1 /*
2 * Copyright (c) Lynne
3 *
4 * Power of two FFT:
5 * Copyright (c) Lynne
6 * Copyright (c) 2008 Loren Merritt
7 * Copyright (c) 2002 Fabrice Bellard
8 * Partly based on libdjbfft by D. J. Bernstein
9 *
10 * This file is part of FFmpeg.
11 *
12 * FFmpeg is free software; you can redistribute it and/or
13 * modify it under the terms of the GNU Lesser General Public
14 * License as published by the Free Software Foundation; either
15 * version 2.1 of the License, or (at your option) any later version.
16 *
17 * FFmpeg is distributed in the hope that it will be useful,
18 * but WITHOUT ANY WARRANTY; without even the implied warranty of
19 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
20 * Lesser General Public License for more details.
21 *
22 * You should have received a copy of the GNU Lesser General Public
23 * License along with FFmpeg; if not, write to the Free Software
24 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
25 */
26
27 #define TABLE_DEF(name, size) \
28 DECLARE_ALIGNED(32, TXSample, TX_TAB(ff_tx_tab_ ##name))[size]
29
30 #define SR_TABLE(len) \
31 TABLE_DEF(len, len/4 + 1)
32
33 /* Power of two tables */
34 SR_TABLE(8);
35 SR_TABLE(16);
36 SR_TABLE(32);
37 SR_TABLE(64);
38 SR_TABLE(128);
39 SR_TABLE(256);
40 SR_TABLE(512);
41 SR_TABLE(1024);
42 SR_TABLE(2048);
43 SR_TABLE(4096);
44 SR_TABLE(8192);
45 SR_TABLE(16384);
46 SR_TABLE(32768);
47 SR_TABLE(65536);
48 SR_TABLE(131072);
49
50 /* Other factors' tables */
51 TABLE_DEF(53, 8);
52 TABLE_DEF( 7, 6);
53 TABLE_DEF( 9, 8);
54
55 typedef struct FFSRTabsInitOnce {
56 void (*func)(void);
57 AVOnce control;
58 int factors[TX_MAX_SUB]; /* Must be sorted high -> low */
59 } FFSRTabsInitOnce;
60
61 #define INIT_FF_SR_TAB(len) \
62 static av_cold void TX_TAB(ff_tx_init_tab_ ##len)(void) \
63 { \
64 double freq = 2*M_PI/len; \
65 TXSample *tab = TX_TAB(ff_tx_tab_ ##len); \
66 \
67 for (int i = 0; i < len/4; i++) \
68 *tab++ = RESCALE(cos(i*freq)); \
69 \
70 *tab = 0; \
71 }
72
73 INIT_FF_SR_TAB(8)
74 INIT_FF_SR_TAB(16)
75 INIT_FF_SR_TAB(32)
76 INIT_FF_SR_TAB(64)
77 INIT_FF_SR_TAB(128)
78 INIT_FF_SR_TAB(256)
79 INIT_FF_SR_TAB(512)
80 INIT_FF_SR_TAB(1024)
81 INIT_FF_SR_TAB(2048)
82 INIT_FF_SR_TAB(4096)
83 INIT_FF_SR_TAB(8192)
84 INIT_FF_SR_TAB(16384)
85 INIT_FF_SR_TAB(32768)
86 INIT_FF_SR_TAB(65536)
87 INIT_FF_SR_TAB(131072)
88
89 static FFSRTabsInitOnce sr_tabs_init_once[] = {
90 { TX_TAB(ff_tx_init_tab_8), AV_ONCE_INIT },
91 { TX_TAB(ff_tx_init_tab_16), AV_ONCE_INIT },
92 { TX_TAB(ff_tx_init_tab_32), AV_ONCE_INIT },
93 { TX_TAB(ff_tx_init_tab_64), AV_ONCE_INIT },
94 { TX_TAB(ff_tx_init_tab_128), AV_ONCE_INIT },
95 { TX_TAB(ff_tx_init_tab_256), AV_ONCE_INIT },
96 { TX_TAB(ff_tx_init_tab_512), AV_ONCE_INIT },
97 { TX_TAB(ff_tx_init_tab_1024), AV_ONCE_INIT },
98 { TX_TAB(ff_tx_init_tab_2048), AV_ONCE_INIT },
99 { TX_TAB(ff_tx_init_tab_4096), AV_ONCE_INIT },
100 { TX_TAB(ff_tx_init_tab_8192), AV_ONCE_INIT },
101 { TX_TAB(ff_tx_init_tab_16384), AV_ONCE_INIT },
102 { TX_TAB(ff_tx_init_tab_32768), AV_ONCE_INIT },
103 { TX_TAB(ff_tx_init_tab_65536), AV_ONCE_INIT },
104 { TX_TAB(ff_tx_init_tab_131072), AV_ONCE_INIT },
105 };
106
TX_TAB(ff_tx_init_tab_53)107 static av_cold void TX_TAB(ff_tx_init_tab_53)(void)
108 {
109 TX_TAB(ff_tx_tab_53)[0] = RESCALE(cos(2 * M_PI / 12));
110 TX_TAB(ff_tx_tab_53)[1] = RESCALE(cos(2 * M_PI / 12));
111 TX_TAB(ff_tx_tab_53)[2] = RESCALE(cos(2 * M_PI / 6));
112 TX_TAB(ff_tx_tab_53)[3] = RESCALE(cos(2 * M_PI / 6));
113 TX_TAB(ff_tx_tab_53)[4] = RESCALE(cos(2 * M_PI / 5));
114 TX_TAB(ff_tx_tab_53)[5] = RESCALE(sin(2 * M_PI / 5));
115 TX_TAB(ff_tx_tab_53)[6] = RESCALE(cos(2 * M_PI / 10));
116 TX_TAB(ff_tx_tab_53)[7] = RESCALE(sin(2 * M_PI / 10));
117 }
118
TX_TAB(ff_tx_init_tab_7)119 static av_cold void TX_TAB(ff_tx_init_tab_7)(void)
120 {
121 TX_TAB(ff_tx_tab_7)[0] = RESCALE(cos(2 * M_PI / 7));
122 TX_TAB(ff_tx_tab_7)[1] = RESCALE(sin(2 * M_PI / 7));
123 TX_TAB(ff_tx_tab_7)[2] = RESCALE(sin(2 * M_PI / 28));
124 TX_TAB(ff_tx_tab_7)[3] = RESCALE(cos(2 * M_PI / 28));
125 TX_TAB(ff_tx_tab_7)[4] = RESCALE(cos(2 * M_PI / 14));
126 TX_TAB(ff_tx_tab_7)[5] = RESCALE(sin(2 * M_PI / 14));
127 }
128
TX_TAB(ff_tx_init_tab_9)129 static av_cold void TX_TAB(ff_tx_init_tab_9)(void)
130 {
131 TX_TAB(ff_tx_tab_9)[0] = RESCALE(cos(2 * M_PI / 3));
132 TX_TAB(ff_tx_tab_9)[1] = RESCALE(sin(2 * M_PI / 3));
133 TX_TAB(ff_tx_tab_9)[2] = RESCALE(cos(2 * M_PI / 9));
134 TX_TAB(ff_tx_tab_9)[3] = RESCALE(sin(2 * M_PI / 9));
135 TX_TAB(ff_tx_tab_9)[4] = RESCALE(cos(2 * M_PI / 36));
136 TX_TAB(ff_tx_tab_9)[5] = RESCALE(sin(2 * M_PI / 36));
137 TX_TAB(ff_tx_tab_9)[6] = TX_TAB(ff_tx_tab_9)[2] + TX_TAB(ff_tx_tab_9)[5];
138 TX_TAB(ff_tx_tab_9)[7] = TX_TAB(ff_tx_tab_9)[3] - TX_TAB(ff_tx_tab_9)[4];
139 }
140
141 static FFSRTabsInitOnce nptwo_tabs_init_once[] = {
142 { TX_TAB(ff_tx_init_tab_53), AV_ONCE_INIT, { 15, 5, 3 } },
143 { TX_TAB(ff_tx_init_tab_9), AV_ONCE_INIT, { 9 } },
144 { TX_TAB(ff_tx_init_tab_7), AV_ONCE_INIT, { 7 } },
145 };
146
TX_TAB(ff_tx_init_tabs)147 av_cold void TX_TAB(ff_tx_init_tabs)(int len)
148 {
149 int factor_2 = ff_ctz(len);
150 if (factor_2) {
151 int idx = factor_2 - 3;
152 for (int i = 0; i <= idx; i++)
153 ff_thread_once(&sr_tabs_init_once[i].control,
154 sr_tabs_init_once[i].func);
155 len >>= factor_2;
156 }
157
158 for (int i = 0; i < FF_ARRAY_ELEMS(nptwo_tabs_init_once); i++) {
159 int f, f_idx = 0;
160
161 if (len <= 1)
162 return;
163
164 while ((f = nptwo_tabs_init_once[i].factors[f_idx++])) {
165 if (f % len)
166 continue;
167
168 ff_thread_once(&nptwo_tabs_init_once[i].control,
169 nptwo_tabs_init_once[i].func);
170 len /= f;
171 break;
172 }
173 }
174 }
175
fft3(TXComplex * out,TXComplex * in,ptrdiff_t stride)176 static av_always_inline void fft3(TXComplex *out, TXComplex *in,
177 ptrdiff_t stride)
178 {
179 TXComplex tmp[2];
180 const TXSample *tab = TX_TAB(ff_tx_tab_53);
181 #ifdef TX_INT32
182 int64_t mtmp[4];
183 #endif
184
185 BF(tmp[0].re, tmp[1].im, in[1].im, in[2].im);
186 BF(tmp[0].im, tmp[1].re, in[1].re, in[2].re);
187
188 out[0*stride].re = in[0].re + tmp[1].re;
189 out[0*stride].im = in[0].im + tmp[1].im;
190
191 #ifdef TX_INT32
192 mtmp[0] = (int64_t)tab[0] * tmp[0].re;
193 mtmp[1] = (int64_t)tab[1] * tmp[0].im;
194 mtmp[2] = (int64_t)tab[2] * tmp[1].re;
195 mtmp[3] = (int64_t)tab[2] * tmp[1].im;
196 out[1*stride].re = in[0].re - (mtmp[2] + mtmp[0] + 0x40000000 >> 31);
197 out[1*stride].im = in[0].im - (mtmp[3] - mtmp[1] + 0x40000000 >> 31);
198 out[2*stride].re = in[0].re - (mtmp[2] - mtmp[0] + 0x40000000 >> 31);
199 out[2*stride].im = in[0].im - (mtmp[3] + mtmp[1] + 0x40000000 >> 31);
200 #else
201 tmp[0].re = tab[0] * tmp[0].re;
202 tmp[0].im = tab[1] * tmp[0].im;
203 tmp[1].re = tab[2] * tmp[1].re;
204 tmp[1].im = tab[2] * tmp[1].im;
205 out[1*stride].re = in[0].re - tmp[1].re + tmp[0].re;
206 out[1*stride].im = in[0].im - tmp[1].im - tmp[0].im;
207 out[2*stride].re = in[0].re - tmp[1].re - tmp[0].re;
208 out[2*stride].im = in[0].im - tmp[1].im + tmp[0].im;
209 #endif
210 }
211
212 #define DECL_FFT5(NAME, D0, D1, D2, D3, D4) \
213 static av_always_inline void NAME(TXComplex *out, TXComplex *in, \
214 ptrdiff_t stride) \
215 { \
216 TXComplex z0[4], t[6]; \
217 const TXSample *tab = TX_TAB(ff_tx_tab_53); \
218 \
219 BF(t[1].im, t[0].re, in[1].re, in[4].re); \
220 BF(t[1].re, t[0].im, in[1].im, in[4].im); \
221 BF(t[3].im, t[2].re, in[2].re, in[3].re); \
222 BF(t[3].re, t[2].im, in[2].im, in[3].im); \
223 \
224 out[D0*stride].re = in[0].re + t[0].re + t[2].re; \
225 out[D0*stride].im = in[0].im + t[0].im + t[2].im; \
226 \
227 SMUL(t[4].re, t[0].re, tab[4], tab[6], t[2].re, t[0].re); \
228 SMUL(t[4].im, t[0].im, tab[4], tab[6], t[2].im, t[0].im); \
229 CMUL(t[5].re, t[1].re, tab[5], tab[7], t[3].re, t[1].re); \
230 CMUL(t[5].im, t[1].im, tab[5], tab[7], t[3].im, t[1].im); \
231 \
232 BF(z0[0].re, z0[3].re, t[0].re, t[1].re); \
233 BF(z0[0].im, z0[3].im, t[0].im, t[1].im); \
234 BF(z0[2].re, z0[1].re, t[4].re, t[5].re); \
235 BF(z0[2].im, z0[1].im, t[4].im, t[5].im); \
236 \
237 out[D1*stride].re = in[0].re + z0[3].re; \
238 out[D1*stride].im = in[0].im + z0[0].im; \
239 out[D2*stride].re = in[0].re + z0[2].re; \
240 out[D2*stride].im = in[0].im + z0[1].im; \
241 out[D3*stride].re = in[0].re + z0[1].re; \
242 out[D3*stride].im = in[0].im + z0[2].im; \
243 out[D4*stride].re = in[0].re + z0[0].re; \
244 out[D4*stride].im = in[0].im + z0[3].im; \
245 }
246
247 DECL_FFT5(fft5, 0, 1, 2, 3, 4)
248 DECL_FFT5(fft5_m1, 0, 6, 12, 3, 9)
249 DECL_FFT5(fft5_m2, 10, 1, 7, 13, 4)
250 DECL_FFT5(fft5_m3, 5, 11, 2, 8, 14)
251
fft7(TXComplex * out,TXComplex * in,ptrdiff_t stride)252 static av_always_inline void fft7(TXComplex *out, TXComplex *in,
253 ptrdiff_t stride)
254 {
255 TXComplex t[6], z[3];
256 const TXComplex *tab = (const TXComplex *)TX_TAB(ff_tx_tab_7);
257 #ifdef TX_INT32
258 int64_t mtmp[12];
259 #endif
260
261 BF(t[1].re, t[0].re, in[1].re, in[6].re);
262 BF(t[1].im, t[0].im, in[1].im, in[6].im);
263 BF(t[3].re, t[2].re, in[2].re, in[5].re);
264 BF(t[3].im, t[2].im, in[2].im, in[5].im);
265 BF(t[5].re, t[4].re, in[3].re, in[4].re);
266 BF(t[5].im, t[4].im, in[3].im, in[4].im);
267
268 out[0*stride].re = in[0].re + t[0].re + t[2].re + t[4].re;
269 out[0*stride].im = in[0].im + t[0].im + t[2].im + t[4].im;
270
271 #ifdef TX_INT32 /* NOTE: it's possible to do this with 16 mults but 72 adds */
272 mtmp[ 0] = ((int64_t)tab[0].re)*t[0].re - ((int64_t)tab[2].re)*t[4].re;
273 mtmp[ 1] = ((int64_t)tab[0].re)*t[4].re - ((int64_t)tab[1].re)*t[0].re;
274 mtmp[ 2] = ((int64_t)tab[0].re)*t[2].re - ((int64_t)tab[2].re)*t[0].re;
275 mtmp[ 3] = ((int64_t)tab[0].re)*t[0].im - ((int64_t)tab[1].re)*t[2].im;
276 mtmp[ 4] = ((int64_t)tab[0].re)*t[4].im - ((int64_t)tab[1].re)*t[0].im;
277 mtmp[ 5] = ((int64_t)tab[0].re)*t[2].im - ((int64_t)tab[2].re)*t[0].im;
278
279 mtmp[ 6] = ((int64_t)tab[2].im)*t[1].im + ((int64_t)tab[1].im)*t[5].im;
280 mtmp[ 7] = ((int64_t)tab[0].im)*t[5].im + ((int64_t)tab[2].im)*t[3].im;
281 mtmp[ 8] = ((int64_t)tab[2].im)*t[5].im + ((int64_t)tab[1].im)*t[3].im;
282 mtmp[ 9] = ((int64_t)tab[0].im)*t[1].re + ((int64_t)tab[1].im)*t[3].re;
283 mtmp[10] = ((int64_t)tab[2].im)*t[3].re + ((int64_t)tab[0].im)*t[5].re;
284 mtmp[11] = ((int64_t)tab[2].im)*t[1].re + ((int64_t)tab[1].im)*t[5].re;
285
286 z[0].re = (int32_t)(mtmp[ 0] - ((int64_t)tab[1].re)*t[2].re + 0x40000000 >> 31);
287 z[1].re = (int32_t)(mtmp[ 1] - ((int64_t)tab[2].re)*t[2].re + 0x40000000 >> 31);
288 z[2].re = (int32_t)(mtmp[ 2] - ((int64_t)tab[1].re)*t[4].re + 0x40000000 >> 31);
289 z[0].im = (int32_t)(mtmp[ 3] - ((int64_t)tab[2].re)*t[4].im + 0x40000000 >> 31);
290 z[1].im = (int32_t)(mtmp[ 4] - ((int64_t)tab[2].re)*t[2].im + 0x40000000 >> 31);
291 z[2].im = (int32_t)(mtmp[ 5] - ((int64_t)tab[1].re)*t[4].im + 0x40000000 >> 31);
292
293 t[0].re = (int32_t)(mtmp[ 6] - ((int64_t)tab[0].im)*t[3].im + 0x40000000 >> 31);
294 t[2].re = (int32_t)(mtmp[ 7] - ((int64_t)tab[1].im)*t[1].im + 0x40000000 >> 31);
295 t[4].re = (int32_t)(mtmp[ 8] + ((int64_t)tab[0].im)*t[1].im + 0x40000000 >> 31);
296 t[0].im = (int32_t)(mtmp[ 9] + ((int64_t)tab[2].im)*t[5].re + 0x40000000 >> 31);
297 t[2].im = (int32_t)(mtmp[10] - ((int64_t)tab[1].im)*t[1].re + 0x40000000 >> 31);
298 t[4].im = (int32_t)(mtmp[11] - ((int64_t)tab[0].im)*t[3].re + 0x40000000 >> 31);
299 #else
300 z[0].re = tab[0].re*t[0].re - tab[2].re*t[4].re - tab[1].re*t[2].re;
301 z[1].re = tab[0].re*t[4].re - tab[1].re*t[0].re - tab[2].re*t[2].re;
302 z[2].re = tab[0].re*t[2].re - tab[2].re*t[0].re - tab[1].re*t[4].re;
303 z[0].im = tab[0].re*t[0].im - tab[1].re*t[2].im - tab[2].re*t[4].im;
304 z[1].im = tab[0].re*t[4].im - tab[1].re*t[0].im - tab[2].re*t[2].im;
305 z[2].im = tab[0].re*t[2].im - tab[2].re*t[0].im - tab[1].re*t[4].im;
306
307 /* It's possible to do t[4].re and t[0].im with 2 multiplies only by
308 * multiplying the sum of all with the average of the twiddles */
309
310 t[0].re = tab[2].im*t[1].im + tab[1].im*t[5].im - tab[0].im*t[3].im;
311 t[2].re = tab[0].im*t[5].im + tab[2].im*t[3].im - tab[1].im*t[1].im;
312 t[4].re = tab[2].im*t[5].im + tab[1].im*t[3].im + tab[0].im*t[1].im;
313 t[0].im = tab[0].im*t[1].re + tab[1].im*t[3].re + tab[2].im*t[5].re;
314 t[2].im = tab[2].im*t[3].re + tab[0].im*t[5].re - tab[1].im*t[1].re;
315 t[4].im = tab[2].im*t[1].re + tab[1].im*t[5].re - tab[0].im*t[3].re;
316 #endif
317
318 BF(t[1].re, z[0].re, z[0].re, t[4].re);
319 BF(t[3].re, z[1].re, z[1].re, t[2].re);
320 BF(t[5].re, z[2].re, z[2].re, t[0].re);
321 BF(t[1].im, z[0].im, z[0].im, t[0].im);
322 BF(t[3].im, z[1].im, z[1].im, t[2].im);
323 BF(t[5].im, z[2].im, z[2].im, t[4].im);
324
325 out[1*stride].re = in[0].re + z[0].re;
326 out[1*stride].im = in[0].im + t[1].im;
327 out[2*stride].re = in[0].re + t[3].re;
328 out[2*stride].im = in[0].im + z[1].im;
329 out[3*stride].re = in[0].re + z[2].re;
330 out[3*stride].im = in[0].im + t[5].im;
331 out[4*stride].re = in[0].re + t[5].re;
332 out[4*stride].im = in[0].im + z[2].im;
333 out[5*stride].re = in[0].re + z[1].re;
334 out[5*stride].im = in[0].im + t[3].im;
335 out[6*stride].re = in[0].re + t[1].re;
336 out[6*stride].im = in[0].im + z[0].im;
337 }
338
fft9(TXComplex * out,TXComplex * in,ptrdiff_t stride)339 static av_always_inline void fft9(TXComplex *out, TXComplex *in,
340 ptrdiff_t stride)
341 {
342 const TXComplex *tab = (const TXComplex *)TX_TAB(ff_tx_tab_9);
343 TXComplex t[16], w[4], x[5], y[5], z[2];
344 #ifdef TX_INT32
345 int64_t mtmp[12];
346 #endif
347
348 BF(t[1].re, t[0].re, in[1].re, in[8].re);
349 BF(t[1].im, t[0].im, in[1].im, in[8].im);
350 BF(t[3].re, t[2].re, in[2].re, in[7].re);
351 BF(t[3].im, t[2].im, in[2].im, in[7].im);
352 BF(t[5].re, t[4].re, in[3].re, in[6].re);
353 BF(t[5].im, t[4].im, in[3].im, in[6].im);
354 BF(t[7].re, t[6].re, in[4].re, in[5].re);
355 BF(t[7].im, t[6].im, in[4].im, in[5].im);
356
357 w[0].re = t[0].re - t[6].re;
358 w[0].im = t[0].im - t[6].im;
359 w[1].re = t[2].re - t[6].re;
360 w[1].im = t[2].im - t[6].im;
361 w[2].re = t[1].re - t[7].re;
362 w[2].im = t[1].im - t[7].im;
363 w[3].re = t[3].re + t[7].re;
364 w[3].im = t[3].im + t[7].im;
365
366 z[0].re = in[0].re + t[4].re;
367 z[0].im = in[0].im + t[4].im;
368
369 z[1].re = t[0].re + t[2].re + t[6].re;
370 z[1].im = t[0].im + t[2].im + t[6].im;
371
372 out[0*stride].re = z[0].re + z[1].re;
373 out[0*stride].im = z[0].im + z[1].im;
374
375 #ifdef TX_INT32
376 mtmp[0] = t[1].re - t[3].re + t[7].re;
377 mtmp[1] = t[1].im - t[3].im + t[7].im;
378
379 y[3].re = (int32_t)(((int64_t)tab[0].im)*mtmp[0] + 0x40000000 >> 31);
380 y[3].im = (int32_t)(((int64_t)tab[0].im)*mtmp[1] + 0x40000000 >> 31);
381
382 mtmp[0] = (int32_t)(((int64_t)tab[0].re)*z[1].re + 0x40000000 >> 31);
383 mtmp[1] = (int32_t)(((int64_t)tab[0].re)*z[1].im + 0x40000000 >> 31);
384 mtmp[2] = (int32_t)(((int64_t)tab[0].re)*t[4].re + 0x40000000 >> 31);
385 mtmp[3] = (int32_t)(((int64_t)tab[0].re)*t[4].im + 0x40000000 >> 31);
386
387 x[3].re = z[0].re + (int32_t)mtmp[0];
388 x[3].im = z[0].im + (int32_t)mtmp[1];
389 z[0].re = in[0].re + (int32_t)mtmp[2];
390 z[0].im = in[0].im + (int32_t)mtmp[3];
391
392 mtmp[0] = ((int64_t)tab[1].re)*w[0].re;
393 mtmp[1] = ((int64_t)tab[1].re)*w[0].im;
394 mtmp[2] = ((int64_t)tab[2].im)*w[0].re;
395 mtmp[3] = ((int64_t)tab[2].im)*w[0].im;
396 mtmp[4] = ((int64_t)tab[1].im)*w[2].re;
397 mtmp[5] = ((int64_t)tab[1].im)*w[2].im;
398 mtmp[6] = ((int64_t)tab[2].re)*w[2].re;
399 mtmp[7] = ((int64_t)tab[2].re)*w[2].im;
400
401 x[1].re = (int32_t)(mtmp[0] + ((int64_t)tab[2].im)*w[1].re + 0x40000000 >> 31);
402 x[1].im = (int32_t)(mtmp[1] + ((int64_t)tab[2].im)*w[1].im + 0x40000000 >> 31);
403 x[2].re = (int32_t)(mtmp[2] - ((int64_t)tab[3].re)*w[1].re + 0x40000000 >> 31);
404 x[2].im = (int32_t)(mtmp[3] - ((int64_t)tab[3].re)*w[1].im + 0x40000000 >> 31);
405 y[1].re = (int32_t)(mtmp[4] + ((int64_t)tab[2].re)*w[3].re + 0x40000000 >> 31);
406 y[1].im = (int32_t)(mtmp[5] + ((int64_t)tab[2].re)*w[3].im + 0x40000000 >> 31);
407 y[2].re = (int32_t)(mtmp[6] - ((int64_t)tab[3].im)*w[3].re + 0x40000000 >> 31);
408 y[2].im = (int32_t)(mtmp[7] - ((int64_t)tab[3].im)*w[3].im + 0x40000000 >> 31);
409
410 y[0].re = (int32_t)(((int64_t)tab[0].im)*t[5].re + 0x40000000 >> 31);
411 y[0].im = (int32_t)(((int64_t)tab[0].im)*t[5].im + 0x40000000 >> 31);
412
413 #else
414 y[3].re = tab[0].im*(t[1].re - t[3].re + t[7].re);
415 y[3].im = tab[0].im*(t[1].im - t[3].im + t[7].im);
416
417 x[3].re = z[0].re + tab[0].re*z[1].re;
418 x[3].im = z[0].im + tab[0].re*z[1].im;
419 z[0].re = in[0].re + tab[0].re*t[4].re;
420 z[0].im = in[0].im + tab[0].re*t[4].im;
421
422 x[1].re = tab[1].re*w[0].re + tab[2].im*w[1].re;
423 x[1].im = tab[1].re*w[0].im + tab[2].im*w[1].im;
424 x[2].re = tab[2].im*w[0].re - tab[3].re*w[1].re;
425 x[2].im = tab[2].im*w[0].im - tab[3].re*w[1].im;
426 y[1].re = tab[1].im*w[2].re + tab[2].re*w[3].re;
427 y[1].im = tab[1].im*w[2].im + tab[2].re*w[3].im;
428 y[2].re = tab[2].re*w[2].re - tab[3].im*w[3].re;
429 y[2].im = tab[2].re*w[2].im - tab[3].im*w[3].im;
430
431 y[0].re = tab[0].im*t[5].re;
432 y[0].im = tab[0].im*t[5].im;
433 #endif
434
435 x[4].re = x[1].re + x[2].re;
436 x[4].im = x[1].im + x[2].im;
437
438 y[4].re = y[1].re - y[2].re;
439 y[4].im = y[1].im - y[2].im;
440 x[1].re = z[0].re + x[1].re;
441 x[1].im = z[0].im + x[1].im;
442 y[1].re = y[0].re + y[1].re;
443 y[1].im = y[0].im + y[1].im;
444 x[2].re = z[0].re + x[2].re;
445 x[2].im = z[0].im + x[2].im;
446 y[2].re = y[2].re - y[0].re;
447 y[2].im = y[2].im - y[0].im;
448 x[4].re = z[0].re - x[4].re;
449 x[4].im = z[0].im - x[4].im;
450 y[4].re = y[0].re - y[4].re;
451 y[4].im = y[0].im - y[4].im;
452
453 out[1*stride] = (TXComplex){ x[1].re + y[1].im, x[1].im - y[1].re };
454 out[2*stride] = (TXComplex){ x[2].re + y[2].im, x[2].im - y[2].re };
455 out[3*stride] = (TXComplex){ x[3].re + y[3].im, x[3].im - y[3].re };
456 out[4*stride] = (TXComplex){ x[4].re + y[4].im, x[4].im - y[4].re };
457 out[5*stride] = (TXComplex){ x[4].re - y[4].im, x[4].im + y[4].re };
458 out[6*stride] = (TXComplex){ x[3].re - y[3].im, x[3].im + y[3].re };
459 out[7*stride] = (TXComplex){ x[2].re - y[2].im, x[2].im + y[2].re };
460 out[8*stride] = (TXComplex){ x[1].re - y[1].im, x[1].im + y[1].re };
461 }
462
fft15(TXComplex * out,TXComplex * in,ptrdiff_t stride)463 static av_always_inline void fft15(TXComplex *out, TXComplex *in,
464 ptrdiff_t stride)
465 {
466 TXComplex tmp[15];
467
468 for (int i = 0; i < 5; i++)
469 fft3(tmp + i, in + i*3, 5);
470
471 fft5_m1(out, tmp + 0, stride);
472 fft5_m2(out, tmp + 5, stride);
473 fft5_m3(out, tmp + 10, stride);
474 }
475
476 #define BUTTERFLIES(a0, a1, a2, a3) \
477 do { \
478 r0=a0.re; \
479 i0=a0.im; \
480 r1=a1.re; \
481 i1=a1.im; \
482 BF(t3, t5, t5, t1); \
483 BF(a2.re, a0.re, r0, t5); \
484 BF(a3.im, a1.im, i1, t3); \
485 BF(t4, t6, t2, t6); \
486 BF(a3.re, a1.re, r1, t4); \
487 BF(a2.im, a0.im, i0, t6); \
488 } while (0)
489
490 #define TRANSFORM(a0, a1, a2, a3, wre, wim) \
491 do { \
492 CMUL(t1, t2, a2.re, a2.im, wre, -wim); \
493 CMUL(t5, t6, a3.re, a3.im, wre, wim); \
494 BUTTERFLIES(a0, a1, a2, a3); \
495 } while (0)
496
497 /* z[0...8n-1], w[1...2n-1] */
TX_NAME(ff_tx_fft_sr_combine)498 static inline void TX_NAME(ff_tx_fft_sr_combine)(TXComplex *z,
499 const TXSample *cos, int len)
500 {
501 int o1 = 2*len;
502 int o2 = 4*len;
503 int o3 = 6*len;
504 const TXSample *wim = cos + o1 - 7;
505 TXUSample t1, t2, t3, t4, t5, t6, r0, i0, r1, i1;
506
507 for (int i = 0; i < len; i += 4) {
508 TRANSFORM(z[0], z[o1 + 0], z[o2 + 0], z[o3 + 0], cos[0], wim[7]);
509 TRANSFORM(z[2], z[o1 + 2], z[o2 + 2], z[o3 + 2], cos[2], wim[5]);
510 TRANSFORM(z[4], z[o1 + 4], z[o2 + 4], z[o3 + 4], cos[4], wim[3]);
511 TRANSFORM(z[6], z[o1 + 6], z[o2 + 6], z[o3 + 6], cos[6], wim[1]);
512
513 TRANSFORM(z[1], z[o1 + 1], z[o2 + 1], z[o3 + 1], cos[1], wim[6]);
514 TRANSFORM(z[3], z[o1 + 3], z[o2 + 3], z[o3 + 3], cos[3], wim[4]);
515 TRANSFORM(z[5], z[o1 + 5], z[o2 + 5], z[o3 + 5], cos[5], wim[2]);
516 TRANSFORM(z[7], z[o1 + 7], z[o2 + 7], z[o3 + 7], cos[7], wim[0]);
517
518 z += 2*4;
519 cos += 2*4;
520 wim -= 2*4;
521 }
522 }
523
TX_NAME(ff_tx_fft_sr_codelet_init)524 static av_cold int TX_NAME(ff_tx_fft_sr_codelet_init)(AVTXContext *s,
525 const FFTXCodelet *cd,
526 uint64_t flags,
527 FFTXCodeletOptions *opts,
528 int len, int inv,
529 const void *scale)
530 {
531 TX_TAB(ff_tx_init_tabs)(len);
532 return ff_tx_gen_ptwo_revtab(s, opts ? opts->invert_lookup : 1);
533 }
534
535 #define DECL_SR_CODELET_DEF(n) \
536 static const FFTXCodelet TX_NAME(ff_tx_fft##n##_ns_def) = { \
537 .name = TX_NAME_STR("fft" #n "_ns"), \
538 .function = TX_NAME(ff_tx_fft##n##_ns), \
539 .type = TX_TYPE(FFT), \
540 .flags = AV_TX_INPLACE | AV_TX_UNALIGNED | \
541 FF_TX_PRESHUFFLE, \
542 .factors[0] = 2, \
543 .min_len = n, \
544 .max_len = n, \
545 .init = TX_NAME(ff_tx_fft_sr_codelet_init), \
546 .cpu_flags = FF_TX_CPU_FLAGS_ALL, \
547 .prio = FF_TX_PRIO_BASE, \
548 };
549
550 #define DECL_SR_CODELET(n, n2, n4) \
551 static void TX_NAME(ff_tx_fft##n##_ns)(AVTXContext *s, void *dst, \
552 void *src, ptrdiff_t stride) \
553 { \
554 TXComplex *z = dst; \
555 const TXSample *cos = TX_TAB(ff_tx_tab_##n); \
556 \
557 TX_NAME(ff_tx_fft##n2##_ns)(s, z, z, stride); \
558 TX_NAME(ff_tx_fft##n4##_ns)(s, z + n4*2, z + n4*2, stride); \
559 TX_NAME(ff_tx_fft##n4##_ns)(s, z + n4*3, z + n4*3, stride); \
560 TX_NAME(ff_tx_fft_sr_combine)(z, cos, n4 >> 1); \
561 } \
562 \
563 DECL_SR_CODELET_DEF(n)
564
TX_NAME(ff_tx_fft2_ns)565 static void TX_NAME(ff_tx_fft2_ns)(AVTXContext *s, void *dst,
566 void *src, ptrdiff_t stride)
567 {
568 TXComplex *z = dst;
569 TXComplex tmp;
570
571 BF(tmp.re, z[0].re, z[0].re, z[1].re);
572 BF(tmp.im, z[0].im, z[0].im, z[1].im);
573 z[1] = tmp;
574 }
575
TX_NAME(ff_tx_fft4_ns)576 static void TX_NAME(ff_tx_fft4_ns)(AVTXContext *s, void *dst,
577 void *src, ptrdiff_t stride)
578 {
579 TXComplex *z = dst;
580 TXSample t1, t2, t3, t4, t5, t6, t7, t8;
581
582 BF(t3, t1, z[0].re, z[1].re);
583 BF(t8, t6, z[3].re, z[2].re);
584 BF(z[2].re, z[0].re, t1, t6);
585 BF(t4, t2, z[0].im, z[1].im);
586 BF(t7, t5, z[2].im, z[3].im);
587 BF(z[3].im, z[1].im, t4, t8);
588 BF(z[3].re, z[1].re, t3, t7);
589 BF(z[2].im, z[0].im, t2, t5);
590 }
591
TX_NAME(ff_tx_fft8_ns)592 static void TX_NAME(ff_tx_fft8_ns)(AVTXContext *s, void *dst,
593 void *src, ptrdiff_t stride)
594 {
595 TXComplex *z = dst;
596 TXSample t1, t2, t3, t4, t5, t6, r0, i0, r1, i1;
597 const TXSample cos = TX_TAB(ff_tx_tab_8)[1];
598
599 TX_NAME(ff_tx_fft4_ns)(s, z, z, stride);
600
601 BF(t1, z[5].re, z[4].re, -z[5].re);
602 BF(t2, z[5].im, z[4].im, -z[5].im);
603 BF(t5, z[7].re, z[6].re, -z[7].re);
604 BF(t6, z[7].im, z[6].im, -z[7].im);
605
606 BUTTERFLIES(z[0], z[2], z[4], z[6]);
607 TRANSFORM(z[1], z[3], z[5], z[7], cos, cos);
608 }
609
TX_NAME(ff_tx_fft16_ns)610 static void TX_NAME(ff_tx_fft16_ns)(AVTXContext *s, void *dst,
611 void *src, ptrdiff_t stride)
612 {
613 TXComplex *z = dst;
614 const TXSample *cos = TX_TAB(ff_tx_tab_16);
615
616 TXSample t1, t2, t3, t4, t5, t6, r0, i0, r1, i1;
617 TXSample cos_16_1 = cos[1];
618 TXSample cos_16_2 = cos[2];
619 TXSample cos_16_3 = cos[3];
620
621 TX_NAME(ff_tx_fft8_ns)(s, z + 0, z + 0, stride);
622 TX_NAME(ff_tx_fft4_ns)(s, z + 8, z + 8, stride);
623 TX_NAME(ff_tx_fft4_ns)(s, z + 12, z + 12, stride);
624
625 t1 = z[ 8].re;
626 t2 = z[ 8].im;
627 t5 = z[12].re;
628 t6 = z[12].im;
629 BUTTERFLIES(z[0], z[4], z[8], z[12]);
630
631 TRANSFORM(z[ 2], z[ 6], z[10], z[14], cos_16_2, cos_16_2);
632 TRANSFORM(z[ 1], z[ 5], z[ 9], z[13], cos_16_1, cos_16_3);
633 TRANSFORM(z[ 3], z[ 7], z[11], z[15], cos_16_3, cos_16_1);
634 }
635
636 DECL_SR_CODELET_DEF(2)
637 DECL_SR_CODELET_DEF(4)
638 DECL_SR_CODELET_DEF(8)
639 DECL_SR_CODELET_DEF(16)
640 DECL_SR_CODELET(32,16,8)
641 DECL_SR_CODELET(64,32,16)
642 DECL_SR_CODELET(128,64,32)
643 DECL_SR_CODELET(256,128,64)
644 DECL_SR_CODELET(512,256,128)
645 DECL_SR_CODELET(1024,512,256)
646 DECL_SR_CODELET(2048,1024,512)
647 DECL_SR_CODELET(4096,2048,1024)
648 DECL_SR_CODELET(8192,4096,2048)
649 DECL_SR_CODELET(16384,8192,4096)
650 DECL_SR_CODELET(32768,16384,8192)
651 DECL_SR_CODELET(65536,32768,16384)
652 DECL_SR_CODELET(131072,65536,32768)
653
TX_NAME(ff_tx_fft_sr_init)654 static av_cold int TX_NAME(ff_tx_fft_sr_init)(AVTXContext *s,
655 const FFTXCodelet *cd,
656 uint64_t flags,
657 FFTXCodeletOptions *opts,
658 int len, int inv,
659 const void *scale)
660 {
661 int ret;
662 int is_inplace = !!(flags & AV_TX_INPLACE);
663 FFTXCodeletOptions sub_opts = { .invert_lookup = !is_inplace };
664
665 flags &= ~FF_TX_OUT_OF_PLACE; /* We want the subtransform to be */
666 flags |= AV_TX_INPLACE; /* in-place */
667 flags |= FF_TX_PRESHUFFLE; /* This function handles the permute step */
668
669 if ((ret = ff_tx_init_subtx(s, TX_TYPE(FFT), flags, &sub_opts, len, inv, scale)))
670 return ret;
671
672 if (is_inplace && (ret = ff_tx_gen_ptwo_inplace_revtab_idx(s)))
673 return ret;
674
675 return 0;
676 }
677
TX_NAME(ff_tx_fft_sr)678 static void TX_NAME(ff_tx_fft_sr)(AVTXContext *s, void *_dst,
679 void *_src, ptrdiff_t stride)
680 {
681 TXComplex *src = _src;
682 TXComplex *dst = _dst;
683 int *map = s->sub[0].map;
684 int len = s->len;
685
686 /* Compilers can't vectorize this anyway without assuming AVX2, which they
687 * generally don't, at least without -march=native -mtune=native */
688 for (int i = 0; i < len; i++)
689 dst[i] = src[map[i]];
690
691 s->fn[0](&s->sub[0], dst, dst, stride);
692 }
693
TX_NAME(ff_tx_fft_sr_inplace)694 static void TX_NAME(ff_tx_fft_sr_inplace)(AVTXContext *s, void *_dst,
695 void *_src, ptrdiff_t stride)
696 {
697 TXComplex *dst = _dst;
698 TXComplex tmp;
699 const int *map = s->sub->map;
700 const int *inplace_idx = s->map;
701 int src_idx, dst_idx;
702
703 src_idx = *inplace_idx++;
704 do {
705 tmp = dst[src_idx];
706 dst_idx = map[src_idx];
707 do {
708 FFSWAP(TXComplex, tmp, dst[dst_idx]);
709 dst_idx = map[dst_idx];
710 } while (dst_idx != src_idx); /* Can be > as well, but was less predictable */
711 dst[dst_idx] = tmp;
712 } while ((src_idx = *inplace_idx++));
713
714 s->fn[0](&s->sub[0], dst, dst, stride);
715 }
716
717 static const FFTXCodelet TX_NAME(ff_tx_fft_sr_def) = {
718 .name = TX_NAME_STR("fft_sr"),
719 .function = TX_NAME(ff_tx_fft_sr),
720 .type = TX_TYPE(FFT),
721 .flags = AV_TX_UNALIGNED | FF_TX_OUT_OF_PLACE,
722 .factors[0] = 2,
723 .min_len = 2,
724 .max_len = TX_LEN_UNLIMITED,
725 .init = TX_NAME(ff_tx_fft_sr_init),
726 .cpu_flags = FF_TX_CPU_FLAGS_ALL,
727 .prio = FF_TX_PRIO_BASE,
728 };
729
730 static const FFTXCodelet TX_NAME(ff_tx_fft_sr_inplace_def) = {
731 .name = TX_NAME_STR("fft_sr_inplace"),
732 .function = TX_NAME(ff_tx_fft_sr_inplace),
733 .type = TX_TYPE(FFT),
734 .flags = AV_TX_UNALIGNED | AV_TX_INPLACE,
735 .factors[0] = 2,
736 .min_len = 2,
737 .max_len = TX_LEN_UNLIMITED,
738 .init = TX_NAME(ff_tx_fft_sr_init),
739 .cpu_flags = FF_TX_CPU_FLAGS_ALL,
740 .prio = FF_TX_PRIO_BASE,
741 };
742
TX_NAME(ff_tx_fft_naive)743 static void TX_NAME(ff_tx_fft_naive)(AVTXContext *s, void *_dst, void *_src,
744 ptrdiff_t stride)
745 {
746 TXComplex *src = _src;
747 TXComplex *dst = _dst;
748 const int n = s->len;
749 double phase = s->inv ? 2.0*M_PI/n : -2.0*M_PI/n;
750
751 for(int i = 0; i < n; i++) {
752 TXComplex tmp = { 0 };
753 for(int j = 0; j < n; j++) {
754 const double factor = phase*i*j;
755 const TXComplex mult = {
756 RESCALE(cos(factor)),
757 RESCALE(sin(factor)),
758 };
759 TXComplex res;
760 CMUL3(res, src[j], mult);
761 tmp.re += res.re;
762 tmp.im += res.im;
763 }
764 dst[i] = tmp;
765 }
766 }
767
768 static const FFTXCodelet TX_NAME(ff_tx_fft_naive_def) = {
769 .name = TX_NAME_STR("fft_naive"),
770 .function = TX_NAME(ff_tx_fft_naive),
771 .type = TX_TYPE(FFT),
772 .flags = AV_TX_UNALIGNED | FF_TX_OUT_OF_PLACE,
773 .factors[0] = TX_FACTOR_ANY,
774 .min_len = 2,
775 .max_len = TX_LEN_UNLIMITED,
776 .init = NULL,
777 .cpu_flags = FF_TX_CPU_FLAGS_ALL,
778 .prio = FF_TX_PRIO_MIN,
779 };
780
TX_NAME(ff_tx_fft_pfa_init)781 static av_cold int TX_NAME(ff_tx_fft_pfa_init)(AVTXContext *s,
782 const FFTXCodelet *cd,
783 uint64_t flags,
784 FFTXCodeletOptions *opts,
785 int len, int inv,
786 const void *scale)
787 {
788 int ret;
789 int sub_len = len / cd->factors[0];
790 FFTXCodeletOptions sub_opts = { .invert_lookup = 0 };
791
792 flags &= ~FF_TX_OUT_OF_PLACE; /* We want the subtransform to be */
793 flags |= AV_TX_INPLACE; /* in-place */
794 flags |= FF_TX_PRESHUFFLE; /* This function handles the permute step */
795
796 if ((ret = ff_tx_init_subtx(s, TX_TYPE(FFT), flags, &sub_opts,
797 sub_len, inv, scale)))
798 return ret;
799
800 if ((ret = ff_tx_gen_compound_mapping(s, cd->factors[0], sub_len)))
801 return ret;
802
803 if (!(s->tmp = av_malloc(len*sizeof(*s->tmp))))
804 return AVERROR(ENOMEM);
805
806 TX_TAB(ff_tx_init_tabs)(len / sub_len);
807
808 return 0;
809 }
810
811 #define DECL_COMP_FFT(N) \
812 static void TX_NAME(ff_tx_fft_pfa_##N##xM)(AVTXContext *s, void *_out, \
813 void *_in, ptrdiff_t stride) \
814 { \
815 const int m = s->sub->len; \
816 const int *in_map = s->map, *out_map = in_map + s->len; \
817 const int *sub_map = s->sub->map; \
818 TXComplex *in = _in; \
819 TXComplex *out = _out; \
820 TXComplex fft##N##in[N]; \
821 \
822 for (int i = 0; i < m; i++) { \
823 for (int j = 0; j < N; j++) \
824 fft##N##in[j] = in[in_map[i*N + j]]; \
825 fft##N(s->tmp + sub_map[i], fft##N##in, m); \
826 } \
827 \
828 for (int i = 0; i < N; i++) \
829 s->fn[0](&s->sub[0], s->tmp + m*i, s->tmp + m*i, sizeof(TXComplex)); \
830 \
831 for (int i = 0; i < N*m; i++) \
832 out[i] = s->tmp[out_map[i]]; \
833 } \
834 \
835 static const FFTXCodelet TX_NAME(ff_tx_fft_pfa_##N##xM_def) = { \
836 .name = TX_NAME_STR("fft_pfa_" #N "xM"), \
837 .function = TX_NAME(ff_tx_fft_pfa_##N##xM), \
838 .type = TX_TYPE(FFT), \
839 .flags = AV_TX_UNALIGNED | FF_TX_OUT_OF_PLACE, \
840 .factors = { N, TX_FACTOR_ANY }, \
841 .min_len = N*2, \
842 .max_len = TX_LEN_UNLIMITED, \
843 .init = TX_NAME(ff_tx_fft_pfa_init), \
844 .cpu_flags = FF_TX_CPU_FLAGS_ALL, \
845 .prio = FF_TX_PRIO_BASE, \
846 };
847
848 DECL_COMP_FFT(3)
849 DECL_COMP_FFT(5)
850 DECL_COMP_FFT(7)
851 DECL_COMP_FFT(9)
852 DECL_COMP_FFT(15)
853
TX_NAME(ff_tx_mdct_naive_init)854 static av_cold int TX_NAME(ff_tx_mdct_naive_init)(AVTXContext *s,
855 const FFTXCodelet *cd,
856 uint64_t flags,
857 FFTXCodeletOptions *opts,
858 int len, int inv,
859 const void *scale)
860 {
861 s->scale_d = *((SCALE_TYPE *)scale);
862 s->scale_f = s->scale_d;
863 return 0;
864 }
865
TX_NAME(ff_tx_mdct_naive_fwd)866 static void TX_NAME(ff_tx_mdct_naive_fwd)(AVTXContext *s, void *_dst,
867 void *_src, ptrdiff_t stride)
868 {
869 TXSample *src = _src;
870 TXSample *dst = _dst;
871 double scale = s->scale_d;
872 int len = s->len;
873 const double phase = M_PI/(4.0*len);
874
875 stride /= sizeof(*dst);
876
877 for (int i = 0; i < len; i++) {
878 double sum = 0.0;
879 for (int j = 0; j < len*2; j++) {
880 int a = (2*j + 1 + len) * (2*i + 1);
881 sum += UNSCALE(src[j]) * cos(a * phase);
882 }
883 dst[i*stride] = RESCALE(sum*scale);
884 }
885 }
886
TX_NAME(ff_tx_mdct_naive_inv)887 static void TX_NAME(ff_tx_mdct_naive_inv)(AVTXContext *s, void *_dst,
888 void *_src, ptrdiff_t stride)
889 {
890 TXSample *src = _src;
891 TXSample *dst = _dst;
892 double scale = s->scale_d;
893 int len = s->len >> 1;
894 int len2 = len*2;
895 const double phase = M_PI/(4.0*len2);
896
897 stride /= sizeof(*src);
898
899 for (int i = 0; i < len; i++) {
900 double sum_d = 0.0;
901 double sum_u = 0.0;
902 double i_d = phase * (4*len - 2*i - 1);
903 double i_u = phase * (3*len2 + 2*i + 1);
904 for (int j = 0; j < len2; j++) {
905 double a = (2 * j + 1);
906 double a_d = cos(a * i_d);
907 double a_u = cos(a * i_u);
908 double val = UNSCALE(src[j*stride]);
909 sum_d += a_d * val;
910 sum_u += a_u * val;
911 }
912 dst[i + 0] = RESCALE( sum_d*scale);
913 dst[i + len] = RESCALE(-sum_u*scale);
914 }
915 }
916
917 static const FFTXCodelet TX_NAME(ff_tx_mdct_naive_fwd_def) = {
918 .name = TX_NAME_STR("mdct_naive_fwd"),
919 .function = TX_NAME(ff_tx_mdct_naive_fwd),
920 .type = TX_TYPE(MDCT),
921 .flags = AV_TX_UNALIGNED | FF_TX_OUT_OF_PLACE | FF_TX_FORWARD_ONLY,
922 .factors = { 2, TX_FACTOR_ANY }, /* MDCTs need an even length */
923 .min_len = 2,
924 .max_len = TX_LEN_UNLIMITED,
925 .init = TX_NAME(ff_tx_mdct_naive_init),
926 .cpu_flags = FF_TX_CPU_FLAGS_ALL,
927 .prio = FF_TX_PRIO_MIN,
928 };
929
930 static const FFTXCodelet TX_NAME(ff_tx_mdct_naive_inv_def) = {
931 .name = TX_NAME_STR("mdct_naive_inv"),
932 .function = TX_NAME(ff_tx_mdct_naive_inv),
933 .type = TX_TYPE(MDCT),
934 .flags = AV_TX_UNALIGNED | FF_TX_OUT_OF_PLACE | FF_TX_INVERSE_ONLY,
935 .factors = { 2, TX_FACTOR_ANY },
936 .min_len = 2,
937 .max_len = TX_LEN_UNLIMITED,
938 .init = TX_NAME(ff_tx_mdct_naive_init),
939 .cpu_flags = FF_TX_CPU_FLAGS_ALL,
940 .prio = FF_TX_PRIO_MIN,
941 };
942
TX_NAME(ff_tx_mdct_sr_init)943 static av_cold int TX_NAME(ff_tx_mdct_sr_init)(AVTXContext *s,
944 const FFTXCodelet *cd,
945 uint64_t flags,
946 FFTXCodeletOptions *opts,
947 int len, int inv,
948 const void *scale)
949 {
950 int ret;
951 FFTXCodeletOptions sub_opts = { .invert_lookup = 0 };
952
953 s->scale_d = *((SCALE_TYPE *)scale);
954 s->scale_f = s->scale_d;
955
956 flags &= ~FF_TX_OUT_OF_PLACE; /* We want the subtransform to be */
957 flags |= AV_TX_INPLACE; /* in-place */
958 flags |= FF_TX_PRESHUFFLE; /* This function handles the permute step */
959
960 if ((ret = ff_tx_init_subtx(s, TX_TYPE(FFT), flags, &sub_opts, len >> 1,
961 inv, scale)))
962 return ret;
963
964 if ((ret = TX_TAB(ff_tx_mdct_gen_exp)(s)))
965 return ret;
966
967 return 0;
968 }
969
TX_NAME(ff_tx_mdct_sr_fwd)970 static void TX_NAME(ff_tx_mdct_sr_fwd)(AVTXContext *s, void *_dst, void *_src,
971 ptrdiff_t stride)
972 {
973 TXSample *src = _src, *dst = _dst;
974 TXComplex *exp = s->exp, tmp, *z = _dst;
975 const int len2 = s->len >> 1;
976 const int len4 = s->len >> 2;
977 const int len3 = len2 * 3;
978 const int *sub_map = s->sub->map;
979
980 stride /= sizeof(*dst);
981
982 for (int i = 0; i < len2; i++) { /* Folding and pre-reindexing */
983 const int k = 2*i;
984 const int idx = sub_map[i];
985 if (k < len2) {
986 tmp.re = FOLD(-src[ len2 + k], src[1*len2 - 1 - k]);
987 tmp.im = FOLD(-src[ len3 + k], -src[1*len3 - 1 - k]);
988 } else {
989 tmp.re = FOLD(-src[ len2 + k], -src[5*len2 - 1 - k]);
990 tmp.im = FOLD( src[-len2 + k], -src[1*len3 - 1 - k]);
991 }
992 CMUL(z[idx].im, z[idx].re, tmp.re, tmp.im, exp[i].re, exp[i].im);
993 }
994
995 s->fn[0](&s->sub[0], z, z, sizeof(TXComplex));
996
997 for (int i = 0; i < len4; i++) {
998 const int i0 = len4 + i, i1 = len4 - i - 1;
999 TXComplex src1 = { z[i1].re, z[i1].im };
1000 TXComplex src0 = { z[i0].re, z[i0].im };
1001
1002 CMUL(dst[2*i1*stride + stride], dst[2*i0*stride], src0.re, src0.im,
1003 exp[i0].im, exp[i0].re);
1004 CMUL(dst[2*i0*stride + stride], dst[2*i1*stride], src1.re, src1.im,
1005 exp[i1].im, exp[i1].re);
1006 }
1007 }
1008
TX_NAME(ff_tx_mdct_sr_inv)1009 static void TX_NAME(ff_tx_mdct_sr_inv)(AVTXContext *s, void *_dst, void *_src,
1010 ptrdiff_t stride)
1011 {
1012 TXComplex *z = _dst, *exp = s->exp;
1013 const TXSample *src = _src, *in1, *in2;
1014 const int len2 = s->len >> 1;
1015 const int len4 = s->len >> 2;
1016 const int *sub_map = s->sub->map;
1017
1018 stride /= sizeof(*src);
1019 in1 = src;
1020 in2 = src + ((len2*2) - 1) * stride;
1021
1022 for (int i = 0; i < len2; i++) {
1023 TXComplex tmp = { in2[-2*i*stride], in1[2*i*stride] };
1024 CMUL3(z[sub_map[i]], tmp, exp[i]);
1025 }
1026
1027 s->fn[0](&s->sub[0], z, z, sizeof(TXComplex));
1028
1029 for (int i = 0; i < len4; i++) {
1030 const int i0 = len4 + i, i1 = len4 - i - 1;
1031 TXComplex src1 = { z[i1].im, z[i1].re };
1032 TXComplex src0 = { z[i0].im, z[i0].re };
1033
1034 CMUL(z[i1].re, z[i0].im, src1.re, src1.im, exp[i1].im, exp[i1].re);
1035 CMUL(z[i0].re, z[i1].im, src0.re, src0.im, exp[i0].im, exp[i0].re);
1036 }
1037 }
1038
1039 static const FFTXCodelet TX_NAME(ff_tx_mdct_sr_fwd_def) = {
1040 .name = TX_NAME_STR("mdct_sr_fwd"),
1041 .function = TX_NAME(ff_tx_mdct_sr_fwd),
1042 .type = TX_TYPE(MDCT),
1043 .flags = AV_TX_UNALIGNED | FF_TX_OUT_OF_PLACE | FF_TX_FORWARD_ONLY,
1044 .factors[0] = 2,
1045 .min_len = 2,
1046 .max_len = TX_LEN_UNLIMITED,
1047 .init = TX_NAME(ff_tx_mdct_sr_init),
1048 .cpu_flags = FF_TX_CPU_FLAGS_ALL,
1049 .prio = FF_TX_PRIO_BASE,
1050 };
1051
1052 static const FFTXCodelet TX_NAME(ff_tx_mdct_sr_inv_def) = {
1053 .name = TX_NAME_STR("mdct_sr_inv"),
1054 .function = TX_NAME(ff_tx_mdct_sr_inv),
1055 .type = TX_TYPE(MDCT),
1056 .flags = AV_TX_UNALIGNED | FF_TX_OUT_OF_PLACE | FF_TX_INVERSE_ONLY,
1057 .factors[0] = 2,
1058 .min_len = 2,
1059 .max_len = TX_LEN_UNLIMITED,
1060 .init = TX_NAME(ff_tx_mdct_sr_init),
1061 .cpu_flags = FF_TX_CPU_FLAGS_ALL,
1062 .prio = FF_TX_PRIO_BASE,
1063 };
1064
TX_NAME(ff_tx_mdct_inv_full_init)1065 static av_cold int TX_NAME(ff_tx_mdct_inv_full_init)(AVTXContext *s,
1066 const FFTXCodelet *cd,
1067 uint64_t flags,
1068 FFTXCodeletOptions *opts,
1069 int len, int inv,
1070 const void *scale)
1071 {
1072 int ret;
1073
1074 s->scale_d = *((SCALE_TYPE *)scale);
1075 s->scale_f = s->scale_d;
1076
1077 flags &= ~AV_TX_FULL_IMDCT;
1078
1079 if ((ret = ff_tx_init_subtx(s, TX_TYPE(MDCT), flags, NULL, len, 1, scale)))
1080 return ret;
1081
1082 return 0;
1083 }
1084
TX_NAME(ff_tx_mdct_inv_full)1085 static void TX_NAME(ff_tx_mdct_inv_full)(AVTXContext *s, void *_dst,
1086 void *_src, ptrdiff_t stride)
1087 {
1088 int len = s->len << 1;
1089 int len2 = len >> 1;
1090 int len4 = len >> 2;
1091 TXSample *dst = _dst;
1092
1093 s->fn[0](&s->sub[0], dst + len4, _src, stride);
1094
1095 stride /= sizeof(*dst);
1096
1097 for (int i = 0; i < len4; i++) {
1098 dst[ i*stride] = -dst[(len2 - i - 1)*stride];
1099 dst[(len - i - 1)*stride] = dst[(len2 + i + 0)*stride];
1100 }
1101 }
1102
1103 static const FFTXCodelet TX_NAME(ff_tx_mdct_inv_full_def) = {
1104 .name = TX_NAME_STR("mdct_inv_full"),
1105 .function = TX_NAME(ff_tx_mdct_inv_full),
1106 .type = TX_TYPE(MDCT),
1107 .flags = AV_TX_UNALIGNED | AV_TX_INPLACE |
1108 FF_TX_OUT_OF_PLACE | AV_TX_FULL_IMDCT,
1109 .factors = { 2, TX_FACTOR_ANY },
1110 .min_len = 2,
1111 .max_len = TX_LEN_UNLIMITED,
1112 .init = TX_NAME(ff_tx_mdct_inv_full_init),
1113 .cpu_flags = FF_TX_CPU_FLAGS_ALL,
1114 .prio = FF_TX_PRIO_BASE,
1115 };
1116
TX_NAME(ff_tx_mdct_pfa_init)1117 static av_cold int TX_NAME(ff_tx_mdct_pfa_init)(AVTXContext *s,
1118 const FFTXCodelet *cd,
1119 uint64_t flags,
1120 FFTXCodeletOptions *opts,
1121 int len, int inv,
1122 const void *scale)
1123 {
1124 int ret, sub_len;
1125 FFTXCodeletOptions sub_opts = { .invert_lookup = 0 };
1126
1127 len >>= 1;
1128 sub_len = len / cd->factors[0];
1129
1130 s->scale_d = *((SCALE_TYPE *)scale);
1131 s->scale_f = s->scale_d;
1132
1133 flags &= ~FF_TX_OUT_OF_PLACE; /* We want the subtransform to be */
1134 flags |= AV_TX_INPLACE; /* in-place */
1135 flags |= FF_TX_PRESHUFFLE; /* This function handles the permute step */
1136
1137 if ((ret = ff_tx_init_subtx(s, TX_TYPE(FFT), flags, &sub_opts,
1138 sub_len, inv, scale)))
1139 return ret;
1140
1141 if ((ret = ff_tx_gen_compound_mapping(s, cd->factors[0], sub_len)))
1142 return ret;
1143
1144 if ((ret = TX_TAB(ff_tx_mdct_gen_exp)(s)))
1145 return ret;
1146
1147 if (!(s->tmp = av_malloc(len*sizeof(*s->tmp))))
1148 return AVERROR(ENOMEM);
1149
1150 TX_TAB(ff_tx_init_tabs)(len / sub_len);
1151
1152 return 0;
1153 }
1154
1155 #define DECL_COMP_IMDCT(N) \
1156 static void TX_NAME(ff_tx_mdct_pfa_##N##xM_inv)(AVTXContext *s, void *_dst, \
1157 void *_src, ptrdiff_t stride) \
1158 { \
1159 TXComplex fft##N##in[N]; \
1160 TXComplex *z = _dst, *exp = s->exp; \
1161 const TXSample *src = _src, *in1, *in2; \
1162 const int len4 = s->len >> 2; \
1163 const int m = s->sub->len; \
1164 const int *in_map = s->map, *out_map = in_map + N*m; \
1165 const int *sub_map = s->sub->map; \
1166 \
1167 stride /= sizeof(*src); /* To convert it from bytes */ \
1168 in1 = src; \
1169 in2 = src + ((N*m*2) - 1) * stride; \
1170 \
1171 for (int i = 0; i < m; i++) { \
1172 for (int j = 0; j < N; j++) { \
1173 const int k = in_map[i*N + j]; \
1174 TXComplex tmp = { in2[-k*stride], in1[k*stride] }; \
1175 CMUL3(fft##N##in[j], tmp, exp[k >> 1]); \
1176 } \
1177 fft##N(s->tmp + sub_map[i], fft##N##in, m); \
1178 } \
1179 \
1180 for (int i = 0; i < N; i++) \
1181 s->fn[0](&s->sub[0], s->tmp + m*i, s->tmp + m*i, sizeof(TXComplex)); \
1182 \
1183 for (int i = 0; i < len4; i++) { \
1184 const int i0 = len4 + i, i1 = len4 - i - 1; \
1185 const int s0 = out_map[i0], s1 = out_map[i1]; \
1186 TXComplex src1 = { s->tmp[s1].im, s->tmp[s1].re }; \
1187 TXComplex src0 = { s->tmp[s0].im, s->tmp[s0].re }; \
1188 \
1189 CMUL(z[i1].re, z[i0].im, src1.re, src1.im, exp[i1].im, exp[i1].re); \
1190 CMUL(z[i0].re, z[i1].im, src0.re, src0.im, exp[i0].im, exp[i0].re); \
1191 } \
1192 } \
1193 \
1194 static const FFTXCodelet TX_NAME(ff_tx_mdct_pfa_##N##xM_inv_def) = { \
1195 .name = TX_NAME_STR("mdct_pfa_" #N "xM_inv"), \
1196 .function = TX_NAME(ff_tx_mdct_pfa_##N##xM_inv), \
1197 .type = TX_TYPE(MDCT), \
1198 .flags = AV_TX_UNALIGNED | FF_TX_OUT_OF_PLACE | FF_TX_INVERSE_ONLY, \
1199 .factors = { N, TX_FACTOR_ANY }, \
1200 .min_len = N*2, \
1201 .max_len = TX_LEN_UNLIMITED, \
1202 .init = TX_NAME(ff_tx_mdct_pfa_init), \
1203 .cpu_flags = FF_TX_CPU_FLAGS_ALL, \
1204 .prio = FF_TX_PRIO_BASE, \
1205 };
1206
1207 DECL_COMP_IMDCT(3)
1208 DECL_COMP_IMDCT(5)
1209 DECL_COMP_IMDCT(7)
1210 DECL_COMP_IMDCT(9)
1211 DECL_COMP_IMDCT(15)
1212
1213 #define DECL_COMP_MDCT(N) \
1214 static void TX_NAME(ff_tx_mdct_pfa_##N##xM_fwd)(AVTXContext *s, void *_dst, \
1215 void *_src, ptrdiff_t stride) \
1216 { \
1217 TXComplex fft##N##in[N]; \
1218 TXSample *src = _src, *dst = _dst; \
1219 TXComplex *exp = s->exp, tmp; \
1220 const int m = s->sub->len; \
1221 const int len4 = N*m; \
1222 const int len3 = len4 * 3; \
1223 const int len8 = s->len >> 2; \
1224 const int *in_map = s->map, *out_map = in_map + N*m; \
1225 const int *sub_map = s->sub->map; \
1226 \
1227 stride /= sizeof(*dst); \
1228 \
1229 for (int i = 0; i < m; i++) { /* Folding and pre-reindexing */ \
1230 for (int j = 0; j < N; j++) { \
1231 const int k = in_map[i*N + j]; \
1232 if (k < len4) { \
1233 tmp.re = FOLD(-src[ len4 + k], src[1*len4 - 1 - k]); \
1234 tmp.im = FOLD(-src[ len3 + k], -src[1*len3 - 1 - k]); \
1235 } else { \
1236 tmp.re = FOLD(-src[ len4 + k], -src[5*len4 - 1 - k]); \
1237 tmp.im = FOLD( src[-len4 + k], -src[1*len3 - 1 - k]); \
1238 } \
1239 CMUL(fft##N##in[j].im, fft##N##in[j].re, tmp.re, tmp.im, \
1240 exp[k >> 1].re, exp[k >> 1].im); \
1241 } \
1242 fft##N(s->tmp + sub_map[i], fft##N##in, m); \
1243 } \
1244 \
1245 for (int i = 0; i < N; i++) \
1246 s->fn[0](&s->sub[0], s->tmp + m*i, s->tmp + m*i, sizeof(TXComplex)); \
1247 \
1248 for (int i = 0; i < len8; i++) { \
1249 const int i0 = len8 + i, i1 = len8 - i - 1; \
1250 const int s0 = out_map[i0], s1 = out_map[i1]; \
1251 TXComplex src1 = { s->tmp[s1].re, s->tmp[s1].im }; \
1252 TXComplex src0 = { s->tmp[s0].re, s->tmp[s0].im }; \
1253 \
1254 CMUL(dst[2*i1*stride + stride], dst[2*i0*stride], src0.re, src0.im, \
1255 exp[i0].im, exp[i0].re); \
1256 CMUL(dst[2*i0*stride + stride], dst[2*i1*stride], src1.re, src1.im, \
1257 exp[i1].im, exp[i1].re); \
1258 } \
1259 } \
1260 \
1261 static const FFTXCodelet TX_NAME(ff_tx_mdct_pfa_##N##xM_fwd_def) = { \
1262 .name = TX_NAME_STR("mdct_pfa_" #N "xM_fwd"), \
1263 .function = TX_NAME(ff_tx_mdct_pfa_##N##xM_fwd), \
1264 .type = TX_TYPE(MDCT), \
1265 .flags = AV_TX_UNALIGNED | FF_TX_OUT_OF_PLACE | FF_TX_FORWARD_ONLY, \
1266 .factors = { N, TX_FACTOR_ANY }, \
1267 .min_len = N*2, \
1268 .max_len = TX_LEN_UNLIMITED, \
1269 .init = TX_NAME(ff_tx_mdct_pfa_init), \
1270 .cpu_flags = FF_TX_CPU_FLAGS_ALL, \
1271 .prio = FF_TX_PRIO_BASE, \
1272 };
1273
1274 DECL_COMP_MDCT(3)
1275 DECL_COMP_MDCT(5)
1276 DECL_COMP_MDCT(7)
1277 DECL_COMP_MDCT(9)
1278 DECL_COMP_MDCT(15)
1279
TX_NAME(ff_tx_rdft_init)1280 static av_cold int TX_NAME(ff_tx_rdft_init)(AVTXContext *s,
1281 const FFTXCodelet *cd,
1282 uint64_t flags,
1283 FFTXCodeletOptions *opts,
1284 int len, int inv,
1285 const void *scale)
1286 {
1287 int ret;
1288 double f, m;
1289 TXSample *tab;
1290
1291 s->scale_d = *((SCALE_TYPE *)scale);
1292 s->scale_f = s->scale_d;
1293
1294 if ((ret = ff_tx_init_subtx(s, TX_TYPE(FFT), flags, NULL, len >> 1, inv, scale)))
1295 return ret;
1296
1297 if (!(s->exp = av_mallocz((8 + (len >> 2) - 1)*sizeof(*s->exp))))
1298 return AVERROR(ENOMEM);
1299
1300 tab = (TXSample *)s->exp;
1301
1302 f = 2*M_PI/len;
1303
1304 m = (inv ? 2*s->scale_d : s->scale_d);
1305
1306 *tab++ = RESCALE((inv ? 0.5 : 1.0) * m);
1307 *tab++ = RESCALE(inv ? 0.5*m : 1.0);
1308 *tab++ = RESCALE( m);
1309 *tab++ = RESCALE(-m);
1310
1311 *tab++ = RESCALE( (0.5 - 0.0) * m);
1312 *tab++ = RESCALE( (0.0 - 0.5) * m);
1313 *tab++ = RESCALE( (0.5 - inv) * m);
1314 *tab++ = RESCALE(-(0.5 - inv) * m);
1315
1316 for (int i = 0; i < len >> 2; i++)
1317 *tab++ = RESCALE(cos(i*f));
1318 for (int i = len >> 2; i >= 0; i--)
1319 *tab++ = RESCALE(cos(i*f) * (inv ? +1.0 : -1.0));
1320
1321 return 0;
1322 }
1323
1324 #define DECL_RDFT(name, inv) \
1325 static void TX_NAME(ff_tx_rdft_ ##name)(AVTXContext *s, void *_dst, \
1326 void *_src, ptrdiff_t stride) \
1327 { \
1328 const int len2 = s->len >> 1; \
1329 const int len4 = s->len >> 2; \
1330 const TXSample *fact = (void *)s->exp; \
1331 const TXSample *tcos = fact + 8; \
1332 const TXSample *tsin = tcos + len4; \
1333 TXComplex *data = inv ? _src : _dst; \
1334 TXComplex t[3]; \
1335 \
1336 if (!inv) \
1337 s->fn[0](&s->sub[0], data, _src, sizeof(TXComplex)); \
1338 else \
1339 data[0].im = data[len2].re; \
1340 \
1341 /* The DC value's both components are real, but we need to change them \
1342 * into complex values. Also, the middle of the array is special-cased. \
1343 * These operations can be done before or after the loop. */ \
1344 t[0].re = data[0].re; \
1345 data[0].re = t[0].re + data[0].im; \
1346 data[0].im = t[0].re - data[0].im; \
1347 data[ 0].re = MULT(fact[0], data[ 0].re); \
1348 data[ 0].im = MULT(fact[1], data[ 0].im); \
1349 data[len4].re = MULT(fact[2], data[len4].re); \
1350 data[len4].im = MULT(fact[3], data[len4].im); \
1351 \
1352 for (int i = 1; i < len4; i++) { \
1353 /* Separate even and odd FFTs */ \
1354 t[0].re = MULT(fact[4], (data[i].re + data[len2 - i].re)); \
1355 t[0].im = MULT(fact[5], (data[i].im - data[len2 - i].im)); \
1356 t[1].re = MULT(fact[6], (data[i].im + data[len2 - i].im)); \
1357 t[1].im = MULT(fact[7], (data[i].re - data[len2 - i].re)); \
1358 \
1359 /* Apply twiddle factors to the odd FFT and add to the even FFT */ \
1360 CMUL(t[2].re, t[2].im, t[1].re, t[1].im, tcos[i], tsin[i]); \
1361 \
1362 data[ i].re = t[0].re + t[2].re; \
1363 data[ i].im = t[2].im - t[0].im; \
1364 data[len2 - i].re = t[0].re - t[2].re; \
1365 data[len2 - i].im = t[2].im + t[0].im; \
1366 } \
1367 \
1368 if (inv) { \
1369 s->fn[0](&s->sub[0], _dst, data, sizeof(TXComplex)); \
1370 } else { \
1371 /* Move [0].im to the last position, as convention requires */ \
1372 data[len2].re = data[0].im; \
1373 data[ 0].im = 0; \
1374 } \
1375 }
1376
1377 DECL_RDFT(r2c, 0)
1378 DECL_RDFT(c2r, 1)
1379
1380 static const FFTXCodelet TX_NAME(ff_tx_rdft_r2c_def) = {
1381 .name = TX_NAME_STR("rdft_r2c"),
1382 .function = TX_NAME(ff_tx_rdft_r2c),
1383 .type = TX_TYPE(RDFT),
1384 .flags = AV_TX_UNALIGNED | AV_TX_INPLACE |
1385 FF_TX_OUT_OF_PLACE | FF_TX_FORWARD_ONLY,
1386 .factors = { 2, TX_FACTOR_ANY },
1387 .min_len = 2,
1388 .max_len = TX_LEN_UNLIMITED,
1389 .init = TX_NAME(ff_tx_rdft_init),
1390 .cpu_flags = FF_TX_CPU_FLAGS_ALL,
1391 .prio = FF_TX_PRIO_BASE,
1392 };
1393
1394 static const FFTXCodelet TX_NAME(ff_tx_rdft_c2r_def) = {
1395 .name = TX_NAME_STR("rdft_c2r"),
1396 .function = TX_NAME(ff_tx_rdft_c2r),
1397 .type = TX_TYPE(RDFT),
1398 .flags = AV_TX_UNALIGNED | AV_TX_INPLACE |
1399 FF_TX_OUT_OF_PLACE | FF_TX_INVERSE_ONLY,
1400 .factors = { 2, TX_FACTOR_ANY },
1401 .min_len = 2,
1402 .max_len = TX_LEN_UNLIMITED,
1403 .init = TX_NAME(ff_tx_rdft_init),
1404 .cpu_flags = FF_TX_CPU_FLAGS_ALL,
1405 .prio = FF_TX_PRIO_BASE,
1406 };
1407
TX_TAB(ff_tx_mdct_gen_exp)1408 int TX_TAB(ff_tx_mdct_gen_exp)(AVTXContext *s)
1409 {
1410 int len4 = s->len >> 1;
1411 double scale = s->scale_d;
1412 const double theta = (scale < 0 ? len4 : 0) + 1.0/8.0;
1413
1414 if (!(s->exp = av_malloc_array(len4, sizeof(*s->exp))))
1415 return AVERROR(ENOMEM);
1416
1417 scale = sqrt(fabs(scale));
1418 for (int i = 0; i < len4; i++) {
1419 const double alpha = M_PI_2 * (i + theta) / len4;
1420 s->exp[i].re = RESCALE(cos(alpha) * scale);
1421 s->exp[i].im = RESCALE(sin(alpha) * scale);
1422 }
1423
1424 return 0;
1425 }
1426
1427 const FFTXCodelet * const TX_NAME(ff_tx_codelet_list)[] = {
1428 /* Split-Radix codelets */
1429 &TX_NAME(ff_tx_fft2_ns_def),
1430 &TX_NAME(ff_tx_fft4_ns_def),
1431 &TX_NAME(ff_tx_fft8_ns_def),
1432 &TX_NAME(ff_tx_fft16_ns_def),
1433 &TX_NAME(ff_tx_fft32_ns_def),
1434 &TX_NAME(ff_tx_fft64_ns_def),
1435 &TX_NAME(ff_tx_fft128_ns_def),
1436 &TX_NAME(ff_tx_fft256_ns_def),
1437 &TX_NAME(ff_tx_fft512_ns_def),
1438 &TX_NAME(ff_tx_fft1024_ns_def),
1439 &TX_NAME(ff_tx_fft2048_ns_def),
1440 &TX_NAME(ff_tx_fft4096_ns_def),
1441 &TX_NAME(ff_tx_fft8192_ns_def),
1442 &TX_NAME(ff_tx_fft16384_ns_def),
1443 &TX_NAME(ff_tx_fft32768_ns_def),
1444 &TX_NAME(ff_tx_fft65536_ns_def),
1445 &TX_NAME(ff_tx_fft131072_ns_def),
1446
1447 /* Standalone transforms */
1448 &TX_NAME(ff_tx_fft_sr_def),
1449 &TX_NAME(ff_tx_fft_sr_inplace_def),
1450 &TX_NAME(ff_tx_fft_pfa_3xM_def),
1451 &TX_NAME(ff_tx_fft_pfa_5xM_def),
1452 &TX_NAME(ff_tx_fft_pfa_7xM_def),
1453 &TX_NAME(ff_tx_fft_pfa_9xM_def),
1454 &TX_NAME(ff_tx_fft_pfa_15xM_def),
1455 &TX_NAME(ff_tx_fft_naive_def),
1456 &TX_NAME(ff_tx_mdct_sr_fwd_def),
1457 &TX_NAME(ff_tx_mdct_sr_inv_def),
1458 &TX_NAME(ff_tx_mdct_pfa_3xM_fwd_def),
1459 &TX_NAME(ff_tx_mdct_pfa_5xM_fwd_def),
1460 &TX_NAME(ff_tx_mdct_pfa_7xM_fwd_def),
1461 &TX_NAME(ff_tx_mdct_pfa_9xM_fwd_def),
1462 &TX_NAME(ff_tx_mdct_pfa_15xM_fwd_def),
1463 &TX_NAME(ff_tx_mdct_pfa_3xM_inv_def),
1464 &TX_NAME(ff_tx_mdct_pfa_5xM_inv_def),
1465 &TX_NAME(ff_tx_mdct_pfa_7xM_inv_def),
1466 &TX_NAME(ff_tx_mdct_pfa_9xM_inv_def),
1467 &TX_NAME(ff_tx_mdct_pfa_15xM_inv_def),
1468 &TX_NAME(ff_tx_mdct_naive_fwd_def),
1469 &TX_NAME(ff_tx_mdct_naive_inv_def),
1470 &TX_NAME(ff_tx_mdct_inv_full_def),
1471 &TX_NAME(ff_tx_rdft_r2c_def),
1472 &TX_NAME(ff_tx_rdft_c2r_def),
1473
1474 NULL,
1475 };
1476