FLOAT32)x1r, W1) + |
828 | 66.0M | ixheaacd_mult32X32float((FLOAT32)x1i, W4)); |
829 | 66.0M | x1r = tmp; |
830 | | |
831 | 66.0M | *ptr_xr = (x0r) - (x1r); |
832 | 66.0M | *(ptr_xr + 1) = (x0i) - (x1i); |
833 | 66.0M | ptr_y -= 32; |
834 | 66.0M | ptr_xr -= 32; |
835 | | |
836 | 66.0M | *ptr_xr = (x0r) + (x1r); |
837 | 66.0M | *(ptr_xr + 1) = (x0i) + (x1i); |
838 | 66.0M | ptr_y += 2; |
839 | 66.0M | ptr_xr += 2; |
840 | | |
841 | 66.0M | x0r = *ptr_z; |
842 | 66.0M | x0i = *(ptr_z + 1); |
843 | 66.0M | ptr_z += 32; |
844 | 66.0M | ptr_xi += 32; |
845 | | |
846 | 66.0M | x1r = *ptr_z; |
847 | 66.0M | x1i = *(ptr_z + 1); |
848 | | |
849 | 66.0M | tmp = (FLOAT32)(ixheaacd_mult32X32float((FLOAT32)x1r, W4) + |
850 | 66.0M | ixheaacd_mult32X32float((FLOAT32)x1i, W1)); |
851 | 66.0M | x1i = (FLOAT32)(-ixheaacd_mult32X32float((FLOAT32)x1r, W1) + |
852 | 66.0M | ixheaacd_mult32X32float((FLOAT32)x1i, W4)); |
853 | 66.0M | x1r = tmp; |
854 | | |
855 | 66.0M | *ptr_xi = (x0r) - (x1r); |
856 | 66.0M | *(ptr_xi + 1) = (x0i) - (x1i); |
857 | 66.0M | ptr_z -= 32; |
858 | 66.0M | ptr_xi -= 32; |
859 | | |
860 | 66.0M | *ptr_xi = (x0r) + (x1r); |
861 | 66.0M | *(ptr_xi + 1) = (x0i) + (x1i); |
862 | 66.0M | ptr_z += 2; |
863 | 66.0M | ptr_xi += 2; |
864 | 66.0M | } |
865 | 8.25M | } |
866 | 8.25M | } |
867 | | |
868 | 21.3k | VOID ixheaacd_mps_complex_fft(FLOAT32 *xr, FLOAT32 *xi, WORD32 nlength) { |
869 | 21.3k | WORD32 i, j, k, n_stages, h2; |
870 | 21.3k | FLOAT32 x0r, x0i, x1r, x1i, x2r, x2i, x3r, x3i; |
871 | 21.3k | WORD32 del, nodespacing, in_loop_cnt; |
872 | 21.3k | WORD32 dig_rev_shift; |
873 | 21.3k | WORD32 not_power_4; |
874 | 21.3k | FLOAT32 ptr_x[256]; |
875 | 21.3k | FLOAT32 y[256]; |
876 | 21.3k | WORD32 npoints = nlength; |
877 | 21.3k | FLOAT32 *ptr_y = y; |
878 | 21.3k | const FLOAT32 *ptr_w; |
879 | 21.3k | dig_rev_shift = ixheaac_norm32(npoints) + 1 - 16; |
880 | 21.3k | n_stages = 30 - ixheaac_norm32(npoints); |
881 | 21.3k | not_power_4 = n_stages & 1; |
882 | | |
883 | 21.3k | n_stages = n_stages >> 1; |
884 | | |
885 | | |
886 | 1.38M | for (i = 0; i<nlength; i++) |
887 | 1.36M | { |
888 | 1.36M | ptr_x[2 * i] = xr[i]; |
889 | 1.36M | ptr_x[2 * i + 1] = xi[i]; |
890 | 1.36M | } |
891 | | |
892 | 21.3k | ptr_w = ixheaacd_twiddle_table_fft; |
893 | | |
894 | 363k | for (i = 0; i<npoints; i += 4) |
895 | 342k | { |
896 | 342k | FLOAT32 *inp = ptr_x; |
897 | | |
898 | 342k | DIG_REV(i, dig_rev_shift, h2); |
899 | 342k | if (not_power_4) |
900 | 0 | { |
901 | 0 | h2 += 1; |
902 | 0 | h2 &= ~1; |
903 | 0 | } |
904 | 342k | inp += (h2); |
905 | | |
906 | 342k | x0r = *inp; |
907 | 342k | x0i = *(inp + 1); |
908 | 342k | inp += (npoints >> 1); |
909 | | |
910 | 342k | x1r = *inp; |
911 | 342k | x1i = *(inp + 1); |
912 | 342k | inp += (npoints >> 1); |
913 | | |
914 | 342k | x2r = *inp; |
915 | 342k | x2i = *(inp + 1); |
916 | 342k | inp += (npoints >> 1); |
917 | | |
918 | 342k | x3r = *inp; |
919 | 342k | x3i = *(inp + 1); |
920 | | |
921 | 342k | x0r = x0r + x2r; |
922 | 342k | x0i = x0i + x2i; |
923 | 342k | x2r = x0r - (x2r * 2); |
924 | 342k | x2i = x0i - (x2i * 2); |
925 | 342k | x1r = x1r + x3r; |
926 | 342k | x1i = x1i + x3i; |
927 | 342k | x3r = x1r - (x3r * 2); |
928 | 342k | x3i = x1i - (x3i * 2); |
929 | | |
930 | 342k | x0r = x0r + x1r; |
931 | 342k | x0i = x0i + x1i; |
932 | 342k | x1r = x0r - (x1r * 2); |
933 | 342k | x1i = x0i - (x1i * 2); |
934 | 342k | x2r = x2r + x3i; |
935 | 342k | x2i = x2i - x3r; |
936 | 342k | x3i = x2r - (x3i * 2); |
937 | 342k | x3r = x2i + (x3r * 2); |
938 | | |
939 | 342k | *ptr_y++ = x0r; |
940 | 342k | *ptr_y++ = x0i; |
941 | 342k | *ptr_y++ = x2r; |
942 | 342k | *ptr_y++ = x2i; |
943 | 342k | *ptr_y++ = x1r; |
944 | 342k | *ptr_y++ = x1i; |
945 | 342k | *ptr_y++ = x3i; |
946 | 342k | *ptr_y++ = x3r; |
947 | 342k | } |
948 | 21.3k | ptr_y -= 2 * npoints; |
949 | 21.3k | del = 4; |
950 | 21.3k | nodespacing = 64; |
951 | 21.3k | in_loop_cnt = npoints >> 4; |
952 | 64.1k | for (i = n_stages - 1; i>0; i--) |
953 | 42.7k | { |
954 | 42.7k | const FLOAT32 *twiddles = ptr_w; |
955 | 42.7k | FLOAT32 *data = ptr_y; |
956 | 42.7k | FLOAT32 w1h, w2h, w3h, w1l, w2l, w3l; |
957 | 42.7k | WORD32 sec_loop_cnt; |
958 | | |
959 | 149k | for (k = in_loop_cnt; k != 0; k--) |
960 | 106k | { |
961 | 106k | x0r = (*data); |
962 | 106k | x0i = (*(data + 1)); |
963 | 106k | data += (del << 1); |
964 | | |
965 | 106k | x1r = (*data); |
966 | 106k | x1i = (*(data + 1)); |
967 | 106k | data += (del << 1); |
968 | | |
969 | 106k | x2r = (*data); |
970 | 106k | x2i = (*(data + 1)); |
971 | 106k | data += (del << 1); |
972 | | |
973 | 106k | x3r = (*data); |
974 | 106k | x3i = (*(data + 1)); |
975 | 106k | data -= 3 * (del << 1); |
976 | | |
977 | 106k | x0r = x0r + x2r; |
978 | 106k | x0i = x0i + x2i; |
979 | 106k | x2r = x0r - (x2r * 2); |
980 | 106k | x2i = x0i - (x2i * 2); |
981 | 106k | x1r = x1r + x3r; |
982 | 106k | x1i = x1i + x3i; |
983 | 106k | x3r = x1r - (x3r * 2); |
984 | 106k | x3i = x1i - (x3i * 2); |
985 | | |
986 | 106k | x0r = x0r + x1r; |
987 | 106k | x0i = x0i + x1i; |
988 | 106k | x1r = x0r - (x1r * 2); |
989 | 106k | x1i = x0i - (x1i * 2); |
990 | 106k | x2r = x2r + x3i; |
991 | 106k | x2i = x2i - x3r; |
992 | 106k | x3i = x2r - (x3i * 2); |
993 | 106k | x3r = x2i + (x3r * 2); |
994 | | |
995 | 106k | *data = x0r; |
996 | 106k | *(data + 1) = x0i; |
997 | 106k | data += (del << 1); |
998 | | |
999 | 106k | *data = x2r; |
1000 | 106k | *(data + 1) = x2i; |
1001 | 106k | data += (del << 1); |
1002 | | |
1003 | 106k | *data = x1r; |
1004 | 106k | *(data + 1) = x1i; |
1005 | 106k | data += (del << 1); |
1006 | | |
1007 | 106k | *data = x3i; |
1008 | 106k | *(data + 1) = x3r; |
1009 | 106k | data += (del << 1); |
1010 | 106k | } |
1011 | 42.7k | data = ptr_y + 2; |
1012 | | |
1013 | 42.7k | sec_loop_cnt = (nodespacing * del); |
1014 | 42.7k | sec_loop_cnt = (sec_loop_cnt / 4) + (sec_loop_cnt / 8) - (sec_loop_cnt / 16) \ |
1015 | 42.7k | + (sec_loop_cnt / 32) - (sec_loop_cnt / 64) + (sec_loop_cnt / 128) \ |
1016 | 42.7k | - (sec_loop_cnt / 256); |
1017 | 42.7k | j = nodespacing; |
1018 | | |
1019 | 171k | for (j = nodespacing; j <= sec_loop_cnt; j += nodespacing) |
1020 | 128k | { |
1021 | 128k | w1h = *(twiddles + 2 * j); |
1022 | 128k | w1l = *(twiddles + 2 * j + 1); |
1023 | 128k | w2h = *(twiddles + 2 * (j << 1)); |
1024 | 128k | w2l = *(twiddles + 2 * (j << 1) + 1); |
1025 | 128k | w3h = *(twiddles + 2 * j + 2 * (j << 1)); |
1026 | 128k | w3l = *(twiddles + 2 * j + 2 * (j << 1) + 1); |
1027 | | |
1028 | 320k | for (k = in_loop_cnt; k != 0; k--) |
1029 | 192k | { |
1030 | 192k | FLOAT32 tmp; |
1031 | 192k | FLOAT32 x0r, x0i, x1r, x1i, x2r, x2i, x3r, x3i; |
1032 | | |
1033 | 192k | data += (del << 1); |
1034 | | |
1035 | 192k | x1r = *data; |
1036 | 192k | x1i = *(data + 1); |
1037 | 192k | data += (del << 1); |
1038 | | |
1039 | 192k | x2r = *data; |
1040 | 192k | x2i = *(data + 1); |
1041 | 192k | data += (del << 1); |
1042 | | |
1043 | 192k | x3r = *data; |
1044 | 192k | x3i = *(data + 1); |
1045 | 192k | data -= 3 * (del << 1); |
1046 | | |
1047 | 192k | tmp = (ixheaacd_mult32X32float(x1r, w1l) - ixheaacd_mult32X32float(x1i, w1h)); |
1048 | 192k | x1i = ixheaacd_mac32X32float(ixheaacd_mult32X32float(x1r, w1h), x1i, w1l); |
1049 | 192k | x1r = tmp; |
1050 | | |
1051 | 192k | tmp = (ixheaacd_mult32X32float(x2r, w2l) - ixheaacd_mult32X32float(x2i, w2h)); |
1052 | 192k | x2i = ixheaacd_mac32X32float(ixheaacd_mult32X32float(x2r, w2h), x2i, w2l); |
1053 | 192k | x2r = tmp; |
1054 | | |
1055 | 192k | tmp = (ixheaacd_mult32X32float(x3r, w3l) - ixheaacd_mult32X32float(x3i, w3h)); |
1056 | 192k | x3i = ixheaacd_mac32X32float(ixheaacd_mult32X32float(x3r, w3h), x3i, w3l); |
1057 | 192k | x3r = tmp; |
1058 | | |
1059 | 192k | x0r = (*data); |
1060 | 192k | x0i = (*(data + 1)); |
1061 | | |
1062 | 192k | x0r = x0r + (x2r); |
1063 | 192k | x0i = x0i + (x2i); |
1064 | 192k | x2r = x0r - (x2r * 2); |
1065 | 192k | x2i = x0i - (x2i * 2); |
1066 | 192k | x1r = x1r + x3r; |
1067 | 192k | x1i = x1i + x3i; |
1068 | 192k | x3r = x1r - (x3r * 2); |
1069 | 192k | x3i = x1i - (x3i * 2); |
1070 | | |
1071 | 192k | x0r = x0r + (x1r); |
1072 | 192k | x0i = x0i + (x1i); |
1073 | 192k | x1r = x0r - (x1r * 2); |
1074 | 192k | x1i = x0i - (x1i * 2); |
1075 | 192k | x2r = x2r + (x3i); |
1076 | 192k | x2i = x2i - (x3r); |
1077 | 192k | x3i = x2r - (x3i * 2); |
1078 | 192k | x3r = x2i + (x3r * 2); |
1079 | | |
1080 | 192k | *data = x0r; |
1081 | 192k | *(data + 1) = x0i; |
1082 | 192k | data += (del << 1); |
1083 | | |
1084 | 192k | *data = x2r; |
1085 | 192k | *(data + 1) = x2i; |
1086 | 192k | data += (del << 1); |
1087 | | |
1088 | 192k | *data = x1r; |
1089 | 192k | *(data + 1) = x1i; |
1090 | 192k | data += (del << 1); |
1091 | | |
1092 | 192k | *data = x3i; |
1093 | 192k | *(data + 1) = x3r; |
1094 | 192k | data += (del << 1); |
1095 | 192k | } |
1096 | 128k | data -= 2 * npoints; |
1097 | 128k | data += 2; |
1098 | 128k | } |
1099 | 128k | for (; j <= (nodespacing * del) >> 1; j += nodespacing) |
1100 | 85.5k | { |
1101 | 85.5k | w1h = *(twiddles + 2 * j); |
1102 | 85.5k | w2h = *(twiddles + 2 * (j << 1)); |
1103 | 85.5k | w3h = *(twiddles + 2 * j + 2 * (j << 1) - 512); |
1104 | 85.5k | w1l = *(twiddles + 2 * j + 1); |
1105 | 85.5k | w2l = *(twiddles + 2 * (j << 1) + 1); |
1106 | 85.5k | w3l = *(twiddles + 2 * j + 2 * (j << 1) - 511); |
1107 | | |
1108 | 235k | for (k = in_loop_cnt; k != 0; k--) |
1109 | 149k | { |
1110 | 149k | FLOAT32 tmp; |
1111 | 149k | FLOAT32 x0r, x0i, x1r, x1i, x2r, x2i, x3r, x3i; |
1112 | | |
1113 | 149k | data += (del << 1); |
1114 | | |
1115 | 149k | x1r = *data; |
1116 | 149k | x1i = *(data + 1); |
1117 | 149k | data += (del << 1); |
1118 | | |
1119 | 149k | x2r = *data; |
1120 | 149k | x2i = *(data + 1); |
1121 | 149k | data += (del << 1); |
1122 | | |
1123 | 149k | x3r = *data; |
1124 | 149k | x3i = *(data + 1); |
1125 | 149k | data -= 3 * (del << 1); |
1126 | | |
1127 | 149k | tmp = (ixheaacd_mult32X32float(x1r, w1l) - ixheaacd_mult32X32float(x1i, w1h)); |
1128 | 149k | x1i = ixheaacd_mac32X32float(ixheaacd_mult32X32float(x1r, w1h), x1i, w1l); |
1129 | 149k | x1r = tmp; |
1130 | | |
1131 | 149k | tmp = (ixheaacd_mult32X32float(x2r, w2l) - ixheaacd_mult32X32float(x2i, w2h)); |
1132 | 149k | x2i = ixheaacd_mac32X32float(ixheaacd_mult32X32float(x2r, w2h), x2i, w2l); |
1133 | 149k | x2r = tmp; |
1134 | | |
1135 | 149k | tmp = (ixheaacd_mult32X32float(x3r, w3h) + ixheaacd_mult32X32float(x3i, w3l)); |
1136 | 149k | x3i = -ixheaacd_mult32X32float(x3r, w3l) + ixheaacd_mult32X32float(x3i, w3h); |
1137 | 149k | x3r = tmp; |
1138 | | |
1139 | 149k | x0r = (*data); |
1140 | 149k | x0i = (*(data + 1)); |
1141 | | |
1142 | 149k | x0r = x0r + (x2r); |
1143 | 149k | x0i = x0i + (x2i); |
1144 | 149k | x2r = x0r - (x2r * 2); |
1145 | 149k | x2i = x0i - (x2i * 2); |
1146 | 149k | x1r = x1r + x3r; |
1147 | 149k | x1i = x1i + x3i; |
1148 | 149k | x3r = x1r - (x3r * 2); |
1149 | 149k | x3i = x1i - (x3i * 2); |
1150 | | |
1151 | 149k | x0r = x0r + (x1r); |
1152 | 149k | x0i = x0i + (x1i); |
1153 | 149k | x1r = x0r - (x1r * 2); |
1154 | 149k | x1i = x0i - (x1i * 2); |
1155 | 149k | x2r = x2r + (x3i); |
1156 | 149k | x2i = x2i - (x3r); |
1157 | 149k | x3i = x2r - (x3i * 2); |
1158 | 149k | x3r = x2i + (x3r * 2); |
1159 | | |
1160 | 149k | *data = x0r; |
1161 | 149k | *(data + 1) = x0i; |
1162 | 149k | data += (del << 1); |
1163 | | |
1164 | 149k | *data = x2r; |
1165 | 149k | *(data + 1) = x2i; |
1166 | 149k | data += (del << 1); |
1167 | | |
1168 | 149k | *data = x1r; |
1169 | 149k | *(data + 1) = x1i; |
1170 | 149k | data += (del << 1); |
1171 | | |
1172 | 149k | *data = x3i; |
1173 | 149k | *(data + 1) = x3r; |
1174 | 149k | data += (del << 1); |
1175 | 149k | } |
1176 | 85.5k | data -= 2 * npoints; |
1177 | 85.5k | data += 2; |
1178 | 85.5k | } |
1179 | 85.5k | for (; j <= sec_loop_cnt * 2; j += nodespacing) |
1180 | 42.7k | { |
1181 | 42.7k | w1h = *(twiddles + 2 * j); |
1182 | 42.7k | w2h = *(twiddles + 2 * (j << 1) - 512); |
1183 | 42.7k | w3h = *(twiddles + 2 * j + 2 * (j << 1) - 512); |
1184 | 42.7k | w1l = *(twiddles + 2 * j + 1); |
1185 | 42.7k | w2l = *(twiddles + 2 * (j << 1) - 511); |
1186 | 42.7k | w3l = *(twiddles + 2 * j + 2 * (j << 1) - 511); |
1187 | | |
1188 | 85.5k | for (k = in_loop_cnt; k != 0; k--) |
1189 | 42.7k | { |
1190 | 42.7k | FLOAT32 tmp; |
1191 | 42.7k | FLOAT32 x0r, x0i, x1r, x1i, x2r, x2i, x3r, x3i; |
1192 | | |
1193 | 42.7k | data += (del << 1); |
1194 | | |
1195 | 42.7k | x1r = *data; |
1196 | 42.7k | x1i = *(data + 1); |
1197 | 42.7k | data += (del << 1); |
1198 | | |
1199 | 42.7k | x2r = *data; |
1200 | 42.7k | x2i = *(data + 1); |
1201 | 42.7k | data += (del << 1); |
1202 | | |
1203 | 42.7k | x3r = *data; |
1204 | 42.7k | x3i = *(data + 1); |
1205 | 42.7k | data -= 3 * (del << 1); |
1206 | | |
1207 | 42.7k | tmp = (ixheaacd_mult32X32float(x1r, w1l) - ixheaacd_mult32X32float(x1i, w1h)); |
1208 | 42.7k | x1i = ixheaacd_mac32X32float(ixheaacd_mult32X32float(x1r, w1h), x1i, w1l); |
1209 | 42.7k | x1r = tmp; |
1210 | | |
1211 | 42.7k | tmp = (ixheaacd_mult32X32float(x2r, w2h) + ixheaacd_mult32X32float(x2i, w2l)); |
1212 | 42.7k | x2i = -ixheaacd_mult32X32float(x2r, w2l) + ixheaacd_mult32X32float(x2i, w2h); |
1213 | 42.7k | x2r = tmp; |
1214 | | |
1215 | 42.7k | tmp = (ixheaacd_mult32X32float(x3r, w3h) + ixheaacd_mult32X32float(x3i, w3l)); |
1216 | 42.7k | x3i = -ixheaacd_mult32X32float(x3r, w3l) + ixheaacd_mult32X32float(x3i, w3h); |
1217 | 42.7k | x3r = tmp; |
1218 | | |
1219 | 42.7k | x0r = (*data); |
1220 | 42.7k | x0i = (*(data + 1)); |
1221 | | |
1222 | 42.7k | x0r = x0r + (x2r); |
1223 | 42.7k | x0i = x0i + (x2i); |
1224 | 42.7k | x2r = x0r - (x2r * 2); |
1225 | 42.7k | x2i = x0i - (x2i * 2); |
1226 | 42.7k | x1r = x1r + x3r; |
1227 | 42.7k | x1i = x1i + x3i; |
1228 | 42.7k | x3r = x1r - (x3r * 2); |
1229 | 42.7k | x3i = x1i - (x3i * 2); |
1230 | | |
1231 | 42.7k | x0r = x0r + (x1r); |
1232 | 42.7k | x0i = x0i + (x1i); |
1233 | 42.7k | x1r = x0r - (x1r * 2); |
1234 | 42.7k | x1i = x0i - (x1i * 2); |
1235 | 42.7k | x2r = x2r + (x3i); |
1236 | 42.7k | x2i = x2i - (x3r); |
1237 | 42.7k | x3i = x2r - (x3i * 2); |
1238 | 42.7k | x3r = x2i + (x3r * 2); |
1239 | | |
1240 | 42.7k | *data = x0r; |
1241 | 42.7k | *(data + 1) = x0i; |
1242 | 42.7k | data += (del << 1); |
1243 | | |
1244 | 42.7k | *data = x2r; |
1245 | 42.7k | *(data + 1) = x2i; |
1246 | 42.7k | data += (del << 1); |
1247 | | |
1248 | 42.7k | *data = x1r; |
1249 | 42.7k | *(data + 1) = x1i; |
1250 | 42.7k | data += (del << 1); |
1251 | | |
1252 | 42.7k | *data = x3i; |
1253 | 42.7k | *(data + 1) = x3r; |
1254 | 42.7k | data += (del << 1); |
1255 | 42.7k | } |
1256 | 42.7k | data -= 2 * npoints; |
1257 | 42.7k | data += 2; |
1258 | 42.7k | } |
1259 | 171k | for (; j<nodespacing * del; j += nodespacing) |
1260 | 128k | { |
1261 | 128k | w1h = *(twiddles + 2 * j); |
1262 | 128k | w2h = *(twiddles + 2 * (j << 1) - 512); |
1263 | 128k | w3h = *(twiddles + 2 * j + 2 * (j << 1) - 1024); |
1264 | 128k | w1l = *(twiddles + 2 * j + 1); |
1265 | 128k | w2l = *(twiddles + 2 * (j << 1) - 511); |
1266 | 128k | w3l = *(twiddles + 2 * j + 2 * (j << 1) - 1023); |
1267 | | |
1268 | 320k | for (k = in_loop_cnt; k != 0; k--) |
1269 | 192k | { |
1270 | 192k | FLOAT32 tmp; |
1271 | 192k | FLOAT32 x0r, x0i, x1r, x1i, x2r, x2i, x3r, x3i; |
1272 | | |
1273 | 192k | data += (del << 1); |
1274 | | |
1275 | 192k | x1r = *data; |
1276 | 192k | x1i = *(data + 1); |
1277 | 192k | data += (del << 1); |
1278 | | |
1279 | 192k | x2r = *data; |
1280 | 192k | x2i = *(data + 1); |
1281 | 192k | data += (del << 1); |
1282 | | |
1283 | 192k | x3r = *data; |
1284 | 192k | x3i = *(data + 1); |
1285 | 192k | data -= 3 * (del << 1); |
1286 | | |
1287 | 192k | tmp = (ixheaacd_mult32X32float(x1r, w1l) - ixheaacd_mult32X32float(x1i, w1h)); |
1288 | 192k | x1i = ixheaacd_mac32X32float(ixheaacd_mult32X32float(x1r, w1h), x1i, w1l); |
1289 | 192k | x1r = tmp; |
1290 | | |
1291 | 192k | tmp = (ixheaacd_mult32X32float(x2r, w2h) + ixheaacd_mult32X32float(x2i, w2l)); |
1292 | 192k | x2i = -ixheaacd_mult32X32float(x2r, w2l) + ixheaacd_mult32X32float(x2i, w2h); |
1293 | 192k | x2r = tmp; |
1294 | | |
1295 | 192k | tmp = (-ixheaacd_mult32X32float(x3r, w3l) + ixheaacd_mult32X32float(x3i, w3h)); |
1296 | 192k | x3i = ixheaacd_mac32X32float(ixheaacd_mult32X32float(x3r, w3h), x3i, w3l); |
1297 | 192k | x3r = tmp; |
1298 | | |
1299 | 192k | x0r = (*data); |
1300 | 192k | x0i = (*(data + 1)); |
1301 | | |
1302 | 192k | x0r = x0r + (x2r); |
1303 | 192k | x0i = x0i + (x2i); |
1304 | 192k | x2r = x0r - (x2r * 2); |
1305 | 192k | x2i = x0i - (x2i * 2); |
1306 | 192k | x1r = x1r + x3r; |
1307 | 192k | x1i = x1i - x3i; |
1308 | 192k | x3r = x1r - (x3r * 2); |
1309 | 192k | x3i = x1i + (x3i * 2); |
1310 | | |
1311 | 192k | x0r = x0r + (x1r); |
1312 | 192k | x0i = x0i + (x1i); |
1313 | 192k | x1r = x0r - (x1r * 2); |
1314 | 192k | x1i = x0i - (x1i * 2); |
1315 | 192k | x2r = x2r + (x3i); |
1316 | 192k | x2i = x2i - (x3r); |
1317 | 192k | x3i = x2r - (x3i * 2); |
1318 | 192k | x3r = x2i + (x3r * 2); |
1319 | | |
1320 | 192k | *data = x0r; |
1321 | 192k | *(data + 1) = x0i; |
1322 | 192k | data += (del << 1); |
1323 | | |
1324 | 192k | *data = x2r; |
1325 | 192k | *(data + 1) = x2i; |
1326 | 192k | data += (del << 1); |
1327 | | |
1328 | 192k | *data = x1r; |
1329 | 192k | *(data + 1) = x1i; |
1330 | 192k | data += (del << 1); |
1331 | | |
1332 | 192k | *data = x3i; |
1333 | 192k | *(data + 1) = x3r; |
1334 | 192k | data += (del << 1); |
1335 | 192k | } |
1336 | 128k | data -= 2 * npoints; |
1337 | 128k | data += 2; |
1338 | 128k | } |
1339 | 42.7k | nodespacing >>= 2; |
1340 | 42.7k | del <<= 2; |
1341 | 42.7k | in_loop_cnt >>= 2; |
1342 | 42.7k | } |
1343 | 21.3k | if (not_power_4) |
1344 | 0 | { |
1345 | 0 | const FLOAT32 *twiddles = ptr_w; |
1346 | 0 | nodespacing <<= 1; |
1347 | |
|
1348 | 0 | for (j = del / 2; j != 0; j--) |
1349 | 0 | { |
1350 | 0 | FLOAT32 w1h = *twiddles; |
1351 | 0 | FLOAT32 w1l = *(twiddles + 1); |
1352 | 0 | FLOAT32 tmp; |
1353 | 0 | twiddles += nodespacing * 2; |
1354 | |
|
1355 | 0 | x0r = *ptr_y; |
1356 | 0 | x0i = *(ptr_y + 1); |
1357 | 0 | ptr_y += (del << 1); |
1358 | |
|
1359 | 0 | x1r = *ptr_y; |
1360 | 0 | x1i = *(ptr_y + 1); |
1361 | |
|
1362 | 0 | tmp = (ixheaacd_mult32X32float(x1r, w1l) - ixheaacd_mult32X32float(x1i, w1h)); |
1363 | 0 | x1i = ixheaacd_mac32X32float(ixheaacd_mult32X32float(x1r, w1h), x1i, w1l); |
1364 | 0 | x1r = tmp; |
1365 | |
|
1366 | 0 | *ptr_y = (x0r) - (x1r); |
1367 | 0 | *(ptr_y + 1) = (x0i) - (x1i); |
1368 | 0 | ptr_y -= (del << 1); |
1369 | |
|
1370 | 0 | *ptr_y = (x0r) + (x1r); |
1371 | 0 | *(ptr_y + 1) = (x0i) + (x1i); |
1372 | 0 | ptr_y += 2; |
1373 | 0 | } |
1374 | 0 | twiddles = ptr_w; |
1375 | 0 | for (j = del / 2; j != 0; j--) |
1376 | 0 | { |
1377 | 0 | FLOAT32 w1h = *twiddles; |
1378 | 0 | FLOAT32 w1l = *(twiddles + 1); |
1379 | 0 | FLOAT32 tmp; |
1380 | 0 | twiddles += nodespacing * 2; |
1381 | |
|
1382 | 0 | x0r = *ptr_y; |
1383 | 0 | x0i = *(ptr_y + 1); |
1384 | 0 | ptr_y += (del << 1); |
1385 | |
|
1386 | 0 | x1r = *ptr_y; |
1387 | 0 | x1i = *(ptr_y + 1); |
1388 | |
|
1389 | 0 | tmp = (ixheaacd_mult32X32float(x1r, w1h) + ixheaacd_mult32X32float(x1i, w1l)); |
1390 | 0 | x1i = -ixheaacd_mult32X32float(x1r, w1l) + ixheaacd_mult32X32float(x1i, w1h); |
1391 | 0 | x1r = tmp; |
1392 | |
|
1393 | 0 | *ptr_y = (x0r) - (x1r); |
1394 | 0 | *(ptr_y + 1) = (x0i) - (x1i); |
1395 | 0 | ptr_y -= (del << 1); |
1396 | |
|
1397 | 0 | *ptr_y = (x0r) + (x1r); |
1398 | 0 | *(ptr_y + 1) = (x0i) + (x1i); |
1399 | 0 | ptr_y += 2; |
1400 | 0 | } |
1401 | 0 | } |
1402 | | |
1403 | 1.38M | for (i = 0; i<nlength; i++) |
1404 | 1.36M | { |
1405 | 1.36M | xr[i] = y[2 * i]; |
1406 | 1.36M | xi[i] = y[2 * i + 1]; |
1407 | 1.36M | } |
1408 | | |
1409 | 21.3k | return; |
1410 | 21.3k | } |
1411 | | |
1412 | | VOID ixheaacd_complex_fft_p2_dec(WORD32 *xr, WORD32 *xi, WORD32 nlength, |
1413 | 13.2M | WORD32 fft_mode, WORD32 *preshift) { |
1414 | 13.2M | WORD32 i, j, k, n_stages; |
1415 | 13.2M | WORD32 h2, x0r, x0i, x1r, x1i, x2r, x2i, x3r, x3i; |
1416 | 13.2M | WORD32 del, nodespacing, in_loop_cnt; |
1417 | 13.2M | WORD32 not_power_4; |
1418 | 13.2M | WORD32 npts, shift; |
1419 | 13.2M | WORD32 dig_rev_shift; |
1420 | 13.2M | WORD32 ptr_x[1024]; |
1421 | 13.2M | WORD32 y[1024]; |
1422 | 13.2M | WORD32 npoints = nlength; |
1423 | 13.2M | WORD32 n = 0; |
1424 | 13.2M | WORD32 *ptr_y = y; |
1425 | 13.2M | const WORD32 *ptr_w; |
1426 | 13.2M | dig_rev_shift = ixheaac_norm32(npoints) + 1 - 16; |
1427 | 13.2M | n_stages = 30 - ixheaac_norm32(npoints); |
1428 | 13.2M | not_power_4 = n_stages & 1; |
1429 | | |
1430 | 13.2M | n_stages = n_stages >> 1; |
1431 | | |
1432 | 13.2M | npts = npoints; |
1433 | 52.9M | while (npts >> 1) { |
1434 | 39.7M | n++; |
1435 | 39.7M | npts = npts >> 1; |
1436 | 39.7M | } |
1437 | | |
1438 | 13.2M | if (n % 2 == 0) |
1439 | 8.39M | shift = ((n + 4)) / 2; |
1440 | 4.85M | else |
1441 | 4.85M | shift = ((n + 3) / 2); |
1442 | | |
1443 | 315M | for (i = 0; i < nlength; i++) { |
1444 | 302M | ptr_x[2 * i] = (xr[i] / (1 << (shift))); |
1445 | 302M | ptr_x[2 * i + 1] = (xi[i] / (1 << (shift))); |
1446 | 302M | } |
1447 | | |
1448 | 13.2M | if (fft_mode == -1) { |
1449 | 12.0M | ptr_w = ixheaacd_twiddle_table_fft_32x32; |
1450 | | |
1451 | 43.0M | for (i = 0; i < npoints; i += 4) { |
1452 | 31.0M | WORD32 *inp = ptr_x; |
1453 | | |
1454 | 31.0M | DIG_REV(i, dig_rev_shift, h2); |
1455 | 31.0M | if (not_power_4) { |
1456 | 18.2M | h2 += 1; |
1457 | 18.2M | h2 &= ~1; |
1458 | 18.2M | } |
1459 | 31.0M | inp += (h2); |
1460 | | |
1461 | 31.0M | x0r = *inp; |
1462 | 31.0M | x0i = *(inp + 1); |
1463 | 31.0M | inp += (npoints >> 1); |
1464 | | |
1465 | 31.0M | x1r = *inp; |
1466 | 31.0M | x1i = *(inp + 1); |
1467 | 31.0M | inp += (npoints >> 1); |
1468 | | |
1469 | 31.0M | x2r = *inp; |
1470 | 31.0M | x2i = *(inp + 1); |
1471 | 31.0M | inp += (npoints >> 1); |
1472 | | |
1473 | 31.0M | x3r = *inp; |
1474 | 31.0M | x3i = *(inp + 1); |
1475 | | |
1476 | 31.0M | x0r = ixheaac_add32_sat(x0r, x2r); |
1477 | 31.0M | x0i = ixheaac_add32_sat(x0i, x2i); |
1478 | 31.0M | x2r = ixheaac_sub32_sat(x0r, ixheaac_shl32_sat(x2r, 1)); |
1479 | 31.0M | x2i = ixheaac_sub32_sat(x0i, ixheaac_shl32_sat(x2i, 1)); |
1480 | 31.0M | x1r = ixheaac_add32_sat(x1r, x3r); |
1481 | 31.0M | x1i = ixheaac_add32_sat(x1i, x3i); |
1482 | 31.0M | x3r = ixheaac_sub32_sat(x1r, ixheaac_shl32_sat(x3r, 1)); |
1483 | 31.0M | x3i = ixheaac_sub32_sat(x1i, ixheaac_shl32_sat(x3i, 1)); |
1484 | | |
1485 | 31.0M | x0r = ixheaac_add32_sat(x0r, x1r); |
1486 | 31.0M | x0i = ixheaac_add32_sat(x0i, x1i); |
1487 | 31.0M | x1r = ixheaac_sub32_sat(x0r, ixheaac_shl32_sat(x1r, 1)); |
1488 | 31.0M | x1i = ixheaac_sub32_sat(x0i, ixheaac_shl32_sat(x1i, 1)); |
1489 | 31.0M | x2r = ixheaac_add32_sat(x2r, x3i); |
1490 | 31.0M | x2i = ixheaac_sub32_sat(x2i, x3r); |
1491 | 31.0M | x3i = ixheaac_sub32_sat(x2r, ixheaac_shl32_sat(x3i, 1)); |
1492 | 31.0M | x3r = ixheaac_add32_sat(x2i, ixheaac_shl32_sat(x3r, 1)); |
1493 | | |
1494 | 31.0M | *ptr_y++ = x0r; |
1495 | 31.0M | *ptr_y++ = x0i; |
1496 | 31.0M | *ptr_y++ = x2r; |
1497 | 31.0M | *ptr_y++ = x2i; |
1498 | 31.0M | *ptr_y++ = x1r; |
1499 | 31.0M | *ptr_y++ = x1i; |
1500 | 31.0M | *ptr_y++ = x3i; |
1501 | 31.0M | *ptr_y++ = x3r; |
1502 | 31.0M | } |
1503 | 12.0M | ptr_y -= 2 * npoints; |
1504 | 12.0M | del = 4; |
1505 | 12.0M | nodespacing = 64; |
1506 | 12.0M | in_loop_cnt = npoints >> 4; |
1507 | 13.6M | for (i = n_stages - 1; i > 0; i--) { |
1508 | 1.64M | const WORD32 *twiddles = ptr_w; |
1509 | 1.64M | WORD32 *data = ptr_y; |
1510 | 1.64M | WORD32 w1h, w2h, w3h, w1l, w2l, w3l; |
1511 | 1.64M | WORD32 sec_loop_cnt; |
1512 | | |
1513 | 6.45M | for (k = in_loop_cnt; k != 0; k--) { |
1514 | 4.81M | x0r = (*data); |
1515 | 4.81M | x0i = (*(data + 1)); |
1516 | 4.81M | data += (del << 1); |
1517 | | |
1518 | 4.81M | x1r = (*data); |
1519 | 4.81M | x1i = (*(data + 1)); |
1520 | 4.81M | data += (del << 1); |
1521 | | |
1522 | 4.81M | x2r = (*data); |
1523 | 4.81M | x2i = (*(data + 1)); |
1524 | 4.81M | data += (del << 1); |
1525 | | |
1526 | 4.81M | x3r = (*data); |
1527 | 4.81M | x3i = (*(data + 1)); |
1528 | 4.81M | data -= 3 * (del << 1); |
1529 | | |
1530 | 4.81M | x0r = ixheaac_add32_sat(x0r, x2r); |
1531 | 4.81M | x0i = ixheaac_add32_sat(x0i, x2i); |
1532 | 4.81M | x2r = ixheaac_sub32_sat(x0r, ixheaac_shl32_sat(x2r, 1)); |
1533 | 4.81M | x2i = ixheaac_sub32_sat(x0i, ixheaac_shl32_sat(x2i, 1)); |
1534 | 4.81M | x1r = ixheaac_add32_sat(x1r, x3r); |
1535 | 4.81M | x1i = ixheaac_add32_sat(x1i, x3i); |
1536 | 4.81M | x3r = ixheaac_sub32_sat(x1r, ixheaac_shl32_sat(x3r, 1)); |
1537 | 4.81M | x3i = ixheaac_sub32_sat(x1i, ixheaac_shl32_sat(x3i, 1)); |
1538 | | |
1539 | 4.81M | x0r = ixheaac_add32_sat(x0r, x1r); |
1540 | 4.81M | x0i = ixheaac_add32_sat(x0i, x1i); |
1541 | 4.81M | x1r = ixheaac_sub32_sat(x0r, ixheaac_shl32_sat(x1r, 1)); |
1542 | 4.81M | x1i = ixheaac_sub32_sat(x0i, ixheaac_shl32_sat(x1i, 1)); |
1543 | 4.81M | x2r = ixheaac_add32_sat(x2r, x3i); |
1544 | 4.81M | x2i = ixheaac_sub32_sat(x2i, x3r); |
1545 | 4.81M | x3i = ixheaac_sub32_sat(x2r, ixheaac_shl32_sat(x3i, 1)); |
1546 | 4.81M | x3r = ixheaac_add32_sat(x2i, ixheaac_shl32_sat(x3r, 1)); |
1547 | | |
1548 | 4.81M | *data = x0r; |
1549 | 4.81M | *(data + 1) = x0i; |
1550 | 4.81M | data += (del << 1); |
1551 | | |
1552 | 4.81M | *data = x2r; |
1553 | 4.81M | *(data + 1) = x2i; |
1554 | 4.81M | data += (del << 1); |
1555 | | |
1556 | 4.81M | *data = x1r; |
1557 | 4.81M | *(data + 1) = x1i; |
1558 | 4.81M | data += (del << 1); |
1559 | | |
1560 | 4.81M | *data = x3i; |
1561 | 4.81M | *(data + 1) = x3r; |
1562 | 4.81M | data += (del << 1); |
1563 | 4.81M | } |
1564 | 1.64M | data = ptr_y + 2; |
1565 | | |
1566 | 1.64M | sec_loop_cnt = (nodespacing * del); |
1567 | 1.64M | sec_loop_cnt = (sec_loop_cnt / 4) + (sec_loop_cnt / 8) - |
1568 | 1.64M | (sec_loop_cnt / 16) + (sec_loop_cnt / 32) - |
1569 | 1.64M | (sec_loop_cnt / 64) + (sec_loop_cnt / 128) - |
1570 | 1.64M | (sec_loop_cnt / 256); |
1571 | 1.64M | j = nodespacing; |
1572 | | |
1573 | 5.47M | for (j = nodespacing; j <= sec_loop_cnt; j += nodespacing) { |
1574 | 3.83M | w1h = *(twiddles + 2 * j); |
1575 | 3.83M | w1l = *(twiddles + 2 * j + 1); |
1576 | 3.83M | w2h = *(twiddles + 2 * (j << 1)); |
1577 | 3.83M | w2l = *(twiddles + 2 * (j << 1) + 1); |
1578 | 3.83M | w3h = *(twiddles + 2 * j + 2 * (j << 1)); |
1579 | 3.83M | w3l = *(twiddles + 2 * j + 2 * (j << 1) + 1); |
1580 | | |
1581 | 11.8M | for (k = in_loop_cnt; k != 0; k--) { |
1582 | 8.03M | WORD32 tmp; |
1583 | 8.03M | WORD32 x0r, x0i, x1r, x1i, x2r, x2i, x3r, x3i; |
1584 | | |
1585 | 8.03M | data += (del << 1); |
1586 | | |
1587 | 8.03M | x1r = *data; |
1588 | 8.03M | x1i = *(data + 1); |
1589 | 8.03M | data += (del << 1); |
1590 | | |
1591 | 8.03M | x2r = *data; |
1592 | 8.03M | x2i = *(data + 1); |
1593 | 8.03M | data += (del << 1); |
1594 | | |
1595 | 8.03M | x3r = *data; |
1596 | 8.03M | x3i = *(data + 1); |
1597 | 8.03M | data -= 3 * (del << 1); |
1598 | | |
1599 | 8.03M | tmp = ixheaac_sub32_sat(ixheaacd_mult32_sat(x1r, w1l), |
1600 | 8.03M | ixheaacd_mult32_sat(x1i, w1h)); |
1601 | 8.03M | x1i = ixheaacd_mac32_sat(ixheaacd_mult32_sat(x1r, w1h), x1i, w1l); |
1602 | 8.03M | x1r = tmp; |
1603 | | |
1604 | 8.03M | tmp = ixheaac_sub32_sat(ixheaacd_mult32_sat(x2r, w2l), |
1605 | 8.03M | ixheaacd_mult32_sat(x2i, w2h)); |
1606 | 8.03M | x2i = ixheaacd_mac32_sat(ixheaacd_mult32_sat(x2r, w2h), x2i, w2l); |
1607 | 8.03M | x2r = tmp; |
1608 | | |
1609 | 8.03M | tmp = ixheaac_sub32_sat(ixheaacd_mult32_sat(x3r, w3l), |
1610 | 8.03M | ixheaacd_mult32_sat(x3i, w3h)); |
1611 | 8.03M | x3i = ixheaacd_mac32_sat(ixheaacd_mult32_sat(x3r, w3h), x3i, w3l); |
1612 | 8.03M | x3r = tmp; |
1613 | | |
1614 | 8.03M | x0r = (*data); |
1615 | 8.03M | x0i = (*(data + 1)); |
1616 | | |
1617 | 8.03M | x0r = ixheaac_add32_sat(x0r, x2r); |
1618 | 8.03M | x0i = ixheaac_add32_sat(x0i, x2i); |
1619 | 8.03M | x2r = ixheaac_sub32_sat(x0r, ixheaac_shl32_sat(x2r, 1)); |
1620 | 8.03M | x2i = ixheaac_sub32_sat(x0i, ixheaac_shl32_sat(x2i, 1)); |
1621 | 8.03M | x1r = ixheaac_add32_sat(x1r, x3r); |
1622 | 8.03M | x1i = ixheaac_add32_sat(x1i, x3i); |
1623 | 8.03M | x3r = ixheaac_sub32_sat(x1r, ixheaac_shl32_sat(x3r, 1)); |
1624 | 8.03M | x3i = ixheaac_sub32_sat(x1i, ixheaac_shl32_sat(x3i, 1)); |
1625 | | |
1626 | 8.03M | x0r = ixheaac_add32_sat(x0r, x1r); |
1627 | 8.03M | x0i = ixheaac_add32_sat(x0i, x1i); |
1628 | 8.03M | x1r = ixheaac_sub32_sat(x0r, ixheaac_shl32_sat(x1r, 1)); |
1629 | 8.03M | x1i = ixheaac_sub32_sat(x0i, ixheaac_shl32_sat(x1i, 1)); |
1630 | 8.03M | x2r = ixheaac_add32_sat(x2r, x3i); |
1631 | 8.03M | x2i = ixheaac_sub32_sat(x2i, x3r); |
1632 | 8.03M | x3i = ixheaac_sub32_sat(x2r, ixheaac_shl32_sat(x3i, 1)); |
1633 | 8.03M | x3r = ixheaac_add32_sat(x2i, ixheaac_shl32_sat(x3r, 1)); |
1634 | | |
1635 | 8.03M | *data = x0r; |
1636 | 8.03M | *(data + 1) = x0i; |
1637 | 8.03M | data += (del << 1); |
1638 | | |
1639 | 8.03M | *data = x2r; |
1640 | 8.03M | *(data + 1) = x2i; |
1641 | 8.03M | data += (del << 1); |
1642 | | |
1643 | 8.03M | *data = x1r; |
1644 | 8.03M | *(data + 1) = x1i; |
1645 | 8.03M | data += (del << 1); |
1646 | | |
1647 | 8.03M | *data = x3i; |
1648 | 8.03M | *(data + 1) = x3r; |
1649 | 8.03M | data += (del << 1); |
1650 | 8.03M | } |
1651 | 3.83M | data -= 2 * npoints; |
1652 | 3.83M | data += 2; |
1653 | 3.83M | } |
1654 | 4.38M | for (; j <= (nodespacing * del) >> 1; j += nodespacing) { |
1655 | 2.73M | w1h = *(twiddles + 2 * j); |
1656 | 2.73M | w2h = *(twiddles + 2 * (j << 1)); |
1657 | 2.73M | w3h = *(twiddles + 2 * j + 2 * (j << 1) - 512); |
1658 | 2.73M | w1l = *(twiddles + 2 * j + 1); |
1659 | 2.73M | w2l = *(twiddles + 2 * (j << 1) + 1); |
1660 | 2.73M | w3l = *(twiddles + 2 * j + 2 * (j << 1) - 511); |
1661 | | |
1662 | 9.16M | for (k = in_loop_cnt; k != 0; k--) { |
1663 | 6.42M | WORD32 tmp; |
1664 | 6.42M | WORD32 x0r, x0i, x1r, x1i, x2r, x2i, x3r, x3i; |
1665 | 6.42M | data += (del << 1); |
1666 | | |
1667 | 6.42M | x1r = *data; |
1668 | 6.42M | x1i = *(data + 1); |
1669 | 6.42M | data += (del << 1); |
1670 | | |
1671 | 6.42M | x2r = *data; |
1672 | 6.42M | x2i = *(data + 1); |
1673 | 6.42M | data += (del << 1); |
1674 | | |
1675 | 6.42M | x3r = *data; |
1676 | 6.42M | x3i = *(data + 1); |
1677 | 6.42M | data -= 3 * (del << 1); |
1678 | | |
1679 | 6.42M | tmp = ixheaac_sub32_sat(ixheaacd_mult32_sat(x1r, w1l), |
1680 | 6.42M | ixheaacd_mult32_sat(x1i, w1h)); |
1681 | 6.42M | x1i = ixheaacd_mac32_sat(ixheaacd_mult32_sat(x1r, w1h), x1i, w1l); |
1682 | 6.42M | x1r = tmp; |
1683 | | |
1684 | 6.42M | tmp = ixheaac_sub32_sat(ixheaacd_mult32_sat(x2r, w2l), |
1685 | 6.42M | ixheaacd_mult32_sat(x2i, w2h)); |
1686 | 6.42M | x2i = ixheaacd_mac32_sat(ixheaacd_mult32_sat(x2r, w2h), x2i, w2l); |
1687 | 6.42M | x2r = tmp; |
1688 | | |
1689 | 6.42M | tmp = ixheaac_add32_sat(ixheaacd_mult32_sat(x3r, w3h), |
1690 | 6.42M | ixheaacd_mult32_sat(x3i, w3l)); |
1691 | 6.42M | x3i = ixheaac_sub32_sat(ixheaacd_mult32_sat(x3i, w3h), |
1692 | 6.42M | ixheaacd_mult32_sat(x3r, w3l)); |
1693 | 6.42M | x3r = tmp; |
1694 | | |
1695 | 6.42M | x0r = (*data); |
1696 | 6.42M | x0i = (*(data + 1)); |
1697 | | |
1698 | 6.42M | x0r = ixheaac_add32_sat(x0r, x2r); |
1699 | 6.42M | x0i = ixheaac_add32_sat(x0i, x2i); |
1700 | 6.42M | x2r = ixheaac_sub32_sat(x0r, ixheaac_shl32_sat(x2r, 1)); |
1701 | 6.42M | x2i = ixheaac_sub32_sat(x0i, ixheaac_shl32_sat(x2i, 1)); |
1702 | 6.42M | x1r = ixheaac_add32_sat(x1r, x3r); |
1703 | 6.42M | x1i = ixheaac_add32_sat(x1i, x3i); |
1704 | 6.42M | x3r = ixheaac_sub32_sat(x1r, ixheaac_shl32_sat(x3r, 1)); |
1705 | 6.42M | x3i = ixheaac_sub32_sat(x1i, ixheaac_shl32_sat(x3i, 1)); |
1706 | | |
1707 | 6.42M | x0r = ixheaac_add32_sat(x0r, x1r); |
1708 | 6.42M | x0i = ixheaac_add32_sat(x0i, x1i); |
1709 | 6.42M | x1r = ixheaac_sub32_sat(x0r, ixheaac_shl32_sat(x1r, 1)); |
1710 | 6.42M | x1i = ixheaac_sub32_sat(x0i, ixheaac_shl32_sat(x1i, 1)); |
1711 | 6.42M | x2r = ixheaac_add32_sat(x2r, x3i); |
1712 | 6.42M | x2i = ixheaac_sub32_sat(x2i, x3r); |
1713 | 6.42M | x3i = ixheaac_sub32_sat(x2r, ixheaac_shl32_sat(x3i, 1)); |
1714 | 6.42M | x3r = ixheaac_add32_sat(x2i, ixheaac_shl32_sat(x3r, 1)); |
1715 | | |
1716 | 6.42M | *data = x0r; |
1717 | 6.42M | *(data + 1) = x0i; |
1718 | 6.42M | data += (del << 1); |
1719 | | |
1720 | 6.42M | *data = x2r; |
1721 | 6.42M | *(data + 1) = x2i; |
1722 | 6.42M | data += (del << 1); |
1723 | | |
1724 | 6.42M | *data = x1r; |
1725 | 6.42M | *(data + 1) = x1i; |
1726 | 6.42M | data += (del << 1); |
1727 | | |
1728 | 6.42M | *data = x3i; |
1729 | 6.42M | *(data + 1) = x3r; |
1730 | 6.42M | data += (del << 1); |
1731 | 6.42M | } |
1732 | 2.73M | data -= 2 * npoints; |
1733 | 2.73M | data += 2; |
1734 | 2.73M | } |
1735 | 2.73M | for (; j <= sec_loop_cnt * 2; j += nodespacing) { |
1736 | 1.09M | w1h = *(twiddles + 2 * j); |
1737 | 1.09M | w2h = *(twiddles + 2 * (j << 1) - 512); |
1738 | 1.09M | w3h = *(twiddles + 2 * j + 2 * (j << 1) - 512); |
1739 | 1.09M | w1l = *(twiddles + 2 * j + 1); |
1740 | 1.09M | w2l = *(twiddles + 2 * (j << 1) - 511); |
1741 | 1.09M | w3l = *(twiddles + 2 * j + 2 * (j << 1) - 511); |
1742 | | |
1743 | 2.70M | for (k = in_loop_cnt; k != 0; k--) { |
1744 | 1.61M | WORD32 tmp; |
1745 | 1.61M | WORD32 x0r, x0i, x1r, x1i, x2r, x2i, x3r, x3i; |
1746 | | |
1747 | 1.61M | data += (del << 1); |
1748 | | |
1749 | 1.61M | x1r = *data; |
1750 | 1.61M | x1i = *(data + 1); |
1751 | 1.61M | data += (del << 1); |
1752 | | |
1753 | 1.61M | x2r = *data; |
1754 | 1.61M | x2i = *(data + 1); |
1755 | 1.61M | data += (del << 1); |
1756 | | |
1757 | 1.61M | x3r = *data; |
1758 | 1.61M | x3i = *(data + 1); |
1759 | 1.61M | data -= 3 * (del << 1); |
1760 | | |
1761 | 1.61M | tmp = ixheaac_sub32_sat(ixheaacd_mult32_sat(x1r, w1l), |
1762 | 1.61M | ixheaacd_mult32_sat(x1i, w1h)); |
1763 | 1.61M | x1i = ixheaacd_mac32_sat(ixheaacd_mult32_sat(x1r, w1h), x1i, w1l); |
1764 | 1.61M | x1r = tmp; |
1765 | | |
1766 | 1.61M | tmp = ixheaac_add32_sat(ixheaacd_mult32_sat(x2r, w2h), |
1767 | 1.61M | ixheaacd_mult32_sat(x2i, w2l)); |
1768 | 1.61M | x2i = ixheaac_sub32_sat(ixheaacd_mult32_sat(x2i, w2h), |
1769 | 1.61M | ixheaacd_mult32_sat(x2r, w2l)); |
1770 | 1.61M | x2r = tmp; |
1771 | | |
1772 | 1.61M | tmp = ixheaac_add32_sat(ixheaacd_mult32_sat(x3r, w3h), |
1773 | 1.61M | ixheaacd_mult32_sat(x3i, w3l)); |
1774 | 1.61M | x3i = ixheaac_sub32_sat(ixheaacd_mult32_sat(x3i, w3h), |
1775 | 1.61M | ixheaacd_mult32_sat(x3r, w3l)); |
1776 | 1.61M | x3r = tmp; |
1777 | | |
1778 | 1.61M | x0r = (*data); |
1779 | 1.61M | x0i = (*(data + 1)); |
1780 | | |
1781 | 1.61M | x0r = ixheaac_add32_sat(x0r, x2r); |
1782 | 1.61M | x0i = ixheaac_add32_sat(x0i, x2i); |
1783 | 1.61M | x2r = ixheaac_sub32_sat(x0r, ixheaac_shl32_sat(x2r, 1)); |
1784 | 1.61M | x2i = ixheaac_sub32_sat(x0i, ixheaac_shl32_sat(x2i, 1)); |
1785 | 1.61M | x1r = ixheaac_add32_sat(x1r, x3r); |
1786 | 1.61M | x1i = ixheaac_add32_sat(x1i, x3i); |
1787 | 1.61M | x3r = ixheaac_sub32_sat(x1r, ixheaac_shl32_sat(x3r, 1)); |
1788 | 1.61M | x3i = ixheaac_sub32_sat(x1i, ixheaac_shl32_sat(x3i, 1)); |
1789 | | |
1790 | 1.61M | x0r = ixheaac_add32_sat(x0r, x1r); |
1791 | 1.61M | x0i = ixheaac_add32_sat(x0i, x1i); |
1792 | 1.61M | x1r = ixheaac_sub32_sat(x0r, ixheaac_shl32_sat(x1r, 1)); |
1793 | 1.61M | x1i = ixheaac_sub32_sat(x0i, ixheaac_shl32_sat(x1i, 1)); |
1794 | 1.61M | x2r = ixheaac_add32_sat(x2r, x3i); |
1795 | 1.61M | x2i = ixheaac_sub32_sat(x2i, x3r); |
1796 | 1.61M | x3i = ixheaac_sub32_sat(x2r, ixheaac_shl32_sat(x3i, 1)); |
1797 | 1.61M | x3r = ixheaac_add32_sat(x2i, ixheaac_shl32_sat(x3r, 1)); |
1798 | | |
1799 | 1.61M | *data = x0r; |
1800 | 1.61M | *(data + 1) = x0i; |
1801 | 1.61M | data += (del << 1); |
1802 | | |
1803 | 1.61M | *data = x2r; |
1804 | 1.61M | *(data + 1) = x2i; |
1805 | 1.61M | data += (del << 1); |
1806 | | |
1807 | 1.61M | *data = x1r; |
1808 | 1.61M | *(data + 1) = x1i; |
1809 | 1.61M | data += (del << 1); |
1810 | | |
1811 | 1.61M | *data = x3i; |
1812 | 1.61M | *(data + 1) = x3r; |
1813 | 1.61M | data += (del << 1); |
1814 | 1.61M | } |
1815 | 1.09M | data -= 2 * npoints; |
1816 | 1.09M | data += 2; |
1817 | 1.09M | } |
1818 | 5.47M | for (; j < nodespacing * del; j += nodespacing) { |
1819 | 3.83M | w1h = *(twiddles + 2 * j); |
1820 | 3.83M | w2h = *(twiddles + 2 * (j << 1) - 512); |
1821 | 3.83M | w3h = *(twiddles + 2 * j + 2 * (j << 1) - 1024); |
1822 | 3.83M | w1l = *(twiddles + 2 * j + 1); |
1823 | 3.83M | w2l = *(twiddles + 2 * (j << 1) - 511); |
1824 | 3.83M | w3l = *(twiddles + 2 * j + 2 * (j << 1) - 1023); |
1825 | | |
1826 | 11.8M | for (k = in_loop_cnt; k != 0; k--) { |
1827 | 8.03M | WORD32 tmp; |
1828 | 8.03M | WORD32 x0r, x0i, x1r, x1i, x2r, x2i, x3r, x3i; |
1829 | | |
1830 | 8.03M | data += (del << 1); |
1831 | | |
1832 | 8.03M | x1r = *data; |
1833 | 8.03M | x1i = *(data + 1); |
1834 | 8.03M | data += (del << 1); |
1835 | | |
1836 | 8.03M | x2r = *data; |
1837 | 8.03M | x2i = *(data + 1); |
1838 | 8.03M | data += (del << 1); |
1839 | | |
1840 | 8.03M | x3r = *data; |
1841 | 8.03M | x3i = *(data + 1); |
1842 | 8.03M | data -= 3 * (del << 1); |
1843 | | |
1844 | 8.03M | tmp = ixheaac_sub32_sat(ixheaacd_mult32_sat(x1r, w1l), |
1845 | 8.03M | ixheaacd_mult32_sat(x1i, w1h)); |
1846 | 8.03M | x1i = ixheaacd_mac32_sat(ixheaacd_mult32_sat(x1r, w1h), x1i, w1l); |
1847 | 8.03M | x1r = tmp; |
1848 | | |
1849 | 8.03M | tmp = ixheaac_add32_sat(ixheaacd_mult32_sat(x2r, w2h), |
1850 | 8.03M | ixheaacd_mult32_sat(x2i, w2l)); |
1851 | 8.03M | x2i = ixheaac_sub32_sat(ixheaacd_mult32_sat(x2i, w2h), |
1852 | 8.03M | ixheaacd_mult32_sat(x2r, w2l)); |
1853 | 8.03M | x2r = tmp; |
1854 | | |
1855 | 8.03M | tmp = ixheaac_sub32_sat(ixheaacd_mult32_sat(x3i, w3h), |
1856 | 8.03M | ixheaacd_mult32_sat(x3r, w3l)); |
1857 | 8.03M | x3i = ixheaacd_mac32_sat(ixheaacd_mult32_sat(x3r, w3h), x3i, w3l); |
1858 | 8.03M | x3r = tmp; |
1859 | | |
1860 | 8.03M | x0r = (*data); |
1861 | 8.03M | x0i = (*(data + 1)); |
1862 | | |
1863 | 8.03M | x0r = ixheaac_add32_sat(x0r, x2r); |
1864 | 8.03M | x0i = ixheaac_add32_sat(x0i, x2i); |
1865 | 8.03M | x2r = ixheaac_sub32_sat(x0r, ixheaac_shl32_sat(x2r, 1)); |
1866 | 8.03M | x2i = ixheaac_sub32_sat(x0i, ixheaac_shl32_sat(x2i, 1)); |
1867 | 8.03M | x1r = ixheaac_add32_sat(x1r, x3r); |
1868 | 8.03M | x1i = ixheaac_sub32_sat(x1i, x3i); |
1869 | 8.03M | x3r = ixheaac_sub32_sat(x1r, ixheaac_shl32_sat(x3r, 1)); |
1870 | 8.03M | x3i = ixheaac_add32_sat(x1i, ixheaac_shl32_sat(x3i, 1)); |
1871 | | |
1872 | 8.03M | x0r = ixheaac_add32_sat(x0r, x1r); |
1873 | 8.03M | x0i = ixheaac_add32_sat(x0i, x1i); |
1874 | 8.03M | x1r = ixheaac_sub32_sat(x0r, ixheaac_shl32_sat(x1r, 1)); |
1875 | 8.03M | x1i = ixheaac_sub32_sat(x0i, ixheaac_shl32_sat(x1i, 1)); |
1876 | 8.03M | x2r = ixheaac_add32_sat(x2r, x3i); |
1877 | 8.03M | x2i = ixheaac_sub32_sat(x2i, x3r); |
1878 | 8.03M | x3i = ixheaac_sub32_sat(x2r, ixheaac_shl32_sat(x3i, 1)); |
1879 | 8.03M | x3r = ixheaac_add32_sat(x2i, ixheaac_shl32_sat(x3r, 1)); |
1880 | | |
1881 | 8.03M | *data = x0r; |
1882 | 8.03M | *(data + 1) = x0i; |
1883 | 8.03M | data += (del << 1); |
1884 | | |
1885 | 8.03M | *data = x2r; |
1886 | 8.03M | *(data + 1) = x2i; |
1887 | 8.03M | data += (del << 1); |
1888 | | |
1889 | 8.03M | *data = x1r; |
1890 | 8.03M | *(data + 1) = x1i; |
1891 | 8.03M | data += (del << 1); |
1892 | | |
1893 | 8.03M | *data = x3i; |
1894 | 8.03M | *(data + 1) = x3r; |
1895 | 8.03M | data += (del << 1); |
1896 | 8.03M | } |
1897 | 3.83M | data -= 2 * npoints; |
1898 | 3.83M | data += 2; |
1899 | 3.83M | } |
1900 | 1.64M | nodespacing >>= 2; |
1901 | 1.64M | del <<= 2; |
1902 | 1.64M | in_loop_cnt >>= 2; |
1903 | 1.64M | } |
1904 | 12.0M | if (not_power_4) { |
1905 | 4.54M | const WORD32 *twiddles = ptr_w; |
1906 | 4.54M | nodespacing <<= 1; |
1907 | 4.54M | shift += 1; |
1908 | | |
1909 | 22.7M | for (j = del / 2; j != 0; j--) { |
1910 | 18.2M | WORD32 w1h = *twiddles; |
1911 | 18.2M | WORD32 w1l = *(twiddles + 1); |
1912 | 18.2M | WORD32 tmp; |
1913 | 18.2M | twiddles += nodespacing * 2; |
1914 | | |
1915 | 18.2M | x0r = *ptr_y; |
1916 | 18.2M | x0i = *(ptr_y + 1); |
1917 | 18.2M | ptr_y += (del << 1); |
1918 | | |
1919 | 18.2M | x1r = *ptr_y; |
1920 | 18.2M | x1i = *(ptr_y + 1); |
1921 | | |
1922 | 18.2M | tmp = ixheaac_sub32_sat(ixheaacd_mult32_sat(x1r, w1l), |
1923 | 18.2M | ixheaacd_mult32_sat(x1i, w1h)); |
1924 | 18.2M | x1i = ixheaacd_mac32_sat(ixheaacd_mult32_sat(x1r, w1h), x1i, w1l); |
1925 | 18.2M | x1r = tmp; |
1926 | | |
1927 | 18.2M | *ptr_y = (x0r) / 2 - (x1r) / 2; |
1928 | 18.2M | *(ptr_y + 1) = (x0i) / 2 - (x1i) / 2; |
1929 | 18.2M | ptr_y -= (del << 1); |
1930 | | |
1931 | 18.2M | *ptr_y = (x0r) / 2 + (x1r) / 2; |
1932 | 18.2M | *(ptr_y + 1) = (x0i) / 2 + (x1i) / 2; |
1933 | 18.2M | ptr_y += 2; |
1934 | 18.2M | } |
1935 | 4.54M | twiddles = ptr_w; |
1936 | 22.7M | for (j = del / 2; j != 0; j--) { |
1937 | 18.2M | WORD32 w1h = *twiddles; |
1938 | 18.2M | WORD32 w1l = *(twiddles + 1); |
1939 | 18.2M | WORD32 tmp; |
1940 | 18.2M | twiddles += nodespacing * 2; |
1941 | | |
1942 | 18.2M | x0r = *ptr_y; |
1943 | 18.2M | x0i = *(ptr_y + 1); |
1944 | 18.2M | ptr_y += (del << 1); |
1945 | | |
1946 | 18.2M | x1r = *ptr_y; |
1947 | 18.2M | x1i = *(ptr_y + 1); |
1948 | | |
1949 | 18.2M | tmp = ixheaac_add32_sat(ixheaacd_mult32_sat(x1r, w1h), |
1950 | 18.2M | ixheaacd_mult32_sat(x1i, w1l)); |
1951 | 18.2M | x1i = ixheaac_sub32_sat(ixheaacd_mult32_sat(x1i, w1h), |
1952 | 18.2M | ixheaacd_mult32_sat(x1r, w1l)); |
1953 | 18.2M | x1r = tmp; |
1954 | | |
1955 | 18.2M | *ptr_y = (x0r) / 2 - (x1r) / 2; |
1956 | 18.2M | *(ptr_y + 1) = (x0i) / 2 - (x1i) / 2; |
1957 | 18.2M | ptr_y -= (del << 1); |
1958 | | |
1959 | 18.2M | *ptr_y = (x0r) / 2 + (x1r) / 2; |
1960 | 18.2M | *(ptr_y + 1) = (x0i) / 2 + (x1i) / 2; |
1961 | 18.2M | ptr_y += 2; |
1962 | 18.2M | } |
1963 | 4.54M | } |
1964 | 12.0M | } |
1965 | | |
1966 | 1.22M | else { |
1967 | 1.22M | ptr_w = ixheaacd_twiddle_table_fft_32x32; |
1968 | | |
1969 | 45.8M | for (i = 0; i < npoints; i += 4) { |
1970 | 44.6M | WORD32 *inp = ptr_x; |
1971 | | |
1972 | 44.6M | DIG_REV(i, dig_rev_shift, h2); |
1973 | 44.6M | if (not_power_4) { |
1974 | 31.7M | h2 += 1; |
1975 | 31.7M | h2 &= ~1; |
1976 | 31.7M | } |
1977 | 44.6M | inp += (h2); |
1978 | | |
1979 | 44.6M | x0r = *inp; |
1980 | 44.6M | x0i = *(inp + 1); |
1981 | 44.6M | inp += (npoints >> 1); |
1982 | | |
1983 | 44.6M | x1r = *inp; |
1984 | 44.6M | x1i = *(inp + 1); |
1985 | 44.6M | inp += (npoints >> 1); |
1986 | | |
1987 | 44.6M | x2r = *inp; |
1988 | 44.6M | x2i = *(inp + 1); |
1989 | 44.6M | inp += (npoints >> 1); |
1990 | | |
1991 | 44.6M | x3r = *inp; |
1992 | 44.6M | x3i = *(inp + 1); |
1993 | | |
1994 | 44.6M | x0r = ixheaac_add32_sat(x0r, x2r); |
1995 | 44.6M | x0i = ixheaac_add32_sat(x0i, x2i); |
1996 | 44.6M | x2r = ixheaac_sub32_sat(x0r, ixheaac_shl32_sat(x2r, 1)); |
1997 | 44.6M | x2i = ixheaac_sub32_sat(x0i, ixheaac_shl32_sat(x2i, 1)); |
1998 | 44.6M | x1r = ixheaac_add32_sat(x1r, x3r); |
1999 | 44.6M | x1i = ixheaac_add32_sat(x1i, x3i); |
2000 | 44.6M | x3r = ixheaac_sub32_sat(x1r, ixheaac_shl32_sat(x3r, 1)); |
2001 | 44.6M | x3i = ixheaac_sub32_sat(x1i, ixheaac_shl32_sat(x3i, 1)); |
2002 | | |
2003 | 44.6M | x0r = ixheaac_add32_sat(x0r, x1r); |
2004 | 44.6M | x0i = ixheaac_add32_sat(x0i, x1i); |
2005 | 44.6M | x1r = ixheaac_sub32_sat(x0r, ixheaac_shl32_sat(x1r, 1)); |
2006 | 44.6M | x1i = ixheaac_sub32_sat(x0i, ixheaac_shl32_sat(x1i, 1)); |
2007 | 44.6M | x2r = ixheaac_sub32_sat(x2r, x3i); |
2008 | 44.6M | x2i = ixheaac_add32_sat(x2i, x3r); |
2009 | 44.6M | x3i = ixheaac_add32_sat(x2r, ixheaac_shl32_sat(x3i, 1)); |
2010 | 44.6M | x3r = ixheaac_sub32_sat(x2i, ixheaac_shl32_sat(x3r, 1)); |
2011 | | |
2012 | 44.6M | *ptr_y++ = x0r; |
2013 | 44.6M | *ptr_y++ = x0i; |
2014 | 44.6M | *ptr_y++ = x2r; |
2015 | 44.6M | *ptr_y++ = x2i; |
2016 | 44.6M | *ptr_y++ = x1r; |
2017 | 44.6M | *ptr_y++ = x1i; |
2018 | 44.6M | *ptr_y++ = x3i; |
2019 | 44.6M | *ptr_y++ = x3r; |
2020 | 44.6M | } |
2021 | 1.22M | ptr_y -= 2 * npoints; |
2022 | 1.22M | del = 4; |
2023 | 1.22M | nodespacing = 64; |
2024 | 1.22M | in_loop_cnt = npoints >> 4; |
2025 | 3.75M | for (i = n_stages - 1; i > 0; i--) { |
2026 | 2.52M | const WORD32 *twiddles = ptr_w; |
2027 | 2.52M | WORD32 *data = ptr_y; |
2028 | 2.52M | WORD32 w1h, w2h, w3h, w1l, w2l, w3l; |
2029 | 2.52M | WORD32 sec_loop_cnt; |
2030 | | |
2031 | 16.8M | for (k = in_loop_cnt; k != 0; k--) { |
2032 | 14.3M | x0r = (*data); |
2033 | 14.3M | x0i = (*(data + 1)); |
2034 | 14.3M | data += (del << 1); |
2035 | | |
2036 | 14.3M | x1r = (*data); |
2037 | 14.3M | x1i = (*(data + 1)); |
2038 | 14.3M | data += (del << 1); |
2039 | | |
2040 | 14.3M | x2r = (*data); |
2041 | 14.3M | x2i = (*(data + 1)); |
2042 | 14.3M | data += (del << 1); |
2043 | | |
2044 | 14.3M | x3r = (*data); |
2045 | 14.3M | x3i = (*(data + 1)); |
2046 | 14.3M | data -= 3 * (del << 1); |
2047 | | |
2048 | 14.3M | x0r = ixheaac_add32_sat(x0r, x2r); |
2049 | 14.3M | x0i = ixheaac_add32_sat(x0i, x2i); |
2050 | 14.3M | x2r = ixheaac_sub32_sat(x0r, ixheaac_shl32_sat(x2r, 1)); |
2051 | 14.3M | x2i = ixheaac_sub32_sat(x0i, ixheaac_shl32_sat(x2i, 1)); |
2052 | 14.3M | x1r = ixheaac_add32_sat(x1r, x3r); |
2053 | 14.3M | x1i = ixheaac_add32_sat(x1i, x3i); |
2054 | 14.3M | x3r = ixheaac_sub32_sat(x1r, ixheaac_shl32_sat(x3r, 1)); |
2055 | 14.3M | x3i = ixheaac_sub32_sat(x1i, ixheaac_shl32_sat(x3i, 1)); |
2056 | | |
2057 | 14.3M | x0r = ixheaac_add32_sat(x0r, x1r); |
2058 | 14.3M | x0i = ixheaac_add32_sat(x0i, x1i); |
2059 | 14.3M | x1r = ixheaac_sub32_sat(x0r, ixheaac_shl32_sat(x1r, 1)); |
2060 | 14.3M | x1i = ixheaac_sub32_sat(x0i, ixheaac_shl32_sat(x1i, 1)); |
2061 | 14.3M | x2r = ixheaac_sub32_sat(x2r, x3i); |
2062 | 14.3M | x2i = ixheaac_add32_sat(x2i, x3r); |
2063 | 14.3M | x3i = ixheaac_add32_sat(x2r, ixheaac_shl32_sat(x3i, 1)); |
2064 | 14.3M | x3r = ixheaac_sub32_sat(x2i, ixheaac_shl32_sat(x3r, 1)); |
2065 | | |
2066 | 14.3M | *data = x0r; |
2067 | 14.3M | *(data + 1) = x0i; |
2068 | 14.3M | data += (del << 1); |
2069 | | |
2070 | 14.3M | *data = x2r; |
2071 | 14.3M | *(data + 1) = x2i; |
2072 | 14.3M | data += (del << 1); |
2073 | | |
2074 | 14.3M | *data = x1r; |
2075 | 14.3M | *(data + 1) = x1i; |
2076 | 14.3M | data += (del << 1); |
2077 | | |
2078 | 14.3M | *data = x3i; |
2079 | 14.3M | *(data + 1) = x3r; |
2080 | 14.3M | data += (del << 1); |
2081 | 14.3M | } |
2082 | 2.52M | data = ptr_y + 2; |
2083 | | |
2084 | 2.52M | sec_loop_cnt = (nodespacing * del); |
2085 | 2.52M | sec_loop_cnt = (sec_loop_cnt / 4) + (sec_loop_cnt / 8) - |
2086 | 2.52M | (sec_loop_cnt / 16) + (sec_loop_cnt / 32) - |
2087 | 2.52M | (sec_loop_cnt / 64) + (sec_loop_cnt / 128) - |
2088 | 2.52M | (sec_loop_cnt / 256); |
2089 | 2.52M | j = nodespacing; |
2090 | | |
2091 | 13.9M | for (j = nodespacing; j <= sec_loop_cnt; j += nodespacing) { |
2092 | 11.4M | w1h = *(twiddles + 2 * j); |
2093 | 11.4M | w2h = *(twiddles + 2 * (j << 1)); |
2094 | 11.4M | w3h = *(twiddles + 2 * j + 2 * (j << 1)); |
2095 | 11.4M | w1l = *(twiddles + 2 * j + 1); |
2096 | 11.4M | w2l = *(twiddles + 2 * (j << 1) + 1); |
2097 | 11.4M | w3l = *(twiddles + 2 * j + 2 * (j << 1) + 1); |
2098 | | |
2099 | 45.8M | for (k = in_loop_cnt; k != 0; k--) { |
2100 | 34.4M | WORD32 tmp; |
2101 | 34.4M | WORD32 x0r, x0i, x1r, x1i, x2r, x2i, x3r, x3i; |
2102 | | |
2103 | 34.4M | data += (del << 1); |
2104 | | |
2105 | 34.4M | x1r = *data; |
2106 | 34.4M | x1i = *(data + 1); |
2107 | 34.4M | data += (del << 1); |
2108 | | |
2109 | 34.4M | x2r = *data; |
2110 | 34.4M | x2i = *(data + 1); |
2111 | 34.4M | data += (del << 1); |
2112 | | |
2113 | 34.4M | x3r = *data; |
2114 | 34.4M | x3i = *(data + 1); |
2115 | 34.4M | data -= 3 * (del << 1); |
2116 | | |
2117 | 34.4M | tmp = ixheaac_add32_sat(ixheaacd_mult32_sat(x1r, w1l), |
2118 | 34.4M | ixheaacd_mult32_sat(x1i, w1h)); |
2119 | 34.4M | x1i = ixheaacd_mac32_sat(-ixheaacd_mult32_sat(x1r, w1h), x1i, w1l); |
2120 | 34.4M | x1r = tmp; |
2121 | | |
2122 | 34.4M | tmp = ixheaac_add32_sat(ixheaacd_mult32_sat(x2r, w2l), |
2123 | 34.4M | ixheaacd_mult32_sat(x2i, w2h)); |
2124 | 34.4M | x2i = ixheaacd_mac32_sat(-ixheaacd_mult32_sat(x2r, w2h), x2i, w2l); |
2125 | 34.4M | x2r = tmp; |
2126 | | |
2127 | 34.4M | tmp = ixheaac_add32_sat(ixheaacd_mult32_sat(x3r, w3l), |
2128 | 34.4M | ixheaacd_mult32_sat(x3i, w3h)); |
2129 | 34.4M | x3i = ixheaacd_mac32_sat(-ixheaacd_mult32_sat(x3r, w3h), x3i, w3l); |
2130 | 34.4M | x3r = tmp; |
2131 | | |
2132 | 34.4M | x0r = (*data); |
2133 | 34.4M | x0i = (*(data + 1)); |
2134 | | |
2135 | 34.4M | x0r = ixheaac_add32_sat(x0r, x2r); |
2136 | 34.4M | x0i = ixheaac_add32_sat(x0i, x2i); |
2137 | 34.4M | x2r = ixheaac_sub32_sat(x0r, ixheaac_shl32_sat(x2r, 1)); |
2138 | 34.4M | x2i = ixheaac_sub32_sat(x0i, ixheaac_shl32_sat(x2i, 1)); |
2139 | 34.4M | x1r = ixheaac_add32_sat(x1r, x3r); |
2140 | 34.4M | x1i = ixheaac_add32_sat(x1i, x3i); |
2141 | 34.4M | x3r = ixheaac_sub32_sat(x1r, ixheaac_shl32_sat(x3r, 1)); |
2142 | 34.4M | x3i = ixheaac_sub32_sat(x1i, ixheaac_shl32_sat(x3i, 1)); |
2143 | | |
2144 | 34.4M | x0r = ixheaac_add32_sat(x0r, x1r); |
2145 | 34.4M | x0i = ixheaac_add32_sat(x0i, x1i); |
2146 | 34.4M | x1r = ixheaac_sub32_sat(x0r, ixheaac_shl32_sat(x1r, 1)); |
2147 | 34.4M | x1i = ixheaac_sub32_sat(x0i, ixheaac_shl32_sat(x1i, 1)); |
2148 | 34.4M | x2r = ixheaac_sub32_sat(x2r, x3i); |
2149 | 34.4M | x2i = ixheaac_add32_sat(x2i, x3r); |
2150 | 34.4M | x3i = ixheaac_add32_sat(x2r, ixheaac_shl32_sat(x3i, 1)); |
2151 | 34.4M | x3r = ixheaac_sub32_sat(x2i, ixheaac_shl32_sat(x3r, 1)); |
2152 | | |
2153 | 34.4M | *data = x0r; |
2154 | 34.4M | *(data + 1) = x0i; |
2155 | 34.4M | data += (del << 1); |
2156 | | |
2157 | 34.4M | *data = x2r; |
2158 | 34.4M | *(data + 1) = x2i; |
2159 | 34.4M | data += (del << 1); |
2160 | | |
2161 | 34.4M | *data = x1r; |
2162 | 34.4M | *(data + 1) = x1i; |
2163 | 34.4M | data += (del << 1); |
2164 | | |
2165 | 34.4M | *data = x3i; |
2166 | 34.4M | *(data + 1) = x3r; |
2167 | 34.4M | data += (del << 1); |
2168 | 34.4M | } |
2169 | 11.4M | data -= 2 * npoints; |
2170 | 11.4M | data += 2; |
2171 | 11.4M | } |
2172 | 9.49M | for (; j <= (nodespacing * del) >> 1; j += nodespacing) { |
2173 | 6.96M | w1h = *(twiddles + 2 * j); |
2174 | 6.96M | w2h = *(twiddles + 2 * (j << 1)); |
2175 | 6.96M | w3h = *(twiddles + 2 * j + 2 * (j << 1) - 512); |
2176 | 6.96M | w1l = *(twiddles + 2 * j + 1); |
2177 | 6.96M | w2l = *(twiddles + 2 * (j << 1) + 1); |
2178 | 6.96M | w3l = *(twiddles + 2 * j + 2 * (j << 1) - 511); |
2179 | | |
2180 | 31.3M | for (k = in_loop_cnt; k != 0; k--) { |
2181 | 24.4M | WORD32 tmp; |
2182 | 24.4M | WORD32 x0r, x0i, x1r, x1i, x2r, x2i, x3r, x3i; |
2183 | | |
2184 | 24.4M | data += (del << 1); |
2185 | | |
2186 | 24.4M | x1r = *data; |
2187 | 24.4M | x1i = *(data + 1); |
2188 | 24.4M | data += (del << 1); |
2189 | | |
2190 | 24.4M | x2r = *data; |
2191 | 24.4M | x2i = *(data + 1); |
2192 | 24.4M | data += (del << 1); |
2193 | | |
2194 | 24.4M | x3r = *data; |
2195 | 24.4M | x3i = *(data + 1); |
2196 | 24.4M | data -= 3 * (del << 1); |
2197 | | |
2198 | 24.4M | tmp = ixheaac_add32_sat(ixheaacd_mult32_sat(x1r, w1l), |
2199 | 24.4M | ixheaacd_mult32_sat(x1i, w1h)); |
2200 | 24.4M | x1i = ixheaacd_mac32_sat(-ixheaacd_mult32_sat(x1r, w1h), x1i, w1l); |
2201 | 24.4M | x1r = tmp; |
2202 | | |
2203 | 24.4M | tmp = ixheaac_add32_sat(ixheaacd_mult32_sat(x2r, w2l), |
2204 | 24.4M | ixheaacd_mult32_sat(x2i, w2h)); |
2205 | 24.4M | x2i = ixheaacd_mac32_sat(-ixheaacd_mult32_sat(x2r, w2h), x2i, w2l); |
2206 | 24.4M | x2r = tmp; |
2207 | | |
2208 | 24.4M | tmp = ixheaac_sub32_sat(ixheaacd_mult32_sat(x3r, w3h), |
2209 | 24.4M | ixheaacd_mult32_sat(x3i, w3l)); |
2210 | 24.4M | x3i = ixheaac_add32_sat(ixheaacd_mult32_sat(x3r, w3l), |
2211 | 24.4M | ixheaacd_mult32_sat(x3i, w3h)); |
2212 | 24.4M | x3r = tmp; |
2213 | | |
2214 | 24.4M | x0r = (*data); |
2215 | 24.4M | x0i = (*(data + 1)); |
2216 | | |
2217 | 24.4M | x0r = ixheaac_add32_sat(x0r, x2r); |
2218 | 24.4M | x0i = ixheaac_add32_sat(x0i, x2i); |
2219 | 24.4M | x2r = ixheaac_sub32_sat(x0r, ixheaac_shl32_sat(x2r, 1)); |
2220 | 24.4M | x2i = ixheaac_sub32_sat(x0i, ixheaac_shl32_sat(x2i, 1)); |
2221 | 24.4M | x1r = ixheaac_add32_sat(x1r, x3r); |
2222 | 24.4M | x1i = ixheaac_add32_sat(x1i, x3i); |
2223 | 24.4M | x3r = ixheaac_sub32_sat(x1r, ixheaac_shl32_sat(x3r, 1)); |
2224 | 24.4M | x3i = ixheaac_sub32_sat(x1i, ixheaac_shl32_sat(x3i, 1)); |
2225 | | |
2226 | 24.4M | x0r = ixheaac_add32_sat(x0r, x1r); |
2227 | 24.4M | x0i = ixheaac_add32_sat(x0i, x1i); |
2228 | 24.4M | x1r = ixheaac_sub32_sat(x0r, ixheaac_shl32_sat(x1r, 1)); |
2229 | 24.4M | x1i = ixheaac_sub32_sat(x0i, ixheaac_shl32_sat(x1i, 1)); |
2230 | 24.4M | x2r = ixheaac_sub32_sat(x2r, x3i); |
2231 | 24.4M | x2i = ixheaac_add32_sat(x2i, x3r); |
2232 | 24.4M | x3i = ixheaac_add32_sat(x2r, ixheaac_shl32_sat(x3i, 1)); |
2233 | 24.4M | x3r = ixheaac_sub32_sat(x2i, ixheaac_shl32_sat(x3r, 1)); |
2234 | | |
2235 | 24.4M | *data = x0r; |
2236 | 24.4M | *(data + 1) = x0i; |
2237 | 24.4M | data += (del << 1); |
2238 | | |
2239 | 24.4M | *data = x2r; |
2240 | 24.4M | *(data + 1) = x2i; |
2241 | 24.4M | data += (del << 1); |
2242 | | |
2243 | 24.4M | *data = x1r; |
2244 | 24.4M | *(data + 1) = x1i; |
2245 | 24.4M | data += (del << 1); |
2246 | | |
2247 | 24.4M | *data = x3i; |
2248 | 24.4M | *(data + 1) = x3r; |
2249 | 24.4M | data += (del << 1); |
2250 | 24.4M | } |
2251 | 6.96M | data -= 2 * npoints; |
2252 | 6.96M | data += 2; |
2253 | 6.96M | } |
2254 | 6.96M | for (; j <= sec_loop_cnt * 2; j += nodespacing) { |
2255 | 4.43M | w1h = *(twiddles + 2 * j); |
2256 | 4.43M | w2h = *(twiddles + 2 * (j << 1) - 512); |
2257 | 4.43M | w3h = *(twiddles + 2 * j + 2 * (j << 1) - 512); |
2258 | 4.43M | w1l = *(twiddles + 2 * j + 1); |
2259 | 4.43M | w2l = *(twiddles + 2 * (j << 1) - 511); |
2260 | 4.43M | w3l = *(twiddles + 2 * j + 2 * (j << 1) - 511); |
2261 | | |
2262 | 14.4M | for (k = in_loop_cnt; k != 0; k--) { |
2263 | 10.0M | WORD32 tmp; |
2264 | 10.0M | WORD32 x0r, x0i, x1r, x1i, x2r, x2i, x3r, x3i; |
2265 | | |
2266 | 10.0M | data += (del << 1); |
2267 | | |
2268 | 10.0M | x1r = *data; |
2269 | 10.0M | x1i = *(data + 1); |
2270 | 10.0M | data += (del << 1); |
2271 | | |
2272 | 10.0M | x2r = *data; |
2273 | 10.0M | x2i = *(data + 1); |
2274 | 10.0M | data += (del << 1); |
2275 | | |
2276 | 10.0M | x3r = *data; |
2277 | 10.0M | x3i = *(data + 1); |
2278 | 10.0M | data -= 3 * (del << 1); |
2279 | | |
2280 | 10.0M | tmp = ixheaac_add32_sat(ixheaacd_mult32_sat(x1r, w1l), |
2281 | 10.0M | ixheaacd_mult32_sat(x1i, w1h)); |
2282 | 10.0M | x1i = ixheaacd_mac32_sat(-ixheaacd_mult32_sat(x1r, w1h), x1i, w1l); |
2283 | 10.0M | x1r = tmp; |
2284 | | |
2285 | 10.0M | tmp = ixheaac_sub32_sat(ixheaacd_mult32_sat(x2r, w2h), |
2286 | 10.0M | ixheaacd_mult32_sat(x2i, w2l)); |
2287 | 10.0M | x2i = ixheaac_add32_sat(ixheaacd_mult32_sat(x2r, w2l), |
2288 | 10.0M | ixheaacd_mult32_sat(x2i, w2h)); |
2289 | 10.0M | x2r = tmp; |
2290 | | |
2291 | 10.0M | tmp = ixheaac_sub32_sat(ixheaacd_mult32_sat(x3r, w3h), |
2292 | 10.0M | ixheaacd_mult32_sat(x3i, w3l)); |
2293 | 10.0M | x3i = ixheaac_add32_sat(ixheaacd_mult32_sat(x3r, w3l), |
2294 | 10.0M | ixheaacd_mult32_sat(x3i, w3h)); |
2295 | 10.0M | x3r = tmp; |
2296 | | |
2297 | 10.0M | x0r = (*data); |
2298 | 10.0M | x0i = (*(data + 1)); |
2299 | | |
2300 | 10.0M | x0r = ixheaac_add32_sat(x0r, x2r); |
2301 | 10.0M | x0i = ixheaac_add32_sat(x0i, x2i); |
2302 | 10.0M | x2r = ixheaac_sub32_sat(x0r, ixheaac_shl32_sat(x2r, 1)); |
2303 | 10.0M | x2i = ixheaac_sub32_sat(x0i, ixheaac_shl32_sat(x2i, 1)); |
2304 | 10.0M | x1r = ixheaac_add32_sat(x1r, x3r); |
2305 | 10.0M | x1i = ixheaac_add32_sat(x1i, x3i); |
2306 | 10.0M | x3r = ixheaac_sub32_sat(x1r, ixheaac_shl32_sat(x3r, 1)); |
2307 | 10.0M | x3i = ixheaac_sub32_sat(x1i, ixheaac_shl32_sat(x3i, 1)); |
2308 | | |
2309 | 10.0M | x0r = ixheaac_add32_sat(x0r, x1r); |
2310 | 10.0M | x0i = ixheaac_add32_sat(x0i, x1i); |
2311 | 10.0M | x1r = ixheaac_sub32_sat(x0r, ixheaac_shl32_sat(x1r, 1)); |
2312 | 10.0M | x1i = ixheaac_sub32_sat(x0i, ixheaac_shl32_sat(x1i, 1)); |
2313 | 10.0M | x2r = ixheaac_sub32_sat(x2r, x3i); |
2314 | 10.0M | x2i = ixheaac_add32_sat(x2i, x3r); |
2315 | 10.0M | x3i = ixheaac_add32_sat(x2r, ixheaac_shl32_sat(x3i, 1)); |
2316 | 10.0M | x3r = ixheaac_sub32_sat(x2i, ixheaac_shl32_sat(x3r, 1)); |
2317 | | |
2318 | 10.0M | *data = x0r; |
2319 | 10.0M | *(data + 1) = x0i; |
2320 | 10.0M | data += (del << 1); |
2321 | | |
2322 | 10.0M | *data = x2r; |
2323 | 10.0M | *(data + 1) = x2i; |
2324 | 10.0M | data += (del << 1); |
2325 | | |
2326 | 10.0M | *data = x1r; |
2327 | 10.0M | *(data + 1) = x1i; |
2328 | 10.0M | data += (del << 1); |
2329 | | |
2330 | 10.0M | *data = x3i; |
2331 | 10.0M | *(data + 1) = x3r; |
2332 | 10.0M | data += (del << 1); |
2333 | 10.0M | } |
2334 | 4.43M | data -= 2 * npoints; |
2335 | 4.43M | data += 2; |
2336 | 4.43M | } |
2337 | 13.9M | for (; j < nodespacing * del; j += nodespacing) { |
2338 | 11.4M | w1h = *(twiddles + 2 * j); |
2339 | 11.4M | w2h = *(twiddles + 2 * (j << 1) - 512); |
2340 | 11.4M | w3h = *(twiddles + 2 * j + 2 * (j << 1) - 1024); |
2341 | 11.4M | w1l = *(twiddles + 2 * j + 1); |
2342 | 11.4M | w2l = *(twiddles + 2 * (j << 1) - 511); |
2343 | 11.4M | w3l = *(twiddles + 2 * j + 2 * (j << 1) - 1023); |
2344 | | |
2345 | 45.8M | for (k = in_loop_cnt; k != 0; k--) { |
2346 | 34.4M | WORD32 tmp; |
2347 | 34.4M | WORD32 x0r, x0i, x1r, x1i, x2r, x2i, x3r, x3i; |
2348 | | |
2349 | 34.4M | data += (del << 1); |
2350 | | |
2351 | 34.4M | x1r = *data; |
2352 | 34.4M | x1i = *(data + 1); |
2353 | 34.4M | data += (del << 1); |
2354 | | |
2355 | 34.4M | x2r = *data; |
2356 | 34.4M | x2i = *(data + 1); |
2357 | 34.4M | data += (del << 1); |
2358 | | |
2359 | 34.4M | x3r = *data; |
2360 | 34.4M | x3i = *(data + 1); |
2361 | 34.4M | data -= 3 * (del << 1); |
2362 | | |
2363 | 34.4M | tmp = ixheaac_add32_sat(ixheaacd_mult32_sat(x1r, w1l), |
2364 | 34.4M | ixheaacd_mult32_sat(x1i, w1h)); |
2365 | 34.4M | x1i = ixheaacd_mac32_sat(-ixheaacd_mult32_sat(x1r, w1h), x1i, w1l); |
2366 | 34.4M | x1r = tmp; |
2367 | | |
2368 | 34.4M | tmp = ixheaac_sub32_sat(ixheaacd_mult32_sat(x2r, w2h), |
2369 | 34.4M | ixheaacd_mult32_sat(x2i, w2l)); |
2370 | 34.4M | x2i = ixheaac_add32_sat(ixheaacd_mult32_sat(x2r, w2l), |
2371 | 34.4M | ixheaacd_mult32_sat(x2i, w2h)); |
2372 | 34.4M | x2r = tmp; |
2373 | | |
2374 | 34.4M | tmp = -ixheaac_add32_sat(ixheaacd_mult32_sat(x3r, w3l), |
2375 | 34.4M | ixheaacd_mult32_sat(x3i, w3h)); |
2376 | 34.4M | x3i = ixheaacd_mac32_sat(-ixheaacd_mult32_sat(x3r, w3h), x3i, w3l); |
2377 | 34.4M | x3r = tmp; |
2378 | | |
2379 | 34.4M | x0r = (*data); |
2380 | 34.4M | x0i = (*(data + 1)); |
2381 | | |
2382 | 34.4M | x0r = ixheaac_add32_sat(x0r, x2r); |
2383 | 34.4M | x0i = ixheaac_add32_sat(x0i, x2i); |
2384 | 34.4M | x2r = ixheaac_sub32_sat(x0r, ixheaac_shl32_sat(x2r, 1)); |
2385 | 34.4M | x2i = ixheaac_sub32_sat(x0i, ixheaac_shl32_sat(x2i, 1)); |
2386 | 34.4M | x1r = ixheaac_add32_sat(x1r, x3r); |
2387 | 34.4M | x1i = ixheaac_sub32_sat(x1i, x3i); |
2388 | 34.4M | x3r = ixheaac_sub32_sat(x1r, ixheaac_shl32_sat(x3r, 1)); |
2389 | 34.4M | x3i = ixheaac_add32_sat(x1i, ixheaac_shl32_sat(x3i, 1)); |
2390 | | |
2391 | 34.4M | x0r = ixheaac_add32_sat(x0r, x1r); |
2392 | 34.4M | x0i = ixheaac_add32_sat(x0i, x1i); |
2393 | 34.4M | x1r = ixheaac_sub32_sat(x0r, ixheaac_shl32_sat(x1r, 1)); |
2394 | 34.4M | x1i = ixheaac_sub32_sat(x0i, ixheaac_shl32_sat(x1i, 1)); |
2395 | 34.4M | x2r = ixheaac_sub32_sat(x2r, x3i); |
2396 | 34.4M | x2i = ixheaac_add32_sat(x2i, x3r); |
2397 | 34.4M | x3i = ixheaac_add32_sat(x2r, ixheaac_shl32_sat(x3i, 1)); |
2398 | 34.4M | x3r = ixheaac_sub32_sat(x2i, ixheaac_shl32_sat(x3r, 1)); |
2399 | | |
2400 | 34.4M | *data = x0r; |
2401 | 34.4M | *(data + 1) = x0i; |
2402 | 34.4M | data += (del << 1); |
2403 | | |
2404 | 34.4M | *data = x2r; |
2405 | 34.4M | *(data + 1) = x2i; |
2406 | 34.4M | data += (del << 1); |
2407 | | |
2408 | 34.4M | *data = x1r; |
2409 | 34.4M | *(data + 1) = x1i; |
2410 | 34.4M | data += (del << 1); |
2411 | | |
2412 | 34.4M | *data = x3i; |
2413 | 34.4M | *(data + 1) = x3r; |
2414 | 34.4M | data += (del << 1); |
2415 | 34.4M | } |
2416 | 11.4M | data -= 2 * npoints; |
2417 | 11.4M | data += 2; |
2418 | 11.4M | } |
2419 | 2.52M | nodespacing >>= 2; |
2420 | 2.52M | del <<= 2; |
2421 | 2.52M | in_loop_cnt >>= 2; |
2422 | 2.52M | } |
2423 | 1.22M | if (not_power_4) { |
2424 | 307k | const WORD32 *twiddles = ptr_w; |
2425 | 307k | nodespacing <<= 1; |
2426 | 307k | shift += 1; |
2427 | 32.0M | for (j = del / 2; j != 0; j--) { |
2428 | 31.7M | WORD32 w1h = *twiddles; |
2429 | 31.7M | WORD32 w1l = *(twiddles + 1); |
2430 | | |
2431 | 31.7M | WORD32 tmp; |
2432 | 31.7M | twiddles += nodespacing * 2; |
2433 | | |
2434 | 31.7M | x0r = *ptr_y; |
2435 | 31.7M | x0i = *(ptr_y + 1); |
2436 | 31.7M | ptr_y += (del << 1); |
2437 | | |
2438 | 31.7M | x1r = *ptr_y; |
2439 | 31.7M | x1i = *(ptr_y + 1); |
2440 | | |
2441 | 31.7M | tmp = ixheaac_add32_sat(ixheaacd_mult32_sat(x1r, w1l), |
2442 | 31.7M | ixheaacd_mult32_sat(x1i, w1h)); |
2443 | 31.7M | x1i = ixheaacd_mac32_sat(-ixheaacd_mult32_sat(x1r, w1h), x1i, w1l); |
2444 | 31.7M | x1r = tmp; |
2445 | | |
2446 | 31.7M | *ptr_y = (x0r) / 2 - (x1r) / 2; |
2447 | 31.7M | *(ptr_y + 1) = (x0i) / 2 - (x1i) / 2; |
2448 | 31.7M | ptr_y -= (del << 1); |
2449 | | |
2450 | 31.7M | *ptr_y = (x0r) / 2 + (x1r) / 2; |
2451 | 31.7M | *(ptr_y + 1) = (x0i) / 2 + (x1i) / 2; |
2452 | 31.7M | ptr_y += 2; |
2453 | 31.7M | } |
2454 | 307k | twiddles = ptr_w; |
2455 | 32.0M | for (j = del / 2; j != 0; j--) { |
2456 | 31.7M | WORD32 w1h = *twiddles; |
2457 | 31.7M | WORD32 w1l = *(twiddles + 1); |
2458 | 31.7M | WORD32 tmp; |
2459 | 31.7M | twiddles += nodespacing * 2; |
2460 | | |
2461 | 31.7M | x0r = *ptr_y; |
2462 | 31.7M | x0i = *(ptr_y + 1); |
2463 | 31.7M | ptr_y += (del << 1); |
2464 | | |
2465 | 31.7M | x1r = *ptr_y; |
2466 | 31.7M | x1i = *(ptr_y + 1); |
2467 | | |
2468 | 31.7M | tmp = ixheaac_sub32_sat(ixheaacd_mult32_sat(x1r, w1h), |
2469 | 31.7M | ixheaacd_mult32_sat(x1i, w1l)); |
2470 | 31.7M | x1i = ixheaac_add32_sat(ixheaacd_mult32_sat(x1r, w1l), |
2471 | 31.7M | ixheaacd_mult32_sat(x1i, w1h)); |
2472 | 31.7M | x1r = tmp; |
2473 | | |
2474 | 31.7M | *ptr_y = (x0r) / 2 - (x1r) / 2; |
2475 | 31.7M | *(ptr_y + 1) = (x0i) / 2 - (x1i) / 2; |
2476 | 31.7M | ptr_y -= (del << 1); |
2477 | | |
2478 | 31.7M | *ptr_y = (x0r) / 2 + (x1r) / 2; |
2479 | 31.7M | *(ptr_y + 1) = (x0i) / 2 + (x1i) / 2; |
2480 | 31.7M | ptr_y += 2; |
2481 | 31.7M | } |
2482 | 307k | } |
2483 | 1.22M | } |
2484 | | |
2485 | 315M | for (i = 0; i < nlength; i++) { |
2486 | 302M | xr[i] = y[2 * i]; |
2487 | 302M | xi[i] = y[2 * i + 1]; |
2488 | 302M | } |
2489 | | |
2490 | 13.2M | *preshift = shift - *preshift; |
2491 | 13.2M | return; |
2492 | 13.2M | } |
2493 | | |
2494 | | static PLATFORM_INLINE void ixheaacd_complex_3point_fft(WORD32 *inp, WORD32 *op, |
2495 | 20.9M | WORD32 sign_dir) { |
2496 | 20.9M | WORD32 add_r, sub_r; |
2497 | 20.9M | WORD32 add_i, sub_i; |
2498 | 20.9M | WORD32 temp_real, temp_imag, temp; |
2499 | | |
2500 | 20.9M | WORD32 p1, p2, p3, p4; |
2501 | | |
2502 | 20.9M | WORD32 sinmu; |
2503 | 20.9M | sinmu = -1859775393 * sign_dir; |
2504 | | |
2505 | 20.9M | temp_real = ixheaac_add32_sat(inp[0], inp[2]); |
2506 | 20.9M | temp_imag = ixheaac_add32_sat(inp[1], inp[3]); |
2507 | | |
2508 | 20.9M | add_r = ixheaac_add32_sat(inp[2], inp[4]); |
2509 | 20.9M | add_i = ixheaac_add32_sat(inp[3], inp[5]); |
2510 | | |
2511 | 20.9M | sub_r = ixheaac_sub32_sat(inp[2], inp[4]); |
2512 | 20.9M | sub_i = ixheaac_sub32_sat(inp[3], inp[5]); |
2513 | | |
2514 | 20.9M | p1 = add_r >> 1; |
2515 | 20.9M | p4 = add_i >> 1; |
2516 | 20.9M | p2 = ixheaac_mult32_shl(sub_i, sinmu); |
2517 | 20.9M | p3 = ixheaac_mult32_shl(sub_r, sinmu); |
2518 | | |
2519 | 20.9M | temp = ixheaac_sub32(inp[0], p1); |
2520 | | |
2521 | 20.9M | op[0] = ixheaac_add32_sat(temp_real, inp[4]); |
2522 | 20.9M | op[1] = ixheaac_add32_sat(temp_imag, inp[5]); |
2523 | 20.9M | op[2] = ixheaac_add32_sat(temp, p2); |
2524 | 20.9M | op[3] = ixheaac_sub32_sat(ixheaac_sub32_sat(inp[1], p3), p4); |
2525 | 20.9M | op[4] = ixheaac_sub32_sat(temp, p2); |
2526 | 20.9M | op[5] = ixheaac_sub32_sat(ixheaac_add32_sat(inp[1], p3), p4); |
2527 | | |
2528 | 20.9M | return; |
2529 | 20.9M | } |
2530 | | |
2531 | | VOID ixheaacd_complex_fft_p3(WORD32 *xr, WORD32 *xi, WORD32 nlength, |
2532 | 2.66M | WORD32 fft_mode, WORD32 *preshift) { |
2533 | 2.66M | WORD32 i, j; |
2534 | 2.66M | WORD32 shift = 0; |
2535 | 2.66M | WORD32 xr_3[384]; |
2536 | 2.66M | WORD32 xi_3[384]; |
2537 | 2.66M | WORD32 x[1024]; |
2538 | 2.66M | WORD32 y[1024]; |
2539 | 2.66M | WORD32 cnfac, npts; |
2540 | 2.66M | WORD32 mpass = nlength; |
2541 | 2.66M | WORD32 n = 0; |
2542 | 2.66M | WORD32 *ptr_x = x; |
2543 | 2.66M | WORD32 *ptr_y = y; |
2544 | | |
2545 | 2.66M | cnfac = 0; |
2546 | 5.32M | while (mpass % 3 == 0) { |
2547 | 2.66M | mpass /= 3; |
2548 | 2.66M | cnfac++; |
2549 | 2.66M | } |
2550 | 2.66M | npts = mpass; |
2551 | | |
2552 | 10.6M | for (i = 0; i < 3 * cnfac; i++) { |
2553 | 70.6M | for (j = 0; j < mpass; j++) { |
2554 | 62.7M | xr_3[j] = xr[3 * j + i]; |
2555 | 62.7M | xi_3[j] = xi[3 * j + i]; |
2556 | 62.7M | } |
2557 | | |
2558 | 7.98M | (*ixheaacd_complex_fft_p2)(xr_3, xi_3, mpass, fft_mode, &shift); |
2559 | | |
2560 | 70.6M | for (j = 0; j < mpass; j++) { |
2561 | 62.7M | xr[3 * j + i] = xr_3[j]; |
2562 | 62.7M | xi[3 * j + i] = xi_3[j]; |
2563 | 62.7M | } |
2564 | 7.98M | } |
2565 | | |
2566 | 8.90M | while (npts >> 1) { |
2567 | 6.24M | n++; |
2568 | 6.24M | npts = npts >> 1; |
2569 | 6.24M | } |
2570 | | |
2571 | 2.66M | if (n % 2 == 0) |
2572 | 2.47M | shift = ((n + 4)) / 2; |
2573 | 187k | else |
2574 | 187k | shift = ((n + 5) / 2); |
2575 | | |
2576 | 2.66M | *preshift = shift - *preshift + 1; |
2577 | | |
2578 | 65.3M | for (i = 0; i < nlength; i++) { |
2579 | 62.7M | ptr_x[2 * i] = (xr[i] >> 1); |
2580 | 62.7M | ptr_x[2 * i + 1] = (xi[i] >> 1); |
2581 | 62.7M | } |
2582 | | |
2583 | 2.66M | { |
2584 | 2.66M | const WORD32 *w1r, *w1i; |
2585 | 2.66M | WORD32 tmp; |
2586 | 2.66M | w1r = ixheaacd_twiddle_table_3pr; |
2587 | 2.66M | w1i = ixheaacd_twiddle_table_3pi; |
2588 | | |
2589 | 2.66M | if (fft_mode < 0) { |
2590 | 19.3M | for (i = 0; i < nlength; i += 3) { |
2591 | 16.7M | w1r++; |
2592 | 16.7M | w1i++; |
2593 | | |
2594 | 16.7M | tmp = ixheaac_sub32_sat(ixheaacd_mult32_sat(ptr_x[2 * i + 2], (*w1r)), |
2595 | 16.7M | ixheaacd_mult32_sat(ptr_x[2 * i + 3], (*w1i))); |
2596 | 16.7M | ptr_x[2 * i + 3] = |
2597 | 16.7M | ixheaac_add32_sat(ixheaacd_mult32_sat(ptr_x[2 * i + 2], (*w1i)), |
2598 | 16.7M | ixheaacd_mult32_sat(ptr_x[2 * i + 3], (*w1r))); |
2599 | 16.7M | ptr_x[2 * i + 2] = tmp; |
2600 | | |
2601 | 16.7M | w1r++; |
2602 | 16.7M | w1i++; |
2603 | | |
2604 | 16.7M | tmp = ixheaac_sub32_sat(ixheaacd_mult32_sat(ptr_x[2 * i + 4], (*w1r)), |
2605 | 16.7M | ixheaacd_mult32_sat(ptr_x[2 * i + 5], (*w1i))); |
2606 | 16.7M | ptr_x[2 * i + 5] = |
2607 | 16.7M | ixheaac_add32_sat(ixheaacd_mult32_sat(ptr_x[2 * i + 4], (*w1i)), |
2608 | 16.7M | ixheaacd_mult32_sat(ptr_x[2 * i + 5], (*w1r))); |
2609 | 16.7M | ptr_x[2 * i + 4] = tmp; |
2610 | | |
2611 | 16.7M | w1r += 3 * (128 / mpass - 1) + 1; |
2612 | 16.7M | w1i += 3 * (128 / mpass - 1) + 1; |
2613 | 16.7M | } |
2614 | 2.58M | } |
2615 | | |
2616 | 74.2k | else { |
2617 | 4.24M | for (i = 0; i < nlength; i += 3) { |
2618 | 4.17M | w1r++; |
2619 | 4.17M | w1i++; |
2620 | | |
2621 | 4.17M | tmp = ixheaac_add32_sat(ixheaacd_mult32_sat(ptr_x[2 * i + 2], (*w1r)), |
2622 | 4.17M | ixheaacd_mult32_sat(ptr_x[2 * i + 3], (*w1i))); |
2623 | 4.17M | ptr_x[2 * i + 3] = |
2624 | 4.17M | ixheaac_sub32_sat(ixheaacd_mult32_sat(ptr_x[2 * i + 3], (*w1r)), |
2625 | 4.17M | ixheaacd_mult32_sat(ptr_x[2 * i + 2], (*w1i))); |
2626 | 4.17M | ptr_x[2 * i + 2] = tmp; |
2627 | | |
2628 | 4.17M | w1r++; |
2629 | 4.17M | w1i++; |
2630 | | |
2631 | 4.17M | tmp = ixheaac_add32_sat(ixheaacd_mult32_sat(ptr_x[2 * i + 4], (*w1r)), |
2632 | 4.17M | ixheaacd_mult32_sat(ptr_x[2 * i + 5], (*w1i))); |
2633 | 4.17M | ptr_x[2 * i + 5] = |
2634 | 4.17M | ixheaac_sub32_sat(ixheaacd_mult32_sat(ptr_x[2 * i + 5], (*w1r)), |
2635 | 4.17M | ixheaacd_mult32_sat(ptr_x[2 * i + 4], (*w1i))); |
2636 | 4.17M | ptr_x[2 * i + 4] = tmp; |
2637 | | |
2638 | 4.17M | w1r += 3 * (128 / mpass - 1) + 1; |
2639 | 4.17M | w1i += 3 * (128 / mpass - 1) + 1; |
2640 | 4.17M | } |
2641 | 74.2k | } |
2642 | 2.66M | } |
2643 | | |
2644 | 23.5M | for (i = 0; i < mpass; i++) { |
2645 | 20.9M | ixheaacd_complex_3point_fft(ptr_x, ptr_y, fft_mode); |
2646 | | |
2647 | 20.9M | ptr_x = ptr_x + 6; |
2648 | 20.9M | ptr_y = ptr_y + 6; |
2649 | 20.9M | } |
2650 | | |
2651 | 2.66M | ptr_y = y; |
2652 | 23.5M | for (i = 0; i < mpass; i++) { |
2653 | 20.9M | xr[i] = *ptr_y++; |
2654 | 20.9M | xi[i] = *ptr_y++; |
2655 | 20.9M | xr[mpass + i] = *ptr_y++; |
2656 | 20.9M | xi[mpass + i] = *ptr_y++; |
2657 | 20.9M | xr[2 * mpass + i] = *ptr_y++; |
2658 | 20.9M | xi[2 * mpass + i] = *ptr_y++; |
2659 | 20.9M | } |
2660 | | |
2661 | 2.66M | return; |
2662 | 2.66M | } |
2663 | | |
2664 | | VOID ixheaacd_complex_fft(WORD32 *data_r, WORD32 *data_i, WORD32 nlength, WORD32 fft_mode, |
2665 | 1.74M | WORD32 *preshift) { |
2666 | 1.74M | if (nlength & (nlength - 1)) { |
2667 | 323k | ixheaacd_complex_fft_p3(data_r, data_i, nlength, fft_mode, preshift); |
2668 | 323k | } else |
2669 | 1.41M | (*ixheaacd_complex_fft_p2)(data_r, data_i, nlength, fft_mode, preshift); |
2670 | | |
2671 | 1.74M | return; |
2672 | 1.74M | } |