Mercurial > libavcodec.hg
annotate fft.c @ 10162:8d369aee733f libavcodec
ARM: NEON optimised MDCT
author | mru |
---|---|
date | Fri, 11 Sep 2009 02:24:22 +0000 |
parents | a349795e8dca |
children | eda985c53dba |
rev | line source |
---|---|
781 | 1 /* |
2 * FFT/IFFT transforms | |
7542 | 3 * Copyright (c) 2008 Loren Merritt |
8629
04423b2f6e0b
cosmetics: Remove pointless period after copyright statement non-sentences.
diego
parents:
8590
diff
changeset
|
4 * Copyright (c) 2002 Fabrice Bellard |
7542 | 5 * Partly based on libdjbfft by D. J. Bernstein |
781 | 6 * |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3746
diff
changeset
|
7 * This file is part of FFmpeg. |
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3746
diff
changeset
|
8 * |
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3746
diff
changeset
|
9 * FFmpeg is free software; you can redistribute it and/or |
781 | 10 * modify it under the terms of the GNU Lesser General Public |
11 * License as published by the Free Software Foundation; either | |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3746
diff
changeset
|
12 * version 2.1 of the License, or (at your option) any later version. |
781 | 13 * |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3746
diff
changeset
|
14 * FFmpeg is distributed in the hope that it will be useful, |
781 | 15 * but WITHOUT ANY WARRANTY; without even the implied warranty of |
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
17 * Lesser General Public License for more details. | |
18 * | |
19 * You should have received a copy of the GNU Lesser General Public | |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3746
diff
changeset
|
20 * License along with FFmpeg; if not, write to the Free Software |
3036
0b546eab515d
Update licensing information: The FSF changed postal address.
diego
parents:
2979
diff
changeset
|
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
781 | 22 */ |
1106 | 23 |
24 /** | |
8718
e9d9d946f213
Use full internal pathname in doxygen @file directives.
diego
parents:
8694
diff
changeset
|
25 * @file libavcodec/fft.c |
1106 | 26 * FFT/IFFT transforms. |
27 */ | |
28 | |
781 | 29 #include "dsputil.h" |
30 | |
7542 | 31 /* cos(2*pi*x/n) for 0<=x<=n/4, followed by its reverse */ |
32 DECLARE_ALIGNED_16(FFTSample, ff_cos_16[8]); | |
33 DECLARE_ALIGNED_16(FFTSample, ff_cos_32[16]); | |
34 DECLARE_ALIGNED_16(FFTSample, ff_cos_64[32]); | |
35 DECLARE_ALIGNED_16(FFTSample, ff_cos_128[64]); | |
36 DECLARE_ALIGNED_16(FFTSample, ff_cos_256[128]); | |
37 DECLARE_ALIGNED_16(FFTSample, ff_cos_512[256]); | |
38 DECLARE_ALIGNED_16(FFTSample, ff_cos_1024[512]); | |
39 DECLARE_ALIGNED_16(FFTSample, ff_cos_2048[1024]); | |
40 DECLARE_ALIGNED_16(FFTSample, ff_cos_4096[2048]); | |
41 DECLARE_ALIGNED_16(FFTSample, ff_cos_8192[4096]); | |
42 DECLARE_ALIGNED_16(FFTSample, ff_cos_16384[8192]); | |
43 DECLARE_ALIGNED_16(FFTSample, ff_cos_32768[16384]); | |
44 DECLARE_ALIGNED_16(FFTSample, ff_cos_65536[32768]); | |
10143
c5e8a5a044c3
The ff_cos_tabs table itself is constant, too, so mark it as such.
reimar
parents:
9177
diff
changeset
|
45 FFTSample * const ff_cos_tabs[] = { |
7542 | 46 ff_cos_16, ff_cos_32, ff_cos_64, ff_cos_128, ff_cos_256, ff_cos_512, ff_cos_1024, |
47 ff_cos_2048, ff_cos_4096, ff_cos_8192, ff_cos_16384, ff_cos_32768, ff_cos_65536, | |
48 }; | |
49 | |
50 static int split_radix_permutation(int i, int n, int inverse) | |
51 { | |
52 int m; | |
53 if(n <= 2) return i&1; | |
54 m = n >> 1; | |
55 if(!(i&m)) return split_radix_permutation(i, m, inverse)*2; | |
56 m >>= 1; | |
57 if(inverse == !(i&m)) return split_radix_permutation(i, m, inverse)*4 + 1; | |
58 else return split_radix_permutation(i, m, inverse)*4 - 1; | |
59 } | |
60 | |
8637 | 61 av_cold int ff_fft_init(FFTContext *s, int nbits, int inverse) |
781 | 62 { |
63 int i, j, m, n; | |
64 float alpha, c1, s1, s2; | |
7542 | 65 int split_radix = 1; |
6503 | 66 int av_unused has_vectors; |
10153 | 67 int revtab_shift = 0; |
2967 | 68 |
7542 | 69 if (nbits < 2 || nbits > 16) |
70 goto fail; | |
781 | 71 s->nbits = nbits; |
72 n = 1 << nbits; | |
73 | |
7542 | 74 s->tmp_buf = NULL; |
8974 | 75 s->exptab = av_malloc((n / 2) * sizeof(FFTComplex)); |
781 | 76 if (!s->exptab) |
77 goto fail; | |
78 s->revtab = av_malloc(n * sizeof(uint16_t)); | |
79 if (!s->revtab) | |
80 goto fail; | |
81 s->inverse = inverse; | |
82 | |
83 s2 = inverse ? 1.0 : -1.0; | |
2967 | 84 |
7542 | 85 s->fft_permute = ff_fft_permute_c; |
8974 | 86 s->fft_calc = ff_fft_calc_c; |
87 s->imdct_calc = ff_imdct_calc_c; | |
88 s->imdct_half = ff_imdct_half_c; | |
10161 | 89 s->mdct_calc = ff_mdct_calc_c; |
8974 | 90 s->exptab1 = NULL; |
781 | 91 |
8590 | 92 #if HAVE_MMX && HAVE_YASM |
6503 | 93 has_vectors = mm_support(); |
8981
dc19e4d7d0eb
Only enable SSE/3DNOW optimizations when they have been enabled at compilation.
diego
parents:
8974
diff
changeset
|
94 if (has_vectors & FF_MM_SSE && HAVE_SSE) { |
7542 | 95 /* SSE for P3/P4/K8 */ |
8974 | 96 s->imdct_calc = ff_imdct_calc_sse; |
97 s->imdct_half = ff_imdct_half_sse; | |
7542 | 98 s->fft_permute = ff_fft_permute_sse; |
8974 | 99 s->fft_calc = ff_fft_calc_sse; |
8981
dc19e4d7d0eb
Only enable SSE/3DNOW optimizations when they have been enabled at compilation.
diego
parents:
8974
diff
changeset
|
100 } else if (has_vectors & FF_MM_3DNOWEXT && HAVE_AMD3DNOWEXT) { |
7542 | 101 /* 3DNowEx for K7 */ |
6503 | 102 s->imdct_calc = ff_imdct_calc_3dn2; |
7263 | 103 s->imdct_half = ff_imdct_half_3dn2; |
8974 | 104 s->fft_calc = ff_fft_calc_3dn2; |
8981
dc19e4d7d0eb
Only enable SSE/3DNOW optimizations when they have been enabled at compilation.
diego
parents:
8974
diff
changeset
|
105 } else if (has_vectors & FF_MM_3DNOW && HAVE_AMD3DNOW) { |
6503 | 106 /* 3DNow! for K6-2/3 */ |
7544 | 107 s->imdct_calc = ff_imdct_calc_3dn; |
108 s->imdct_half = ff_imdct_half_3dn; | |
8974 | 109 s->fft_calc = ff_fft_calc_3dn; |
6503 | 110 } |
9177
4b1736ba9f2f
Remove long unused ALTIVEC_USE_REFERENCE_C_CODE ifdef; all other references
conrad
parents:
8981
diff
changeset
|
111 #elif HAVE_ALTIVEC |
6503 | 112 has_vectors = mm_support(); |
8104
0d108ec85620
Remove duplicated MM_* macros for CPU capabilities from dsputil.h.
rathann
parents:
7547
diff
changeset
|
113 if (has_vectors & FF_MM_ALTIVEC) { |
6503 | 114 s->fft_calc = ff_fft_calc_altivec; |
7542 | 115 split_radix = 0; |
6503 | 116 } |
10153 | 117 #elif HAVE_NEON |
118 s->fft_permute = ff_fft_permute_neon; | |
119 s->fft_calc = ff_fft_calc_neon; | |
120 s->imdct_calc = ff_imdct_calc_neon; | |
121 s->imdct_half = ff_imdct_half_neon; | |
10162 | 122 s->mdct_calc = ff_mdct_calc_neon; |
10153 | 123 revtab_shift = 3; |
6503 | 124 #endif |
781 | 125 |
7542 | 126 if (split_radix) { |
127 for(j=4; j<=nbits; j++) { | |
128 int m = 1<<j; | |
129 double freq = 2*M_PI/m; | |
130 FFTSample *tab = ff_cos_tabs[j-4]; | |
131 for(i=0; i<=m/4; i++) | |
132 tab[i] = cos(i*freq); | |
133 for(i=1; i<m/4; i++) | |
134 tab[m/2-i] = tab[i]; | |
135 } | |
136 for(i=0; i<n; i++) | |
10153 | 137 s->revtab[-split_radix_permutation(i, n, s->inverse) & (n-1)] = |
138 i << revtab_shift; | |
7542 | 139 s->tmp_buf = av_malloc(n * sizeof(FFTComplex)); |
140 } else { | |
6504 | 141 int np, nblocks, np2, l; |
142 FFTComplex *q; | |
2967 | 143 |
7542 | 144 for(i=0; i<(n/2); i++) { |
145 alpha = 2 * M_PI * (float)i / (float)n; | |
146 c1 = cos(alpha); | |
147 s1 = sin(alpha) * s2; | |
148 s->exptab[i].re = c1; | |
149 s->exptab[i].im = s1; | |
150 } | |
151 | |
6504 | 152 np = 1 << nbits; |
153 nblocks = np >> 3; | |
154 np2 = np >> 1; | |
155 s->exptab1 = av_malloc(np * 2 * sizeof(FFTComplex)); | |
156 if (!s->exptab1) | |
157 goto fail; | |
158 q = s->exptab1; | |
159 do { | |
160 for(l = 0; l < np2; l += 2 * nblocks) { | |
161 *q++ = s->exptab[l]; | |
162 *q++ = s->exptab[l + nblocks]; | |
975
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
971
diff
changeset
|
163 |
6504 | 164 q->re = -s->exptab[l].im; |
165 q->im = s->exptab[l].re; | |
166 q++; | |
167 q->re = -s->exptab[l + nblocks].im; | |
168 q->im = s->exptab[l + nblocks].re; | |
169 q++; | |
170 } | |
171 nblocks = nblocks >> 1; | |
172 } while (nblocks != 0); | |
173 av_freep(&s->exptab); | |
781 | 174 |
7543 | 175 /* compute bit reverse table */ |
176 for(i=0;i<n;i++) { | |
177 m=0; | |
178 for(j=0;j<nbits;j++) { | |
179 m |= ((i >> j) & 1) << (nbits-j-1); | |
180 } | |
181 s->revtab[i]=m; | |
781 | 182 } |
7542 | 183 } |
184 | |
781 | 185 return 0; |
186 fail: | |
187 av_freep(&s->revtab); | |
188 av_freep(&s->exptab); | |
189 av_freep(&s->exptab1); | |
7542 | 190 av_freep(&s->tmp_buf); |
781 | 191 return -1; |
192 } | |
193 | |
7542 | 194 void ff_fft_permute_c(FFTContext *s, FFTComplex *z) |
781 | 195 { |
196 int j, k, np; | |
197 FFTComplex tmp; | |
198 const uint16_t *revtab = s->revtab; | |
7542 | 199 np = 1 << s->nbits; |
200 | |
201 if (s->tmp_buf) { | |
202 /* TODO: handle split-radix permute in a more optimal way, probably in-place */ | |
203 for(j=0;j<np;j++) s->tmp_buf[revtab[j]] = z[j]; | |
204 memcpy(z, s->tmp_buf, np * sizeof(FFTComplex)); | |
205 return; | |
206 } | |
2967 | 207 |
781 | 208 /* reverse */ |
209 for(j=0;j<np;j++) { | |
210 k = revtab[j]; | |
211 if (k < j) { | |
212 tmp = z[k]; | |
213 z[k] = z[j]; | |
214 z[j] = tmp; | |
215 } | |
216 } | |
217 } | |
218 | |
8687 | 219 av_cold void ff_fft_end(FFTContext *s) |
781 | 220 { |
221 av_freep(&s->revtab); | |
222 av_freep(&s->exptab); | |
223 av_freep(&s->exptab1); | |
7542 | 224 av_freep(&s->tmp_buf); |
225 } | |
226 | |
227 #define sqrthalf (float)M_SQRT1_2 | |
228 | |
229 #define BF(x,y,a,b) {\ | |
230 x = a - b;\ | |
231 y = a + b;\ | |
232 } | |
233 | |
234 #define BUTTERFLIES(a0,a1,a2,a3) {\ | |
235 BF(t3, t5, t5, t1);\ | |
236 BF(a2.re, a0.re, a0.re, t5);\ | |
237 BF(a3.im, a1.im, a1.im, t3);\ | |
238 BF(t4, t6, t2, t6);\ | |
239 BF(a3.re, a1.re, a1.re, t4);\ | |
240 BF(a2.im, a0.im, a0.im, t6);\ | |
241 } | |
242 | |
243 // force loading all the inputs before storing any. | |
244 // this is slightly slower for small data, but avoids store->load aliasing | |
245 // for addresses separated by large powers of 2. | |
246 #define BUTTERFLIES_BIG(a0,a1,a2,a3) {\ | |
247 FFTSample r0=a0.re, i0=a0.im, r1=a1.re, i1=a1.im;\ | |
248 BF(t3, t5, t5, t1);\ | |
249 BF(a2.re, a0.re, r0, t5);\ | |
250 BF(a3.im, a1.im, i1, t3);\ | |
251 BF(t4, t6, t2, t6);\ | |
252 BF(a3.re, a1.re, r1, t4);\ | |
253 BF(a2.im, a0.im, i0, t6);\ | |
254 } | |
255 | |
256 #define TRANSFORM(a0,a1,a2,a3,wre,wim) {\ | |
257 t1 = a2.re * wre + a2.im * wim;\ | |
258 t2 = a2.im * wre - a2.re * wim;\ | |
259 t5 = a3.re * wre - a3.im * wim;\ | |
260 t6 = a3.im * wre + a3.re * wim;\ | |
261 BUTTERFLIES(a0,a1,a2,a3)\ | |
262 } | |
263 | |
264 #define TRANSFORM_ZERO(a0,a1,a2,a3) {\ | |
265 t1 = a2.re;\ | |
266 t2 = a2.im;\ | |
267 t5 = a3.re;\ | |
268 t6 = a3.im;\ | |
269 BUTTERFLIES(a0,a1,a2,a3)\ | |
270 } | |
271 | |
272 /* z[0...8n-1], w[1...2n-1] */ | |
273 #define PASS(name)\ | |
274 static void name(FFTComplex *z, const FFTSample *wre, unsigned int n)\ | |
275 {\ | |
276 FFTSample t1, t2, t3, t4, t5, t6;\ | |
277 int o1 = 2*n;\ | |
278 int o2 = 4*n;\ | |
279 int o3 = 6*n;\ | |
280 const FFTSample *wim = wre+o1;\ | |
281 n--;\ | |
282 \ | |
283 TRANSFORM_ZERO(z[0],z[o1],z[o2],z[o3]);\ | |
284 TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1]);\ | |
285 do {\ | |
286 z += 2;\ | |
287 wre += 2;\ | |
288 wim -= 2;\ | |
289 TRANSFORM(z[0],z[o1],z[o2],z[o3],wre[0],wim[0]);\ | |
290 TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1]);\ | |
291 } while(--n);\ | |
781 | 292 } |
293 | |
7542 | 294 PASS(pass) |
295 #undef BUTTERFLIES | |
296 #define BUTTERFLIES BUTTERFLIES_BIG | |
297 PASS(pass_big) | |
298 | |
299 #define DECL_FFT(n,n2,n4)\ | |
300 static void fft##n(FFTComplex *z)\ | |
301 {\ | |
302 fft##n2(z);\ | |
303 fft##n4(z+n4*2);\ | |
304 fft##n4(z+n4*3);\ | |
305 pass(z,ff_cos_##n,n4/2);\ | |
306 } | |
307 | |
308 static void fft4(FFTComplex *z) | |
309 { | |
310 FFTSample t1, t2, t3, t4, t5, t6, t7, t8; | |
311 | |
312 BF(t3, t1, z[0].re, z[1].re); | |
313 BF(t8, t6, z[3].re, z[2].re); | |
314 BF(z[2].re, z[0].re, t1, t6); | |
315 BF(t4, t2, z[0].im, z[1].im); | |
316 BF(t7, t5, z[2].im, z[3].im); | |
317 BF(z[3].im, z[1].im, t4, t8); | |
318 BF(z[3].re, z[1].re, t3, t7); | |
319 BF(z[2].im, z[0].im, t2, t5); | |
320 } | |
321 | |
322 static void fft8(FFTComplex *z) | |
323 { | |
324 FFTSample t1, t2, t3, t4, t5, t6, t7, t8; | |
325 | |
326 fft4(z); | |
327 | |
328 BF(t1, z[5].re, z[4].re, -z[5].re); | |
329 BF(t2, z[5].im, z[4].im, -z[5].im); | |
330 BF(t3, z[7].re, z[6].re, -z[7].re); | |
331 BF(t4, z[7].im, z[6].im, -z[7].im); | |
332 BF(t8, t1, t3, t1); | |
333 BF(t7, t2, t2, t4); | |
334 BF(z[4].re, z[0].re, z[0].re, t1); | |
335 BF(z[4].im, z[0].im, z[0].im, t2); | |
336 BF(z[6].re, z[2].re, z[2].re, t7); | |
337 BF(z[6].im, z[2].im, z[2].im, t8); | |
338 | |
339 TRANSFORM(z[1],z[3],z[5],z[7],sqrthalf,sqrthalf); | |
340 } | |
341 | |
8590 | 342 #if !CONFIG_SMALL |
7542 | 343 static void fft16(FFTComplex *z) |
344 { | |
345 FFTSample t1, t2, t3, t4, t5, t6; | |
346 | |
347 fft8(z); | |
348 fft4(z+8); | |
349 fft4(z+12); | |
350 | |
351 TRANSFORM_ZERO(z[0],z[4],z[8],z[12]); | |
352 TRANSFORM(z[2],z[6],z[10],z[14],sqrthalf,sqrthalf); | |
353 TRANSFORM(z[1],z[5],z[9],z[13],ff_cos_16[1],ff_cos_16[3]); | |
354 TRANSFORM(z[3],z[7],z[11],z[15],ff_cos_16[3],ff_cos_16[1]); | |
355 } | |
356 #else | |
357 DECL_FFT(16,8,4) | |
358 #endif | |
359 DECL_FFT(32,16,8) | |
360 DECL_FFT(64,32,16) | |
361 DECL_FFT(128,64,32) | |
362 DECL_FFT(256,128,64) | |
363 DECL_FFT(512,256,128) | |
8590 | 364 #if !CONFIG_SMALL |
7542 | 365 #define pass pass_big |
366 #endif | |
367 DECL_FFT(1024,512,256) | |
368 DECL_FFT(2048,1024,512) | |
369 DECL_FFT(4096,2048,1024) | |
370 DECL_FFT(8192,4096,2048) | |
371 DECL_FFT(16384,8192,4096) | |
372 DECL_FFT(32768,16384,8192) | |
373 DECL_FFT(65536,32768,16384) | |
374 | |
375 static void (*fft_dispatch[])(FFTComplex*) = { | |
376 fft4, fft8, fft16, fft32, fft64, fft128, fft256, fft512, fft1024, | |
377 fft2048, fft4096, fft8192, fft16384, fft32768, fft65536, | |
378 }; | |
379 | |
380 void ff_fft_calc_c(FFTContext *s, FFTComplex *z) | |
381 { | |
382 fft_dispatch[s->nbits-2](z); | |
383 } | |
384 |