Mercurial > libavcodec.hg
annotate fft.c @ 10153:7a63015e4627 libavcodec
ARM: NEON optimised FFT and MDCT
Vorbis and AC3 ~3x faster.
Parts by Naotoshi Nojiri, naonoj gmail
author | mru |
---|---|
date | Thu, 10 Sep 2009 08:50:03 +0000 |
parents | c5e8a5a044c3 |
children | a349795e8dca |
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; | |
89 s->exptab1 = NULL; | |
781 | 90 |
8590 | 91 #if HAVE_MMX && HAVE_YASM |
6503 | 92 has_vectors = mm_support(); |
8981
dc19e4d7d0eb
Only enable SSE/3DNOW optimizations when they have been enabled at compilation.
diego
parents:
8974
diff
changeset
|
93 if (has_vectors & FF_MM_SSE && HAVE_SSE) { |
7542 | 94 /* SSE for P3/P4/K8 */ |
8974 | 95 s->imdct_calc = ff_imdct_calc_sse; |
96 s->imdct_half = ff_imdct_half_sse; | |
7542 | 97 s->fft_permute = ff_fft_permute_sse; |
8974 | 98 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
|
99 } else if (has_vectors & FF_MM_3DNOWEXT && HAVE_AMD3DNOWEXT) { |
7542 | 100 /* 3DNowEx for K7 */ |
6503 | 101 s->imdct_calc = ff_imdct_calc_3dn2; |
7263 | 102 s->imdct_half = ff_imdct_half_3dn2; |
8974 | 103 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
|
104 } else if (has_vectors & FF_MM_3DNOW && HAVE_AMD3DNOW) { |
6503 | 105 /* 3DNow! for K6-2/3 */ |
7544 | 106 s->imdct_calc = ff_imdct_calc_3dn; |
107 s->imdct_half = ff_imdct_half_3dn; | |
8974 | 108 s->fft_calc = ff_fft_calc_3dn; |
6503 | 109 } |
9177
4b1736ba9f2f
Remove long unused ALTIVEC_USE_REFERENCE_C_CODE ifdef; all other references
conrad
parents:
8981
diff
changeset
|
110 #elif HAVE_ALTIVEC |
6503 | 111 has_vectors = mm_support(); |
8104
0d108ec85620
Remove duplicated MM_* macros for CPU capabilities from dsputil.h.
rathann
parents:
7547
diff
changeset
|
112 if (has_vectors & FF_MM_ALTIVEC) { |
6503 | 113 s->fft_calc = ff_fft_calc_altivec; |
7542 | 114 split_radix = 0; |
6503 | 115 } |
10153 | 116 #elif HAVE_NEON |
117 s->fft_permute = ff_fft_permute_neon; | |
118 s->fft_calc = ff_fft_calc_neon; | |
119 s->imdct_calc = ff_imdct_calc_neon; | |
120 s->imdct_half = ff_imdct_half_neon; | |
121 revtab_shift = 3; | |
6503 | 122 #endif |
781 | 123 |
7542 | 124 if (split_radix) { |
125 for(j=4; j<=nbits; j++) { | |
126 int m = 1<<j; | |
127 double freq = 2*M_PI/m; | |
128 FFTSample *tab = ff_cos_tabs[j-4]; | |
129 for(i=0; i<=m/4; i++) | |
130 tab[i] = cos(i*freq); | |
131 for(i=1; i<m/4; i++) | |
132 tab[m/2-i] = tab[i]; | |
133 } | |
134 for(i=0; i<n; i++) | |
10153 | 135 s->revtab[-split_radix_permutation(i, n, s->inverse) & (n-1)] = |
136 i << revtab_shift; | |
7542 | 137 s->tmp_buf = av_malloc(n * sizeof(FFTComplex)); |
138 } else { | |
6504 | 139 int np, nblocks, np2, l; |
140 FFTComplex *q; | |
2967 | 141 |
7542 | 142 for(i=0; i<(n/2); i++) { |
143 alpha = 2 * M_PI * (float)i / (float)n; | |
144 c1 = cos(alpha); | |
145 s1 = sin(alpha) * s2; | |
146 s->exptab[i].re = c1; | |
147 s->exptab[i].im = s1; | |
148 } | |
149 | |
6504 | 150 np = 1 << nbits; |
151 nblocks = np >> 3; | |
152 np2 = np >> 1; | |
153 s->exptab1 = av_malloc(np * 2 * sizeof(FFTComplex)); | |
154 if (!s->exptab1) | |
155 goto fail; | |
156 q = s->exptab1; | |
157 do { | |
158 for(l = 0; l < np2; l += 2 * nblocks) { | |
159 *q++ = s->exptab[l]; | |
160 *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
|
161 |
6504 | 162 q->re = -s->exptab[l].im; |
163 q->im = s->exptab[l].re; | |
164 q++; | |
165 q->re = -s->exptab[l + nblocks].im; | |
166 q->im = s->exptab[l + nblocks].re; | |
167 q++; | |
168 } | |
169 nblocks = nblocks >> 1; | |
170 } while (nblocks != 0); | |
171 av_freep(&s->exptab); | |
781 | 172 |
7543 | 173 /* compute bit reverse table */ |
174 for(i=0;i<n;i++) { | |
175 m=0; | |
176 for(j=0;j<nbits;j++) { | |
177 m |= ((i >> j) & 1) << (nbits-j-1); | |
178 } | |
179 s->revtab[i]=m; | |
781 | 180 } |
7542 | 181 } |
182 | |
781 | 183 return 0; |
184 fail: | |
185 av_freep(&s->revtab); | |
186 av_freep(&s->exptab); | |
187 av_freep(&s->exptab1); | |
7542 | 188 av_freep(&s->tmp_buf); |
781 | 189 return -1; |
190 } | |
191 | |
7542 | 192 void ff_fft_permute_c(FFTContext *s, FFTComplex *z) |
781 | 193 { |
194 int j, k, np; | |
195 FFTComplex tmp; | |
196 const uint16_t *revtab = s->revtab; | |
7542 | 197 np = 1 << s->nbits; |
198 | |
199 if (s->tmp_buf) { | |
200 /* TODO: handle split-radix permute in a more optimal way, probably in-place */ | |
201 for(j=0;j<np;j++) s->tmp_buf[revtab[j]] = z[j]; | |
202 memcpy(z, s->tmp_buf, np * sizeof(FFTComplex)); | |
203 return; | |
204 } | |
2967 | 205 |
781 | 206 /* reverse */ |
207 for(j=0;j<np;j++) { | |
208 k = revtab[j]; | |
209 if (k < j) { | |
210 tmp = z[k]; | |
211 z[k] = z[j]; | |
212 z[j] = tmp; | |
213 } | |
214 } | |
215 } | |
216 | |
8687 | 217 av_cold void ff_fft_end(FFTContext *s) |
781 | 218 { |
219 av_freep(&s->revtab); | |
220 av_freep(&s->exptab); | |
221 av_freep(&s->exptab1); | |
7542 | 222 av_freep(&s->tmp_buf); |
223 } | |
224 | |
225 #define sqrthalf (float)M_SQRT1_2 | |
226 | |
227 #define BF(x,y,a,b) {\ | |
228 x = a - b;\ | |
229 y = a + b;\ | |
230 } | |
231 | |
232 #define BUTTERFLIES(a0,a1,a2,a3) {\ | |
233 BF(t3, t5, t5, t1);\ | |
234 BF(a2.re, a0.re, a0.re, t5);\ | |
235 BF(a3.im, a1.im, a1.im, t3);\ | |
236 BF(t4, t6, t2, t6);\ | |
237 BF(a3.re, a1.re, a1.re, t4);\ | |
238 BF(a2.im, a0.im, a0.im, t6);\ | |
239 } | |
240 | |
241 // force loading all the inputs before storing any. | |
242 // this is slightly slower for small data, but avoids store->load aliasing | |
243 // for addresses separated by large powers of 2. | |
244 #define BUTTERFLIES_BIG(a0,a1,a2,a3) {\ | |
245 FFTSample r0=a0.re, i0=a0.im, r1=a1.re, i1=a1.im;\ | |
246 BF(t3, t5, t5, t1);\ | |
247 BF(a2.re, a0.re, r0, t5);\ | |
248 BF(a3.im, a1.im, i1, t3);\ | |
249 BF(t4, t6, t2, t6);\ | |
250 BF(a3.re, a1.re, r1, t4);\ | |
251 BF(a2.im, a0.im, i0, t6);\ | |
252 } | |
253 | |
254 #define TRANSFORM(a0,a1,a2,a3,wre,wim) {\ | |
255 t1 = a2.re * wre + a2.im * wim;\ | |
256 t2 = a2.im * wre - a2.re * wim;\ | |
257 t5 = a3.re * wre - a3.im * wim;\ | |
258 t6 = a3.im * wre + a3.re * wim;\ | |
259 BUTTERFLIES(a0,a1,a2,a3)\ | |
260 } | |
261 | |
262 #define TRANSFORM_ZERO(a0,a1,a2,a3) {\ | |
263 t1 = a2.re;\ | |
264 t2 = a2.im;\ | |
265 t5 = a3.re;\ | |
266 t6 = a3.im;\ | |
267 BUTTERFLIES(a0,a1,a2,a3)\ | |
268 } | |
269 | |
270 /* z[0...8n-1], w[1...2n-1] */ | |
271 #define PASS(name)\ | |
272 static void name(FFTComplex *z, const FFTSample *wre, unsigned int n)\ | |
273 {\ | |
274 FFTSample t1, t2, t3, t4, t5, t6;\ | |
275 int o1 = 2*n;\ | |
276 int o2 = 4*n;\ | |
277 int o3 = 6*n;\ | |
278 const FFTSample *wim = wre+o1;\ | |
279 n--;\ | |
280 \ | |
281 TRANSFORM_ZERO(z[0],z[o1],z[o2],z[o3]);\ | |
282 TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1]);\ | |
283 do {\ | |
284 z += 2;\ | |
285 wre += 2;\ | |
286 wim -= 2;\ | |
287 TRANSFORM(z[0],z[o1],z[o2],z[o3],wre[0],wim[0]);\ | |
288 TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1]);\ | |
289 } while(--n);\ | |
781 | 290 } |
291 | |
7542 | 292 PASS(pass) |
293 #undef BUTTERFLIES | |
294 #define BUTTERFLIES BUTTERFLIES_BIG | |
295 PASS(pass_big) | |
296 | |
297 #define DECL_FFT(n,n2,n4)\ | |
298 static void fft##n(FFTComplex *z)\ | |
299 {\ | |
300 fft##n2(z);\ | |
301 fft##n4(z+n4*2);\ | |
302 fft##n4(z+n4*3);\ | |
303 pass(z,ff_cos_##n,n4/2);\ | |
304 } | |
305 | |
306 static void fft4(FFTComplex *z) | |
307 { | |
308 FFTSample t1, t2, t3, t4, t5, t6, t7, t8; | |
309 | |
310 BF(t3, t1, z[0].re, z[1].re); | |
311 BF(t8, t6, z[3].re, z[2].re); | |
312 BF(z[2].re, z[0].re, t1, t6); | |
313 BF(t4, t2, z[0].im, z[1].im); | |
314 BF(t7, t5, z[2].im, z[3].im); | |
315 BF(z[3].im, z[1].im, t4, t8); | |
316 BF(z[3].re, z[1].re, t3, t7); | |
317 BF(z[2].im, z[0].im, t2, t5); | |
318 } | |
319 | |
320 static void fft8(FFTComplex *z) | |
321 { | |
322 FFTSample t1, t2, t3, t4, t5, t6, t7, t8; | |
323 | |
324 fft4(z); | |
325 | |
326 BF(t1, z[5].re, z[4].re, -z[5].re); | |
327 BF(t2, z[5].im, z[4].im, -z[5].im); | |
328 BF(t3, z[7].re, z[6].re, -z[7].re); | |
329 BF(t4, z[7].im, z[6].im, -z[7].im); | |
330 BF(t8, t1, t3, t1); | |
331 BF(t7, t2, t2, t4); | |
332 BF(z[4].re, z[0].re, z[0].re, t1); | |
333 BF(z[4].im, z[0].im, z[0].im, t2); | |
334 BF(z[6].re, z[2].re, z[2].re, t7); | |
335 BF(z[6].im, z[2].im, z[2].im, t8); | |
336 | |
337 TRANSFORM(z[1],z[3],z[5],z[7],sqrthalf,sqrthalf); | |
338 } | |
339 | |
8590 | 340 #if !CONFIG_SMALL |
7542 | 341 static void fft16(FFTComplex *z) |
342 { | |
343 FFTSample t1, t2, t3, t4, t5, t6; | |
344 | |
345 fft8(z); | |
346 fft4(z+8); | |
347 fft4(z+12); | |
348 | |
349 TRANSFORM_ZERO(z[0],z[4],z[8],z[12]); | |
350 TRANSFORM(z[2],z[6],z[10],z[14],sqrthalf,sqrthalf); | |
351 TRANSFORM(z[1],z[5],z[9],z[13],ff_cos_16[1],ff_cos_16[3]); | |
352 TRANSFORM(z[3],z[7],z[11],z[15],ff_cos_16[3],ff_cos_16[1]); | |
353 } | |
354 #else | |
355 DECL_FFT(16,8,4) | |
356 #endif | |
357 DECL_FFT(32,16,8) | |
358 DECL_FFT(64,32,16) | |
359 DECL_FFT(128,64,32) | |
360 DECL_FFT(256,128,64) | |
361 DECL_FFT(512,256,128) | |
8590 | 362 #if !CONFIG_SMALL |
7542 | 363 #define pass pass_big |
364 #endif | |
365 DECL_FFT(1024,512,256) | |
366 DECL_FFT(2048,1024,512) | |
367 DECL_FFT(4096,2048,1024) | |
368 DECL_FFT(8192,4096,2048) | |
369 DECL_FFT(16384,8192,4096) | |
370 DECL_FFT(32768,16384,8192) | |
371 DECL_FFT(65536,32768,16384) | |
372 | |
373 static void (*fft_dispatch[])(FFTComplex*) = { | |
374 fft4, fft8, fft16, fft32, fft64, fft128, fft256, fft512, fft1024, | |
375 fft2048, fft4096, fft8192, fft16384, fft32768, fft65536, | |
376 }; | |
377 | |
378 void ff_fft_calc_c(FFTContext *s, FFTComplex *z) | |
379 { | |
380 fft_dispatch[s->nbits-2](z); | |
381 } | |
382 |